Stochastic differential equations · Part 3 of 10

Building the Itô integral

· 4 min read

Many models of random dynamics have the form “rate of change equals drift plus noise”. Written as an integral equation, such a model reads

Xt=X0+∫0tb(Xs) ds+∫0tσ(Xs) dBs,X_t = X_0 + \int_0^t b(X_s)\,ds + \int_0^t \sigma(X_s)\,dB_s,

and everything hinges on the last term. This post defines it.

Why the obvious definition fails

For a smooth integrator AA, the integral ∫0tHs dAs\int_0^t H_s\,dA_s is a limit of Riemann–Stieltjes sums ∑iHτi(Ati+1−Ati)\sum_i H_{\tau_i}(A_{t_{i+1}} - A_{t_i}), and it does not matter where in [ti,ti+1][t_i, t_{i+1}] the point τi\tau_i is chosen. Brownian motion has infinite variation, and suddenly the choice matters. Compare left and right endpoints in sums for ∫B dB\int B\,dB:

∑iBti+1(Bti+1−Bti)−∑iBti(Bti+1−Bti)=∑i(Bti+1−Bti)2⟶t.\sum_i B_{t_{i+1}}(B_{t_{i+1}} - B_{t_i}) - \sum_i B_{t_i}(B_{t_{i+1}} - B_{t_i}) = \sum_i (B_{t_{i+1}} - B_{t_i})^2 \longrightarrow t.

The two kinds of sums converge to limits that differ by tt. Itô’s choice is the left endpoint. It is the natural one for modeling, because the integrand at each step is fixed before the next increment of noise arrives, the way a trading strategy has to be chosen before prices move.

Simple adapted integrands

Fix a filtration (Ft)(\mathcal{F}_t) to which BB is adapted, with Bt−BsB_t - B_s independent of Fs\mathcal{F}_s for s<ts < t. A simple adapted process is a step function

Ht=∑i=0n−1hi 1(ti,ti+1](t),H_t = \sum_{i=0}^{n-1} h_i\, \mathbf{1}_{(t_i, t_{i+1}]}(t),

where each hih_i is a bounded random variable that is Fti\mathcal{F}_{t_i}-measurable, that is, known at the start of its step. For such processes there is only one reasonable definition:

∫0THt dBt=∑i=0n−1hi (Bti+1−Bti).\int_0^T H_t\,dB_t = \sum_{i=0}^{n-1} h_i\,\big(B_{t_{i+1}} - B_{t_i}\big).

The Itô isometry

The key identity is that this map preserves the L2L^2 norm:

E[(∫0THt dBt)2]=E∫0THt2 dt.\mathbb{E}\left[\left(\int_0^T H_t\,dB_t\right)^2\right] = \mathbb{E}\int_0^T H_t^2\,dt.

To see why, write Δi=Bti+1−Bti\Delta_i = B_{t_{i+1}} - B_{t_i} and expand the square. For i<ji < j, the increment Δj\Delta_j is independent of hihjΔih_i h_j \Delta_i, which is known at time tjt_j, so the cross term E[hihjΔiΔj]\mathbb{E}[h_i h_j \Delta_i \Delta_j] is zero. For a diagonal term, Δi\Delta_i is independent of hih_i, so E[hi2Δi2]=E[hi2] (ti+1−ti)\mathbb{E}[h_i^2 \Delta_i^2] = \mathbb{E}[h_i^2]\,(t_{i+1} - t_i). Summing gives exactly E∫0THt2 dt\mathbb{E}\int_0^T H_t^2\,dt. Adaptedness is what makes every step of this argument work.

Extending by approximation

Now let HH be any adapted process with E∫0THt2 dt<∞\mathbb{E}\int_0^T H_t^2\,dt < \infty (and a mild measurability condition). Such processes can be approximated by simple adapted processes H(n)H^{(n)}, in the sense that E∫0T(Ht−Ht(n))2 dt→0\mathbb{E}\int_0^T (H_t - H^{(n)}_t)^2\,dt \to 0. By the isometry, the integrals ∫H(n) dB\int H^{(n)}\,dB form a Cauchy sequence in L2(P)L^2(\mathbb{P}), so they converge, and the limit does not depend on the approximating sequence. That limit is the Itô integral of HH.

Viewed as a process in its upper limit, It=∫0tHs dBsI_t = \int_0^t H_s\,dB_s inherits three properties from the simple case:

  • it has a version with continuous paths;
  • it is a martingale with mean zero;
  • its variance is E[It2]=E∫0tHs2 ds\mathbb{E}[I_t^2] = \mathbb{E}\int_0^t H_s^2\,ds.

If HH only satisfies ∫0tHs2 ds<∞\int_0^t H_s^2\,ds < \infty almost surely, a localization argument (stopping before the integral gets too large) still defines ItI_t, but then it may only be a local martingale. The difference matters, and it comes back in the post on martingales.

Quadratic variation and covariation

For a continuous process XX, the quadratic variation ⟨X⟩t\langle X \rangle_t is the limit in probability of ∑i(Xti+1−Xti)2\sum_i (X_{t_{i+1}} - X_{t_i})^2 as the mesh of a partition of [0,t][0, t] goes to zero. The previous post showed that ⟨B⟩t=t\langle B \rangle_t = t. For an Itô process

Xt=X0+∫0tbs ds+∫0tσs dBs,X_t = X_0 + \int_0^t b_s\,ds + \int_0^t \sigma_s\,dB_s,

the drift contributes nothing, because it has finite variation, and

⟨X⟩t=∫0tσs2 ds.\langle X \rangle_t = \int_0^t \sigma_s^2\,ds.

The covariation of two processes is defined by polarization, ⟨X,Y⟩=14(⟨X+Y⟩−⟨X−Y⟩)\langle X, Y \rangle = \tfrac{1}{4}\big(\langle X + Y \rangle - \langle X - Y \rangle\big). In differential shorthand, all of this is captured by a multiplication table:

dBt⋅dBt=dt,dBt⋅dt=0,dt⋅dt=0,dB_t \cdot dB_t = dt, \qquad dB_t \cdot dt = 0, \qquad dt \cdot dt = 0,

and dBt⋅dWt=0dB_t \cdot dW_t = 0 for independent Brownian motions BB and WW. These rules are exactly what Itô’s formula needs.

The example that explains everything

Return to ∫0tBs dBs\int_0^t B_s\,dB_s. Using left-endpoint sums and the algebraic identity a(b−a)=12(b2−a2)−12(b−a)2a(b - a) = \tfrac{1}{2}(b^2 - a^2) - \tfrac{1}{2}(b - a)^2,

∑iBti(Bti+1−Bti)=12Bt2−12∑i(Bti+1−Bti)2⟶12Bt2−12t.\sum_i B_{t_i}(B_{t_{i+1}} - B_{t_i}) = \frac{1}{2}B_t^2 - \frac{1}{2}\sum_i (B_{t_{i+1}} - B_{t_i})^2 \longrightarrow \frac{1}{2}B_t^2 - \frac{1}{2}t.

So

∫0tBs dBs=Bt2−t2,\int_0^t B_s\,dB_s = \frac{B_t^2 - t}{2},

not 12Bt2\tfrac{1}{2}B_t^2 as ordinary calculus would suggest. The correction −t/2-t/2 is exactly what keeps the right side a martingale with mean zero. Evaluating at midpoints instead gives the Stratonovich integral, which obeys the ordinary chain rule but is not a martingale. The next post turns this example into a general rule.

← All posts in Stochastic differential equations