Lévy Areas in Noncommutative Milstein Schemes

Why a Second Brownian Increment Is Not Enough

For an SDE driven by several Brownian motions, the first terms in a Milstein update use the Brownian increments over a time step. If the diffusion fields commute, the cross terms can be written using products of those increments. When they do not commute, an additional random quantity remains: the signed area traced by the pair of Brownian coordinates during the step. This Lévy area is not determined by the two endpoint increments.

Work on a filtered probability space \((\Omega,\mathcal{F},(\mathcal{F}_t)_{t\ge0},\mathbb{P})\) carrying a two-dimensional Brownian motion \(\mathbf{B}_t=(B_t^1,B_t^2)\), whose components are independent Brownian motions relative to the filtration. Every Ito integral below is with respect to this filtration and has a progressively measurable integrand \(H_s\) satisfying \(\mathbb{E}\int_0^T H_s^2\,ds<\infty\) on the interval of integration. These are the square-integrable Ito integrands, denoted by \(\mathcal{L}^2\).

Fix a step \([t,t+h]\), where \(h>0\), and write \(\Delta B^j=B^j_{t+h}-B^j_t\). For \(j,k\in\{1,2\}\), define the iterated Ito integral with inner index \(j\) and outer index \(k\) by

The inner integrand \(B_s^j-B_t^j\) is adapted and progressively measurable, and its expected squared integral is \(h^2/2\), so it belongs to \(\mathcal{L}^2\). The two cross-integrals \(I_{12}\) and \(I_{21}\) can be separated into a symmetric part, determined by the endpoint increments, and an antisymmetric part, the Lévy area.

The Cross-Integral Identity

The identities in the definition follow from the Ito product rule. Here the rule can be obtained directly from Ito's formula for \(f(x,y)=xy\). Its partial derivatives are \(f_x=y\), \(f_y=x\), \(f_{xx}=f_{yy}=0\), and \(f_{xy}=1\). The second-order term is therefore the cross-variation term \(d[B^1,B^2]_s\), which is zero for independent Brownian components. Thus \(d(B^1_sB^2_s)=B^1_s\,dB^2_s+B^2_s\,dB^1_s\). Applying this to the increments from time \(t\) gives the result:

Proof. Apply the stated Ito formula to \(f(x,y)=xy\) with \(x=B_s^1-B_t^1\) and \(y=B_s^2-B_t^2\). The first-order terms are \( (B_s^1-B_t^1)dB_s^2+(B_s^2-B_t^2)dB_s^1\), and the mixed second-order term is \(d[B^1,B^2]_s=0\). Integrating from \(t\) to \(t+h\) yields \(\Delta B^1\Delta B^2=I_{12}+I_{21}\). Combining this sum with the definition \(A_{t,t+h}=\tfrac12(I_{12}-I_{21})\) gives the two expressions. \(\square\)

The identity concerns two different quantities: the product of endpoint increments gives only the sum of the iterated integrals, not either integral separately. The area measures the missing antisymmetric part. Since the Brownian components have zero cross-variation, the cross-integral \(\int_t^{t+h}(B_s^1-B_t^1)\circ dB_s^2\), if written in Stratonovich form, equals its Ito version: the Ito-to-Stratonovich correction is half the cross-variation of the integrand with \(B^2\), which is zero. The Milstein formulas here use Ito integrals.

How the Area Enters the Milstein Update

Consider a \(d\)-dimensional state \(X_t\) satisfying the two-noise Ito SDE \(dX_t=\mu(X_t)\,dt+\sigma_1(X_t)\,dB_t^1+\sigma_2(X_t)\,dB_t^2\). Here \(\mu\) is the drift and \(\sigma_j\) is the diffusion coefficient for noise \(j\), each taking values in \(\mathbb{R}^d\). Assume the coefficients have the differentiability and growth needed for the Ito expansion below; for a standard global Milstein convergence result, one imposes global Lipschitz and linear growth conditions together with sufficiently bounded derivatives. The area algebra itself does not require those global conditions.

For a differentiable vector field \(g:\mathbb{R}^d\to\mathbb{R}^d\), write \(Dg(x)v\) for its derivative at \(x\) applied to \(v\). In the diffusion part of the Ito-Taylor expansion, the first variation of \(\sigma_k(X_s)\) along noise \(j\) is \(D\sigma_k(X_t)\sigma_j(X_t)\,dB^j\). Integrating that variation against \(dB^k\) produces the coefficient \(D\sigma_k\,\sigma_j\) multiplying \(I_{jk}\). Thus the diffusion terms of the Milstein update are

The diagonal integrals are determined by increments: Ito's formula for \(x^2\), with \(f'(x)=2x\), \(f''(x)=2\), and correction \(\tfrac12 f''dt=dt\), gives \(I_{jj}=\tfrac12((\Delta B^j)^2-h)\). For the cross terms, set \(C_1=D\sigma_2\,\sigma_1\) and \(C_2=D\sigma_1\,\sigma_2\), evaluated at \(X_t\). Substitution of the area identities yields

The area over a step has size of order \(h\), just like products of Brownian increments. It cannot generally be discarded while retaining the full Milstein diffusion correction. In simulation, using only \(\Delta B^1\) and \(\Delta B^2\) does not specify the noncommutative update: the iterated cross-integrals, or equivalently their signed area, must also be supplied or approximated.

Mean and Variance of the Area

The area is a random variable beyond the endpoint increments. Its basic moments follow from the Ito isometry. For this calculation, each iterated integral is a square-integrable martingale increment: its integrand is progressively measurable and square-integrable, and the Ito integral has conditional mean zero given \(\mathcal{F}_t\). In particular, the area has conditional mean zero given the information at the start of the step.

Proof. Brownian increments after time \(t\) are independent of \(\mathcal{F}_t\). For \(I_{12}\), the Ito isometry gives conditional second moment \(\int_t^{t+h}\mathbb{E}[(B_s^1-B_t^1)^2\mid\mathcal{F}_t]\,ds=\int_t^{t+h}(s-t)\,ds=h^2/2\). The same computation gives conditional second moment \(h^2/2\) for \(I_{21}\). Their conditional covariance is zero: the two outer integrators are different Brownian components, whose cross-variation is zero, so the cross-isometry gives \(\mathbb{E}[I_{12}I_{21}\mid\mathcal{F}_t]=0\). Both integrals have conditional mean zero. Consequently, for \(A_{t,t+h}=\tfrac12(I_{12}-I_{21})\), its conditional mean is zero and its conditional second moment is \(\tfrac14(h^2/2+h^2/2)=h^2/4\). These finite second moments also verify integrability. \(\square\)

Common Pitfall: Replacing Each Cross-Integral by Half the Product

A tempting but incorrect step is to set \(I_{12}=I_{21}=\tfrac12\Delta B^1\Delta B^2\). The product identity proves only that their sum is \(\Delta B^1\Delta B^2\); equality of the two integrals would require the signed area to be zero. For the concrete values \(\Delta B^1=\tfrac12\), \(\Delta B^2=-\tfrac13\), and \(A_{t,t+h}=\tfrac1{10}\), the incorrect calculation gives \(I_{12}=-\tfrac1{12}\). The correct value is \(I_{12}=-\tfrac1{12}+\tfrac1{10}=\tfrac1{60}\). In the worked SDE, this error would give the second coordinate \(3-\tfrac23-\tfrac1{12}=\tfrac{27}{12}=\tfrac94\), rather than the checked value \(\tfrac{47}{20}\).

The issue is not a small algebraic adjustment: when diffusion fields do not commute, the area is multiplied by a nonzero vector-field bracket. Ignoring it changes the step itself. In the commuting case the coefficient of the area vanishes, which is why the endpoint increments then suffice for the cross correction.