Deriving the Hessian-Covariance Correction

From Covariance to the Second-Order Term

In one dimension, the second-order term in Itô’s formula is one half of the second derivative multiplied by the instantaneous variance. In several dimensions, the same principle applies to every pair of coordinates: the Hessian records how the function curves in each pair of directions, while the covariance records how the process fluctuates jointly in those directions. Their contraction is the Hessian-covariance correction.

Work on a filtered probability space \((\Omega,\mathcal{F},(\mathcal{F}_t)_{t\ge0},\mathbb{P})\) carrying an \(m\)-dimensional standard Brownian motion \(\mathbf{B}_t=(B_t^1,\ldots,B_t^m)\). Each component is Brownian motion relative to the filtration, and the components have independent increments and zero mutual covariation. Let \(X_t\) take values in \(\mathbb{R}^n\) and satisfy

Here \(b_t\) is the drift vector and \(A_t\) is the diffusion matrix. Assume the coefficients are progressively measurable, \(\int_0^T |b_t|\,dt<\infty\) almost surely, and each diffusion entry \(A_{ik}\) is in \(\mathcal{L}^2\) on \([0,T]\), meaning \(\mathbb{E}\int_0^T A_{ik}(t)^2\,dt<\infty\). Thus every displayed stochastic integral is an Itô integral with an adapted, progressively measurable, square-integrable integrand. For a \(C^2\) function \(f:\mathbb{R}^n\to\mathbb{R}\), the Hessian is the matrix \(D^2f=(f_{ij})\), where \(f_{ij}=\partial^2f/\partial x_i\partial x_j\).

The drift contributes first-order change, \(\sum_i f_i b^i\,dt\). The correction comes from the quadratic part of the local change in \(f\). To see its structure, write the second-order Taylor term for a small vector increment \(\Delta X\) as \(\tfrac12\sum_{i,j}f_{ij}\Delta X_i\Delta X_j\). On the stochastic time scale, the products of diffusion increments contribute at order \(dt\); drift-increment products and products of two drift increments do not. The multiplication rules are \(dB_t^k\,dB_t^\ell=\mathbf{1}_{\{k=\ell\}}dt\), \(dB_t^k\,dt=0\), and \(dt\,dt=0\).

Why the Trace Formula Has This Form

For each pair of coordinates, the diffusion part of \(dX^i\,dX^j\) is

Thus the quadratic Taylor term contributes \(\tfrac12\sum_{i,j}f_{ij}C_{ij}\,dt\). The following algebraic identity explains why this is the trace expression and why mixed partials appear twice in the full sum.

Proof. By the definition of matrix multiplication, \(\operatorname{tr}(CH)=\sum_{i=1}^n(CH)_{ii}=\sum_{i=1}^n\sum_{j=1}^n C_{ij}H_{ji}\). Since \(H\) is symmetric, \(H_{ji}=H_{ij}\), giving the first equality. For the second, \(C_{ij}=\sum_k A_{ik}A_{jk}\), so

All sums are finite, so their order can be interchanged. The identities also hold when \(A\), \(C\), or \(H\) is zero; no invertibility assumption is needed. \(\square\)

For a Hessian, \(H\) is symmetric when \(f\) is \(C^2\). Expanding the sum by diagonal and off-diagonal entries gives the coordinate form

The factor one half remains on each diagonal term. For an off-diagonal pair, the full double sum contains both \(C_{ij}f_{ij}\) and \(C_{ji}f_{ji}\). Symmetry makes these equal, and the factor one half turns their sum into one copy. In particular, a mixed derivative is not omitted merely because it is not on the diagonal.

The multidimensional Itô formula for scalar functions, stated here in the notation needed for the correction, is

The middle term is a sum of Itô integrals. On any interval where its integrands satisfy the \(\mathcal{L}^2\) condition, those integrals are defined in the usual square-integrable sense. The formula above identifies the correction; the full partition and remainder argument for the multidimensional formula is a separate proof. Its second-order content is precisely the quadratic-covariation computation and matrix identity given here.

Correlated Brownian Drivers

The matrix \(A\) need not itself display the correlations in a model. Suppose instead \(dX=b\,dt+\Sigma\,d\mathbf{W}\), where the components of \(\mathbf{W}\) have instantaneous correlation matrix \(R\): \(d[W^k,W^\ell]_t=R_{k\ell}\,dt\). Then the coordinate covariation is \(d[X^i,X^j]_t=(\Sigma R\Sigma^{\mathsf T})_{ij}\,dt\). Consequently, the correction uses \(C=\Sigma R\Sigma^{\mathsf T}\), not merely \(\Sigma\Sigma^{\mathsf T}\) unless \(R\) is the identity matrix. This statement concerns covariation; one may equivalently represent the correlated drivers using independent Brownian components and a matrix factor of \(R\).

A Common Counting Error

A frequent mistake is to write the trace as a sum over \(i\le j\) while retaining the factor one half on every term. In the correlated-product example, that incorrect calculation would keep only \(\tfrac12 C_{12}f_{12}=\tfrac{3}{10}\) for the off-diagonal contribution. It misses the equal term \(\tfrac12 C_{21}f_{21}=\tfrac{3}{10}\). The correct mixed contribution is their sum, \(\tfrac35\), as confirmed directly by \(dX^1dX^2=\tfrac35dt\). When using a sum over unordered pairs, the correct coordinate form is one full \(C_{ij}f_{ij}\) for each \(i<j\), alongside the half-weighted diagonal terms.

The correction depends on both curvature and covariance. If \(C_{ij}=0\), that coordinate pair contributes no mixed term, even if \(f_{ij}\ne0\). If the covariance is nonzero but \(f_{ij}=0\), it likewise contributes nothing. The matrix form packages these cases without requiring independent coordinates: it is the diffusion covariance, rather than independence of the coordinates of \(X\), that determines the second-order term.