Diagnosing Missing Corrections in Ito Calculations

A Correction Ledger for Ito Calculations

Many errors in Ito calculations are not errors of differentiation. They arise because one derivative term, one quadratic covariation, or one correlation coefficient is lost when the formula is assembled. A reliable check is to keep separate accounts for the time derivative, the drift, the diffusion, and every second-order contribution. This tutorial develops that check for scalar and multidimensional processes.

Work on a filtered probability space $(\Omega,\mathcal{F},(\mathcal{F}_t)_{t\ge0},\mathbb{P})$ carrying a Brownian motion $B_t$, or a vector Brownian motion $\mathbf{B}_t$ when there are several drivers. For a fixed finite horizon $T$, an Ito integral $\int_0^t H_s\,dB_s$ below has a progressively measurable integrand $H$ in $\mathcal{L}^2$, meaning $\mathbb{E}\int_0^T H_s^2\,ds<\infty$. For a vector integral, each component integrand satisfies this condition. The Brownian components may be correlated; their instantaneous covariations must be included in the calculation.

Definition: For $dX_t=\mu_t\,dt+\sigma_t\,dB_t$, the drift coefficient is $\mu_t$ and the diffusion coefficient is $\sigma_t$. If $f(t,x)$ is once continuously differentiable in time and twice continuously differentiable in space, Ito's formula gives $$ df(t,X_t)=\left(f_t+\mu_t f_x+\tfrac12\sigma_t^2f_{xx}\right)(t,X_t)\,dt+f_x(t,X_t)\sigma_t\,dB_t. $$ The term $\tfrac12\sigma_t^2 f_{xx}\,dt$ is the scalar squared-diffusion correction. The formula uses $dB_t\,dB_t=dt$, $dB_t\,dt=0$, and $dt\,dt=0$; these are Ito multiplication rules, not ordinary differential identities.

In more than one dimension, write $dX_t^i=b_t^i\,dt+\sum_{k=1}^m \sigma_t^{ik}\,dB_t^k$, for $i=1,\ldots,d$. Suppose $d[B^k,B^\ell]_t=R_{k\ell}\,dt$, where $R$ is the Brownian drivers' correlation matrix. The state covariance matrix per unit time is $A_t=\sigma_t R\sigma_t^\mathsf{T}$, so $d[X^i,X^j]_t=(A_t)_{ij}\,dt$. The multidimensional Ito formula for a scalar function $f(t,x)$ is

df(t,\mathbf{X}_t)=\left(f_t+\sum_{i=1}^d b_t^i f_{x_i} +\tfrac12\sum_{i=1}^d\sum_{j=1}^d(A_t)_{ij}f_{x_ix_j}\right)dt +\sum_{k=1}^m\left(\sum_{i=1}^d f_{x_i}\sigma_t^{ik}\right)dB_t^k.

Here and below derivatives are evaluated at $(t,\mathbf{X}_t)$. The displayed formulas are the standard scalar and multidimensional Ito formulas, established in the earlier material on Ito's formula. Whenever an Ito integral is displayed, assume its integrand is progressively measurable and in $\mathcal{L}^2$ on the finite time horizon, or that the calculation has been stopped so this condition holds. The process $f(t,\mathbf{X}_t)$ must also be defined on the region visited by $\mathbf{X}$.

Two Audits That Isolate Missing Terms

Proposition: Under the scalar formula's hypotheses, omitting the squared-diffusion correction changes the proposed drift by $\tfrac12\sigma_t^2 f_{xx}(t,X_t)$. The correction vanishes at a point precisely when $\sigma_t=0$ or $f_{xx}(t,X_t)=0$.

Proof. The full drift from the scalar Ito formula is $f_t+\mu_t f_x+\tfrac12\sigma_t^2 f_{xx}$. The drift obtained by deleting the second-order term is $f_t+\mu_t f_x$. Subtracting the latter from the former gives $\tfrac12\sigma_t^2 f_{xx}$. Since a product of real numbers is zero exactly when at least one factor is zero, this correction vanishes exactly when $\sigma_t^2=0$ or $f_{xx}=0$, equivalently when $\sigma_t=0$ or $f_{xx}=0$. This is a pointwise statement; a correction can also integrate to zero through cancellation without being zero at every time. $\square$

Proposition: In the multidimensional formula, the two off-diagonal terms associated with coordinates $i\ne j$ combine to $(A_t)_{ij}f_{x_ix_j}\,dt$. In particular, if $A_t=\sigma_tR\sigma_t^\mathsf{T}$, mixed corrections depend on the Brownian correlation coefficients in $R$.

Proof. The second-order contribution is $\tfrac12\sum_{i,j}(A_t)_{ij}f_{x_ix_j}\,dt$. The covariance matrix $A_t$ is symmetric because $R$ is symmetric, and the Hessian is symmetric because $f$ has continuous second derivatives. For a fixed pair $i\ne j$, the summands indexed by $(i,j)$ and $(j,i)$ are therefore equal: $(A_t)_{ij}f_{x_ix_j}=(A_t)_{ji}f_{x_jx_i}$. Their sum, including the factor $\tfrac12$, is $(A_t)_{ij}f_{x_ix_j}\,dt$. Since $A_t=\sigma_tR\sigma_t^\mathsf{T}$, its entries are $(A_t)_{ij}=\sum_{k,\ell}\sigma_t^{ik}R_{k\ell}\sigma_t^{j\ell}$; thus changing a correlation coefficient in $R$ can change this mixed correction. If $i=j$, there is only the diagonal contribution $\tfrac12(A_t)_{ii}f_{x_ix_i}\,dt$. $\square$

These propositions give a compact audit: first compare each proposed drift term with the formula, then expand the covariance matrix before simplifying mixed derivatives. A zero mixed derivative or zero covariance removes a particular mixed term; it does not justify deleting the other second-order terms.

Worked Scalar Checks

Worked Example: The Missing Squared Diffusion

Let $X_t=t+2B_t$, and set $f(t,x)=x^2+3t$. Find the Ito differential and compare it with the result from omitting the squared-diffusion correction.

Here $\mu_t=1$, $\sigma_t=2$, $f_t=3$, $f_x=2x$, and $f_{xx}=2$. Substituting each quantity into the scalar formula gives drift $3+1(2X_t)+\tfrac12(2^2)(2)=3+2X_t+4=7+2X_t$, and diffusion coefficient $2X_t\cdot2=4X_t$. Thus

d(X_t^2+3t)=(7+2X_t)\,dt+4X_t\,dB_t.

Deleting the correction would instead give $(3+2X_t)\,dt+4X_t\,dB_t$, missing $4\,dt$. The stochastic integrand $4X_t$ is progressively measurable, and $\mathbb{E}\int_0^T(4X_s)^2\,ds=16\int_0^T(s^2+4s)\,ds<\infty$, since $X_s=s+2B_s$ and $\mathbb{E}[X_s^2]=s^2+4s$.

Check: Directly, $d(X_t^2)=2X_t\,dX_t+d[X]_t$, where $d[X]_t=2^2dt=4dt$. Adding $d(3t)=3dt$ gives $2X_t(dt+2dB_t)+4dt+3dt=(7+2X_t)dt+4X_tdB_t$, as above. Also $\mathbb{E}[X_t^2+3t]=t^2+4t+3t=t^2+7t$; the drift's expectation integrates to $\int_0^t(7+2s)\,ds=7t+t^2$, independently confirming the correction.

Worked Example: A Mixed Term with Correlated Drivers

Let $B^1,B^2$ be Brownian drivers with $d[B^1,B^2]_t=\tfrac12dt$, and define $X_t=B_t^1$, $Y_t=2B_t^2$, both starting at zero. Compute $d(X_tY_t)$.

For $f(x,y)=xy$, the first derivatives are $f_x=y$, $f_y=x$, and the second derivatives are $f_{xx}=f_{yy}=0$, $f_{xy}=f_{yx}=1$. The state covariation is $d[X,Y]_t=2\,d[B^1,B^2]_t=dt$. The two mixed Hessian terms combine to $d[X,Y]_t$. Therefore

d(X_tY_t)=Y_t\,dB_t^1+2X_t\,dB_t^2+dt.

Both stochastic integrands meet the required condition: $\mathbb{E}\int_0^T Y_s^2ds=4\int_0^T s\,ds<\infty$ and $\mathbb{E}\int_0^T(2X_s)^2ds=4\int_0^T s\,ds<\infty$. If the correlation were incorrectly treated as zero, the $dt$ correction would be omitted.

Check: The direct expectation is $\mathbb{E}[X_tY_t]=\mathbb{E}[B_t^1(2B_t^2)]=2(\tfrac12t)=t$. In the differential, the stochastic integrals have mean zero because their integrands are in $\mathcal{L}^2$, and the $dt$ term has expectation $t$. Omitting the mixed correction would predict expectation zero, contradicting the direct covariance calculation.

Worked Example: The Time Derivative Is Not Optional

Let $X_t=B_t$ and $f(t,x)=tx$. Determine $d(tB_t)$, and quantify what is lost if $f_t$ is omitted.

The derivatives are $f_t=x$, $f_x=t$, and $f_{xx}=0$. The process has drift zero and diffusion coefficient one. Hence

d(tB_t)=B_t\,dt+t\,dB_t.

The integral $\int_0^t s\,dB_s$ has a progressively measurable integrand with $\mathbb{E}\int_0^T s^2ds=T^3/3<\infty$; also $\mathbb{E}\int_0^T B_s^2ds=T^2/2<\infty$ for the time integral. Omitting $f_t$ leaves only $t\,dB_t$ and loses $\int_0^t B_s\,ds$.

Check: The product rule for a finite-variation factor $t$ gives $d(tB_t)=t\,dB_t+B_t\,dt$, matching Ito's formula. The omitted term is not identically zero: its second moment is $\mathbb{E}[(\int_0^T B_sds)^2]=\int_0^T\int_0^T\min(s,u)\,ds\,du=T^3/3$, which is positive for $T>0$. At $T=2$, this is $8/3$. Thus the time derivative contributes a genuine random term, even though its expectation is zero.

Separating Ito and Stratonovich Calculations

A further source of missing corrections is switching between Ito and Stratonovich notation midway through a calculation. The Ito integral $\int_0^t H_s\,dB_s$ is defined with left-endpoint adapted integrands. When the Stratonovich integral is defined for a continuous semimartingale integrand with the required covariation, its conversion rule is $\int_0^t H_s\circ dB_s=\int_0^t H_s\,dB_s+\tfrac12[H,B]_t$. Thus the ordinary chain rule for a Stratonovich differential does not mean that the Ito squared-diffusion correction can be discarded: the conversion term accounts for it. The examples here are all Ito calculations.

For a final check, inspect each line of a proposed result and ask four questions: Is $f_t$ present when $f$ depends on time? Is $\tfrac12 f_{xx}\sigma^2$ present in the scalar drift? Have all distinct covariations been included? Was the correlation matrix used to compute those covariations? These checks do not replace the hypotheses of Ito's formula, but they expose the common algebraic omissions once those hypotheses hold.

Common Mistakes

Using independent-driver rules for correlated Brownian motions. In the second example, replacing $d[B^1,B^2]_t=\tfrac12dt$ by zero incorrectly gives $d(X_tY_t)=Y_t\,dB_t^1+2X_t\,dB_t^2$. The correction is $d[X,Y]_t=2(\tfrac12dt)=dt$, so the correct differential includes $+dt$.

Counting the mixed term twice, or not at all. The Hessian sum has both $(i,j)$ and $(j,i)$ entries. Its leading factor $\tfrac12$ means the pair combines to one mixed term, not two. Conversely, deleting both entries because the covariance matrix is symmetric loses that term entirely.

Confusing absence of drift with absence of correction. A process with zero drift can still have nonzero diffusion, and the diffusion generates quadratic variation. In the scalar formula, the relevant correction is controlled by $\sigma^2 f_{xx}$, not by $\mu$.

Audit item	Term to verify	When it vanishes
Time dependence	$f_t\,dt$	$f$ has no time dependence
Scalar diffusion curvature	$\tfrac12\sigma^2f_{xx}\,dt$	$\sigma=0$ or $f_{xx}=0$
Mixed coordinates	$d[X^i,X^j]f_{x_ix_j}$	zero covariance or zero mixed derivative
Driver correlation	$A=\sigma R\sigma^\mathsf{T}$	use the actual correlation matrix

Check Your Understanding

For each question, identify the relevant derivative or covariance before simplifying.

For $dX_t=3\,dB_t$ and $f(x)=x^4$, what is the second-order drift correction?
If $d[B^1,B^2]_t=\rho\,dt$, $X=B^1$, and $Y=B^2$, what is the mixed correction in $d(XY)$?
For $f(t,x)=e^t x$, which term is lost if the time derivative is omitted?
In the multidimensional Hessian sum, explain why the $(i,j)$ and $(j,i)$ terms combine to one mixed term.
Give the two distinct conditions under which the scalar squared-diffusion correction vanishes at a point.

Audit item	Term to verify	When it vanishes
Time dependence	\(f_t\,dt\)	\(f\) has no time dependence
Scalar diffusion curvature	\(\tfrac12\sigma^2f_{xx}\,dt\)	\(\sigma=0\) or \(f_{xx}=0\)
Mixed coordinates	\(d[X^i,X^j]f_{x_ix_j}\)	zero covariance or zero mixed derivative
Driver correlation	\(A=\sigma R\sigma^\mathsf{T}\)	use the actual correlation matrix