The Quotient Rule for Ito Processes

Why a Quotient Has Ito Corrections

For differentiable functions, the quotient rule differentiates the numerator and denominator and combines the results. For Ito processes, that first-order calculation is not enough: the denominator can fluctuate, and its quadratic variation contributes a second-order drift. If numerator and denominator are driven by common Brownian noise, their covariation contributes another correction. The quotient rule must retain both terms.

Work on a filtered probability space $(\Omega,\mathcal{F},(\mathcal{F}_t),\mathbb{P})$ carrying a $d$-dimensional Brownian motion $\mathbf{B}_t=(B_t^1,\ldots,B_t^d)$. Let $X_t$ and $Y_t$ be continuous Ito processes on $[0,T]$, with dynamics $dX_t=a_t\,dt+\sum_{j=1}^d \sigma_t^j\,dB_t^j$ and $dY_t=b_t\,dt+\sum_{j=1}^d \tau_t^j\,dB_t^j$. Here $a_t,b_t$ are the drift coefficients, and $\sigma_t^j,\tau_t^j$ are diffusion coefficients. Assume these coefficients are progressively measurable and, for every $T<\infty$, each is in $\mathcal{L}^2$: its expected time integral of the square is finite. The stochastic integrals in these dynamics are Ito integrals with these $\mathcal{L}^2$ integrands.

Write $\sigma_t=(\sigma_t^1,\ldots,\sigma_t^d)$ and $\tau_t=(\tau_t^1,\ldots,\tau_t^d)$. The quadratic-variation and covariation densities are $q_t=\sum_j(\tau_t^j)^2$ and $c_t=\sum_j\sigma_t^j\tau_t^j$: thus $d[Y]_t=q_t\,dt$ and $d[X,Y]_t=c_t\,dt$. These formulas use $dB_t^j\,dB_t^k=\mathbf{1}_{\{j=k\}}dt$, $dB_t^j\,dt=0$, and $dt\,dt=0$.

Definition: The quotient $R_t=X_t/Y_t$ is defined at times when $Y_t\ne0$. We say the denominator stays nonzero on $[0,T]$ if, almost surely, $Y_t\ne0$ for every $t\in[0,T]$. The quotient formula applies on this interval; if zeros are possible, it applies up to stopping times that keep the denominator away from zero.

The Quotient Rule

Theorem: Under the assumptions above, if $Y_t\ne0$ on $[0,T]$, then the quotient has the local Ito differential $$ d\left(\frac{X_t}{Y_t}\right) = \left( \frac{a_t}{Y_t} -\frac{X_tb_t}{Y_t^2} +\frac{X_tq_t}{Y_t^3} -\frac{c_t}{Y_t^2} \right)dt + \sum_{j=1}^d \left( \frac{\sigma_t^j}{Y_t} -\frac{X_t\tau_t^j}{Y_t^2} \right)dB_t^j. $$ The $X_tq_t/Y_t^3$ term is the correction from denominator variance; the $-c_t/Y_t^2$ term is the correction from covariation.

Proof. Apply the multidimensional Ito formula to $f(x,y)=x/y$, on the open set $y\ne0$. Its first and second partial derivatives are $f_x=1/y$, $f_y=-x/y^2$, $f_{xx}=0$, $f_{xy}=-1/y^2$, and $f_{yy}=2x/y^3$. In the Ito formula the first-order terms are $f_x\,dX_t+f_y\,dY_t$; the second-order terms are $\tfrac12 f_{xx}\,d[X]_t+f_{xy}\,d[X,Y]_t+\tfrac12 f_{yy}\,d[Y]_t$. Substitution gives the drift $a_t/Y_t-X_tb_t/Y_t^2-c_t/Y_t^2+X_tq_t/Y_t^3$. The coefficient of $dB_t^j$ is $f_x\sigma_t^j+f_y\tau_t^j=\sigma_t^j/Y_t-X_t\tau_t^j/Y_t^2$, giving the stated formula. To apply Ito’s formula when $Y$ may approach zero, stop first when $|Y|\le\varepsilon$, for $\varepsilon>0$; on the stopped interval the derivatives are defined and bounded on each bounded range of $X$. Localizing also when $|X|$ or the accumulated coefficient integrals become large justifies the formula there. If $Y$ stays nonzero on the compact interval, continuity gives a positive path-dependent lower bound for $|Y|$, so these stopped identities give the formula throughout $[0,T]$. $\square$

The displayed stochastic integrals are understood locally: their integrands are progressively measurable, and after stopping where $|Y|\ge\varepsilon$ and the other coefficients and processes are bounded, they belong to $\mathcal{L}^2$. On an unstopped interval, a quotient integrand need not have finite expected integral of its square merely because the original coefficients do. If the quotient integrands themselves are in $\mathcal{L}^2$, the corresponding Ito integrals are square-integrable martingales by the Ito isometry. Without that condition, the formula asserts a local stochastic integral, which need not be a true martingale.

A Second Derivation: Reciprocal and Product

The same correction terms can be checked by writing $X/Y=X(1/Y)$. This method makes the separate roles of denominator variance and covariation visible. The reciprocal is a function of $Y$ alone, so its Ito differential uses the $f''/2$ correction.

Proof. Set $g(y)=1/y$. Then $g'(y)=-1/y^2$ and $g''(y)=2/y^3$; the one-dimensional Ito formula gives $d(1/Y_t)=(-b_t/Y_t^2+q_t/Y_t^3)dt-\sum_j(\tau_t^j/Y_t^2)dB_t^j$. The Ito product rule is $d(UV)=U\,dV+V\,dU+d[U,V]$. For $U=X$ and $V=1/Y$, the covariation density is $-c_t/Y_t^2$, since the diffusion coefficients of $1/Y$ are $-\tau_t^j/Y_t^2$. Therefore $d(X/Y)=(1/Y)dX+X\,d(1/Y)+d[X,1/Y]$. Substituting the differentials yields drift $a_t/Y_t-X_tb_t/Y_t^2+X_tq_t/Y_t^3-c_t/Y_t^2$ and diffusion coefficients $\sigma_t^j/Y_t-X_t\tau_t^j/Y_t^2$, as claimed. As before, stop away from zero and localize if necessary; the product rule and Ito integrals are then valid on the stopped intervals. $\square$

Worked Applications

Worked Example: A Shared Brownian Driver

Let $B_t$ be a one-dimensional Brownian motion and set $X_t=1+2B_t$, $Y_t=2+B_t$. Work up to the first time $Y$ reaches zero; on any interval stopped earlier, compute $d(X_t/Y_t)$.

Here $a_t=b_t=0$, $\sigma_t=2$, and $\tau_t=1$. Thus $q_t=1$ and $c_t=2$. Substitution gives drift $-2/Y_t^2+X_t/Y_t^3=(-2Y_t+X_t)/Y_t^3$. Since $Y_t=2+B_t$ and $X_t=1+2B_t$, the numerator is $-2(2+B_t)+(1+2B_t)=-3$. The diffusion coefficient is $2/Y_t-X_t/Y_t^2=(2Y_t-X_t)/Y_t^2$; its numerator is $2(2+B_t)-(1+2B_t)=3$. Therefore

d(X_t/Y_t)=-\frac{3}{Y_t^3}\,dt+\frac{3}{Y_t^2}\,dB_t.

Check: $X_t=2Y_t-3$, so $X_t/Y_t=2-3/Y_t$. For $g(y)=2-3/y$, $g'(y)=3/y^2$ and $g''(y)=-6/y^3$. Since $dY_t=dB_t$, Ito’s formula gives drift $\tfrac12(-6/Y_t^3)=-3/Y_t^3$ and diffusion $3/Y_t^2$, exactly as above.

Worked Example: Independent Noise in the Denominator

Let $B^1,B^2$ be independent Brownian motions, $X_t=3$, and $Y_t=2+B_t^2$. Compute the quotient differential before $Y$ reaches zero.

The coefficients are $a_t=b_t=0$, $\sigma_t^1=\sigma_t^2=0$, $\tau_t^1=0$, and $\tau_t^2=1$. Hence $q_t=1$ and $c_t=0$. The quotient rule gives drift $3/Y_t^3$ and diffusion coefficient $-3/Y_t^2$ against $dB_t^2$; the $dB_t^1$ coefficient is zero. Thus

d(3/Y_t)=\frac{3}{Y_t^3}\,dt-\frac{3}{Y_t^2}\,dB_t^2.

Check: for $g(y)=3/y$, $g'(y)=-3/y^2$ and $g''(y)=6/y^3$. Ito’s correction is $\tfrac12g''(Y_t)d[Y]_t=3/Y_t^3\,dt$, and the first-order term is $-3/Y_t^2\,dB_t^2$, confirming both coefficients. Independence makes the covariation term zero; it does not remove the denominator’s variance correction.

Worked Example: A Deterministic Denominator

Let $X_t=B_t$ and $Y_t=2+t$, for $0\le t\le T$. Compute the differential of $R_t=B_t/(2+t)$.

Here $a_t=b_t=0$, $\sigma_t=1$, and $\tau_t=0$, so $q_t=c_t=0$. The rule gives drift zero and diffusion $1/(2+t)$, hence $dR_t=dB_t/(2+t)$. The denominator is at least $2$, so it never vanishes.

Check: the product representation is $R_t=g(t)B_t$, where $g(t)=1/(2+t)$ has derivative $g'(t)=-1/(2+t)^2$. The time-dependent product rule gives $dR_t=g(t)dB_t+g'(t)B_tdt$ only if differentiating the factor is included; in this case, that would appear to give a drift. Directly applying the time-dependent Ito formula to $f(t,x)=x/(2+t)$, however, gives $f_t=-x/(2+t)^2$, $f_x=1/(2+t)$, and $f_{xx}=0$. Thus the correct differential is $-B_t/(2+t)^2\,dt+dB_t/(2+t)$. The quotient rule gives the same result only after including the time derivative of the denominator: $dY_t=dt$, so $b_t=1$, and its drift term is $-X_t/(2+t)^2$. The corrected answer is therefore

dR_t=-\frac{B_t}{(2+t)^2}\,dt+\frac{1}{2+t}\,dB_t.

The initial zero drift calculation omitted $b_t=1$; the independent time-dependent calculation detects the omission.

Conditions, Interpretation, and a Common Mistake

The phrase “the denominator stays nonzero” matters. If a continuous process $Y$ starts away from zero but can hit zero, the quotient is not defined at that hitting time. For example, one may define $\tau_\varepsilon=\inf\{t\ge0:|Y_t|\le\varepsilon\}$ and apply the formula up to $T\wedge\tau_\varepsilon$, then consider the limit only on intervals where the denominator remains nonzero. A formal expression with powers of $1/Y_t$ does not extend the quotient through a zero.

A common mistake is to retain only the ordinary first-order quotient terms, $(a_t/Y_t-X_tb_t/Y_t^2)dt$, and the ordinary-looking diffusion terms. In Worked Example 1 this would predict zero drift. The correct drift is $-3/Y_t^3$, as the independent calculation from $2-3/Y_t$ verifies. The omitted drift is not arbitrary: the denominator variance contributes $X_tq_t/Y_t^3$, while shared noise contributes $-c_t/Y_t^2$. These may reinforce or cancel one another, but neither can be discarded without checking the coefficients.

In one dimension, the covariation correction is zero when the two processes have no common Brownian component, as in Worked Example 2. With several drivers it is the dot product of their diffusion vectors, so the drivers need not be identical for covariation to be nonzero. A deterministic denominator has $q_t=0$; the usual time-dependent quotient drift still includes $-X_tb_t/Y_t^2$.

Takeaway: For a quotient, compute the denominator’s quadratic-variation density $q_t$ and the numerator-denominator covariation density $c_t$ before simplifying. The second-order drift is $X_tq_t/Y_t^3-c_t/Y_t^2$. The formula is local near nonzero denominators, and its stochastic-integral terms are true martingales only when the required square-integrability holds.

Check Your Understanding

For each question, assume the coefficients are adapted and the relevant stochastic integrals are defined; when a quotient is used, restrict to times on which the denominator is nonzero.

For $dX_t=a_tdt+\sigma_tdB_t$ and $dY_t=b_tdt+\tau_tdB_t$, identify $q_t$ and $c_t$.
Which correction term remains when $X_t$ is deterministic but $Y_t$ has nonzero diffusion?
Why does independent Brownian noise in $X$ and $Y$ make $c_t=0$ but not necessarily $q_t=0$?
Apply the quotient rule to $X_t=1+B_t$, $Y_t=3+B_t$, up to the first zero of $Y$.
What integrability condition ensures that a quotient’s Ito integral is a square-integrable martingale rather than merely a local martingale?