Smoothness and Domain Conditions for Ito's Formula

Why Smoothness and Domain Conditions Matter

Brownian Itô formula combines a first derivative with a second-derivative correction. To use it, the function must have enough regularity for both derivatives to exist and behave continuously, and the Brownian path must remain where the function is defined. These are separate issues: a function can be smooth on its domain but singular at the domain boundary, while a globally smooth function can have derivatives whose growth requires care with stochastic-integral integrability.

Work on a filtered probability space $(\Omega,\mathcal{F},(\mathcal{F}_t)_{t\ge0},\mathbb{P})$ carrying a standard Brownian motion $B_t$, adapted to $(\mathcal{F}_t)$, with $B_0=0$. For a fixed starting point $x$, write $X_t=x+B_t$. Thus $X_0=x$; it is the shifted process $X$, not the standard Brownian motion $B$, that starts at $x$.

Definition: For an open interval $U\subseteq\mathbb{R}$, $f\in C^2(U)$ means that $f$, $f'$, and $f''$ exist on $U$ and are continuous there. For a function $f:[0,T]\times U\to\mathbb{R}$, $f\in C^{1,2}([0,T]\times U)$ means that $f$, $f_t$, $f_x$, and $f_{xx}$ are continuous, with the time derivative interpreted one-sided at the endpoints of $[0,T]$.

The Brownian Itô formula, as established in the earlier material on Itô formula, says that if $f\in C^2(\mathbb{R})$, then for $X_t=x+B_t$, $f(X_t)=f(x)+\int_0^t f'(X_s)\,dB_s+\frac12\int_0^t f''(X_s)\,ds$. For $f\in C^{1,2}([0,T]\times\mathbb{R})$, the time-dependent version is $f(t,X_t)=f(0,x)+\int_0^t f_x(s,X_s)\,dB_s+\int_0^t(f_t(s,X_s)+\frac12 f_{xx}(s,X_s))\,ds$. The second-derivative term is required because $dB_t\,dB_t=dt$; the ordinary chain rule does not apply to Brownian paths.

For these displayed formulas, the stochastic integral is the Itô integral. Whenever it is used as an ordinary square-integrable integral on $[0,T]$, its integrand must be progressively measurable and belong to $\mathcal{L}^2$, meaning $\mathbb{E}\int_0^T H_s^2\,ds<\infty$. If that global integrability has not been established, the formula is understood after stopping so that each integral is in $\mathcal{L}^2$; the resulting unstopped integral is then defined consistently as a local stochastic integral. In that case it is a local martingale, not necessarily a true martingale.

Local Smoothness and Stopping on Compact Subdomains

Global smoothness on all of $\mathbb{R}$ is stronger than necessary. Suppose $U$ is an open interval, $x\in U$, and $f\in C^2(U)$. The formula can be applied while $X$ remains in $U$. To make this precise, choose increasing open intervals $D_n=(a_n,b_n)$ such that $x\in D_n$, $\overline{D_n}$ is a compact subset of $U$, and $\bigcup_nD_n=U$. Define the exit times $\tau_n=\inf\{t\ge0:X_t\notin D_n\}$ and $\tau_U=\inf\{t\ge0:X_t\notin U\}$, with the infimum of the empty set defined as $+\infty$.

Theorem: Under these assumptions, for each $n$ and finite $t\ge0$, $$ f(X_{t\wedge\tau_n})=f(x)+\int_0^t \mathbf{1}_{\{s<\tau_n\}}f'(X_s)\,dB_s+\frac12\int_0^t \mathbf{1}_{\{s<\tau_n\}}f''(X_s)\,ds. $$ Both integrands in the stochastic integral are progressively measurable, and that integrand belongs to $\mathcal{L}^2$ on every finite time interval. Consequently, for $t<\tau_U$, the same identity holds with integrals taken up to $t$, without stopping at $\tau_n$, in the consistent local sense.

Proof. Fix $n$. Since $\overline{D_n}$ is compact and is contained in $U$, continuity of $f'$ and $f''$ implies that each is bounded on $\overline{D_n}$. Apply the standard Brownian Itô formula to the stopped process, or equivalently apply it up to $\tau_n$. The stochastic-integral integrand is $H_s=\mathbf{1}_{\{s<\tau_n\}}f'(X_s)$. It is progressively measurable because $X$ is adapted and continuous and the stopped indicator is progressively measurable. If $\sup_{y\in\overline{D_n}}|f'(y)|=C_n<\infty$, then for every finite $T$, $$ \mathbb{E}\int_0^T H_s^2\,ds\le C_n^2T<\infty. $$ Thus this stopped integral is an Itô integral with an $\mathcal{L}^2$ integrand, and the stopped formula follows. The intervals $D_n$ increase, so their exit times increase and $\tau_n\uparrow\tau_U$: before exit from $U$, continuity of the path on any compact time interval keeps it inside some $D_n$. For a fixed $t<\tau_U$, it follows that $t<\tau_n$ for all sufficiently large $n$. The stopped identities then agree on $[0,t]$, giving the asserted local identity. This argument does not assert a formula at $\tau_U$ if $f$ has no defined boundary value there. $\square$

The time-dependent statement has the same localization principle. If $f\in C^{1,2}([0,T]\times U)$, apply the time-dependent Brownian Itô formula on $[0,T]\times\overline{D_n}$. The derivatives are bounded there by continuity and compactness. In particular, the stopped stochastic-integral integrand $\mathbf{1}_{\{s<\tau_n\}}f_x(s,X_s)$ is progressively measurable and satisfies the same $\mathcal{L}^2$ bound with $C_n=\sup_{[0,T]\times\overline{D_n}}|f_x|$.

Worked Applications of the Conditions

Worked Example: A Globally Smooth Polynomial

Apply Itô formula to $f(y)=y^3$ and $X_t=2+B_t$, then check the expected value of the result.

Here $f'(y)=3y^2$ and $f''(y)=6y$, both continuous on $\mathbb{R}$. Substituting $X_0=2$ and these derivatives gives $$ (2+B_t)^3=8+\int_0^t 3(2+B_s)^2\,dB_s+\frac12\int_0^t 6(2+B_s)\,ds =8+\int_0^t 3(2+B_s)^2\,dB_s+\int_0^t 3(2+B_s)\,ds. $$ The stochastic integrand is progressively measurable. For every finite $T$, its squared-integrability follows from $\mathbb{E}\int_0^T9(2+B_s)^4\,ds=9\int_0^T(16+24s+3s^2)\,ds<\infty$; here the fourth moment is obtained by expanding the fourth power of a normal variable with mean $2$ and variance $s$.

Check: the stochastic integral has mean zero because it is an Itô integral with an $\mathcal{L}^2$ integrand. The expected correction integral is $3\int_0^t\mathbb{E}[2+B_s]\,ds=3\int_0^t2\,ds=6t$, so the formula gives $\mathbb{E}[(2+B_t)^3]=8+6t$. Directly, expanding the cube and using $\mathbb{E}[B_t]=0$, $\mathbb{E}[B_t^2]=t$, and $\mathbb{E}[B_t^3]=0$ gives $8+6t$ as well.

Worked Example: The Logarithm Before a Boundary

Let $X_t=2+B_t$, $U=(0,\infty)$, and $\tau=\inf\{t\ge0:X_t\notin(1,3)\}$. Apply Itô formula to $f(y)=\log y$ up to $\tau$, and determine the drift at the starting point.

The function is smooth on $U$, with $f'(y)=1/y$ and $f''(y)=-1/y^2$. The interval $[1,3]$ is a compact subset of $U$, and before $\tau$, $1\le X_s\le3$. Thus the stopped integrand $\mathbf{1}_{\{s<\tau\}}/X_s$ is progressively measurable and satisfies $\mathbb{E}\int_0^T\mathbf{1}_{\{s<\tau\}}X_s^{-2}\,ds\le T$. The stopped formula is $$ \log(X_{t\wedge\tau})=\log 2+\int_0^t\frac{\mathbf{1}_{\{s<\tau\}}}{X_s}\,dB_s-\frac12\int_0^t\frac{\mathbf{1}_{\{s<\tau\}}}{X_s^2}\,ds. $$ At $X_0=2$, the first-derivative coefficient is $f'(2)=1/2$, and the drift correction is $\frac12f''(2)=\frac12(-1/4)=-1/8$.

Check: the drift at any $y\in(1,3)$, computed from the generator for Brownian motion, is $\frac12 f''(y)=-1/(2y^2)$; substituting $y=2$ again gives $-1/8$. The identity is valid at the stopping time because $\log$ is also defined at the exit values $1$ and $3$. It does not justify applying $\log(X_t)$ after $X$ reaches zero, where the function is undefined.

Worked Example: Time Dependence and Spatial Smoothness

For $X_t=1+B_t$, apply the time-dependent formula to $f(t,y)=ty+y^2$, and check the expectation.

The derivatives are $f_t(t,y)=y$, $f_x(t,y)=t+2y$, and $f_{xx}(t,y)=2$. The function is in $C^{1,2}([0,T]\times\mathbb{R})$. The integrand $H_s=s+2(1+B_s)$ is progressively measurable, and $\mathbb{E}\int_0^T H_s^2\,ds=\int_0^T((s+2)^2+4s)\,ds<\infty$. The time-dependent formula therefore gives $$ t(1+B_t)+(1+B_t)^2 =1+\int_0^t(s+2+2B_s)\,dB_s+\int_0^t\big((1+B_s)+1\big)\,ds. $$ Indeed, $f(0,1)=1$, and the drift integrand is $f_t+\frac12f_{xx}=y+1$, which at $y=X_s$ equals $2+B_s$.

Check: the expected value of the left side is $t\mathbb{E}[1+B_t]+\mathbb{E}[(1+B_t)^2]=t+(1+t)=1+2t$. On the right, the stochastic integral has mean zero by its $\mathcal{L}^2$ condition, and the drift has expectation $\int_0^t\mathbb{E}[2+B_s]\,ds=2t$. The two calculations agree.

Common Mistakes and Practical Checks

Omitting the second derivative is not an innocuous simplification. For $f(y)=y^2$ and $X_t=B_t$, the ordinary chain rule would incorrectly give $B_t^2=2\int_0^tB_s\,dB_s$. Itô formula instead gives $B_t^2=2\int_0^tB_s\,dB_s+t$, since $f'(y)=2y$ and $\frac12f''(y)=1$. Taking expectations checks the correction: the left side has expectation $t$, while the stochastic integral has mean zero. The incorrect expression would give expectation zero.

Do not evaluate a function outside its domain. For $\log y$, first identify an interval contained in $(0,\infty)$, then stop before leaving it. A formula valid up to a stopping time does not automatically remain valid after the process reaches a singular boundary. If the function extends continuously to the boundary, the stopped identity may include the exit value; otherwise, state the formula only strictly before exit.

Continuous derivatives do not by themselves guarantee a true martingale. The Itô integrand must meet the required integrability condition for the integral to be square-integrable and hence a martingale. On a compact stopped region, bounded $f'$ supplies that condition on every finite horizon. Without stopping, growth of $f'$ can make $\mathbb{E}\int_0^T f'(X_s)^2\,ds$ infinite; the formula then requires localization, and its stochastic-integral term is only asserted to be a local martingale.

In applying the formula, check these points in order:

Is the function $C^2$, or $C^{1,2}$ when it depends on time, on an open set containing the process values under consideration?
Does the process remain in that set throughout the time interval, or must it be stopped at an exit time?
On the stopped region, are the derivatives bounded so the stochastic-integral integrand belongs to $\mathcal{L}^2$?
Have both the time derivative, when present, and the $\frac12 f_{xx}$ correction been included?

Check Your Understanding

For each question, specify the relevant domain and integrability conditions, not just the formal derivative calculation.

For $f(y)=1/y$, what open interval contains the starting point $x=2$, and what stopping rule would ensure the derivatives remain bounded?
State the first- and second-derivative terms in Itô formula for $f(y)=y^4$ and $X_t=x+B_t$.
Why does $f\in C^2((0,\infty))$ not justify applying Itô formula to $\log(X_t)$ after $X_t$ reaches zero?
What condition on $H$ makes $\int_0^T H_s\,dB_s$ a square-integrable Itô integral?
For a time-dependent $f$, which three derivatives appear in the Brownian drift, and how are they combined?

Smoothness and Domain Conditions for Ito's Formula

What You'll Learn