Itô's Lemma: Derivation and Applications

Setup

We work on $(\Omega, \mathcal{F}, \mathbb{P})$ with filtration $\mathbb{F} = (\mathcal{F}_t)_{t \geq 0}$ satisfying the usual conditions. Let $(W_t)_{t \geq 0}$ be a standard $\mathbb{P}$-Brownian motion.

An Itô process is a continuous adapted process of the form: $X_t = X_0 + \int_0^t \mu_s \, ds + \int_0^t \sigma_s \, dW_s,$ where $\mu$ (drift) and $\sigma$ (diffusion) are $\mathbb{F}$-progressively measurable and satisfy: $\int_0^T |\mu_s| \, ds < \infty \quad \text{and} \quad \int_0^T \sigma_s^2 \, ds < \infty \quad \text{a.s.}$

In differential notation: $dX_t = \mu_t \, dt + \sigma_t \, dW_t$.

The Itô Integral

The stochastic integral $\int_0^T \sigma_s \, dW_s$ is defined as the $L^2$ limit of non-anticipating Riemann sums: $\int_0^T \sigma_s \, dW_s = L^2\text{-}\!\lim_{|\pi| \to 0} \sum_{i} \sigma_{t_{i-1}} \bigl(W_{t_i} - W_{t_{i-1}}\bigr).$

The requirement to evaluate $\sigma$ at the left endpoint $t_{i-1}$ — not the midpoint — is what makes the integral Itô. Left-endpoint evaluation ensures $\sigma_{t_{i-1}}$ is $\mathcal{F}_{t_{i-1}}$-measurable (non-anticipating). The choice of evaluation point matters: different conventions yield different integrals (see Itô vs Stratonovich below).

Itô isometry. For square-integrable adapted $\sigma$: $\mathbb{E}\!\left[\left(\int_0^T \sigma_s \, dW_s\right)^2\right] = \mathbb{E}\!\left[\int_0^T \sigma_s^2 \, ds\right].$

This is the key $L^2$ norm identity. It follows from expanding the square and using independence of non-overlapping Brownian increments: cross terms $\mathbb{E}[\sigma_{t_{i-1}}(W_{t_i} - W_{t_{i-1}}) \cdot \sigma_{t_{j-1}}(W_{t_j} - W_{t_{j-1}})] = 0$ for $i \neq j$.

A square-integrable adapted integrand produces a martingale: $\mathbb{E}\!\left[\int_0^t \sigma_s \, dW_s \,\Big|\, \mathcal{F}_r\right] = \int_0^r \sigma_s \, dW_s, \quad r \leq t.$

Itô's Lemma: Statement

Let $f: \mathbb{R}_+ \times \mathbb{R} \to \mathbb{R}$ be $C^{1,2}$ (once continuously differentiable in $t$, twice in $x$). Define $Y_t = f(t, X_t)$. Then $Y$ is again an Itô process and: $dY_t = \frac{\partial f}{\partial t}(t, X_t) \, dt + \frac{\partial f}{\partial x}(t, X_t) \, dX_t + \frac{1}{2} \frac{\partial^2 f}{\partial x^2}(t, X_t) \, (dX_t)^2,$ where $(dX_t)^2$ is evaluated using the Itô multiplication table: $dt \cdot dt = 0, \qquad dt \cdot dW_t = 0, \qquad dW_t \cdot dW_t = dt.$

Substituting $dX_t = \mu_t \, dt + \sigma_t \, dW_t$ and $(dX_t)^2 = \sigma_t^2 \, dt$: $\boxed{dY_t = \left(\frac{\partial f}{\partial t} + \mu_t \frac{\partial f}{\partial x} + \frac{1}{2}\sigma_t^2 \frac{\partial^2 f}{\partial x^2}\right) dt + \sigma_t \frac{\partial f}{\partial x} \, dW_t.}$

The term $\frac{1}{2}\sigma_t^2 f_{xx}$ is the Itô correction. It has no analogue in classical calculus and arises solely from the non-zero quadratic variation of Brownian motion.

Derivation

Apply Taylor's theorem to $f(t + dt, X_{t+dt})$ around $(t, X_t)$: $df = f_t \, dt + f_x \, dX + \frac{1}{2}f_{xx}(dX)^2 + \frac{1}{2}f_{tt}(dt)^2 + f_{xt} \, dt \, dX + \cdots$

Substitute $dX = \mu \, dt + \sigma \, dW$ and expand $(dX)^2$: $(dX)^2 = \mu^2(dt)^2 + 2\mu\sigma \, dt \, dW + \sigma^2(dW)^2.$

Apply the Itô multiplication table. The key step: $(dW)^2 = dt$ is the quadratic variation result $[W]_t = t$ in differential form. The other products vanish as $o(dt)$: $(dX)^2 = \sigma^2 \, dt + O\!\left((dt)^{3/2}\right).$

Similarly, $(dt)^2 = 0$ and $dt \cdot dW = 0$ in the $L^2$ sense. Retaining only terms of order $dt$: $df = f_t \, dt + f_x(\mu \, dt + \sigma \, dW) + \frac{1}{2}f_{xx}\sigma^2 \, dt.$

Collecting drift and diffusion terms: $df = \left(f_t + \mu f_x + \frac{1}{2}\sigma^2 f_{xx}\right) dt + \sigma f_x \, dW.$

This is Itô's lemma. The heuristic argument is exact in identifying the correct terms; the rigorous version replaces the Taylor remainder analysis with an $L^2$ convergence argument using the Itô isometry.

Application: Solving the GBM SDE

Let $S_t$ satisfy the geometric Brownian motion SDE: $dS_t = \mu S_t \, dt + \sigma S_t \, dW_t.$

Apply Itô's lemma to $f(S_t) = \ln S_t$, with $f_x = 1/x$, $f_{xx} = -1/x^2$, $f_t = 0$: $d(\ln S_t) = \frac{1}{S_t} \, dS_t - \frac{1}{2} \cdot \frac{1}{S_t^2} \cdot \sigma^2 S_t^2 \, dt = \left(\mu - \frac{\sigma^2}{2}\right) dt + \sigma \, dW_t.$

Integrating from 0 to $T$: $\ln S_T - \ln S_0 = \left(\mu - \frac{\sigma^2}{2}\right)T + \sigma W_T.$

Therefore: $S_T = S_0 \exp\!\left(\left(\mu - \frac{\sigma^2}{2}\right)T + \sigma W_T\right).$

The Role of the Itô Correction

The term $-\sigma^2/2$ is not a typo. It is the Itô correction from $f_{xx} = -1/S^2$. Without it — if one naively wrote $\ln S_T = \ln S_0 + \mu T + \sigma W_T$ — the expectation would be wrong: $\mathbb{E}[S_T] = S_0 e^{\mu T}, \quad \text{correct from the SDE:} \quad d\mathbb{E}[S_t] = \mu \mathbb{E}[S_t] \, dt.$

But $\mathbb{E}[e^{(\mu + \sigma^2/2)T + \sigma W_T}] = e^{(\mu + \sigma^2/2)T} \cdot e^{\sigma^2 T/2} = e^{(\mu + \sigma^2)T} \neq e^{\mu T}$. The Itô correction $-\sigma^2/2$ is what reconciles the arithmetic mean growth $\mu$ of the SDE with the geometric mean growth $\mu - \sigma^2/2$ of the log process.

Itô vs Stratonovich

The Stratonovich integral evaluates the integrand at the midpoint: $\int_0^T \sigma_s \circ dW_s = L^2\text{-}\!\lim_{|\pi| \to 0} \sum_i \frac{\sigma_{t_i} + \sigma_{t_{i-1}}}{2}\bigl(W_{t_i} - W_{t_{i-1}}\bigr).$

The Stratonovich integral obeys the classical chain rule — no correction term: $df(W_t) = f'(W_t) \circ dW_t.$

The two integrals are related by: $\int_0^T \sigma_s \circ dW_s = \int_0^T \sigma_s \, dW_s + \frac{1}{2}\int_0^T \frac{\partial \sigma}{\partial x}(X_s) \sigma_s \, ds.$

The Stratonovich convention is preferred in physics and differential geometry (because it obeys the classical chain rule and behaves well under coordinate changes). The Itô convention is standard in financial mathematics because:

1. The Itô integral of an adapted square-integrable process is a martingale — essential for risk-neutral pricing.
2. The Stratonovich integral anticipates future information through the midpoint evaluation, which is economically meaningless.

Multidimensional Itô Lemma

For a vector of Itô processes $X = (X^1, \ldots, X^d)$ driven by correlated Brownian motions $W^1, \ldots, W^m$ with $d\langle W^i, W^j\rangle_t = \rho_{ij} \, dt$, and a function $f \in C^{1,2}$: $df(t, X_t) = f_t \, dt + \sum_{i=1}^d f_{x_i} \, dX_t^i + \frac{1}{2}\sum_{i,j=1}^d f_{x_i x_j} \, d\langle X^i, X^j\rangle_t.$

Here $d\langle X^i, X^j\rangle_t = \sum_{k=1}^m \sigma^{ik}\sigma^{jk} \, dt$ is the quadratic covariation, where $\sigma^{ik}$ is the diffusion coefficient of $X^i$ with respect to $W^k$.

Limitations

Regularity. Itô's lemma requires $f \in C^{1,2}$. For payoffs with kinks — the call payoff $(x - K)^+$ has $f_{xx} = \delta_{x=K}$ as a distribution — the formula breaks down. The correct generalization uses Tanaka's formula and local time.

Pathwise inapplicability. The Itô integral cannot be defined sample-path by sample-path: the paths of $W$ have infinite total variation. The $L^2$ construction is inherently probabilistic. All Itô calculus identities hold in the almost-sure or $L^2$ sense, not pointwise in $\omega$.

Itô correction in parameter estimation. If one observes a log-price process and estimates $\mu$ from the empirical mean of log-returns, the estimate targets $\mu - \sigma^2/2$, not $\mu$. The distinction matters for maximum likelihood estimation of drift in GBM models.

Interview Angle

L1: State Itô's lemma. Apply it to $f(S_t) = \ln S_t$ for $dS_t = \mu S_t \, dt + \sigma S_t \, dW_t$. What is the economic meaning of the $-\sigma^2/2$ term?

Statement. For $f \in C^{1,2}$ and $dX_t = \mu_t \, dt + \sigma_t \, dW_t$: $df(t, X_t) = \left(f_t + \mu_t f_x + \tfrac{1}{2}\sigma_t^2 f_{xx}\right) dt + \sigma_t f_x \, dW_t.$

Application to $\ln S_t$. With $f(x) = \ln x$: $f_x = 1/x$, $f_{xx} = -1/x^2$, $f_t = 0$, $\mu_t = \mu S_t$, $\sigma_t = \sigma S_t$: $d(\ln S_t) = \frac{1}{S_t} \cdot \mu S_t \, dt - \frac{1}{2} \cdot \frac{1}{S_t^2} \cdot \sigma^2 S_t^2 \, dt + \frac{\sigma S_t}{S_t} \, dW_t = \left(\mu - \frac{\sigma^2}{2}\right) dt + \sigma \, dW_t.$ Integrating: $S_T = S_0 \exp\!\left((\mu - \sigma^2/2)T + \sigma W_T\right)$.

Economic meaning of $-\sigma^2/2$. This is the volatility drag — the gap between arithmetic and geometric growth. The SDE $dS = \mu S \, dt + \sigma S \, dW$ implies $\mathbb{E}[S_T] = S_0 e^{\mu T}$: the stock grows arithmetically at rate $\mu$. But $\mathbb{E}[\ln(S_T/S_0)] = (\mu - \sigma^2/2)T$: log-returns grow at the lower rate $\mu - \sigma^2/2$. The gap $\sigma^2/2$ arises from the convexity of $\exp$ (Jensen's inequality): because log-returns are normally distributed, the average of the exponent exceeds the exponent of the average by exactly $\sigma^2/2$. In practice: a fund with annualised vol $\sigma = 20\%$ suffers 2% drag annually — its compound growth rate is 2% below its average return. Practitioners who confuse arithmetic and geometric returns misprice long-horizon options and misreport expected performance.

L2: Derive Itô's lemma from a Taylor expansion. Why does $(dW)^2 = dt$ in the Itô table? What is the Itô isometry, and why does it imply the stochastic integral is a martingale?

Derivation. Apply the second-order Taylor expansion to $f(t + dt, X_{t+dt})$ around $(t, X_t)$: $df = f_t \, dt + f_x \, dX + \frac{1}{2}f_{xx}(dX)^2 + O(dt^{3/2}).$ Expand $(dX)^2 = (\mu \, dt + \sigma \, dW)^2 = \mu^2 (dt)^2 + 2\mu\sigma \, dt \cdot dW + \sigma^2(dW)^2$. Apply the multiplication table: $(dt)^2 = 0$, $dt \cdot dW = 0$, and $(dW)^2 = dt$. Only the last term survives, giving $(dX)^2 = \sigma^2 \, dt$. Substituting: $df = f_t \, dt + f_x(\mu \, dt + \sigma \, dW) + \frac{1}{2}f_{xx}\sigma^2 \, dt = \left(f_t + \mu f_x + \tfrac{1}{2}\sigma^2 f_{xx}\right)dt + \sigma f_x \, dW. \qquad \square$

Why $(dW)^2 = dt$. This is the differential form of $[W]_t = t$. On a partition of mesh $h$, the squared increment $(W_{t+h} - W_t)^2$ has mean $h$ and variance $2h^2$. Summing over $n = T/h$ intervals: the total has mean $T$ and variance $2T^2/n \to 0$ as $n \to \infty$. The sum concentrates on $T$ in $L^2$ (not just in expectation) — the fluctuations vanish, and the quadratic variation is deterministically equal to $t$. In the Taylor expansion this means $(dW)^2$ contributes a deterministic correction of size $dt$, not a random term: it is absorbed into the drift, not the martingale part.

Itô isometry. For square-integrable adapted $\sigma$: $\mathbb{E}\!\left[\left(\int_0^T \sigma_s \, dW_s\right)^2\right] = \mathbb{E}\!\left[\int_0^T \sigma_s^2 \, ds\right].$ Proof. Expand the square of the Riemann-sum approximant $\sum_i \sigma_{t_{i-1}} \Delta W_i$: $\mathbb{E}\!\left[\left(\sum_i \sigma_{t_{i-1}} \Delta W_i\right)^2\right] = \sum_{i,j} \mathbb{E}[\sigma_{t_{i-1}}\sigma_{t_{j-1}} \Delta W_i \Delta W_j].$ For $i \neq j$ (say $i < j$): $\Delta W_j$ is independent of $\mathcal{F}_{t_{j-1}}$, which contains $\sigma_{t_{i-1}}, \sigma_{t_{j-1}}, \Delta W_i$. So $\mathbb{E}[\sigma_{t_{i-1}}\sigma_{t_{j-1}} \Delta W_i \Delta W_j] = \mathbb{E}[\sigma_{t_{i-1}}\sigma_{t_{j-1}} \Delta W_i] \cdot \mathbb{E}[\Delta W_j] = 0$. Only diagonal terms contribute: $\sum_i \mathbb{E}[\sigma_{t_{i-1}}^2 (\Delta W_i)^2] = \sum_i \mathbb{E}[\sigma_{t_{i-1}}^2] \cdot h \to \mathbb{E}[\int_0^T \sigma_s^2 \, ds]$.

Why the stochastic integral is a martingale. For an adapted square-integrable integrand, $M_t = \int_0^t \sigma_s \, dW_s$ is a martingale: $\mathbb{E}[M_t \mid \mathcal{F}_r] = M_r, \quad r \leq t.$ The key is the non-anticipating (left-endpoint) evaluation: $\sigma_{t_{i-1}}$ is $\mathcal{F}_{t_{i-1}}$-measurable and hence independent of $\Delta W_i = W_{t_i} - W_{t_{i-1}}$. So $\mathbb{E}[\sigma_{t_{i-1}} \Delta W_i \mid \mathcal{F}_{t_{i-1}}] = \sigma_{t_{i-1}} \mathbb{E}[\Delta W_i \mid \mathcal{F}_{t_{i-1}}] = 0$. Summing over future steps: $\mathbb{E}[M_t - M_r \mid \mathcal{F}_r] = 0$. This is why the Itô convention (left-endpoint) is essential for finance: the resulting stochastic integral represents the gains from a non-anticipating trading strategy, and the martingale property ensures no-arbitrage.

L3: Compare Itô and Stratonovich conventions. Why is the Stratonovich integral not a martingale? State the multidimensional Itô lemma and apply it to a two-factor stochastic volatility model where $dS$ and $d\nu$ are correlated.

Itô vs Stratonovich. The Itô integral evaluates the integrand at the left endpoint of each interval; the Stratonovich integral at the midpoint: $\int_0^T \sigma_s \circ dW_s = L^2\text{-}\lim \sum_i \frac{\sigma_{t_{i-1}} + \sigma_{t_i}}{2}(W_{t_i} - W_{t_{i-1}}).$ The midpoint value $\frac{1}{2}(\sigma_{t_{i-1}} + \sigma_{t_i})$ partially anticipates the future: $\sigma_{t_i}$ is $\mathcal{F}_{t_i}$-measurable, and there is a non-trivial correlation $\mathbb{E}[(\sigma_{t_i} - \sigma_{t_{i-1}})(W_{t_i} - W_{t_{i-1}})] \neq 0$ when $\sigma$ is itself driven by $W$. Concretely, if $d\sigma = a \, dt + b \, dW$, then: $\mathbb{E}\!\left[\frac{\sigma_{t_{i-1}} + \sigma_{t_i}}{2}(W_{t_i} - W_{t_{i-1}})\right] \approx \frac{1}{2} b \cdot h \neq 0.$ So $\mathbb{E}[M^{\mathrm{Strat}}_t - M^{\mathrm{Strat}}_r \mid \mathcal{F}_r] \neq 0$: the conditional increment has a non-zero drift, and the Stratonovich integral is not a martingale. The conversion formula makes this precise: $\int_0^T \sigma_s \circ dW_s = \int_0^T \sigma_s \, dW_s + \frac{1}{2}\int_0^T \frac{\partial\sigma}{\partial x}(X_s)\sigma_s \, ds.$ The correction term $\frac{1}{2}\int \frac{\partial\sigma}{\partial x}\sigma \, ds$ is a finite-variation process — it is the martingale-killing drift added by midpoint evaluation. Stratonovich is standard in physics (where it obeys the classical chain rule and is appropriate for ODEs perturbed by smooth noise) and differential geometry, but it is the wrong convention for financial modelling.

Multidimensional Itô lemma. For Itô processes $X^1, \ldots, X^d$ driven by correlated Brownian motions $W^1, \ldots, W^m$ with $d\langle W^i, W^j\rangle_t = \rho_{ij} \, dt$, and $f \in C^{1,2}$: $df(t, X_t) = f_t \, dt + \sum_i f_{x_i} \, dX^i_t + \frac{1}{2}\sum_{i,j} f_{x_i x_j} \, d\langle X^i, X^j\rangle_t.$

Application: Heston-type model. Let $S_t$ (spot) and $\nu_t$ (instantaneous variance) satisfy: $dS_t = \mu S_t \, dt + \sqrt{\nu_t} S_t \, dW^1_t, \qquad d\nu_t = \kappa(\bar\nu - \nu_t) \, dt + \xi\sqrt{\nu_t} \, dW^2_t, \qquad d\langle W^1, W^2\rangle_t = \rho \, dt.$ The quadratic covariations are: $d\langle S, S\rangle_t = \nu_t S_t^2 \, dt, \quad d\langle \nu, \nu\rangle_t = \xi^2 \nu_t \, dt, \quad d\langle S, \nu\rangle_t = \rho\xi\nu_t S_t \, dt.$ Applying the multidimensional Itô lemma to $V(t, S_t, \nu_t)$: $dV = \underbrace{\left[V_t + \mu S V_S + \kappa(\bar\nu - \nu) V_\nu + \tfrac{1}{2}\nu S^2 V_{SS} + \rho\xi\nu S \, V_{S\nu} + \tfrac{1}{2}\xi^2\nu V_{\nu\nu}\right]}_{\text{drift}} dt + \sqrt{\nu} S V_S \, dW^1 + \xi\sqrt{\nu} V_\nu \, dW^2.$ Under the risk-neutral measure $\mathbb{Q}$, we replace $\mu \to r$ and the two Brownian terms constitute the option's hedge portfolio. Setting the discounted option price to be a $\mathbb{Q}$-martingale forces the drift to equal $rV$, giving the Heston PDE: $V_t + rS V_S + \kappa(\bar\nu - \nu) V_\nu + \tfrac{1}{2}\nu S^2 V_{SS} + \rho\xi\nu S \, V_{S\nu} + \tfrac{1}{2}\xi^2\nu V_{\nu\nu} - rV = 0.$ Note the cross-derivative term $\rho\xi\nu S \, V_{S\nu}$: it is absent in Black-Scholes and directly encodes the vol-of-vol and spot-vol correlation that generates the implied vol skew.