Itô's Lemma & Stochastic Differential Equations

You can already simulate a stock by chaining random draws — that’s Monte Carlo, and the engine inside it was geometric Brownian motion. But where did that engine come from? Why does the GBM step subtract a mysterious $\tfrac{1}{2}\sigma^2$? Why does the price turn out lognormal rather than normal? This lesson is the conceptual peak of the whole topic: we write down the stochastic differential equation (SDE) that defines how a price moves in continuous time, watch ordinary calculus break the instant it touches a Brownian path, and meet Itô’s lemma — the one tool that repairs it, manufactures the $\tfrac{1}{2}\sigma^2$ correction, and hands you GBM as a solution rather than a recipe you took on faith.

What a stochastic differential equation says

A stochastic differential equation is a recipe for the next infinitesimally small move. Written out, the SDE for a process $X_t$ is:

$dX_t = \mu(X_t, t)\,dt + \sigma(X_t, t)\,dW_t$

Read it left to right as a sentence. Over the next tiny instant $dt$, the change $dX_t$ is the sum of two pieces:

• The drift term $\mu\,dt$ — a predictable pull, proportional to the length of time $dt$. This is where the process wants to go: the steady, deterministic lean.
• The diffusion term $\sigma\,dW_t$ — a random Brownian kick, proportional to the increment $dW_t$ of a Wiener process (the continuous-time random walk from the last lesson). This is the noise: unpredictable, mean-zero, scaling with the volatility $\sigma$.

Analogy. Picture a sailboat crossing a bay. There’s a steady current carrying it in one direction — that’s the drift, smooth and predictable, it acts the same every second. And there are gusty winds shoving it left and right at random — that’s the diffusion, jittery and mean-zero. Where the boat ends up is the current’s steady push plus the accumulated buffeting of a thousand random gusts. The SDE is just that sentence written in math: position next instant = steady push + random shove.

Worked example — arithmetic Brownian motion. The simplest SDE takes constant coefficients:

$dX_t = \mu\,dt + \sigma\,dW_t$

Read it aloud: “over the next instant, $X$ drifts up by $\mu\,dt$ (a fixed lean) and gets jostled by $\sigma\,dW_t$ (a random kick scaled by volatility).” There’s no $X$ on the right-hand side, so the drift and kick don’t depend on the current level — the process wanders the whole real line and can go negative. That’s fine for an interest-rate spread or a temperature, but as you’ll see, it’s wrong for a price, which is why finance reaches for a level-dependent SDE instead.

Why ordinary calculus breaks

The chain rule you learned in school — $df = f'(x)\,dx$ — has a hidden assumption baked in: the path $x(t)$ is smooth. Smooth means that over a tiny step, the change is tiny in proportion ($dx$), and anything quadratic ($dx^2$) is negligibly tinier still, so you throw it away. That’s the whole trick of differential calculus: second-order terms vanish.

Brownian motion shatters that assumption. As you saw last lesson, a Wiener path is continuous but nowhere differentiable — you cannot draw a tangent line to it — and it has the strange property of quadratic variation equal to elapsed time. In the shorthand of stochastic calculus, that’s the box-calculus rule:

$(dW)^2 = dt, \qquad (dt)^2 = 0, \qquad dt\,dW = 0$

That first equation is the entire revolution. On a smooth path, $(dx)^2$ is second-order and dies. On a Brownian path, $(dW)^2$ is first-order — it equals $dt$, the same order as the drift — so it refuses to vanish. When you Taylor-expand a function $f$ of a Brownian path, the second-order term

$\tfrac{1}{2}f''(W)\,(dW)^2 = \tfrac{1}{2}f''(W)\,dt$

does not disappear. It survives, contributing a real $dt$-sized drift that ordinary calculus never sees. That surviving term is Itô’s correction. Everything else in this lesson is a consequence of those two characters: $(dW)^2 = dt$.

Itô’s lemma and the ½σ² correction

Itô’s lemma is the chain rule for stochastic paths. In words: when you transform a process $X_t$ through a smooth function $f$, the change $df$ is everything ordinary calculus gives you plus one extra term — the surviving second-order piece $\tfrac{1}{2}\sigma^2 f''\,dt$. For $f(X_t)$ where $dX_t = \mu\,dt + \sigma\,dW_t$:

$df = \underbrace{\left(\mu f' + \tfrac{1}{2}\sigma^2 f''\right)dt}_{\text{drift}} + \underbrace{\sigma f'\,dW_t}_{\text{diffusion}}$

Compare it to the ordinary chain rule $df = f'\,dX = (\mu f')dt + (\sigma f')dW$. Identical except for the bonus $\tfrac{1}{2}\sigma^2 f''\,dt$ glued onto the drift — the fingerprint of $(dW)^2 = dt$.

The headline consequence for finance. Suppose a price $S$ follows the SDE $dS = \mu S\,dt + \sigma S\,dW$ (drift $\mu$, volatility $\sigma$, both scaling with the price level). Now ask: how does $\log S$ move? Apply Itô’s lemma with $f(S) = \log S$, so $f' = 1/S$ and $f'' = -1/S^2$. The drift of $\log S$ becomes $\mu - \tfrac{1}{2}\sigma^2$. In words:

A price with drift $\mu$ does not give a log-price that drifts at $\mu$. The log drifts at $\mu - \tfrac{1}{2}\sigma^2$.

The curvature of the log function (it’s concave — its second derivative $f''$ is negative) bends the random kicks downward on average, and Itô’s term measures exactly how much: half the variance per unit time. The chart below makes it visible — slide the volatility $\sigma$ up and watch the log-price drift sink below the price drift, with the shaded wedge between them growing quadratically as the $\tfrac{1}{2}\sigma^2$ correction balloons.

Worked numeric example. Take $\mu = 8\%$ and $\sigma = 40\%$. The correction is $\tfrac{1}{2}\sigma^2 = \tfrac{1}{2}(0.40)^2 = \tfrac{1}{2}(0.16) = 0.08 = 8\%$. So the log-price drift is $\mu - \tfrac{1}{2}\sigma^2 = 0.08 - 0.08 = 0$. Read that again: the price is expected to grow at 8% a year, yet its log drifts at 0% — the typical (median) path is expected to go nowhere. The price’s average is propped up by a few lucky exponential winners, but the median investor, riding the typical path, treads water. That entire gap is the volatility tax, and Itô’s lemma is what put a number on it.

Geometric Brownian motion as the solution

Now everything snaps together. The SDE for a stock price is geometric Brownian motion:

$dS_t = \mu S_t\,dt + \sigma S_t\,dW_t$

Drift and diffusion both scale with the price $S_t$ — a percentage move, not a fixed-dollar move, which is exactly why prices stay positive and behave the same on a $5 stock and a $5000 stock. Solving this SDE (apply Itô’s lemma to $\log S$, then integrate the constant-drift result) gives the closed form:

$S_t = S_0 \exp\!\Big(\big(\mu - \tfrac{1}{2}\sigma^2\big)t + \sigma W_t\Big)$

Look at what’s inside the exponential. The log-price $\log S_t = \log S_0 + (\mu - \tfrac{1}{2}\sigma^2)t + \sigma W_t$ is just arithmetic Brownian motion with drift $\mu - \tfrac{1}{2}\sigma^2$ — the same correction the chart showed. And because $\log S_t$ is normal (a constant plus a scaled Wiener value), the price $S_t$ is lognormal: always positive, right-skewed, with a long upside tail and a floor at zero. That’s precisely the fan of paths you generated in Monte Carlo — now you know it falls straight out of the SDE.

Worked example — mean versus median. For a lognormal price, the expected price grows like $e^{\mu t}$ (the average, lifted by the lucky right-tail winners), but the median price grows like $e^{(\mu - \tfrac{1}{2}\sigma^2)t}$ (the typical path). With $\mu = 8\%$, $\sigma = 40\%$ over one year: the expected price multiplies by $e^{0.08} \approx 1.083$ (up 8.3%), while the median multiplies by $e^{0} = 1.000$ (flat). The mean rises while the median stays put — the unmistakable signature of a right-skewed lognormal, and the reason “expected return” and “what you’ll typically get” are not the same number.

The volatility drag, restated

Step back from the calculus and the same $\tfrac{1}{2}\sigma^2$ shows up wearing a different hat: volatility drag, the gap between the arithmetic mean return and the geometric (compounded) growth rate. They are the same number arriving from two directions.

The connection. Your wealth compounds at the geometric mean, not the arithmetic mean — and the geometric mean is lower by approximately $\tfrac{1}{2}\sigma^2$. So a stock with an 8% arithmetic average return and 40% volatility compounds at only about $8\% - 8\% = 0\%$. The arithmetic average is the drift $\mu$; the rate your money actually grows at is $\mu - \tfrac{1}{2}\sigma^2$ — the log-drift, the median path. This is the exact same ledger you’ll see again in Kelly bet-sizing and CAGR: maximizing long-run growth means maximizing $\mu - \tfrac{1}{2}\sigma^2$, not $\mu$.

Pitfall 1 — quoting the arithmetic mean and expecting to compound it. People average a fund’s yearly returns, get (say) 8%, and assume $100 grows to $108-ish a year on average over the long haul. It won’t. Volatility silently skims $\tfrac{1}{2}\sigma^2$ off every period’s compounding. The more a return series bounces, the wider the gap between the cheerful arithmetic average they quote and the duller geometric rate their account balance actually tracks.

Pitfall 2 — treating Itô’s correction as a “fudge factor.” The $\tfrac{1}{2}\sigma^2$ is not an arbitrary haircut someone bolted on to make the math work. It is the necessary second-order term that appears whenever you push a random process through a curved (nonlinear) transform — here, the log. On a linear transform ($f'' = 0$) the correction is exactly zero; it only bites on curvature. Far from being a fudge, it’s the part of the answer ordinary calculus is structurally incapable of seeing.

Putting it together

A stochastic differential equation $dX = \mu\,dt + \sigma\,dW$ is a recipe for the next infinitesimal move: a deterministic drift ($\propto dt$) plus a random diffusion ($\propto dW$) — a steady current plus gusty wind. Ordinary calculus breaks on a Brownian path because $(dW)^2 = dt$ refuses to vanish, so the second-order Taylor term survives. Itô’s lemma is the chain rule that keeps that term, adding $\tfrac{1}{2}\sigma^2 f''\,dt$ to the drift. The headline: a price with drift $\mu$ has a log-price that drifts at $\mu - \tfrac{1}{2}\sigma^2$. Solving the GBM SDE $dS = \mu S\,dt + \sigma S\,dW$ gives $S_t = S_0\exp((\mu - \tfrac{1}{2}\sigma^2)t + \sigma W_t)$ — log-price is arithmetic Brownian motion, so the price is lognormal, with the mean ($e^{\mu t}$) riding above the median ($e^{(\mu - \frac{1}{2}\sigma^2)t}$). And that same $\tfrac{1}{2}\sigma^2$ is volatility drag: the tax compounding pays for bouncing around.

Next, in mean reversion and jumps, we leave the constant-drift world: the Ornstein–Uhlenbeck SDE adds a pull back toward a long-run level (perfect for interest rates and spreads), and Merton’s jump-diffusion bolts sudden gaps onto smooth GBM to capture crashes and fat tails — both written as SDEs you now know how to read.