Girsanov's Theorem
Girsanov's theorem is the drift-ectomy. A stock's real-world dynamics carry
an awkward drift \mu — every investor's private, unknowable guess
about the future. Girsanov is the surgical tool that removes it: reweight the probabilities of
the paths by exactly the right exponential factor, and a Brownian motion with drift
becomes a Brownian motion without one. No path is bent, no wiggle is smoothed — the
operation touches only the weights — and yet the world you wake up in after surgery
is the risk-neutral one where everything grows at the riskless rate and options have prices.
The
equivalent martingale measure
was promised but never built. Girsanov's theorem is the construction. It answers a precise
question: if I reweight my probabilities by a clever density, what happens to a
Brownian motion? The astonishing answer is that
changing measure changes the drift, and nothing else. The randomness — the Brownian
wiggle — survives untouched; only the tilt of the path shifts.
Concretely, let W be a Brownian motion under
\mathbb{P}, fix an adapted process
\theta_t, and use as the
density
the
exponential martingale
Z_T = \frac{d\mathbb{Q}}{d\mathbb{P}} = \exp\!\left(-\int_0^T \theta_s\, dW_s - \tfrac{1}{2}\int_0^T \theta_s^2\, ds\right).
Then under the new measure \mathbb{Q} the shifted process
\tilde{W}_t = W_t + \int_0^t \theta_s\, ds
is itself a Brownian motion. The drift \int_0^t \theta_s\,ds that
\mathbb{Q} grafts on under \mathbb{P} is
exactly the drift it strips off going the other way: under
\mathbb{Q}, \tilde{W} is driftless and
clean. In the workhorse case \theta is a constant and
everything collapses to three memorable lines:
Z_t = e^{-\theta W_t - \theta^2 t/2}, then
d\mathbb{Q} = Z_T\,d\mathbb{P}, then
\tilde{W}_t = W_t + \theta t is
\mathbb{Q}-Brownian. That constant-\theta
version is the one we now put to work, twice: first on a bare drifting Brownian motion, then
on the stock itself.
The simplest surgery: killing a constant drift
Before the stock, operate on the simplest patient. Under
\mathbb{P}, let
X_t = \mu t + \sigma W_t
— a Brownian motion scaled by \sigma and tilted by a constant drift
\mu. We want a measure under which X is
driftless.
Choose the dose. Take \theta = \mu/\sigma. Then the
Girsanov-shifted process is
\tilde{W}_t = W_t + \theta t = W_t + (\mu/\sigma)\,t, and substituting
W_t = \tilde{W}_t - (\mu/\sigma)t into X:
X_t = \mu t + \sigma\!\left(\tilde{W}_t - \tfrac{\mu}{\sigma}t\right) = \mu t - \mu t + \sigma \tilde{W}_t = \sigma \tilde{W}_t.
Under \mathbb{Q}, X is a pure scaled
Brownian motion — the drift has been excised completely. But this only counts if
\tilde{W} genuinely is a Brownian motion under
\mathbb{Q}. Check the axioms:
-
Starts at zero, continuous paths. Free: \tilde{W}_0 = W_0 = 0,
and \tilde{W} is a continuous path plus a straight line. Changing
measure never edits a path, so these path properties carry over untouched.
-
Gaussian increments with the right law. This is where the density earns its
keep — see the computation below.
-
Independent increments. Follows from the martingale structure of
Z (the honest proof runs through Lévy's characterisation of
Brownian motion: continuous martingale + quadratic variation t
⇒ Brownian motion — we take it on trust here).
The key computation — complete the square. Under
\mathbb{P}, W_T \sim N(0, T) with density
\varphi(w) = \tfrac{1}{\sqrt{2\pi T}}e^{-w^2/2T}. Under
\mathbb{Q}, every outcome is reweighted by
Z_T = e^{-\theta w - \theta^2 T/2}, so the
\mathbb{Q}-density of W_T is the product:
e^{-\theta w - \frac{\theta^2 T}{2}}\cdot\frac{1}{\sqrt{2\pi T}}\,e^{-\frac{w^2}{2T}} = \frac{1}{\sqrt{2\pi T}}\,\exp\!\left(-\frac{w^2 + 2\theta T w + \theta^2 T^2}{2T}\right) = \frac{1}{\sqrt{2\pi T}}\,e^{-\frac{(w + \theta T)^2}{2T}}.
That is exactly the N(-\theta T,\, T) density: under
\mathbb{Q}, W_T is centred at
-\theta T — so the shifted variable
\tilde{W}_T = W_T + \theta T \sim N(0, T), precisely what a Brownian
motion demands at time T. The exponential weight
e^{-\theta w} boosts the outcomes where W
came out low and shrinks the ones where it came out high — recentring the bell curve
by pure reweighting, with the variance (the width of the bell) left exactly where it was.
The centrepiece: sending the stock risk-neutral, line by line
Now the operation the whole theory was built for: turn the real-world stock
dynamics into the risk-neutral dynamics. Start with
geometric Brownian motion
under \mathbb{P}, where the stock drifts at its real rate
\mu:
dS_t = \mu S_t\, dt + \sigma S_t\, dW_t.
Step 1 — invert the Girsanov shift. Take
\theta constant. The relation
\tilde{W}_t = W_t + \theta t differentiates to
d\tilde{W}_t = dW_t + \theta\,dt, so the
\mathbb{P}-Brownian increment is
dW_t = d\tilde{W}_t - \theta\, dt.
Step 2 — substitute into the SDE. Replace every
dW_t with d\tilde{W}_t - \theta\,dt:
dS_t = \mu S_t\, dt + \sigma S_t\,\big(d\tilde{W}_t - \theta\, dt\big).
Step 3 — expand and collect the dt terms. Multiply
out the bracket and gather the two drift pieces:
dS_t = \mu S_t\, dt - \sigma\theta S_t\, dt + \sigma S_t\, d\tilde{W}_t = (\mu - \sigma\theta)\,S_t\, dt + \sigma S_t\, d\tilde{W}_t.
The diffusion coefficient \sigma S_t is unchanged — only
the drift moved, exactly as Girsanov advertised. The new drift is a knob we can turn by our
choice of \theta.
Step 4 — choose \theta to hit the riskless rate. We
want the stock to drift at r under
\mathbb{Q}, so demand
\mu - \sigma\theta = r and solve for
\theta:
\theta = \frac{\mu - r}{\sigma}.
This special \theta is the market price of risk —
the excess return \mu - r per unit of volatility
\sigma. One Greek letter, and it is the same one for every asset
driven by this Brownian motion: in an arbitrage-free market all such assets must offer the same
compensation per unit of risk, or you could sell the stingy one and buy the generous one.
Step 5 — verify the drift collapses to r. Substitute
\theta = (\mu - r)/\sigma back into the drift coefficient and watch
the \sigma cancel:
\mu - \sigma\theta = \mu - \sigma\cdot\frac{\mu - r}{\sigma} = \mu - (\mu - r) = r.
Step 6 — read off the risk-neutral dynamics. The drift is now exactly the
riskless rate:
\boxed{\,dS_t = r S_t\, dt + \sigma S_t\, d\tilde{W}_t.\,}
Under \mathbb{Q} the stock earns r on
average — the real-world premium \mu has been reweighted clean away
— and \tilde{W} is the new Brownian motion. The discounted price
\tilde{S}_t = e^{-rt}S_t is then a
\mathbb{Q}-martingale, so \mathbb{Q} is
the equivalent martingale measure we were missing. Black–Scholes pricing is now a matter of
computing one expectation under \mathbb{Q} — and notice what has
quietly happened to \mu: it appears nowhere in the boxed
SDE. The one parameter nobody can agree on has been surgically removed from the pricing
problem. That, more than anything, is why option prices are objective while stock forecasts
are opinions.
Let W be a \mathbb{P}-Brownian motion
and \theta_t an adapted process (with
\int_0^T \theta_s^2\,ds < \infty). Then:
-
the density
Z_T = \exp\!\big(-\int_0^T \theta_s\,dW_s - \tfrac12\int_0^T \theta_s^2\,ds\big)
is the exponential martingale; it is non-negative with
\mathbb{E}_{\mathbb{P}}[Z_T] = 1, so
d\mathbb{Q} = Z_T\, d\mathbb{P} defines a measure
\mathbb{Q} \sim \mathbb{P}.
-
for constant \theta this reads
Z_t = \exp\!\big(-\theta W_t - \tfrac12\theta^2 t\big).
-
under \mathbb{Q}, the shifted process
\tilde{W}_t = W_t + \int_0^t \theta_s\,ds
(= W_t + \theta t for constant \theta)
is a Brownian motion.
-
the change of measure shifts the drift by
\sigma\theta and leaves the diffusion
\sigma untouched.
-
choosing the market price of risk
\theta = (\mu - r)/\sigma makes the discounted stock
e^{-rt}S_t a \mathbb{Q}-martingale.
The quantity \theta = (\mu - r)/\sigma is the market price
of risk (or Sharpe ratio): how much extra expected return the market pays for each
extra unit of volatility carried. Girsanov reweights probability until that price is exactly
zero — under \mathbb{Q} risk is not compensated, which is why
everything earns the same r.
One technical hazard: the density Z_T is only a
local martingale in general, and could secretly lose mass
(\mathbb{E}[Z_T] < 1), which would break the change of measure.
Novikov's condition rules this out — if
\mathbb{E}_{\mathbb{P}}\!\left[\exp\!\left(\tfrac12\int_0^T \theta_s^2\, ds\right)\right] < \infty,
then Z_T is a genuine martingale with
\mathbb{E}[Z_T] = 1 and Girsanov goes through. For constant
\theta the integral is just
\tfrac12\theta^2 T, finite always — so the textbook
Black–Scholes case is safe.
Finally, recognise the density itself. With
X_t = -\int_0^t \theta_s\,dW_s, its quadratic variation is
\langle X\rangle_t = \int_0^t \theta_s^2\,ds, and
Z_t = \exp\!\big(X_t - \tfrac12\langle X\rangle_t\big)
is precisely the stochastic exponential of
-\int\theta\,dW — the very
exponential martingale
\exp(\sigma W_t - \tfrac12\sigma^2 t) we built with
Itô's lemma, now with the integrand
-\theta in the role of \sigma. The
-\tfrac12\int\theta^2 ds in the exponent is the same drift
compensator, doing the same job: keeping the mean pinned at
1 so that Z_T is a legitimate density.
Why the -\tfrac12\theta^2 t in the exponent?
It is not decoration — it is the normalisation that makes
Z_t a legitimate density. A density must average to
1: total probability is conserved, only shuffled between outcomes.
But the naive exponential tilt e^{-\theta W_t} does not
average to 1. Since -\theta W_t \sim N(0, \theta^2 t),
the tilt is lognormal, and the lognormal mean formula
\mathbb{E}[e^X] = e^{m + s^2/2} (for
X \sim N(m, s^2)) gives
\mathbb{E}_{\mathbb{P}}\!\left[e^{-\theta W_t}\right] = e^{0 + \theta^2 t/2} = e^{\theta^2 t/2} > 1.
The convexity of e^x means the exponential of a symmetric wiggle
gains more on the upside than it loses on the downside — Jensen's inequality made concrete.
The tilt manufactures extra mass at the rate e^{\theta^2 t/2}, so we
must divide it back out:
Z_t = \frac{e^{-\theta W_t}}{\mathbb{E}[e^{-\theta W_t}]} = e^{-\theta W_t}\cdot e^{-\theta^2 t/2} = e^{-\theta W_t - \theta^2 t/2}, \qquad \mathbb{E}_{\mathbb{P}}[Z_t] = 1.
You have met this correction before, wearing a different hat: it is the same
-\tfrac12\sigma^2 t that appears in the GBM solution
S_t = S_0 e^{(\mu - \sigma^2/2)t + \sigma W_t}, and the same
+\sigma^2/2 in the lognormal mean
e^{\mu + \sigma^2/2}. Wherever a Gaussian is fed through an
exponential, the half-variance term shows up to balance the books — here it pins
\mathbb{E}[Z_t] at exactly 1 for
every t, which is what makes Z a
martingale rather than a mere positive process.
Sanity check: reweighting on two outcomes
The formula d\mathbb{Q} = Z_T\,d\mathbb{P} is used through one
identity — for any payoff X,
\mathbb{E}_{\mathbb{Q}}[X] = \mathbb{E}_{\mathbb{P}}[Z_T\, X].
Strip away the continuum and check it on a toy world with just two outcomes,
up and down, each with \mathbb{P}-probability
\tfrac12. Choose the density values
Z(\text{up}) = 1.6 and Z(\text{down}) = 0.4
— note \mathbb{E}_{\mathbb{P}}[Z] = \tfrac12(1.6) + \tfrac12(0.4) = 1,
the discrete stand-in for the -\theta^2t/2 normalisation. The new
probabilities are the old ones times the density:
\mathbb{Q}(\text{up}) = \tfrac12 \times 1.6 = 0.8, \qquad \mathbb{Q}(\text{down}) = \tfrac12 \times 0.4 = 0.2.
Now price a payoff worth 10 on up and 0
on down, both ways:
\mathbb{E}_{\mathbb{Q}}[X] = 0.8 \times 10 + 0.2 \times 0 = 8, \qquad \mathbb{E}_{\mathbb{P}}[Z X] = \tfrac12(1.6 \times 10) + \tfrac12(0.4 \times 0) = 8.
Same answer, as it must be — computing a \mathbb{Q}-expectation
is computing a Z-weighted
\mathbb{P}-expectation. Girsanov's continuous version does exactly
this, outcome by outcome, with Z_T = e^{-\theta W_T - \theta^2 T/2}
playing the role of the numbers 1.6 and
0.4: paths where W wandered low get
weight above 1, paths where it soared get weight below, and the
Gaussian recentres.
-
Girsanov changes the drift only — the volatility \sigma
is measure-invariant. Quadratic variation is a path property: it is
computed wiggle by wiggle along a single realised path, with no probabilities in sight. A
change of measure reweights how likely each path is — it never edits a path — so anything
you can read off one path, \sigma included, cannot move. This is
why implied volatility is meaningfully comparable to historical volatility: they live under
different measures, but \sigma is the same object under both.
-
Novikov's condition guards the theorem. The exponential
Z must be a true martingale with
\mathbb{E}[Z_T] = 1; a wild enough
\theta makes it a strict local martingale that leaks mass, and
then \mathbb{Q} is not a probability measure at all. Constant
\theta always passes; exotic stochastic-volatility models must
be checked.
-
Under \mathbb{Q} the world isn't "wrong." The
same paths exist, the same events are possible and impossible
(\mathbb{Q} \sim \mathbb{P} — that's equivalence) — only the
weights differ. Nobody claims stocks really drift at r;
\mathbb{Q} is an accounting device under which fair prices
become plain expectations. Forecast under \mathbb{P}, price
under \mathbb{Q} — mixing them up is the classic quant blunder.
Igor Vladimirovich Girsanov (1934–1967) was one of the brightest of the postwar Moscow
probabilists — a student in the circle of Dynkin at Moscow State University, working on
diffusion processes and stochastic differential equations. The theorem that carries his name
appeared in his 1960 paper on transforming stochastic processes by an
absolutely continuous substitution of measures — building on earlier work of
Cameron and Martin, who had handled the deterministic-shift case in the 1940s. To Girsanov
it was a clean piece of pure mathematics about Wiener space. Finance is mentioned nowhere.
He never learned what it would become. A passionate mountaineer and ski instructor, Girsanov
died in March 1967, aged just 32, in an avalanche while leading a ski expedition in the Sayan
Mountains. Black–Scholes–Merton was six years away; Harrison, Kreps and Pliska's
martingale-pricing framework — the work that made "apply Girsanov, price under
\mathbb{Q}" the daily grammar of quantitative finance — a decade
beyond that. Today his theorem runs silently inside the pricing engines of derivative markets
measured in the hundreds of trillions of dollars notional: arguably the largest gap ever
between what a mathematician lived to see and what his theorem went on to do.
One outcome, two drifts
Here is Girsanov made visible. A single Brownian outcome
\omega drives both curves: the
\mathbb{P} stock climbing at its real drift
\mu, and the \mathbb{Q} stock climbing at
the riskless r. They share every wiggle — only the average tilt
differs, the difference being \sigma\theta. Drag
\theta to retune the market price of risk and watch the
\mathbb{Q} path's drift slide; Refresh for a fresh
\omega.
Two things to notice as you play. First, at \theta = 0 the curves
coincide exactly — no reweighting, no surgery. Second, however far you push
\theta, the two curves never differ in texture: every kink,
every micro-wiggle, every burst of volatility is identical, because quadratic variation is a
path property that no change of measure can touch. What you are steering is not the path but
the exponential weight Z hanging invisibly on it — this one
\omega is more probable under one measure than the other, and the
drift you see is the population-level echo of that reweighting across all the
\omegas you'd get by pressing Refresh forever.