The Radon–Nikodym Derivative
Two people can disagree about probabilities without disagreeing about what is
possible. A pessimist and an optimist may price the same coin differently, yet both
agree the coin can land heads and can land tails. When two
measures
\mathbb{P} and \mathbb{Q} agree on
exactly which events are impossible — the same null sets — we call them
equivalent and write \mathbb{P} \sim \mathbb{Q}.
Equivalence is the precise licence to translate one measure into the other. The dictionary is
a single non-negative random variable, the density (or
Radon–Nikodym derivative)
Z = \frac{d\mathbb{Q}}{d\mathbb{P}} \ge 0,
which reweights \mathbb{P} into
\mathbb{Q} outcome by outcome. Where Z > 1
the optimist counts an outcome as more likely than the pessimist; where
0 < Z < 1, less likely. Formally,
\mathbb{Q} measures a set A by summing
Z over it under \mathbb{P}:
\mathbb{Q}(A) = \int_A Z \, d\mathbb{P} \qquad \text{for every event } A.
It helps to think of Z as a currency exchange rate between
measures. Just as a currency converter turns a price in dollars into a price in
euros by multiplying by the day's rate, Z turns a
\mathbb{P}-expectation into a
\mathbb{Q}-expectation by multiplying, outcome by outcome, before
averaging. Every risk-neutral price you will ever compute is secretly this trick in disguise:
somewhere behind the scenes sits a density Z converting the
real-world measure into a fictional one in which discounted asset prices are fair games —
and Girsanov's theorem
is precisely the machine that manufactures Z for continuous-time
models.
From sets to expectations, derived line by line
The defining relation \mathbb{Q}(A) = \int_A Z\,d\mathbb{P} only
talks about probabilities of sets. We will upgrade it into the full
change-of-measure formula for
expectations,
\mathbb{E}_{\mathbb{Q}}[X] = \mathbb{E}_{\mathbb{P}}[XZ],
by the standard ladder of integration theory: start with indicators, climb to simple
functions, then to limits.
Step 1 — read the definition as an expectation. The integral of
Z over A is exactly the
\mathbb{P}-expectation of Z masked by the
indicator \mathbf{1}_A (which is 1 on
A and 0 off it):
\mathbb{Q}(A) = \int_A Z\,d\mathbb{P} = \mathbb{E}_{\mathbb{P}}[\mathbf{1}_A\, Z].
Step 2 — the indicator case of the formula. But the
\mathbb{Q}-probability of A is itself the
\mathbb{Q}-expectation of its indicator,
\mathbb{Q}(A) = \mathbb{E}_{\mathbb{Q}}[\mathbf{1}_A]. Comparing
with Step 1, the formula \mathbb{E}_{\mathbb{Q}}[X] = \mathbb{E}_{\mathbb{P}}[XZ]
already holds for every X = \mathbf{1}_A:
\mathbb{E}_{\mathbb{Q}}[\mathbf{1}_A] = \mathbb{E}_{\mathbb{P}}[\mathbf{1}_A\, Z].
Step 3 — extend to simple functions by linearity. A
simple random variable is a finite sum
X = \sum_{k} c_k\,\mathbf{1}_{A_k}. Both expectations are linear, so
applying Step 2 term by term,
\mathbb{E}_{\mathbb{Q}}[X] = \sum_k c_k\,\mathbb{E}_{\mathbb{Q}}[\mathbf{1}_{A_k}] = \sum_k c_k\,\mathbb{E}_{\mathbb{P}}[\mathbf{1}_{A_k} Z] = \mathbb{E}_{\mathbb{P}}\Big[\Big(\textstyle\sum_k c_k \mathbf{1}_{A_k}\Big) Z\Big] = \mathbb{E}_{\mathbb{P}}[XZ].
Step 4 — pass to the limit. Any non-negative measurable
X is an increasing limit of simple functions
X_n \uparrow X. Step 3 holds for each
X_n, and monotone convergence lets us take the limit inside both
expectations (note X_n Z \uparrow XZ since
Z \ge 0):
\mathbb{E}_{\mathbb{Q}}[X] = \lim_{n\to\infty} \mathbb{E}_{\mathbb{Q}}[X_n] = \lim_{n\to\infty} \mathbb{E}_{\mathbb{P}}[X_n Z] = \mathbb{E}_{\mathbb{P}}[XZ].
Splitting a general X = X^+ - X^- into positive and negative parts
extends it to any integrable X. The formula is proved.
The density integrates to one
Step 5 — take A = \Omega. The whole-space case of
the defining relation is the normalisation that pins Z down. Since
\mathbf{1}_\Omega = 1,
\mathbb{E}_{\mathbb{P}}[Z] = \int_\Omega Z\,d\mathbb{P} = \mathbb{Q}(\Omega) = 1,
because \mathbb{Q}, being a probability measure, assigns total mass
1. So a valid density is any non-negative
Z with \mathbb{P}-mean exactly
1: it is a reweighting that neither creates nor destroys total
probability. For the new measure to be equivalent (not merely absolutely continuous)
we need a touch more — Z > 0 almost surely — so that the dictionary
runs both ways and 1/Z = d\mathbb{P}/d\mathbb{Q}.
Let \mathbb{P} and \mathbb{Q} be
probability measures on (\Omega, \mathcal{F}). Then:
-
\mathbb{Q} is absolutely continuous with
respect to \mathbb{P} (written
\mathbb{Q} \ll \mathbb{P}, meaning
\mathbb{P}(A) = 0 \Rightarrow \mathbb{Q}(A) = 0) iff
there is a non-negative density
Z = d\mathbb{Q}/d\mathbb{P} with
\mathbb{Q}(A) = \int_A Z\,d\mathbb{P}. The two measures are
equivalent (\mathbb{P} \sim \mathbb{Q}) exactly
when Z > 0 almost surely.
-
The density is normalised:
\mathbb{E}_{\mathbb{P}}[Z] = 1.
-
Change of measure: for every integrable
X,
\mathbb{E}_{\mathbb{Q}}[X] = \mathbb{E}_{\mathbb{P}}[XZ].
The two notions are easy to confuse but do different jobs.
\mathbb{Q} \ll \mathbb{P} (absolute continuity) is one-directional:
anything \mathbb{P} rules out, \mathbb{Q}
rules out too — but \mathbb{Q} may have extra impossible
events of its own (those are the places where Z = 0).
Equivalence is the two-sided version,
\mathbb{Q} \ll \mathbb{P} and
\mathbb{P} \ll \mathbb{Q} together: the same null sets, no
exceptions, so Z > 0 everywhere and you can divide by it.
Statisticians know Z by another name: a
likelihood ratio. If \mathbb{P} and
\mathbb{Q} have densities p and
q against some common reference, then
Z(\omega) = \frac{d\mathbb{Q}}{d\mathbb{P}}(\omega) = \frac{q(\omega)}{p(\omega)},
the ratio of how plausible \omega is under the two hypotheses.
Every time you compute a likelihood ratio you are silently writing down a Radon–Nikodym
derivative.
Keep this dictionary in your back pocket: a "likelihood ratio," a "Radon–Nikodym derivative,"
and "the density that reweights \mathbb{P} into
\mathbb{Q}" are three names for exactly the same object.
Worked example 1 — a two-outcome coin, checked by hand
Let \Omega = \{H, T\} with
\mathbb{P}(H) = 0.6, \mathbb{P}(T) = 0.4,
and a second measure \mathbb{Q}(H) = 0.75,
\mathbb{Q}(T) = 0.25. Both measures put positive weight on both
outcomes, so \mathbb{P} \sim \mathbb{Q}.
Step 1 — build the density outcome by outcome. On a finite space the density
is simply the ratio of the two probabilities at each point:
Z(H) = \frac{\mathbb{Q}(H)}{\mathbb{P}(H)} = \frac{0.75}{0.6} = 1.25, \qquad Z(T) = \frac{\mathbb{Q}(T)}{\mathbb{P}(T)} = \frac{0.25}{0.4} = 0.625.
Step 2 — check the normalisation \mathbb{E}_{\mathbb{P}}[Z] = 1.
\mathbb{E}_{\mathbb{P}}[Z] = \mathbb{P}(H)Z(H) + \mathbb{P}(T)Z(T) = (0.6)(1.25) + (0.4)(0.625) = 0.75 + 0.25 = 1. \checkmark
Step 3 — verify the change-of-measure formula on a real random variable. Let
X(H) = 10, X(T) = 2 (say, a payoff).
Compute both sides independently:
\mathbb{E}_{\mathbb{Q}}[X] = (0.75)(10) + (0.25)(2) = 7.5 + 0.5 = 8,
\mathbb{E}_{\mathbb{P}}[XZ] = (0.6)(10)(1.25) + (0.4)(2)(0.625) = 7.5 + 0.5 = 8. \checkmark
The two computations, done completely independently, land on the same number. That is the
entire content of the theorem, seen with no machinery at all: reweighting the outcomes with
Z before averaging under \mathbb{P} gives
exactly the same answer as averaging under \mathbb{Q} directly.
Worked example 2 — Girsanov's kernel, derived in one completed square
Let \mathbb{P} give X a standard normal
density p(x) = \tfrac{1}{\sqrt{2\pi}} e^{-x^2/2}, and let
\mathbb{Q} give it the shifted density
q(x) = \tfrac{1}{\sqrt{2\pi}} e^{-(x-\theta)^2/2} — the same bell
curve, slid sideways to mean \theta. What is
Z = d\mathbb{Q}/d\mathbb{P}?
Step 1 — take the ratio of densities (the constants
1/\sqrt{2\pi} cancel):
Z(x) = \frac{q(x)}{p(x)} = \exp\!\left(-\frac{(x-\theta)^2}{2} + \frac{x^2}{2}\right).
Step 2 — expand the square in the exponent.
(x-\theta)^2 = x^2 - 2\theta x + \theta^2, so
-\frac{(x-\theta)^2}{2} + \frac{x^2}{2} = -\frac{x^2 - 2\theta x + \theta^2}{2} + \frac{x^2}{2} = \theta x - \frac{\theta^2}{2}.
Step 3 — read off the density.
Z(x) = e^{\theta x - \theta^2/2}.
This is exactly the tilt used in the chart below, and it is precisely the exponential
martingale that Girsanov's theorem
builds in continuous time to shift a Brownian motion's drift. The whole of Girsanov's theorem
is, in essence, this one completed-square calculation done at every instant along a path.
When the derivative fails to exist
The theorem's \mathbb{Q} \ll \mathbb{P} hypothesis is not red tape —
drop it and there is no density to speak of. Let \Omega = [0, 1],
let \mathbb{P} be the uniform (Lebesgue) measure, and let
\mathbb{Q} put all its probability on the single point
\{0.5\} (a point mass, or "Dirac" measure). Take
A = \{0.5\}: since a single point has zero length,
\mathbb{P}(A) = 0 — but by construction
\mathbb{Q}(A) = 1 \ne 0. A \mathbb{P}-null
set is not \mathbb{Q}-null, so
\mathbb{Q} \ll \mathbb{P} fails outright, and no function
Z with \mathbb{Q}(A) = \int_A Z\, d\mathbb{P}
can possibly exist — any such integral over a single point is 0, never
1. Some pairs of measures simply refuse to be exchanged.
Four points about the density that are easy to state and easy to misuse:
-
Absolute continuity is the entry ticket, not a footnote. No
\mathbb{Q} \ll \mathbb{P}, no density — full stop. The Dirac-versus-Lebesgue
example above is not a pathological curiosity; it is exactly the failure mode the hypothesis
is there to rule out.
-
Get the direction of "null" right.
\mathbb{Q} \ll \mathbb{P} means every
\mathbb{P}-null set is \mathbb{Q}-null —
not the other way round. It is \mathbb{P}'s impossibilities that
bind \mathbb{Q}; \mathbb{Q} is free to
declare extra events impossible of its own (exactly where Z=0).
-
Z \ge 0 and \mathbb{E}_{\mathbb{P}}[Z]=1
are not optional extras — they hold for every valid density, always,
because \mathbb{Q} is a probability measure that can't go negative
or fail to total 1. If an algebra slip produces a
Z that goes negative or averages to something other than
1 under \mathbb{P}, go back and find the
mistake.
-
"Derivative" is a suggestive name, not a literal one. Nothing is being
differentiated with respect to anything — there is no limit of difference quotients anywhere
in sight. The name is an analogy to
dx/du from ordinary calculus because
Z plays the same *role* (a local rescaling factor), but the
construction is entirely measure-theoretic.
Suppose you need to estimate the probability that a stock crashes 40% in a day, or that a
server farm gets hit by an outage during its rarest failure mode — events so unlikely under
the real-world measure \mathbb{P} that a plain Monte Carlo
simulation would need billions of draws before seeing even one. The trick, called
importance sampling, is a direct application of this page: simulate instead
from a friendlier measure \mathbb{Q} under which the rare event is
common, then undo the distortion by reweighting each simulated path with the density
running the other way,
\mathbb{E}_{\mathbb{P}}[h(X)] = \mathbb{E}_{\mathbb{Q}}\!\left[h(X)\,\frac{d\mathbb{P}}{d\mathbb{Q}}(X)\right] = \mathbb{E}_{\mathbb{Q}}\!\left[h(X)\, \frac{1}{Z(X)}\right].
Concretely: to estimate \mathbb{P}(X > 4) for a standard normal
X, don't draw a billion standard normals hoping a handful exceed
4. Draw from \mathbb{Q} = N(4, 1)
instead — now crossing 4 is an everyday event — and multiply each
hit by 1/Z(X) to convert the count back into a
\mathbb{P}-probability. Same idea, same formula, wildly fewer
simulations.
Reweighting a density, live
Take a base measure \mathbb{P} with a standard-normal density and
tilt it by an exponential factor
Z(x) = e^{\theta x - \theta^2/2} — the very density derived by
completing the square in worked example 2 above, and exactly the shape Girsanov will use. This
Z has \mathbb{P}-mean
1 (so total probability is preserved), and multiplying it onto the
\mathbb{P} bell produces the \mathbb{Q}
density — a normal shifted to mean \theta. Slide
\theta to watch the mass slosh sideways while the area stays
1.
The faint fixed curve is the original \mathbb{P} density; the bold
curve is \mathbb{Q} = Z\cdot\mathbb{P}. At
\theta = 0 the tilt is Z \equiv 1 and the
two coincide — no reweighting at all.