Radon-Nikodym Theorem

Motivation: why this matters in quant finance

The Radon-Nikodym theorem is the measure-theoretic engine that underlies every change of measure in quantitative finance. It is the fact that, whenever two probability measures

\mathbb{P}

and

\mathbb{Q}

on the same

\sigma

-algebra agree on "what is possible" (absolute continuity), there exists a non-negative random variable

Z = d\mathbb{Q}/d\mathbb{P}

such that

\mathbb{E}^{\mathbb{Q}}[X] = \mathbb{E}^{\mathbb{P}}[Z\cdot X]

for every integrable

X

. The function

Z

is the Radon-Nikodym derivative, the density of

\mathbb{Q}

with respect to

\mathbb{P}

Concrete quant-finance applications:

Risk-neutral pricing. The fundamental theorem of asset pricing guarantees (under no-arbitrage) the existence of an equivalent martingale measure $\mathbb{Q}$ . The Radon-Nikodym derivative $d\mathbb{Q}/d\mathbb{P}$ encodes the "risk-neutralisation" that turns real-world drift into the risk-free rate.
Girsanov's theorem. The specific formula for $d\mathbb{Q}/d\mathbb{P}$ when changing the drift of a Brownian motion — a Doléans-Dade exponential. Radon-Nikodym is the existence statement; Girsanov is the constructive formula.
Importance sampling in Monte Carlo. To price a deep OTM option, sample paths under $\mathbb{Q}$ with higher probability in the interesting region, then re-weight by $d\mathbb{Q}/d\mathbb{P}$ to keep the answer unbiased. Variance reduction; classical variance-reduction trick.
Esscher transforms. In actuarial and insurance pricing, the Esscher transform $d\mathbb{Q}/d\mathbb{P} \propto e^{\theta X}$ provides a tractable family of pricing measures for jump processes and Lévy models.
Density conversions. Going from a cumulative-distribution statement to a density statement, or from one parameterisation of a distribution to another, is always a Radon-Nikodym computation in disguise.

The theorem sits at an abstract level but its consequences are bread-and-butter. Every time you write " $\mathbb{E}^{\mathbb{Q}}[\cdot] = \mathbb{E}^{\mathbb{P}}[Z\cdot]$ ", you are invoking Radon-Nikodym.

The informal idea

Two probability measures

\mathbb{P}

and

\mathbb{Q}

on a measurable space

(\Omega, \mathcal{F})

can have very different probabilities assigned to each event. But if

\mathbb{Q}

is absolutely continuous with respect to

\mathbb{P}

(written

\mathbb{Q} \ll \mathbb{P}

) — meaning every

\mathbb{P}

-null event is also a

\mathbb{Q}

-null event — then

\mathbb{Q}

can be encoded by a density

Z: \Omega \to [0, \infty)

relative to

\mathbb{P}

\mathbb{Q}(A) = \int_A Z\,d\mathbb{P} = \mathbb{E}^{\mathbb{P}}[Z\cdot \mathbf{1}_A] \quad \forall A \in \mathcal{F}.

The function

Z

is the density of

\mathbb{Q}

relative to

\mathbb{P}

. It's the same mental model as a continuous density

f_X

relative to Lebesgue measure — but generalised to arbitrary measures.

Why absolute continuity? If some

\mathbb{P}

-null set

A

has

\mathbb{Q}(A) > 0

, then no density

Z

relative to

\mathbb{P}

can produce that mass:

\int_A Z\,d\mathbb{P} = 0

because

\mathbb{P}(A) = 0

. So absolute continuity is the exact condition needed for

Z

to exist.

Formal statement

Radon-Nikodym Theorem

Let

(\Omega, \mathcal{F})

be a measurable space and let

\mu, \nu

be $\sigma$ -finite measures on

\mathcal{F}

with

\nu \ll \mu

(i.e.

\mu(A) = 0 \Rightarrow \nu(A) = 0

). Then there exists a measurable function

f: \Omega \to [0, \infty)

, called the Radon-Nikodym derivative of

\nu

with respect to

\mu

, such that:

\nu(A) = \int_A f\,d\mu \quad \forall A \in \mathcal{F}.

The function $f$ is unique up to $\mu$ -null sets. We write $f = d\nu/d\mu$ .

In the probabilistic setting

\mathbb{P}, \mathbb{Q}

are probability measures on

(\Omega, \mathcal{F})

with

\mathbb{Q} \ll \mathbb{P}

, then there exists a non-negative random variable

Z

with

\mathbb{E}^{\mathbb{P}}[Z] = 1

such that:

\mathbb{Q}(A) = \mathbb{E}^{\mathbb{P}}[Z\cdot \mathbf{1}_A] \quad \forall A \in \mathcal{F}.

Equivalently, for any random variable $X \ge 0$ (or $X \in L^1(\mathbb{Q})$ ):

\mathbb{E}^{\mathbb{Q}}[X] = \mathbb{E}^{\mathbb{P}}[Z\cdot X].

$Z$ is denoted $d\mathbb{Q}/d\mathbb{P}$ .

Equivalence of measures. If additionally

\mathbb{P} \ll \mathbb{Q}

, we say

\mathbb{P}

and

\mathbb{Q}

are equivalent (write

\mathbb{P} \sim \mathbb{Q}

). In this case

Z > 0

a.s. and

d\mathbb{P}/d\mathbb{Q} = 1/Z

Standard properties

Property 1 — Chain rule

If $\mathbb{R} \ll \mathbb{Q} \ll \mathbb{P}$ , then:

\frac{d\mathbb{R}}{d\mathbb{P}} = \frac{d\mathbb{R}}{d\mathbb{Q}}\cdot \frac{d\mathbb{Q}}{d\mathbb{P}} \quad \mathbb{P}\text{-a.s.}

So densities compose by multiplication — the natural calculus rule.

Property 2 — Inverse

If $\mathbb{Q} \sim \mathbb{P}$ (equivalence), then $Z = d\mathbb{Q}/d\mathbb{P}$ is strictly positive $\mathbb{P}$ -a.s. and:

\frac{d\mathbb{P}}{d\mathbb{Q}} = \frac{1}{Z} \quad \mathbb{Q}\text{-a.s.}

Property 3 — Conditional expectation under change of measure (Bayes' formula)

If $\mathbb{Q} \ll \mathbb{P}$ with density $Z$ , and $\mathcal{G} \subset \mathcal{F}$ is a sub- $\sigma$ -algebra, then for any $X \in L^1(\mathbb{Q})$ :

\mathbb{E}^{\mathbb{Q}}[X \mid \mathcal{G}] = \frac{\mathbb{E}^{\mathbb{P}}[Z\cdot X \mid \mathcal{G}]}{\mathbb{E}^{\mathbb{P}}[Z \mid \mathcal{G}]} \quad \mathbb{Q}\text{-a.s.}

This Bayes formula for conditional expectations is the workhorse of change-of-measure pricing. It converts

\mathbb{Q}

-conditional expectations into

\mathbb{P}

-conditional expectations weighted by

Z

Proof sketch

The proof is not elementary but the idea is constructive. Define:

f := \sup\{g: g \ge 0 \text{ measurable}, \int_A g\,d\mu \le \nu(A)\ \forall A\}.

Show that the supremum is attained (MCT on an increasing sequence achieving the supremum). Show that $\int_A f\,d\mu = \nu(A)$ — if equality failed for some $A$ (strict inequality), one could find a "larger" candidate, contradicting supremum maximality. $\sigma$ -finiteness ensures this construction doesn't break due to infinite measure issues.

The full proof uses the Hahn decomposition of signed measures or the Jordan decomposition. See any graduate text (Folland, Royden, Billingsley).

Canonical examples

Example 1 — Densities on $\mathbb{R}$

If $X$ has density $f_X$ with respect to Lebesgue measure on $\mathbb{R}$ , then the law of $X$ , $\mu_X(B) := \mathbb{P}(X \in B)$ , is absolutely continuous w.r.t. Lebesgue measure, and the density is $f_X$ :

\frac{d\mu_X}{d(\text{Lebesgue})} = f_X.

This is the only Radon-Nikodym derivative most students have seen, disguised as "density."

Example 2 — Exponential tilting (Esscher transform)

Let $X \sim \mathcal{N}(0, 1)$ under $\mathbb{P}$ . Define $\mathbb{Q}$ by:

\frac{d\mathbb{Q}}{d\mathbb{P}} = Z := \exp\!\left(\theta X - \tfrac{1}{2}\theta^2\right).

Note $\mathbb{E}^{\mathbb{P}}[Z] = e^{-\theta^2/2}\cdot \mathbb{E}^{\mathbb{P}}[e^{\theta X}] = e^{-\theta^2/2}\cdot e^{\theta^2/2} = 1$ . ✓ ( $\mathbb{Q}$ is a probability measure.)

Compute the distribution of $X$ under $\mathbb{Q}$ :

\mathbb{Q}(X \le x) = \mathbb{E}^{\mathbb{P}}[Z\cdot \mathbf{1}_{X \le x}] = \int_{-\infty}^x e^{\theta t - \theta^2/2}\cdot \frac{1}{\sqrt{2\pi}}e^{-t^2/2}\,dt.

The integrand is $\frac{1}{\sqrt{2\pi}}\exp(-(t - \theta)^2/2)$ — density of $\mathcal{N}(\theta, 1)$ . So under $\mathbb{Q}$ , $X \sim \mathcal{N}(\theta, 1)$ .

Interpretation. Multiplying the

\mathbb{P}

-density by

e^{\theta X - \theta^2/2}

shifts the mean from

0

\theta

without changing the variance. This is the discrete-time Girsanov theorem for a single normal random variable: change of drift for gaussian variables is an explicit exponential tilt.

Example 3 — Change of measure in a one-period binomial model

A stock goes up to $uS_0$ with probability $p$ or down to $dS_0$ with probability $1 - p$ under $\mathbb{P}$ . The risk-free rate over one period is $r$ . The risk-neutral probability is $q := (e^r - d)/(u - d)$ .

The Radon-Nikodym derivative is:

\frac{d\mathbb{Q}}{d\mathbb{P}}(\omega) = \begin{cases} q/p & \omega = \text{up} \\ (1 - q)/(1 - p) & \omega = \text{down} \end{cases}.

Check: $\mathbb{E}^{\mathbb{P}}[d\mathbb{Q}/d\mathbb{P}] = p\cdot q/p + (1-p)\cdot (1-q)/(1-p) = q + 1 - q = 1$ . ✓

The Radon-Nikodym derivative converts the real-world probability $p$ (which reflects drift and risk preferences) to the risk-neutral probability $q$ (which reflects no-arbitrage pricing).

Example 4 — Girsanov sketch

If $(W_t)$ is a $\mathbb{P}$ -Brownian motion and $\theta$ is a (sufficiently regular) process, define:

Z_t := \exp\!\left(-\int_0^t \theta_s\,dW_s - \tfrac{1}{2}\int_0^t \theta_s^2\,ds\right) \quad\text{(a Doléans-Dade exponential)}.

Under suitable integrability (Novikov's condition),

Z_t

is a

\mathbb{P}

-martingale with

\mathbb{E}^{\mathbb{P}}[Z_t] = 1

, so

\mathbb{Q}_t

defined by

d\mathbb{Q}_t/d\mathbb{P} = Z_t

is a probability measure. The Girsanov theorem says that under

\mathbb{Q}_t

, the process

\tilde W_t := W_t + \int_0^t \theta_s\,ds

is a Brownian motion.

This is the machinery that enables changing from the real-world measure to the risk-neutral measure in continuous-time models. Full treatment in the Girsanov's Theorem lesson (if authored).

Common pitfalls

" $d\mathbb{Q}/d\mathbb{P}$ is a single number." No — it's a random variable (or function on

\Omega

). The density varies depending on the

\omega

"Radon-Nikodym requires $\mathbb{P}$ and $\mathbb{Q}$ to have densities." No — the theorem constructs the density. The requirement is absolute continuity of one with respect to the other.

"If $\mathbb{P}$ and $\mathbb{Q}$ are different, one must be absolutely continuous w.r.t. the other." False. Counter-example: take

\mathbb{P}

supported on

\{0, 1\}

and

\mathbb{Q}

supported on

\{2, 3\}

. Each is singular w.r.t. the other; no Radon-Nikodym derivative exists in either direction. (Lebesgue's decomposition theorem handles the general case: any

\mathbb{Q}

can be written as

\mathbb{Q} = \mathbb{Q}_{\text{ac}} + \mathbb{Q}_{\text{sing}}

where

\mathbb{Q}_{\text{ac}} \ll \mathbb{P}

and

\mathbb{Q}_{\text{sing}} \perp \mathbb{P}

" $\mathbb{E}^{\mathbb{Q}}[Z] = 1$ ."

\mathbb{E}^{\mathbb{P}}[Z] = 1

, not under

\mathbb{Q}

. Under

\mathbb{Q}

\mathbb{E}^{\mathbb{Q}}[Z] = \mathbb{E}^{\mathbb{P}}[Z^2]

— which can differ from

1

"Girsanov is the same as Radon-Nikodym." Girsanov is a specific instance of Radon-Nikodym: it provides the explicit formula for

d\mathbb{Q}/d\mathbb{P}

when changing the drift of a Brownian motion. Radon-Nikodym is the general abstract existence theorem.

Where this goes next

Girsanov's Theorem: The continuous-time change-of-drift formula, the single most important instance of Radon-Nikodym in quant finance.
Martingale Measures: Risk-neutral pricing relies on Radon-Nikodym's guarantee of existence of an equivalent martingale measure under no-arbitrage.
Conditional Expectation: Bayes' formula for conditional expectations under change of measure uses Radon-Nikodym explicitly.
Importance Sampling: Monte Carlo variance-reduction via Radon-Nikodym tilts (a specific application of the change-of-measure idea).
The Derivation of the Black-Scholes Formula: The transition from the real-world SDE to the risk-neutral SDE uses the Radon-Nikodym derivative from Girsanov's theorem.

Radon-Nikodym Theorem

Motivation: why this matters in quant finance

The informal idea

Formal statement

Radon-Nikodym Theorem

In the probabilistic setting

Standard properties

Property 1 — Chain rule

Property 2 — Inverse

Property 3 — Conditional expectation under change of measure (Bayes' formula)

Proof sketch

Canonical examples

Example 1 — Densities on R\mathbb{R}R

Example 2 — Exponential tilting (Esscher transform)

Example 3 — Change of measure in a one-period binomial model

Example 4 — Girsanov sketch

Common pitfalls

Where this goes next

Exercises

Example 1 — Densities on $\mathbb{R}$