The Stochastic Exponential¶
In the unifying framework of this section, the stochastic exponential is the transformation enabling control — it converts additive local martingales into multiplicative densities for measure change.
The stochastic exponential (or Doleans-Dade exponential) is the stochastic analogue of the ordinary exponential function. It is the fundamental tool for constructing Radon-Nikodym derivatives in measure change, and therefore central to Girsanov's theorem and risk-neutral pricing.
Prerequisites
This section assumes familiarity with:
Definition¶
SDE Definition¶
Given a continuous semimartingale \(X_t\), the stochastic exponential \(\mathcal{E}(X)_t\) is defined as the unique solution to:
Note that the initial condition is always \(\mathcal{E}(X)_0 = 1\), regardless of the value of \(X_0\).
Explicit Formula¶
For continuous semimartingales:
where \(\langle X \rangle_t\) is the quadratic variation of \(X\).
When \(X_0 = 0\) (the most common case in applications), this simplifies to:
Why "Stochastic Exponential"?¶
Analogy with Ordinary Exponential¶
For a deterministic function \(x(t)\), the ordinary exponential \(e^{x(t)}\) satisfies:
In differential form: \(d(e^{x}) = e^x\,dx\).
The Stochastic Version¶
The stochastic exponential satisfies the same multiplicative structure:
The Crucial Difference¶
In ordinary calculus: \(e^{x(t)} = \exp\left(\int_0^t dx(s)\right) = \exp(x(t) - x(0))\)
In stochastic calculus: \(\mathcal{E}(X)_t = \exp\left(X_t - X_0 - \frac{1}{2}\langle X \rangle_t\right)\)
The term \(-\frac{1}{2}\langle X \rangle_t\) is the Itô correction arising from quadratic variation. This correction is essential: without it, the exponential would have a drift term and fail to be a local martingale.
Derivation via Itô's Lemma¶
Goal: Verify that \(Z_t = \exp(X_t - X_0 - \frac{1}{2}\langle X \rangle_t)\) satisfies \(dZ_t = Z_t\,dX_t\).
Setup¶
Let \(X_t\) be a continuous local martingale with \(dX_t = \sigma_t\,dW_t\). Then \(d\langle X \rangle_t = \sigma_t^2\,dt\).
Define \(Z_t = \exp(X_t - X_0 - \frac{1}{2}\langle X \rangle_t)\).
Apply Itô's Lemma¶
Using the Itô formula with \(f(x) = e^x\) and the process \(Y_t = X_t - X_0 - \frac{1}{2}\langle X \rangle_t\):
Now compute \(dY_t\) and \((dY_t)^2\):
(using the Itô multiplication rules: \((dW_t)^2 = dt\), \(dt \cdot dW_t = 0\), \((dt)^2 = 0\))
Substitute¶
The correction terms cancel exactly, leaving no drift. \(\square\)
Key Properties¶
Property 1: Positivity¶
The stochastic exponential is always strictly positive (exponential of a real number).
Property 2: Local Martingale Property¶
If \(X_t\) is a continuous local martingale, then \(\mathcal{E}(X)_t\) is also a local martingale.
Proof: From \(d\mathcal{E}(X)_t = \mathcal{E}(X)_t\,dX_t\), we see there is no \(dt\) term—the SDE has zero drift. \(\square\)
Property 3: Multiplication Rule¶
For two continuous semimartingales \(X\) and \(Y\):
Compare to the ordinary rule: \(e^x \cdot e^y = e^{x+y}\). The stochastic version has an extra covariation term.
Proof: Let \(Z_t = \mathcal{E}(X)_t\) and \(W_t = \mathcal{E}(Y)_t\). By the Itô product rule:
Since \(dZ = Z\,dX\) and \(dW = W\,dY\):
This shows \(ZW = \mathcal{E}(X)\mathcal{E}(Y)\) satisfies the SDE for \(\mathcal{E}(X + Y + \langle X, Y \rangle)\). \(\square\)
Property 4: Reciprocal¶
The reciprocal of a stochastic exponential is:
Important: The reciprocal is generally not a stochastic exponential itself. We have:
This follows from direct computation:
so
Property 5: Unit Expectation (Under Conditions)¶
If \(\mathcal{E}(X)\) is a true martingale (not just a local martingale):
This property is essential for measure changes but requires verification via Novikov or Kazamaki conditions.
Special Cases¶
Case 1: X_t = σ W_t (Constant Volatility)¶
For \(\sigma \in \mathbb{R}\) constant:
This is the fundamental exponential martingale. It satisfies \(\mathbb{E}[\mathcal{E}(\sigma W)_t] = 1\) for all \(t\) (Novikov is trivially satisfied since \(\sigma\) is constant).
Case 2: X_t = ∫_0^t σ_s dW_s (Stochastic Integrand)¶
This is the general form used in Girsanov's theorem. Whether it's a true martingale depends on \(\sigma\)—see Novikov and Kazamaki Conditions.
Case 3: Geometric Brownian Motion¶
The solution to \(dS_t = \mu S_t\,dt + \sigma S_t\,dW_t\) with \(S_0 > 0\) is:
This can be written as:
The term \(\mu - \frac{\sigma^2}{2}\) is the drift-adjusted growth rate—the Itô correction \(-\frac{\sigma^2}{2}\) appears naturally.
Connection to Girsanov's Theorem¶
The stochastic exponential is the Radon–Nikodym derivative for measure change.
Measure Change Setup¶
To change from measure \(\mathbb{P}\) to measure \(\mathbb{Q}\):
where the density process is:
Here \(\theta_t\) is the Girsanov kernel (market price of risk in finance).
Girsanov's Result¶
Under \(\mathbb{Q}\), the process:
is a standard Brownian motion.
Validity: For this measure change to be valid (i.e., \(\mathbb{Q}\) is a probability measure equivalent to \(\mathbb{P}\)), we need \(Z_t\) to be a true martingale with \(\mathbb{E}^{\mathbb{P}}[Z_T] = 1\). This is guaranteed by Novikov's or Kazamaki's condition.
See Girsanov's Theorem for the complete treatment.
Why the Correction Term Matters¶
Without Correction: e^W_t Has Drift¶
If we naively define \(Z_t = e^{W_t}\), then by Itô's lemma:
This has a positive drift term \(\frac{1}{2}Z_t\,dt\). The process \(e^{W_t}\) is a submartingale, not a martingale:
With Correction: E(W)_t Is a Martingale¶
Then:
No drift term—this is a true martingale with \(\mathbb{E}[\mathcal{E}(W)_t] = 1\).
The correction \(-\frac{1}{2}\langle X \rangle_t\) precisely removes the drift introduced by Itô's lemma.
When Is E(X) a True Martingale?¶
The stochastic exponential of a local martingale is always a local martingale, but may be a strict local martingale (local martingale that is not a true martingale).
Sufficient Conditions¶
Novikov's Condition: If
then \(\mathcal{E}(X)\) is a true martingale on \([0,T]\).
Kazamaki's Condition (weaker): If \(\mathcal{E}(X/2)\) is a submartingale, then \(\mathcal{E}(X)\) is a true martingale.
For details and proofs, see Novikov and Kazamaki Conditions.
When Conditions Fail¶
If neither condition holds, \(\mathcal{E}(X)\) may satisfy \(\mathbb{E}[\mathcal{E}(X)_T] < 1\). In finance, this corresponds to asset price bubbles. See Local Martingales for examples.
Extension to Discontinuous Processes¶
For a general semimartingale \(X\) with jumps, the Doléans-Dade exponential has the more complex form:
where:
- \(X^c\) is the continuous martingale part of \(X\)
- \(\Delta X_s = X_s - X_{s-}\) are the jumps
For \(\mathcal{E}(X)_t\) to remain positive, we need \(\Delta X_s > -1\) for all \(s\) (jumps cannot be too negative).
Applications in Finance¶
1. Risk-Neutral Measure Construction¶
The density process for the risk-neutral measure \(\mathbb{Q}\):
where \((\mu - r)/\sigma\) is the market price of risk (Sharpe ratio).
2. Change of Numéraire¶
When changing from numéraire \(M\) to numéraire \(N\):
This ratio involves stochastic exponentials of the volatility difference between the two numéraires.
3. Forward Measure¶
The \(T\)-forward measure uses the zero-coupon bond \(P(t,T)\) as numéraire. The density involves stochastic exponentials of bond volatilities.
See Forward Measure for details.
Summary¶
| Property | Statement |
|---|---|
| Definition | Unique solution to \(d\mathcal{E} = \mathcal{E}\,dX\), \(\mathcal{E}_0 = 1\) |
| Positivity | \(\mathcal{E}(X)_t > 0\) always |
| Local martingale | If \(X\) is a continuous local martingale, so is \(\mathcal{E}(X)\) |
| Multiplication | \(\mathcal{E}(X)\mathcal{E}(Y) = \mathcal{E}(X + Y + \langle X,Y\rangle)\) |
| True martingale | Requires Novikov or Kazamaki condition |
| Finance application | Radon–Nikodym derivative for measure change |
Key Takeaway
The stochastic exponential is the "correct" way to exponentiate stochastic processes, accounting for the Itô correction term \(-\frac{1}{2}\langle X \rangle_t\). This correction removes the drift that would otherwise arise, making \(\mathcal{E}(X)\) a local martingale. The stochastic exponential is essential for all measure-change arguments in mathematical finance.
Exercises¶
Exercise 1. Compute the stochastic exponential \(\mathcal{E}(\sigma W)_t\) explicitly for \(\sigma = 2\). Verify that \(\mathbb{E}[\mathcal{E}(2W)_t] = 1\) using the moment generating function of the normal distribution. What is \(\mathrm{Var}(\mathcal{E}(2W)_t)\)?
Solution to Exercise 1
For \(\sigma = 2\):
To verify \(\mathbb{E}[\mathcal{E}(2W)_t] = 1\): since \(W_t \sim N(0, t)\), we have \(2W_t \sim N(0, 4t)\). Using the MGF of a normal, \(\mathbb{E}[e^{aW_t}] = e^{a^2 t/2}\):
For the variance, first compute \(\mathbb{E}[\mathcal{E}(2W)_t^2]\):
Therefore:
Exercise 2. Let \(X_t = \int_0^t \sigma_s\,dW_s\) with \(\sigma_s = s\). Write the explicit formula for \(\mathcal{E}(X)_t\) and compute its quadratic variation \(\langle \mathcal{E}(X) \rangle_t\). Verify Novikov's condition for finite \(T\) and conclude that \(\mathcal{E}(X)\) is a true martingale on \([0, T]\).
Solution to Exercise 2
With \(X_t = \int_0^t s\,dW_s\), the quadratic variation is:
The stochastic exponential is:
The quadratic variation of \(\mathcal{E}(X)\) is computed from \(d\mathcal{E}(X)_t = \mathcal{E}(X)_t\,dX_t = \mathcal{E}(X)_t \cdot t\,dW_t\):
Novikov verification: For finite \(T\):
Since \(\sigma_s = s\) is deterministic, \(\langle X \rangle_T = T^3/3\) is deterministic, and the exponential moment is trivially finite. By Novikov's theorem, \(\mathcal{E}(X)\) is a true martingale on \([0, T]\) for any finite \(T\).
Exercise 3. Apply Ito's lemma to \(Z_t = e^{W_t}\) (without the correction) and show that \(dZ_t = Z_t\,dW_t + \frac{1}{2}Z_t\,dt\). Then apply Ito's lemma to \(\mathcal{E}(W)_t = e^{W_t - t/2}\) and show that the \(dt\) term vanishes. Explain why the Ito correction \(-\frac{1}{2}\langle W \rangle_t = -t/2\) is exactly what is needed to remove the drift.
Solution to Exercise 3
Without correction: Let \(Z_t = e^{W_t}\). By Itô's lemma with \(f(x) = e^x\):
The \(\frac{1}{2}Z_t\,dt\) drift makes \(Z_t\) a submartingale (systematically increasing in expectation).
With correction: Let \(\mathcal{E}(W)_t = e^{W_t - t/2}\). Define \(Y_t = W_t - t/2\), so \(\mathcal{E}(W)_t = e^{Y_t}\). By Itô's lemma:
Now \(dY_t = dW_t - \frac{1}{2}\,dt\) and \((dY_t)^2 = (dW_t)^2 = dt\). Substituting:
The \(dt\) terms cancel exactly. The Itô correction \(-\frac{1}{2}\langle W \rangle_t = -t/2\) compensates precisely for the second-order term \(\frac{1}{2}f''(W_t)\,dt\) in Itô's formula, removing the drift and producing a martingale.
Exercise 4. Prove the multiplication rule \(\mathcal{E}(X)_t \cdot \mathcal{E}(Y)_t = \mathcal{E}(X + Y + \langle X, Y \rangle)_t\) by applying the Ito product rule to \(Z_t = \mathcal{E}(X)_t\) and \(U_t = \mathcal{E}(Y)_t\). Identify where the covariation term \(\langle X, Y \rangle\) enters.
Solution to Exercise 4
Let \(Z_t = \mathcal{E}(X)_t\) and \(U_t = \mathcal{E}(Y)_t\). By the Itô product rule:
Since \(dZ_t = Z_t\,dX_t\) and \(dU_t = U_t\,dY_t\):
For the covariation: \(d\langle Z, U \rangle_t = Z_t U_t\,d\langle X, Y \rangle_t\) (since the diffusion coefficients of \(Z\) and \(U\) are \(Z_t\) times \(dX_t\) and \(U_t\) times \(dY_t\)).
Therefore:
This means \(Z_t U_t\) solves the SDE \(d\mathcal{E} = \mathcal{E}\,d(X + Y + \langle X, Y \rangle)\) with initial condition \(Z_0 U_0 = 1\). By uniqueness of the stochastic exponential SDE:
The covariation \(\langle X, Y \rangle\) enters through the cross-term \(d\langle Z, U \rangle_t\) in the Itô product rule — this is the stochastic calculus analogue of the fact that the product of two exponentials involves the sum of exponents, but with an additional correction from quadratic covariation.
Exercise 5. In the Black-Scholes model, the stock price is \(S_t = S_0 e^{(\mu - \sigma^2/2)t + \sigma W_t} = S_0 e^{\mu t} \mathcal{E}(\sigma W)_t\). Show that the discounted price \(e^{-rt}S_t\) can be written as \(S_0 e^{(\mu - r)t}\mathcal{E}(\sigma W)_t\) and explain why this is a martingale under \(\mathbb{Q}\) but not under \(\mathbb{P}\) (unless \(\mu = r\)).
Solution to Exercise 5
The stock price is \(S_t = S_0 e^{(\mu - \sigma^2/2)t + \sigma W_t} = S_0 e^{\mu t}\mathcal{E}(\sigma W)_t\). The discounted price is:
Under \(\mathbb{P}\): The factor \(e^{(\mu - r)t}\) introduces a deterministic drift. Since \(\mathcal{E}(\sigma W)_t\) is a \(\mathbb{P}\)-martingale with \(\mathbb{E}^{\mathbb{P}}[\mathcal{E}(\sigma W)_t] = 1\):
So \(e^{-rt}S_t\) is not a \(\mathbb{P}\)-martingale when \(\mu \neq r\).
Under \(\mathbb{Q}\): Girsanov's theorem replaces \(W_t\) by \(W_t^{\mathbb{Q}} - \frac{\mu - r}{\sigma}t\), so \(\sigma W_t = \sigma W_t^{\mathbb{Q}} - (\mu - r)t\). Then:
The factor \(e^{(\mu-r)t}\) cancels with \(e^{-(\mu-r)t}\) from the Girsanov shift, and the discounted price becomes \(S_0\mathcal{E}(\sigma W^{\mathbb{Q}})_t\), which is a \(\mathbb{Q}\)-martingale with unit expectation.
Exercise 6. For the Radon-Nikodym derivative \(Z_t = \mathcal{E}(-\int_0^{\cdot}\theta_s\,dW_s)_t\), show that \(1/Z_t\) is not equal to \(\mathcal{E}(\int_0^{\cdot}\theta_s\,dW_s)_t\) in general. Compute the ratio explicitly and identify the extra multiplicative factor involving \(\langle M \rangle_t\) where \(M_t = \int_0^t \theta_s\,dW_s\).
Solution to Exercise 6
Let \(M_t = -\int_0^t \theta_s\,dW_s\), so \(Z_t = \mathcal{E}(M)_t = \exp(M_t - \frac{1}{2}\langle M \rangle_t)\) and \(\langle M \rangle_t = \int_0^t \theta_s^2\,ds\).
The reciprocal is:
On the other hand, \(\mathcal{E}(-M)_t = \mathcal{E}(\int_0^\cdot \theta_s\,dW_s)_t = \exp(\int_0^t \theta_s\,dW_s - \frac{1}{2}\int_0^t \theta_s^2\,ds)\).
Comparing:
The extra factor is \(\exp(\langle M \rangle_t) = \exp(\int_0^t \theta_s^2\,ds)\), which is always \(\geq 1\). So \(1/Z_t > \mathcal{E}(-M)_t\) unless \(\theta \equiv 0\). This discrepancy arises because the reciprocal operation interacts with the Itô correction term: flipping the sign of \(M\) changes the sign of the first-order term but not the sign of the quadratic variation, creating an asymmetry.
Exercise 7. Consider two independent Brownian motions \(W_t^1\) and \(W_t^2\) and define \(X_t = \sigma_1 W_t^1 + \sigma_2 W_t^2\). Compute \(\langle X \rangle_t\) and write \(\mathcal{E}(X)_t\) explicitly. Then verify using the multiplication rule that \(\mathcal{E}(\sigma_1 W^1)_t \cdot \mathcal{E}(\sigma_2 W^2)_t = \mathcal{E}(\sigma_1 W^1 + \sigma_2 W^2)_t\) (since \(\langle W^1, W^2 \rangle = 0\)).
Solution to Exercise 7
With \(X_t = \sigma_1 W_t^1 + \sigma_2 W_t^2\) and \(W^1, W^2\) independent:
The stochastic exponential is:
Now verify the multiplication rule. With \(\langle \sigma_1 W^1, \sigma_2 W^2 \rangle_t = \sigma_1\sigma_2\langle W^1, W^2 \rangle_t = 0\) (independence):
Explicitly:
The key point is that when the cross-covariation vanishes (\(\langle W^1, W^2 \rangle = 0\)), the multiplication rule reduces to the ordinary exponential identity \(e^a \cdot e^b = e^{a+b}\).