Martingale Representation Theorem
1 Martingale Representation Theorem
The martingale representation theorem states that every square-integrable martingale in the filtration generated by a Brownian motion is an Itô integral with respect to that Brownian motion. Equivalently, the only source of randomness in such a filtration is the Brownian motion itself, and all information is carried by its increments.
In Quantitative Finance this is the reason that a market driven by Brownian motion is complete: every contingent claim can be written as a stochastic integral, which is the gain of a self-financing hedging strategy.
2 Setup
Let \(W=(W^1,\dots,W^n)\) be an \(n\)-dimensional Brownian motion on \((\Omega,\mathcal{F},\mathbb{P})\) and let \(\mathbb{F}=(\mathcal{F}_t)\) be its natural filtration, augmented by the \(\mathbb{P}\)-null sets (see Filtrations and Adapted Processes). Fix \(T>0\) and write \(\mathcal{L}^2\) for the space of progressively measurable processes \(g=(g^1,\dots,g^n)\) with
\[ \mathbb{E}\int_0^T |g_s|^2\,ds < \infty. \]
For \(g\in\mathcal{L}^2\) the Itô integral \(\int_0^t g_s\,dW_s=\sum_{i=1}^n\int_0^t g^i_s\,dW^i_s\) is a square-integrable martingale and satisfies the Itô isometry
\[ \mathbb{E}\left[\left(\int_0^T g_s\,dW_s\right)^2\right] = \mathbb{E}\int_0^T |g_s|^2\,ds. \]
3 The Theorem
For every \(F\in L^2(\Omega,\mathcal{F}_T,\mathbb{P})\) there is a unique \(g\in\mathcal{L}^2\) such that
\[ F = \mathbb{E}[F] + \int_0^T g_s\,dW_s \quad \text{a.s.} \]
Let \(M=(M_t)_{t\in[0,T]}\) be an \(\mathbb{F}\)-martingale with \(\mathbb{E}[M_T^2]<\infty\). Then there is a unique \(g\in\mathcal{L}^2\) such that
\[ M_t = M_0 + \int_0^t g_s\,dW_s \quad \text{for all } t\in[0,T], \text{ a.s.} \]
In particular \(M_0=\mathbb{E}[M_T]\) is constant and \(M\) has a continuous version.
If \(M\) is square integrable on \([0,t]\) for every \(t\ge 0\), the processes \(g\) obtained on different intervals agree by uniqueness, which gives a single \(g\) on \([0,\infty)\).
4 Proof
The martingale representation theorem follows from the Itô representation theorem, which we prove first. The strategy is to show that the claim holds for a family of random variables whose span is dense in \(L^2(\mathcal{F}_T)\), and then pass to the limit using the Itô isometry.
4.1 Step 1: Exponentials are Total
For a bounded deterministic function \(h:[0,T]\to\mathbb{R}^n\) define
\[ Y^h_t = \exp\left( \int_0^t h_s\cdot dW_s - \frac12\int_0^t |h_s|^2\,ds \right). \]
The linear span of \(\{Y^h_T : h \text{ a bounded deterministic step function}\}\) is dense in \(L^2(\Omega,\mathcal{F}_T,\mathbb{P})\).
Proof. Since the span is a linear subspace, it is dense if its orthogonal complement is \(\{0\}\). Let \(g\in L^2(\mathcal{F}_T)\) be orthogonal to every \(Y^h_T\), that is \(\mathbb{E}[g\,Y^h_T]=0\). Since \(Y^h_T\) is \(\exp\left(\int_0^T h\cdot dW\right)\) times a positive constant, this says
\[ \mathbb{E}\left[g\,\exp\left(\int_0^T h_s\cdot dW_s\right)\right] = 0. \]
Fix times \(0=t_0<t_1<\dots<t_m\le T\) and let \(Z\) be the vector of increments \(W_{t_j}-W_{t_{j-1}}\), \(j=1,\dots,m\), which lies in \(\mathbb{R}^N\) with \(N=nm\). Taking \(h=\sum_j\lambda_j\mathbf{1}_{(t_{j-1},t_j]}\) with \(\lambda_j\in\mathbb{R}^n\) gives
\[ \phi(\lambda) := \mathbb{E}\left[g\,e^{\lambda\cdot Z}\right] = 0 \quad \text{for all } \lambda\in\mathbb{R}^N. \]
Because \(Z\) is Gaussian it has exponential moments of all orders, so by Cauchy-Schwarz the expectation is finite for complex \(\lambda\in\mathbb{C}^N\) and \(\phi\) is an analytic function on \(\mathbb{C}^N\). It vanishes on \(\mathbb{R}^N\), hence on all of \(\mathbb{C}^N\). In particular \(\mathbb{E}[g\,e^{iy\cdot Z}]=0\) for all \(y\in\mathbb{R}^N\). This says that the finite signed measure \(\nu(A)=\mathbb{E}[g\,\mathbf{1}_{\{Z\in A\}}]\) on \(\mathbb{R}^N\) has zero Fourier transform, so \(\nu=0\). Therefore \(\mathbb{E}[g\,\mathbf{1}_A]=0\) for every \(A\in\sigma(Z)\).
The \(\sigma\)-algebras \(\sigma(Z)\), over all choices of times, form a \(\pi\)-system that generates \(\mathcal{F}_T\) (up to null sets). By the \(\pi\)-\(\lambda\) theorem, \(\mathbb{E}[g\,\mathbf{1}_A]=0\) for all \(A\in\mathcal{F}_T\), and hence \(g=0\) almost surely.
4.2 Step 2: Representation of Exponentials
Let \(h\) be a bounded deterministic step function. By Itô Formula applied to \(\exp\) of the process \(\int_0^t h\cdot dW-\frac12\int_0^t|h|^2ds\),
\[ dY^h_t = Y^h_t\,h_t\cdot dW_t, \qquad\text{so}\qquad Y^h_T = 1 + \int_0^T Y^h_s\,h_s\cdot dW_s. \]
Here \(Y^h_s h_s\in\mathcal{L}^2\), since \(h\) is bounded and \(\mathbb{E}[(Y^h_s)^2]=\exp\left(\int_0^s|h|^2\right)\) is bounded on \([0,T]\). Since \(\mathbb{E}[Y^h_T]=1\), this is the representation in the theorem for \(F=Y^h_T\). By linearity, the representation also holds for every finite linear combination of the \(Y^h_T\).
4.3 Step 3: Passage to the Limit
Let \(F\in L^2(\mathcal{F}_T)\). By Lemma 1 there are finite linear combinations \(F_k\to F\) in \(L^2\), and by Step 2
\[ F_k = \mathbb{E}[F_k] + \int_0^T g^{(k)}_s\,dW_s, \qquad g^{(k)}\in\mathcal{L}^2. \]
The stochastic integral has mean zero, so by the Itô isometry
\[ \mathbb{E}\left[(F_k-F_l)^2\right] = \left(\mathbb{E}[F_k-F_l]\right)^2 + \mathbb{E}\int_0^T |g^{(k)}_s-g^{(l)}_s|^2\,ds. \]
The left side tends to \(0\) as \(k,l\to\infty\), so \((g^{(k)})\) is Cauchy in \(\mathcal{L}^2\), which is complete, and converges to some \(g\in\mathcal{L}^2\). Also \(\mathbb{E}[F_k]\to\mathbb{E}[F]\). Passing to the limit in \(L^2\), using the isometry again,
\[ F = \lim_{k\to\infty}F_k = \mathbb{E}[F] + \int_0^T g_s\,dW_s. \]
Uniqueness. If \(\int_0^T g_s\,dW_s=\int_0^T \tilde g_s\,dW_s\), then by the isometry \(\mathbb{E}\int_0^T|g_s-\tilde g_s|^2ds=0\), so \(g=\tilde g\) \(dt\otimes d\mathbb{P}\)-almost everywhere. This proves the Itô representation theorem.
4.4 Step 4: Martingales
Let \(M\) be a square-integrable \(\mathbb{F}\)-martingale. Apply the Itô representation theorem to \(F=M_T\in L^2(\mathcal{F}_T)\) to get \(M_T=\mathbb{E}[M_T]+\int_0^T g_s\,dW_s\). Since \(M\) is a martingale, \(\mathbb{E}[M_T]=M_0\) and, taking conditional expectations and using that the stochastic integral is a martingale,
\[ M_t = \mathbb{E}[M_T\mid\mathcal{F}_t] = M_0 + \int_0^t g_s\,dW_s, \qquad t\in[0,T]. \]
Uniqueness follows from the isometry as above. The right side is a continuous process, which shows that \(M\) has a continuous version. \(\blacksquare\)
5 Application: Hedging
Consider the Black-Scholes-Merton Model and a claim with payoff \(h(S_T)\in L^2(\mathbb{Q})\). Under the risk-neutral measure \(\mathbb{Q}\) of Risk-Neutral Pricing the process
\[ M_t = e^{-rT}\,\mathbb{E}_{\mathbb{Q}}[h(S_T)\mid\mathcal{F}_t] \]
is a square-integrable \(\mathbb{Q}\)-martingale in the filtration of the \(\mathbb{Q}\)-Brownian motion \(W^{\mathbb{Q}}\) (which equals the filtration of \(S\), since \(S\) and \(W^{\mathbb{Q}}\) determine each other). By the theorem,
\[ M_t = M_0 + \int_0^t g_s\,dW^{\mathbb{Q}}_s \]
for some \(g\). The discounted stock price satisfies \(d(e^{-rt}S_t)=\sigma e^{-rt}S_t\,dW^{\mathbb{Q}}_t\), so the strategy that holds \(\Delta_t = g_t/(\sigma e^{-rt}S_t)\) shares has discounted gains exactly \(M_t-M_0\). Starting from capital \(M_0\), which is the price of the claim, this self-financing strategy replicates the payoff. This shows that every claim can be hedged, so the model is complete, and identifies the hedge ratio.
6 Remarks
- The theorem depends on the filtration. It holds for the Brownian filtration, but not for a larger one: if \(\mathcal{F}_t\) carries information that is independent of \(W\) (for example a second independent source of randomness or a Poisson process), martingales need not be stochastic integrals against \(W\).
- The integrand \(g\) can be identified explicitly by the Clark-Ocone formula, \(g_t=\mathbb{E}[D_tF\mid\mathcal{F}_t]\), where \(D_tF\) is the Malliavin derivative, when \(F\) is sufficiently smooth.
- The converse direction is immediate: every Itô integral of an \(\mathcal{L}^2\) integrand is a square-integrable martingale.
8 References
- Øksendal, B. Stochastic Differential Equations: An Introduction with Applications, Chapter 4 (the Itô representation theorem and the martingale representation theorem).
- Karatzas, I. and Shreve, S. E. Brownian Motion and Stochastic Calculus, Section 3.4.