Martingale Representation: Hedging as Projection

Risk-neutral pricing says derivatives are discounted expectations under $\mathbb{Q}$. But there is a deeper question: how do you replicate a derivative by trading in the underlying? This is the question of martingale representation. The answer is one of the most elegant results in stochastic analysis: every square-integrable claim that can be priced by no-arbitrage can be written as a stochastic integral of a predictable process against the underlying Brownian motion. The integrand is the trading strategy.

This article derives the result for the Brownian setting, gives the intuition via the Cameron-Martin space, and explains what it means for hedging in practice. We will see why delta hedging works, why exotic options are harder to hedge than vanillas, and why markets with multiple sources of randomness need multiple hedging instruments.

The Setting

We work on a probability space $(\Omega, \mathcal{F}, \mathbb{Q})$ supporting a Brownian motion $W_t$. The filtration $\mathcal{F}_t$ is the natural filtration of $W_t$ augmented by null sets. We assume the discounted price process $\tilde{S}_t = e^{-rt} S_t$ is a $\mathbb{Q}$-martingale. Under GBM,

A contingent claim is an $\mathcal{F}_T$-measurable random variable $V_T$. We say $V_T$ is attainable if there is a self-financing trading strategy $\phi_t$ (the number of shares held at time $t$) such that the wealth process $V_t = \phi_t \tilde{S}_t + \psi_t B_t$ (with $\psi_t$ in the bond) satisfies $V_T = V_T$. The initial cost is the price.

The martingale representation theorem asks: for which $V_T$ does such a $\phi_t$ exist?

The Representation Theorem

Theorem (Brownian Martingale Representation). Let $W_t$ be a standard Brownian motion under $\mathbb{Q}$. Let $V_T$ be a square-integrable $\mathcal{F}_T$-measurable random variable. Then there exists a unique predictable process $\phi_t$ such that

The process $\phi_t$ is the integrand or replicating strategy. The representation is unique: if two integrands give the same claim, they agree almost surely.

This theorem is the rigorous foundation of complete-market hedging. The discounted wealth process $\tilde{V}_t = e^{-rt} V_t$ is a martingale. By the theorem, any martingale in a Brownian filtration can be written as a constant plus a stochastic integral against $W_t$. The integrand is the position in the risky asset.

Equivalently, in terms of undiscounted prices, the value process of any self-financing trading strategy satisfies

and the position $\phi_t$ in shares is determined by the representation theorem applied to $V_T$.

Why the Theorem Matters

The theorem is the bridge between pricing and hedging. Pricing says: $V_0 = e^{-rT} \mathbb{E}^{\mathbb{Q}}[V_T]$. The theorem says: there exists a strategy that achieves $V_T$ starting from $V_0$. Together they imply no-arbitrage pricing — the price equals the cost of replication.

This is also the formal content of Black-Scholes-Merton: the existence of a replicating strategy for European options in a complete market. The delta hedge $\phi_t = \partial V / \partial S$ is exactly the integrand predicted by the theorem. We will see how to compute it shortly.

For claims that are not in $L^2$ (very heavy-tailed payoffs), the theorem fails. For those, you cannot replicate by bounded trading strategies, and pricing may not be unique. We will skip those technicalities here.

Computing the Integrand: The Clark-Ocone Formula

The martingale representation theorem is existence-and-uniqueness. To compute $\phi_t$, you need the Clark-Ocone formula:

where $D_t$ is the Malliavin derivative (a stochastic derivative, like a directional derivative in a function space).

The Malliavin derivative is defined for smooth functionals of Brownian motion. If $V_T = f(W_{t_1}, W_{t_2}, \ldots, W_{t_n})$ for a smooth function $f$, then

The derivative picks up the sensitivity of the payoff to each Brownian increment. The Clark-Ocone formula then takes the conditional expectation to project onto the predictable (non-anticipating) integrand.

For a European call $V_T = (S_T - K)^+$, the computation gives the famous delta:

in units of shares per option. The Malliavin derivative and the delta coincide for this smooth, monotone payoff.

The Intuition: Projections in Function Space

The martingale representation theorem has a clean linear-algebra reading. The space $L^2(\Omega, \mathcal{F}_T, \mathbb{Q})$ of square-integrable claims is a Hilbert space with inner product $\mathbb{E}^{\mathbb{Q}}[XY]$. The subspace $M^2$ of $\mathbb{Q}$-martingales of the form $\int_0^T \phi_s \, dW_s$ is a closed subspace. Every claim $V_T \in L^2$ can be decomposed uniquely as

where $M_T \in M^2$. The component $M_T$ is the martingale part, and the projection onto $M^2$ gives the integrand.

In this view, hedging is the orthogonal projection of the payoff onto the subspace of attainable claims. The mean is the constant term, the stochastic integral is the projection. The theorem guarantees that every claim has such a projection. The integrand $\phi_t$ is the predictable projection of $D_t V_T$.

For European options, this projection can be computed in closed form. For path-dependent or exotic payoffs, the projection typically requires numerical methods. Monte Carlo, PDE solvers, or Malliavin-based simulation.

Multi-Asset Generalization

When there are $d$ Brownian motions and $d$ risky assets, the representation is multi-dimensional:

The hedging strategy has $d$ components (one per asset), one for each source of randomness. This is the multi-asset complete market: you need one hedging instrument per source of randomness to span all claims.

If there are fewer assets than sources of randomness, the market is incomplete: not every claim can be replicated. The set of attainable claims is a strict subspace. Pricing is no longer unique; there is a range of no-arbitrage prices.

For example, an interest-rate derivative driven by two factors (level and slope) cannot be perfectly hedged with a single bond position. You need at least two bonds (or a bond and a swap) to span the two factors.

The Markovian Case and the Black-Scholes PDE

For payoffs of the form $V_T = g(S_T)$, the value process is a function of $S_t$ and $t$: $V_t = V(S_t, t)$. By Itô’s lemma,

Setting the discounted process $\tilde{V}_t$ to be a martingale and using $dS_t = r S_t \, dt + \sigma S_t \, dW_t$, we get

plus the drift term, which must vanish:

Multiplying by $e^{rt}$ and simplifying gives the Black-Scholes PDE:

with terminal condition $V(S_T, T) = g(S_T)$. The PDE is the pricing equation. Its solution is the unique arbitrage-free price of the European claim. The delta hedge is $V_S = \partial V / \partial S$, which is the integrand from the representation theorem.

This is the most useful application of martingale representation for practitioners: for European claims, the PDE gives you both the price and the hedge. For path-dependent claims, the PDE becomes path-dependent itself or has extra state variables (running max, integral of price, etc.).

The Hedging Error

In practice, you cannot trade continuously. You rebalance discretely. This introduces hedging error, which is the difference between the realized payoff of the discrete hedge and the desired terminal payoff.

For a delta hedge rebalanced at times $t_0, t_1, \ldots, t_n$ with $\Delta t = T/n$, the hedging error scales as $O(\sqrt{\Delta t})$ for a European option. More frequent rebalancing gives smaller error, but at higher transaction cost. The optimal rebalancing frequency balances these two.

For exotic options, the scaling is worse. Discretely monitored barriers have knock-out probabilities that depend on the monitoring frequency. Discrete Asian options have averaging error. The replication is never exact in practice, but it can be very good with sufficient rebalancing.

The martingale representation theorem gives the ideal hedge in continuous time. Real-world hedges are approximations. Quantifying the gap is an active research area, especially in regimes with discrete trading, transaction costs, and model misspecification.

A Code Illustration

import numpy as np

def delta_hedge_call(S0, K, r, sigma, T, n_steps, n_paths=10000, seed=0):
    rng = np.random.default_rng(seed)
    dt = T / n_steps
    discount = np.exp(-r * T)

    # Precompute the Black-Scholes delta for the call
    def call_delta(t, S):
        tau = T - t
        if tau <= 0:
            return 1.0 if S > K else 0.0
        d1 = (np.log(S/K) + (r + 0.5*sigma**2)*tau) / (sigma*np.sqrt(tau))
        from scipy.stats import norm
        return norm.cdf(d1)

    # Simulate paths and apply the hedge
    S = np.full(n_paths, S0)
    cash = np.zeros(n_paths)
    cash += call_delta(0.0, S) * S
    for i in range(n_steps):
        t = i * dt
        delta = call_delta(t, S)
        cash -= delta * S
        dW = rng.standard_normal(n_paths) * np.sqrt(dt)
        S = S * np.exp((r - 0.5*sigma**2)*dt + sigma*dW)
        cash *= np.exp(r * dt)
        cash += delta * S

    payoff = np.maximum(S - K, 0)
    hedge_pnl = cash + call_delta(T, S) * S - payoff
    print(f"Mean hedge error: {hedge_pnl.mean():.4f}")
    print(f"Std hedge error: {hedge_pnl.std():.4f}")

The mean hedge error should be close to zero (since the strategy is self-financing in continuous time), and the std gives the variance of the discrete-rebalancing error. With many steps and paths, the std shrinks as $O(1/\sqrt{n_{\text{steps}}})$, confirming the convergence to perfect replication.

Why Path-Dependent Claims Are Harder

For an Asian option with payoff $g(\frac{1}{T} \int_0^T S_t \, dt)$, the natural state variables are $(S_t, A_t, t)$ where $A_t = \int_0^t S_u \, du$ is the running average. The pricing PDE is three-dimensional. The hedging strategy depends on both $S_t$ and $A_t$:

Computing this delta in real time requires tracking the running average, which is straightforward in simulation but expensive in production trading systems where you need fast Greeks.

For a barrier option, you also need to track the running maximum $M_t = \max_{u \leq t} S_u$. The PDE is again three-dimensional. The delta can change discontinuously near the barrier, which creates practical hedging challenges.

In all cases, the martingale representation theorem still holds — the integrand exists and is unique. Computing it in closed form is harder, and computing it in real time is sometimes infeasible. That is why some exotic options trade with wider bid-ask spreads than vanillas.

The Hedging Strategy as Insurance

A useful way to think about delta hedging: it is buying insurance against the price moves of the underlying. The position $\phi_t \sigma S_t \, dW_t$ represents the infinitesimal P&L of the hedge. Summing over all infinitesimal moves, the hedge absorbs the diffusion risk. What remains is the drift (which is $r$ in the risk-neutral world) plus any model error.

For a delta-neutral hedge, the next-order Greek is gamma: $\partial^2 V / \partial S^2$. Gamma measures how fast delta changes with the underlying. High gamma means the hedge has to be rebalanced frequently. Low gamma means a static hedge is approximately correct.

Vega ($\partial V / \partial \sigma$) is the sensitivity to volatility. A vega-neutral hedge eliminates small parallel shifts in the vol surface. This is harder to maintain, because vol is not directly tradable. You approximate vega hedges by trading options at different strikes.

What to Take Away

The martingale representation theorem says: every contingent claim in a Brownian filtration can be written as its expectation plus a stochastic integral. The integrand is the hedging strategy. The Clark-Ocone formula gives it as the conditional expectation of the Malliavin derivative.

For Markovian claims, the hedge is the delta $\partial V / \partial S$, computable from the Black-Scholes PDE. For path-dependent claims, the hedge requires more state variables. For multi-asset claims, you need one hedging instrument per source of randomness.

The practical takeaway: pricing by risk-neutral expectation is justified because hedging exists. The two are the same statement, viewed from different angles. The integrand from representation theory is the position you take in the underlying. You compute it once per rebalance, hold it until the next rebalance, and absorb the price moves.

References: Øksendal (2003), Chapter 7; Shreve (2008), Stochastic Calculus for Finance II, Chapter 6; Björk (2009), Chapter 12. Glasserman (2004) discusses the connection between representation and Monte Carlo in Section 6. The full proof of the Clark-Ocone formula requires Malliavin calculus, which we have glossed over. For practical work, the Black-Scholes PDE and the delta formula are what you use.

Malliavin Calculus in One Paragraph

The Malliavin derivative $D_t$ is the stochastic analogue of the directional derivative. For a random variable $F = f(W(h_1), \ldots, W(h_n))$ where $W(h_i) = \int_0^T h_i(s) \, dW_s$ are Wiener integrals of deterministic functions $h_i$, the Malliavin derivative is

In words: $D_s F$ measures the sensitivity of $F$ to a perturbation of the Brownian path at time $s$. It is a process indexed by $s \in [0, T]$. For a simple example, if $F = W_T$, then $D_s F = 1$ for $s \leq T$ and 0 otherwise. Brownian motion at the terminal time is sensitive to every increment along the way.

The key property: $D_s F$ is adapted to the reverse filtration (it depends on future values of $W$ if $s < T$). The Clark-Ocone formula projects $D_s F$ onto the natural filtration by conditioning on $\mathcal{F}_s$. This projection is the integrand $\phi_s$.

Computing the Malliavin derivative for a European option gives the same answer as computing the delta — they coincide. The deeper value of the formula is for path-dependent or non-smooth payoffs, where classical Greeks may not exist but the Malliavin derivative does. For digital options, lookback payoffs, or barrier options, the Malliavin approach yields robust integrands even where finite differences break down.

Incomplete Markets: A Different World

If the market has more sources of randomness than traded assets, not every claim can be replicated. The set of attainable claims is a strict subspace of $L^2(\mathcal{F}_T)$. Pricing is no longer unique.

For example, consider a single stock driven by two independent Brownian motions: $dS_t = r S_t \, dt + \sigma_1 S_t \, dW_t^1 + \sigma_2 S_t \, dW_t^2$. There is only one traded asset, but two random sources. The martingale representation in this filtration would have two integrands:

But we only have one risky asset, so we can only specify one trading strategy. We cannot independently control $\phi^1$ and $\phi^2$. The set of attainable claims is the subset where $\phi_t^2 = 0$ for all $t$ — i.e., claims whose payoff is a function of the stock path alone, which automatically lies in the one-dimensional subspace spanned by $dW^1$.

The price range for unattainable claims is determined by the dual variables (state-price densities) that span the orthogonal complement. This is the basis of utility-indifference pricing: choose a utility function, find the price that makes the hedger and the unhedgeable risk jointly optimal.

Connection to PDE Methods

The martingale representation theorem and the Black-Scholes PDE give the same answers for European claims. Why two formulations?

• The PDE gives the price as a function of state variables. It is useful for computing greeks, for visualizing price surfaces, and for understanding how prices depend on parameters.
• The martingale representation gives the price as an expectation. It is useful for Monte Carlo simulation, for path-dependent payoffs, and for high-dimensional problems where PDEs become infeasible.

For an option with one underlying and one source of randomness, both methods work. For basket options or Asian options, PDE methods struggle (the state space becomes too large), and Monte Carlo with martingale representation wins. For high-dimensional interest-rate derivatives, only Monte Carlo is practical.

The two methods are dual. The PDE comes from conditioning the expectation on the current state; the representation comes from decomposing the payoff in the filtration. They give the same price for the same claim. Use whichever is computationally tractable.

The Predictable Representation Property

A market is said to have the predictable representation property if every square-integrable martingale is representable as a stochastic integral against the traded assets’ driving Brownian motions. The Brownian case satisfies this property. Markets driven by jump-diffusions or Lévy processes generally do not, unless there are enough traded assets to span all sources of randomness.

If the predictable representation property fails, the martingale representation theorem may not apply. You may have martingales that are not stochastic integrals against traded assets. In that case, no-arbitrage pricing is still valid (the FTAP says an equivalent measure exists), but replication is not always possible.

This is the formal foundation of why some markets are complete and others are not. Complete markets have the predictable representation property. Incomplete markets do not.

Numerical Computation of the Clark-Ocone Integrand

For high-dimensional or non-smooth payoffs, computing the Clark-Ocone integrand numerically is an active research area. The basic algorithm is:

1. Compute the Malliavin derivative $D_t F$ along simulated paths.
2. Take the conditional expectation $\mathbb{E}^{\mathbb{Q}}[D_t F | \mathcal{F}_t]$.
3. Use this as the integrand $\phi_t$.

Step 1 can be done by finite differences in the Brownian increments (analogous to bumping the underlying). Step 2 requires non-parametric regression on the simulated paths (kernel regression, neural networks, etc.). Step 3 gives the integrand, which you can use for hedging.

The technique is called the Malliavin-Monte Carlo method. It is sometimes more efficient than direct delta computation, especially for payoffs that are not smooth enough for finite differences. The references include Fournié et al. (1999) and the survey by Kohatsu-Higa and Yano.

A Hedging Paradox

Here is a paradox worth thinking about. The martingale representation theorem says every claim is replicable in continuous time with continuous rebalancing. But real trading happens in discrete time with discrete prices. So in practice, replication is never exact. How, then, can the price be unique?

The answer is in the limit. As the rebalancing frequency goes to infinity, the discrete hedge converges to the continuous hedge in $L^2$, and the cost converges to the price. The martingale representation theorem gives the asymptotic ideal. The bid-ask spread reflects the cost of getting close to the ideal with finite rebalancing.

For European options on liquid underlyings, the spread is tiny — well within the precision of the model. For exotic options, the spread can be substantial. The pricing error from discrete hedging is part of the cost of trading.

What the Theorem Doesn’t Tell You

The martingale representation theorem assumes:

• Continuous price paths (no jumps).
• Continuous-time trading (no frictions).
• Continuous Brownian motion (no discrete ticks).
• The model is correct.

In practice, all four assumptions fail. Prices jump (8% overnight gaps happen). Trading has bid-ask spreads and commissions. Prices tick in discrete units. Models are wrong. The theorem is a limit, and the limit may not be tight.

For a working quant, the question is not “what is the perfect hedge” but “what is the best hedge given my constraints”. The martingale representation gives you the target. Practical hedging gets you close. The gap is your edge — or your cost.