Stochastic differential equations · Part 8 of 10

Martingales

· 5 min read

Martingales have appeared throughout this collection: Brownian motion, Itô integrals and the processes u(t,Xt)u(t, X_t) in the post on PDEs are all martingales. This post collects the general theory that makes them so useful.

Definitions

An adapted, integrable process MM is a martingale if E[Mt∣Fs]=Ms\mathbb{E}[M_t \mid \mathcal{F}_s] = M_s for all s≤ts \le t, a submartingale if E[Mt∣Fs]≥Ms\mathbb{E}[M_t \mid \mathcal{F}_s] \ge M_s, and a supermartingale if the inequality goes the other way. The names are easy to mix up: a supermartingale is a game that is unfavorable to the player. If MM is a martingale and ϕ\phi is convex, then ϕ(Mt)\phi(M_t) is a submartingale by Jensen’s inequality, so ∣Mt∣|M_t| and Mt2M_t^2 are submartingales.

In discrete time, a betting strategy HnH_n that is decided before round nn turns a martingale into the martingale transform ∑k≤nHk(Mk−Mk−1)\sum_{k \le n} H_k (M_k - M_{k-1}), which is again a martingale. In other words, no strategy can beat a fair game. The Itô integral is the continuous-time version of this statement.

Convergence and uniform integrability

The martingale convergence theorem. A martingale, or more generally a submartingale, with sup⁡tE∣Mt∣<∞\sup_t \mathbb{E}|M_t| < \infty converges almost surely to a finite limit M∞M_\infty.

The proof counts upcrossings. A path that oscillated infinitely often between two levels would let a “buy low, sell high” strategy win without bound, which is impossible in a fair game.

Almost sure convergence does not imply E[M∞]=E[M0]\mathbb{E}[M_\infty] = \mathbb{E}[M_0], though. The exponential martingale exp⁡(Bt−t/2)\exp(B_t - t/2) has mean one at every time but tends to zero almost surely. The missing ingredient is uniform integrability. For a martingale, the following are equivalent:

  • MM is uniformly integrable;
  • MtM_t converges in L1L^1;
  • there is an integrable random variable M∞M_\infty with Mt=E[M∞∣Ft]M_t = \mathbb{E}[M_\infty \mid \mathcal{F}_t] for all tt.

Optional stopping

A fair game stays fair at a random time, provided the time is a stopping time, one that does not peek into the future, and some control is in place.

The optional stopping theorem. If τ\tau is a stopping time and either τ\tau is bounded or MM is uniformly integrable, then E[Mτ]=E[M0]\mathbb{E}[M_\tau] = \mathbb{E}[M_0].

Apply it to a Brownian motion started at zero and stopped at τ\tau, the first time it hits a>0a > 0 or −b<0-b < 0. The martingale BtB_t gives P(Bτ=a)=ba+b\mathbb{P}(B_\tau = a) = \frac{b}{a + b}, and the martingale Bt2−tB_t^2 - t gives E[τ]=ab\mathbb{E}[\tau] = ab.

Without the conditions, the theorem fails. Let σ\sigma be the first time BB hits 11. Then Bσ=1B_\sigma = 1 while B0=0B_0 = 0. The reason is that σ\sigma, although finite, has infinite mean, and the stopped martingale is not uniformly integrable.

Local martingales

An Itô integral ∫0tHs dBs\int_0^t H_s\,dB_s with only ∫0tHs2 ds<∞\int_0^t H_s^2\,ds < \infty almost surely need not be a martingale. It is a local martingale: there are stopping times τn↑∞\tau_n \uparrow \infty such that each stopped process Mt∧τnM_{t \wedge \tau_n} is a martingale. By Fatou’s lemma, a nonnegative local martingale is a supermartingale, but it can be a strict one. A classic example: if BB is a three-dimensional Brownian motion started away from the origin, then 1/∣Bt∣1/|B_t| is a local martingale, because 1/∣x∣1/|x| is harmonic away from zero, yet its expectation decreases to zero over time. Such strict local martingales are used to model price bubbles in finance.

Maximal inequalities

Martingales give strong control over their running maximum Mt∗=sup⁡s≤t∣Ms∣M_t^* = \sup_{s \le t}|M_s|:

  • Doob’s maximal inequality: P(Mt∗≥λ)≤E∣Mt∣/λ\mathbb{P}(M_t^* \ge \lambda) \le \mathbb{E}|M_t| / \lambda for λ>0\lambda > 0.
  • Doob’s LpL^p inequality: E[(Mt∗)p]≤(pp−1)p E∣Mt∣p\mathbb{E}[(M_t^*)^p] \le \big(\tfrac{p}{p - 1}\big)^p\, \mathbb{E}|M_t|^p for p>1p > 1.
  • The Burkholder–Davis–Gundy inequalities: for continuous local martingales starting at zero and any p>0p > 0, E[(Mt∗)p]\mathbb{E}[(M_t^*)^p] is bounded above and below by constant multiples of E[⟨M⟩tp/2]\mathbb{E}\big[\langle M \rangle_t^{p/2}\big].

The L2L^2 case of Doob’s inequality is what made Picard iteration converge uniformly in the existence theorem for SDEs.

Lévy’s characterization and time change

Lévy’s characterization. A continuous local martingale MM with M0=0M_0 = 0 and ⟨M⟩t=t\langle M \rangle_t = t is a Brownian motion. In dd dimensions, the condition is ⟨Mi,Mj⟩t=δij t\langle M^i, M^j \rangle_t = \delta_{ij}\,t.

This is remarkable: the Gaussian distribution of the increments follows from the martingale property and the quadratic variation alone. A consequence is the Dambis–Dubins–Schwarz theorem. Every continuous local martingale with ⟨M⟩∞=∞\langle M \rangle_\infty = \infty can be written as

Mt=W⟨M⟩tM_t = W_{\langle M \rangle_t}

for some Brownian motion WW. Every continuous martingale is a Brownian motion running on a random clock, whose speed is set by the quadratic variation.

Time changes also build solutions of SDEs. To solve dXt=σ(Xt) dBtdX_t = \sigma(X_t)\,dB_t, run a Brownian motion WW and slow it down where σ\sigma is small: with At=∫0tσ(Ws)−2 dsA_t = \int_0^t \sigma(W_s)^{-2}\,ds and its inverse TtT_t, the process Xt=WTtX_t = W_{T_t} is a weak solution. This needs no Lipschitz condition at all, only that σ\sigma stays away from zero.

Martingale representation

The martingale representation theorem. If Ft\mathcal{F}_t is the filtration generated by a Brownian motion, every square-integrable martingale can be written as

Mt=M0+∫0tHs dBsM_t = M_0 + \int_0^t H_s\,dB_s

for some adapted process HH with E∫0tHs2 ds<∞\mathbb{E}\int_0^t H_s^2\,ds < \infty.

In the Brownian world, Itô integrals are therefore all the martingales there are. In finance, this is the mathematical reason the Black–Scholes market is complete: every payoff can be replicated by trading the stock, and HH is the hedging strategy.

← All posts in Stochastic differential equations