ChungMG

Martingale

A martingale is a stochastic process that represents a “fair game” - the expected future value, given all past information, equals the current value. Martingales are fundamental in stochastic calculus, financial mathematics, and the theory of [[Stochastic Differential Equation (SDE)|Stochastic Differential Equations]].

---

1. Intuitive Understanding

1.1 Fair Game Analogy

Imagine a fair coin toss game:

• You win $1 if heads, lose $1 if tails
• Your expected wealth after the next toss equals your current wealth
• No matter what happened in the past, the game remains “fair”

This is the essence of a martingale: the best prediction of tomorrow’s value is today’s value.

[!NOTE] Key Intuition
A martingale has no predictable trend or drift. All changes are purely random with zero expected value.

---

2. Formal Definition

2.1 Discrete-Time Martingale

A stochastic process $\{M_n{\}}_{n\ge 0}$ is a martingale with respect to filtration $\{{\mathcal{F}}_n\}$ if:

1. Integrability: $\mathbb{E}[|M_n|]<\mathrm{\infty }$ for all $n$
2. Adaptedness: $M_n$ is ${\mathcal{F}}_n$ -measurable
3. Martingale Property: $\mathbb{E}[M_{n+1}\mid {\mathcal{F}}_n]=M_n$

2.2 Continuous-Time Martingale

A stochastic process $\{M_t{\}}_{t\ge 0}$ is a martingale with respect to filtration $\{{\mathcal{F}}_t\}$ if:

1. Integrability: $\mathbb{E}[|M_t|]<\mathrm{\infty }$ for all $t$
2. Adaptedness: $M_t$ is ${\mathcal{F}}_t$ -measurable
3. Martingale Property: $\mathbb{E}[M_t\mid {\mathcal{F}}_s]=M_s$ for all $s<t$

[!QUOTE] Martingale Property

$$\mathbb{E}[M_t\mid {\mathcal{F}}_s]=M_s,\mathrm{\forall }s<t$$

Given all information up to time $s$ , the expected value at time $t$ is exactly the current value $M_s$ .

---

3. Related Processes

3.1 Submartingale and Supermartingale

Type	Definition	Intuition
Martingale	$\mathbb{E}[M_t\mid {\mathcal{F}}_s]=M_s$	Fair game
Submartingale	$\mathbb{E}[X_t\mid {\mathcal{F}}_s]\ge X_s$	Favorable game (upward drift)
Supermartingale	$\mathbb{E}[Y_t\mid {\mathcal{F}}_s]\le Y_s$	Unfavorable game (downward drift)

3.2 Local Martingale

A process $M_t$ is a local martingale if there exists a sequence of stopping times ${\tau }_n\uparrow \mathrm{\infty }$ such that the stopped process $M_{t\wedge {\tau }_n}$ is a martingale for each $n$ .

[!WARNING] Important Distinction
Not all local martingales are true martingales. Local martingales may have “explosive” behavior that violates integrability conditions.

---

4. Key Examples

4.1 [[Wiener Process|Wiener Process]]

The [[Wiener Process|Wiener Process]] $W_t$ is a martingale:

$$\mathbb{E}[W_t\mid {\mathcal{F}}_s]=W_s,\mathrm{\forall }s<t$$

Proof: Since $W_t-W_s\sim \mathcal{N}(0,t-s)$ and is independent of ${\mathcal{F}}_s$ :

$$\mathbb{E}[W_t\mid {\mathcal{F}}_s]=\mathbb{E}[W_s+(W_t-W_s)\mid {\mathcal{F}}_s]=W_s+0=W_s$$

4.2 Exponential Martingale

For [[Wiener Process|Wiener Process]] $W_t$ , the process:

$$M_t=\exp (\theta W_t-\frac{1}{2}{\theta }^{2}t)$$

is a martingale for any constant $\theta \in \mathbb{R}$ .

4.3 Stochastic Integral

If $H_t$ is adapted and $W_t$ is a [[Wiener Process|Wiener Process]], then:

$$M_t={\int }_0^t{H}_s\mathrm{d}{W}_s$$

is a local martingale (and a true martingale under appropriate conditions on $H_t$ ).

4.4 Geometric [[Wiener Process|Brownian Motion]] (Adjusted)

While $S_t=S_0\exp (\mu t+\sigma W_t)$ is not a martingale, the discounted process:

$${\tilde{S}}_t=e^{-rt}{S}_t=S_0\exp ((\mu -r)t+\sigma W_t)$$

is a martingale under the risk-neutral measure when $\mu =r$ .

---

5. Fundamental Properties

5.1 Optional Stopping Theorem

Theorem: If $M_t$ is a martingale and $\tau$ is a bounded stopping time, then:

$$\mathbb{E}[M_{\tau }]=\mathbb{E}[M_0]$$

Applications:

• Gambling strategies cannot beat a fair game
• Pricing of American options
• Exit time analysis

5.2 Doob’s Inequality

For a martingale $M_t$ and any $\lambda >0$ :

$$\mathbb{P}(\underset{0\le s\le t}{sup}|M_s|\ge \lambda )\le \frac{\mathbb{E}[|M_t|]}{\lambda }$$

This provides bounds on the maximum value of a martingale.

5.3 Martingale Convergence Theorem

Theorem: If $M_t$ is a martingale bounded in $L^1$ (i.e., $\underset{t}{sup}\mathbb{E}[|M_t|]<\mathrm{\infty }$ ), then:

$$M_t\to M_{\mathrm{\infty }}\text{almost surely as }t\to \mathrm{\infty }$$

5.4 Quadratic Variation

For a continuous martingale $M_t$ , the quadratic variation $[M{]}_t$ exists and:

$$M_t^2-[M{]}_t\text{is a martingale}$$

For [[Wiener Process|Wiener Process]]: $[W{]}_t=t$ .

---

6. Martingales in SDEs

6.1 Drift Condition

An Itô process:

$$\mathrm{d}{X}_t=\mu (t,X_t)\mathrm{d}t+\sigma (t,X_t)\mathrm{d}{W}_t$$

is a martingale if and only if $\mu (t,x)=0$ for all $(t,x)$ .

[!TIP] Key Insight
Martingales have zero drift. The deterministic component must vanish.

6.2 Itô Integral as Martingale

The [[Itô Integral|Itô integral]]:

$$M_t={\int }_0^t\sigma (s,X_s)\mathrm{d}{W}_s$$

is a martingale (under square-integrability conditions).

6.3 Martingale Representation Theorem

Theorem: Any martingale $M_t$ adapted to the filtration generated by [[Wiener Process|Wiener Process]] $W_t$ can be represented as:

$$M_t=M_0+{\int }_0^t{\varphi }_s\mathrm{d}{W}_s$$

for some adapted process ${\varphi }_t$ .

---

7. Applications

7.1 Financial Mathematics

Application	Description
Risk-Neutral Pricing	Asset prices are martingales under risk-neutral measure
Fundamental Theorem	No arbitrage $⟺$ existence of equivalent martingale measure
Option Pricing	Black-Scholes formula derived using martingale methods
Hedging	Construct replicating portfolios via martingale representation

7.2 Optimal Stopping

Problems of the form:

$$V=\underset{\tau }{sup}\mathbb{E}[g(X_{\tau })]$$

where $\tau$ is a stopping time, are solved using martingale techniques (Snell envelope).

7.3 Statistical Inference

• Likelihood ratio tests: Likelihood ratios form martingales
• Sequential analysis: Wald’s sequential probability ratio test
• Bootstrap methods: Martingale-based resampling

7.4 Diffusion Models

In [[Diffusion Model|Diffusion Models]]:

• The noise process $W_t$ is a martingale
• [[Score Function|Score matching]] objectives relate to martingale estimating functions
• Reverse-time processes involve martingale decompositions

Martingale Role in Training:

1. Forward process as martingale: The forward [[Stochastic Differential Equation (SDE)|SDE]] $\mathrm{d}x=f(t)x\mathrm{d}t+g(t)\mathrm{d}{W}_t$ has a martingale component $\int g(t)\mathrm{d}{W}_t$

2. Score matching as martingale estimating equation: The score matching loss can be interpreted as a martingale estimating function, ensuring unbiased gradient estimates

3. ELBO decomposition: The variational bound in [[Diffusion Model|DDPM]] decomposes into KL divergences, each involving martingale properties of the forward process

4. Reverse-time martingales: Under time reversal, certain processes become martingales, enabling tractable reverse sampling

[!NOTE] Why Martingale Matters for Diffusion
The martingale property of $W_t$ ensures that the noise injection in the forward process has no predictable drift — it’s purely random. This is what makes the reverse process learnable: the model only needs to predict the deterministic drift, not the random component.

---

8. Testing for Martingale Property

8.1 Practical Checklist

To verify if a process $X_t$ is a martingale:

1. ✓ Check integrability: $\mathbb{E}[|X_t|]<\mathrm{\infty }$
2. ✓ Verify adaptedness: $X_t$ depends only on information up to time $t$
3. ✓ Compute conditional expectation: $\mathbb{E}[X_t\mid {\mathcal{F}}_s]\stackrel{?}{=}{X}_s$
4. ✓ For Itô processes: Check if drift term is zero

8.2 Common Pitfalls

Process	Martingale?	Reason
$W_t$	✓ Yes	Zero drift
$W_t^2$	✗ No	$\mathbb{E}[W_t^2\mid {\mathcal{F}}_s]=W_s^2+(t-s)$
$W_t^2-t$	✓ Yes	Compensated quadratic variation
$e^{W_t}$	✗ No	$\mathbb{E}[e^{W_t}]=e^{t/2}$
$e^{W_t-t/2}$	✓ Yes	Exponential martingale

---

9. Core Formula Cards

[!QUOTE] Martingale Property

$$\mathbb{E}[M_t\mid {\mathcal{F}}_s]=M_s,\mathrm{\forall }s<t$$

[!QUOTE] Optional Stopping Theorem

$$\mathbb{E}[M_{\tau }]=\mathbb{E}[M_0]\text{(for bounded stopping time }\tau )$$

[!QUOTE] Exponential Martingale

$$M_t=\exp (\theta W_t-\frac{1}{2}{\theta }^{2}t)$$

[!QUOTE] Itô Integral Martingale

$$M_t={\int }_0^t{H}_s\mathrm{d}{W}_s\text{is a martingale}$$

[!QUOTE] Quadratic Variation Martingale

$$M_t^2-[M{]}_t\text{is a martingale}$$

---

Related Concepts

• [[Stochastic Process]]
• [[Wiener Process|Wiener Process]]
• [[Stochastic Differential Equation (SDE)]]
• [[Itô Integral]]
• [[Itô’s Lemma]]
• [[Markov Process]]
• [[Brownian Motion]]
• [[Filtration]]
• [[Stopping Time]]
• [[Doob’s Decomposition]]
• [[Diffusion Model]]
• [[Score Function]]

---

Dataview Query

LIST
FROM #martingale OR #stochastic_process
SORT file.ctime DESC

---

References

• Book: [[Wiener Process|Brownian Motion]] and Stochastic Calculus - Karatzas & Shreve
• Book: Stochastic Differential Equations - Bernt Øksendal
• Book: Probability with Martingales - David Williams
• Course: MIT 18.S096 Topics in Mathematics with Applications in Finance
• Paper: Martingale Methods in Financial Modelling - Musiela & Rutkowski