Introductory Martingales

The Central Idea

Let

be a stochastic process, and let

represent the information available through each time.

The process is a martingale when

Once the current information is known, the expected next value equals the current value.

Why Conditional Expectation Is Essential

A martingale is not defined by the unconditional equation

alone.

The defining condition is stronger:

The process must have zero predictable drift after conditioning on all information currently available.

What Is \(\mathcal{F}_n\)?

The symbol

denotes an information set, formally a sigma-algebra, that represents what is known by time \(n\).

In an introductory setting, it is useful to think of \(\mathcal{F}_n\) as containing the observable history

together with any other modeled information revealed by that time.

Information Grows Over Time

A filtration satisfies

As time advances, information may increase, but previously known information is not forgotten.

Adapted Processes

A process \(X_n\) is adapted to the filtration \(\mathcal{F}_n\) when the value \(X_n\) is known at time \(n\).

Informally,

may depend on information revealed up to time \(n\), but not on future information.

The Full Martingale Conditions

A discrete-time process \(\{X_n\}\) is a martingale with respect to \(\{\mathcal{F}_n\}\) when:

1. \(X_n\) is adapted to \(\mathcal{F}_n\);
2. \(E[|X_n|]<\infty\) for every \(n\);
3. \(E[X_{n+1}\mid\mathcal{F}_n]=X_n\).

Each condition is part of the definition.

Fair-Game Intuition

Suppose \(X_n\) represents a player's wealth after \(n\) rounds of a fair game.

If

then, conditional on the entire history so far, the expected wealth after one more round is exactly the current wealth.

No systematic gain or loss is predictable from the available information.

A Martingale Need Not Stay Constant

The equation

concerns conditional expectation, not actual pathwise equality.

Individual sample paths can fluctuate substantially even when the process is a martingale.

Simple Symmetric Random Walk

Let

where the increments are independent and satisfy

Then

Why the Symmetric Random Walk Is a Martingale

Since

and the next increment is independent of the current information,

Because

we obtain

Biased Random Walks

Suppose

Then

Unless

the raw position process has nonzero drift and is not a martingale.

Centering a Biased Random Walk

Let

Define

Then

Since the centered increment has conditional mean zero,

Removing predictable drift produces a martingale.

Martingale Differences

Define the increment

If \(X_n\) is a martingale, then

Such increments are called martingale differences.

Building a Martingale from Zero-Mean Increments

Suppose \(D_1,D_2,\ldots\) satisfy

Then the cumulative process

is a martingale under the usual integrability and adaptedness conditions.

Constant Processes

The simplest martingale is

for every \(n\).

Then

Conditional Expectation as Best Prediction

Conditional expectation can be interpreted as the best mean-square prediction based on currently available information.

Thus, the martingale condition

says that the current value already contains the predictable component of the next value.

Unconditional Expectations Stay Constant

If \(X_n\) is a martingale, then by the tower property,

Therefore,

Hence,

whenever the martingale is integrable.

The Tower Property

If

and \(m\le n\), then

This nested-conditioning rule is one of the central tools behind martingale calculations.

Multi-Step Martingale Prediction

Repeated use of the tower property gives, for \(m \[ \boxed{ E[X_n\mid\mathcal{F}_m]=X_m }. \]

Thus, the current value is not only the conditional expectation of the next value; it is the conditional expectation of every later value.

Submartingales

A process is a submartingale when

The process has nonnegative predictable drift.

The prefix “sub” refers to the current value lying below or equal to the conditional expected future value.

Supermartingales

A process is a supermartingale when

The process has nonpositive predictable drift.

Martingale, Submartingale, or Supermartingale

Example: Biased Walk as a Submartingale

If

then the random-walk increment has positive mean:

Therefore,

The position process is a submartingale.

Example: Biased Walk as a Supermartingale

If

then

Hence,

Convex Functions of Martingales

If \(X_n\) is a martingale and \(\varphi\) is convex under suitable integrability conditions, then Jensen's inequality gives

Therefore,

Thus \(\varphi(X_n)\) is often a submartingale.

Example: Squared Symmetric Random Walk

Let \(S_n\) be a simple symmetric random walk.

Then

with

Expanding,

A Corrected Square Martingale

Conditioning on the present gives

Therefore \(S_n^2\) is a submartingale, not a martingale.

However,

satisfies

Subtracting the predictable variance growth produces a martingale.

Predictable Drift

Many stochastic processes can be decomposed informally into

Martingale methods often work by removing the predictable drift so that the remaining process has conditional mean-zero increments.

Predictable Multipliers

Suppose \(D_{n+1}\) is a martingale difference and \(H_n\) is determined by information available at time \(n\).

Then

This shows why decisions based only on currently available information cannot create positive conditional expectation from fair increments.

Fair Betting Strategies

In a fair betting model, a gambler may choose a stake based on all previous outcomes.

If the next gain has conditional mean zero and the stake is chosen before the next outcome is revealed, the resulting gain still has conditional mean zero.

Adaptivity alone does not manufacture favorable expected value from a fair game.

Future Information Is Not Allowed

A betting or decision rule must be based on

before the next random outcome occurs.

A strategy that uses future information is not predictable and falls outside the usual martingale framework.

Stopping Times

A random time

is a stopping time if the decision to stop by time \(n\) can be made using only information available by time \(n\).

Informally, one may know whether

without looking into the future.

Example: First Hitting Time

For a random walk \(S_n\), define

This is a stopping time because at time \(n\) one can determine whether the walk has already reached \(5\) using only the observed path through time \(n\).

A Nonexample of a Stopping Time

Suppose \(\tau\) is defined as the time of the last visit to \(0\) before a fixed future horizon.

At an intermediate time, one cannot know whether the current visit will be the last without observing the future.

Such a rule is generally not a stopping time.

Stopped Processes

Given a process \(X_n\) and stopping time \(\tau\), define the stopped process

where

The process follows \(X_n\) until stopping and then remains at the stopped value.

Optional Stopping Intuition

Under appropriate technical conditions, stopping a martingale does not create expected gain:

This principle is associated with the optional stopping theorem.

The technical conditions matter and must not be ignored.

Why Optional Stopping Needs Conditions

A martingale can have rare but extremely large values or a stopping time can have an uncontrolled tail.

In such cases, simply writing

may be invalid.

Bounded stopping times and other integrability conditions provide common safe settings.

Bounded Stopping Times

If a stopping time satisfies

for some fixed finite \(N\), then optional-stopping arguments are particularly well behaved under standard integrability assumptions.

Gambler's Ruin Revisited

Consider a simple symmetric random walk on

stopped when it first reaches \(0\) or \(N\).

The stopped position can be used with martingale reasoning to derive the probability of reaching \(N\) before \(0\).

Boundary-Hitting Probability from Martingale Balance

Suppose the walk begins at \(i\), and let

At stopping, the position is either \(0\) or \(N\), so

Under a valid optional-stopping argument,

Therefore,

Martingales and Markov Chains Are Different Concepts

A Markov process is defined by a conditional-dependence structure: the present state contains the information needed for the next transition.

A martingale is defined by a conditional-expectation relation.

A process can be:

• both Markov and martingale;
• Markov but not martingale;
• martingale without being described through a Markov state model.

Martingales and Independence Are Different Concepts

A martingale does not require the values

to be independent.

In fact, successive values are often strongly dependent because

is built from \(X_n\) plus a new innovation.

Martingales and Constant Means

Every integrable martingale has constant unconditional mean.

The converse is false.

A process can satisfy

for every \(n\) while failing the conditional martingale property.

Example: Constant Mean Without Martingale Structure

Suppose \(X_0\) is equally likely to be \(1\) or \(-1\), and define

Then

But

which is not equal to \(X_0\) unless \(X_0=0\).

Constant mean alone does not imply martingale behavior.

Doob Martingales

Suppose \(Y\) is an integrable random variable and

Define

Then

As information accumulates, the conditional estimate of a fixed future quantity evolves as a martingale.

Why Doob Martingales Are Natural

The tower property gives

Therefore,

Information Revelation Example

Suppose \(Y\) is the final total of several future random contributions.

As each contribution is revealed, the best conditional forecast of the final total changes.

The sequence of forecasts

forms a martingale.

Martingales in Finance

Martingale ideas appear in mathematical finance when properly adjusted asset prices are modeled under probability measures for which predictable excess gains are removed.

Introductory martingale theory provides the conditional-expectation language behind these more advanced models.

Martingales in Statistics

Martingale differences appear in sequential estimation, adaptive experiments, stochastic approximation, and time-series analysis.

They provide a framework for random errors whose conditional mean is zero even when independence is too strong an assumption.

Martingales in Algorithms

Randomized algorithms often reveal information one step at a time.

Conditional expectations of a final quantity can form martingales, allowing probabilistic bounds on deviations and stopping behavior.

Martingales and Concentration

Advanced probability uses martingales to prove concentration inequalities for dependent processes.

The key advantage is that conditional mean-zero innovations can be controlled even when the entire process is not independent.

Martingale Convergence Intuition

Under suitable boundedness or integrability conditions, martingales can converge as time grows.

Martingale convergence theorems are major results in probability, but their full technical statements are beyond this introductory lesson.

Simulation of Martingales

A martingale can be explored computationally by simulating many sample paths.

For a symmetric random walk:

1. start each path at the same \(X_0\);
2. generate independent \(\pm1\) increments;
3. record \(X_n\) across time;
4. average the simulated values at each time;
5. compare the empirical mean with \(X_0\).

What Simulation Can and Cannot Verify

Simulation can provide evidence that sample averages remain approximately stable.

But finite simulation cannot prove the martingale property.

The defining condition is analytical:

Common Error: Checking Only the Ordinary Mean

Showing

does not prove that \(X_n\) is a martingale.

The conditional expectation must be checked.

Common Error: Ignoring the Filtration

Martingale status depends on what information is available.

A process can behave differently relative to different filtrations.

Always state or understand the information structure being used.

Common Error: Assuming Martingale Means Independent

Martingale values are generally not independent.

The requirement concerns conditional expectation, not independence of successive observations.

Common Error: Assuming Martingale Means No Variability

A martingale can fluctuate widely.

The symmetric random walk has

but

Zero drift does not imply small uncertainty.

Common Error: Assuming Fair Means Safe

A fair game in expectation can still produce large losses on particular sample paths.

Expected value does not measure risk by itself.

Common Error: Using Future Information

Martingale decisions must be based on current or past information.

A rule that depends on a future outcome is not an admissible predictable strategy in the standard framework.

Common Error: Applying Optional Stopping Without Conditions

It is not valid to assume automatically that

for every martingale and every stopping time.

Optional-stopping theorems require conditions controlling the process, stopping time, or integrability.

A Reliable Procedure for Checking a Martingale

1. Identify the process \(X_n\).
2. Identify the filtration \(\mathcal{F}_n\).
3. Check that \(X_n\) is adapted.
4. Check that \(E[|X_n|]<\infty\).
5. Write \(X_{n+1}\) in terms of \(X_n\) and the new random increment.
6. Condition on \(\mathcal{F}_n\).
7. Use independence or known conditional expectations for the new increment.
8. Verify whether the result equals, exceeds, or falls below \(X_n\).

A Reliable Procedure for Centering a Process

1. Find the predictable expected increment.
2. Accumulate the predictable drift through time.
3. Subtract that drift from the original process.
4. Check that the remaining increments have conditional mean zero.
5. Verify the martingale conditions for the centered process.

A Reliable Procedure for Stopping-Time Reasoning

1. Define the proposed random stopping time \(\tau\).
2. Check that the decision to stop can be made using only current and past information.
3. Identify the martingale being stopped.
4. Check whether an optional-stopping theorem applies under the available conditions.
5. Determine the possible values of the stopped process.
6. Use expectation only after the theorem's conditions have been justified.

Example: Full Martingale Workflow

Let

where the \(Y_k\) are independent and

First compute the expected increment:

Therefore,

Hence,

Practice 1: Symmetric Random Walk

Explain why a simple symmetric random walk is a martingale with respect to its natural filtration.

Practice 2: Positive Drift

Suppose

and

Is \(X_n\) a martingale, submartingale, or supermartingale?

Practice 3: Negative Drift

Suppose

Classify the resulting cumulative process.

Practice 4: Centering

A process satisfies

with independent increments having mean

Give a centered martingale constructed from \(X_n\).

Practice 5: Martingale Differences

If \(M_n\) is a martingale, what is

Practice 6: Constant Mean

Does

for every \(n\) prove that \(X_n\) is a martingale? Explain.

Practice 7: Square of a Symmetric Random Walk

If \(S_n\) is a simple symmetric random walk, is \(S_n^2\) a martingale? What corrected process is a martingale?

Practice 8: Stopping Time

Let

Is \(\tau\) a stopping time with respect to the natural filtration of \(X_n\)?

Practice 9: Multi-Step Prediction

If \(M_n\) is a martingale and \(m \[ E[M_n\mid\mathcal{F}_m]? \]

Practice 10: Identify the Error

A student says: “Because martingales are fair, a clever strategy for choosing when to stop always preserves expected wealth.”

Explain the problem with this statement.

Practice Answers

Practice 1

For a symmetric random walk,

where the next increment is independent of the past and has mean zero. Therefore,

Hence the walk is a martingale.

Practice 2

Therefore,

Practice 3

Therefore,

Practice 4

Since the predictable mean drift after \(n\) steps is \(3n\), define

Then the increments of \(M_n\) have mean zero.

Practice 5

Practice 6

No. Constant unconditional mean is necessary for an integrable martingale but is not sufficient.

One must verify

Practice 7

No. For a simple symmetric random walk,

so \(S_n^2\) is a submartingale.

The corrected process

is a martingale.

Practice 8

Yes. At time \(n\), one can determine whether \(X_k\ge10\) has occurred for some \(k\le n\) using only the observed history.

Therefore, \(\tau\) is a stopping time.

Practice 9

This follows from repeated conditioning and the tower property.

Practice 10

Optional-stopping conclusions require technical conditions.

A stopping time can have an uncontrolled tail or interact with an unbounded martingale in ways that invalidate the simple expectation identity.

One must verify the conditions of an appropriate optional-stopping theorem before concluding

What to Remember

A martingale is a stochastic process whose conditional expected future value equals its present value:

• \(\mathcal{F}_n\) represents the information available by time \(n\).
• A martingale must be adapted and integrable.
• Martingale increments have conditional mean zero.
• Every integrable martingale has constant unconditional expectation.
• Constant unconditional expectation alone does not imply a martingale.
• The simple symmetric random walk is a martingale.
• A biased random walk can be centered by subtracting its predictable drift.
• Submartingales have nonnegative predictable drift.
• Supermartingales have nonpositive predictable drift.
• For a symmetric random walk, \(S_n^2-n\) is a martingale.
• A stopping time must be identifiable using only present and past information.
• Optional-stopping results require conditions and cannot be applied automatically.
• Markov and martingale properties describe different aspects of stochastic processes.
• Doob martingales arise naturally from conditional expectations of a fixed future quantity.
• Simulation can illustrate martingale behavior but cannot replace conditional-expectation proof.

The central principle is: after conditioning on everything currently known, a martingale has no predictable drift.

Where We Go Next

Random sequences, Markov chains, Poisson processes, queues, reliability models, and martingales all describe stochastic systems from different structural viewpoints. The final lesson in this section turns those models into computational experiments.

The next lesson, Simulation of Stochastic Systems, develops methods for generating sample paths, estimating probabilities and expectations, checking convergence, comparing models, and diagnosing simulation error.