Convergence in distribution is the basic language of limit laws. It is weak enough to describe asymptotic shapes of random variables without requiring a pointwise coupling, yet strong enough to support stable consequences such as convergence of probabilities of continuity sets and convergence of expectations of bounded continuous test functions. The most direct tool for proving convergence in distribution is the characteristic function, the Fourier transform of a probability measure. Characteristic functions exist for every random variable, multiply under sums of independent variables, and determine the law uniquely.

This article explains weak convergence, why characteristic functions determine distributions, and how they yield a clean proof of the central limit theorem (CLT). The emphasis is on the logic: what needs to be shown, what is automatic, and where moment assumptions enter.

Convergence in distribution and the portmanteau viewpoint

Write $X_n \Rightarrow X$ for convergence in distribution. By definition, this means

for all continuity points $t$ of the distribution function of $X$.

An equivalent and often more usable characterization is: for every bounded continuous function $f$,

This is one form of the portmanteau theorem. It clarifies that weak convergence is about convergence of integrals against a large class of test functions. This is exactly why Fourier transforms enter: they are integrals against the bounded continuous functions $x\mapsto e^{itx}$.

Weak convergence is weaker than convergence in probability. It cannot distinguish sequences that differ on sets of small probability, and it does not control moments unless additional uniform integrability-type bounds are supplied. Its strength is that it identifies limiting laws.

Tightness: preventing mass from escaping

A sequence of laws $\mu_n$ on $\mathbb{R}$ can fail to have convergent subsequences if mass escapes to infinity. Tightness rules this out: $\{\mu_n\}$ is tight if for every $\varepsilon>0$ there exists $M$ such that $\mu_n([-M,M])\ge 1-\varepsilon$ for all $n$. On $\mathbb{R}$, tightness plus pointwise convergence of distribution functions at continuity points is a robust route to weak convergence.

In many limit theorems, tightness is obtained from variance bounds via Chebyshev’s inequality or from uniform moment bounds.

Characteristic functions: definition and core properties

For a real-valued random variable $X$, the characteristic function is

It exists without moment assumptions because $|e^{itX}|=1$.

Key properties:

• $\varphi_X(0)=1$, and $|\varphi_X(t)|\le 1$.
• $\varphi_X$ is uniformly continuous.
• Scaling: $\varphi_{aX}(t)=\varphi_X(at)$.
• Independence: if $X$ and $Y$ are independent, $\varphi_{X+Y}(t)=\varphi_X(t)\varphi_Y(t)$.

The multiplicative property is the main algebraic advantage: sums become products.

Uniqueness: characteristic functions determine the law

If $\varphi_X(t)=\varphi_Y(t)$ for all $t$, then $X$ and $Y$ have the same distribution. There are several proofs; one route uses Fourier inversion for probability measures and approximations by smooth test functions. The important practical takeaway is that identifying the pointwise limit of characteristic functions identifies the limiting distribution.

Lévy’s continuity theorem: the convergence bridge

Lévy’s continuity theorem states:

• If $\varphi_{X_n}(t)\to \varphi(t)$ pointwise for all $t$, and $\varphi$ is continuous at 0, then $\varphi$ is the characteristic function of some random variable $X$, and $X_n\Rightarrow X$.

Conversely, if $X_n\Rightarrow X$, then $\varphi_{X_n}(t)\to \varphi_X(t)$ for all $t$.

This theorem is what makes characteristic functions a method for proving weak convergence: prove an analytic limit and recognize it.

The central limit theorem in this language

Let $X_1,X_2,\dots$ be independent identically distributed with $\mathbb{E}[X_i]=0$ and $\mathrm{Var}(X_i)=\sigma^2\in(0,\infty)$. Define

The CLT claims $S_n\Rightarrow Z$ where $Z$ is standard normal.

The characteristic function of $S_n$ is

Thus everything reduces to understanding $\varphi_{X_1}(u)$ near $u=0$.

The small-$u$ expansion

Using $e^{iuX}=1+iuX-\frac{u^2X^2}{2}+R(uX)$, where the remainder satisfies $|R(y)|\le C|y|^3$ for small $y$ and $|R(y)|\le 2$ globally, one can show under $\mathbb{E}[X^2]<\infty$ that

The mean-zero condition eliminates the linear term, and the variance supplies the quadratic term. The $o(u^2)$ term is where integrability and truncation arguments enter: one splits $X$ into a bounded part and a tail part and uses dominated convergence on the bounded part while controlling the tail using the finite second moment.

Passing to the $n$-th power

Set $u=t/(\sigma\sqrt{n})$. Then

Therefore,

The limit $e^{-t^2/2}$ is the characteristic function of the standard normal distribution. By Lévy’s theorem, $S_n\Rightarrow Z$.

This proof isolates the universality mechanism: after normalization by $\sqrt{n}$, only the second moment contributes at leading order to the characteristic function near 0.

Quantitative refinements and what weak convergence does not give

The characteristic-function proof identifies the limit but does not provide an error bound on $P(S_n\le t)-\Phi(t)$. Such quantitative control requires additional input (for instance, bounds involving third moments and smoothing inequalities). The gap is structural: weak convergence is a qualitative statement. To make it quantitative, one needs uniform control of the approximation error in Fourier space and a way to translate that control back to distribution functions.

It is still valuable to know what weak convergence does guarantee. If $f$ is bounded and continuous, $\mathbb{E}[f(S_n)]\to \mathbb{E}[f(Z)]$. Many asymptotic statistics are built by applying continuous mappings \to $S_n$; the continuous mapping theorem then transfers weak convergence through continuous transformations.

Why characteristic functions remain a practical tool

Even outside the CLT, characteristic functions offer a repeatable method:

• express the quantity of interest as a sum of independent components or as a scaled transformation,
• compute or estimate the characteristic function,
• identify its limit,
• invoke Lévy’s theorem.

This method excels for sums, for stable limits, and for problems where distribution functions are hard to handle directly but Fourier transforms are tractable.

Characteristic functions do not replace other methods, but they provide one of the clearest pipelines from probabilistic structure to limiting law: algebraic manipulation in the exponent, analytic limits, then an inversion theorem to return to probability.

Inversion and why continuity at zero matters

The statement “characteristic functions determine distributions” is not just a slogan; it comes from an inversion principle. For sufficiently nice densities $f$, Fourier inversion says

For general probability measures, one uses smoothed approximations and shows that integrals of test functions can be recovered from $\varphi_X$. The continuity-at-zero condition in Lévy’s theorem is a minimal regularity requirement that rules out pathological pointwise limits that fail to correspond to probability measures. Intuitively, continuity at zero ensures the limiting function behaves like an average of unit-modulus exponentials and therefore can be a characteristic function.

In many applications, continuity at zero is automatic because each $\varphi_{X_n}$ is continuous and the convergence is controlled well enough to preserve continuity at zero.

Slutsky and continuous mapping: how weak limits propagate

Two structural facts make weak convergence useful in applied probability.

• Slutsky’s theorem: If $X_n\Rightarrow X$ and $Y_n\to c$ in probability, then $X_n+Y_n\Rightarrow X+c$ and $X_nY_n\Rightarrow cX$.
• Continuous mapping theorem: If $X_n\Rightarrow X$ and $g$ is continuous, then $g(X_n)\Rightarrow g(X)$.

These results are why one can prove a CLT for sums and then immediately obtain a CLT for standardized statistics, studentized quantities under suitable conditions, and many functionals built from the sum.

Characteristic functions are compatible with these theorems. For example, convergence in probability \to a constant corresponds to characteristic functions converging pointwise \to $e^{itc}$, and products and compositions behave as expected.

Triangular arrays and the Lindeberg idea

The i.i.d. CLT is a baseline. Many applications involve sums of independent but not identically distributed terms. Consider a triangular array $X_{n,1},\dots,X_{n,n}$ independent with mean zero and variances $\sigma_{n,k}^2$, and let $s_n^2=\sum_{k=1}^n \sigma_{n,k}^2$. One studies

A central condition ensuring a normal limit is a Lindeberg-type tail control: for every $\varepsilon>0$,

This condition prevents rare large summands from dominating the normalized sum. In the characteristic-function proof, this is exactly what is needed to justify a uniform small-$u$ expansion: the quadratic term from variance is stable, and the remainder terms from large deviations vanish in aggregate.

The conceptual message is stable across formulations: normal limits arise when the sum has many small contributors and no single contributor carries a macroscopic fraction of the variance.

Tightness from variance bounds

In the CLT setting, tightness is straightforward: $\mathrm{Var}(S_n)=1$, so Chebyshev implies

Thus the laws of $S_n$ are tight. Tightness ensures that pointwise convergence of characteristic functions is not hiding mass escape. On $\mathbb{R}$, Lévy’s theorem already packages the needed compactness, but it is useful to remember that second-moment normalization carries a built-in tightness guarantee.

Why characteristic functions are more than a proof device

Characteristic functions provide a practical calculus of limits:

• They turn convolution (sums of independent variables) into multiplication.
• They turn scaling into a simple argument rescaling.
• They provide access to stable limits when densities or distribution functions are not tractable.

They are also a diagnostic tool. If you can compute or approximate $\log \varphi_{X_n}(t)$, then weak limits often become limits of exponentials. In many problems, $\log \varphi$ has an additive decomposition, which mirrors independence at the level of cumulants.

A common pitfall: weak convergence does not imply moment convergence

Even when $X_n\Rightarrow X$, it need not be true that $\mathbb{E}[X_n]\to \mathbb{E}[X]$ or that variances converge. Moment convergence requires uniform integrability or explicit moment bounds. In CLT contexts, normalization often forces variance to be controlled, but higher moments can still fail to converge.

A safe rule is:

• weak convergence plus uniform integrability of $|X_n|$ implies convergence of expectations,
• weak convergence plus uniform integrability of $X_n^2$ implies convergence of second moments.

This is one reason quantitative CLT bounds often assume finite third absolute moment: it controls tails strongly enough to translate Fourier error estimates into distribution-function error bounds.

The central picture

Weak convergence is convergence of laws. Characteristic functions encode laws via Fourier transforms and are stable under sums of independent components. Lévy’s continuity theorem makes pointwise characteristic-function limits equivalent to weak limits. The CLT follows from a second-order expansion near zero and a product-\to-exponential limit. Extensions to non-identical summands require a tail condition that keeps the quadratic variance term dominant. With these pieces, many limiting-distribution problems become analytic limit problems with a clear algebraic structure.

Martingales are the most efficient language for “no predictable drift.” They formalize fair games, but their reach is broader: they govern many stochastic processes, provide clean proofs of limit theorems, and yield sharp bounds on fluctuations. The power of martingales comes from two facts. First, the defining identity is conditional: it tracks what can be predicted from available information. Second, there are deep inequalities that control the maximum of a martingale in terms of its final value or its quadratic variation. Together, these convert dynamic questions into static estimates.

This article presents martingales and stopping \times as a working system: how to recognize martingales, what optional stopping truly allows, why maximal inequalities are indispensable, and how convergence is proved in practice.

Filtrations and adapted processes

A filtration $(\mathcal{F}_n)_{n\ge 0}$ is an increasing family of σ-algebras representing information revealed over time. A process $X_n$ is adapted if $X_n$ is $\mathcal{F}_n$-measurable for each $n$, meaning the value at time $n$ is determined by information available at time $n$.

In many models, $\mathcal{F}_n$ is the σ-algebra generated by the first $n$ observations. This “natural filtration” is usually the right one unless extra information is introduced.

Martingales, submartingales, supermartingales

An integrable adapted process $(X_n,\mathcal{F}_n)$ is a martingale if

It is a submartingale if the conditional expectation is at least $X_n$, and a supermartingale if it is at most $X_n$.

These definitions are precise versions of “fair,” “favorable,” and “unfavorable” relative to the information flow.

Canonical examples

• If $\xi_1,\xi_2,\dots$ are independent with $\mathbb{E}[\xi_k]=0$, then $S_n=\sum_{k=1}^n \xi_k$ is a martingale.
• If $Y\in L^1$, then $X_n=\mathbb{E}[Y\mid \mathcal{F}_n]$ is a martingale. This explains why martingales appear whenever information is revealed progressively: conditional expectations are martingales by construction.
• For simple symmetric random walk, $S_n^2-n$ is a martingale. This reflects the accumulation of variance and is a prototype for quadratic-variation ideas.

The last example is a common technique: construct martingales by finding functions $f$ such that $f(X_n)-\text{compensator}$ has zero conditional drift.

Stopping \times: admissible decision rules

A stopping time $\tau$ is a random time such that $\{\tau\le n\}\in \mathcal{F}_n$ for each $n$. This means the decision to stop by time $n$ can be made using information available at time $n$. It rules out anticipation.

Typical stopping \times include first hitting \times:

and threshold crossing \times.

Stopping \times let one encode questions like “when does the walk hit zero?” or “when do we first exceed a safety limit?” in a way compatible with conditional expectation.

The stopped process remains a martingale

Given a martingale $X_n$ and a stopping time $\tau$, the stopped process

is also a martingale. This closure property is fundamental. It lets one freeze the process after a stopping event while keeping martingale structure. Most optional stopping arguments are proofs about the stopped process, not about $X_\tau$ directly.

Optional stopping: the theorem and its real hypotheses

A tempting but false statement is: “for any stopping time $\tau$, $\mathbb{E}[X_\tau]=\mathbb{E}[X_0]$.” The failure mode is simple: a stopping time can select rare but huge deviations, biasing the average. Optional stopping is therefore a theorem about controlling tails and integrability.

A safe version is:

• If $\tau$ is bounded (there exists $N$ with $\tau\le N$ a.s.), then $\mathbb{E}[X_\tau]=\mathbb{E}[X_0]$.

More general versions hold under conditions such as:

• $\tau$ has finite expectation and the martingale increments are uniformly bounded, or
• the family $\{X_{n\wedge \tau}\}$ is uniformly integrable, or
• $\mathbb{E}[\sup_{k\le \tau}|X_k|]<\infty$.

Rather than memorizing variants, it is more useful to remember the mechanism: one proves $\mathbb{E}[X_{n\wedge \tau}]=\mathbb{E}[X_0]$ for each $n$ and then passes to the limit $n\to\infty$. The step from $n\wedge \tau$ \to $\tau$ is exactly where integrability is needed.

A standard computation: gambler’s ruin probability

Let $S_n$ be simple random walk on $\{0,1,\dots,N\}$ absorbed at 0 and $N$. Starting at $S_0=i$, define $\tau=\inf\{n:\ S_n\in\{0,N\}\}$. The process $S_n$ is a martingale. Under mild justification for optional stopping on the bounded state space, one gets

But $S_\tau\in\{0,N\}$, so $\mathbb{E}[S_\tau]=N\,P(S_\tau=N)$. Hence $P(S_\tau=N)=i/N$. The point is not the specific model; it is the pattern: choose a martingale whose stopped value takes only a few outcomes, then solve for the probability.

A parallel computation with a quadratic martingale yields the expected absorption time, illustrating how martingales convert time-\to-hit questions into algebra.

Doob’s maximal inequality: controlling the maximum of a path

Martingale convergence and many tail estimates require bounds on $\max_{k\le n} X_k$. Doob’s inequalities provide exactly this. For a nonnegative submartingale $X_k$,

This inequality is conceptually simple: a large maximum forces a large final expectation, and the submartingale property prevents the process from “hiding” large values without paying in expectation.

For $p>1$, the $L^p$ maximal inequality gives

This is one of the main technical levers in martingale theory: it turns a bound on the terminal value into a bound on the whole path.

Quadratic variation and predictable compensators

Many martingales are built by subtracting the predictable drift. For random walk, $S_n^2-n$ is a martingale because

The term $n$ is the compensator that removes the predictable increase in $S_n^2$. In general, quadratic variation measures accumulated variance and is the natural quantity in inequalities that control fluctuations.

This idea scales: if a process $M_n$ has increments with conditional mean zero, then sums of conditional variances provide the right control scale for deviations. Many concentration inequalities for martingales are built on this structure.

Martingale convergence: a standard route

A central theorem says that a nonnegative supermartingale converges almost surely. The proof uses two ingredients:

• supermartingales have decreasing expectations, providing $L^1$ control,
• maximal inequalities control oscillations, preventing infinite up-and-down movement.

In practice, one often proves convergence by identifying a quantity $X_n\ge 0$ that is a supermartingale. Then:

• $\mathbb{E}[X_n]$ decreases and is bounded below, so it converges,
• the process itself converges almost surely.

Uniform integrability refines this, giving $L^1$ convergence and interchange of limits and expectations.

A disciplined martingale workflow

Martingale arguments are repeatable:

• Choose the filtration that represents information.
• Construct a process with zero conditional drift, often via conditional expectation or by subtracting a compensator.
• Identify a stopping time that captures the event or time of interest.
• Work with the stopped process $X_{n\wedge \tau}$ \to avoid integrability issues.
• Verify optional stopping hypotheses before passing limits.
• Use maximal inequalities for convergence and tail control.

The main misuse is also repeatable: applying optional stopping directly \to $X_\tau$ without verifying that the limit from $n\wedge \tau$ is justified. When that check is done carefully, martingales become one of the most reliable reasoning tools in probability.

Doob decomposition: separating drift from noise

Submartingales often arise when a process has a systematic trend plus random fluctuation. Doob’s decomposition makes this precise in discrete time: if $X_n$ is an integrable submartingale, then it can be written as

where $M_n$ is a martingale and $A_n$ is an adapted increasing predictable process (meaning $A_{n+1}-A_n$ is $\mathcal{F}_n$-measurable and nonnegative). The increment $A_{n+1}-A_n$ captures the conditional drift:

This decomposition explains why martingales are the “noise-only” part of many models: once the predictable drift is subtracted, what remains has zero conditional mean.

In practice, this is a construction tool. If you can identify the predictable drift, you can build a martingale by subtracting it. That is exactly how compensated Poisson processes, centered counting processes, and many likelihood-ratio processes are formed.

Upcrossing inequality: why bounded supermartingales converge

A key reason martingales are useful is that they often converge. One mechanism behind convergence is the upcrossing inequality. Fix real numbers $a

Doob’s upcrossing inequality bounds the expected number of upcrossings in terms of the negative part of the process and its terminal value. A standard consequence is:

• If $X_n$ is a supermartingale bounded in $L^1$, then $X_n$ converges almost surely.

The conceptual point is that supermartingale structure prevents endless profitable oscillation: the process cannot keep crossing upward without paying in expectation. Upcrossing control is the technical bridge from an $L^1$ bound to almost sure convergence.

A concentration tool: Azuma–Hoeffding for bounded differences

Martingales also provide sharp tail bounds when increments are controlled. Suppose $M_n$ is a martingale with differences $D_k = M_k-M_{k-1}$ satisfying $|D_k|\le c_k$ almost surely. Then Azuma–Hoeffding gives

and similarly for $P(M_n-M_0 \le -t)$. This is one of the cleanest examples of martingales providing quantitative stability: bounded conditional increments imply subgaussian tails for deviations.

The inequality is used far beyond gambling models. It applies to randomized algorithms, sampling without replacement (with modifications), and functions of independent variables revealed sequentially, where the natural filtration is “reveal one coordinate at a time.”

Optional stopping revisited: a useful checklist

When applying optional stopping, it helps to explicitly verify the passage from $n\wedge \tau$ \to $\tau$. A reliable checklist is:

• Prove $\mathbb{E}[X_{n\wedge \tau}]=\mathbb{E}[X_0]$ for each $n$ using the stopped process.
• Show $X_{n\wedge \tau}\to X_\tau$ almost surely as $n\to\infty$.
• Justify exchanging limit and expectation using dominated convergence, uniform integrability, or an integrable bound on $\sup_{k\le \tau}|X_k|$.

The last step is where many mistakes occur. If $\tau$ can be large and the martingale can grow with $\tau$, then $X_\tau$ may fail to be integrable, and $\mathbb{E}[X_\tau]$ may not even exist. In such cases, the correct object is often $\mathbb{E}[X_{n\wedge \tau}]$ with an explicit limit procedure rather than a direct statement about $\mathbb{E}[X_\tau]$.

Why this toolkit scales

Martingales are effective because they respect the information structure of probability. Conditional expectation identities survive under approximation, stopping, and limit operations. Inequalities like Doob’s maximal inequality and Azuma–Hoeffding turn those identities into quantitative control. Convergence tools such as upcrossing and uniform integrability then convert boundedness into limits.

Once these pieces are in place, a wide range of results become variants of a single theme: identify the right martingale or supermartingale, stop it at the right time, and use a stability inequality to justify the limit or bound you need.

Probability theory becomes conceptually complete when it is formulated as measure theory with total mass one. The benefit is not abstraction for its own sake. The measure-theoretic framework tells you exactly which sets can be assigned probabilities without contradiction, which functions can be treated as random variables, and why the operations that dominate probability—limits, conditioning, and independence—work reliably. It also unifies discrete and continuous models under one definition of expectation: integration.

This article develops the core objects of measure-theoretic probability and shows, by example, how each definition is forced by the kinds of problems probability routinely asks.

Why “all subsets” cannot be events

Start with a set of outcomes $\Omega$. An “event” is meant to be a \subset $A\subseteq \Omega$ \to which a probability is assigned. If $\Omega$ is finite, defining a probability on all subsets is easy: specify weights on points and sum them. For uncountable spaces, trying to assign probabilities to all subsets breaks essential properties such as translation invariance and countable additivity. In the real line, for example, there exist nonmeasurable sets for which no consistent “length” can be defined while keeping the usual symmetries. Probability inherits the same obstruction.

The correct repair is to specify a collection $\mathcal{F}$ of subsets called the measurable sets. This is not a loss of generality in practice: $\mathcal{F}$ is chosen large enough to contain all events you can build from the model’s basic observables by countable operations.

σ-Algebras: stability under limits of events

A σ-algebra $\mathcal{F}$ on $\Omega$ is a collection of subsets satisfying:

• $\Omega\in\mathcal{F}$,
• $A\in\mathcal{F}\Rightarrow A^c\in\mathcal{F}$,
• $A_1,A_2,\dots\in\mathcal{F}\Rightarrow \bigcup_{n=1}^\infty A_n\in\mathcal{F}$.

From these axioms, $\mathcal{F}$ is also closed under countable intersections and set differences. The countability requirement is the point. Many probabilistic constructions use sequences: events like “$X_n$ eventually stays below 1,” “a random walk hits a state infinitely often,” or “$X_n\to X$” are built from countable unions and intersections. Without countable closure, those events might not even be defined.

A probability measure $P$ on $(\Omega,\mathcal{F})$ is a function $P:\mathcal{F}\to[0,1]$ such that:

• $P(\Omega)=1$,
• for pairwise disjoint events $A_1,A_2,\dots\in\mathcal{F}$,

The triple $(\Omega,\mathcal{F},P)$ is a probability space.

Two continuity properties are immediate consequences of countable additivity and are used constantly:

• If $A_n\uparrow A$ (increasing sequence), then $P(A_n)\uparrow P(A)$.
• If $A_n\downarrow A$ and $P(A_1)<\infty$ (automatic here), then $P(A_n)\downarrow P(A)$.

These are the measure-theoretic version of “probability respects limits of events.”

Random variables: measurable functions

A random variable is a measurable function $X:(\Omega,\mathcal{F})\to(\mathbb{R},\mathcal{B})$, where $\mathcal{B}$ is the Borel σ-algebra of $\mathbb{R}$. Measurability means:

Equivalently, it is enough to check sets of the form $(-\infty,t]$, because these generate $\mathcal{B}$:

This definition is not just formal. It is exactly what ensures that events defined by $X$—like $\{X\le t\}$, $\{X\in[a,b]\}$, or $\{X\in B\}$ for complicated $B$—are measurable and therefore have probabilities.

The distribution (law) of $X$ is the pushforward measure $P_X$ on $(\mathbb{R},\mathcal{B})$ defined by

This separates “how $X$ behaves” from “how it is represented on $\Omega$.” Many different probability spaces can support random variables with the same distribution.

Expectation as integration: the unification

In a finite probability space with outcomes $\omega_k$ of probability $p_k$, expectation is a weighted sum $\sum x(\omega_k)p_k$. In a continuous model with density $f$, expectation is $\int x f(x)\,dx$. Measure theory unifies these: expectation is the Lebesgue integral with respect \to $P$:

when $X$ is integrable.

To see why this is forced, start with indicator variables. If $X=\mathbf{1}_A$, then the expectation should be $P(A)$. The integral does exactly that:

For a simple random variable $S=\sum_{k=1}^m a_k\mathbf{1}_{A_k}$ with disjoint $A_k$, linearity gives

Now every nonnegative measurable $X$ can be approximated from below by an increasing sequence of simple functions $S_n\uparrow X$. The Lebesgue integral is defined by

This definition is built to make limits work. That is the essential reason expectation is an integral: probability arguments continually take monotone limits and dominated limits.

For general $X$, write $X=X^+-X^-$ where $X^+=\max(X,0)$ and $X^-=\max(-X,0)$. One defines $\mathbb{E}[X]$ when both $\mathbb{E}[X^+]$ and $\mathbb{E}[X^-]$ are finite.

Convergence theorems: the real pay-off

The following theorems are the workhorses of probability because they justify passing limits through expectations.

Monotone convergence theorem

If $0\le X_n\uparrow X$ almost surely, then

A standard use is truncation: $X_n=X\wedge n$. Many “infinite expectation” phenomena are proved by computing $\mathbb{E}[X\wedge n]$ and letting $n\to\infty$.

Dominated convergence theorem

If $X_n\to X$ almost surely and $|X_n|\le Y$ for an integrable $Y$, then

This is used whenever a parameter is sent \to a limit inside an integral-like expectation, and it clarifies why uniform integrability conditions are needed when domination fails.

Fatou’s lemma

For nonnegative $X_n$,

Fatou provides inequalities when full convergence hypotheses are unavailable.

These are not peripheral results. They are the mechanism by which probabilistic limits become rigorous.

Almost sure statements and modification on null sets

Measure theory distinguishes between holding everywhere and holding outside a probability-zero set. A property holds almost surely if it fails only on a null set. Many constructions in probability produce objects defined only up to null sets. For example, conditional expectations are defined as equivalence classes in $L^1$: changing $\mathbb{E}[X|\mathcal{G}]$ on a null set does not change its defining property.

This matters because “pathwise” regularity of stochastic processes can often be obtained only after modifying the process on a null set. Theorems about versions and modifications rely on this flexibility.

Conditional expectation: conditioning as a σ-algebra projection

Conditioning on an event $B$ is simple: $\mathbb{E}[X|B]$ is the average of $X$ restricted \to $B$. Conditioning on information is subtler. A σ-algebra $\mathcal{G}\subseteq\mathcal{F}$ represents the information available. The conditional expectation $\mathbb{E}[X|\mathcal{G}]$ is the $\mathcal{G}$-measurable random variable satisfying

This definition guarantees that $\mathbb{E}[X|\mathcal{G}]$ reproduces the averages of $X$ on every event in $\mathcal{G}$. When $\mathcal{G}$ is generated by a partition, it reduces to the familiar “average on each cell.” In general, it is the unique object that behaves like an information-based average.

Key properties follow quickly:

• Linearity: $\mathbb{E}[aX+bY|\mathcal{G}]=a\mathbb{E}[X|\mathcal{G}]+b\mathbb{E}[Y|\mathcal{G}]$.
• Tower property: if $\mathcal{H}\subseteq \mathcal{G}$, then $\mathbb{E}[\mathbb{E}[X|\mathcal{G}]|\mathcal{H}]=\mathbb{E}[X|\mathcal{H}]$.
• Taking out what is known: if $Y$ is $\mathcal{G}$-measurable and integrable, then $\mathbb{E}[XY|\mathcal{G}]=Y\mathbb{E}[X|\mathcal{G}]$ under appropriate integrability.

These properties are the algebra behind martingales and stochastic processes.

Independence via generated σ-algebras

Events $A$ and $B$ are independent if $P(A\cap B)=P(A)P(B)$. σ-algebras $\mathcal{A}$ and $\mathcal{B}$ are independent if every event in $\mathcal{A}$ is independent of every event in $\mathcal{B}$. Random variables $X$ and $Y$ are independent if the σ-algebras they generate, $\sigma(X)$ and $\sigma(Y)$, are independent.

This language clarifies two common confusions:

• Pairwise independence of variables is weaker than mutual independence, because independence must hold for all finite intersections of generated events.
• Independence is a statement about events (hence σ-algebras), not about numeric values directly.

The practical message

Measure-theoretic probability provides a stable foundation for the operations probability uses most:

• σ-algebras ensure limits of events remain events.
• Random variables are exactly the measurable functions that generate measurable events.
• Expectation is an integral designed to commute with limits under precise hypotheses.
• Conditional expectation is conditioning on information, not just on a single event.
• Independence is cleanly expressed via generated σ-algebras.

Once these pieces are in place, many advanced results become recombinations of a small set of robust principles: measurability, integrability, convergence theorems, and conditioning.