Introduction

In previous lessons we studied several notions of convergence:

• Pointwise convergence
• Uniform convergence
• Almost Everywhere convergence
• Convergence of integrals

At first, these notions may appear sufficient.

However, as probability theory and functional analysis developed, mathematicians discovered that another notion of convergence was needed.

This notion would be:

• stronger than some forms of convergence,
• weaker than others,
• extremely useful in probability theory.

This concept is called Convergence in Measure.

It is one of the central notions of modern measure theory.

In probability theory it becomes:

Convergence in Probability

which is one of the most important concepts in statistics and machine learning.

---

Motivation

Suppose:

$$f_n(x)\to f(x)$$

for most points.

Perhaps a few points behave badly.

Or perhaps a small region behaves badly.

Should we consider the sequence convergent?

Measure theory suggests:

Maybe we should not care about a tiny exceptional set.

Instead of examining individual points, we examine the size of the region where the approximation is poor.

This leads naturally to convergence in measure.

---

Definition

Let:

$$\left(X,\mathcal F,\mu\right)$$

be a measure space.

We say:

$$f_n \to f$$

in measure if for every:

$$\varepsilon>0$$

we have:

$$\mu\left({x:|f_n(x)-f(x)|>\varepsilon}\right)\to0$$

as:

$$n\to\infty$$

---

Understanding the Definition

The set:

$${x:|f_n(x)-f(x)|>\varepsilon}$$

contains points where the approximation error exceeds:

$$\varepsilon$$

Convergence in measure says:

For large n, the set of badly approximated points becomes very small.

The errors may not disappear everywhere.

The bad region merely shrinks toward measure zero.

---

Intuition

Pointwise convergence asks:

Does every point eventually behave?

Convergence in measure asks:

Does the set of badly behaved points become negligible?

This is a much more global viewpoint.

---

Example 1

Consider:

$$f_n(x)=\frac{x}{n}$$

on:

$$[0,1]$$

Let:

$$f(x)=0$$

Then:

$$|f_n(x)-0|=\frac{x}{n}$$

For any:

$$\varepsilon>0$$

eventually:

$$\frac{x}{n}<\varepsilon$$

for every:

$$x\in[0,1]$$

Therefore:

$$\mu\left({x:|f_n(x)|>\varepsilon}\right)=0$$

for sufficiently large:

$$n$$

Hence:

$$f_n\to0$$

in measure.

---

Example 2

Consider:

$$f_n(x)=\mathbf1_{[0,\frac1n]}(x)$$

on:

$$[0,1]$$

For every:

$$x>0$$

we have:

$$f_n(x)\to0$$

Pointwise convergence occurs.

Now examine convergence in measure.

For:

$$\varepsilon=\frac12$$

the bad set is:

$$[0,\frac1n]$$

whose measure equals:

$$\frac1n$$

Since:

$$\frac1n\to0$$

we obtain:

$$f_n\to0$$

in measure.

---

Example 3

A Sequence That Does Not Converge Pointwise

Define:

$$f_n=\mathbf1_{I_n}$$

where:

$$I_n$$

moves around:

$$[0,1]$$

so that every point is visited infinitely often.

Pointwise convergence may fail completely.

However, if:

$$\lambda(I_n)\to0$$

then:

$$f_n\to0$$

in measure.

This is an important observation.

Convergence in measure can occur even when pointwise convergence fails.

---

Why This Is Surprising

Most students initially believe:

Pointwise convergence is the weakest possible convergence.

This is false.

Convergence in measure can be even weaker.

A sequence may converge in measure while failing to converge pointwise.

---

Relationship with Almost Everywhere Convergence

Suppose:

$$f_n\to f$$

almost everywhere.

Does convergence in measure follow?

Not always.

However:

If:

$$\mu(X)<\infty$$

then:

$$f_n\to f \quad \text{a.e.}$$

implies:

$$f_n\to f \quad \text{in measure}$$

This theorem is extremely important.

---

Why Finite Measure Matters

If:

$$X$$

has finite measure, then exceptional sets cannot hide in infinitely large regions.

This allows almost everywhere convergence to control convergence in measure.

Without finite measure, counterexamples exist.

---

Relationship with Uniform Convergence

Uniform convergence implies convergence in measure.

Indeed:

If:

$$\sup_{x\in X}|f_n(x)-f(x)|\to0$$

then eventually:

$$|f_n(x)-f(x)|<\varepsilon$$

for every:

$$x$$

Therefore the bad set becomes empty.

Thus:

$$\text{Uniform Convergence} \implies \text{Convergence in Measure}$$

---

Relationship with L¹ Convergence

Suppose:

$$\int |f_n-f|,d\mu\to0$$

Then:

$$f_n\to f$$

in measure.

This follows from Markov’s inequality.

Thus:

$$L^1 \text{ convergence} \implies \text{ Convergence in Measure}$$

---

The Hierarchy So Far

Generally:

$$\text{Uniform Convergence} \implies L^1 \text{ Convergence} \implies \text{Convergence in Measure}$$

and:

$$\text{Almost Everywhere Convergence} \implies \text{Convergence in Measure}$$

for finite measure spaces.

The reverse implications usually fail.

---

Why Analysts Love Convergence in Measure

Convergence in measure behaves well under:

• addition
• multiplication
• approximation
• probability limits

Many difficult sequences fail to converge pointwise but converge in measure.

Thus it provides a more flexible framework.

---

Convergence in Measure and Probability

Suppose:

$$\mu=P$$

is a probability measure.

Then:

$$\mu(X)=1$$

The definition becomes:

$$P(|X_n-X|>\varepsilon)\to0$$

This is exactly the definition of:

Convergence in Probability

Thus convergence in probability is simply convergence in measure on a probability space.

This observation is one of the most important bridges between measure theory and statistics.

---

Example from Statistics

Suppose:

$$\bar X_n$$

is the sample mean.

The Weak Law of Large Numbers states:

$$P(|\bar X_n-\mu|>\varepsilon)\to0$$

This means:

$$\bar X_n\to\mu$$

in probability.

Measure theorists would say:

$$\bar X_n\to\mu$$

in measure.

The concepts are identical.

---

Why Machine Learning Uses This

Many machine learning results involve showing:

$$\hat f_n \to f$$

in probability.

Examples include:

• Consistency of estimators
• Empirical risk minimization
• Bayesian posterior consistency

Underneath these results lies convergence in measure.

---

A Fundamental Theorem

One of the most beautiful results states:

If:

$$f_n\to f$$

in measure,

then there exists a subsequence:

$$f_{n_k}$$

such that:

$$f_{n_k}\to f$$

almost everywhere.

This theorem is remarkable.

Even weak convergence in measure contains hidden pointwise convergence along suitable subsequences.

---

Why This Matters

Many proofs proceed as follows:

1. Prove convergence in measure.
2. Extract an almost everywhere convergent subsequence.
3. Apply stronger theorems.

This strategy appears throughout modern analysis.

---

Connection to Functional Analysis

When we later study:

$$L^p$$

spaces,

convergence in measure becomes an important intermediate concept.

Many compactness theorems and approximation results use it.

---

Connection to Alain Connes

Convergence in measure introduces a recurring theme:

Exact pointwise behavior is often less important than behavior on most of the space.

This philosophy becomes increasingly important in modern analysis.

Later:

• functions become equivalence classes,
• measures become states,
• pointwise geometry becomes operator-theoretic geometry.

Connes’ work repeatedly emphasizes global structures over individual points.

Convergence in measure is one of the first places where this philosophical shift becomes visible.

---

Key Concepts Learned

By the end of this lesson you should understand:

• Convergence in measure means:

$$\mu\left({x:|f_n-f|>\varepsilon}\right)\to0$$

for every:

$$\varepsilon>0$$

• The set of badly approximated points shrinks to measure zero.
• Convergence in measure may occur without pointwise convergence.
• Uniform convergence implies convergence in measure.
• $$L^1$$ convergence implies convergence in measure.
• On finite measure spaces, almost everywhere convergence implies convergence in measure.
• Convergence in probability is simply convergence in measure on a probability space.
• Every sequence converging in measure contains an almost everywhere convergent subsequence.

---

Looking Ahead

In the next lesson:

Measure Theory Lesson 26: Egoroff’s Theorem

we will prove one of the most elegant results in measure theory. Egoroff’s Theorem shows that on finite measure spaces, almost everywhere convergence is “almost” uniform convergence. This theorem creates a surprising bridge between pointwise behavior and uniform behavior and is one of the hidden gems of classical analysis.