Analysis II — Chapter-by-Chapter Outline

Author: Terence Tao First published: 2006 (Hindustan Book Agency) Edition covered: Fourth Edition (Springer, 2022; ISBN 978-981-19-7284-3). Earlier editions numbered chapters starting at 12 (continuing from Analysis I); the Third and Fourth editions renumber them 1–8 as a self-contained volume. The Fourth Edition incorporates corrections accumulated since the Third Edition and adds further exercises; the chapter and section structure is otherwise identical to the Third Edition.

---

Central thesis

Real analysis cannot stop at the real line. Once the foundations of single-variable analysis are secure, the subject demands a general framework: abstract spaces where distance makes sense, functions between them, and an integration theory powerful enough to handle limits that Riemann's construction cannot. Analysis II builds that framework systematically, carrying the reader from the abstract geometry of metric spaces through multivariable calculus to the Lebesgue integral.

The book's central intellectual wager is that generality and rigor go together: by working in metric spaces from the start, the classical theorems of real analysis (uniform continuity, uniform convergence, the inverse function theorem) are not just proved but understood, because the right level of abstraction reveals what each proof actually depends on. The Lebesgue integral appears at the end not as an afterthought but as the inevitable answer to the question: what integration theory is closed under limits?

What does it mean for a sequence of functions to converge, and what properties does the limit inherit? How should we integrate functions that are themselves limits of other functions?

---

Chapter 1 — Metric Spaces

Central question

How can we talk about convergence, limits, and continuity in spaces other than the real line, using only the notion of distance?

Main argument

Defining the metric space. Tao opens by abstracting the one property of the real line that drives all of Analysis I: the distance function. A metric space is a set X together with a function d : X × X → ℝ satisfying positivity, symmetry, and the triangle inequality. The definition is parsimonious — it captures just enough structure to define convergence. Concrete examples ground the abstraction immediately: the real line ℝ with |x − y|, Euclidean space ℝⁿ with the Euclidean metric, the discrete metric (d(x,y) = 1 if x ≠ y), and the space of continuous functions with the supremum metric. A sequence (xₙ) in X converges to a limit x if d(xₙ, x) → 0.

Point-set topology of metric spaces. Open balls B(x₀, r) = {x : d(x, x₀) < r} generate the topological vocabulary. Tao defines open and closed sets in terms of balls, interior points, boundary points, and limit points — and proves their basic properties: arbitrary unions of open sets are open; finite intersections of open sets are open. The concept of adherent point (a point every neighbourhood of which meets the set) is treated carefully to avoid the informal errors common in first encounters. The real line's familiar topology is now a special case of a general theory.

Relative topology. A subset Y of a metric space (X, d) inherits a metric by restriction. Open and closed sets in Y are the traces of open and closed sets in X: a set is open in Y iff it equals the intersection of Y with some open set in X. This seemingly technical point becomes essential when studying continuous functions on domains — a set can be simultaneously open in a subspace and not open in the ambient space.

Cauchy sequences and completeness. A sequence in a metric space is Cauchy if its terms become arbitrarily close to one another — not necessarily to any particular limit. Every convergent sequence is Cauchy, but the converse requires an additional hypothesis: completeness. A complete metric space is one in which every Cauchy sequence converges within the space. ℝ is complete; ℚ is not. Tao also introduces the completion of a metric space: every metric space embeds isometrically and densely in a unique complete metric space, constructed by a quotient of the space of Cauchy sequences.

Compact metric spaces. A metric space is compact if every sequence in it has a convergent subsequence (sequential compactness). Equivalently, every open cover has a finite subcover. Tao proves the Heine-Borel theorem in ℝⁿ: a subset is compact if and only if it is closed and bounded. Compact spaces enjoy two properties proved in Chapter 2: continuous functions on them are uniformly continuous, and real-valued continuous functions on them attain their maximum and minimum.

Key ideas

• The metric axioms (positivity, symmetry, triangle inequality) are the minimal data required to define convergence and continuity in an abstract space.
• Open sets, closed sets, and closure are defined purely in terms of the metric; no additional algebraic structure is assumed.
• Cauchy sequences capture the intuition of convergence without presupposing a limit point in the space, making completeness a precise and checkable property.
• A complete metric space has "no gaps": sequences that look like they should converge actually do converge within the space.
• Sequential compactness (every sequence has a convergent subsequence) is equivalent to the Heine-Borel covering characterisation in metric spaces.
• In ℝⁿ, compact is equivalent to closed and bounded (Heine-Borel); this fails in infinite-dimensional spaces.
• The relative topology shows that "open" and "closed" are always relative to an ambient space.

Key takeaway

Metric spaces provide the minimal abstract framework — just a distance function obeying three axioms — in which convergence, completeness, and compactness can be defined rigorously and generalized far beyond the real line.

---

Chapter 2 — Continuous Functions on Metric Spaces

Central question

What does continuity mean in metric spaces, and what structural properties does it preserve?

Main argument

Defining continuity. A function f : X → Y between metric spaces is continuous at a point x₀ if for every ε > 0 there exists δ > 0 such that dX(x, x₀) < δ implies dY(f(x), f(x₀)) < ε. Tao establishes three equivalent formulations: the epsilon-delta definition, sequential continuity (f(xₙ) → f(x₀) whenever xₙ → x₀), and the topological definition (the preimage of every open set in Y is open in X). The equivalence of these three characterisations is a fundamental theorem, showing that each definition captures the same concept from a different angle.

Continuity and product spaces. The product X × Y of two metric spaces carries a natural metric. A function f : Z → X × Y is continuous if and only if each coordinate projection is continuous. This reduces questions about vector-valued functions to their component functions.

Continuity and compactness. Two fundamental theorems connect continuity with compactness from Chapter 1. First: the continuous image of a compact set is compact. Second: a continuous function on a compact metric space is uniformly continuous — the δ can be chosen independently of the base point x₀. The extreme value theorem follows at once: a continuous real-valued function on a compact metric space attains its maximum and minimum. These results lift classical theorems from closed bounded intervals in ℝ to the full generality of compact metric spaces.

Continuity and connectedness. A metric space is connected if it cannot be written as the union of two disjoint non-empty open sets. Tao proves that the continuous image of a connected set is connected. On the real line, the connected sets are precisely the intervals, and the intermediate value theorem is an immediate corollary: a continuous function that takes both positive and negative values on a connected set must vanish somewhere on it. The proof is elegant — connectedness is preserved, and a disconnecting of the image would lift to a disconnecting of the domain.

Topological spaces (optional). An optional section introduces the more general notion of a topological space: a set equipped with a collection of "open sets" satisfying three axioms (whole space and empty set are open; arbitrary unions of open sets are open; finite intersections of open sets are open). Every metric space is a topological space. This section is optional — the rest of the book does not depend on it — but it plants the seed of topology as a discipline.

Key ideas

• The three equivalent definitions of continuity (epsilon-delta, sequential, topological) illuminate the concept from complementary perspectives; the topological formulation is most powerful for proving preservation theorems.
• Uniform continuity requires the same δ everywhere; pointwise continuity allows δ to depend on the base point. Compact domains force uniform continuity.
• Compactness is preserved under continuous maps (the continuous image of compact is compact), making compact spaces the natural setting for optimization.
• Connectedness is preserved under continuous maps; on the real line, connectedness of an interval gives the intermediate value theorem.
• The topological definition of continuity shows that continuity depends only on the open-set structure, not the particular metric.

Key takeaway

Continuous functions between metric spaces preserve the two key topological properties — compactness and connectedness — which immediately yield the extreme value theorem and the intermediate value theorem as corollaries.

---

Chapter 3 — Uniform Convergence

Central question

When does a sequence of functions converge in a sense that preserves continuity and allows interchange of limits with integrals and derivatives?

Main argument

Pointwise versus uniform convergence. Given a sequence fₙ : X → ℝ, pointwise convergence means fₙ(x) → f(x) for each fixed x independently. But pointwise convergence is too weak: the sequence fₙ(x) = xⁿ on [0, 1] converges pointwise to a discontinuous function (0 for x < 1, 1 at x = 1), even though each fₙ is continuous. Tao introduces uniform convergence: fₙ → f uniformly if for every ε > 0 there exists N (independent of x) such that n ≥ N implies |fₙ(x) − f(x)| < ε for all x. The key is that N may not depend on x.

The metric of uniform convergence. On the space of bounded functions X → ℝ, the supremum metric d∞(f, g) = sup{x ∈ X} |f(x) − g(x)| measures the largest pointwise discrepancy. Uniform convergence is precisely convergence in this metric: fₙ → f uniformly iff d_∞(fₙ, f) → 0. This reframes uniform convergence as convergence in a metric space, connecting this chapter to the theory of Chapter 1.

Uniform convergence and continuity. The key theorem: the uniform limit of continuous functions is continuous. Tao's proof is the archetype of an "3ε" argument: to bound |f(x) − f(x₀)|, add and subtract fₙ(x) and fₙ(x₀) for a suitably chosen n, splitting the gap into three pieces each bounded by ε/3 — one controlled by uniform convergence, two by continuity of fₙ. This is the fundamental reason for insisting on uniform over pointwise convergence.

The Weierstrass M-test. For a series of functions Σfₙ, the Weierstrass M-test gives a practical sufficient condition for uniform convergence: if |fₙ(x)| ≤ Mₙ for all x and ΣMₙ < ∞, then the series Σfₙ converges uniformly (and absolutely). The Mₙ are summable bounds independent of x. This test is the main tool for verifying that a power series or Fourier series converges to a continuous function.

Uniform convergence and integration. If fₙ → f uniformly on [a, b] and each fₙ is Riemann integrable, then f is Riemann integrable and ∫ₐᵇ fₙ → ∫ₐᵇ f. This limit-integral exchange fails for pointwise convergence. Tao's proof bounds the difference |∫fₙ − ∫f| ≤ ∫|fₙ − f| ≤ (b−a)·d_∞(fₙ, f), which tends to 0 by uniform convergence.

Uniform convergence and derivatives. The situation for derivatives is more delicate: uniform convergence of fₙ does not in general imply fₙ' → f'. However, if fₙ → f pointwise and fₙ' → g uniformly, then f is differentiable and f' = g. The key hypothesis is on the derivatives, not the functions themselves.

Uniform approximation by polynomials. Tao proves the Weierstrass approximation theorem: every continuous function on a closed interval [a, b] can be uniformly approximated by polynomials. The proof uses Bernstein polynomials: for f : [0, 1] → ℝ, define Bₙf(x) = Σ_{k=0}^n f(k/n) · C(n,k) · xᵏ · (1−x)^{n−k}. This is the expected value of f at a Binomial(n, x) random variable; the law of large numbers argument shows Bₙf → f uniformly. The theorem says polynomials are dense in C([a, b]) under the supremum norm — a prototype for approximation theory.

Key ideas

• Pointwise convergence preserves almost no analytic property; uniform convergence preserves continuity and allows limit-integral and limit-derivative exchange.
• The supremum metric transforms uniform convergence into convergence in a metric space, allowing the completeness results of Chapter 1 to apply.
• The Weierstrass M-test is the standard tool for uniform convergence of series, used repeatedly in Chapters 4 and 5.
• Limit-integral exchange holds for uniformly convergent sequences but can fail spectacularly for pointwise convergent ones.
• Derivative-limit exchange requires uniform convergence of the derivatives, not just of the functions.
• The Weierstrass approximation theorem shows that continuous functions on compact intervals can always be approximated by polynomials; this is one of the first "density in a function space" results.
• Bernstein polynomials provide a probabilistically motivated constructive proof of the Weierstrass theorem.

Key takeaway

Uniform convergence is the correct notion of function-sequence convergence because it — unlike pointwise convergence — preserves continuity, allows interchange of limits with integrals, and underlies the Weierstrass approximation theorem.

---

Chapter 4 — Power Series

Central question

Which functions can be expressed as power series, and what analytic and algebraic properties follow from such a representation?

Main argument

Formal power series and radius of convergence. Tao begins with the algebraic object: a formal power series Σ{n=0}^∞ aₙ(x − x₀)ⁿ, initially just a formal expression. The radius of convergence is determined by the Cauchy-Hadamard formula: 1/R = limsup{n→∞} |aₙ|^{1/n}. The series converges absolutely for |x − x₀| < R and diverges for |x − x₀| > R; behaviour at the endpoints |x − x₀| = R requires separate analysis in each case. Within the radius of convergence the Weierstrass M-test (Chapter 3) guarantees uniform convergence on any closed sub-interval.

Real analytic functions. A function f is real analytic at a point x₀ if it equals a convergent power series in some open interval around x₀. Real analyticity is strictly stronger than infinite differentiability. Tao proves that a real analytic function can be differentiated and integrated term by term within its radius of convergence: the derivative of Σaₙxⁿ is Σnaₙxⁿ⁻¹ with the same radius of convergence. The Taylor coefficients aₙ = f^(n)(x₀)/n! are uniquely determined by f.

Abel's theorem. If the power series Σaₙxⁿ converges at the right endpoint x = R, then the sum function is continuous at x = R from the left: lim_{x→R⁻} Σaₙxⁿ = ΣaₙRⁿ. Abel's theorem is delicate because uniform convergence may fail at the endpoint; the proof uses summation by parts (Abel summation) to handle the boundary carefully. This theorem allows one to evaluate series such as log(2) = 1 − 1/2 + 1/3 − ... by approaching the boundary of convergence.

Multiplication of power series. The Cauchy product of Σaₙxⁿ and Σbₙxⁿ is Σcₙxⁿ where cₙ = Σ_{k=0}^n aₖbₙ₋ₖ. Tao proves (Mertens' theorem) that if both series converge and at least one converges absolutely, then their Cauchy product converges to the product of their sums. This justifies the formal algebraic manipulation of power series.

The exponential and logarithm functions. The exponential function is defined by its power series: exp(x) = Σ_{n=0}^∞ xⁿ/n!. Properties are derived from this definition: the functional equation exp(x + y) = exp(x)exp(y) follows from the Cauchy product; exp'(x) = exp(x) follows by term-by-term differentiation; positivity and strict monotonicity follow from the series representation. The natural logarithm is defined as the inverse function of exp, and its power series log(1 + x) = x − x²/2 + x³/3 − ... converges for x ∈ (−1, 1] by Abel's theorem.

A digression on complex numbers. An optional section introduces ℂ as ordered pairs (a, b) with the standard arithmetic. This minimal digression is sufficient to define complex exponentials, state Euler's formula e^{iθ} = cos θ + i sin θ, and derive the definitions of sine and cosine.

Trigonometric functions. The real power series sin(x) = Σ{n=0}^∞ (−1)ⁿ x^{2n+1}/(2n+1)! and cos(x) = Σ{n=0}^∞ (−1)ⁿ x^{2n}/(2n)! define sine and cosine with infinite radii of convergence. All standard identities (Pythagorean identity, addition formulas, derivative formulas sin' = cos, cos' = −sin, periodicity) are derived from these power series definitions without appeal to geometry. The number π is defined analytically as twice the smallest positive zero of cosine.

Key ideas

• The Cauchy-Hadamard formula determines the radius of convergence from the coefficients alone, marking a sharp convergence/divergence boundary.
• Within the radius of convergence, power series converge uniformly on compact sub-intervals and can be differentiated and integrated term by term.
• Abel's theorem extends evaluation to the boundary of convergence when the series converges there, using a summation-by-parts argument.
• The Cauchy product multiplies power series algebraically when both series converge absolutely.
• The exponential function is most cleanly defined by its power series; the functional equation exp(x+y) = exp(x)exp(y) then follows from the Cauchy product.
• Defining sin and cos by power series puts them on rigorous footing and avoids circular appeals to geometry.
• π is defined as twice the smallest positive zero of cosine, an analytical definition consistent with all geometric usages.

Key takeaway

Power series are the rigorous vehicle for defining the elementary transcendental functions — exponential, logarithm, sine, cosine — because all their fundamental properties follow from power-series manipulations, making geometry-free proofs possible.

---

Chapter 5 — Fourier Series

Central question

Can an arbitrary periodic function be represented as an infinite sum of complex exponentials, and in what sense does such a representation converge?

Main argument

Periodic functions. The chapter studies functions f : ℝ → ℂ that are periodic with period 2π, equivalently functions on the circle ℝ/2πℤ. The complex exponentials e^{inx} for n ∈ ℤ serve as basis functions; working with complex exponentials streamlines the algebra considerably compared to using sines and cosines separately, and the two descriptions are equivalent via Euler's formula from Chapter 4.

Inner products on periodic functions. Tao defines the L² inner product ⟨f, g⟩ = (1/2π) ∫_{−π}^π f(x) g̅(x) dx. This turns the space of square-integrable periodic functions into an inner product space, enabling the geometric language of orthogonality and projection. The central fact: the exponentials e^{inx} form an orthonormal system under this inner product — ⟨e^{imx}, e^{inx}⟩ = 1 if m = n, and 0 otherwise. Orthonormality is verified directly by integrating e^{i(m−n)x} over a full period.

Trigonometric polynomials. A trigonometric polynomial is a finite linear combination Σ{n=−N}^N cₙ e^{inx}. The best approximation of f by a trigonometric polynomial of degree ≤ N in the L² sense is the partial Fourier sum Sₙf(x) = Σ{n=−N}^N ĉₙ e^{inx}, where the Fourier coefficients are ĉₙ = ⟨f, e^{inx}⟩ = (1/2π) ∫ f(x) e^{−inx} dx. The projection theorem of Hilbert space theory guarantees that Sₙf minimises the L² distance to f among all degree-≤N trigonometric polynomials.

Periodic convolutions. The convolution of two periodic functions is (f * g)(x) = (1/2π) ∫{−π}^π f(y) g(x − y) dy. The key algebraic fact: the Fourier coefficients of f * g are the pointwise products of the coefficients of f and g separately. This convolution theorem means that Fourier series converts convolution (a complicated integral operation) into pointwise multiplication (a simple algebraic operation). The Dirichlet kernel Dₙ(x) = Σ{k=−N}^N e^{ikx} satisfies Sₙf = f * Dₙ; its oscillatory behaviour near the boundary of the interval explains the Gibbs phenomenon.

The Fourier and Plancherel theorems. The chapter culminates in two completeness results. First, trigonometric polynomials are dense in C(ℝ/2πℤ) under the uniform norm — a consequence of the Weierstrass approximation theorem for periodic functions, proved via the Fejér kernel (Cesàro averages of the Dirichlet sums). Second, the Plancherel theorem: ‖f‖²{L²} = Σ{n=−∞}^∞ |ĉₙ|². This says the sum of the squared Fourier coefficients equals the squared L²-norm of f — the Fourier transform is an isometry from L² to ℓ². As a consequence, Fourier series converge to f in L², meaning ‖Sₙf − f‖_{L²} → 0.

Key ideas

• Complex exponentials e^{inx} form an orthonormal system in L²; Fourier coefficients are the inner products of f with these basis elements.
• The partial Fourier sum Sₙf is the best L² approximation to f by trigonometric polynomials of degree ≤ N.
• The convolution theorem converts convolution into pointwise multiplication of Fourier coefficients, revealing the spectral structure of convolution operators.
• L² convergence of Fourier series holds for all square-integrable functions (Plancherel theorem), but pointwise convergence everywhere requires additional smoothness.
• The Plancherel theorem ‖f‖² = Σ|ĉₙ|² is the isometric completeness statement, making the Fourier transform an isometry L² → ℓ².
• Fourier series are a prototype for the spectral theory of operators on Hilbert spaces — the generalisation that Chapter 5 prepares.
• The Fejér kernel (Cesàro sums) converges uniformly for continuous f even where the Dirichlet kernel may not.

Key takeaway

Fourier series decompose periodic functions into orthogonal exponential components; the Plancherel theorem shows this decomposition is isometric in L², making Fourier series the prototype for spectral decompositions in Hilbert spaces.

---

Chapter 6 — Several Variable Differential Calculus

Central question

What is the correct notion of the derivative for a function f : ℝⁿ → ℝᵐ, and what does it imply about the local structure of such functions?

Main argument

Linear transformations as the derivative. The derivative of f : ℝⁿ → ℝᵐ at a point x₀ is not a number but a linear transformation T : ℝⁿ → ℝᵐ satisfying lim_{h→0} ‖f(x₀+h) − f(x₀) − Th‖/‖h‖ = 0. T is the best linear approximation to f near x₀. Tao first reviews the necessary linear algebra: matrices, linear maps, norms of linear maps (the operator norm ‖T‖ = sup{‖Tv‖ : ‖v‖ ≤ 1}). The total derivative (Fréchet derivative) Df(x₀) = T, when it exists, is unique. In one dimension this reduces to the usual derivative f'(x₀), with T being multiplication by the scalar f'(x₀).

Partial and directional derivatives. The partial derivative ∂f/∂xᵢ(x₀) is the rate of change along the i-th coordinate direction. A directional derivative Dᵥf(x₀) is the rate along any direction v. If f is differentiable at x₀ then Dᵥf(x₀) = Df(x₀) · v (the total derivative applied to v). Tao establishes that differentiability implies existence of all directional derivatives, but not conversely: a function can have all partial derivatives at a point without being differentiable there (with an explicit counterexample). The correct sufficient condition is that the partial derivatives exist in a neighborhood and are continuous at the point.

The several-variable chain rule. If f : ℝⁿ → ℝᵐ is differentiable at x₀ and g : ℝᵐ → ℝᵖ is differentiable at f(x₀), then g ∘ f is differentiable at x₀ and D(g ∘ f)(x₀) = Dg(f(x₀)) ∘ Df(x₀). In matrix terms: the Jacobian of the composition is the product of the Jacobians. This is the central rule for computing derivatives of composite functions and underlies all subsequent results.

Double derivatives and Clairaut's theorem. The second-order mixed partial derivatives ∂²f/∂xᵢ∂xⱼ and ∂²f/∂xⱼ∂xᵢ need not be equal for an arbitrary f. Clairaut's theorem (Schwarz's theorem) guarantees their equality when both mixed partials are continuous at the point. Tao provides both a careful proof and an explicit counterexample demonstrating that the continuity hypothesis cannot be dropped.

The contraction mapping theorem. A map f : X → X on a metric space is a contraction if there exists L < 1 with d(f(x), f(y)) ≤ L · d(x, y) for all x, y. The Banach fixed-point theorem: every contraction on a complete metric space has a unique fixed point, approached by iteration from any starting point. This theorem belongs to pure metric space theory (Chapter 1) but is proved here because it is the engine driving the next two theorems.

The inverse function theorem. If f : ℝⁿ → ℝⁿ is continuously differentiable and Df(x₀) is invertible, then f is a local diffeomorphism near x₀: there exist open sets U ∋ x₀ and V ∋ f(x₀) such that f|_U : U → V is a bijection with a continuously differentiable inverse. The proof: define a contraction on ℝⁿ whose unique fixed point is the desired preimage. The invertibility of Df(x₀) ensures the contraction condition holds near x₀. This theorem converts a linear-algebraic hypothesis (the Jacobian is invertible) into a geometric conclusion (the function is locally invertible).

The implicit function theorem. Given F : ℝⁿ⁺ᵐ → ℝᵐ with F(a, b) = 0 and the partial Jacobian D_y F(a, b) invertible (where y ∈ ℝᵐ denotes the last m variables), there exists a neighborhood U of a and a unique continuously differentiable function g : U → ℝᵐ such that F(x, g(x)) = 0 and g(a) = b. The proof reduces to the inverse function theorem applied to an augmented map. The implicit function theorem converts an implicit constraint into an explicit parametrisation, underpinning the local geometry of level sets and the theory of Lagrange multipliers.

Key ideas

• The correct derivative for f : ℝⁿ → ℝᵐ is a linear map (the total derivative), represented by the m×n Jacobian matrix; this generalises the one-variable derivative from a number to a linear transformation.
• Existence of all partial derivatives does not imply differentiability; continuous partial derivatives in a neighborhood suffice (but are not necessary).
• The chain rule in several variables: the Jacobian of a composition is the product of the Jacobians — matrix multiplication reflects composition of linear approximations.
• Mixed partial derivatives commute when they are continuous (Clairaut's theorem), but pathological counterexamples exist without the continuity hypothesis.
• The contraction mapping theorem (Banach fixed-point theorem) is the analytic engine underlying both the inverse and implicit function theorems.
• The inverse function theorem: invertibility of the linear approximation (nonzero Jacobian determinant) implies local invertibility of the nonlinear function.
• The implicit function theorem converts F(x,y) = 0 from an implicit constraint into an explicit function y = g(x) whenever the partial Jacobian in y is invertible.

Key takeaway

The right derivative for maps f : ℝⁿ → ℝᵐ is the total derivative — a linear map — and once this is established, the contraction mapping theorem gives a unified proof of the inverse and implicit function theorems, the two deepest structural results in multivariable calculus.

---

Chapter 7 — Lebesgue Measure

Central question

Can we assign a "size" to every subset of ℝⁿ that agrees with length/area/volume for rectangles, is countably additive over disjoint pieces, and is defined for as large a class of sets as possible?

Main argument

The goal: Lebesgue measure. The Riemann integral is built on the elementary measure of boxes (length of intervals, area of rectangles). But the Riemann theory cannot handle limits of functions without strong uniform convergence assumptions. Tao motivates Lebesgue measure by posing a requirement: a measure on subsets of ℝⁿ that (1) agrees with Jordan measure on elementary sets, (2) is countably additive (not just finitely additive), and (3) is translation-invariant and defined on as large a class as possible.

First attempt: outer measure. Lebesgue outer measure m(E) is the infimum of the total volume of countably many open boxes covering E: m(E) = inf{Σ vol(Bₖ) : E ⊆ ∪ Bₖ, each Bₖ an open box}. The key modification from Jordan outer measure is that the covering collection can be countably infinite. Tao proves outer measure is monotone (E ⊆ F implies m(E) ≤ m(F)), countably subadditive (m(∪Eₙ) ≤ Σm(Eₙ)), and agrees with volume on boxes.

Outer measure is not additive. The critical obstacle: outer measure is not countably additive on all pairs of disjoint sets. Tao constructs the Vitali set — using the axiom of choice, one selects one representative from each equivalence class of ℝ modulo ℚ in [0,1]. The resulting set has the property that the outer measure of the union of its countably many rational translates (which partition [0,1]) is neither 0 nor 1, a contradiction with both additivity and bounded outer measure. This proves that no translation-invariant measure can be defined on all subsets of ℝ.

Measurable sets. The resolution: restrict attention to sets for which outer measure behaves well. Tao defines E to be Lebesgue measurable if for every ε > 0 there exists an open set U ⊇ E with m(U \ E) < ε (Littlewood's first principle — measurable sets are "almost open"). On measurable sets, the Lebesgue measure m(E) equals the outer measure m(E). The fundamental structural theorem: the collection of measurable sets is a σ-algebra — closed under countable unions, intersections, and complements. Moreover, m is countably additive on this σ-algebra: for disjoint measurable Eₙ, m(∪Eₙ) = Σm(Eₙ). All open sets, closed sets, compact sets, countable sets, and null sets (sets of outer measure 0) are measurable.

Measurable functions. A function f : ℝⁿ → ℝ is Lebesgue measurable if the preimage f⁻¹((a, ∞)) is a Lebesgue measurable set for every a ∈ ℝ. Continuous functions are measurable. Pointwise limits of measurable functions are measurable. The class of measurable functions is closed under arithmetic operations, absolute values, and limits — making it the correct domain for the Lebesgue integral.

Key ideas

• Outer measure extends Jordan measure by using countable rather than finite covers; this is the key modification enabling the theory to handle countably infinite constructions.
• Countable subadditivity holds for all sets; strict countable additivity fails on arbitrary disjoint sets (Vitali set).
• The Vitali construction (using the axiom of choice) proves non-measurable sets exist; this is not a deficiency of the theory but a fundamental feature of set theory.
• Measurable sets form a σ-algebra on which outer measure is countably additive — exactly the structure needed for a well-behaved integration theory.
• All sets that arise naturally in analysis (open, closed, compact, null) are measurable; the non-measurable sets cannot be explicitly constructed without the axiom of choice.
• Measurable functions are those whose level sets are measurable; the class is closed under all standard limit operations, making it stable under the operations of analysis.

Key takeaway

Lebesgue measure is obtained by restricting outer measure to the σ-algebra of measurable sets — large enough to contain every set arising in analysis, small enough to exclude pathological non-measurable sets — on which countable additivity holds.

---

Chapter 8 — Lebesgue Integration

Central question

How should we integrate Lebesgue measurable functions, and why is this integral superior to the Riemann integral under limiting operations?

Main argument

Simple functions. The Lebesgue integral is built in stages. A simple function is a measurable function taking finitely many values: s = Σ{k=1}^K aₖ · 1{Eₖ}, where 1_{Eₖ} is the indicator function of the measurable set Eₖ and Σm(Eₖ) < ∞. The integral of a simple function is defined as ∫s = Σ aₖ · m(Eₖ). Tao proves this definition is well-posed (independent of the representation chosen), and that the integral of simple functions is linear (∫(αs + βt) = α∫s + β∫t) and monotone (s ≤ t implies ∫s ≤ ∫t). Simple functions play the role that step functions play in the Riemann theory, but they approximate from below using measure-theoretic sets rather than intervals.

Integration of non-negative measurable functions. For a non-negative measurable function f : ℝⁿ → [0, ∞], the Lebesgue integral is defined as the supremum of integrals of non-negative simple functions dominated by f: ∫f = sup{∫s : 0 ≤ s ≤ f, s simple}. The fundamental convergence theorem for this stage is the monotone convergence theorem: if 0 ≤ f₁ ≤ f₂ ≤ ... is an increasing sequence of non-negative measurable functions converging pointwise to f, then ∫fₙ → ∫f. No uniform convergence is required. This is a first demonstration of the integral's stability under limits.

Integration of absolutely integrable functions. A measurable function f is absolutely integrable (or Lebesgue integrable, denoted f ∈ L¹) if ∫|f| < ∞. Write f = f⁺ − f⁻ where f⁺ = max(f, 0) and f⁻ = max(−f, 0) are the positive and negative parts (both non-negative); then ∫f = ∫f⁺ − ∫f⁻. The integral is linear and satisfies the triangle inequality |∫f| ≤ ∫|f|. The critical theorem is the dominated convergence theorem: if fₙ → f pointwise and |fₙ| ≤ g with g ∈ L¹, then ∫fₙ → ∫f. The dominating function g controls the "size" of the functions, making uniform convergence unnecessary. This theorem is the workhorse of modern analysis — it is used repeatedly in functional analysis, probability, and partial differential equations.

Comparison with the Riemann integral. Every Riemann integrable function on [a, b] is Lebesgue integrable and the two integrals agree. The Lebesgue integral is strictly more powerful: the indicator function of the rationals 1_ℚ is not Riemann integrable but is Lebesgue integrable (with integral 0, since ℚ has measure 0). Tao establishes Lebesgue's characterisation of Riemann integrability: a bounded function on [a, b] is Riemann integrable if and only if its set of discontinuities has Lebesgue measure zero. This result makes Lebesgue measure the precise diagnostic for when Riemann integration fails.

Fubini's theorem. For an absolutely integrable function f : ℝⁿ⁺ᵐ → ℝ, Fubini's theorem states that the double integral equals the iterated integral in either order: ∫{ℝⁿ⁺ᵐ} f(x, y) d(x,y) = ∫{ℝⁿ}(∫{ℝᵐ} f(x,y) dy) dx = ∫{ℝᵐ}(∫_{ℝⁿ} f(x,y) dx) dy. Tonelli's theorem provides the non-negative analogue: for non-negative measurable f, the iterated integrals are always equal (possibly ∞). Together these theorems allow multidimensional integrals to be computed slice by slice, the measure-theoretic generalisation of "integrate one variable at a time." The proof proceeds through the stages: indicator functions, simple functions, non-negative functions (Tonelli), then absolutely integrable functions (Fubini).

Key ideas

• The Lebesgue integral is defined in three stages: simple functions (finite linear combinations of indicators), non-negative functions (supremum over simple functions), absolutely integrable functions (decompose into positive and negative parts).
• The monotone convergence theorem allows limit-integral exchange for increasing sequences with no uniformity assumption.
• The dominated convergence theorem allows limit-integral exchange for any sequence bounded by an integrable function — impossible in the Riemann theory without uniform convergence.
• Lebesgue's characterisation: a bounded function is Riemann integrable iff it is continuous almost everywhere (its discontinuity set has measure zero).
• Fubini's theorem: absolute integrability justifies reversing the order of integration; Tonelli's theorem handles non-negative functions unconditionally.
• The space L¹ of absolutely integrable functions with norm ‖f‖ = ∫|f| is a Banach space (complete normed vector space) — a fact whose proof uses the dominated convergence theorem essentially.
• The Lebesgue integral provides the foundation for L² spaces (Chapter 5's square-integrable functions), for modern probability theory (expectation as a Lebesgue integral), and for the theory of distributions.

Key takeaway

The Lebesgue integral, built from simple functions upward through the supremum construction, is superior to the Riemann integral because the monotone and dominated convergence theorems allow interchange of limits and integrals under mild hypotheses — the engine underlying modern analysis, probability, and Fourier theory.

---

The book's overall argument

1. Chapter 1 (Metric Spaces) — establishes the abstract framework: any set with a distance function supports convergence, completeness, and compactness, generalizing the real line and providing the structural vocabulary for all subsequent chapters.

2. Chapter 2 (Continuous Functions on Metric Spaces) — studies how continuity behaves in the abstract setting, proving that continuous maps preserve both compactness (yielding the extreme value theorem) and connectedness (yielding the intermediate value theorem), lifting classical single-variable theorems to full generality.

3. Chapter 3 (Uniform Convergence) — exposes the inadequacy of pointwise convergence for preserving analytic properties, introduces uniform convergence as the correct notion, proves it preserves continuity and allows limit-integral exchange, and establishes the Weierstrass approximation theorem as the prototype density result in function spaces.

4. Chapter 4 (Power Series) — harnesses uniform convergence to define and rigorously establish the elementary transcendental functions (exp, log, sin, cos) by their power series, grounding them in analysis rather than geometry and defining π analytically.

5. Chapter 5 (Fourier Series) — equips the space of periodic functions with an inner product, identifies e^{inx} as an orthonormal system, and proves via the Plancherel theorem that Fourier series converge isometrically in L², showing that any square-integrable function can be decomposed into frequency components.

6. Chapter 6 (Several Variable Differential Calculus) — generalises differentiation to linear maps (the total derivative), uses the contraction mapping theorem as the central analytic engine, and proves the inverse and implicit function theorems — the deepest structural results in multivariable calculus — as corollaries.

7. Chapter 7 (Lebesgue Measure) — constructs Lebesgue measure by restricting outer measure to the σ-algebra of measurable sets, proving countable additivity on this domain, establishing that non-measurable sets exist but cannot be explicitly constructed, and identifying measurable functions as the correct input class for integration.

8. Chapter 8 (Lebesgue Integration) — builds the Lebesgue integral stage by stage, proves the monotone and dominated convergence theorems that the Riemann integral cannot provide, characterises Riemann integrability in terms of measure, and establishes Fubini's theorem — completing the analytic infrastructure for modern analysis.

---

Common misunderstandings

Misunderstanding: The pointwise limit of continuous functions is continuous.

False. The sequence fₙ(x) = xⁿ on [0, 1] is continuous for each n but converges pointwise to a discontinuous function (0 on [0,1), 1 at x = 1). The correct statement requires uniform convergence. Chapter 3 makes this distinction central to the entire subject.

Misunderstanding: If all partial derivatives of f exist, then f is differentiable.

False. Tao provides an explicit counterexample in Chapter 6: f(x, y) = xy/(x² + y²) for (x,y) ≠ 0 has directional derivatives in every direction at the origin but is not differentiable there. Differentiability requires a linear approximation valid in all directions simultaneously, not merely along coordinate axes.

Misunderstanding: The Lebesgue integral is merely a technical generalisation useful only for pathological functions.

The Lebesgue integral's advantage is not about exotic functions but about the behaviour of limits. The dominated convergence theorem — only available in the Lebesgue theory — is essential for modern analysis, probability theory, and Fourier theory. Every Riemann integral is a Lebesgue integral; the converse fails, and the Riemann theory cannot exchange limits and integrals without uniform convergence.

Misunderstanding: Outer measure is the same as Lebesgue measure.

Outer measure is defined for all subsets of ℝⁿ but is not countably additive on all of them; the Vitali set is a non-measurable set on which countable additivity fails. Lebesgue measure is the restriction of outer measure to the σ-algebra of measurable sets, where countable additivity holds. The two agree on measurable sets but the restriction is foundational.

Misunderstanding: The Fourier series of a continuous function converges pointwise everywhere to that function.

This is false: du Bois-Reymond constructed a continuous function whose Fourier series diverges at a point. Chapter 5 proves L² convergence (via the Plancherel theorem) and uniform approximation by Fejér sums (Cesàro averages of the partial sums), which are the correct convergence results at the level of this text. Pointwise convergence of Fourier series everywhere for continuous functions (Carleson's theorem, 1966) is far beyond the scope of the book.

Misunderstanding: Analysis II is independent of Analysis I.

Analysis II begins immediately in the generality of metric spaces, assuming the reader is fluent with the single-variable results from Analysis I — real numbers, sequences, series, limits, continuity on ℝ, differentiation, and Riemann integration. The two volumes form a single conceptual arc.

---

Central paradox / key insight

The book's deepest insight is the incompatibility between universal measurability and countable additivity. Every student's intuition says: any set should have a size, and the size of a countable disjoint union should be the sum of the individual sizes. The Vitali construction in Chapter 7 shows that both demands cannot be simultaneously satisfied for all subsets of ℝ: the axiom of choice produces a set to which no translation-invariant, countably additive measure can be assigned consistently. This is not a defect of the Lebesgue theory; it is a fundamental obstruction in set theory.

The resolution is to restrict the domain: instead of measuring all sets, one identifies the σ-algebra of Lebesgue measurable sets — large enough to contain every set that arises naturally in analysis (open, closed, compact, null sets), yet small enough to exclude the pathological ones. On this restricted domain, countable additivity holds, and the theory is complete and powerful.

To make integration work under limits, you must accept that not every set has a size — but the sets that lack a size are precisely those that analysis never needs.

This is the decisive move that makes the dominated convergence theorem and Fubini's theorem possible, and in turn makes all of modern analysis, probability theory, and L²-Fourier theory coherent.

---

Important concepts

Metric space

A set X with a distance function d : X × X → ℝ satisfying positivity (d(x,y) ≥ 0 with equality iff x = y), symmetry (d(x,y) = d(y,x)), and the triangle inequality (d(x,z) ≤ d(x,y) + d(y,z)). The minimal structure for convergence and continuity.

Complete metric space

A metric space in which every Cauchy sequence converges within the space. The real line ℝ is complete; ℚ is not. Completeness is required for the contraction mapping theorem and for function spaces like C([a,b]) with the supremum norm.

Compact metric space

A metric space in which every sequence has a convergent subsequence. In ℝⁿ, a set is compact iff it is closed and bounded (Heine-Borel theorem). Compact domains force uniform continuity and the attainment of extrema.

Uniform convergence

A sequence fₙ : X → Y converges uniformly to f if for every ε > 0 there exists N independent of x such that n ≥ N implies d(fₙ(x), f(x)) < ε for all x. Stronger than pointwise convergence; preserves continuity and permits limit-integral exchange.

Weierstrass M-test

If |fₙ(x)| ≤ Mₙ for all x and ΣMₙ < ∞, then Σfₙ converges uniformly and absolutely. The standard tool for series of functions.

Radius of convergence

For the power series Σaₙxⁿ, R = 1/limsup_{n→∞} |aₙ|^{1/n} (Cauchy-Hadamard formula). The series converges absolutely for |x| < R and diverges for |x| > R.

Real analytic function

A function that locally equals a convergent power series. Strictly stronger than being infinitely differentiable; Taylor coefficients are uniquely determined by the function.

Fourier coefficients

For a 2π-periodic function f, the coefficients ĉₙ = (1/2π) ∫_{−π}^π f(x) e^{−inx} dx for n ∈ ℤ. They measure the amplitude of the frequency-n oscillation in f.

Plancherel theorem

For a square-integrable 2π-periodic function f: ‖f‖²{L²} = Σ{n=−∞}^∞ |ĉₙ|². The Fourier transform is an isometric isomorphism from L² to ℓ².

Total derivative (Fréchet derivative)

For f : ℝⁿ → ℝᵐ, the unique linear map Df(x₀) (if it exists) satisfying lim_{h→0} ‖f(x₀+h) − f(x₀) − Df(x₀)h‖/‖h‖ = 0. Represented by the m×n Jacobian matrix of partial derivatives.

Contraction mapping

A function f : X → X with Lipschitz constant L < 1: d(f(x), f(y)) ≤ L·d(x,y). The Banach fixed-point theorem guarantees a unique fixed point, found by iteration.

σ-algebra

A collection of subsets of ℝⁿ closed under complementation and countable unions (hence also countable intersections). The Lebesgue measurable sets form a σ-algebra; any measure must be defined on a σ-algebra to guarantee countable additivity.

Lebesgue outer measure

m*(E) = inf{Σ vol(Bₖ) : E ⊆ ∪Bₖ, each Bₖ an open box}. Defined for all subsets of ℝⁿ; monotone and countably subadditive but not countably additive on arbitrary sets.

Lebesgue measurable set

A set E for which every ε > 0 admits an open U ⊇ E with m*(U \ E) < ε. On this σ-algebra, the Lebesgue measure m (= outer measure restricted here) is countably additive.

Simple function

A measurable function taking finitely many values: s = Σ aₖ 1_{Eₖ}, with each Eₖ a measurable set. The building blocks of the Lebesgue integral; analogous to step functions in the Riemann theory.

Dominated convergence theorem

If fₙ → f pointwise, |fₙ| ≤ g, and g is Lebesgue integrable (∫g < ∞), then ∫fₙ → ∫f. The central limit theorem of Lebesgue integration, replacing the uniform convergence hypothesis of the Riemann theory.

Fubini's theorem

For f ∈ L¹(ℝⁿ⁺ᵐ): ∫{ℝⁿ⁺ᵐ} f = ∫{ℝⁿ}(∫_{ℝᵐ} f(x,y) dy) dx. Iterated integrals over product spaces can be computed in either order when f is absolutely integrable.

---

References and Web Links

Primary book and edition information

• Tao, Terence. Analysis II. Fourth Edition. Texts and Readings in Mathematics, Vol. 38. Hindustan Book Agency / Springer Singapore, 2022. ISBN 978-981-19-7284-3.
  • Analysis II: Fourth Edition — Springer Nature
  • Analysis II: Fourth Edition — Google Books
• Tao, Terence. Analysis II. Third Edition. Texts and Readings in Mathematics, Vol. 38. Hindustan Book Agency / Springer Singapore, 2016. ISBN 978-981-10-1804-6.
  • Analysis II: Third Edition — Springer Nature
  • Analysis II: Third Edition — Google Books
• Analysis II: First Edition (2006) — Internet Archive
• Analysis II — Open Library

Author's official page and errata

• Analysis II — Terence Tao's official book page with errata
• Analysis I — companion volume

Background and overview

• Springer Professional — Analysis II chapter listing

Lebesgue measure and integration (Tao's own lecture notes)

• 245A Notes 1: Lebesgue Measure — terrytao.wordpress.com
• 245A Notes 2: The Lebesgue Integral — terrytao.wordpress.com
• An Introduction to Measure Theory — Tao's book page

Fourier series and harmonic analysis (Tao's notes)

• 245C Notes 2: The Fourier Transform — terrytao.wordpress.com

Point-set topology (Tao's notes)

• 245B Notes 8: A quick review of point-set topology — terrytao.wordpress.com

Additional study resources

These are secondary summaries and should be used alongside, rather than instead of, the original book.

• Analysis, Part II: Continuous Functions on Metric Spaces — Jay Havaldar's notes
• Terence Tao's books — MacTutor History of Mathematics