The Baire Category Theorem

The Baire Category Theorem is a load-bearing wall of functional analysis: it is the engine behind the Uniform Boundedness Principle (Banach–Steinhaus), the Open Mapping Theorem, and the Closed Graph Theorem, all of which are proved by applying Baire's theorem to a Banach space decomposed as a countable union of carefully chosen closed sets. These three theorems, in turn, underlie essentially all of linear PDE theory, spectral theory of operators, and the mathematical foundations of quantum mechanics, wherever unbounded or bounded linear operators on complete normed spaces are analyzed. In descriptive set theory, the notions of meager and comeager sets, together with their measure-theoretic analogues, structure the modern classification of subsets of Polish spaces (complete separable metric spaces), feeding into logic, set theory, and the study of definable sets of reals. In dynamical systems, genericity arguments via category show that "typical" dynamical systems (in a topological sense) exhibit chaotic or otherwise exotic behavior, complementing measure-theoretic ("almost every," in the sense of a natural invariant measure) genericity statements — the two notions can disagree, famously, on which behaviors are "typical." In numerical analysis and approximation theory, category arguments establish the existence of functions with pathological approximation properties (e.g. functions with divergent Fourier series at a point, first shown to exist this way before explicit examples were later constructed). In theoretical computer science and complexity theory, an analogous "resource-bounded" Baire category has been developed to argue about typical behavior of algorithms and complexity classes. Even in cryptography and coding theory, generic existence arguments over complete metric spaces of codes or parameters occasionally invoke the same category philosophy to show that "generic" instances avoid degenerate structure.

Ask the tutor: How does the Baire Category Theorem's nested-ball proof adapt to show locally compact Hausdorff spaces are also Baire spaces, even without a metric?

Think of nowhere dense sets as sets so sparse that they can't even fill up a small ball: no matter which open ball you pick, you can always find a smaller ball entirely missing the set. A single point is nowhere dense in $\mathbb{R}$; so is a line segment sitting inside $\mathbb{R}^2$; so is the Cantor set inside $\mathbb{R}$. A meager set is a countable union of such sparse sets — still, intuitively, "small" — but a priori a countable union could conceivably add up to everything, the way countably many measure-zero sets can have positive measure union in the wrong setting, or the way $\mathbb{Q}$, a countable union of singletons, is dense in $\mathbb{R}$. The theorem's content is that in a complete metric space this cannot happen: no matter how you choose countably many nowhere dense sets, their union never exhausts the whole space, and in fact its complement remains dense. The proof is a nested-ball construction that manufactures a point avoiding all the given nowhere dense sets simultaneously, using completeness exactly at the moment we need a Cauchy sequence of shrinking balls to converge to an actual point of $X$. Picture a target space and a sequence of "obstacles" $A_1, A_2, A_3,\dots$, each nowhere dense. Starting from any open ball, we can always dodge $A_1$ by shrinking into a sub-ball missing $\overline{A_1}$; then dodge $A_2$ inside that; and so on, forever. Completeness guarantees the infinite dodging process converges to a genuine point that has escaped every obstacle.

Example 1: $\mathbb{R}$ is uncountable, proved by category (no diagonal argument). We show that $\mathbb{R}$, with its usual metric, is not countable — using only the Baire Category Theorem, as an illustration of its power as an existence tool. $\mathbb{R}$ is a complete metric space (Cauchy sequences of reals converge, a basic fact from real analysis). Suppose toward contradiction that $\mathbb{R}$ were countable, say $\mathbb{R} = \{a_1, a_2, a_3, \dots\}$, an enumeration of all real numbers. For each $n$, let $A_n = \{a_n\}$, a singleton. A singleton in $\mathbb{R}$ is closed (since $\mathbb{R}$ is Hausdorff/metric) and has empty interior (it contains no open interval), so each $A_n$ is nowhere dense. Then

exhibits $\mathbb{R}$ as a countable union of nowhere dense sets, i.e. $\mathbb{R}$ is meager in itself. This directly contradicts the Corollary above (a nonempty complete metric space cannot be meager in itself). Hence $\mathbb{R}$ is not countable. $\blacksquare$ This is worth pausing on: the existence of an uncountability proof via category, entirely different in flavor from Cantor's diagonal argument, shows the theorem is not just an abstract curiosity — it directly controls cardinality-adjacent phenomena through topology.

Example 2: existence of a continuous function that is nowhere differentiable (existence, not construction). Let $X = C[0,1]$, the space of continuous real-valued functions on $[0,1]$, equipped with the sup-norm metric

This is a complete metric space: a standard fact is that $(C[0,1], \|\cdot\|_\infty)$ is a Banach space, hence complete (uniform limits of continuous functions on a compact interval are continuous, and uniform Cauchy sequences converge uniformly). For each positive integer $n$, define

Intuitively, $E_n$ consists of functions that have "some point where the difference quotients up to scale $1/n$ stay bounded by $n$" — a strong finite surrogate for having a bounded derivative somewhere. Any function $f$ that is differentiable at even a single point $x_0 \in [0,1)$ lies in $E_n$ for all sufficiently large $n$ (take $n$ larger than $|f'(x_0)|+1$ and larger than $1/(1-x_0)$, using the definition of derivative to bound difference quotients near $x_0$). Hence

Claim: each $E_n$ is closed and nowhere dense in $C[0,1]$. Closedness. If $f_k \in E_n$ with $f_k \to f$ uniformly, let $x_k \in [0,1-1/n]$ witness $f_k \in E_n$. By compactness of $[0,1-1/n]$, a subsequence $x_{k_j} \to x_0 \in [0,1-1/n]$. A uniform limit argument (using uniform convergence to pass difference-quotient bounds to the limit, together with continuity of $f$) shows $x_0$ witnesses $f \in E_n$; so $E_n$ is closed. (This step uses a standard but slightly technical uniform-convergence estimate; it is a well-known component of this classical construction.) Empty interior. Given any $f \in C[0,1]$ and $\varepsilon > 0$, we must find $g$ with $\|f - g\|_\infty < \varepsilon$ and $g \notin E_n$. First approximate $f$ uniformly within $\varepsilon/2$ by a piecewise-linear function $p$ (possible since $f$ is uniformly continuous on the compact interval $[0,1]$). Then perturb $p$ by adding a small sawtooth (zig-zag) function $z$ of amplitude less than $\varepsilon/2$ but with slope magnitude greater than $2n$ everywhere it is differentiable, giving $g = p+z$. Since $p$ has bounded slope on each linear piece and $z$'s slope dominates and oscillates, every difference quotient of $g$ over intervals of length $\le 1/n$ at any point exceeds $n$ in absolute value at some sub-scale, so $g \notin E_n$, while $\|f-g\|_\infty \le \|f-p\|_\infty + \|z\|_\infty < \varepsilon$. Hence every open ball in $C[0,1]$ contains a point outside $E_n$, i.e. $E_n$ has empty interior; combined with closedness, $\overline{E_n}=E_n$ has empty interior, so $E_n$ is nowhere dense. By the Baire Category Theorem, $\bigcup_n E_n \neq C[0,1]$, and in fact $C[0,1] \setminus \bigcup_n E_n$ is dense (indeed residual). Any $f$ in this residual complement is differentiable at no point of $[0,1)$ (since differentiability at any point would force $f \in E_n$ for some $n$, contradiction), and a symmetric argument with left-derivatives handles the endpoint $t=1$. Hence:

Note the method: we never wrote down such a function explicitly (unlike Weierstrass's famous trigonometric series construction); we proved that residually many continuous functions have this property, which is topologically much stronger — it says nowhere-differentiable functions are "generic," while differentiable-somewhere functions are the meager exception.