Here is a fact that ought to be impossible. Pick a continuous function on [0,1] "at random" — whatever that should mean. It is almost certainly nowhere differentiable: a jagged, fractal monster with no tangent line at a single point. The smooth, well-behaved functions you drew all through calculus are not the norm — they are a vanishingly rare exception. Nobody proves this by exhibiting the typical monster. Instead one shows that the set of "nice" functions is topologically negligible, and the engine that makes "negligible" a rigorous, usable idea is a single theorem about complete metric spaces: the Baire category theorem.

That same theorem is the quiet workhorse under the three cornerstone results of linear analysis on Banach spaces — the uniform boundedness principle, the open mapping theorem, and the closed graph theorem. Each looks like a statement about linear operators; each is, at heart, the observation that a complete space is too "big" to be built out of countably many thin pieces. Learn Baire once and all three fall out of the same move. This page teaches that one idea — what it says, why completeness is the whole game, and how it powers functional analysis.

We work throughout in a metric space (X, d); the deep applications need it to be complete (every Cauchy sequence converges in X). For the operator theorems X and Y will be Banach spaces — complete normed vector spaces.

Step 1 — nowhere dense: sets with no room inside

The first building block measures a very strong kind of smallness. A subset A \subseteq X is nowhere dense if its closure has empty interior:

Unpack it. Taking the closure \overline{A} first is essential — it fills in all the limit points, so we cannot cheat by pointing at a "gappy" set that is secretly dense. Empty interior then says: no open ball fits inside \overline{A}. Equivalently, the complement X \setminus \overline{A} is a dense open set — around every point, arbitrarily close, there are points that miss A entirely. A nowhere-dense set is one you can always dodge.

• A single point in \mathbb{R} is nowhere dense — its closure is itself, and no interval fits in a point.
• A line inside \mathbb{R}^2, or the integers \mathbb{Z} inside \mathbb{R}: closed already, and containing no ball. Nowhere dense.
• The middle-thirds Cantor set C \subseteq [0,1]: closed (it is an intersection of closed sets) and containing no interval (at stage n it is covered by 2^n intervals of total length (2/3)^n \to 0). Nowhere dense — yet uncountable. Small in this topological sense has nothing to do with cardinality.
• The rationals \mathbb{Q} \subseteq \mathbb{R} are not nowhere dense: \overline{\mathbb{Q}} = \mathbb{R}, whose interior is all of \mathbb{R}. But — crucially — \mathbb{Q} is a countable union of the nowhere-dense singletons \{q\}. That distinction is the next idea.

Step 2 — meagre vs comeagre: a topological notion of "size"

Nowhere-dense is the atom of smallness. Close it under countable unions and you get the working notion. A set is meagre (older name: of the first category) if it is a countable union of nowhere-dense sets:

Anything not meagre is of the second category; the complement of a meagre set is comeagre (or residual). Think of meagre as "topologically negligible" and comeagre as "topologically almost everything." So \mathbb{Q}, a countable union of points, is meagre in \mathbb{R}, and the irrationals \mathbb{R} \setminus \mathbb{Q} are comeagre.

Why insist on countable unions? Two reasons. First, meagreness is meant to be a \sigma-ideal — closed under countable unions, so that "a countable list of negligible things is still negligible" (a monotone-class-style closure that mirrors the measure-zero sets). Second, an arbitrary union of nowhere-dense sets is useless: every point \{x\} is nowhere dense, and their uncountable union is all of X. The line between countable and uncountable is exactly where the content lives — and Baire's theorem is the statement that, in a complete space, that line cannot be crossed.

Step 3 — the theorem

Everything above is definitions; here is the one theorem, in the three interchangeable forms you will meet in the wild. All three say the same thing about a complete metric space.

The three forms are De Morgan duals: complement the dense open G_n to get nowhere-dense closed F_n = X \setminus G_n, and "the intersection stays dense" becomes "the union misses points," i.e. cannot be all of X. The category form is the one that fuels functional analysis: write your complete space as a countable union of closed pieces, and one piece must be fat enough to contain a ball.

Why completeness does all the work: the nested-ball proof

Take the intersection form. Let G_1, G_2, \dots be dense open, and let B_0 be any open ball; we must show B_0 \cap \bigcap_n G_n \ne \varnothing. Because G_1 is dense and open, it meets B_0 in a nonempty open set, so we can find a closed ball

Now repeat inside B_1 using G_2: density and openness give a closed ball \overline{B_2} \subseteq B_1 \cap G_2 of radius < \tfrac{1}{4}. Continuing, we get a nested sequence

with \overline{B_n} \subseteq G_n. The centres (x_n) form a Cauchy sequence — for m, n \ge N both centres lie in \overline{B_N}, so d(x_m, x_n) < 2^{-N+1}. Here, and only here, we spend completeness: the limit x = \lim x_n exists in X. Since the balls are closed and nested, x \in \overline{B_n} for every n, hence x \in \bigcap_n G_n, and also x \in B_0. Done.

Remove completeness and the sequence of shrinking demands can converge to a "hole" that isn't a point of the space — the whole argument leaks out. That is not a technicality; it is the theorem's entire content. Watch it happen in the figure below.

Seeing Baire: a dense open set you can shrink to nothing

Enumerate the rationals in [0,1] as q_1, q_2, q_3, \dots and, given a budget \varepsilon > 0, put an open interval of length \varepsilon \, 2^{-k} around q_k. Their union

is open and dense (it contains every rational, and the rationals are dense), yet its total length is at most \sum_k \varepsilon\,2^{-k} = \varepsilon — as small as we please. Step through the construction, and slide \varepsilon toward 0: the covered set thins out but never stops being dense. Now Baire says that intersecting countably many such dense open sets \bigcap_m G_{1/m} still leaves a dense residual — a comeagre set of measure zero on which "most" of the interesting behaviour lives.

The engine: one move, three theorems

Every one of the three big theorems runs the same play. You have a Banach space X — complete, hence non-meagre in itself. You write it as a countable union of closed sets X = \bigcup_n F_n built from the operators in question. Baire's category form forces one F_N to contain an open ball. A dose of linearity (scaling and translating that ball, using symmetry and convexity) then spreads the local control to the whole space. That is the entire strategy; only the choice of F_n changes. Keep it in mind as the three theorems go by.

Theorem A — the Uniform Boundedness Principle (Banach–Steinhaus)

How Baire drives it. For n \in \mathbb{N} set

Each \{x : \|T_i x\| \le n\} is closed (preimage of a closed set under the continuous map x \mapsto \|T_i x\|), so F_n is closed. Pointwise boundedness says every x lands in some F_n, i.e. X = \bigcup_n F_n. By Baire, some F_N contains a ball B(x_0, r). For \|z\| \le r we then have x_0, x_0 + z \in F_N, so by the triangle inequality \|T_i z\| \le \|T_i(x_0+z)\| + \|T_i x_0\| \le 2N for every i. Rescaling, \|T_i\| \le 2N/r for all i. The pointwise data has become a uniform bound — that is the Baire miracle in three lines.

Theorem B — the Open Mapping Theorem (and bounded inverse)

How Baire drives it. It suffices to show the image T\big(B_X(0,1)\big) of the open unit ball contains a ball around 0 in Y. Since T is onto and X = \bigcup_n B_X(0,n),

Y is complete, so by Baire one of these closed sets has nonempty interior; by scaling, \overline{T(B_X(0,1))} contains a ball. A second step — a successive-approximation argument that uses completeness of X to sum a rapidly converging series of preimages — upgrades the closure \overline{T(B_X(0,1))} to the genuine image T(B_X(0,2)), removing the bar. Both completenesses appear: Baire needs Y complete, the series needs X complete.

Theorem C — the Closed Graph Theorem

How Baire drives it (via B). The "only if" is easy. For "if", equip the product X \times Y with the norm \|(x,y)\| = \|x\| + \|y\| — it is a Banach space. The graph \Gamma(T), being a closed subspace of a Banach space, is itself Banach. The projection \pi_X : \Gamma(T) \to X, (x, Tx) \mapsto x, is a bounded linear bijection between Banach spaces, so by the bounded inverse theorem its inverse x \mapsto (x, Tx) is bounded — whence \|Tx\| \le \|(x,Tx)\| \le C\|x\| and T is bounded. So the closed graph theorem is the open mapping theorem in disguise, and both are Baire.

The practical payoff: to prove an operator continuous you no longer have to produce the limit of Tx_n from thin air — the hardest half of continuity is granted for free. You merely verify that if the two limits exist they agree, which is usually a one-line computation.

The one hypothesis you must never drop

All three theorems carry the word Banach for a reason. Baire is a theorem about complete spaces, so every result it powers inherits completeness as a load-bearing wall — of both spaces. Kick it out and each theorem fails, often dramatically.