What if "distance" didn't mean a ruler? A taxi in Manhattan cannot cut diagonally through a city block — for it, the distance between two corners is the number of blocks driven, not the crow-flies length. A chess king measures distance in moves: three squares east and two north is three moves away, because a diagonal step covers both directions at once. A spell-checker measures the distance between recieve and receive in edit operations. And in uniform convergence you have already measured the distance between two functions as the biggest gap between their graphs.

Four rulers, four utterly different notions of "how far" — and yet every argument in analysis you have met so far leans on just a handful of properties of the humble |x - y|. Distil those properties into axioms, and something remarkable happens: limits, continuity, convergence, completeness — the whole machinery — runs unchanged in every one of these worlds. Prove a theorem once, from the axioms, and it holds for numbers, for points of the plane under any of these distances, for spaces of functions, for words in a dictionary. That abstraction is the metric space, and it is the gateway from calculus into modern analysis.

A metric on a set X is a function d : X \times X \to [0, \infty) assigning a distance to each pair of points, subject to three axioms. For all x, y, z \in X:

The pair (X, d) is a metric space. Notice what the axioms do not say: nothing about coordinates, angles, straight lines, or even numbers. X is a bare set — its "points" may be words, sequences, functions, matrices. If a rule for measuring gaps obeys (M1)–(M3), all of analysis is open for business on it.

Earning each axiom: what breaks without it

Axioms should never be a shopping list to memorise — each one is there because dropping it destroys something you need.

• Without identity (M1): try d(x, y) = |x^2 - y^2| on \mathbb{R}. It is symmetric and even obeys the triangle inequality — but d(1, -1) = 0 although 1 \ne -1. Two different points at distance zero means limits stop being unique: a sequence could "converge" to both 1 and -1 at once. (Weaken this axiom deliberately and you get a pseudometric — useful, but you pay in ambiguity.)
• Without symmetry (M2): think of travel time in a city of one-way streets, or walking uphill versus downhill. "Time from x to y" can genuinely differ from the reverse — a legitimate object (a quasimetric), but now "the ball around x" splits into a ball of points you can reach and a ball of points that can reach you, and every definition doubles.
• Without the triangle inequality (M3): try d(x, y) = (x - y)^2 on \mathbb{R}. Positive, zero exactly on the diagonal, symmetric — it looks innocent. But d(0, 2) = 4 while d(0, 1) + d(1, 2) = 1 + 1 = 2: the detour is shorter than the direct trip. With that, "close to close is close" fails, and the workhorse estimate of all analysis — bound d(x, z) by hopping through an intermediate point — is gone. No triangle inequality, no limit theorems.

One pleasant economy: you never need to check d \ge 0 separately. From the other axioms, 0 = d(x, x) \le d(x, y) + d(y, x) = 2\,d(x, y) — non-negativity is free.

A gallery of metrics

• \mathbb{R} with the absolute value: d(x, y) = |x - y|. The ordinary number line — the special case all the others generalise.
• \mathbb{R}^n with the Euclidean metric: d_2(x, y) = \sqrt{\sum_{i=1}^{n} (x_i - y_i)^2}. Straight-line ("as the crow flies") distance.
• \mathbb{R}^2 with the taxicab metric: d_1(x, y) = |x_1 - y_1| + |x_2 - y_2|. Blocks east plus blocks north — no diagonals allowed.
• \mathbb{R}^2 with the Chebyshev (max) metric: d_\infty(x, y) = \max\big(|x_1 - y_1|,\, |x_2 - y_2|\big). Exactly the number of moves a chess king needs, since one move fixes a step in both coordinates.
• Functions with the sup metric: on the continuous functions on [a, b], d_\infty(f, g) = \sup_{x} |f(x) - g(x)|. Two functions are "close" when their graphs are everywhere close — exactly the uniform distance, now recognised as a genuine metric on a space whose points are functions.
• Any set X with the discrete metric: d(x, y) = 0 if x = y and d(x, y) = 1 otherwise. Weird but perfectly legal — see the worked example below.

Worked example 1: the taxicab metric really is a metric

Claim: d_1(x, y) = |x_1 - y_1| + |x_2 - y_2| satisfies (M1)–(M3) on \mathbb{R}^2.

(M1). A sum of two absolute values is \ge 0, and it equals 0 only when both terms vanish — that is, when x_1 = y_1 and x_2 = y_2, i.e. x = y. (Note the "and": one matching coordinate is not enough, which is exactly why the identity axiom quantifies over the whole point.)

(M2). |x_i - y_i| = |y_i - x_i| coordinate by coordinate, so the sum is symmetric.

(M3). The one-dimensional triangle inequality |a + b| \le |a| + |b| gives, in each coordinate,

Add the two coordinate inequalities and regroup:

Done — the grid-walking distance is a genuine metric, and the same argument works in \mathbb{R}^n for any n. Notice the strategy, because it recurs throughout analysis: reduce a many-dimensional triangle inequality to the one-dimensional one, coordinate by coordinate, then sum.

Worked example 2: the Euclidean metric obeys the triangle inequality

For d_2, positivity and symmetry are again immediate (a square root of a sum of squares is \ge 0 and vanishes only when every coordinate matches; swapping x and y changes nothing). The real work is the triangle inequality — and this time no coordinate-by-coordinate shortcut exists, because the square root tangles the coordinates together. Write the claim with displacement vectors u = x - y and v = y - z, so that x - z = u + v and the claim becomes \|u + v\| \le \|u\| + \|v\|.

Step 1 — square the target. Both sides are non-negative, so it suffices to compare their squares. Expand using the dot product:

Step 2 — bound the cross term by Cauchy–Schwarz. The one fact we borrow is the Cauchy–Schwarz inequality |u \cdot v| \le \|u\|\,\|v\|. In particular u \cdot v \le \|u\|\,\|v\|, so

Step 3 — recognise a perfect square. The right-hand side is (\|u\| + \|v\|)^2:

Step 4 — take square roots. Both sides are non-negative, and the square root is increasing, so

Step 5 — translate back to distances. Substituting u = x - y and v = y - z gives exactly (M3):

So Euclidean distance is a genuine metric. Geometrically this is the schoolroom fact that any side of a triangle is at most the sum of the other two — a detour through y is never shorter than the straight path.

Worked example 3: the discrete metric — weird but legal

On any set X whatsoever, define

(M1) and (M2) hold by construction. For (M3), check d(x, z) \le d(x, y) + d(y, z) by cases: if x = z the left side is 0 and there is nothing to prove. If x \ne z, the left side is 1 — and y cannot equal both x and z, so at least one term on the right is 1. Every case checks out: a metric.

Its geometry is gloriously strange. Every point sits in splendid isolation: the ball B(x, \tfrac{1}{2}) contains only x itself, so every singleton — indeed every subset — is an open set. Every function out of a discrete space is continuous. And, as the "Watch out!" box warned, a sequence converges only by eventually sitting still. The discrete metric is the standard stress test: whenever you conjecture "surely every metric space behaves like \mathbb{R}^n…", run it past the discrete metric first.

The shape of "within radius r": the iconic picture

Here is the picture every analyst carries in their head. Fix a centre and a radius r, and draw the open ball B(0, r) = \{\, y : d(0, y) < r \,\} in three different metrics on the same plane. In the Euclidean metric d_2 it is the familiar round disc. In the taxicab metric d_1 it is a diamond — the set |x| + |y| < r, whose corners poke out along the axes because moving along one axis is "cheap". In the Chebyshev metric d_\infty it is an axis-aligned square — the set \max(|x|, |y|) < r, everything a king reaches in fewer than r moves.

Slide the radius and watch all three grow together. Notice the nesting, \text{diamond} \subset \text{disc} \subset \text{square}, which encodes the inequalities d_\infty \le d_2 \le d_1 — the same pair of points is "closest" in the king's metric and "farthest" in the taxi's. Because each ball fits inside a scaled copy of the others (e.g. d_1 \le 2\,d_\infty in the plane), the three metrics declare exactly the same sequences convergent and the same sets open: they are equivalent metrics. Different geometry, identical analysis.

The vocabulary, generalised — analysis runs verbatim

With a metric in hand, the basic notions of analysis transplant word for word — every |x - y| becomes d(x, y), and nothing else changes.

• Open ball: B(x, r) = \{\, y \in X : d(x, y) < r \,\} — all points within radius r of x. In C[a,b] with the sup metric, B(f, \varepsilon) is a tube of width 2\varepsilon around the graph of f.
• Open set: a set U such that every point of U has some ball B(x, r) \subseteq U around it — the metric's version of "no boundary points included".
• Convergence: x_n \to x means d(x_n, x) \to 0; spelled out, for every \varepsilon > 0 there is an N with d(x_n, x) < \varepsilon for all n \ge N — the tail of the sequence enters every \varepsilon-ball. In (C[a,b], d_\infty) this is exactly uniform convergence.
• Continuity: f : (X, d_X) \to (Y, d_Y) is continuous at a when for every \varepsilon > 0 there is a \delta > 0 such that d_X(x, a) < \delta \Rightarrow d_Y\big(f(x), f(a)\big) < \varepsilon. Compare this with the \varepsilon–\delta definition on \mathbb{R}: it is the same sentence with the absolute values renamed — and now it covers maps between any two metric spaces, even with different metrics on each side.

Cauchy = convergent becomes completeness. Recall that on the real line a sequence converges iff it is Cauchy — its terms eventually crowd together. A general (x_n) is Cauchy when d(x_m, x_n) \to 0 as m, n \to \infty. Crowding together does not by itself guarantee a limit inside the space — the rationals are Cauchy toward \sqrt{2}, which is not rational. A metric space in which every Cauchy sequence does converge to a point of the space is called complete:

"\mathbb{R} is complete" is precisely the no-holes property of the reals, now seen as one instance of a property any metric space may or may not have. The space (C[a,b], d_\infty) is complete too — that is the uniform-limit theorem wearing its metric-space clothes — and the discrete metric is trivially complete, since its Cauchy sequences are eventually constant.