Euclidean Space and Norms

How do you measure the distance between two points? This could be in the plane, in space, or in the abstract 1717-dimensional space of tuning parameters for a machine learning model. The question sounds naive. The honest answer requires building an entire structure from scratch. You need a notion of length, a notion of angle, and a proof these notions behave consistently with algebra. Euclidean space Rn\mathbb{R}^n is the arena for almost all of multivariable calculus, differential geometry, and optimization theory. Its geometry is entirely encoded in a single algebraic gadget: the inner product. This matters because calculus in one variable leans on the order structure of R\mathbb{R}. You compare numbers, take limits using ∣a−b∣|a-b|, and everything is linear along a single axis. In Rn\mathbb{R}^n there is no natural order. So distance itself must be manufactured from the vector space structure. The inner product does exactly this. From it we extract the Euclidean norm ∥x∥\|x\|, which plays the role ∣a−b∣|a-b| played in single-variable analysis. The subtlety is that the norm must satisfy a triangle inequality to deserve the name "distance." Proving this inequality is genuinely nontrivial. It hinges on the celebrated Cauchy–Schwarz inequality. The big picture is this: Rn\mathbb{R}^n is simultaneously three things. It is a vector space, allowing addition and scalar multiplication. It is an inner product space, allowing lengths and angles. It is also a metric space, allowing limits, continuity, open sets. This article builds the bridge between the first and the third via the second. We also address a question that seems merely technical. Yet it turns out to be foundational for all of analysis in finite dimensions. Does it matter which norm you use to measure distance? The answer — no, up to constants, in finite dimensions — is the equivalence of norms theorem. It is what allows topology in Rn\mathbb{R}^n to be defined without reference to a particular norm at all. By the end of this article you will have a complete, rigorous construction. This includes the Euclidean inner product and norm. You will also have a full proof of Cauchy–Schwarz and the triangle inequality. Finally, you will see a proof that all norms on a finite-dimensional real vector space are equivalent. This groundwork is what makes the later study of Open and Closed Sets in R^n and Limits and Continuity in Several Variables possible.

Key takeaways
  • Rn\mathbb{R}^n carries a standard inner product ⟨x,y⟩=∑xiyi\langle x,y\rangle = \sum x_i y_i, from which the Euclidean norm ∥x∥=⟨x,x⟩\|x\| = \sqrt{\langle x,x\rangle} is defined.
  • The Cauchy–Schwarz inequality ∣⟨x,y⟩∣≤∥x∥∥y∥|\langle x,y\rangle| \le \|x\|\|y\| is the technical engine behind almost everything geometric in Rn\mathbb{R}^n.
  • The triangle inequality ∥x+y∥≤∥x∥+∥y∥\|x+y\| \le \|x\| + \|y\| follows from Cauchy–Schwarz and makes ∥⋅∥\|\cdot\| a genuine distance-measuring tool.
  • A norm is any function satisfying positivity, homogeneity, and the triangle inequality; Rn\mathbb{R}^n admits infinitely many norms (ℓ1\ell^1, ℓ2\ell^2, ℓ∞\ell^\infty, ...).
  • In finite dimensions, all norms are equivalent: they induce the same convergent sequences, the same open sets, the same notion of continuity.
  • This equivalence is exactly what makes it safe to say "the topology of Rn\mathbb{R}^n" without specifying a norm.
  • Key idea: distance in Rn\mathbb{R}^n is built from an inner product, and while the specific numbers depend on the norm chosen, the underlying notion of "close" and "far" does not.

Where this lives today

The Euclidean norm and Cauchy–Schwarz inequality are structural load-bearing walls across mathematics and its applications. In statistics and machine learning, the Cauchy–Schwarz inequality is the reason the correlation coefficient between two random variables always lies in [−1,1][-1,1] — correlation is literally an inner product of centered, normalized random variables. Regularization terms in regression (ridge regression, LASSO) are built directly from the ℓ2\ell^2 and ℓ1\ell^1 norms discussed here, and the choice between them has real consequences (sparsity versus smoothness) precisely because these norms, while topologically equivalent, are numerically and geometrically different — their unit balls (illustrated above) have different shapes that induce different optimization behavior. In numerical analysis and optimization, norm equivalence guarantees that convergence of an iterative algorithm can be checked in whichever norm is computationally convenient (ℓ∞\ell^\infty is cheap, requiring only a max), while theoretical guarantees are often proved in ℓ2\ell^2. The constants in the equivalence bounds (as in Example 2) matter quantitatively for convergence rate estimates, especially in high dimensions where these constants can grow like n\sqrt n or nn. In physics, the Euclidean inner product underlies the notion of work as F⃗⋅d⃗\vec F \cdot \vec d and generalizes to the Minkowski inner product of special relativity (a signature-indefinite cousin where positive-definiteness is deliberately dropped). In cryptography, lattice-based cryptographic schemes (e.g., those underlying post-quantum cryptography standards) rely on the geometry of Rn\mathbb{R}^n under the Euclidean norm to define hard computational problems like the Shortest Vector Problem. In pure mathematics, this article's machinery is the direct prerequisite for defining Open and Closed Sets in R^n, for making sense of Limits and Continuity in Several Variables, and eventually for the entire theory of differentiability of multivariable functions, where the norm is what makes "h→0h \to 0" meaningful for vector-valued increments.

:::quiz[In the counterexample on C[0,1]C[0,1], why does no constant CC satisfy ∥f∥∞≤C∥f∥1\|f\|_\infty \le C\|f\|_1 for all ff?]

  • Because C[0,1]C[0,1] is not a vector space
  • Because ∥⋅∥1\|\cdot\|_1 is not actually a norm on C[0,1]C[0,1]
  • Because the spike functions fnf_n have fixed sup-norm 11 but ℓ1\ell^1-norm shrinking like 1/(2n)1/(2n) ✓
  • Because continuous functions cannot be integrated

Ask the tutor: How does the Euclidean norm's role here extend to defining open balls and continuity in Limits and Continuity in Several Variables?

DefinitionEuclidean space, inner product, and norm

Let n∈Nn \in \mathbb{N}. Euclidean space Rn\mathbb{R}^n is the real vector space of nn-tuples x=(x1,…,xn)x = (x_1, \dots, x_n) with coordinatewise addition and scalar multiplication. The standard (Euclidean) inner product on Rn\mathbb{R}^n is the function ⟨⋅,⋅⟩:Rn×Rn→R\langle \cdot,\cdot\rangle : \mathbb{R}^n \times \mathbb{R}^n \to \mathbb{R} defined by

⟨x,y⟩=∑i=1nxiyi.\langle x, y \rangle = \sum_{i=1}^n x_i y_i.

It satisfies, for all x,y,z∈Rnx,y,z \in \mathbb{R}^n and α∈R\alpha \in \mathbb{R}: 1. Symmetry: ⟨x,y⟩=⟨y,x⟩\langle x,y\rangle = \langle y,x\rangle. 2. Bilinearity: ⟨αx+z,y⟩=α⟨x,y⟩+⟨z,y⟩\langle \alpha x + z, y\rangle = \alpha\langle x,y\rangle + \langle z,y\rangle (and similarly in the second slot, by symmetry). 3. Positive-definiteness: ⟨x,x⟩≥0\langle x,x\rangle \ge 0, with equality if and only if x=0x = 0. The Euclidean norm of x∈Rnx \in \mathbb{R}^n is

∥x∥=⟨x,x⟩=(∑i=1nxi2)1/2.\|x\| = \sqrt{\langle x, x\rangle} = \left(\sum_{i=1}^n x_i^2\right)^{1/2}.

More generally, a norm on a real vector space VV is a function ∥⋅∥:V→R\|\cdot\| : V \to \mathbb{R} satisfying, for all x,y∈Vx,y \in V and α∈R\alpha \in \mathbb{R}:

  • (N1) ∥x∥≥0\|x\| \ge 0, and ∥x∥=0  ⟺  x=0\|x\| = 0 \iff x = 0 (positivity);
  • (N2) ∥αx∥=∣α∣ ∥x∥\|\alpha x\| = |\alpha|\,\|x\| (absolute homogeneity);
  • (N3) ∥x+y∥≤∥x∥+∥y∥\|x + y\| \le \|x\| + \|y\| (triangle inequality).

A vector space equipped with a norm is called a normed space. Given a norm, one defines a distance function (metric) d(x,y)=∥x−y∥d(x,y) = \|x - y\|.

Intuition

Think of ⟨x,y⟩\langle x, y \rangle as an algebraic stand-in for "how much xx and yy point in the same direction, scaled by their lengths." When n=2n=2 or n=3n=3, this matches the familiar formula ⟨x,y⟩=∥x∥∥y∥cos⁡θ\langle x,y\rangle = \|x\|\|y\|\cos\theta from geometry, but the definition ∑xiyi\sum x_i y_i makes sense in any dimension without any prior notion of angle — indeed, we will define angle in terms of the inner product once Cauchy–Schwarz guarantees the relevant ratio lies in [−1,1][-1,1]. The norm ∥x∥\|x\| generalizes the absolute value ∣x∣|x| from R\mathbb{R}. Just as ∣a−b∣|a - b| measures distance on the line, ∥x−y∥\|x - y\| measures distance in Rn\mathbb{R}^n. The triangle inequality is the geometric fact that going from xx to zz directly is never longer than going from xx to yy and then yy to zz — a statement that looks visually obvious in the plane but requires real algebraic work to prove for arbitrary vectors in Rn\mathbb{R}^n. The subtlety this article resolves is that ∥⋅∥2\|\cdot\|_2 (Euclidean) is not the only reasonable way to measure size. The taxicab norm ∥x∥1=∑∣xi∣\|x\|_1 = \sum |x_i| and the max norm ∥x∥∞=max⁡i∣xi∣\|x\|_\infty = \max_i |x_i| are equally legitimate norms. They give different numerical distances, but — crucially — they agree on which sets are open, which sequences converge, and which functions are continuous. This is the content of the equivalence-of-norms theorem, illustrated below for n=2n=2: the unit balls of ∥⋅∥1,∥⋅∥2,∥⋅∥∞\|\cdot\|_1, \|\cdot\|_2, \|\cdot\|_\infty are nested inside constant multiples of one another.

Figure
Unit balls of the ℓ1\ell^1, ℓ2\ell^2, ℓ∞\ell^\infty norms in R2\mathbb{R}^2, showing mutual containment
‖x‖∞ ≤ 1 ‖x‖₂ ≤ 1 ‖x‖₁ ≤ 1 −1 1 1 −1 Containments: diamond ⊆ circle ⊆ square ‖x‖∞ ≤ ‖x‖₂ ≤ ‖x‖₁ ‖x‖₁ ≤ √2·‖x‖₂ (n=2) ‖x‖₂ ≤ √2·‖x‖∞ (n=2)
TheoremCauchy–Schwarz Inequality

Let x,y∈Rnx, y \in \mathbb{R}^n. Then

∣⟨x,y⟩∣≤∥x∥ ∥y∥.|\langle x, y \rangle| \le \|x\| \, \|y\|.

Equality holds if and only if xx and yy are linearly dependent (i.e. one is a scalar multiple of the other, including the trivial case where either is 00).

Proof

Technique: analysis of a quadratic polynomial's discriminant. If y=0y = 0, then ⟨x,y⟩=0\langle x, y\rangle = 0 and ∥y∥=0\|y\| = 0, so both sides of the inequality equal 00; equality holds, and indeed y=0y = 0 is (trivially) a scalar multiple of xx (with scalar 00, or vice versa). So assume y≠0y \ne 0. Define f:R→Rf : \mathbb{R} \to \mathbb{R} by

f(t)=∥x−ty∥2=⟨x−ty,x−ty⟩.f(t) = \|x - t y\|^2 = \langle x - ty, x - ty\rangle.

By bilinearity and symmetry of the inner product,

f(t)=⟨x,x⟩−2t⟨x,y⟩+t2⟨y,y⟩=∥y∥2t2−2⟨x,y⟩t+∥x∥2.f(t) = \langle x,x\rangle - 2t\langle x,y\rangle + t^2 \langle y,y\rangle = \|y\|^2 t^2 - 2\langle x,y\rangle t + \|x\|^2.

By positive-definiteness (property 3 in the definition of the inner product), f(t)=∥x−ty∥2≥0f(t) = \|x - ty\|^2 \ge 0 for every t∈Rt \in \mathbb{R}, since it is the squared norm of a vector. Thus ff is a real quadratic polynomial in tt with leading coefficient ∥y∥2>0\|y\|^2 > 0 (nonzero since y≠0y \ne 0, using positive-definiteness again) that is nonnegative for all real tt. A quadratic at2+bt+cat^2+bt+c with a>0a>0 is nonnegative for all tt if and only if its discriminant satisfies b2−4ac≤0b^2 - 4ac \le 0. (Proof of this elementary fact: complete the square, at2+bt+c=a(t+b2a)2+4ac−b24aat^2+bt+c = a\left(t+\frac{b}{2a}\right)^2 + \frac{4ac-b^2}{4a}; this is ≥0\ge 0 for all tt iff the constant term 4ac−b24a≥0\frac{4ac-b^2}{4a} \ge 0, i.e. iff b2≤4acb^2 \le 4ac.) Applying this with a=∥y∥2a = \|y\|^2, b=−2⟨x,y⟩b = -2\langle x,y\rangle, c=∥x∥2c = \|x\|^2, the discriminant condition reads

4⟨x,y⟩2−4∥y∥2∥x∥2≤0,4\langle x,y\rangle^2 - 4\|y\|^2\|x\|^2 \le 0,

which rearranges to

⟨x,y⟩2≤∥x∥2∥y∥2.\langle x, y\rangle^2 \le \|x\|^2 \|y\|^2.

Taking square roots of both nonnegative quantities gives ∣⟨x,y⟩∣≤∥x∥∥y∥|\langle x,y\rangle| \le \|x\|\|y\|, as required.

Equality case. Suppose y≠0y \ne 0 and equality holds: ∣⟨x,y⟩∣=∥x∥∥y∥|\langle x,y\rangle| = \|x\|\|y\|. Then the discriminant of ff is exactly 00, so ff has a (unique) real root t0t_0, meaning f(t0)=∥x−t0y∥2=0f(t_0) = \|x - t_0 y\|^2 = 0. By positive-definiteness, x−t0y=0x - t_0 y = 0, i.e. x=t0yx = t_0 y, so x,yx,y are linearly dependent. Conversely, if x=λyx = \lambda y for some scalar λ\lambda (with y≠0y\ne0), then ⟨x,y⟩=λ⟨y,y⟩=λ∥y∥2\langle x,y\rangle = \lambda \langle y,y\rangle = \lambda \|y\|^2 and ∥x∥=∣λ∣∥y∥\|x\| = |\lambda| \|y\|, so

∣⟨x,y⟩∣=∣λ∣∥y∥2=(∣λ∣∥y∥)∥y∥=∥x∥∥y∥,|\langle x,y\rangle| = |\lambda| \|y\|^2 = (|\lambda|\|y\|)\|y\| = \|x\|\|y\|,

giving equality. This covers all linearly dependent pairs (the case y=0y=0 was handled at the start), completing the proof. ■\blacksquare

∎
TheoremProperties of the Euclidean norm

The function ∥⋅∥:Rn→R\|\cdot\| : \mathbb{R}^n \to \mathbb{R}, ∥x∥=⟨x,x⟩\|x\| = \sqrt{\langle x,x\rangle}, is a norm on Rn\mathbb{R}^n: it satisfies (N1) positivity, (N2) homogeneity, and (N3) the triangle inequality.

Proof

Technique: direct verification, using Cauchy–Schwarz for (N3).

(N1) Positivity. Since ⟨x,x⟩≥0\langle x,x\rangle \ge 0 (positive-definiteness), ∥x∥=⟨x,x⟩\|x\| = \sqrt{\langle x,x\rangle} is a well-defined nonnegative real number. If x=0x = 0, then ⟨x,x⟩=∑i02=0\langle x,x\rangle = \sum_i 0^2 = 0, so ∥x∥=0\|x\| = 0. Conversely if ∥x∥=0\|x\| = 0 then ⟨x,x⟩=0\langle x,x\rangle = 0, and positive-definiteness of the inner product forces x=0x = 0.

(N2) Homogeneity. For α∈R\alpha \in \mathbb{R} and x∈Rnx \in \mathbb{R}^n, bilinearity gives

∥αx∥2=⟨αx,αx⟩=α2⟨x,x⟩=α2∥x∥2.\|\alpha x\|^2 = \langle \alpha x, \alpha x\rangle = \alpha^2 \langle x, x\rangle = \alpha^2 \|x\|^2.

Taking (nonnegative) square roots, ∥αx∥=α2 ∥x∥=∣α∣ ∥x∥\|\alpha x\| = \sqrt{\alpha^2}\,\|x\| = |\alpha|\,\|x\|.

(N3) Triangle inequality. For x,y∈Rnx, y \in \mathbb{R}^n, expand using bilinearity:

∥x+y∥2=⟨x+y,x+y⟩=⟨x,x⟩+2⟨x,y⟩+⟨y,y⟩=∥x∥2+2⟨x,y⟩+∥y∥2.\|x+y\|^2 = \langle x+y, x+y\rangle = \langle x,x\rangle + 2\langle x,y\rangle + \langle y,y\rangle = \|x\|^2 + 2\langle x,y\rangle + \|y\|^2.

By the Cauchy–Schwarz Inequality, ⟨x,y⟩≤∣⟨x,y⟩∣≤∥x∥∥y∥\langle x,y\rangle \le |\langle x,y\rangle| \le \|x\|\|y\|. Substituting,

∥x+y∥2≤∥x∥2+2∥x∥∥y∥+∥y∥2=(∥x∥+∥y∥)2.\|x+y\|^2 \le \|x\|^2 + 2\|x\|\|y\| + \|y\|^2 = (\|x\| + \|y\|)^2.

Since both ∥x+y∥\|x+y\| and ∥x∥+∥y∥\|x\|+\|y\| are nonnegative real numbers, and the square of the former is at most the square of the latter, we may take square roots (the function t↦tt \mapsto \sqrt{t} is increasing on [0,∞)[0,\infty)) to conclude

∥x+y∥≤∥x∥+∥y∥.■\|x + y\| \le \|x\| + \|y\|. \qquad\blacksquare
∎
CorollaryReverse triangle inequality

For all x,y∈Rnx, y \in \mathbb{R}^n,

∣ ∥x∥−∥y∥ ∣≤∥x−y∥.\big|\,\|x\| - \|y\|\,\big| \le \|x - y\|.
Proof

By the triangle inequality applied to x=(x−y)+yx = (x-y) + y:

∥x∥≤∥x−y∥+∥y∥  ⟹  ∥x∥−∥y∥≤∥x−y∥.\|x\| \le \|x - y\| + \|y\| \implies \|x\| - \|y\| \le \|x-y\|.

Symmetrically, applying the triangle inequality to y=(y−x)+xy = (y-x)+x and using ∥y−x∥=∣−1∣ ∥x−y∥=∥x−y∥\|y-x\| = |-1|\,\|x-y\| = \|x-y\| (by homogeneity, N2):

∥y∥−∥x∥≤∥y−x∥=∥x−y∥.\|y\| - \|x\| \le \|y - x\| = \|x - y\|.

Combining the two inequalities, −∥x−y∥≤∥x∥−∥y∥≤∥x−y∥-\|x-y\| \le \|x\|-\|y\| \le \|x-y\|, which is exactly ∣ ∥x∥−∥y∥ ∣≤∥x−y∥|\,\|x\|-\|y\|\,| \le \|x-y\|. ■\blacksquare

∎
TheoremEquivalence of Norms in Finite Dimensions

Let VV be a finite-dimensional real vector space, dim⁡V=n\dim V = n, and let ∥⋅∥a\|\cdot\|_a and ∥⋅∥b\|\cdot\|_b be any two norms on VV. Then there exist constants c,C>0c, C > 0 (depending on the norms and on VV, but not on the vector) such that

c ∥x∥a≤∥x∥b≤C ∥x∥afor all x∈V.c \, \|x\|_a \le \|x\|_b \le C \, \|x\|_a \qquad \text{for all } x \in V.
Proof

Technique: reduce to comparing an arbitrary norm with the Euclidean norm, then use compactness of the unit sphere (Bolzano–Weierstrass / extreme value theorem) together with continuity. Fix a basis e1,…,ene_1, \dots, e_n of VV, giving coordinates: every x∈Vx \in V is uniquely x=∑i=1nxieix = \sum_{i=1}^n x_i e_i. This identifies VV with Rn\mathbb{R}^n as vector spaces, and we let ∥x∥2=(∑xi2)1/2\|x\|_2 = \left(\sum x_i^2\right)^{1/2} denote the Euclidean norm in these coordinates. It suffices to show every norm ∥⋅∥\|\cdot\| on VV is equivalent to ∥⋅∥2\|\cdot\|_2: if ∥⋅∥a\|\cdot\|_a is equivalent to ∥⋅∥2\|\cdot\|_2 (constants c1,C1c_1,C_1) and ∥⋅∥b\|\cdot\|_b is equivalent to ∥⋅∥2\|\cdot\|_2 (constants c2,C2c_2, C_2), then chaining the inequalities

c2C1∥x∥a≤∥x∥b≤C2c1∥x∥a\frac{c_2}{C_1}\|x\|_a \le \|x\|_b \le \frac{C_2}{c_1}\|x\|_a

gives equivalence of ∥⋅∥a\|\cdot\|_a and ∥⋅∥b\|\cdot\|_b directly. (Verify: from c1∥x∥2≤∥x∥a≤C1∥x∥2c_1\|x\|_2 \le \|x\|_a \le C_1\|x\|_2 we get ∥x∥2≥∥x∥a/C1\|x\|_2 \ge \|x\|_a/C_1 and ∥x∥2≤∥x∥a/c1\|x\|_2 \le \|x\|_a/c_1; combine with c2∥x∥2≤∥x∥b≤C2∥x∥2c_2\|x\|_2\le\|x\|_b\le C_2\|x\|_2.)

Step 1: an upper bound ∥x∥≤C∥x∥2\|x\| \le C\|x\|_2. Let ∥⋅∥\|\cdot\| be an arbitrary norm on VV. For x=∑ixieix = \sum_i x_i e_i, the triangle inequality (N3) applied repeatedly and homogeneity (N2) give

∥x∥=∥∑i=1nxiei∥≤∑i=1n∥xiei∥=∑i=1n∣xi∣ ∥ei∥.\|x\| = \left\|\sum_{i=1}^n x_i e_i\right\| \le \sum_{i=1}^n \|x_i e_i\| = \sum_{i=1}^n |x_i|\,\|e_i\|.

By the Cauchy–Schwarz inequality applied to the vectors (∣x1∣,…,∣xn∣)(|x_1|,\dots,|x_n|) and (∥e1∥,…,∥en∥)(\|e_1\|,\dots,\|e_n\|) in Rn\mathbb{R}^n,

∑i=1n∣xi∣∥ei∥≤(∑ixi2)1/2(∑i∥ei∥2)1/2=M∥x∥2,\sum_{i=1}^n |x_i|\|e_i\| \le \left(\sum_i x_i^2\right)^{1/2}\left(\sum_i \|e_i\|^2\right)^{1/2} = M \|x\|_2,

where M:=(∑i=1n∥ei∥2)1/2M := \left(\sum_{i=1}^n \|e_i\|^2\right)^{1/2} is a finite constant depending only on the basis and the norm ∥⋅∥\|\cdot\| (not on xx). Thus ∥x∥≤M∥x∥2\|x\| \le M \|x\|_2 for all x∈Vx \in V, establishing the upper bound with C=MC = M.

Step 2: continuity of ∥⋅∥\|\cdot\| with respect to ∥⋅∥2\|\cdot\|_2. The inequality from Step 1, combined with the reverse-triangle-type estimate for a general norm (proved exactly as the Corollary above, using only (N1)–(N3), which holds for any norm, not just Euclidean),

∣ ∥x∥−∥y∥ ∣≤∥x−y∥≤M∥x−y∥2,\big|\,\|x\| - \|y\|\,\big| \le \|x - y\| \le M\|x-y\|_2,

shows that ∥⋅∥:(V,∥⋅∥2)→R\|\cdot\| : (V, \|\cdot\|_2) \to \mathbb{R} is continuous: given ε>0\varepsilon>0, take δ=ε/M\delta = \varepsilon/M; then ∥x−y∥2<δ\|x-y\|_2 < \delta implies ∣∥x∥−∥y∥∣≤M∥x−y∥2<ε|\|x\|-\|y\|| \le M\|x-y\|_2 < \varepsilon.

Step 3: a positive lower bound via compactness. Consider the Euclidean unit sphere S={x∈V:∥x∥2=1}S = \{x \in V : \|x\|_2 = 1\}. Under the coordinate identification with Rn\mathbb{R}^n, SS corresponds to {(x1,…,xn):∑xi2=1}\{(x_1,\dots,x_n) : \sum x_i^2 = 1\}, which is closed and bounded in Rn\mathbb{R}^n, hence compact by the Heine–Borel theorem (itself resting on completeness of R\mathbb{R}). The function ∥⋅∥:S→R\|\cdot\| : S \to \mathbb{R} is continuous (Step 2, restricted to SS) on the compact set SS, so by the Extreme Value Theorem it attains a minimum value m=∥x0∥m = \|x_0\| for some x0∈Sx_0 \in S. Since x0∈Sx_0 \in S means ∥x0∥2=1≠0\|x_0\|_2 = 1 \ne 0, we have x0≠0x_0 \ne 0, so by positive-definiteness (N1) of the norm ∥⋅∥\|\cdot\|, m=∥x0∥>0m = \|x_0\| > 0. Thus ∥x∥≥m\|x\| \ge m for every x∈Sx \in S, i.e. for every xx with ∥x∥2=1\|x\|_2 = 1.

Step 4: scale to all of VV. Let x∈Vx \in V, x≠0x \ne 0. Then x/∥x∥2∈Sx/\|x\|_2 \in S (since ∥x/∥x∥2∥2=∥x∥2/∥x∥2=1\|x/\|x\|_2\|_2 = \|x\|_2/\|x\|_2 = 1 by homogeneity of ∥⋅∥2\|\cdot\|_2), so by Step 3,

∥x∥x∥2∥≥m.\left\|\frac{x}{\|x\|_2}\right\| \ge m.

By homogeneity (N2) of the norm ∥⋅∥\|\cdot\|, the left side equals 1∥x∥2∥x∥\frac{1}{\|x\|_2}\|x\|, so

∥x∥≥m∥x∥2.\|x\| \ge m \|x\|_2.

This also holds trivially (as 0≥00 \ge 0) when x=0x = 0. Combined with Step 1, we obtain

m∥x∥2≤∥x∥≤M∥x∥2for all x∈V,m \|x\|_2 \le \|x\| \le M \|x\|_2 \qquad \text{for all } x \in V,

with m,M>0m, M > 0. This proves ∥⋅∥\|\cdot\| is equivalent to ∥⋅∥2\|\cdot\|_2, and by the reduction at the start of the proof, any two norms on VV are equivalent to each other. ■\blacksquare

∎

Worked examples

Example 1: Verifying Cauchy–Schwarz and the equality condition numerically. Let x=(3,−1,4)∈R3x = (3, -1, 4) \in \mathbb{R}^3 and y=(1,2,−2)∈R3y = (1, 2, -2) \in \mathbb{R}^3. We compute both sides of Cauchy–Schwarz explicitly and check the inequality is strict (as expected, since x,yx, y are not parallel). First, the inner product:

⟨x,y⟩=(3)(1)+(−1)(2)+(4)(−2)=3−2−8=−7.\langle x, y\rangle = (3)(1) + (-1)(2) + (4)(-2) = 3 - 2 - 8 = -7.

So ∣⟨x,y⟩∣=7|\langle x,y\rangle| = 7. Next the norms:

∥x∥=32+(−1)2+42=9+1+16=26,\|x\| = \sqrt{3^2 + (-1)^2 + 4^2} = \sqrt{9+1+16} = \sqrt{26},
∥y∥=12+22+(−2)2=1+4+4=9=3.\|y\| = \sqrt{1^2+2^2+(-2)^2} = \sqrt{1+4+4} = \sqrt{9} = 3.

Thus ∥x∥∥y∥=326\|x\|\|y\| = 3\sqrt{26}. Numerically 26≈5.0990\sqrt{26} \approx 5.0990, so ∥x∥∥y∥≈15.297\|x\|\|y\| \approx 15.297. Since 7<15.2977 < 15.297, Cauchy–Schwarz holds, strictly. To confirm it must be strict, note xx and yy are not scalar multiples of each other: if x=λyx = \lambda y then comparing first coordinates 3=λ3 = \lambda, but then the second coordinate would require −1=λ⋅2=6-1 = \lambda \cdot 2 = 6, a contradiction. So the equality condition (linear dependence) correctly predicts strict inequality here. We can also verify the triangle inequality numerically for these same vectors. Compute

x+y=(4,1,2),∥x+y∥=16+1+4=21≈4.583.x + y = (4, 1, 2), \qquad \|x+y\| = \sqrt{16+1+4} = \sqrt{21} \approx 4.583.

And ∥x∥+∥y∥=26+3≈5.099+3=8.099\|x\| + \|y\| = \sqrt{26} + 3 \approx 5.099 + 3 = 8.099. Indeed 4.583≤8.0994.583 \le 8.099, consistent with (N3), with plenty of room since x+yx+y involves substantial cancellation in direction.

Example 2: Explicit equivalence constants between ∥⋅∥1\|\cdot\|_1, ∥⋅∥2\|\cdot\|_2, ∥⋅∥∞\|\cdot\|_\infty on R3\mathbb{R}^3, and testing them on a specific vector. We derive concrete constants (not just existence) for n=3n=3, then verify on x=(5,−3,0)x = (5, -3, 0). Claim: for all x∈Rnx \in \mathbb{R}^n,

∥x∥∞≤∥x∥2≤n ∥x∥∞,∥x∥∞≤∥x∥1≤n∥x∥∞,∥x∥2≤∥x∥1≤n∥x∥2.\|x\|_\infty \le \|x\|_2 \le \sqrt{n}\,\|x\|_\infty, \qquad \|x\|_\infty \le \|x\|_1 \le n\|x\|_\infty, \qquad \|x\|_2 \le \|x\|_1 \le \sqrt n \|x\|_2.

Proof sketch of these bounds: Let M=∥x∥∞=max⁡i∣xi∣M = \|x\|_\infty = \max_i |x_i|.

  • ∥x∥22=∑xi2≤∑M2=nM2\|x\|_2^2 = \sum x_i^2 \le \sum M^2 = nM^2, so ∥x∥2≤nM\|x\|_2 \le \sqrt n M. Also M2≤∑xi2M^2 \le \sum x_i^2 trivially (one term of the sum), so M≤∥x∥2M \le \|x\|_2. This gives the first pair.
  • ∥x∥1=∑∣xi∣≤∑M=nM\|x\|_1 = \sum|x_i| \le \sum M = nM, and M≤∑∣xi∣=∥x∥1M \le \sum |x_i| = \|x\|_1 trivially. This gives the second pair.
  • ∥x∥12=(∑∣xi∣)2≥∑xi2=∥x∥22\|x\|_1^2 = \left(\sum |x_i|\right)^2 \ge \sum x_i^2 = \|x\|_2^2 (expanding the square gives extra nonnegative cross terms), so ∥x∥2≤∥x∥1\|x\|_2 \le \|x\|_1. And by Cauchy–Schwarz applied to (∣x1∣,…,∣xn∣)(|x_1|,\ldots,|x_n|) against (1,…,1)(1,\ldots,1): ∥x∥1=∑∣xi∣⋅1≤∥x∥2⋅n\|x\|_1 = \sum |x_i|\cdot 1 \le \|x\|_2 \cdot \sqrt n.

Now test on x=(5,−3,0)∈R3x = (5,-3,0) \in \mathbb{R}^3 (so n=3n=3):

∥x∥∞=max⁡(5,3,0)=5,∥x∥1=5+3+0=8,∥x∥2=25+9+0=34≈5.831.\|x\|_\infty = \max(5,3,0) = 5, \qquad \|x\|_1 = 5+3+0 = 8, \qquad \|x\|_2 = \sqrt{25+9+0} = \sqrt{34} \approx 5.831.

Check each predicted inequality:

  • ∥x∥∞≤∥x∥2≤3∥x∥∞\|x\|_\infty \le \|x\|_2 \le \sqrt3 \|x\|_\infty: is 5≤5.831≤1.732⋅5=8.6605 \le 5.831 \le 1.732\cdot 5 = 8.660? Yes: 5≤5.831≤8.6605 \le 5.831 \le 8.660. ✓.
  • ∥x∥∞≤∥x∥1≤3∥x∥∞\|x\|_\infty \le \|x\|_1 \le 3\|x\|_\infty: is 5≤8≤155 \le 8 \le 15? Yes. ✓.
  • ∥x∥2≤∥x∥1≤3∥x∥2\|x\|_2 \le \|x\|_1 \le \sqrt3 \|x\|_2: is 5.831≤8≤1.732×5.831≈10.0985.831 \le 8 \le 1.732 \times 5.831 \approx 10.098? Yes: 5.831≤8≤10.0985.831 \le 8 \le 10.098. ✓.

All three chains of inequalities hold on this asymmetric example, confirming the equivalence-of-norms theorem concretely, with sharp constants (n\sqrt n and nn) matching the general nn-dependent bounds derived above — illustrating that while the constants in norm equivalence genuinely depend on nn (they degrade as dimension grows), equivalence itself holds for every fixed finite nn.

CounterexampleEquivalence of norms fails in infinite dimensions

The equivalence-of-norms theorem crucially used finite-dimensionality in Step 3, where compactness of the unit sphere SS (via Heine–Borel) was invoked. Drop this hypothesis and equivalence can fail. Let V=C[0,1]V = C[0,1], the vector space of continuous real-valued functions on [0,1][0,1], an infinite-dimensional real vector space. Define two norms:

∥f∥∞=sup⁡t∈[0,1]∣f(t)∣,∥f∥1=∫01∣f(t)∣ dt.\|f\|_\infty = \sup_{t \in [0,1]} |f(t)|, \qquad \|f\|_1 = \int_0^1 |f(t)|\,dt.

Both are genuine norms on C[0,1]C[0,1] (positivity, homogeneity, and the triangle inequality are routine to check using properties of sup⁡\sup and of the integral). One direction of equivalence does hold: ∥f∥1=∫01∣f(t)∣ dt≤∫01∥f∥∞ dt=∥f∥∞\|f\|_1 = \int_0^1 |f(t)|\,dt \le \int_0^1 \|f\|_\infty \, dt = \|f\|_\infty, so ∥f∥1≤∥f∥∞\|f\|_1 \le \|f\|_\infty always. The question is whether a reverse bound ∥f∥∞≤C∥f∥1\|f\|_\infty \le C\|f\|_1 can hold for some constant CC independent of ff. Consider the family of "spike" functions fn(t)=max⁡(0,1−nt)f_n(t) = \max(0, 1 - n t) for t∈[0,1]t \in [0,1], n=1,2,3,…n = 1, 2, 3, \dots Each fnf_n is continuous, piecewise linear, equal to 11 at t=0t=0, decreasing linearly to 00 at t=1/nt = 1/n, and identically 00 for t≥1/nt \ge 1/n. Compute both norms:

∥fn∥∞=fn(0)=1for every n,\|f_n\|_\infty = f_n(0) = 1 \quad \text{for every } n,
∥fn∥1=∫01/n(1−nt) dt=[t−nt22]01/n=1n−n2⋅1n2=1n−12n=12n.\|f_n\|_1 = \int_0^{1/n} (1-nt)\,dt = \left[t - \frac{nt^2}{2}\right]_0^{1/n} = \frac1n - \frac{n}{2}\cdot\frac{1}{n^2} = \frac1n - \frac1{2n} = \frac{1}{2n}.

If a constant CC existed with ∥f∥∞≤C∥f∥1\|f\|_\infty \le C \|f\|_1 for all f∈C[0,1]f \in C[0,1], then applying this to fnf_n would give

1=∥fn∥∞≤C∥fn∥1=C2nfor every n,1 = \|f_n\|_\infty \le C \|f_n\|_1 = \frac{C}{2n} \quad \text{for every } n,

i.e. C≥2nC \ge 2n for every n∈Nn \in \mathbb{N}. No finite constant CC satisfies this for all nn, so no such CC exists: the norms ∥⋅∥∞\|\cdot\|_\infty and ∥⋅∥1\|\cdot\|_1 are not equivalent on C[0,1]C[0,1]. This shows the finite-dimensionality hypothesis in the theorem is not a technical convenience but is genuinely necessary: the compactness argument (Step 3 of the proof) has no analogue when the unit sphere is not compact, which is always the case in infinite-dimensional normed spaces (a fact known as Riesz's lemma / theorem on non-compactness of the unit ball in infinite dimensions).

Scope, edge cases & common traps

SituationWhat holdsWhat can fail / must be checked
Rn\mathbb{R}^n, nn finite, any two normsEquivalent, constants existConstants c,Cc, C typically depend on nn and blow up as n→∞n \to \infty
Infinite-dimensional normed spacesNothing guaranteedNorms can fail to be equivalent (see counterexample)
Cauchy–Schwarz equality caseIff x,yx,y linearly dependentDon't forget the trivial case x=0x=0 or y=0y=0 always gives equality
Triangle inequalityAlways holds for any norm by definitionA candidate "norm" that violates (N3) is not a norm at all, even if it satisfies (N1)-(N2)
General inner product spaces (not just Rn\mathbb{R}^n)Cauchy–Schwarz proof (quadratic-in-tt trick) generalizes verbatimRequires only symmetry, bilinearity, positive-definiteness — works over any real inner product space

The mathematician behind it

Origins · 1 minAugustin-Louis Cauchy and how this was discovered
AC
Augustin-Louis Cauchy
1789–1857 · early 19th-century French analysis

Cauchy was one of the founders of rigorous mathematical analysis, responsible for precise definitions of limits, continuity, and convergence that replaced the looser 18th-century treatment of calculus. His 1821 textbook Cours d'Analyse set a new standard for rigor in French mathematics education. Cauchy discovered the discrete inequality that bears his name (for sums, in the context of algebra) in work related to error bounds and inequalities among real numbers; Hermann Amandus Schwarz later extended the inequality to integrals in his study of surface area minimization problems, and Viktor Bunyakovsky independently proved the integral version even earlier than Schwarz — which is why the inequality is sometimes called Cauchy–Bunyakovsky–Schwarz in full.

History
- 1821 — Cauchy publishes Cours d'Analyse, proving the discrete inequality (∑aibi)2≤(∑ai2)(∑bi2)\left(\sum a_ib_i\right)^2 \le \left(\sum a_i^2\right)\left(\sum b_i^2\right) for finite sums.
- 1859 — Viktor Bunyakovsky publishes the integral analogue of the inequality.
- 1885 — Hermann Schwarz independently rediscovers and applies the integral version in the context of minimal surfaces.
- Late 19th century — David Hilbert and others develop abstract inner product spaces, recognizing Cauchy–Schwarz as a general feature of any positive-definite bilinear form.
- Early 20th century — Stefan Banach and collaborators formalize general normed spaces, making explicit that finite-dimensionality is exactly the hypothesis needed for norm equivalence, contrasting sharply with the infinite-dimensional theory Banach himself pioneered.
Why does the inner product need to be positive-definite, and what breaks without it?

Positive-definiteness (⟨x,x⟩≥0\langle x,x\rangle \ge 0, with equality only at x=0x=0) is what guarantees ∥x∥=⟨x,x⟩\|x\| = \sqrt{\langle x,x\rangle} is a well-defined real number (the quantity under the square root is never negative) and that ∥x∥=0\|x\|=0 forces x=0x=0, which is exactly requirement (N1) for a norm. It is also essential in the Cauchy–Schwarz proof: the argument relies on f(t)=∥x−ty∥2≥0f(t) = \|x-ty\|^2 \ge 0 for all real tt, which is precisely the statement that squared norms are nonnegative. Without positive-definiteness — for instance, with an indefinite bilinear form like the Minkowski form in relativity — vectors can have "zero length" without being zero, or even negative squared length, and the entire triangle-inequality argument collapses; indeed, the Minkowski "norm" famously violates the triangle inequality in its usual form.

Is the Euclidean norm the only norm that comes from an inner product?

No, but it is a distinguished one. Any inner product ⟨⋅,⋅⟩\langle\cdot,\cdot\rangle on a real vector space induces a norm via ∥x∥=⟨x,x⟩\|x\| = \sqrt{\langle x,x\rangle}, and the proof given here (Cauchy–Schwarz, then the norm axioms) works verbatim for any symmetric, bilinear, positive-definite form, not just the standard dot product. Different choices of inner product (e.g., a weighted inner product ⟨x,y⟩W=∑wixiyi\langle x,y\rangle_W = \sum w_i x_iy_i with wi>0w_i>0) give different induced norms, all "Euclidean-like" in that they satisfy the parallelogram law ∥x+y∥2+∥x−y∥2=2∥x∥2+2∥y∥2\|x+y\|^2+\|x-y\|^2 = 2\|x\|^2+2\|y\|^2. Norms like ℓ1\ell^1 or ℓ∞\ell^\infty do not satisfy the parallelogram law and hence are not induced by any inner product — this is a clean test (the Jordan–von Neumann theorem) for whether a given norm comes from an inner product.

Why does the equivalence-of-norms proof fail in infinite dimensions, precisely?

The proof's crux is Step 3: continuity of an arbitrary norm on the Euclidean unit sphere SS, combined with compactness of SS, guarantees a positive minimum value m>0m>0. In infinite dimensions, the closed and bounded unit sphere is never compact (a theorem due to Riesz), so the extreme value theorem cannot be invoked, and a continuous function on SS need not attain a positive infimum — it can have infimum 00 without attaining it, exactly as the counterexample's spike functions fnf_n demonstrate (∥fn∥∞=1\|f_n\|_\infty=1 fixed while ∥fn∥1→0\|f_n\|_1\to0). There is no way to patch the proof; the conclusion genuinely becomes false, as shown concretely with C[0,1]C[0,1].

How does Cauchy–Schwarz let us define the angle between two vectors in Rn\mathbb{R}^n?

Cauchy–Schwarz guarantees −1≤⟨x,y⟩∥x∥∥y∥≤1-1 \le \dfrac{\langle x,y\rangle}{\|x\|\|y\|} \le 1 whenever x,y≠0x,y \ne 0, since ∣⟨x,y⟩∣≤∥x∥∥y∥|\langle x,y\rangle| \le \|x\|\|y\| divided through by the positive quantity ∥x∥∥y∥\|x\|\|y\| gives exactly this bound. Because the cosine function is a bijection from [0,π][0,\pi] onto [−1,1][-1,1], there is a unique θ∈[0,π]\theta \in [0,\pi] with cos⁡θ=⟨x,y⟩/(∥x∥∥y∥)\cos\theta = \langle x,y\rangle/(\|x\|\|y\|); this θ\theta is defined to be the angle between xx and yy. Without Cauchy–Schwarz, this ratio could exceed 11 in absolute value, in which case no such angle would exist — so the inequality is not just a technical tool but the logical precondition for the geometric concept of angle to make sense in Rn\mathbb{R}^n for n≥2n \ge 2.

Do all norms on Rn\mathbb{R}^n give the same open sets, and why does this matter?

Yes. Because all norms on Rn\mathbb{R}^n (finite nn) are equivalent, a set is open with respect to ∥⋅∥a\|\cdot\|_a if and only if it is open with respect to ∥⋅∥b\|\cdot\|_b, for any two norms — a ball in one norm contains and is contained in balls of the other norm (with rescaled radii), by the defining inequalities c∥x∥a≤∥x∥b≤C∥x∥ac\|x\|_a \le \|x\|_b \le C\|x\|_a. This matters because it means concepts like open set, closed set, compactness, continuity, and convergence of sequences in Rn\mathbb{R}^n can be stated without ever specifying which norm is used — they are genuinely properties of the vector space Rn\mathbb{R}^n itself, not artifacts of a particular distance formula. This is exactly the fact that lets analysts move freely between ℓ1\ell^1, ℓ2\ell^2, ℓ∞\ell^\infty when convenient for a computation.

Why is the Cauchy–Schwarz proof using a quadratic in tt not circular, given that it uses the norm we're trying to establish properties of?

The proof uses only the three inner product axioms (symmetry, bilinearity, positive-definiteness) directly — it does not presuppose the triangle inequality or any norm property being proved. The quantity ∥x−ty∥2\|x-ty\|^2 is shorthand for ⟨x−ty,x−ty⟩\langle x-ty, x-ty\rangle, expanded purely via bilinearity into a real quadratic in tt; the only norm-like fact used is ⟨v,v⟩≥0\langle v,v\rangle \ge 0 for any vector vv, which is precisely the positive-definiteness axiom, assumed from the start, not derived. So the logical order is: inner product axioms ⇒\Rightarrow Cauchy–Schwarz ⇒\Rightarrow triangle inequality for the induced norm — no circularity, since the triangle inequality (the thing being established) is never invoked in its own proof.

What's the practical difference between choosing ℓ1\ell^1, ℓ2\ell^2, or ℓ∞\ell^\infty norms in an application, given that they're all equivalent?

Equivalence guarantees qualitative outcomes agree (a sequence converges in one iff it converges in all, a set is bounded in one iff bounded in all), but numeric values, geometric shapes of balls, and optimization behavior differ substantially. In statistics, ℓ1\ell^1-regularized regression (LASSO) tends to produce sparse solutions (many coefficients exactly zero) because the ℓ1\ell^1 ball has corners aligned with the coordinate axes, while ℓ2\ell^2-regularization (ridge) shrinks coefficients smoothly without inducing exact sparsity, because the ℓ2\ell^2 ball is smooth with no corners. In computational contexts, ℓ∞\ell^\infty is often cheapest to evaluate (just a max). So while topology doesn't care which norm you pick, algorithms, geometry, and statistics very much do — equivalence is a statement about limits and convergence, not about numerical or geometric identity.

Where this leads