The Triangle Inequality and Norm Properties

How do you measure the "size" of a vector? Why should that measurement behave like distance in ordinary space? In the plane, the length of a vector satisfies a familiar fact. The length of one side of a triangle never exceeds the sum of the other two. This article asks whether this geometric fact survives abstraction. This happens when vectors become polynomials, functions, or matrices. In these cases, "length" is defined abstractly through an inner product. This matters because the triangle inequality is the load-bearing wall of analysis on vector spaces. Without it, notions like convergence, continuity, and distance collapse. With it, an Inner Product Space becomes a genuine Normed Vector Space. All the machinery of metric spaces then becomes available. The triangle inequality is not an independent axiom we impose by hand. Remarkably, it is a consequence of the Cauchy-Schwarz Inequality. That inequality is itself a consequence of the inner product axioms alone. The big picture is a chain of implications. An inner product produces a norm via ∥v∥=⟨v,v⟩\|v\| = \sqrt{\langle v, v\rangle}. Cauchy–Schwarz controls the cross terms that appear when expanding ∥u+v∥2\|u+v\|^2. This control forces the triangle inequality. Along the way we uncover the parallelogram law. It is an identity so rigid that it exactly characterizes which norms come from an inner product at all. This is a genuine converse, not a mere analogy. This article delivers a complete, self-contained derivation of the triangle inequality from Cauchy–Schwarz. It also gives a full proof of the parallelogram law. It presents the (perhaps surprising) theorem that the parallelogram law is both necessary and sufficient. That condition determines whether a norm arises from an inner product. It closes by setting up the coordinate geometry — orthonormal bases — that this norm theory makes possible.

Key takeaways
  • The norm induced by an inner product is ∥v∥=⟨v,v⟩\|v\| = \sqrt{\langle v,v\rangle}; it is always well-defined since ⟨v,v⟩≥0\langle v,v \rangle \geq 0.
  • The Cauchy-Schwarz Inequality ∣⟨u,v⟩∣≤∥u∥∥v∥|\langle u,v\rangle| \le \|u\|\|v\| is the single algebraic fact that makes the triangle inequality work.
  • The triangle inequality ∥u+v∥≤∥u∥+∥v∥\|u+v\| \le \|u\| + \|v\| follows by expanding ∥u+v∥2\|u+v\|^2 and bounding the cross term.
  • Equality in the triangle inequality holds exactly when u,vu,v point in the same direction (one is a nonnegative multiple of the other).
  • The parallelogram law ∥u+v∥2+∥u−v∥2=2∥u∥2+2∥v∥2\|u+v\|^2+\|u-v\|^2 = 2\|u\|^2+2\|v\|^2 holds automatically for inner-product norms.
  • Key idea: a norm on a vector space comes from some inner product if and only if it satisfies the parallelogram law.

Where this lives today

The triangle inequality underlies the definition of a metric d(u,v)=∥u−v∥d(u,v) = \|u-v\| on any inner product space, which is the starting point for all of analysis: convergence of sequences of functions, continuity of linear operators, and completeness (leading to Hilbert spaces) all rest on it. In statistics and machine learning, the triangle inequality justifies distance-based algorithms (k-nearest neighbors, clustering) whenever the underlying similarity measure is induced by an inner product, such as cosine similarity or kernel methods in support vector machines. In physics, the parallelogram law appears directly in the analysis of wave superposition and in quantum mechanics, where state vectors live in a Hilbert space and the norm-squared gives probability; the Jordan–von Neumann theorem explains why quantum state spaces are Hilbert spaces rather than general Banach spaces. In numerical linear algebra and engineering, norm inequalities of this type control error propagation and stability estimates for iterative algorithms. The characterization theorem (part C) is also a template for a broader pattern in mathematics: identifying a simple, checkable algebraic identity that exactly detects a richer geometric structure, a theme that recurs when building Orthonormal Bases and Coordinates, where the inner product's extra structure (beyond a mere norm) becomes indispensable for defining orthogonal projections and coordinate expansions.

Ask the tutor: Why does the equality condition in the triangle inequality require a nonnegative scalar multiple rather than just linear dependence?

DefinitionNorm induced by an inner product

Let VV be a vector space over F\mathbb{F} (where F=R\mathbb{F} = \mathbb{R} or F=C\mathbb{F} = \mathbb{C}) equipped with an inner product ⟨⋅,⋅⟩:V×V→F\langle \cdot,\cdot\rangle : V \times V \to \mathbb{F}, satisfying for all u,v,w∈Vu,v,w \in V and λ∈F\lambda \in \mathbb{F}: 1. Conjugate symmetry: ⟨u,v⟩=⟨v,u⟩‾\langle u,v\rangle = \overline{\langle v,u\rangle}. 2. Linearity in the first slot: ⟨λu+w,v⟩=λ⟨u,v⟩+⟨w,v⟩\langle \lambda u + w, v\rangle = \lambda\langle u,v\rangle + \langle w,v\rangle. 3. Positive-definiteness: ⟨v,v⟩≥0\langle v,v\rangle \geq 0 for all vv, with ⟨v,v⟩=0\langle v,v\rangle = 0 if and only if v=0v = 0. The norm induced by ⟨⋅,⋅⟩\langle\cdot,\cdot\rangle is the function ∥⋅∥:V→R≥0\|\cdot\| : V \to \mathbb{R}_{\geq 0} defined by

∥v∥:=⟨v,v⟩.\|v\| := \sqrt{\langle v,v\rangle}.

More generally, an abstract norm on a vector space VV is any function ∥⋅∥:V→R≥0\|\cdot\|: V \to \mathbb{R}_{\geq 0} satisfying: (N1) ∥v∥=0  ⟺  v=0\|v\| = 0 \iff v = 0 (positive definiteness), (N2) ∥λv∥=∣λ∣ ∥v∥\|\lambda v\| = |\lambda|\,\|v\| for all scalars λ\lambda (homogeneity), (N3) ∥u+v∥≤∥u∥+∥v∥\|u+v\| \le \|u\|+\|v\| for all u,v∈Vu,v\in V (the triangle inequality). A norm satisfying the additional identity

∥u+v∥2+∥u−v∥2=2∥u∥2+2∥v∥2for all u,v∈V\|u+v\|^2 + \|u-v\|^2 = 2\|u\|^2 + 2\|v\|^2 \quad \text{for all } u,v \in V

is said to satisfy the parallelogram law.

Intuition

Picture a triangle in the plane with one vertex at the origin, and sides represented by vectors uu, vv, and u+vu+v. Walking along uu and then along vv gets you to the same endpoint as walking directly along u+vu+v; the "shortcut" u+vu+v can never be longer than the two-leg path, and it is strictly shorter unless uu and vv point in exactly the same direction. This is the geometric content of the triangle inequality, and the entire difficulty in proving it algebraically is controlling the cross term 2Re⁡⟨u,v⟩2\operatorname{Re}\langle u,v\rangle that appears when you expand ∥u+v∥2\|u+v\|^2 using bilinearity. Cauchy–Schwarz is precisely the tool that bounds this cross term. The parallelogram law has its own vivid picture: take a parallelogram with sides uu and vv. Its two diagonals are u+vu+v and u−vu-v. The identity says that the sum of the squares of the diagonal lengths equals the sum of the squares of all four side lengths (2∥u∥2+2∥v∥22\|u\|^2 + 2\|v\|^2 counts each side length twice). This is a rigidity condition — most norms (for instance, "taxicab" distance) do not satisfy it, and this failure is exactly what prevents such norms from being induced by any inner product.

Figure
Triangle inequality: the direct path is never longer than the two-leg path
0 u v u+v ‖u+v‖ ≤ ‖u‖ + ‖v‖
Figure
Parallelogram law: sum of squared diagonals equals sum of squared sides
0 v u u+v u+v u−v ‖u+v‖² + ‖u−v‖² = 2‖u‖² + 2‖v‖²
TheoremTriangle Inequality, Parallelogram Law, and the Inner-Product Norm Characterization

Let VV be an inner product space over F∈{R,C}\mathbb{F} \in \{\mathbb{R},\mathbb{C}\}, with induced norm ∥v∥=⟨v,v⟩\|v\| = \sqrt{\langle v,v\rangle}.

(A) Triangle Inequality. For all u,v∈Vu,v \in V,

∥u+v∥≤∥u∥+∥v∥,\|u+v\| \le \|u\| + \|v\|,

with equality if and only if one of u,vu,v is a nonnegative real scalar multiple of the other.

(B) Parallelogram Law. For all u,v∈Vu,v \in V,

∥u+v∥2+∥u−v∥2=2∥u∥2+2∥v∥2.\|u+v\|^2 + \|u-v\|^2 = 2\|u\|^2 + 2\|v\|^2.

(C) Characterization (Jordan–von Neumann). Suppose ∥⋅∥\|\cdot\| is any norm on a real or complex vector space VV (satisfying N1–N3, not assumed a priori to come from an inner product). Then ∥⋅∥\|\cdot\| is induced by some inner product on VV if and only if it satisfies the parallelogram law.

Proof

We prove (A), (B), and (C) in turn. Throughout, ∥v∥2=⟨v,v⟩\|v\|^2 = \langle v,v\rangle, and we use conjugate-symmetry and linearity in the first slot without further comment (these are the inner product axioms, assumed known from Cauchy-Schwarz Inequality).

Proof of (A), triangle inequality. Expand using linearity in the first argument and conjugate-linearity in the second (which follows from conjugate symmetry plus linearity in the first slot):

∥u+v∥2=⟨u+v,u+v⟩=⟨u,u⟩+⟨u,v⟩+⟨v,u⟩+⟨v,v⟩.\|u+v\|^2 = \langle u+v, u+v\rangle = \langle u,u\rangle + \langle u,v\rangle + \langle v,u\rangle + \langle v,v\rangle.

Since ⟨v,u⟩=⟨u,v⟩‾\langle v,u\rangle = \overline{\langle u,v\rangle}, the middle two terms sum to ⟨u,v⟩+⟨u,v⟩‾=2Re⁡⟨u,v⟩\langle u,v\rangle + \overline{\langle u,v\rangle} = 2\operatorname{Re}\langle u,v\rangle. Hence

∥u+v∥2=∥u∥2+2Re⁡⟨u,v⟩+∥v∥2.(∗)\|u+v\|^2 = \|u\|^2 + 2\operatorname{Re}\langle u,v\rangle + \|v\|^2. \tag{$\ast$}

For any complex number zz, Re⁡(z)≤∣z∣\operatorname{Re}(z) \le |z|. Applying this to z=⟨u,v⟩z = \langle u,v\rangle and then invoking the Cauchy-Schwarz Inequality, ∣⟨u,v⟩∣≤∥u∥∥v∥|\langle u,v\rangle| \le \|u\|\|v\|, gives

Re⁡⟨u,v⟩≤∣⟨u,v⟩∣≤∥u∥∥v∥.\operatorname{Re}\langle u,v\rangle \le |\langle u,v\rangle| \le \|u\|\|v\|.

Substituting into (∗)(\ast):

∥u+v∥2≤∥u∥2+2∥u∥∥v∥+∥v∥2=(∥u∥+∥v∥)2.\|u+v\|^2 \le \|u\|^2 + 2\|u\|\|v\| + \|v\|^2 = (\|u\|+\|v\|)^2.

Both sides of the original inequality ∥u+v∥≤∥u∥+∥v∥\|u+v\| \le \|u\|+\|v\| are nonnegative real numbers, and the function t↦t2t \mapsto t^2 is strictly increasing on [0,∞)[0,\infty); therefore taking square roots preserves the inequality, giving ∥u+v∥≤∥u∥+∥v∥\|u+v\| \le \|u\|+\|v\|, as desired.

Equality analysis. Equality in the final inequality forces equality in every step above (since each step was a ≤\le between nonnegative quantities feeding into a strictly increasing square). Concretely, equality holds if and only if:

Re⁡⟨u,v⟩=∣⟨u,v⟩∣and∣⟨u,v⟩∣=∥u∥∥v∥.\operatorname{Re}\langle u,v\rangle = |\langle u,v\rangle| \quad \text{and} \quad |\langle u,v\rangle| = \|u\|\|v\|.

The first condition says ⟨u,v⟩\langle u,v\rangle is a nonnegative real number. The second is the equality case of Cauchy–Schwarz, which holds if and only if u,vu,v are linearly dependent (this equality characterization is part of the Cauchy–Schwarz theorem itself). If v=0v = 0, the desired conclusion ("one is a nonnegative multiple of the other") holds trivially. If v≠0v \neq 0, linear dependence gives u=λvu = \lambda v for some scalar λ\lambda; then ⟨u,v⟩=λ∥v∥2\langle u,v\rangle = \lambda \|v\|^2, and since ∥v∥2>0\|v\|^2 > 0 is real and positive, ⟨u,v⟩≥0\langle u,v\rangle \ge 0 forces λ≥0\lambda \ge 0. Conversely, if u=λvu = \lambda v with λ≥0\lambda \ge 0 real (or v=0v=0), direct computation shows ∥u+v∥=(1+λ)∥v∥=∥u∥+∥v∥\|u+v\| = (1+\lambda)\|v\| = \|u\|+\|v\|, so equality holds. This confirms the stated equality condition in both directions.

Proof of (B), parallelogram law. Using (∗)(\ast) and the analogous expansion for ∥u−v∥2\|u-v\|^2 (replace vv by −v-v, noting ⟨u,−v⟩=−⟨u,v⟩\langle u,-v\rangle = -\langle u,v\rangle by linearity):

∥u−v∥2=∥u∥2−2Re⁡⟨u,v⟩+∥v∥2.\|u-v\|^2 = \|u\|^2 - 2\operatorname{Re}\langle u,v\rangle + \|v\|^2.

Adding this to (∗)(\ast):

∥u+v∥2+∥u−v∥2=2∥u∥2+2∥v∥2,\|u+v\|^2 + \|u-v\|^2 = 2\|u\|^2 + 2\|v\|^2,

since the cross terms 2Re⁡⟨u,v⟩2\operatorname{Re}\langle u,v\rangle and −2Re⁡⟨u,v⟩-2\operatorname{Re}\langle u,v\rangle cancel exactly. This proves (B).

Proof of (C), characterization. (Necessity) If ∥⋅∥\|\cdot\| is induced by an inner product, part (B) shows the parallelogram law holds. This direction is done. (Sufficiency, real case) Suppose ∥⋅∥\|\cdot\| is a norm on a real vector space VV satisfying the parallelogram law. Define

⟨u,v⟩:=14(∥u+v∥2−∥u−v∥2).\langle u,v\rangle := \frac{1}{4}\Big(\|u+v\|^2 - \|u-v\|^2\Big).

This is the polarization identity; we verify it defines an inner product whose induced norm is ∥⋅∥\|\cdot\|. Symmetry: ⟨u,v⟩=⟨v,u⟩\langle u,v\rangle = \langle v,u\rangle is immediate since ∥u−v∥=∥(−1)(v−u)∥=∥v−u∥\|u-v\| = \|(-1)(v-u)\| = \|v-u\| by (N2), and ∥u+v∥=∥v+u∥\|u+v\|=\|v+u\|. Positive-definiteness: Setting v=uv=u, ⟨u,u⟩=14(∥2u∥2−0)=∥u∥2≥0\langle u,u\rangle = \frac{1}{4}(\|2u\|^2 - 0) = \|u\|^2 \ge 0 by (N2), and this is 00 iff ∥u∥=0\|u\|=0 iff u=0u=0 by (N1). This also confirms the induced norm recovers ∥⋅∥\|\cdot\|. Additivity in the first slot: We must show ⟨u1+u2,v⟩=⟨u1,v⟩+⟨u2,v⟩\langle u_1+u_2, v\rangle = \langle u_1,v\rangle + \langle u_2,v\rangle. Apply the parallelogram law to the pairs (u1+v,u2)(u_1+v, u_2) and (u1−v,u2)(u_1-v,u_2):

∥u1+u2+v∥2+∥u1−u2+v∥2=2∥u1+v∥2+2∥u2∥2,\|u_1+u_2+v\|^2 + \|u_1-u_2+v\|^2 = 2\|u_1+v\|^2+2\|u_2\|^2,
∥u1+u2−v∥2+∥u1−u2−v∥2=2∥u1−v∥2+2∥u2∥2.\|u_1+u_2-v\|^2 + \|u_1-u_2-v\|^2 = 2\|u_1-v\|^2+2\|u_2\|^2.

Subtract the second from the first:

(∥u1+u2+v∥2−∥u1+u2−v∥2)+(∥u1−u2+v∥2−∥u1−u2−v∥2)=2∥u1+v∥2−2∥u1−v∥2.\big(\|u_1+u_2+v\|^2 - \|u_1+u_2-v\|^2\big) + \big(\|u_1-u_2+v\|^2-\|u_1-u_2-v\|^2\big) = 2\|u_1+v\|^2 - 2\|u_1-v\|^2.

Dividing by 44 and recognizing each bracket as 4⟨⋅,v⟩4\langle\cdot,v\rangle by the polarization definition:

⟨u1+u2,v⟩+⟨u1−u2,v⟩=2⟨u1,v⟩.(†)\langle u_1+u_2,v\rangle + \langle u_1-u_2,v\rangle = 2\langle u_1,v\rangle. \tag{$\dagger$}

Setting u2=u1u_2 = u_1 in (†)(\dagger) gives ⟨2u1,v⟩+⟨0,v⟩=2⟨u1,v⟩\langle 2u_1,v\rangle + \langle 0,v\rangle = 2\langle u_1,v\rangle; since ⟨0,v⟩=14(∥v∥2−∥v∥2)=0\langle 0,v\rangle = \frac14(\|v\|^2-\|v\|^2)=0, this yields ⟨2u1,v⟩=2⟨u1,v⟩\langle 2u_1,v\rangle = 2\langle u_1,v\rangle. More generally, replacing u1→a+b2u_1 \to \tfrac{a+b}{2}, u2→a−b2u_2 \to \tfrac{a-b}{2} in (†)(\dagger) (valid for any a,b∈Va,b\in V, since every pair (u1,u2)(u_1,u_2) arises this way) gives exactly

⟨a,v⟩+⟨b,v⟩=2⟨a+b2,v⟩.\langle a,v\rangle + \langle b,v\rangle = 2\left\langle \tfrac{a+b}{2}, v\right\rangle.

Combined with the scaling fact ⟨2w,v⟩=2⟨w,v⟩\langle 2w,v\rangle = 2\langle w,v\rangle applied to w=a+b2w = \tfrac{a+b}2, this gives ⟨a,v⟩+⟨b,v⟩=⟨a+b,v⟩\langle a,v\rangle+\langle b,v\rangle = \langle a+b,v\rangle, which is precisely additivity. Homogeneity over Q\mathbb{Q}, then R\mathbb{R}: Additivity gives ⟨nw,v⟩=n⟨w,v⟩\langle nw,v\rangle = n\langle w,v\rangle for all positive integers nn by induction, and ⟨−w,v⟩=−⟨w,v⟩\langle -w,v\rangle = -\langle w,v\rangle follows from ∥−w+v∥=∥w−v∥\|{-w}+v\| = \|w-v\| and ∥−w−v∥=∥w+v∥\|{-w}-v\|=\|w+v\| (by (N2) with λ=−1\lambda=-1), which swaps the two terms in the polarization formula up to sign. Hence ⟨qw,v⟩=q⟨w,v⟩\langle qw,v\rangle = q\langle w,v\rangle for all q∈Qq \in \mathbb{Q}. The map λ↦⟨λw,v⟩\lambda \mapsto \langle \lambda w, v\rangle is continuous in λ\lambda (since λ↦∥λw±v∥\lambda \mapsto \|\lambda w \pm v\| is continuous, being a composition of the continuous norm with continuous vector-space operations), and it agrees with the continuous linear function λ↦λ⟨w,v⟩\lambda \mapsto \lambda \langle w,v\rangle on the dense subset Q⊂R\mathbb{Q} \subset \mathbb{R}. Two continuous functions agreeing on a dense set agree everywhere, so ⟨λw,v⟩=λ⟨w,v⟩\langle \lambda w,v\rangle = \lambda\langle w,v\rangle for all λ∈R\lambda \in \mathbb{R}. Together with additivity, this establishes full linearity in the first slot. By symmetry, ⟨⋅,⋅⟩\langle\cdot,\cdot\rangle is then bilinear. All inner product axioms hold, so ⟨⋅,⋅⟩\langle\cdot,\cdot\rangle is an inner product on VV inducing ∥⋅∥\|\cdot\|, completing the real case. (Sufficiency, complex case) For a complex vector space, define

⟨u,v⟩:=14∑k=03ik∥u+ikv∥2=14(∥u+v∥2−∥u−v∥2+i∥u+iv∥2−i∥u−iv∥2).\langle u,v\rangle := \frac{1}{4}\sum_{k=0}^{3} i^k \|u + i^k v\|^2 = \frac14\Big(\|u+v\|^2 - \|u-v\|^2 + i\|u+iv\|^2 - i\|u-iv\|^2\Big).

The real part of this expression equals the real polarization formula applied to ∥⋅∥\|\cdot\| viewed as a norm on VV regarded as a real vector space, so by the real case argument (applied to real-linear combinations), Re⁡⟨⋅,⋅⟩\operatorname{Re}\langle\cdot,\cdot\rangle is R\mathbb{R}-bilinear, symmetric, and positive-definite in the sense needed. A direct computation (substituting iviv for vv and using ∥iw∥=∥w∥\|iw\|=\|w\| from (N2) with ∣λ∣=∣i∣=1|\lambda|=|i|=1) shows ⟨u,iv⟩=i⟨u,v⟩\langle u, iv\rangle = i\langle u,v\rangle and ⟨iu,v⟩=−i⟨u,v⟩\langle iu,v\rangle = -i\langle u,v\rangle, from which conjugate symmetry ⟨u,v⟩=⟨v,u⟩‾\langle u,v\rangle = \overline{\langle v,u\rangle} and full sesquilinearity follow by combining the real-linear structure with these identities. Setting v=uv=u collapses the imaginary terms (since ∥u+iu∥=∣1+i∣ ∥u∥=∥u−iu∥\|u+iu\| = |1+i|\,\|u\| = \|u-iu\|, these two terms cancel), leaving ⟨u,u⟩=∥u∥2\langle u,u\rangle = \|u\|^2, matching the real case and confirming positive-definiteness and that the induced norm is ∥⋅∥\|\cdot\|. This completes the proof of (C). ■\blacksquare

∎

Worked examples

Example 1: Equality in the triangle inequality, R3\mathbb{R}^3 with a non-obvious dependent pair. Let u=(2,−1,3)u = (2,-1,3) and v=(−4,2,−6)v = (-4,2,-6) in R3\mathbb{R}^3 with the standard inner product ⟨x,y⟩=∑xiyi\langle x,y\rangle = \sum x_iy_i. Notice v=−2uv = -2u, so these are linearly dependent but with a negative scalar — a case that should fail the equality condition, testing the theorem's sharpness. Compute:

∥u∥=4+1+9=14,∥v∥=16+4+36=56=214.\|u\| = \sqrt{4+1+9} = \sqrt{14}, \qquad \|v\| = \sqrt{16+4+36} = \sqrt{56} = 2\sqrt{14}.
u+v=(2−4, −1+2, 3−6)=(−2,1,−3),∥u+v∥=4+1+9=14.u+v = (2-4,\,-1+2,\,3-6) = (-2,1,-3), \qquad \|u+v\| = \sqrt{4+1+9}=\sqrt{14}.

Check the inequality: ∥u+v∥=14\|u+v\| = \sqrt{14} versus ∥u∥+∥v∥=14+214=314\|u\|+\|v\| = \sqrt{14}+2\sqrt{14} = 3\sqrt{14}. Indeed 14≤314\sqrt{14} \le 3\sqrt{14}, and the inequality is strict, exactly as the theorem predicts, since v=−2uv = -2u has negative scalar −2<0-2 < 0, so u,vu,v do not point in the same direction (equality requires a nonnegative multiple). Now test a case that does achieve equality: let u′=(2,−1,3)u' = (2,-1,3) and v′=3u′=(6,−3,9)v' = 3u' = (6,-3,9). Then

∥u′∥=14,∥v′∥=314,u′+v′=(8,−4,12)=4u′,∥u′+v′∥=414.\|u'\| = \sqrt{14}, \quad \|v'\| = 3\sqrt{14}, \quad u'+v' = (8,-4,12) = 4u', \quad \|u'+v'\| = 4\sqrt{14}.

Indeed ∥u′+v′∥=414=14+314=∥u′∥+∥v′∥\|u'+v'\| = 4\sqrt{14} = \sqrt{14}+3\sqrt{14} = \|u'\|+\|v'\|: equality holds, consistent with v′=3u′v' = 3u', 3≥03 \ge 0.

Example 2: Verifying the parallelogram law fails for a non-inner-product norm, using R2\mathbb{R}^2 with the sup-norm. Let V=R2V=\mathbb{R}^2 with ∥x∥∞:=max⁡(∣x1∣,∣x2∣)\|x\|_\infty := \max(|x_1|,|x_2|), a legitimate norm (satisfies N1–N3, easily checked) that is not induced by any inner product. Take u=(3,1)u=(3,1) and v=(1,−2)v=(1,-2), an asymmetric pair to avoid accidentally hitting a degenerate case. Compute:

u+v=(4,−1),∥u+v∥∞=max⁡(4,1)=4.u+v = (4,-1), \quad \|u+v\|_\infty = \max(4,1) = 4.
u−v=(2,3),∥u−v∥∞=max⁡(2,3)=3.u-v = (2,3), \quad \|u-v\|_\infty = \max(2,3) = 3.
∥u+v∥∞2+∥u−v∥∞2=16+9=25.\|u+v\|_\infty^2 + \|u-v\|_\infty^2 = 16+9 = 25.
∥u∥∞=max⁡(3,1)=3,∥v∥∞=max⁡(1,2)=2,2∥u∥∞2+2∥v∥∞2=2(9)+2(4)=26.\|u\|_\infty = \max(3,1)=3, \quad \|v\|_\infty = \max(1,2)=2, \quad 2\|u\|_\infty^2+2\|v\|_\infty^2 = 2(9)+2(4)=26.

Since 25≠2625 \ne 26, the parallelogram law fails for this pair. By part (C) of the theorem (contrapositive of sufficiency), ∥⋅∥∞\|\cdot\|_\infty cannot arise from any inner product on R2\mathbb{R}^2 — a fact confirmed by a single numerical check rather than an abstract argument.

CounterexampleDropping Cauchy–Schwarz breaks the triangle inequality proof — and a genuine bilinear form without positive-definiteness fails (N3)

The proof of part (A) used the Cauchy-Schwarz Inequality in one indispensable step: bounding Re⁡⟨u,v⟩≤∥u∥∥v∥\operatorname{Re}\langle u,v\rangle \le \|u\|\|v\|. If we instead work with a symmetric bilinear form that is not positive-definite (violating axiom 3 of the inner product definition, hence invalidating Cauchy–Schwarz itself), the triangle inequality for the associated "norm" can fail outright — indeed the associated quantity may not even be a norm. Let V=R2V = \mathbb{R}^2 with the indefinite symmetric bilinear form

⟨x,y⟩:=x1y1−x2y2(Minkowski-type form).\langle x,y\rangle := x_1y_1 - x_2y_2 \quad (\text{Minkowski-type form}).

This satisfies conjugate symmetry and linearity but not positive-definiteness: for x=(1,2)x=(1,2), ⟨x,x⟩=1−4=−3<0\langle x,x\rangle = 1-4=-3<0. So ⟨x,x⟩\sqrt{\langle x,x\rangle} is not even real for such xx, and no norm is induced on all of VV. Restrict attention to vectors where ⟨x,x⟩≥0\langle x,x\rangle \ge 0 (the "timelike or lightlike" cone) and define ∥x∥:=⟨x,x⟩\|x\| := \sqrt{\langle x,x\rangle} there, to see concretely how the argument breaks down. Take u=(2,1)u=(2,1) and v=(2,−1)v=(2,-1): both have ⟨u,u⟩=4−1=3≥0\langle u,u\rangle = 4-1=3\ge0 and ⟨v,v⟩=4−1=3≥0\langle v,v\rangle=4-1=3\ge 0, so ∥u∥=∥v∥=3\|u\|=\|v\|=\sqrt3. Now u+v=(4,0)u+v=(4,0), so ⟨u+v,u+v⟩=16−0=16\langle u+v,u+v\rangle = 16-0=16, giving ∥u+v∥=4\|u+v\|=4. Check: is ∥u+v∥≤∥u∥+∥v∥\|u+v\| \le \|u\|+\|v\|? We need 4≤3+3=23≈3.464 \le \sqrt3+\sqrt3 = 2\sqrt3 \approx 3.46. But 4>3.464 > 3.46 — the inequality fails. The reason is structural: Cauchy–Schwarz itself requires positive-definiteness, which fails here, so the crucial bound Re⁡⟨u,v⟩≤∥u∥∥v∥\operatorname{Re}\langle u,v\rangle \le \|u\|\|v\| used in the proof of (A) is simply unavailable, and indeed here ⟨u,v⟩=4+1=5>∥u∥∥v∥=3\langle u,v\rangle = 4+1=5 > \|u\|\|v\| = 3. This shows positive-definiteness is not a cosmetic hypothesis; removing it breaks the triangle inequality's proof and its conclusion simultaneously.

Scope, edge cases & common traps

SituationWhat happensWhy
uu or v=0v = 0Triangle inequality holds with equality trivially when the nonzero vector is compared with 00∥0+v∥=∥v∥=∥0∥+∥v∥\|0+v\| = \|v\| = \|0\|+\|v\|
F=C\mathbb{F}=\mathbb{C}Proof identical, but must track Re⁡⟨u,v⟩\operatorname{Re}\langle u,v\rangle, not ⟨u,v⟩\langle u,v\rangle itselfCross term in (∗)(\ast) is real only via ⟨u,v⟩+⟨u,v⟩‾\langle u,v\rangle+\overline{\langle u,v\rangle}
Equality caseNeeds ⟨u,v⟩\langle u,v\rangle real and nonnegative, not just linear dependenceA negative real multiple gives linear dependence but strict inequality (see Example 1)
Non-inner-product norms (ℓ1\ell^1, ℓ∞\ell^\infty)Satisfy triangle inequality (by direct proof, not via Cauchy–Schwarz) but fail parallelogram lawThese norms are not induced by any inner product
Infinite-dimensional spacesAll statements (A)–(C) hold verbatim for any inner product space, finite or infinite dimensionalProofs use only the inner product axioms, no dimension-counting

The mathematician behind it

Origins · 1 minJohn von Neumann and how this was discovered
JV
John von Neumann
1903–1957 · Institute for Advanced Study, Princeton, foundational work in operator theory and Hilbert space theory

Von Neumann, working alongside Pascual Jordan, formalized the abstract axiomatic treatment of Hilbert spaces in the early 1930s as part of his rigorous mathematical foundation for quantum mechanics. The question of exactly which norms could be "geometrized" by an inner product — giving angles, orthogonality, and projections — was central to distinguishing genuinely Euclidean-like structures (Hilbert spaces) from the more general Banach spaces being studied by Banach and his school in Poland at the same time. The Jordan–von Neumann theorem, identifying the parallelogram law as the exact criterion, resolved this question completely and remains one of the cleanest converse theorems in all of functional analysis.

History
- 1901–1908: Erhard Schmidt and David Hilbert develop early "Hilbert space" ideas from integral equations, implicitly using inner-product geometry.
- 1920s: Stefan Banach and collaborators develop the general theory of normed spaces, later called Banach spaces.
- 1932: Pascual Jordan and John von Neumann publish "On inner products in linear metric spaces," proving the parallelogram law characterization.
- 1932: Von Neumann's Mathematical Foundations of Quantum Mechanics cements the Hilbert space axiomatic framework using exactly this norm theory.
- Mid-20th century: The triangle inequality and parallelogram law become standard first tools in every treatment of Hilbert and Banach space theory.
Why is the triangle inequality not simply assumed as an axiom for inner product norms?

It could be assumed as part of the definition of "norm" in general (indeed it is, as axiom N3), but the point of this article is that when a norm specifically comes from an inner product, the triangle inequality is not an independent fact you need to separately verify or postulate — it is forced by the three inner product axioms (conjugate symmetry, linearity, positive-definiteness) via Cauchy–Schwarz. This is valuable because it means checking that something is an inner product automatically certifies that its induced norm is a genuine norm, streamlining the verification of new examples such as function spaces with integral inner products.

Does the triangle inequality require the vector space to be finite-dimensional?

No. The proof given uses only the three inner product axioms and the Cauchy–Schwarz inequality, neither of which references a basis, dimension, or any finiteness assumption. The same proof applies verbatim to infinite-dimensional inner product spaces, such as L2L^2 function spaces or ℓ2\ell^2 sequence spaces, which is precisely why these spaces support a full theory of distance and convergence (Hilbert space theory).

What exactly fails when the parallelogram law fails, geometrically?

Geometrically, the parallelogram law encodes a very specific relationship between the diagonals and sides of a parallelogram that is special to Euclidean-type geometry generated by an inner product. Norms like ∥⋅∥1\|\cdot\|_1 or ∥⋅∥∞\|\cdot\|_\infty describe genuinely different, non-Euclidean notions of distance (e.g., "taxicab" distance) where this diagonal-side relationship simply does not hold numerically, as shown in Example 2. The failure means there is no way to recover angles or orthogonality consistent with that norm; you can measure length and satisfy the triangle inequality, but you cannot meaningfully talk about vectors being "perpendicular" in a way compatible with the norm.

Is the polarization identity used in the proof valid for any norm, or only inner-product norms?

The polarization formula ⟨u,v⟩=14(∥u+v∥2−∥u−v∥2)\langle u,v\rangle = \frac14(\|u+v\|^2 - \|u-v\|^2) can be written down for any norm, but it only defines a genuine inner product (bilinear, symmetric, positive-definite) when the norm satisfies the parallelogram law. That is exactly the content of part (C) of the theorem: the formula is a candidate inner product, and the parallelogram law is precisely the condition needed to prove it satisfies bilinearity. Without the parallelogram law, the formula still produces some number for each pair (u,v)(u,v), but that number generally fails to be additive in uu, so it is not an inner product at all.

Can equality in the triangle inequality occur for linearly dependent vectors with a negative scalar?

No, and this is the most easily misremembered subtlety. If u=λvu = \lambda v with λ<0\lambda < 0 (and v≠0v \ne 0), the vectors are linearly dependent but point in opposite directions, and one checks directly that ∥u+v∥=∣1+λ∣ ∥v∥<(1+∣λ∣)∥v∥=∥u∥+∥v∥\|u+v\| = |1+\lambda|\,\|v\| < (1+|\lambda|)\|v\| = \|u\|+\|v\| whenever λ<0\lambda<0, so the inequality is strict. Equality requires λ≥0\lambda \ge 0 specifically, i.e., the vectors must point in the same direction (or one must be zero), not merely lie on the same line.

How does this relate to building orthonormal bases?

Once a norm and the notion of angle (via Cauchy–Schwarz, cos⁡θ=⟨u,v⟩/(∥u∥∥v∥)\cos\theta = \langle u,v\rangle/(\|u\|\|v\|)) are established, the next natural step is to find bases where basis vectors have norm 11 and are pairwise orthogonal (⟨ei,ej⟩=0\langle e_i,e_j\rangle=0 for i≠ji\ne j). This is the subject of Orthonormal Bases and Coordinates, and it depends crucially on having a well-behaved norm; the triangle inequality guarantees that coordinate expansions in such bases produce a consistent, finite notion of distance and length.

Why does the complex case need Re⁡⟨u,v⟩\operatorname{Re}\langle u,v \rangle instead of just ⟨u,v⟩\langle u,v\rangle?

Over C\mathbb{C}, ⟨u,v⟩\langle u,v \rangle can be any complex number, so it does not make sense to compare it directly to the real quantity ∥u∥∥v∥\|u\|\|v\| using ≤\le. Expanding ∥u+v∥2\|u+v\|^2 using conjugate symmetry produces the sum ⟨u,v⟩+⟨u,v⟩‾\langle u,v\rangle + \overline{\langle u,v\rangle}, which always equals 2Re⁡⟨u,v⟩2\operatorname{Re}\langle u,v\rangle, a real number. This real quantity is what Cauchy–Schwarz's modulus bound ∣⟨u,v⟩∣≤∥u∥∥v∥|\langle u,v\rangle|\le \|u\|\|v\| can legitimately be compared against, via Re⁡(z)≤∣z∣\operatorname{Re}(z)\le |z|.

Before this

Where this leads