Inner Products and Inner Product Spaces

How do you measure angles and lengths in a vector space with no built-in geometry? A vector space, as defined by its axioms, only knows how to add vectors and scale them. It cannot tell you whether two vectors are perpendicular. It also cannot tell you how "long" a vector is. Nor can it tell you how close two vectors are to each other. Yet these geometric ideas are indispensable. They matter in physics, statistics, signal processing, and numerical analysis. They all trace back to a single algebraic gadget: the inner product. An inner product is a rule that takes two vectors and returns a scalar. It is designed so this scalar behaves like the dot product on Rn\mathbb{R}^n that you already know. From this one rule, an entire geometric universe unfolds. This includes lengths (norms), angles, orthogonality, and projections. It also includes the notion of the "best approximation" of one vector by another. This article sets up a remarkable fact, which later articles will exploit. Three simple axioms are enough to force all of this rich geometric structure to exist. The big picture is this: a vector space equipped with an inner product is called an inner product space. It is the natural home for doing geometry in linear algebra. Every finite-dimensional inner product space, real or complex, behaves the same way. Up to a change of basis, it behaves exactly like Rn\mathbb{R}^n with the dot product. Alternatively, it behaves like Cn\mathbb{C}^n with its Hermitian analogue. But the same axioms apply just as well to infinite-dimensional spaces. This includes spaces of functions, sequences, and signals. That is where the theory shows its full power. This article gives the precise axiomatic definition of an inner product over both R\mathbb{R} and C\mathbb{C}. It explains why each axiom is shaped the way it is. It verifies the standard examples in complete detail. It also proves the basic algebraic consequences that make the definition workable. It sets the stage — without yet proving — for the Norm Induced by an Inner Product, the Cauchy-Schwarz Inequality, and Orthogonality and Orthogonal Sets.

Key takeaways
  • An inner product ⟨⋅,⋅⟩\langle \cdot,\cdot\rangle on a vector space VV over F∈{R,C}\mathbb{F} \in \{\mathbb{R},\mathbb{C}\} assigns a scalar to each pair of vectors.
  • It must satisfy three axioms: conjugate symmetry, linearity in the first argument, and positive-definiteness.
  • Over R\mathbb{R} conjugate symmetry reduces to ordinary symmetry; over C\mathbb{C} the conjugate is essential to make ⟨v,v⟩\langle v,v\rangle real.
  • The axioms force conjugate-linearity in the second slot, so complex inner products are not bilinear but sesquilinear.
  • Standard models: the dot product on Rn\mathbb{R}^n, the Hermitian form ∑xiyi‾\sum x_i\overline{y_i} on Cn\mathbb{C}^n, and integral inner products on function spaces.
  • A vector space paired with a chosen inner product is called an inner product space; this structure underlies norms, angles, and orthogonality.
  • Key idea: three algebraic axioms are precisely what is needed to manufacture a full theory of length and angle inside an abstract vector space.

Where this lives today

Inner products are the algebraic seed from which an enormous amount of applied and pure mathematics grows. In quantum mechanics, physical states are vectors in a complex Hilbert space, and the Hermitian inner product ⟨ψ,ϕ⟩\langle \psi,\phi\rangle computes transition amplitudes whose squared magnitude gives observable probabilities. In statistics and machine learning, the covariance of two random variables is itself an inner product on the vector space of mean-zero random variables with finite variance, and this single observation underlies least-squares regression, principal component analysis, and the geometry of Gaussian processes. In signal processing, the L2L^2 inner product on function spaces (case (d) above) is exactly what makes Fourier analysis work: Fourier coefficients are inner products of a signal against sinusoidal basis functions. In numerical linear algebra, weighted inner products like case (c) arise naturally when discretizing differential equations, since the natural pairing between grid functions typically carries mesh-dependent weights. In cryptography and coding theory, inner products over finite fields (a further generalization beyond R\mathbb{R} and C\mathbb{C}, requiring different axioms) underlie lattice-based cryptographic schemes and error-correcting codes. And in computer graphics, the dot product on R3\mathbb{R}^3 computes angles between surface normals and light directions for shading calculations, a direct descendant of Example (a) above.

:::quiz[On C[a,b]C[a,b] with ⟨f,g⟩=∫abf(t)g(t) dt\langle f,g\rangle = \int_a^b f(t)g(t)\,dt, why does ⟨f,f⟩=0\langle f,f\rangle=0 force f≡0f\equiv 0?]

  • Because every continuous function integrates to zero
  • Because if f(t0)≠0f(t_0)\neq0 for some t0t_0, continuity guarantees f2f^2 stays bounded away from 00 on a whole subinterval, giving a strictly positive integral ✓
  • Because the Riemann integral is defined to be zero only for the zero function, by convention
  • Because ff must be a polynomial

Ask the tutor: Why does the complex inner product need conjugate symmetry instead of plain symmetry, and what would break if we tried to use plain symmetry over C\mathbb{C}?

DefinitionInner Product and Inner Product Space

Let VV be a vector space over a field F\mathbb{F}, where F\mathbb{F} is either R\mathbb{R} or C\mathbb{C}. An inner product on VV is a function

⟨⋅,⋅⟩:V×V→F\langle \cdot,\cdot\rangle : V \times V \to \mathbb{F}

that satisfies the following three axioms for all u,v,w∈Vu,v,w \in V and all scalars λ∈F\lambda \in \mathbb{F}: 1. Conjugate symmetry: ⟨v,w⟩=⟨w,v⟩‾\langle v,w\rangle = \overline{\langle w,v\rangle}. 2. Linearity in the first argument: ⟨u+v, w⟩=⟨u,w⟩+⟨v,w⟩\langle u+v,\,w\rangle = \langle u,w\rangle + \langle v,w\rangle and ⟨λv,w⟩=λ⟨v,w⟩\langle \lambda v, w\rangle = \lambda\langle v,w\rangle. 3. Positive-definiteness: ⟨v,v⟩≥0\langle v,v\rangle \ge 0 for all vv, with ⟨v,v⟩=0\langle v,v\rangle = 0 if and only if v=0v = 0. (When F=R\mathbb{F}=\mathbb{R}, the bar in axiom 1 is the identity map, so it reads ⟨v,w⟩=⟨w,v⟩\langle v,w\rangle=\langle w,v\rangle, ordinary symmetry. Axiom 3 requires ⟨v,v⟩∈R\langle v,v\rangle\in\mathbb{R}, which is automatic in the real case and a consequence of axiom 1 in the complex case, since axiom 1 with w=vw=v gives ⟨v,v⟩=⟨v,v⟩‾\langle v,v\rangle = \overline{\langle v,v\rangle}, forcing ⟨v,v⟩\langle v,v\rangle to be real.) A vector space VV together with a chosen inner product ⟨⋅,⋅⟩\langle\cdot,\cdot\rangle is called an inner product space. A real inner product space uses F=R\mathbb{F}=\mathbb{R}; a complex (or Hermitian) inner product space uses F=C\mathbb{F}=\mathbb{C}.

Intuition

Think of the dot product on R2\mathbb{R}^2: ⟨(x1,x2),(y1,y2)⟩=x1y1+x2y2\langle (x_1,x_2),(y_1,y_2)\rangle = x_1y_1+x_2y_2. This single formula already encodes length (⟨v,v⟩=∥v∥2\langle v,v\rangle = \|v\|^2) and the cosine of the angle between vectors (via ⟨v,w⟩=∥v∥∥w∥cos⁡θ\langle v,w\rangle = \|v\|\|w\|\cos\theta, which will be proved once The Cauchy-Schwarz Inequality is established). The axioms in the definition above are simply a distillation of the algebraic properties that make the dot product behave this way, stripped of any reference to coordinates, so that the same three rules can be imposed on wildly different spaces — polynomials, matrices, random variables, continuous functions — and still yield a working geometry. Symmetry (axiom 1) says the order of the two vectors does not matter, up to conjugation; this is what will let us later define angles symmetrically. Linearity in the first slot (axiom 2) says the inner product interacts with vector addition and scalar multiplication the way any "linear measurement" should. Positive-definiteness (axiom 3) is the geometric heart of the matter: it guarantees that ⟨v,v⟩\langle v,v\rangle deserves to be called a squared length, since it is never negative and vanishes only for the zero vector — the one vector with no length at all. The complex case needs the conjugate because without it, ⟨v,v⟩\langle v,v\rangle need not be real, and "squared length" would not even make sense as a nonnegative number. The figure below shows how the standard inner product on R2\mathbb{R}^2 encodes both length and angle from a single algebraic rule, and previews (without proof) the geometric payoff that later articles establish rigorously.

Figure
The inner product ⟨v,w⟩=∥v∥ ∥w∥cos⁡θ\langle v,w\rangle = \|v\|\,\|w\|\cos\theta encodes both length and angle
θ v = (3, 4) w = (5, 0) ⟨v, w⟩ = 3·5 + 4·0 = 15 ∥v∥ = 5, ∥w∥ = 5, cosθ = 15/25 = 0.6

Basic algebraic consequences

Before stating the central theorem, three immediate consequences of the axioms are worth isolating, since every worked example below relies on them.

LemmaConjugate-linearity in the second argument

Let VV be an inner product space over F∈{R,C}\mathbb{F}\in\{\mathbb{R},\mathbb{C}\}. For all u,v,w∈Vu,v,w\in V and λ∈F\lambda\in\mathbb{F},

⟨u, v+w⟩=⟨u,v⟩+⟨u,w⟩,⟨u,λv⟩=λ‾ ⟨u,v⟩.\langle u,\,v+w\rangle = \langle u,v\rangle + \langle u,w\rangle, \qquad \langle u,\lambda v\rangle = \overline{\lambda}\,\langle u,v\rangle.
Proof

By conjugate symmetry (axiom 1) and linearity in the first slot (axiom 2):

⟨u,v+w⟩=⟨v+w,u⟩‾=⟨v,u⟩+⟨w,u⟩‾=⟨v,u⟩‾+⟨w,u⟩‾=⟨u,v⟩+⟨u,w⟩,\langle u,v+w\rangle = \overline{\langle v+w,u\rangle} = \overline{\langle v,u\rangle + \langle w,u\rangle} = \overline{\langle v,u\rangle} + \overline{\langle w,u\rangle} = \langle u,v\rangle + \langle u,w\rangle,

using that complex conjugation distributes over sums, a+b‾=a‾+b‾\overline{a+b}=\overline a+\overline b, a basic property of C\mathbb{C} (or the identity map, if F=R\mathbb{F}=\mathbb{R}). Similarly,

⟨u,λv⟩=⟨λv,u⟩‾=λ⟨v,u⟩‾=λ‾ ⟨v,u⟩‾=λ‾ ⟨u,v⟩,\langle u,\lambda v\rangle = \overline{\langle \lambda v,u\rangle} = \overline{\lambda\langle v,u\rangle} = \overline{\lambda}\,\overline{\langle v,u\rangle} = \overline{\lambda}\,\langle u,v\rangle,

using λa‾=λ‾ a‾\overline{\lambda a} = \overline{\lambda}\,\overline{a}, again a basic property of complex conjugation. ■\blacksquare

∎

This lemma explains why a complex inner product is called sesquilinear ("one-and-a-half linear"): linear in the first slot, conjugate-linear in the second. (Some texts adopt the opposite convention, linear in the second slot; the definition above follows Axler's convention.) When F=R\mathbb{F}=\mathbb{R}, conjugation is trivial, so the inner product is bilinear: linear in both slots.

LemmaVanishing against everything forces the zero vector

Let VV be an inner product space. If v∈Vv\in V satisfies ⟨v,w⟩=0\langle v,w\rangle = 0 for every w∈Vw\in V, then v=0v=0.

Proof

Apply the hypothesis with w=vw=v: ⟨v,v⟩=0\langle v,v\rangle = 0. By the positive-definiteness axiom, ⟨v,v⟩=0\langle v,v\rangle = 0 implies v=0v=0. ■\blacksquare

∎

This lemma is exactly what is needed later to show, for instance, that a vector orthogonal to a spanning set of VV must be zero — a fact used repeatedly in the theory of Orthogonality and Orthogonal Sets.

TheoremWell-definedness and consistency of the standard models

The following are inner products, in the sense of the definition above:

(a) The Euclidean inner product on Rn\mathbb{R}^n. For x=(x1,…,xn)x=(x_1,\dots,x_n), y=(y1,…,yn)∈Rny=(y_1,\dots,y_n) \in \mathbb{R}^n,

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

(b) The standard Hermitian inner product on Cn\mathbb{C}^n. For x=(x1,…,xn)x=(x_1,\dots,x_n), y=(y1,…,yn)∈Cny=(y_1,\dots,y_n)\in\mathbb{C}^n,

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

(c) A weighted inner product on Cn\mathbb{C}^n. Fix positive real numbers c1,…,cn>0c_1,\dots,c_n>0. For x,y∈Cnx,y\in\mathbb{C}^n,

⟨x,y⟩:=∑i=1nci xiyi‾.\langle x,y\rangle := \sum_{i=1}^n c_i\, x_i\overline{y_i}.

(d) The L2L^2 inner product on C[a,b]C[a,b]. Let V=C[a,b]V = C[a,b] be the vector space of continuous real-valued functions on [a,b][a,b], with a<ba<b. For f,g∈Vf,g\in V,

⟨f,g⟩:=∫abf(t)g(t) dt.\langle f,g\rangle := \int_a^b f(t)g(t)\,dt.
Proof

We verify each of the three axioms for each construction. Throughout, "linearity in the first argument" means additivity plus homogeneity, and by the Lemma above it suffices to check axioms 1–3 directly; conjugate-linearity in the second slot then follows automatically.

(a) Euclidean inner product on Rn\mathbb{R}^n. Conjugate symmetry: Since F=R\mathbb{F}=\mathbb{R}, conjugation is the identity, and ⟨x,y⟩=∑xiyi=∑yixi=⟨y,x⟩\langle x,y\rangle = \sum x_iy_i = \sum y_ix_i = \langle y,x\rangle by commutativity of multiplication in R\mathbb{R}. Linearity in the first slot: For x,x′,y∈Rnx,x',y\in\mathbb{R}^n,

⟨x+x′,y⟩=∑i=1n(xi+xi′)yi=∑i=1nxiyi+∑i=1nxi′yi=⟨x,y⟩+⟨x′,y⟩,\langle x+x', y\rangle = \sum_{i=1}^n (x_i+x_i')y_i = \sum_{i=1}^n x_iy_i + \sum_{i=1}^n x_i'y_i = \langle x,y\rangle + \langle x',y\rangle,

using distributivity of multiplication over addition in R\mathbb{R} and associativity/commutativity of the finite sum. For λ∈R\lambda\in\mathbb{R},

⟨λx,y⟩=∑i=1n(λxi)yi=λ∑i=1nxiyi=λ⟨x,y⟩.\langle \lambda x, y\rangle = \sum_{i=1}^n (\lambda x_i)y_i = \lambda\sum_{i=1}^n x_iy_i = \lambda\langle x,y\rangle.

Positive-definiteness: ⟨x,x⟩=∑i=1nxi2≥0\langle x,x\rangle = \sum_{i=1}^n x_i^2 \ge 0 since each xi2≥0x_i^2\ge 0 in R\mathbb{R}, and this is a sum of finitely many nonnegative reals, hence nonnegative. If ⟨x,x⟩=0\langle x,x\rangle = 0, then ∑xi2=0\sum x_i^2 = 0; since each term is ≥0\ge 0, every term must be 00 (a sum of nonnegative reals is zero only if each summand is zero), so xi2=0x_i^2=0 for all ii, hence xi=0x_i=0 for all ii, i.e. x=0x=0. Conversely if x=0x=0 then trivially ⟨x,x⟩=0\langle x,x\rangle=0.

(b) Standard Hermitian inner product on Cn\mathbb{C}^n. Conjugate symmetry: For z=w‾z=\overline{w} any complex number, recall z‾‾=z\overline{\overline z}=z and zw‾=z‾ w‾\overline{zw}=\overline z\,\overline w. Then

⟨y,x⟩‾=∑i=1nyixi‾‾=∑i=1nyixi‾‾=∑i=1nyi‾ xi‾‾=∑i=1nxiyi‾=⟨x,y⟩,\overline{\langle y,x\rangle} = \overline{\sum_{i=1}^n y_i\overline{x_i}} = \sum_{i=1}^n \overline{y_i\overline{x_i}} = \sum_{i=1}^n \overline{y_i}\,\overline{\overline{x_i}} = \sum_{i=1}^n x_i\overline{y_i} = \langle x,y\rangle,

so ⟨x,y⟩=⟨y,x⟩‾\langle x,y\rangle = \overline{\langle y,x\rangle}, which is exactly axiom 1. Linearity in the first slot: For x,x′,y∈Cnx,x',y\in\mathbb{C}^n,

⟨x+x′,y⟩=∑i=1n(xi+xi′)yi‾=∑xiyi‾+∑xi′yi‾=⟨x,y⟩+⟨x′,y⟩,\langle x+x',y\rangle = \sum_{i=1}^n (x_i+x_i')\overline{y_i} = \sum x_i\overline{y_i} + \sum x_i'\overline{y_i} = \langle x,y\rangle+\langle x',y\rangle,

and for λ∈C\lambda\in\mathbb{C}, ⟨λx,y⟩=∑i=1n(λxi)yi‾=λ∑i=1nxiyi‾=λ⟨x,y⟩\langle \lambda x,y\rangle = \sum_{i=1}^n (\lambda x_i)\overline{y_i} = \lambda\sum_{i=1}^n x_i\overline{y_i} = \lambda\langle x,y\rangle. Positive-definiteness: ⟨x,x⟩=∑i=1nxixi‾=∑i=1n∣xi∣2\langle x,x\rangle = \sum_{i=1}^n x_i\overline{x_i} = \sum_{i=1}^n |x_i|^2, using the identity zz‾=∣z∣2z\overline z = |z|^2 for z∈Cz\in\mathbb{C}. Each ∣xi∣2≥0|x_i|^2\ge 0 is a nonnegative real number, so the finite sum is ≥0\ge 0; it is real (not merely a complex number that happens to be real), consistent with the requirement flagged in the Definition. If ⟨x,x⟩=0\langle x,x\rangle=0 then ∑∣xi∣2=0\sum|x_i|^2=0 forces each ∣xi∣2=0|x_i|^2=0 (sum of nonnegative reals vanishes only if every term vanishes), so xi=0x_i=0 for all ii, i.e., x=0x=0; the converse is immediate.

(c) Weighted inner product on Cn\mathbb{C}^n. The argument mirrors (b) with the extra positive weights cic_i. Conjugate symmetry: ⟨y,x⟩‾=∑ciyixi‾‾=∑ciyi‾ xi=∑cixiyi‾=⟨x,y⟩\overline{\langle y,x\rangle} = \overline{\sum c_i y_i\overline{x_i}} = \sum c_i \overline{y_i}\,x_i = \sum c_i x_i\overline{y_i} = \langle x,y\rangle, using that ci∈Rc_i\in\mathbb{R} so ci‾=ci\overline{c_i}=c_i, and yixi‾‾=yi‾ xi\overline{y_i\overline{x_i}}=\overline{y_i}\,x_i. Linearity in the first slot: Additivity and homogeneity in xx follow exactly as in (b), since cic_i are fixed constants that commute with all the manipulations: ⟨x+x′,y⟩=∑ci(xi+xi′)yi‾=⟨x,y⟩+⟨x′,y⟩\langle x+x',y\rangle = \sum c_i(x_i+x_i')\overline{y_i} = \langle x,y\rangle + \langle x',y\rangle, and ⟨λx,y⟩=∑ci(λxi)yi‾=λ⟨x,y⟩\langle \lambda x,y\rangle = \sum c_i(\lambda x_i)\overline{y_i} = \lambda\langle x,y\rangle. Positive-definiteness: ⟨x,x⟩=∑i=1nci∣xi∣2\langle x,x\rangle = \sum_{i=1}^n c_i|x_i|^2. Since ci>0c_i>0 and ∣xi∣2≥0|x_i|^2\ge 0, each term ci∣xi∣2≥0c_i|x_i|^2 \ge 0, so the sum is ≥0\ge0. If the sum is 00, then since every summand is ≥0\ge 0, each summand ci∣xi∣2=0c_i|x_i|^2=0; since ci>0c_i>0, this forces ∣xi∣2=0|x_i|^2=0, i.e. xi=0x_i=0, for every ii. This step is precisely where ci>0c_i>0 (strict positivity, not merely ci≥0c_i\ge 0) is used; the converse direction is immediate.

(d) L2L^2 inner product on C[a,b]C[a,b]. Conjugate symmetry: Since F=R\mathbb{F}=\mathbb{R}, conjugation is trivial, and ⟨f,g⟩=∫abf(t)g(t) dt=∫abg(t)f(t) dt=⟨g,f⟩\langle f,g\rangle = \int_a^b f(t)g(t)\,dt = \int_a^b g(t)f(t)\,dt = \langle g,f\rangle by commutativity of multiplication of real numbers, which holds pointwise and is preserved by the (linear) Riemann integral. Linearity in the first slot: For f,f′,g∈C[a,b]f,f',g\in C[a,b], both ff and f′f' (hence f+f′f+f') are continuous, so all integrals below exist. By linearity of the Riemann integral,

⟨f+f′,g⟩=∫ab(f(t)+f′(t))g(t) dt=∫abf(t)g(t) dt+∫abf′(t)g(t) dt=⟨f,g⟩+⟨f′,g⟩,\langle f+f',g\rangle = \int_a^b (f(t)+f'(t))g(t)\,dt = \int_a^b f(t)g(t)\,dt + \int_a^b f'(t)g(t)\,dt = \langle f,g\rangle + \langle f',g\rangle,

and for λ∈R\lambda \in \mathbb{R}, ⟨λf,g⟩=∫abλf(t)g(t) dt=λ∫abf(t)g(t) dt=λ⟨f,g⟩\langle \lambda f,g\rangle = \int_a^b \lambda f(t)g(t)\,dt = \lambda\int_a^b f(t)g(t)\,dt = \lambda \langle f,g\rangle. Positive-definiteness: ⟨f,f⟩=∫abf(t)2 dt\langle f,f\rangle = \int_a^b f(t)^2\,dt. The integrand f(t)2≥0f(t)^2 \ge 0 for every t∈[a,b]t\in[a,b], and the integral of a nonnegative continuous function over [a,b][a,b] (with a<ba<b) is ≥0\ge 0; this is a standard property of the Riemann integral (monotonicity: if h≥0h\ge 0 on [a,b][a,b] then ∫abh≥0\int_a^b h \ge 0, since every lower Riemann sum is ≥0\ge 0). If f=0f=0 (the zero function) then clearly ⟨f,f⟩=0\langle f,f\rangle = 0. Conversely, suppose ⟨f,f⟩=∫abf(t)2 dt=0\langle f,f\rangle = \int_a^b f(t)^2\,dt = 0 but f≠0f\neq 0; then f(t0)≠0f(t_0)\neq 0 for some t0∈[a,b]t_0\in[a,b], so f(t0)2=ε>0f(t_0)^2 = \varepsilon > 0. By continuity of ff (hence of f2f^2) at t0t_0, there is an interval [c,d]⊆[a,b][c,d]\subseteq[a,b] of positive length δ=d−c>0\delta = d-c>0 containing t0t_0 on which f(t)2≥ε/2f(t)^2 \ge \varepsilon/2. Then

∫abf(t)2 dt≥∫cdf(t)2 dt≥ε2δ>0,\int_a^b f(t)^2\,dt \ge \int_c^d f(t)^2\,dt \ge \frac{\varepsilon}{2}\delta > 0,

using monotonicity of the integral twice (dropping a nonnegative integrand outside [c,d][c,d] only decreases the total, and bounding f2≥ε/2f^2\ge \varepsilon/2 on [c,d][c,d] from below). This contradicts ⟨f,f⟩=0\langle f,f\rangle=0. Hence f=0f=0. ■\blacksquare

∎

Notice the proof technique used throughout: direct verification from the axioms, reducing each claim to elementary properties of real/complex arithmetic (commutativity, distributivity, zz‾=∣z∣2z\overline z=|z|^2) or to elementary properties of the Riemann integral (linearity, monotonicity). No result depending on inner products was used to prove that inner products exist — the argument is entirely bottom-up.

Worked examples

Example 1 — an asymmetric vector computation in C3\mathbb{C}^3. Let V=C3V=\mathbb{C}^3 with the standard Hermitian inner product ⟨x,y⟩=∑i=13xiyi‾\langle x,y\rangle = \sum_{i=1}^3 x_i\overline{y_i}. Take

x=(2+i,  −1,  3i),y=(1−i,  2i,  4).x = (2+i,\; -1,\; 3i), \qquad y = (1-i,\; 2i,\; 4).

Compute ⟨x,y⟩\langle x,y\rangle and ⟨y,x⟩\langle y,x\rangle, and check conjugate symmetry explicitly. First, y1‾=1−i‾=1+i\overline{y_1} = \overline{1-i} = 1+i, y2‾=2i‾=−2i\overline{y_2} = \overline{2i} = -2i, y3‾=4‾=4\overline{y_3} = \overline{4} = 4. Then

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

Compute each term: (2+i)(1+i)=2+2i+i+i2=2+3i−1=1+3i(2+i)(1+i) = 2+2i+i+i^2 = 2+3i-1 = 1+3i; (−1)(−2i)=2i(-1)(-2i) = 2i; (3i)(4)=12i(3i)(4)=12i. Summing,

⟨x,y⟩=(1+3i)+2i+12i=1+17i.\langle x,y\rangle = (1+3i) + 2i + 12i = 1 + 17i.

Now compute ⟨y,x⟩=∑yixi‾\langle y,x\rangle = \sum y_i\overline{x_i}. We have x1‾=2−i\overline{x_1}=2-i, x2‾=−1\overline{x_2}=-1, x3‾=−3i\overline{x_3}=-3i. So

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

Compute: (1−i)(2−i)=2−i−2i+i2=2−3i−1=1−3i(1-i)(2-i) = 2 -i -2i + i^2 = 2-3i-1 = 1-3i; (2i)(−1)=−2i(2i)(-1) = -2i; (4)(−3i)=−12i(4)(-3i) = -12i. Summing,

⟨y,x⟩=(1−3i)−2i−12i=1−17i.\langle y,x\rangle = (1-3i) - 2i - 12i = 1 - 17i.

Indeed ⟨y,x⟩‾=1−17i‾=1+17i=⟨x,y⟩\overline{\langle y,x\rangle} = \overline{1-17i} = 1+17i = \langle x,y\rangle, confirming conjugate symmetry numerically. Also, ⟨x,x⟩=∣2+i∣2+∣−1∣2+∣3i∣2=5+1+9=15\langle x,x\rangle = |2+i|^2+|-1|^2+|3i|^2 = 5+1+9=15, a nonnegative real number as guaranteed by positive-definiteness, even though the individual coordinates of xx are non-real.

Example 2 — a weighted inner product distinguishing directions in R2\mathbb{R}^2. Let V=R2V=\mathbb{R}^2 with the weighted inner product ⟨x,y⟩=4x1y1+x2y2\langle x,y\rangle = 4x_1y_1 + x_2y_2 (this is case (c) above with F=R\mathbb{F}=\mathbb{R}, c1=4c_1=4, c2=1c_2=1; the same proof of the theorem applies verbatim with the conjugates dropped). Take

u=(1,2),w=(3,−1).u = (1,2), \qquad w = (3,-1).

Compute ⟨u,w⟩\langle u,w\rangle:

⟨u,w⟩=4(1)(3)+(2)(−1)=12−2=10.\langle u,w\rangle = 4(1)(3) + (2)(-1) = 12 - 2 = 10.

Compare with the ordinary dot product, u⋅w=(1)(3)+(2)(−1)=3−2=1u\cdot w = (1)(3)+(2)(-1) = 3-2=1. The two numbers, 1010 versus 11, differ substantially: this weighted inner product is a genuinely different geometric structure on the same underlying vector space R2\mathbb{R}^2, one that "stretches" the first coordinate's contribution by a factor of 44. Next compute the "length" of uu under this inner product:

⟨u,u⟩=4(1)2+(2)2=4+4=8,\langle u,u\rangle = 4(1)^2+(2)^2 = 4+4=8,

which is positive since u≠0u\neq0, consistent with positive-definiteness; note 8≠8 \ne the ordinary squared Euclidean length 12+22=51^2+2^2=5, again illustrating that a different (but still valid) inner product genuinely changes the geometry, foreshadowing why the induced Norm Induced by an Inner Product can differ from the familiar Euclidean norm even on R2\mathbb{R}^2.

CounterexampleWhy positive-definiteness cannot be weakened to positive-semi-definiteness

Consider V=R2V=\mathbb{R}^2 and define

⟨x,y⟩∗:=x1y1\langle x,y\rangle_* := x_1y_1

(dropping the second coordinate entirely). This satisfies conjugate symmetry (trivial, since F=R\mathbb{F}=\mathbb{R}: x1y1=y1x1x_1y_1=y_1x_1) and linearity in the first argument (immediate, since it is literally the ordinary product of first coordinates, which is linear in x1x_1): ⟨x+x′,y⟩∗=(x1+x1′)y1=⟨x,y⟩∗+⟨x′,y⟩∗\langle x+x',y\rangle_* = (x_1+x_1')y_1 = \langle x,y\rangle_* + \langle x',y\rangle_* and ⟨λx,y⟩∗=(λx1)y1=λ⟨x,y⟩∗\langle \lambda x,y\rangle_* = (\lambda x_1)y_1 = \lambda\langle x,y\rangle_*. It even satisfies ⟨x,x⟩∗=x12≥0\langle x,x\rangle_* = x_1^2 \ge 0 for all xx, so it is positive semi-definite. But it fails the "only if v=0v=0" half of positive-definiteness: take v=(0,1)≠0v=(0,1)\neq 0. Then

⟨v,v⟩∗=02=0,\langle v,v\rangle_* = 0^2 = 0,

even though v≠0v\neq 0. So ⟨⋅,⋅⟩∗\langle\cdot,\cdot\rangle_* is not an inner product. The failure is not merely formal bookkeeping — it destroys the geometry the axiom is meant to guarantee. With this bilinear form, the nonzero vector (0,1)(0,1) would have to be assigned "length" ⟨v,v⟩∗=0\sqrt{\langle v,v\rangle_*}=0, so a nonzero vector could have zero length, and the eventual Cauchy-Schwarz Inequality and the notion of Orthogonality and Orthogonal Sets built on top of the inner product would both break down: for instance, (0,1)(0,1) would be "orthogonal" to every vector in R2\mathbb{R}^2, including itself, collapsing the useful distinctions the theory is built to make. This shows precisely why the strict inequality "⟨v,v⟩=0  ⟹  v=0\langle v,v\rangle=0 \implies v=0" — not just ⟨v,v⟩≥0\langle v,v\rangle \ge 0 — is an indispensable part of the definition.

Scope, edge cases & common traps

SituationWhat to checkCommon mistake
Real vector spaceConjugate symmetry reduces to plain symmetry ⟨v,w⟩=⟨w,v⟩\langle v,w\rangle = \langle w,v\rangleWriting an unnecessary conjugate bar over real numbers
Complex vector spaceSecond-slot conjugate-linearity is a consequence, not a separate axiom to impose from scratchAssuming the inner product is bilinear (linear in both slots) over C\mathbb{C}
Weighted sums ∑cixiyi‾\sum c_i x_i\overline{y_i}Need ci>0c_i > 0 strictly, for every iiAllowing some ci=0c_i = 0, which destroys positive-definiteness (Example: the counterexample above is exactly this failure with c1=1,c2=0c_1=1, c_2=0)
Function spaces like C[a,b]C[a,b]Continuity of ff (or at least Riemann integrability with the "vanishing integral of a nonnegative function" argument) is what makes positive-definiteness provableWorking with merely integrable (not continuous) functions, where ∫f2=0\int f^2 = 0 does not imply f≡0f\equiv 0 pointwise (only "almost everywhere"), so positive-definiteness can genuinely fail on that larger space unless one passes to equivalence classes
Infinite-dimensional spacesThe three axioms are stated exactly the same way; no extra completeness assumption is needed to have an inner productConfusing "inner product space" with "Hilbert space" (the latter additionally requires completeness in the induced norm)

The mathematician behind it

Origins · 1 minHermann Weyl and how this was discovered
HW
Hermann Weyl
1885–1955 · early 20th-century mathematical physics and Hilbert-space theory

Hermann Weyl was a German mathematician whose work spanned representation theory, differential geometry, mathematical logic, and the foundations of quantum mechanics. While the axiomatic notion of an inner product space (and, in its completed form, a Hilbert space) is usually credited jointly to David Hilbert's school and to John von Neumann's later axiomatization, Weyl played a decisive role in popularizing and applying the Hermitian inner-product formalism — precisely the sesquilinear structure at the heart of this article — to quantum mechanics, where state vectors live in a complex inner product space and physical observables correspond to operators compatible with that inner product. His 1928 book Gruppentheorie und Quantenmechanik helped cement the complex Hermitian inner product as indispensable machinery in physics, not merely an algebraic curiosity.

History
- 1844: Hermann Grassmann develops an early, largely algebraic notion of an inner-product-like pairing in his Ausdehnungslehre.
- 1901–1910: David Hilbert and Erhard Schmidt study ℓ2\ell^2 sequence spaces and integral equations, implicitly using the inner product now named after Hilbert.
- 1928: Hermann Weyl's Gruppentheorie und Quantenmechanik cements the complex Hermitian inner product as the language of quantum state spaces.
- 1932: John von Neumann gives the modern axiomatic definition of a Hilbert space in Mathematische Grundlagen der Quantenmechanik, formalizing the inner product axioms essentially as stated today.
- 1940s–present: the abstract finite-dimensional theory (real and complex inner product spaces) becomes a standard chapter of linear algebra curricula, exactly as presented here.
Is every bilinear (or sesquilinear) form on a vector space an inner product?

No. Bilinearity/sesquilinearity is only one of the three required axioms. A form can be perfectly bilinear yet fail conjugate symmetry (e.g., ⟨x,y⟩=x1y2\langle x,y\rangle = x_1y_2 on R2\mathbb{R}^2 is bilinear but not symmetric, since swapping xx and yy generally changes the value), or it can be symmetric and bilinear yet fail positive-definiteness (e.g., ⟨x,y⟩=x1y1−x2y2\langle x,y\rangle = x_1y_1 - x_2y_2, which gives ⟨(0,1),(0,1)⟩=−1<0\langle(0,1),(0,1)\rangle = -1 < 0). Such forms are still useful — they're called symmetric or Hermitian bilinear/sesquilinear forms, and appear in relativity and indefinite geometry — but they are not inner products unless all three axioms hold.

Why does the definition require linearity in the first slot specifically, rather than the second?

This is purely a convention, and different textbooks choose oppositely. Axler's convention (adopted here) puts linearity in the first argument and conjugate-linearity in the second; some physics texts (following Dirac's bra-ket notation) do the reverse, with the inner product conjugate-linear in the first slot. Both conventions describe the same underlying geometric structure; formulas for norms, Cauchy–Schwarz, and orthogonality carry over with only the placement of conjugate bars changing. The important invariant is that exactly one slot is linear and the other is conjugate-linear (over C\mathbb{C}), tied together by conjugate symmetry.

Does every vector space admit an inner product?

Every finite-dimensional vector space over R\mathbb{R} or C\mathbb{C} admits at least one inner product: choose any basis, declare it "orthonormal," and transport the standard Hermitian/Euclidean inner product from coordinates. For infinite-dimensional vector spaces the situation is more delicate; using a Hamel basis (which exists by the axiom of choice for any vector space) the same coordinate transport still produces a valid inner product, so in this abstract sense every vector space over R\mathbb{R} or C\mathbb{C} admits one. What is not guaranteed is that a specific, naturally arising bilinear/sesquilinear pairing on a given space of interest (say, all real sequences) is automatically an inner product; that has to be checked axiom by axiom, exactly as done in the theorem above.

What goes wrong if we only require ⟨v,v⟩≥0\langle v,v\rangle \geq 0 without the "only if v=0v=0" clause?

You get a positive semi-definite sesquilinear/bilinear form rather than an inner product, and the counterexample above shows this is a real distinction, not a technicality. Such a form still defines a "seminorm" ⟨v,v⟩\sqrt{\langle v,v\rangle}, but this seminorm can vanish on nonzero vectors, so it cannot detect all differences between vectors. Geometrically, "distance" collapses along an entire subspace, and orthogonality relations become degenerate (a nonzero vector can be orthogonal to itself). Many results downstream — Cauchy–Schwarz still holds in a weakened form, but the strict triangle inequality and unique orthogonal decompositions typically fail.

Can an inner product take negative values?

Yes, for two different vectors u≠vu\neq v; only ⟨v,v⟩\langle v,v\rangle (a vector paired with itself) is required to be nonnegative. For instance, with the standard dot product on R2\mathbb{R}^2, ⟨(1,0),(−1,0)⟩=−1\langle(1,0),(-1,0)\rangle = -1, a negative number, reflecting that these two vectors point in opposite directions. Negative inner products between distinct vectors are exactly what will later correspond to obtuse angles once the cosine formula is established via The Cauchy-Schwarz Inequality.

How does the inner product axiom system relate to norms and orthogonality mentioned at the end of the article?

The inner product is the primitive notion; norms and orthogonality are defined in terms of it, not vice versa. Once ⟨⋅,⋅⟩\langle\cdot,\cdot\rangle is fixed, one defines the induced norm ∥v∥:=⟨v,v⟩\|v\| := \sqrt{\langle v,v\rangle} (well-defined and real by positive-definiteness) and declares u,vu,v orthogonal when ⟨u,v⟩=0\langle u,v\rangle = 0. All of the further structure explored in Norm Induced by an Inner Product and Orthogonality and Orthogonal Sets — the Pythagorean theorem, orthogonal projections, orthonormal bases — is logically downstream of the three axioms proved and used in this article, and nowhere circularly assumes them in advance.

Why does the definition insist F\mathbb{F} be R\mathbb{R} or C\mathbb{C}, rather than an arbitrary field?

Because axiom 1 (conjugate symmetry) and axiom 3 (positive-definiteness, which requires an ordering to make sense of "≥0\ge 0") both rely on structure that a general field need not have. R\mathbb{R} has a natural total order compatible with its arithmetic, letting "⟨v,v⟩≥0\langle v,v\rangle \ge 0" mean something; C\mathbb{C} borrows this order via the real numbers together with a conjugation automorphism z↦z‾z\mapsto\overline z fixing R\mathbb{R} pointwise. A general field, such as a finite field Fp\mathbb{F}_p, has neither a compatible order nor typically a conjugation automorphism, so "positive-definite Hermitian form" does not translate; analogous but different structures (nondegenerate bilinear/quadratic forms) are used there instead.

Where this leads