Inner Product Space

The four axioms

An inner product space is a pair (V, ⟨·,·⟩) where V is a vector space over the field 𝔽 (either ℝ or ℂ) and ⟨·,·⟩ : V × V → 𝔽 is a map satisfying four axioms for all u, v, w ∈ V and all scalars α ∈ 𝔽:

1. Linearity in the first argument — ⟨αu + w, v⟩ = α⟨u, v⟩ + ⟨w, v⟩.
2. Conjugate symmetry — ⟨u, v⟩ = ⟨v, u⟩. Over ℝ the conjugate vanishes and the pairing is plainly symmetric; over ℂ the conjugate is essential.
3. Positive-definiteness — ⟨v, v⟩ ≥ 0, with equality iff v = 0. This is what guarantees a real, non-negative length.
4. Non-degeneracy — implied by positive-definiteness above; if ⟨v, w⟩ = 0 for all w, then v = 0.

Combining axioms 1 and 2 gives conjugate-linearity in the second slot: ⟨v, αw⟩ = ᾱ⟨v, w⟩. The pairing is then called sesquilinear over ℂ. Over ℝ it is plainly bilinear.

From algebra to geometry — what the pairing buys you

The four axioms look modest. Their consequences are not. Once you have an inner product, you immediately recover the entire toolbox of Euclidean geometry:

• Length — define ‖v‖ ≔ √⟨v, v⟩. Positive-definiteness makes the square root real and the length zero only at the zero vector.
• Angle — for non-zero u, v in a real inner product space, the Cauchy-Schwarz inequality bounds ⟨u, v⟩ / (‖u‖·‖v‖) within [−1, 1], so we can define θ ∈ [0, π] by cos θ = ⟨u, v⟩ / (‖u‖·‖v‖).
• Orthogonality — write u ⊥ v when ⟨u, v⟩ = 0. The Pythagorean theorem follows directly: ‖u + v‖² = ‖u‖² + ‖v‖² whenever u ⊥ v.
• Projection — the projection of u onto span(v) is proj_v(u) = (⟨u, v⟩ / ⟨v, v⟩) v. Subtracting it gives the orthogonal residual.
• Orthonormal bases — a basis e₁, …, eₙ with ⟨eᵢ, eⱼ⟩ = δᵢⱼ; coordinates become inner products: v = Σ ⟨v, eᵢ⟩ eᵢ.

Proof of the Cauchy-Schwarz inequality

For all u, v in an inner product space, |⟨u, v⟩| ≤ ‖u‖ · ‖v‖, with equality iff u, v are linearly dependent.

Proof (real case). If v = 0 the inequality reads 0 ≤ 0; assume v ≠ 0. Define the function f(t) = ⟨u − tv, u − tv⟩ for t ∈ ℝ. By positive-definiteness, f(t) ≥ 0 for every t. Expanding:

f(t) = ‖u‖² − 2t⟨u, v⟩ + t²‖v‖² ≥ 0.

This is a quadratic in t with leading coefficient ‖v‖² > 0 that is everywhere non-negative. A non-negative quadratic has non-positive discriminant:

Δ = (2⟨u, v⟩)² − 4‖v‖²‖u‖² ≤ 0 ⟹ ⟨u, v⟩² ≤ ‖u‖²‖v‖².

Take square roots and you have |⟨u, v⟩| ≤ ‖u‖·‖v‖. Equality forces Δ = 0, meaning the quadratic has a (double) root t₀ where u − t₀v = 0, so u and v are linearly dependent. The complex case substitutes t = ⟨u, v⟩ / ‖v‖² and runs the same argument with conjugates.

The triangle inequality, almost for free

From Cauchy-Schwarz, the induced norm automatically satisfies the triangle inequality. Compute:

‖u + v‖² = ⟨u + v, u + v⟩ = ‖u‖² + 2 Re⟨u, v⟩ + ‖v‖²
≤ ‖u‖² + 2|⟨u, v⟩| + ‖v‖²
≤ ‖u‖² + 2‖u‖·‖v‖ + ‖v‖²
= (‖u‖ + ‖v‖)².

Square-rooting: ‖u + v‖ ≤ ‖u‖ + ‖v‖. So an inner product not only gives a length, it gives a length that obeys the geometric intuition that detours are longer than direct paths.

Inner product vs normed vs metric spaces

	Inner product space	Normed space	Metric space
Primitive structure	⟨u, v⟩ ∈ 𝔽	‖v‖ ∈ ℝ≥0	d(x, y) ∈ ℝ≥0
Defines lengths	Yes (‖v‖ = √⟨v,v⟩)	Yes (primitive)	Indirectly via d(x, 0)
Defines angles	Yes (cos θ = ⟨u,v⟩/(‖u‖‖v‖))	No	No
Defines orthogonality	Yes (⟨u, v⟩ = 0)	No general notion	No
Parallelogram law holds	Always	Iff norm comes from an inner product	N/A
Required vector-space structure	Yes	Yes	No (any set will do)
Strength ordering	Strongest (richest)	Middle	Weakest (most general)
Canonical complete example	L²(ℝ), ℓ²	L^p, ℓ^p (p ≠ 2)	(ℝ, |·|)

The hierarchy is strict: every inner product space is a normed space (use the induced norm), and every normed space is a metric space (use d(x, y) = ‖x − y‖). Going the other way fails — see the next section.

ℓᵖ vs Lᵖ norms — which come from an inner product?

Norm	Formula	Comes from inner product?	Why / why not
ℓ¹	Σ |xᵢ|	No	Fails parallelogram law: with x = (1,0), y = (0,1) the LHS is 4 but RHS is 4 — okay; try x = (1,1), y = (1,−1): LHS = 4 + 4 = 8, RHS = 2·4 + 2·4 = 16.
ℓ²	√Σ |xᵢ|²	Yes	Inner product is the standard dot product Σ xᵢ ȳᵢ. The unique p for which ℓᵖ is Hilbert.
ℓᵖ (1 < p < ∞, p ≠ 2)	(Σ |xᵢ|ᵖ)^{1/p}	No	Parallelogram law fails for any p ≠ 2; ℓᵖ is uniformly convex but not Hilbert.
ℓ^∞	sup |xᵢ|	No	Sphere is a cube, not round; parallelogram law fails dramatically.
L¹([a,b])	∫|f|	No	Same parallelogram failure as ℓ¹.
L²([a,b])	√∫|f|²	Yes	Inner product ⟨f, g⟩ = ∫ f(x) g(x) dx. The setting for Fourier series.
Lᵖ([a,b]) (p ≠ 2)	(∫|f|ᵖ)^{1/p}	No	Banach but not Hilbert.

Jordan-von Neumann theorem (1935). A norm comes from an inner product if and only if it satisfies the parallelogram law ‖u+v‖² + ‖u−v‖² = 2‖u‖² + 2‖v‖². This is the cleanest test: any single counter-example pair (u, v) rules out the inner product origin instantly.

Canonical examples

• ℝⁿ with the dot product. ⟨u, v⟩ = u₁v₁ + … + uₙvₙ. The induced norm is the Euclidean length and angles match high school trigonometry.
• ℂⁿ with the Hermitian dot product. ⟨u, v⟩ = Σ uᵢ vᵢ. The conjugate is essential — without it ⟨v, v⟩ wouldn't be real.
• ℓ², the space of square-summable sequences. All real (or complex) sequences (xₙ) with Σ|xₙ|² < ∞, with ⟨x, y⟩ = Σ xₙ yₙ. The infinite-dimensional Hilbert space.
• L²([a, b]). Square-integrable functions with ⟨f, g⟩ = ∫ₐᵇ f(x) g(x) dx. The orthonormal system {1, cos x, sin x, cos 2x, sin 2x, …} (suitably normalized) gives the Fourier expansion of any L² function.
• Matrices with the Frobenius inner product. ⟨A, B⟩ = tr(A* B) = Σᵢⱼ Aᵢⱼ Bᵢⱼ. The induced norm ‖A‖_F is the entrywise ℓ² norm.
• Polynomials of degree ≤ n with ⟨p, q⟩ = ∫₋₁¹ p(x) q(x) dx. Gram-Schmidt on 1, x, x², … gives the Legendre polynomials.

Why it matters in physics, statistics, and analysis

• Quantum mechanics. The state space of a quantum system is a complex Hilbert space ℋ. A unit vector |ψ⟩ ∈ ℋ encodes a state; the inner product ⟨ψ|φ⟩ encodes a probability amplitude; |⟨ψ|φ⟩|² is the probability that measuring system in state |ψ⟩ produces outcome φ. The Pauli, Schrödinger, and Heisenberg pictures are all built on this single primitive.
• Fourier analysis. Decomposing a periodic function into sines and cosines is exactly the projection of f ∈ L²([0, 2π]) onto an orthonormal basis. The Fourier coefficient ĉₙ = ⟨f, eⁱⁿˣ⟩/(2π) is an inner product.
• Statistics and least squares. Linear regression minimizes ‖y − Xβ‖² in an inner product space; the optimal β̂ is the orthogonal projection of y onto the column space of X. Correlation coefficients are normalized inner products of mean-centered variables.
• Machine learning. The kernel trick replaces explicit feature maps φ(x) with kernels K(x, y) = ⟨φ(x), φ(y)⟩, computing inner products in (sometimes infinite-dimensional) feature spaces without ever materializing the features.
• Numerical analysis. Conjugate-gradient and Krylov methods rely on A-orthogonality: ⟨u, v⟩_A = uᵀAv with A symmetric positive-definite.

The parallelogram and polarization identities

Two identities tie norms to inner products. For any inner product space:

Parallelogram law: ‖u + v‖² + ‖u − v‖² = 2‖u‖² + 2‖v‖².

Polarization (real case): ⟨u, v⟩ = ¼ (‖u + v‖² − ‖u − v‖²).

Polarization (complex case): ⟨u, v⟩ = ¼ Σ_{k=0}^{3} iᵏ ‖u + iᵏv‖².

The parallelogram identity is geometric: in a flat Euclidean plane, the sum of the squared diagonals of a parallelogram equals twice the sum of the squared sides. Polarization is striking: it says the inner product is fully recoverable from the norm. So the norm carries all the inner-product data, but only when the parallelogram law holds.

Bessel's inequality and Parseval's identity

Suppose {e₁, e₂, …} is an orthonormal sequence in an inner product space V. For any v ∈ V:

Bessel: Σₙ |⟨v, eₙ⟩|² ≤ ‖v‖².

That is, the energy in the projections never exceeds the energy of v. If the orthonormal system is also complete (its span is dense in V), Bessel becomes an equality:

Parseval: Σₙ |⟨v, eₙ⟩|² = ‖v‖².

Parseval is the Pythagorean theorem in infinite dimensions and the conservation-of-energy statement of Fourier analysis: the L² norm of a signal equals the ℓ² norm of its Fourier coefficients.

Common mistakes

• Assuming every norm comes from an inner product. Only the ℓ² and L² families do. The taxicab (ℓ¹) and supremum (ℓ^∞) norms are perfectly valid norms but generate no inner product, no angles, no orthogonality.
• Forgetting the conjugate over ℂ. Without it, ⟨v, v⟩ is generally complex and you can't take a real square root. The conjugate is what makes positive-definiteness make sense over ℂ.
• Treating "inner product space" and "Hilbert space" as synonyms. Every Hilbert space is an inner product space, but the converse fails in infinite dimensions: completeness is an extra hypothesis, and most natural pre-Hilbert constructions (continuous functions with the L² inner product) are not complete.
• Using non-symmetric bilinear forms as inner products. The Minkowski form ⟨u, v⟩_M = u₀v₀ − u₁v₁ − u₂v₂ − u₃v₃ on Minkowski spacetime is symmetric but indefinite (some non-zero vectors have ⟨v, v⟩ ≤ 0). It is an "inner product" in the indefinite-form sense of relativity, not the positive-definite sense of this article.
• Forgetting positive-definiteness when constructing weighted inner products. uᵀAv is an inner product iff A is symmetric and positive-definite. If A has a zero or negative eigenvalue, you get a degenerate or indefinite form, not an inner product.