Bessel and Parseval Identities

How much of a vector can you capture using only part of an orthonormal system? This is the concrete question behind Bessel's inequality. If you measure a vector against finitely or infinitely many mutually orthonormal directions, something follows. The sum of the squared measurements can never exceed the vector's own squared length. It sounds almost obvious once stated. Yet it is the technical engine behind Fourier series, least-squares approximation, and the very possibility of infinite-dimensional orthonormal expansions converging at all. The idea matters because it separates two regimes cleanly. When an orthonormal family is merely partial — it does not "see" every direction in the space — Bessel's inequality gives a strict accounting. Some of the vector's length is lost, unseen by the family. When the family is complete, meaning it spans the whole space (or is dense enough in an infinite-dimensional setting), that lost length vanishes entirely. Bessel's inequality then sharpens into an equality: Parseval's identity. This equality is a Pythagorean theorem in disguise. It asserts that a vector's squared length equals the sum of squares of its coordinates in an orthonormal basis, no more and no less. The big picture is that these two identities form a single continuum. Bessel's inequality holds for any orthonormal set, finite or infinite, with no further assumptions. It is unconditionally true, and its proof is a two-line algebraic identity. Parseval's identity is what happens when Bessel's inequality becomes tight. Completeness of the orthonormal family is exactly the missing ingredient. It forces the "leftover" term to be zero. Understanding why the two are logically the same computation, differing only by one hypothesis, illuminates something important. It shows how orthonormal bases let us treat vectors as literally interchangeable with their coordinate lists. This article gives the precise statements and complete proofs from first principles. These proofs use only the inner product axioms and orthogonal projection algebra. It also offers worked numerical examples in both finite- and infinite-dimensional settings. It gives a careful account of what fails when completeness is dropped. It closes by connecting these identities to their role in Orthogonal Projections. There, Bessel's inequality reappears as the statement that a projection can only shrink length.

Key takeaways
  • Bessel's inequality: for any orthonormal family (ei)i∈I(e_i)_{i\in I} in an inner product space VV and any v∈Vv \in V, the sum ∑i∣⟨v,ei⟩∣2≤∥v∥2\sum_i |\langle v, e_i\rangle|^2 \le \|v\|^2, always.
  • The proof is a single computation: expand ∥v−∑⟨v,ei⟩ei∥2≥0\big\|v - \sum \langle v,e_i\rangle e_i\big\|^2 \ge 0 and simplify using orthonormality.
  • Parseval's identity is Bessel's inequality upgraded to equality, valid precisely when (ei)(e_i) is a complete orthonormal system (an orthonormal basis).
  • In finite dimensions, completeness just means the family spans VV; the count of vectors equals dim⁡V\dim V.
  • Parseval's identity is a direct generalization of the Pythagorean theorem to arbitrarily many mutually perpendicular directions.
  • Key idea: the gap between ∥v∥2\|v\|^2 and ∑i∣⟨v,ei⟩∣2\sum_i |\langle v,e_i\rangle|^2 measures exactly the squared length of the part of vv missed by the orthonormal family; completeness forces that gap to zero.

Where this lives today

Bessel's inequality and Parseval's identity are the algebraic backbone of Fourier analysis: every statement about how much energy a signal's Fourier coefficients carry, and whether truncating a Fourier series loses information, is a direct instance of these two results applied to the orthonormal family {einx}\{e^{inx}\} or {sin⁡(nx),cos⁡(nx)}\{\sin(nx),\cos(nx)\}. In signal processing and information theory, Parseval's identity underlies the equivalence of energy computed in the time domain and in the frequency domain, which justifies designing filters and compression schemes (such as JPEG's use of the discrete cosine transform) directly in frequency space without loss of total signal energy. In numerical analysis and approximation theory, Bessel's inequality quantifies the error of truncated expansions in orthogonal polynomial bases (Legendre, Chebyshev, Hermite), giving a priori bounds on how well a finite-term approximation can perform. In quantum mechanics, state vectors expanded in a complete orthonormal basis of eigenstates obey Parseval's identity as the statement that total probability sums to one; Bessel's inequality governs the probability captured by an incomplete measurement basis. In statistics and machine learning, principal component analysis and least-squares regression both rest on the same underlying identity: projecting data onto an orthonormal set of directions never increases total variance captured, with equality exactly when the directions span the full data space — directly foreshadowing the theory of Orthogonal Projections.

Ask the tutor: How does the proof of Bessel's inequality change if the orthonormal family is replaced by a merely orthogonal (not unit-length) family?

DefinitionOrthonormal family, Bessel's inequality, completeness, Parseval's identity

Let VV be an inner product space over F∈{R,C}\mathbb{F} \in \{\mathbb{R}, \mathbb{C}\}, with inner product ⟨⋅,⋅⟩\langle \cdot,\cdot\rangle (linear in the first slot, conjugate-linear in the second) and induced norm ∥v∥=⟨v,v⟩\|v\| = \sqrt{\langle v,v\rangle}. A family (ei)i∈I(e_i)_{i \in I} of vectors in VV, indexed by an arbitrary (possibly infinite) index set II, is orthonormal if

⟨ei,ej⟩=δij={1i=j0i≠j\langle e_i, e_j \rangle = \delta_{ij} = \begin{cases} 1 & i = j \\ 0 & i \ne j \end{cases}

for all i,j∈Ii, j \in I. For v∈Vv \in V and an orthonormal family (ei)i∈I(e_i)_{i\in I}, the scalars v^i:=⟨v,ei⟩\hat v_i := \langle v, e_i \rangle are called the Fourier coefficients of vv with respect to (ei)i∈I(e_i)_{i\in I}.

Bessel's inequality. For every finite subset F⊆IF \subseteq I,

∑i∈F∣⟨v,ei⟩∣2  ≤  ∥v∥2.\sum_{i \in F} |\langle v, e_i\rangle|^2 \;\le\; \|v\|^2.

Consequently the family of nonnegative reals (∣⟨v,ei⟩∣2)i∈I(|\langle v,e_i\rangle|^2)_{i \in I} is summable (its net of finite partial sums is bounded), and

∑i∈I∣⟨v,ei⟩∣2  ≤  ∥v∥2,\sum_{i \in I} |\langle v, e_i \rangle|^2 \;\le\; \|v\|^2,

where the left side denotes the supremum over finite partial sums. An orthonormal family (ei)i∈I(e_i)_{i\in I} is complete (an orthonormal basis of VV) if the only vector orthogonal to every eie_i is 00; equivalently, in finite dimensions, if span⁡{ei:i∈I}=V\operatorname{span}\{e_i : i \in I\} = V.

Parseval's identity. If (ei)i∈I(e_i)_{i \in I} is a complete orthonormal family in a finite-dimensional inner product space VV (or more generally an orthonormal basis of a Hilbert space VV), then for every v∈Vv \in V,

∑i∈I∣⟨v,ei⟩∣2  =  ∥v∥2,\sum_{i \in I} |\langle v, e_i \rangle|^2 \;=\; \|v\|^2,

and moreover v=∑i∈I⟨v,ei⟩ eiv = \sum_{i\in I} \langle v, e_i\rangle\, e_i (with the sum interpreted as a limit of finite partial sums when II is infinite).

Intuition

Picture V=R3V = \mathbb{R}^3 with the standard basis vectors, but suppose you are handed only two of the three axes — say e1e_1 and e2e_2 — and asked to describe a vector vv using only those. You compute the two "shadows" ⟨v,e1⟩\langle v, e_1\rangle and ⟨v,e2⟩\langle v, e_2 \rangle, square and add them. Bessel's inequality says this sum never exceeds ∥v∥2\|v\|^2, because you are only measuring the part of vv lying in the e1e_1-e2e_2 plane; whatever component of vv sticks out along the missing third direction is invisible to your measurement but still contributes to ∥v∥2\|v\|^2. The shortfall ∥v∥2−∑∣⟨v,ei⟩∣2\|v\|^2 - \sum |\langle v,e_i\rangle|^2 is exactly the squared length of that invisible leftover piece. Now suppose you are handed all three axes. Nothing is invisible anymore: every component of vv is accounted for, the leftover piece is the zero vector, and the inequality becomes the exact equality of Parseval's identity — precisely the three-dimensional Pythagorean theorem, ∥v∥2=v12+v22+v32\|v\|^2 = v_1^2+v_2^2+v_3^2. The transition from inequality to equality is not a new phenomenon; it is the same computation with the "missing direction" term forced to vanish because there is no missing direction left.

Figure
Bessel's inequality as a shadow with leftover length
e₁ e₂ v projection: Σ⟨v,eᵢ⟩eᵢ leftover, length² = ‖v‖² − Σ|⟨v,eᵢ⟩|² Bessel: shadow² ≤ ‖v‖² (equality iff nothing left over)

Central results

TheoremBessel's Inequality

Let VV be an inner product space over F∈{R,C}\mathbb{F} \in \{\mathbb{R},\mathbb{C}\}, let (e1,…,en)(e_1,\dots,e_n) be a finite orthonormal list in VV, and let v∈Vv \in V be arbitrary. Then

∑k=1n∣⟨v,ek⟩∣2  ≤  ∥v∥2,\sum_{k=1}^n |\langle v, e_k\rangle|^2 \;\le\; \|v\|^2,

with equality if and only if v∈span⁡(e1,…,en)v \in \operatorname{span}(e_1,\dots,e_n). More generally, if (ei)i∈I(e_i)_{i \in I} is an orthonormal family indexed by an arbitrary set II, then for every finite F⊆IF \subseteq I,

∑i∈F∣⟨v,ei⟩∣2≤∥v∥2,\sum_{i \in F} |\langle v, e_i \rangle|^2 \le \|v\|^2,

so the nonnegative reals {∣⟨v,ei⟩∣2}i∈I\{|\langle v, e_i\rangle|^2\}_{i \in I} have bounded finite partial sums, hence at most countably many are nonzero, and

∑i∈I∣⟨v,ei⟩∣2≤∥v∥2.\sum_{i \in I} |\langle v, e_i\rangle|^2 \le \|v\|^2.
Proof

It suffices to prove the finite-list statement, since the general statement follows by applying it to each finite subset F⊆IF \subseteq I (relabeled as e1,…,ene_1,\dots,e_n) and then taking the supremum over FF; a bounded increasing net of finite sums of nonnegative terms has at most countably many nonzero terms and a well-defined supremum equal to the sum, which is then bounded by ∥v∥2\|v\|^2.

Technique: direct algebraic expansion using orthonormality (no circularity — uses only the inner product axioms). Fix v∈Vv \in V and an orthonormal list e1,…,ene_1,\dots,e_n. Define

w:=v−∑k=1n⟨v,ek⟩ek.w := v - \sum_{k=1}^n \langle v, e_k \rangle e_k.

This ww is simply the difference between vv and its "orthogonal projection guess" onto span⁡(e1,…,en)\operatorname{span}(e_1,\dots,e_n); we verify algebraically that it is genuinely orthogonal to each eje_j, without presupposing any projection theorem. Step 1: w⊥ejw \perp e_j for each j=1,…,nj = 1,\dots,n. Using linearity of the inner product in the first argument and conjugate-linearity in the second,

⟨w,ej⟩=⟨v,ej⟩−∑k=1n⟨v,ek⟩⟨ek,ej⟩.\langle w, e_j \rangle = \langle v, e_j\rangle - \sum_{k=1}^n \langle v,e_k\rangle \langle e_k, e_j \rangle.

By orthonormality, ⟨ek,ej⟩=δkj\langle e_k, e_j\rangle = \delta_{kj}, so the sum collapses to the single term k=jk=j:

⟨w,ej⟩=⟨v,ej⟩−⟨v,ej⟩⋅1=0.\langle w, e_j \rangle = \langle v, e_j \rangle - \langle v, e_j\rangle \cdot 1 = 0.

Step 2: expand ∥v∥2\|v\|^2 using v=w+∑k⟨v,ek⟩ekv = w + \sum_k \langle v,e_k\rangle e_k. Write p:=∑k=1n⟨v,ek⟩ekp := \sum_{k=1}^n \langle v, e_k\rangle e_k, so v=w+pv = w + p. Then

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

Since pp is a linear combination of e1,…,ene_1,\dots,e_n and ⟨w,ej⟩=0\langle w, e_j\rangle = 0 for every jj (Step 1), linearity in the first slot gives ⟨p,w⟩=∑k⟨v,ek⟩‾⟨ek,w⟩\langle p, w\rangle = \sum_k \overline{\langle v,e_k\rangle}\langle e_k, w\rangle; using conjugate symmetry ⟨ek,w⟩=⟨w,ek⟩‾=0\langle e_k,w\rangle = \overline{\langle w,e_k\rangle} = 0, both cross terms vanish:

⟨w,p⟩=⟨p,w⟩‾=0,⟨p,w⟩=0.\langle w, p\rangle = \overline{\langle p, w \rangle} = 0, \qquad \langle p, w\rangle = 0.

Hence

∥v∥2=∥w∥2+∥p∥2.\|v\|^2 = \|w\|^2 + \|p\|^2.

Step 3: compute ∥p∥2\|p\|^2 using orthonormality.

∥p∥2=⟨∑k=1n⟨v,ek⟩ek, ∑j=1n⟨v,ej⟩ej⟩=∑k=1n∑j=1n⟨v,ek⟩⟨v,ej⟩‾ ⟨ek,ej⟩.\|p\|^2 = \Big\langle \sum_{k=1}^n \langle v,e_k\rangle e_k,\ \sum_{j=1}^n \langle v,e_j\rangle e_j \Big\rangle = \sum_{k=1}^n\sum_{j=1}^n \langle v,e_k\rangle \overline{\langle v,e_j\rangle}\, \langle e_k, e_j\rangle.

Orthonormality kills every term with k≠jk \ne j, leaving

∥p∥2=∑k=1n⟨v,ek⟩⟨v,ek⟩‾=∑k=1n∣⟨v,ek⟩∣2.\|p\|^2 = \sum_{k=1}^n \langle v,e_k\rangle \overline{\langle v,e_k\rangle} = \sum_{k=1}^n |\langle v,e_k\rangle|^2.

Step 4: conclude. Combining Steps 2 and 3,

∥v∥2=∥w∥2+∑k=1n∣⟨v,ek⟩∣2.\|v\|^2 = \|w\|^2 + \sum_{k=1}^n |\langle v,e_k\rangle|^2.

Since ∥w∥2≥0\|w\|^2 \ge 0 by positive-definiteness of the inner product, we get

∑k=1n∣⟨v,ek⟩∣2≤∥v∥2.\sum_{k=1}^n |\langle v,e_k\rangle|^2 \le \|v\|^2.

This proves the inequality. Equality case. Equality holds if and only if ∥w∥2=0\|w\|^2 = 0, which by positive-definiteness of the inner product holds if and only if w=0w = 0, i.e. v=∑k=1n⟨v,ek⟩ek∈span⁡(e1,…,en)v = \sum_{k=1}^n \langle v,e_k\rangle e_k \in \operatorname{span}(e_1,\dots,e_n). Conversely, if v∈span⁡(e1,…,en)v \in \operatorname{span}(e_1,\dots,e_n), write v=∑kckekv = \sum_k c_k e_k; taking the inner product with eje_j and using orthonormality gives ⟨v,ej⟩=cj\langle v, e_j\rangle = c_j, so v=∑k⟨v,ek⟩ekv = \sum_k \langle v,e_k\rangle e_k, i.e. w=0w=0, giving equality. ■\blacksquare

∎
CorollaryBessel's inequality for infinite orthonormal sequences

If (en)n=1∞(e_n)_{n=1}^\infty is an orthonormal sequence in an inner product space VV and v∈Vv \in V, then the series ∑n=1∞∣⟨v,en⟩∣2\sum_{n=1}^\infty |\langle v, e_n \rangle|^2 converges and

∑n=1∞∣⟨v,en⟩∣2≤∥v∥2.\sum_{n=1}^\infty |\langle v, e_n\rangle|^2 \le \|v\|^2.
Proof

By Bessel's Inequality applied to the finite list e1,…,eNe_1,\dots,e_N for each NN, the partial sums SN:=∑n=1N∣⟨v,en⟩∣2S_N := \sum_{n=1}^N |\langle v,e_n\rangle|^2 satisfy SN≤∥v∥2S_N \le \|v\|^2 for every NN. The sequence (SN)(S_N) is nondecreasing (each term adds a nonnegative summand) and bounded above by ∥v∥2\|v\|^2, so by the monotone convergence theorem for real sequences it converges to a limit S≤∥v∥2S \le \|v\|^2. This limit is by definition the sum of the series. ■\blacksquare

∎
TheoremParseval's Identity

Let VV be a finite-dimensional inner product space with dim⁡V=n\dim V = n, and let (e1,…,en)(e_1,\dots,e_n) be an orthonormal basis of VV (equivalently, a complete orthonormal family in the sense of the definition above). Then for every v∈Vv \in V,

v=∑k=1n⟨v,ek⟩ekv = \sum_{k=1}^n \langle v, e_k\rangle e_k

and

∥v∥2=∑k=1n∣⟨v,ek⟩∣2.\|v\|^2 = \sum_{k=1}^n |\langle v, e_k\rangle|^2.

More generally, for u,v∈Vu, v \in V,

⟨u,v⟩=∑k=1n⟨u,ek⟩⟨v,ek⟩‾.\langle u, v \rangle = \sum_{k=1}^n \langle u, e_k\rangle \overline{\langle v, e_k \rangle}.

(In the infinite-dimensional Hilbert space setting, if (ei)i∈I(e_i)_{i \in I} is a complete orthonormal family — meaning the only vector orthogonal to all eie_i is 00 — the same conclusions hold with the sum interpreted as a norm-convergent series, provided VV is complete with respect to ∥⋅∥\|\cdot\|.)

Proof

Technique: reduce to Bessel's inequality's equality case via a dimension count, then polarize. Step 1: every v∈Vv \in V lies in span⁡(e1,…,en)\operatorname{span}(e_1,\dots,e_n). Since (e1,…,en)(e_1,\dots,e_n) is an orthonormal list, it is linearly independent: if ∑kckek=0\sum_k c_k e_k = 0, taking the inner product with eje_j gives cj=⟨∑kckek,ej⟩=⟨0,ej⟩=0c_j = \langle \sum_k c_k e_k, e_j\rangle = \langle 0, e_j \rangle = 0 for each jj, using orthonormality exactly as in Step 1 of the previous proof. A linearly independent list of n=dim⁡Vn = \dim V vectors in an nn-dimensional space is automatically a basis (this is a standard fact of finite-dimensional linear algebra: a linearly independent list of length equal to the dimension spans the space). Hence span⁡(e1,…,en)=V\operatorname{span}(e_1,\dots,e_n) = V, so every v∈Vv \in V lies in this span. Step 2: apply the equality case of Bessel's inequality. By the equality clause of Bessel's Inequality, since v∈span⁡(e1,…,en)v \in \operatorname{span}(e_1,\dots,e_n), we have equality:

∑k=1n∣⟨v,ek⟩∣2=∥v∥2,\sum_{k=1}^n |\langle v,e_k\rangle|^2 = \|v\|^2,

and moreover the proof of that equality clause showed explicitly that v=∑k=1n⟨v,ek⟩ekv = \sum_{k=1}^n \langle v,e_k\rangle e_k. This proves the first two displayed claims. Step 3: the polarized (bilinear) form. Fix u,v∈Vu,v \in V. By Step 2 applied to uu, u=∑k=1n⟨u,ek⟩eku = \sum_{k=1}^n \langle u,e_k\rangle e_k. Take the inner product of both sides with vv:

⟨u,v⟩=⟨∑k=1n⟨u,ek⟩ek, v⟩=∑k=1n⟨u,ek⟩⟨ek,v⟩,\langle u, v\rangle = \Big\langle \sum_{k=1}^n \langle u,e_k\rangle e_k,\ v\Big\rangle = \sum_{k=1}^n \langle u,e_k\rangle \langle e_k, v\rangle,

using linearity of the inner product in its first argument. Since ⟨ek,v⟩=⟨v,ek⟩‾\langle e_k, v\rangle = \overline{\langle v, e_k\rangle} by conjugate symmetry of the inner product, this is exactly

⟨u,v⟩=∑k=1n⟨u,ek⟩ ⟨v,ek⟩‾.\langle u,v\rangle = \sum_{k=1}^n \langle u, e_k\rangle\,\overline{\langle v, e_k\rangle}.

Setting u=vu = v recovers the norm identity of Step 2 as the special case ⟨v,v⟩=∥v∥2\langle v,v\rangle = \|v\|^2. ■\blacksquare

Infinite-dimensional remark (stated with proof sketch, not full functional-analytic detail). If II is infinite and (ei)i∈I(e_i)_{i\in I} is a complete orthonormal family in a Hilbert space VV, the argument above needs one extra ingredient beyond pure algebra: completeness of VV as a metric space, so that the partial sums pF=∑i∈F⟨v,ei⟩eip_F = \sum_{i \in F}\langle v,e_i\rangle e_i (over finite FF) form a Cauchy net and hence converge to some p∈Vp \in V. By Bessel's Inequality the coefficients are square-summable, which makes (pF)(p_F) Cauchy: for finite F⊆GF \subseteq G, ∥pG−pF∥2=∑i∈G∖F∣⟨v,ei⟩∣2\|p_G - p_F\|^2 = \sum_{i \in G\setminus F} |\langle v,e_i\rangle|^2 can be made arbitrarily small since the full sum converges. The vector w:=v−pw := v - p is then checked, exactly as in Step 1 of the Bessel proof (now taking a limit over finite FF), to satisfy ⟨w,ei⟩=0\langle w, e_i\rangle = 0 for every i∈Ii \in I. Completeness of the orthonormal family forces w=0w = 0, giving v=p=∑i⟨v,ei⟩eiv = p = \sum_i \langle v,e_i\rangle e_i, and Parseval's identity follows exactly as in Step 2 above. This is the standard route in the theory of Hilbert spaces; it is flagged here as the natural extension but the fully rigorous treatment of convergence of nets belongs to that more advanced setting.

∎

Worked examples

Example 1 (finite-dimensional, incomplete family — Bessel strict). Work in V=R3V = \mathbb{R}^3 with the standard inner product. Let

e1=(12,12,0),e2=(12,−12,0).e_1 = \left(\tfrac{1}{\sqrt2}, \tfrac{1}{\sqrt2}, 0\right), \qquad e_2 = \left(\tfrac{1}{\sqrt2}, -\tfrac{1}{\sqrt2}, 0\right).

Check orthonormality: ∥e1∥2=12+12+0=1\|e_1\|^2 = \tfrac12+\tfrac12+0 = 1, ∥e2∥2=12+12+0=1\|e_2\|^2 = \tfrac12+\tfrac12+0=1, and ⟨e1,e2⟩=12−12+0=0\langle e_1,e_2\rangle = \tfrac12 - \tfrac12 + 0 = 0. So (e1,e2)(e_1,e_2) is orthonormal but not complete in R3\mathbb{R}^3, since it only spans the xyxy-plane {(x,y,0)}\{(x,y,0)\}. Take the asymmetric vector v=(1,4,2)v = (1, 4, 2). Its norm squared is

∥v∥2=12+42+22=1+16+4=21.\|v\|^2 = 1^2 + 4^2 + 2^2 = 1 + 16 + 4 = 21.

Compute the Fourier coefficients:

⟨v,e1⟩=12(1)+12(4)+0=52,⟨v,e2⟩=12(1)−12(4)+0=−32.\langle v, e_1\rangle = \frac{1}{\sqrt2}(1) + \frac{1}{\sqrt2}(4) + 0 = \frac{5}{\sqrt2}, \qquad \langle v, e_2\rangle = \frac{1}{\sqrt2}(1) - \frac{1}{\sqrt2}(4) + 0 = \frac{-3}{\sqrt2}.

Then

∣⟨v,e1⟩∣2+∣⟨v,e2⟩∣2=252+92=342=17.|\langle v,e_1\rangle|^2 + |\langle v,e_2\rangle|^2 = \frac{25}{2} + \frac{9}{2} = \frac{34}{2} = 17.

Indeed 17≤2117 \le 21, confirming Bessel's inequality, with a strict gap of 21−17=421 - 17 = 4. This gap is exactly ∥w∥2\|w\|^2 where w=v−pw = v - p; here p=⟨v,e1⟩e1+⟨v,e2⟩e2p = \langle v,e_1\rangle e_1 + \langle v,e_2\rangle e_2. Direct computation gives

p=52(12,12,0)+−32(12,−12,0)=(52−32, 52+32, 0)=(1,4,0).p = \frac{5}{\sqrt2}\left(\tfrac1{\sqrt2},\tfrac1{\sqrt2},0\right) + \frac{-3}{\sqrt2}\left(\tfrac1{\sqrt2},-\tfrac1{\sqrt2},0\right) = \left(\tfrac52-\tfrac32,\ \tfrac52+\tfrac32,\ 0\right) = (1,4,0).

So w=v−p=(0,0,2)w = v - p = (0,0,2), and ∥w∥2=4\|w\|^2 = 4, matching the gap exactly. The missing third direction e3=(0,0,1)e_3 = (0,0,1) (which together with e1,e2e_1,e_2 would complete the family) carries exactly the "lost" length, since ⟨v,e3⟩=2\langle v,e_3\rangle = 2 and 22=42^2 = 4.

Example 2 (finite-dimensional, completed family — Parseval equality). Extend the previous family by adding e3=(0,0,1)e_3 = (0,0,1). One checks ⟨e3,e1⟩=0\langle e_3,e_1\rangle = 0, ⟨e3,e2⟩=0\langle e_3,e_2\rangle=0, ∥e3∥=1\|e_3\|=1, so (e1,e2,e3)(e_1,e_2,e_3) is orthonormal, and since it consists of 3=dim⁡R33 = \dim \mathbb{R}^3 linearly independent vectors, it is a complete orthonormal basis of R3\mathbb{R}^3. With the same v=(1,4,2)v = (1,4,2), we now also have ⟨v,e3⟩=2\langle v, e_3\rangle = 2, so

∑k=13∣⟨v,ek⟩∣2=17+22=17+4=21=∥v∥2,\sum_{k=1}^3 |\langle v,e_k\rangle|^2 = 17 + 2^2 = 17 + 4 = 21 = \|v\|^2,

exactly matching ∥v∥2\|v\|^2 computed above — Parseval's identity holds with equality, as guaranteed by Parseval's Identity. As a further check of the reconstruction claim v=∑k⟨v,ek⟩ekv = \sum_k \langle v,e_k\rangle e_k:

p+⟨v,e3⟩e3=(1,4,0)+2(0,0,1)=(1,4,2)=v,p + \langle v,e_3\rangle e_3 = (1,4,0) + 2(0,0,1) = (1,4,2) = v,

confirming the coordinate expansion recovers vv exactly.

Example 3 (infinite-dimensional flavor, foreshadowing Fourier series). Let V=C[−π,π]V = C[-\pi,\pi] with inner product ⟨f,g⟩=1π∫−ππf(x)g(x) dx\langle f,g\rangle = \frac{1}{\pi}\int_{-\pi}^\pi f(x)g(x)\,dx, and consider the orthonormal sequence en(x)=sin⁡(nx)e_n(x) = \sin(nx) for n=1,2,3,…n = 1,2,3,\dots (orthonormality with this normalization is a standard trigonometric integral computation: 1π∫−ππsin⁡(nx)sin⁡(mx) dx=δnm\frac1\pi\int_{-\pi}^\pi \sin(nx)\sin(mx)\,dx = \delta_{nm}). Take f(x)=xf(x) = x on [−π,π][-\pi,\pi], an asymmetric-looking but odd function. The Fourier coefficients are ⟨f,en⟩=1π∫−ππxsin⁡(nx) dx=2(−1)n+1n\langle f, e_n\rangle = \frac1\pi \int_{-\pi}^\pi x\sin(nx)\,dx = \frac{2(-1)^{n+1}}{n} (a standard integration-by-parts computation). By the Bessel inequality corollary for infinite orthonormal sequences,

∑n=1∞(2(−1)n+1n)2=4∑n=1∞1n2≤∥f∥2=1π∫−ππx2 dx=2π23.\sum_{n=1}^\infty \left(\frac{2(-1)^{n+1}}{n}\right)^2 = 4\sum_{n=1}^\infty \frac{1}{n^2} \le \|f\|^2 = \frac1\pi \int_{-\pi}^\pi x^2\,dx = \frac{2\pi^2}{3}.

Since ∑n=1∞1n2=π26\sum_{n=1}^\infty \frac1{n^2} = \frac{\pi^2}{6}, the left side is 4⋅π26=2π234 \cdot \frac{\pi^2}{6} = \frac{2\pi^2}{3}, matching the right side exactly — because the family (sin⁡(nx))n≥1(\sin(nx))_{n\ge1} turns out to be complete among odd functions on this interval, Bessel's inequality becomes Parseval's identity here, giving the classical evaluation ∑1/n2=π2/6\sum 1/n^2 = \pi^2/6 as a byproduct.

CounterexampleDropping completeness breaks Parseval, even when the count "looks right"

The completeness hypothesis in Parseval's identity cannot be dropped — an orthonormal family can fail to be complete while still consisting of a "full-looking" number of vectors if one is careless, and the identity then genuinely fails. Work again in V=R3V = \mathbb{R}^3. Consider the single orthonormal vector e1=(1,0,0)e_1 = (1,0,0) together with a duplicate direction disguised differently — this is not possible with genuinely orthonormal vectors, so instead consider the honest incomplete pair from Example 1: e1=(12,12,0)e_1 = \left(\tfrac1{\sqrt2},\tfrac1{\sqrt2},0\right), e2=(12,−12,0)e_2 = \left(\tfrac1{\sqrt2},-\tfrac1{\sqrt2},0\right). This is a bona fide orthonormal family in R3\mathbb{R}^3, but it is not complete: the vector e3=(0,0,1)e_3 = (0,0,1) is orthogonal to both e1e_1 and e2e_2 (direct computation: ⟨e3,e1⟩=0\langle e_3,e_1\rangle = 0, ⟨e3,e2⟩=0\langle e_3,e_2\rangle=0) yet e3≠0e_3 \ne 0. This directly violates the completeness definition, which requires the only vector orthogonal to every family member to be 00. Take v=(1,4,2)v = (1,4,2) exactly as before. We computed

∑k=12∣⟨v,ek⟩∣2=17,∥v∥2=21.\sum_{k=1}^2 |\langle v,e_k\rangle|^2 = 17, \qquad \|v\|^2 = 21.

These are unequal: 17≠2117 \ne 21. Parseval's identity fails for this incomplete family, even though the family has 22 vectors and one might naively think "two orthonormal vectors in a 3-dimensional space" is close enough. The shortfall is not a rounding artifact; it is exactly ∥w∥2=4>0\|w\|^2 = 4 > 0 where w=(0,0,2)w=(0,0,2) is the genuine nonzero component of vv orthogonal to both e1,e2e_1,e_2. This is precisely the phenomenon the completeness hypothesis exists to rule out: as long as any nonzero vector remains orthogonal to the whole family, Parseval's identity is false for vectors with a component along that missing direction.

Scope, edge cases & common traps

SituationWhat holdsWhat can fail
Any orthonormal family (finite or infinite), no completeness assumedBessel's inequality always holdsEquality generally fails; strict inequality is typical
Finite orthonormal list of length dim⁡V\dim VAutomatically complete (linear independence + dimension count)N/A — completeness is automatic here
Finite orthonormal list of length <dim⁡V< \dim VBessel's inequality, possibly strictParseval's identity fails whenever vv has a nonzero component orthogonal to the span
Infinite orthonormal sequence in a Hilbert spaceBessel's inequality; coefficients are square-summableConvergence of ∑⟨v,ei⟩ei\sum \langle v,e_i\rangle e_i to vv requires completeness, not just orthonormality
Inner product space that is not complete (e.g. polynomials with L2L^2 inner product, no completion)Bessel's inequality still holdsThe reconstruction v=∑⟨v,ei⟩eiv = \sum \langle v,e_i\rangle e_i may not converge within the space at all

The mathematician behind it

Origins · 1 minFriedrich Wilhelm Bessel and how this was discovered
FW
Friedrich Wilhelm Bessel
1784–1846 · German astronomer and mathematician, Königsberg Observatory

Bessel began his career as an apprentice in a mercantile firm before turning to astronomy, driven by a fascination with navigation and celestial mechanics. He became the first person to measure the parallax of a star other than the Sun (61 Cygni, 1838), a landmark achievement requiring extraordinarily precise reduction of observational data. This practical demand for handling approximation errors in astronomical measurement led him to the mathematical study of series expansions and what are now called Bessel functions, arising from problems of planetary perturbation. The inequality bearing his name emerged from exactly this concern: how much information is captured, and how much is unavoidably lost, when a signal or observation is projected onto a finite set of basis functions.

History
- 1784 — Friedrich Wilhelm Bessel born in Minden, Westphalia.
- 1810 — Appointed director of the newly founded Königsberg Observatory at age 26, on Wilhelm von Humboldt's recommendation.
- 1817–1825 — Develops Bessel functions in the course of studying perturbations in planetary motion, laying groundwork for later orthogonal function theory.
- 1838 — Measures the parallax of the star 61 Cygni, the first stellar distance ever determined, showcasing the era's demand for rigorous error and approximation analysis.
- Late 19th–early 20th century — Marc-Antoine Parseval's 1799 theorem on trigonometric series is recognized as the equality companion to Bessel's inequality, and the two results are unified within the emerging abstract theory of Hilbert spaces developed by Hilbert, Schmidt, and Riesz.
Is Bessel's inequality only true for real inner product spaces, or also complex ones?

It holds identically in both cases. The proof uses only the axioms of an inner product — linearity in the first slot, conjugate-linearity (or ordinary linearity, in the real case) in the second, conjugate symmetry, and positive-definiteness — none of which distinguishes real from complex scalars. The only notational difference is that ∣⟨v,ek⟩∣2|\langle v,e_k\rangle|^2 means the modulus squared of a possibly complex number in the complex case, versus the square of a real number in the real case; in both cases this quantity is a nonnegative real number, and the entire proof by expanding ∥v−∑⟨v,ek⟩ek∥2≥0\|v - \sum \langle v,e_k\rangle e_k\|^2 \ge 0 goes through verbatim.

Why does completeness of the family matter more than completeness of the space in the finite-dimensional case?

In finite dimensions every inner product space is automatically complete as a metric space (a standard fact following from equivalence of all norms on finite-dimensional spaces), so metric completeness is never the obstruction there. What can fail is completeness of the orthonormal family itself — whether it spans the whole space. A finite-dimensional space always has complete orthonormal families (e.g. via Gram–Schmidt applied to any basis), but a given family you are handed might only span a proper subspace, and that is the hypothesis Parseval's identity genuinely needs.

Can Bessel's inequality ever be an equality for a proper (non-complete) subfamily?

Yes, but only for special vectors vv, not for every vector in the space. The equality clause of Bessel's inequality states precisely that equality holds for a given vv if and only if vv lies in the span of the finite list used. So if (e1,e2)(e_1,e_2) spans only a plane in R3\mathbb{R}^3, any vector vv lying exactly in that plane achieves equality, while any vector with a nonzero component perpendicular to the plane gives strict inequality. Parseval's identity is the strictly stronger statement that equality holds for every vector in the space, which forces the family to span the entire space.

What is the relationship between Parseval's identity and the Pythagorean theorem?

Parseval's identity is a direct, arbitrary-dimensional generalization of the Pythagorean theorem. In two dimensions, if e1,e2e_1,e_2 are perpendicular unit vectors and v=ae1+be2v = ae_1+be_2, the Pythagorean theorem gives ∥v∥2=a2+b2\|v\|^2 = a^2+b^2, and here a=⟨v,e1⟩a = \langle v,e_1\rangle, b=⟨v,e2⟩b=\langle v,e_2\rangle, so this is exactly the n=2n=2 case of Parseval's identity. The identity simply asserts that this same decomposition of squared length into a sum of squared "legs" continues to hold along however many mutually perpendicular directions are needed to span the space, or even infinitely many in a Hilbert space.

Does Bessel's inequality require the family (ei)(e_i) to be finite, or can II be uncountable?

The inequality itself makes sense for any index set II, even an uncountable one, once one interprets the sum ∑i∈I∣⟨v,ei⟩∣2\sum_{i\in I}|\langle v,e_i\rangle|^2 as a supremum over finite partial sums. The proof shows this supremum is bounded by ∥v∥2\|v\|^2 for every finite subfamily, and a classical fact about sums of nonnegative reals indexed by an arbitrary set is that if all finite partial sums are bounded, then at most countably many terms are nonzero and the supremum equals an honest countable series sum. So although II may be uncountable in principle, only countably many Fourier coefficients of any fixed vv can be nonzero.

How does Parseval's identity foreshadow orthogonal projections?

The proof of Bessel's inequality secretly constructs the orthogonal projection of vv onto the span of the orthonormal family, namely p=∑k⟨v,ek⟩ekp = \sum_k \langle v,e_k\rangle e_k, and shows v−pv - p is orthogonal to that span. This is exactly the defining property of an orthogonal projection studied more fully under Orthogonal Projections. Bessel's inequality is then the statement ∥p∥2≤∥v∥2\|p\|^2 \le \|v\|^2, i.e. that projecting never increases length, while Parseval's identity is the special case where the projection is onto the whole space, so p=vp = v and no length is lost at all.

If a family is orthonormal and complete, is it automatically a basis in the linear-algebra sense (i.e. every vector is a finite linear combination)?

In finite dimensions, yes: completeness there means the family spans VV, and combined with orthonormality (which guarantees linear independence), this makes it a genuine vector space basis in the usual sense, with every vector a finite linear combination of the eie_i. In infinite-dimensional Hilbert spaces, completeness only guarantees that every vector is a possibly infinite convergent series ∑i⟨v,ei⟩ei\sum_i \langle v,e_i\rangle e_i, not a finite linear combination; such a family is called a complete orthonormal system or Hilbert basis, a strictly weaker and different notion from a Hamel (finite linear combination) basis, which such spaces almost never admit in a countable, well-behaved form.

Before this

Where this leads