The Riesz Representation Theorem

The Riesz Representation Theorem identifies the dual of a Hilbert space with the space itself: every continuous linear functional

f

on a Hilbert space

H

is given by inner product with a unique vector

y_{f} \in H

, i.e.

f (x) = ⟨ x, y_{f} ⟩

. Applies to all (real or complex) Hilbert spaces; foundational in functional analysis, PDE, quantum mechanics, and probability. Proved by Frigyes Riesz and Maurice Fréchet (1907); a separate measure-theoretic "Riesz representation theorem" for

C (X)^{*}

is due to Riesz (1909) and refined by Markov–Kakutani–Banach.

The Riesz Representation Theorem (in its Hilbert-space form) asserts that for a Hilbert space $H$ over $K = R$ or $C$ , the map sending $y \in H$ to the functional $x \mapsto ⟨ x, y ⟩$ is a (conjugate-)linear isometric bijection onto the continuous dual $H^{*}$ . This makes $H$ essentially self-dual, a property that fails badly for general Banach spaces and is the structural reason Hilbert spaces are so tractable. The theorem underlies the existence of adjoint operators, the spectral theorem, weak solutions to PDE (via Lax–Milgram), and the bra–ket formalism of quantum mechanics. A surprising fact: although the proof is short and elementary once one has orthogonal projection, the result is equivalent in strength to the completeness of the inner product space — it can fail on incomplete pre-Hilbert spaces. There is also a distinct measure-theoretic Riesz theorem identifying positive functionals on $C_{c} (X)$ with Radon measures, which shares only the name and the originator.

The Core Insight

A Hilbert space possesses orthogonal projection: for any closed subspace $M$ , every vector splits uniquely as $x = P x + (x - P x)$ with $P x \in M$ and $x - P x ⊥ M$ . Given a nonzero continuous functional $f$ , its kernel is a closed hyperplane. The one-dimensional orthogonal complement of the kernel singles out a direction; rescaling a unit vector in that direction by $\overline{f}$ of itself produces the representing vector. Geometry (perpendicularity) thus encodes the entirely algebraic notion of a linear functional.

Formal Statement

Continuous Dual and Inner Product Conventions

Let $H$ be a Hilbert space over $K \in {R, C}$ with inner product $⟨ \cdot, \cdot ⟩$ , linear in the first argument and conjugate-linear in the second (the physics convention is opposite; see the warning). The continuous (topological) dual $H^{*}$ is the space of bounded linear functionals $f : H \to K$ with operator norm

∥ f ∥ = ∥ x ∥ \leq 1 sup ∣ f (x) ∣.

⟳

∥ f ∥

is the smallest constant

C

with

∣ f (x) ∣ \leq C ∥ x ∥

for all

x

; boundedness is equivalent to continuity for linear maps.

The norm on $H$ is induced by the inner product:

∥ x ∥ = ⟨ x, x ⟩^{1/2} .

⟳positive-definiteness of

⟨ \cdot, \cdot ⟩

guarantees

⟨ x, x ⟩ \geq 0

with equality iff

x = 0

; the square root defines a genuine norm via Cauchy–Schwarz and the parallelogram law.

Riesz–Fréchet Representation Theorem

Let $H$ be a Hilbert space. For every $f \in H^{*}$ there exists a unique $y_{f} \in H$ such that

f (x) = ⟨ x, y_{f} ⟩ for all x \in H,

and moreover

∥ f ∥_{H^{*}} = ∥ y_{f} ∥_{H} .

The map $Φ : H^{*} \to H$ , $Φ (f) = y_{f}$ , is a conjugate-linear, isometric bijection. In the real case $Φ$ is linear, and $H ≅ H^{*}$ as Banach spaces; in the complex case $H^{*}$ is conjugate-linearly isometric to $H$ .

Derivation hint for [3]–[4]: existence via orthogonal projection onto $(ker f)^{⊥}$ ; uniqueness because $⟨ x, y ⟩ = ⟨ x, y^{'} ⟩ \forall x$ forces $y - y^{'} ⊥ H$ ; the norm identity follows from Cauchy–Schwarz with equality attained at $x = y_{f}$ .

Why It Holds / Derivation

Existence, Uniqueness, and Isometry

Let $f \in H^{*}$ .

Step 1 — Trivial case. If

f = 0

, take

y_{f} = 0

; then

f (x) = ⟨ x, 0 ⟩ = 0

. Assume henceforth

f \neq = 0

. Step 2 — The kernel is a proper closed subspace. Since

f

is continuous,

M := ker f = f^{- 1} ({0})

is closed; since

f \neq = 0

M \neq = H

. By the projection theorem (valid because

H

is complete),

H = M \oplus M^{⊥}

with

M^{⊥} \neq = {0}

. Step 3 — The complement is one-dimensional. Pick

z \in M^{⊥}

with

z \neq = 0

; rescale so

f (z) = 1

is unnecessary, but note for any

x \in H

u := f (x) z - f (z) x \in M, since f (u) = f (x) f (z) - f (z) f (x) = 0.

⟳linearity of

f

collapses the combination to zero; this is the standard trick exhibiting any

x

modulo the kernel as a multiple of

z

Step 4 — Project onto the kernel. Because

u \in M

and

z \in M^{⊥}

, we have

⟨ u, z ⟩ = 0

. Expanding:

0 = ⟨ f (x) z - f (z) x, z ⟩ = f (x) ∥ z ∥^{2} - f (z) ⟨ x, z ⟩ .

⟳use linearity of the inner product in the first slot and

⟨ z, z ⟩ = ∥ z ∥^{2}

Step 5 — Solve for $f (x)$ . Rearranging [6],

f (x) = \frac{f ( z )}{∥ z ∥ ^{2}} ⟨ x, z ⟩ = ⟨ x, \frac{f ( z )}{∥ z ∥ ^{2}} z ⟩ .

⟳divide by

∥ z ∥^{2} \neq = 0

; the scalar

f (z) /∥ z ∥^{2}

moves into the conjugate-linear second argument as its complex conjugate.

Define

y_{f} := \frac{f ( z )}{∥ z ∥ ^{2}} z .

⟳this is the unique representing vector read off from [7]; it lies in

M^{⊥}

, the direction orthogonal to the kernel.

Then $f (x) = ⟨ x, y_{f} ⟩$ for all $x$ , proving existence.

Step 6 — Uniqueness. If

⟨ x, y_{f} ⟩ = ⟨ x, y^{'} ⟩

for all

x

, then

⟨ x, y_{f} - y^{'} ⟩ = 0

for all

x

; choosing

x = y_{f} - y^{'}

gives

∥ y_{f} - y^{'} ∥^{2} = 0

, hence

y_{f} = y^{'}

. Step 7 — Isometry [4]. By Cauchy–Schwarz,

∣ f (x) ∣ = ∣ ⟨ x, y_{f} ⟩ ∣ \leq ∥ x ∥ ∥ y_{f} ∥

, so

∥ f ∥ \leq ∥ y_{f} ∥

. Conversely, taking

x = y_{f}

(assuming

y_{f} \neq = 0

∥ f ∥ \geq \frac{∣ f ( y _{f} ) ∣}{∥ y _{f} ∥} = \frac{⟨ y _{f} , y _{f} ⟩}{∥ y _{f} ∥} = ∥ y_{f} ∥.

⟳the supremum defining

∥ f ∥

is attained (or approached) precisely along the representing direction, forcing equality.

Hence $∥ f ∥ = ∥ y_{f} ∥$ .

Step 8 — Conjugate-linearity of $Φ$ . For

f, g \in H^{*}

and scalar

α

Φ (α f + g)

satisfies

⟨ x, Φ (α f + g)⟩ = α f (x) + g (x) = ⟨ x, \overline{α} y_{f} + y_{g} ⟩

wait — carefully:

α f (x) = α ⟨ x, y_{f} ⟩ = ⟨ x, \overline{α} y_{f} ⟩

. Thus

Φ (α f + g) = \overline{α} y_{f} + y_{g}

, i.e.

Φ

is conjugate-linear. Surjectivity is immediate since every

y \in H

defines

f_{y} (x) = ⟨ x, y ⟩ \in H^{*}

■

Completeness is essential

Step 2 invokes the projection theorem, which requires $H$ to be complete. On a non-complete inner product space the theorem fails: there exist bounded functionals with no representing vector. Completeness is not merely a convenience — it is logically necessary.

Frigyes Riesz | 1880–1956

Hungarian mathematician, a founder of functional analysis. In 1907 Riesz and, independently, Maurice Fréchet established the representation of continuous linear functionals on $L^{2}$ via inner products — the prototype of the modern theorem. Riesz also proved the $C (X)^{*}$ representation by measures (1909), introduced the spaces now bearing the names $L^{p}$ , and developed the theory of compact (completely continuous) operators. His textbook with Béla Szőkefalvi-Nagy, Leçons d'analyse fonctionnelle (1952), shaped generations. The breadth of his contributions means "the Riesz theorem" can refer to several distinct results; context disambiguates.

Building Intuition

Think of a linear functional as a "measuring device" that assigns a number to each vector. In finite dimensions every such device is dotting against some fixed vector — its gradient. Riesz says this picture survives intact in infinite dimensions provided the space is complete: the functional still has a gradient, a single arrow $y_{f}$ , and the functional reads off the (signed) length of the projection of $x$ onto that arrow, scaled by $∥ y_{f} ∥$ . The kernel of $f$ is the hyperplane perpendicular to $y_{f}$ ; the further $x$ sticks out beyond that hyperplane, the larger $f (x)$ . The norm identity $∥ f ∥ = ∥ y_{f} ∥$ says: a steeper functional corresponds to a longer gradient arrow.

Geometry of Riesz representation: the functional as projection onto its gradient direction $y_f$

Worked Examples

Sequence space $\ell^2$

Given. Define

f : ℓ^{2} (C) \to C

f (x) = \sum_{n = 1}^{\infty} \frac{x _{n}}{n}

for

x = (x_{n})

. Step 1 — Boundedness. By Cauchy–Schwarz,

∣ f (x) ∣ \leq (\sum ∣ x_{n} ∣^{2})^{1/2} (\sum 1/ n^{2})^{1/2}

, and

\sum 1/ n^{2} = π^{2} /6

. So

f

is bounded with

∥ f ∥ \leq π / 6

. Step 2 — Identify $y_{f}$ . Write

f (x) = ⟨ x, y_{f} ⟩ = \sum x_{n} \overline{(y_{f})_{n}}

. Matching coefficients,

\overline{(y_{f})_{n}} = 1/ n

, i.e.

(y_{f})_{n} = 1/ n

(real). Then

y_{f} = (1, \frac{1}{2}, \frac{1}{3}, \dots) \in ℓ^{2}

since

\sum 1/ n^{2} < \infty

. Step 3 — Norm.

∥ y_{f} ∥ = (\sum 1/ n^{2})^{1/2} = π / 6

, and by [4],

∥ f ∥ = ∥ y_{f} ∥

(the Cauchy–Schwarz bound is achieved).

y_{f} = (\frac{1}{n})_{n \geq 1}, ∥ f ∥ = ∥ y_{f} ∥ = \frac{π}{6} \approx 1.2825

:::example[Function space $L^{2} [0, 1]$ — point-evaluation surrogate] Given. On $H = L^{2} [0, 1]$ , consider $f (g) = \int_{0}^{1} g (t) t^{2} d t$ .

Step 1 — Boundedness.

∣ f (g) ∣ \leq ∥ g ∥_{2} ∥ t^{2} ∥_{2}

by Cauchy–Schwarz. Step 2 — Representing function.

f (g) = ⟨ g, y_{f} ⟩ = \int_{0}^{1} g (t) \overline{y_{f} (t)} d t

. Hence

\overline{y_{f} (t)} = t^{2}

, so

y_{f} (t) = t^{2}

(real-valued). Step 3 — Norm.

∥ y_{f} ∥_{2}^{2} = \int_{0}^{1} t^{4} d t = \frac{1}{5}

, so

∥ f ∥ = ∥ y_{f} ∥_{2} = 1/ 5

y_{f} (t) = t^{2}, ∥ f ∥ = \frac{1}{5} \approx 0.4472

Caveat. True point-evaluation

g \mapsto g (t_{0})

is not bounded on

L^{2}

, so it has no Riesz representative there — but it does in a reproducing kernel Hilbert space, where

y_{f}

is the kernel section

K (\cdot, t_{0})

Common Misconceptions

Linear vs. conjugate-linear; convention clash

With the inner product linear in the first slot, $Φ : H^{*} \to H$ is conjugate-linear. Physicists use $⟨ \cdot, \cdot ⟩$ linear in the second slot (Dirac), which swaps where the conjugation lands; then the bra $⟨ ψ ∣$ is the Riesz functional of the ket $∣ ψ ⟩$ , and the map ket $\mapsto$ bra is antilinear. Both communities are consistent internally — always state your convention. A second pitfall: $∥ f ∥ = ∥ y_{f} ∥$ is an isometry, but $Φ$ is not an isometric linear isomorphism over $C$ (it reverses scalar multiplication by conjugation).

Failure on an incomplete inner product space

Let $H_{0}$ be the inner product space of finitely-supported sequences (a dense, non-complete subspace of $ℓ^{2}$ ), with the $ℓ^{2}$ inner product. Define $f (x) = \sum_{n} x_{n} / n$ , bounded with $∥ f ∥ = π / 6$ . Its only candidate representative is $y_{f} = (1/ n)_{n}$ , which is not finitely supported, hence $y_{f} \in / H_{0}$ . So $f$ is a bounded functional on $H_{0}$ with no representing vector in $H_{0}$ . Completeness is genuinely required; Riesz can fail on pre-Hilbert spaces.

Connections

Where Riesz sits in functional analysis

The theorem is the linchpin connecting several pillars. It yields adjoint operators: given bounded $T : H \to H$ , the functional $x \mapsto ⟨ T x, y ⟩$ is represented by $T^{*} y$ , defining $T^{*}$ — the gateway to the Spectral Theorem for self-adjoint and normal operators. It powers the Lax–Milgram theorem and hence existence/uniqueness of weak solutions to elliptic Partial Differential Equations. Within Banach Space Duality it shows Hilbert spaces are reflexive and self-dual, a sharp contrast with $L^{1}$ , $C (X)$ , etc. It also undergirds Reproducing Kernel Hilbert Spaces, where bounded point evaluation forces a kernel function, and connects to the Radon–Nikodym Theorem (Riesz functional ⟷ density), which von Neumann used to prove Radon–Nikodym via Riesz.

Real-World Applications

Named uses across science

Quantum mechanics (Dirac bra–ket). Every bra $⟨ ϕ ∣$ is the Riesz dual of a ket $∣ ϕ ⟩$ in $L^{2} (R^{3})$ ; observables' expectation values $⟨ ψ ∣ A ∣ ψ ⟩$ rely on adjoints defined via Riesz.
Finite Element Method (Galerkin) for PDE. The Lax–Milgram existence theory for weak solutions of Poisson, elasticity, and Stokes equations is a Riesz/coercivity argument.
Signal processing & sampling. Reproducing kernel structure (Shannon sampling, band-limited spaces) realizes samples $f (t_{0})$ as inner products with reproducing kernels — Riesz representatives of evaluation.
Kernel methods / SVM (machine learning). RKHS theory, central to support vector machines and Gaussian processes, depends on bounded evaluation functionals having Riesz representers.
Probability — conditional expectation. $E [\cdot ∣ G]$ is the orthogonal projection in $L^{2}$ ; the existence of densities (Radon–Nikodym) flows from the Riesz argument.
Optimal control & calculus of variations. Representing the derivative of a cost functional as a gradient vector (the "adjoint state") is a Riesz identification, enabling gradient-based optimization.
Electromagnetics / boundary integral equations. Variational (weak) formulations of Maxwell and Helmholtz problems use Riesz–Lax–Milgram to guarantee solvability in Sobolev spaces.

Historical Development

1902–1906 — Lebesgue's integration theory creates the function spaces ( $L^{2}$ ) on which the theorem will live.
1907 — Frigyes Riesz and, independently, Maurice Fréchet publish the representation of continuous linear functionals on $L^{2}$ as integrals against an $L^{2}$ function — the original Riesz–Fréchet theorem.
1909 — Riesz proves the measure-theoretic representation: positive linear functionals on $C [0, 1]$ are integrals against (Stieltjes) measures.
1910–1913 — Riesz develops $L^{p}$ spaces and the duality $L^{p}^{*} ≅ L^{q}$ ; the $p = 2$ case is the inner-product theorem.
1927 — John von Neumann axiomatizes Hilbert space; the abstract projection theorem makes the modern, geometry-based proof possible and universal.
1932 — Stefan Banach's Théorie des opérations linéaires systematizes dual spaces, placing Riesz within general duality theory.
1940s–50s — Kakutani, Markov, Banach generalize the $C (X)^{*}$ theorem to locally compact Hausdorff $X$ with Radon measures (the modern "Riesz–Markov–Kakutani" theorem).
1950 — Lax & Milgram extend the representation to bounded coercive sesquilinear forms, enabling the variational theory of PDE.
1955 — Riesz–Szőkefalvi-Nagy textbook canonizes the proof and terminology used today.

Subtleties & Limits

Boundaries of the principle

The Hilbert-space theorem is complete and not open, but its boundaries are subtle and active. (i) Non-Hilbert duality is irregular: $L^{1} (μ)^{*} = L^{\infty} (μ)$ only under $σ$ -finiteness; $C (X)^{*}$ requires the full Radon-measure machinery; spaces like $L^{\infty}$ and $ℓ^{\infty}$ are not even reflexive, and their precise duals invoke finitely-additive measures and Stone–Čech compactifications — describing these "exotic" functionals concretely is genuinely hard. (ii) Riesz on non-self-adjoint / indefinite forms. Krein-space and Pontryagin-space analogues replace positive-definiteness with an indefinite metric; the representing element exists but the clean isometry [4] degrades, and open questions persist about operator representations on indefinite inner-product spaces. (iii) Quantitative/effective versions. In computable and constructive analysis, the non-constructive projection step means the representative $y_{f}$ may not be effectively computable from $f$ without additional modulus-of-continuity data — characterizing exactly when it is computable remains a research theme. (iv) Operator-valued and rigged-Hilbert-space extensions (Gelfand triples) handle unbounded functionals like point evaluation and the Dirac delta, but the precise distributional duality is delicate.

The Riesz Representation Theorem

The Core Insight

Formal Statement

Continuous Dual and Inner Product Conventions

Riesz–Fréchet Representation Theorem

Why It Holds / Derivation

Existence, Uniqueness, and Isometry

Completeness is essential

Frigyes Riesz | 1880–1956

Building Intuition

Geometry of Riesz representation: the functional as projection onto its gradient direction $y_f$

Worked Examples

Sequence space $\ell^2$

Common Misconceptions

Linear vs. conjugate-linear; convention clash

Failure on an incomplete inner product space

Connections

Where Riesz sits in functional analysis

Real-World Applications

Named uses across science

Historical Development

Subtleties & Limits

Boundaries of the principle

See Also

Before this

Where this leads