Dual Spaces and Linear Functionals

Duality stopped being abstract the moment quantum mechanics made the bra–ket its native grammar, and it reappears wherever "linear measurement" is the operative idea. In scattering theory the plane-wave and delta-function "states" are not in the Hilbert space at all — they live in the dual of a dense space of smooth test functions, formalized as the rigged Hilbert space or Gelfand triple $\Phi\subset H\subset\Phi^*$, which is what makes Dirac's $\langle x\vert p\rangle=e^{ipx}$ rigorous. In numerical PDE, the weak formulation behind every finite-element solver recasts a differential equation as "find $u$ such that a bilinear form equals a given functional for all test functions," and the Lax–Milgram theorem — a direct descendant of Riesz — guarantees a unique solution; this is the mathematics inside structural, fluid, and electromagnetic simulation codes. Machine learning leans on the same result: a reproducing kernel Hilbert space is defined by demanding that evaluation-at-a-point be a bounded functional, so Riesz hands you a representing "kernel," and the representer theorem behind support-vector machines and kernel ridge regression is Riesz duality doing the heavy lifting. Even gravitational-wave detection is dual-space work — matched filtering applies a functional (the inner product of data with a template) and reads its value. Wherever a gradient, a momentum, a bra, or a test-function pairing appears, $V^*$ is quietly in charge.

Picture a vector as an arrow and a covector as a stack of parallel, evenly spaced rulings — the level surfaces of a linear function. The functional's value on a vector is simply the number of rulings the arrow pierces, a "pierce count" that needs no coordinates and no length: it is defined by counting, not measuring. Where the covector is "strong" the rulings crowd together and a fixed arrow pierces more of them. Vectors point; covectors slice. The first figure makes this concrete for the simplest covector, the dual-basis element that just reads off a coordinate.

A basis labels each vector by its components $v^i$, and the dual basis is nothing more exotic than the operation "read off the $i$-th component," $e^i(v)=v^i$. So covectors are the measurement devices that extract coordinates — logically prior to the numbers in a column, not derived from them. But notice what is missing: nothing so far tells you which covector "is" a given vector. The arrow lives in $V$, the rulings live in $V^*$, and the two spaces face each other across a gap that no amount of pure linear algebra can bridge. An inner product is exactly the structure that glues them, and the Riesz vector is the vector whose "count my rulings" role reproduces $\langle v,\cdot\rangle$; in Dirac's language, the inner product is what turns a ket into a bra. The second figure lays out the three spaces of duality and the maps between them, marking which are free and which cost you a choice.

A linear functional is defined by exactly two demands: it lands in the scalar field, and it respects linear combinations. Collecting every such map gives the dual space, the primary object of this article.

where: $V$ — a vector space over a field $\mathbb{F}$ (here $\mathbb{R}$ or $\mathbb{C}$); $\varphi$ — a linear functional, also called a covector or one-form; $\alpha,\beta$ — scalars in $\mathbb{F}$; $V^*$ — the dual space of all such functionals. Read it as: $V^*$ collects every linear "measuring device" you can point at $V$. The dual space is not merely a set — it inherits its own linear structure by combining functionals pointwise, value by value on each input vector, which makes "the space of measurements" a vector space in its own right.

where: $\varphi,\psi$ — functionals in $V^*$; $\lambda$ — a scalar; $x$ — any vector in $V$. Read it as: covectors add and scale exactly like ordinary functions, so every vector-space axiom holds automatically. Once $V$ has a basis, $V^*$ acquires a canonical partner basis. The dual basis $\{e^i\}$ is fixed by demanding that $e^i$ return $1$ on $e_i$ and $0$ on every other basis vector, and every functional then expands in it with its own values as coordinates.

where: $\{e_j\}$ — a basis of $V$; $\{e^i\}$ — the induced dual basis of $V^*$; $\delta^i{}_j$ — the Kronecker delta ($1$ if $i=j$, else $0$); $\varphi(e_i)$ — the (covariant) components of $\varphi$. Read it as: $e^i$ is the "extract coordinate $i$" functional, and any $\varphi$ is a weighted sum of coordinate-extractors. Change the basis of $V$ and the dual basis must move too — but oppositely, so that the pairing $e^i(e_j)=\delta^i{}_j$ is preserved. If the new basis is built from the old one by a matrix $M$, the new dual basis is built by the inverse of $M$.

where: $M^j{}_i$ — the change-of-basis matrix on $V$; $(M^{-1})^i{}_j$ — its inverse, which drives the dual basis; $\tilde e_i,\tilde e^i$ — the new basis and its new dual. Read it as: bases and their duals transform contragrediently — this opposite behavior is the entire content of the words covariant and contravariant. Duality can be applied twice. The dual of $V^*$ is the double dual $V^{**}$, and one map into it needs no arbitrary choice at all: send a vector $x$ to the operation "evaluate any covector at $x$."

where: $V^{**}$ — the dual of $V^*$, i.e. functionals acting on covectors; $J(x)$ — the "evaluation at $x$" element of $V^{**}$. Read it as: a vector is also a machine that eats covectors, and this reinterpretation is automatic — no basis, no inner product. Adding an inner product changes the game. On any Hilbert space (in particular any finite-dimensional inner-product space) every bounded functional is secretly an inner product against one fixed vector, and that vector is unique. This is the Riesz representation theorem, the bridge from $V^*$ back to $V$.

where: $\langle\cdot,\cdot\rangle$ — the inner product (taken linear in the second slot in the complex case); $v$ — the unique Riesz vector representing $\varphi$; $\|\varphi\|$ — the operator norm of the functional. Read it as: an inner product supplies the missing canonical bridge $V^*\to V$, and it is even an isometry. In an orthonormal basis the Riesz vector is written down instantly: its components are the complex conjugates of the functional's components. This is Riesz made fully concrete.

where: $\overline{(\cdot)}$ — complex conjugation (drop it in the real case); $e_i$ — an orthonormal basis vector. Read it as: in an orthonormal basis, "raising the index" of a covector is just conjugating its components — nothing else happens. Dirac's notation is Riesz in disguise. A ket is a vector in the Hilbert space; the matching bra is its Riesz functional, and the bra–ket product is exactly the inner product.

where: $\lvert\psi\rangle$ — a ket in the Hilbert space $H$; $\langle\psi\rvert$ — the bra, the Riesz functional $\langle\psi,\cdot\rangle\in H^*$; $\dagger$ — the conjugate transpose or Hermitian adjoint. Read it as: the dagger is the Riesz map, and because Riesz is conjugate-linear over $\mathbb{C}$, scaling a ket by $\lambda$ scales its bra by $\bar\lambda$. In component language, an inner product is a metric $g_{ij}$, and Riesz becomes the familiar act of lowering an index with $g$ and raising it back with the inverse metric.

where: $g_{ij}=\langle e_i,e_j\rangle$ — the metric (inner-product) components; $g^{ij}$ — the inverse metric; $v^j$ — contravariant (vector) components; $\omega_i$ — covariant (covector) components. Read it as: raising and lowering indices is not bookkeeping magic — it is the Riesz isomorphism written out in coordinates.

How duality behaves as you strip away or add structure:

1 · Dual bases transform contragrediently (fully worked)

In a 2D space with basis $\{e_1,e_2\}$, shear to a new basis $\tilde e_1=e_1+e_2$, $\tilde e_2=e_2$. Find the new dual basis, then track the components of the functional $\varphi$ with $\varphi(e_1)=2$, $\varphi(e_2)=-3$ in both bases and confirm they move oppositely to the basis vectors.

Step 1. Read off $M$ from $\tilde e_i=\sum_j M^j{}_i e_j$: the columns are the new basis vectors' old components, so $M=\begin{pmatrix}1&0\\1&1\end{pmatrix}$ (with $M^1{}_1=1,M^2{}_1=1,M^1{}_2=0,M^2{}_2=1$). Its inverse is $M^{-1}=\begin{pmatrix}1&0\\-1&1\end{pmatrix}$.

Step 2. Build the new dual basis from equation (4), $\tilde e^i=\sum_j (M^{-1})^i{}_j e^j$: this gives $\tilde e^1=e^1$ and $\tilde e^2=-e^1+e^2$. Sanity-check duality directly — $\tilde e^2(\tilde e_1)=(-e^1+e^2)(e_1+e_2)=-1+1=0$ and $\tilde e^2(\tilde e_2)=(-e^1+e^2)(e_2)=1$, and likewise $\tilde e^1(\tilde e_1)=1$, $\tilde e^1(\tilde e_2)=0$. So $\tilde e^i(\tilde e_j)=\delta^i{}_j$ holds. New dual basis: $\tilde e^1=e^1,\ \tilde e^2=e^2-e^1$.

Step 3. Now the components of $\varphi$. In the new basis $\tilde\varphi_i=\varphi(\tilde e_i)$: $\tilde\varphi_1=\varphi(e_1+e_2)=2+(-3)=-1$ and $\tilde\varphi_2=\varphi(e_2)=-3$. Compare with the covariant transformation law $\tilde\varphi_i=\sum_j M^j{}_i\varphi_j$: $\tilde\varphi_1=1\cdot2+1\cdot(-3)=-1$ and $\tilde\varphi_2=0\cdot2+1\cdot(-3)=-3$ — a perfect match. Covector components transform with $M$ (covariantly); the dual basis vectors transform with $M^{-1}$ — opposite directions, which is the whole point of the word contragredient.

2 · The Riesz vector under a non-standard metric (fully worked)

Equip $\mathbb{R}^2$ with the inner product whose metric is $g=\begin{pmatrix}2&1\\1&2\end{pmatrix}$. Find the Riesz vector $v$ representing the functional $\varphi(x)=x^1+x^2$, i.e. the covector with components $\omega=(1,1)$, and verify $\langle v,x\rangle=\varphi(x)$.

Step 1. Raise the index with the inverse metric, $v^i=g^{ij}\omega_j$. Since $\det g=3$, $g^{-1}=\frac13\begin{pmatrix}2&-1\\-1&2\end{pmatrix}$. Then $v^1=\tfrac13(2\cdot1-1\cdot1)=\tfrac13$ and $v^2=\tfrac13(-1\cdot1+2\cdot1)=\tfrac13$. Riesz vector: $v=\big(\tfrac13,\tfrac13\big)$.

Step 2. Verify by lowering back, $\omega_i=g_{ij}v^j$: $\omega_1=2\cdot\tfrac13+1\cdot\tfrac13=1$ and $\omega_2=1\cdot\tfrac13+2\cdot\tfrac13=1$, recovering $\omega=(1,1)$. Equivalently, for any $x=(a,b)$ the inner product is $\langle v,x\rangle=v^{\mathsf T} g\, x=a+b=\varphi(x)$. The representing vector is not $(1,1)$ — the metric skews it to $(\tfrac13,\tfrac13)$, which is exactly why the Riesz map depends on the inner product and is invisible to bare linear algebra.

3 · Conjugate-linearity of the bra map

In $\mathbb{C}^2$ with the standard inner product, take $\lvert\psi\rangle=(1,\,i)$; write its bra and the bra of $i\lvert\psi\rangle$.

Step 1. The bra is the conjugate transpose, $\langle\psi\rvert=(1,\,-i)$, with $\langle\psi\vert\psi\rangle=1\cdot1+(-i)(i)=2=\|\psi\|^2$. Scaling the ket by $\lambda=i$ gives $i\lvert\psi\rangle=(i,-1)$, whose bra is $(-i,\,-1)=-i\,(1,\,-i)=\bar\lambda\,\langle\psi\rvert$. Scaling a ket by $\lambda$ scales its bra by $\bar\lambda$: the Riesz/dagger map is conjugate-linear, as equation (8) demands.

The dual space is where three strands of twentieth-century mathematics braid together — the integral representations Hadamard and Fréchet were chasing, the abstract functional analysis Banach would systematize, and the operator formalism Dirac needed for quantum mechanics. The knot that ties them is one theorem, and it carries the name of the Hungarian analyst who proved it.

Texts. Stone and Goldbart, Mathematics for Physics, chapter on linear algebra and dual spaces · Arfken, Weber and Harris, Mathematical Methods for Physicists, 7th ed., Ch. 2–3 (tensors, covariant and contravariant indices) · Hassani, Mathematical Physics, Ch. 1–2 and the chapter on Hilbert spaces (Riesz–Fréchet) · Nakahara, Geometry, Topology and Physics, §2.5 (cotangent spaces and one-forms) · Byron and Fuller, Mathematics of Classical and Quantum Physics, Ch. 3–5 · Halmos, Finite-Dimensional Vector Spaces, §13–17 · Reed and Simon, Methods of Modern Mathematical Physics I, §II.2 (Riesz lemma).

Provenance. The dual-basis construction and its spanning and independence proofs follow Halmos §13–17 and Stone and Goldbart; the contragredient transformation law and index raising and lowering follow Arfken–Weber–Harris Ch. 2–3 and Byron and Fuller; the canonical embedding into the double dual follows Halmos §16; the Riesz representation theorem and the adjoint it spawns follow Reed and Simon Vol. I §II.2 and Hassani's Hilbert-space chapter; the cotangent and one-form language follows Nakahara §2.5. A verifier should recompute: the spanning identity and independence argument; the contragredient relation and shear example ($\tilde e^2=e^2-e^1$, covector components $(-1,-3)$); the injectivity of $J$ by basis extension; the Riesz construction $v=\overline{\varphi(z)}z/\|z\|^2$, especially the conjugation and completeness steps; and the metric example $v=(\tfrac13,\tfrac13)$ for $g=\big(\begin{smallmatrix}2&1\\1&2\end{smallmatrix}\big)$. No empirical constants are used.

Prerequisites. Vector Spaces and Subspaces · Linear Independence, Basis, and Dimension · Linear Operators and Matrix Representations · Inner-Product Spaces and Norms. Leads to. Adjoint and Hermitian Operators · Change of Basis and Similarity Transformations · Orthogonality and the Gram–Schmidt Process.

Ask the tutor: Ask me to walk through the Riesz construction for a functional of your choosing, or to show exactly how a bra becomes a row of conjugated components in your favorite basis.