Dual Space

Why dual space matters

• Differential forms. The cotangent space T_p* M of a manifold is a dual space; differential forms are sections of cotangent bundles. Integration on manifolds, de Rham cohomology, and Stokes' theorem are formulated entirely in terms of dual spaces — without them, "integration of a function over a curve" has no coordinate-free meaning.
• Weak topology. The weak topology on V is defined as the coarsest topology making every functional in V* continuous. It's strictly weaker than the norm topology in infinite dimensions and crucial in functional analysis: bounded sets are weakly precompact (Banach-Alaoglu), enabling existence proofs in PDE and optimization.
• Distributions. Generalized functions like the Dirac delta live in the dual of test functions — δ ∈ S' is defined by δ(φ) = φ(0). Distribution theory (Schwartz, 1950s) gives a rigorous framework for "functions" that are too singular to be functions, by reinterpreting them as functionals on smooth test functions.
• Optimization Lagrangians. Convex duality theory writes a primal problem and a dual problem related by Legendre-Fenchel transform. The dual variable lives naturally in V* (the "co-vector" space). Karush-Kuhn-Tucker conditions, linear programming duality, and convex conjugates all live there.
• Tensors. A tensor of type (p, q) is a multilinear map V* × … × V* × V × … × V → F (with p copies of V* and q copies of V). Without dual spaces, you can't write a clean coordinate-free definition of tensors — and physics (general relativity, fluid mechanics) becomes a notational mess.
• Quantum mechanics. Bra-ket notation: |ψ⟩ is a vector in Hilbert space H, ⟨ψ| is the corresponding element of H*. The pairing ⟨φ|ψ⟩ is the dual pairing — an inner product via Riesz representation. Operators have adjoints A* : H* → H* defined by ⟨φ| (A|ψ⟩) = (⟨φ|A) |ψ⟩, which only makes sense once you have V*.
• Functional analysis. Banach space theory studies continuous duals X* of normed spaces; the reflexive condition X = X** is central. Hahn-Banach theorem (extending functionals) is a workhorse. Conjugate exponents 1/p + 1/q = 1 give (Lp)* = Lq, generalizing inner product duality to general p.

Common misconceptions

• V* is concrete. V* is abstract until you pick a basis. After fixing a basis of V (and hence the dual basis of V*), V* is concretely isomorphic to V via row-vector representation. Without the basis choice, there's no canonical way to view a functional as a vector — they're different kinds of objects.
• V = V*. Same dimension in finite-dim, but no canonical isomorphism without extra structure (an inner product or a chosen basis). The "natural" isomorphism people sometimes assume — sending eⱼ to its dual eʲ — is basis-dependent: change basis and the map changes. Distinguishing V from V* protects against bugs in covariant/contravariant tensor computations.
• Always reflexive. Only finite-dim spaces and Hilbert spaces are automatically reflexive. Famous non-reflexive: ℓ¹ (its dual is ℓ^∞, but the dual of ℓ^∞ is much bigger than ℓ¹), c₀ (dual is ℓ¹, but dual of ℓ¹ is ℓ^∞ ⊋ c₀), L¹(ℝ). Reflexivity is a strong condition with major consequences for compactness.
• Continuous = algebraic dual. In topology, V* usually means continuous dual (functionals that are continuous w.r.t. the norm). The algebraic dual (all linear functionals, possibly discontinuous) is much larger in infinite dimensions. For Banach spaces, the relevant V* is always the continuous dual; algebraic dual is rarely used.
• Dual basis is "the same" basis. The dual basis {eⁱ} consists of functionals, not vectors of V. Concrete: for V = ℝ² with basis e₁ = (1,0), e₂ = (0,1), the dual basis vectors are e¹ = (the function (x,y) ↦ x) and e² = (the function (x,y) ↦ y). They live in V*, not V — different space.
• Dual is dimension-doubling. No — dim V* = dim V in finite-dim. Confusion may come from mixing V and V*; their direct sum V ⊕ V* has dimension 2n, but each individually has dimension n. Tangent + cotangent bundles together (the symplectic phase space) have dimension 2n, but each tangent space alone has dimension n.

Worked example

Take V = ℝ³ with standard basis e₁ = (1,0,0), e₂ = (0,1,0), e₃ = (0,0,1). The dual space V* consists of linear functionals ℝ³ → ℝ. The dual basis e¹, e², e³ extracts coordinates: e¹(x,y,z) = x, e²(x,y,z) = y, e³(x,y,z) = z.

A general functional φ(x,y,z) = 2x − 3y + 7z can be written 2e¹ − 3e² + 7e³. As a row vector in the dual basis: [2, −3, 7]. Its action on a vector v = (a,b,c) is the matrix product [2 −3 7][a b c]^T = 2a − 3b + 7c. This is why functionals are often drawn as "row vectors" while elements of V are "column vectors" — they pair via matrix multiplication.

Now change basis. Let f₁ = (1,1,0), f₂ = (1,−1,0), f₃ = (0,0,1) — still a basis of V. The dual basis {f¹, f², f³} is determined by fⁱ(fⱼ) = δⁱⱼ. Solving: f¹(x,y,z) = (x+y)/2, f²(x,y,z) = (x−y)/2, f³(x,y,z) = z. Note these are different functionals than e¹, e², e³ — even though both are dual bases, they're dual to different bases of V. This shows there's no basis-free identification of V with V*.

Double dual: V** consists of functionals on V*. The map ev: V → V** sends v to the functional "evaluate at v": ev(v)(φ) = φ(v). For v = (1,2,3): ev(v)(φ) = φ(1,2,3). If φ = 2e¹ − 3e² + 7e³, then ev(v)(φ) = 2 − 6 + 21 = 17. The map ev is a vector-space isomorphism (in finite-dim), and it's the canonical one — definable without choosing a basis.