Reflections on Duality
Published:
Every duality result answers one question: given a space of states, what are all the ways to assign weights to those states and collapse them into a single number? The dual space \(V^*\) is the totality of such weighting schemes. The bilinear pairing \(\langle \varphi, x \rangle = \varphi(x)\) is the aggregation step.
In finite dimensions this is trivial — \(V^* \cong V\), every functional is an inner product with a fixed vector, and the aggregation operator is summation. In infinite dimensions, the weighting schemes grow more singular, the aggregation operators require more careful construction, and the representation theorems that concretely identify \(V^*\) become genuine results.
The Progression
The dual spaces of successively more general function spaces form a chain:
| Primal Space | Dual Space | Aggregation | Key Theorem |
|---|---|---|---|
| \(\mathbb{R}^n\) | \(\mathbb{R}^n\) | Finite sum | Linear algebra |
| \(\ell^p\) | \(\ell^q\) | Series | Hölder’s inequality |
| \(L^p(\mu)\) | \(L^q(\mu)\) | \(\int f g \, d\mu\) | Radon–Nikodym |
| \(C_0(X)\) | \(\mathcal{M}(X)\) | \(\int f \, d\mu\) | Riesz–Markov |
| \(C_c^\infty\) | Distributions | \(\langle T, \phi \rangle\) | Distribution theory |
The aggregation mechanism generalizes at each step: finite sums \(\to\) series \(\to\) integration against densities \(\to\) integration against measures \(\to\) distributional pairings. The representation theorems in the rightmost column are the substance — they establish that the abstractly defined dual (all bounded linear functionals) is isomorphic to something concrete.
Measures Are Not Functionals
A linear functional eats a function and returns a scalar. A measure eats a measurable set and returns a scalar. These are different types of maps on different types of inputs, and a \(\sigma\)-algebra is not a vector space.
The bridge is integration. Given a measure \(\mu\), one constructs a functional:
\[\varphi_\mu(f) = \int f \, d\mu\]The map \(\mu \mapsto \varphi_\mu\) converts a set function into a linear functional on a function space. The content of Riesz–Markov is that this map is a bijection. So measures and functionals are identified by theorem, not by definition. Integration is the morphism that changes the category — from objects acting on sets to objects acting on functions.
This also pins down the role of Lebesgue’s construction: Riemann integration converts some measures into some functionals, but only Lebesgue integration is general enough to make the correspondence exhaustive.
When Measures Have Densities
The \(L^p\) dualities give densities — functions \(g \in L^q\) such that \(\varphi(f) = \int fg \, d\mu\). But this requires the representing measure to be absolutely continuous with respect to some reference measure. Absolute continuity of \(\nu\) with respect to \(\mu\) (\(\nu \ll \mu\)) means precisely that every \(\mu\)-null set is also \(\nu\)-null — the two measures agree on which sets are negligible. When this holds, Radon–Nikodym guarantees a density \(g\) such that \(d\nu = g \, d\mu\). When it fails — as with the Dirac mass \(\delta_x\) or the Cantor measure relative to Lebesgue measure — no such density exists.
The hierarchy:
\[\text{linear functional} \xleftarrow{\text{Riesz–Markov}} \text{measure} \xrightarrow[\text{when abs. cont.}]{\text{Radon–Nikodym}} \text{density}\]Radon–Nikodym bridges the measure and density levels. When it applies, the weighting scheme admits a pointwise representation. When it doesn’t, the measure is the most concrete form available.
But “no density” is relative to the function space. The Dirac mass has no density in \(L^q\), yet in the space of distributions — the dual of \(C_c^\infty\) — it is a density, the Dirac delta \(\delta_x\). The move to distributions is precisely what allows every measure (and much more besides) to be treated as a generalized density. This is part of why the distributional level sits at the bottom of the table above: it is the setting where the density representation finally becomes universal.
Summary
Duality across linear algebra, functional analysis, and measure theory is not a chain of analogies. It is a single question — what are the weighting schemes, and how do they aggregate? — refracted through spaces of increasing generality, with representation theorems serving as the prisms. The row-column convention from finite dimensions is a useful mnemonic for the pairing structure, but the content lives in the pairing \(\langle \cdot, \cdot \rangle : V^* \times V \to \mathbb{R}\) and in the specific theorems (Radon–Nikodym, Riesz–Markov) that make the identification between abstract functionals and concrete objects constructive.