Distributions and Sobolev Spaces

Why distributions?¶

Classical analysis works with functions that assign values to points. This breaks down in at least three ways.

Differentiation destroys regularity. The absolute value function $|x|$ is continuous but not differentiable at the origin. Its “derivative” should be the sign function, and the “derivative” of the sign function should be $2\delta_0$, but neither of these operations makes sense in the classical framework. More dramatically, solutions to the wave equation develop shock discontinuities, and the fundamental solution of Laplace’s equation has a point singularity. We need a derivative that always exists.

Point evaluation is too rigid. The Dirac delta “function” $\delta(x)$, which physicists write as $\int \delta(x) \varphi(x)\, dx = \varphi(0)$, is not a function: no locally integrable function can reproduce point evaluation via integration. Yet $\delta$ appears naturally as the identity for convolution, the fundamental solution of the identity operator, and the density of a point mass. We need a framework where $\delta$ is a legitimate object.

Measures are not enough. The dual of $C_c(\mathbb{R}^d)$ (continuous, compactly supported functions) is the space of Radon measures, by the Riesz representation theorem. Measures handle $\delta$ just fine (it is a point mass), but the derivative $\delta'$, defined by $\langle \delta', \varphi \rangle = -\varphi'(0)$, is not a measure. It is a genuinely new object that only exists when the test functions are differentiable. To get a framework closed under differentiation, we must test against $C_c^\infty$ rather than just $C_c$. Write $\mathcal{D}(\Omega) = C_c^\infty(\Omega)$ for this space of test functions.

Distribution theory resolves all three issues in one stroke: define a distribution to be a continuous linear functional on $\mathcal{D}(\Omega)$. Every locally integrable function, every Radon measure, and every derivative of every distribution is again a distribution. The price is that we must carefully topologize $\mathcal{D}(\Omega)$.

The two topologies¶

The word “continuous” in the definition of a distribution requires a topology on $\mathcal{D}(\Omega)$. The right topology turns out to be surprisingly subtle: no single norm can capture both “all derivatives converge uniformly” and “supports stay bounded.” The solution is an inductive limit: we decompose $\mathcal{D}(\Omega)$ into Fréchet spaces $\mathcal{D}_K$ (one for each compact $K \subset \Omega$), each equipped with the seminorms $\|\cdot\|_{C^k(K)}$, and take the finest locally convex topology making every inclusion continuous. The resulting convergence is very strong: a sequence $\varphi_n \to \varphi$ only if all supports sit in a single compact set and all derivatives converge uniformly. The details of this construction are in the next section.

Once we have $\mathcal{D}(\Omega)$ topologized, its dual $\mathcal{D}'(\Omega)$ carries the weak-$*$ topology, exactly the construction from the duality chapter. Convergence $T_n \to T$ in $\mathcal{D}'(\Omega)$ means $\langle T_n, \varphi \rangle \to \langle T, \varphi \rangle$ for every test function $\varphi$. This is convergence in the sense of distributions, one of the weakest convergence notions in analysis, yet limits are unique (the weak-$*$ topology is Hausdorff) and every Cauchy sequence converges.

The hierarchy of generalized functions¶

Distribution theory fits into a chain of increasing generality:

where $\mathcal{M}(\Omega)$ denotes Radon measures. Each inclusion is strict:

• The sign function is in $L^1_{\mathrm{loc}}$ but not continuous.

• The Dirac delta $\delta$ is a measure but not in $L^1_{\mathrm{loc}}$.

• The derivative $\delta'$ is a distribution but not a measure.

The embedding $L^1_{\mathrm{loc}}(\Omega) \hookrightarrow \mathcal{D}'(\Omega)$ given by $f \mapsto T_f$, where $\langle T_f, \varphi \rangle = \int f \varphi\, dx$, is injective and continuous. This justifies identifying a function with the distribution it defines, so we write $\langle f, \varphi \rangle$ instead of $\langle T_f, \varphi \rangle$ without ambiguity.

What comes next¶

With the overview in hand, we now develop the details:

1. Test functions and their topology: the inductive limit construction, sequential convergence, and key properties of $\mathcal{D}(\Omega)$.

2. Distributions as continuous linear functionals: the formal definition, the weak-$*$ topology, separation properties, and the fundamental examples.

3. Operations on distributions: differentiation, multiplication by smooth functions, and convolution, all defined by duality.

4. Sobolev spaces $W^{k,p}(\Omega)$: where distributional derivatives meet $L^p$ integrability, giving Banach spaces suited to PDE theory.