Distributions and Sobolev Spaces Introduction

Distributions (generalized functions) extend the notion of a function so that every distribution is infinitely differentiable, while Sobolev spaces

W^{k, p}

measure the integrability of a function together with its weak (distributional) derivatives up to order

k

. Together they form the analytic backbone of modern PDE theory: weak formulations, existence/uniqueness via Hilbert-space methods, and regularity estimates. They apply wherever derivatives must be interpreted past the limits of classical calculus — shocks, point sources, finite elements, quantum mechanics. Foundational work by Sergei Sobolev (1930s) and Laurent Schwartz (1940s, Fields Medal 1950).

A distribution on an open set $Ω \subseteq R^{n}$ is a continuous linear functional on the space $D (Ω) = C_{c}^{\infty} (Ω)$ of smooth, compactly supported "test functions." Because differentiation can be transferred onto the test function by integration by parts, distributions can always be differentiated, infinitely often — even the Dirac delta and discontinuous functions. Sobolev spaces $W^{k, p} (Ω)$ collect those functions in $L^{p}$ whose distributional derivatives up to order $k$ also lie in $L^{p}$ ; they are Banach spaces (Hilbert when $p = 2$ , written $H^{k}$ ). This framework matters because most physically relevant PDEs have no classical solutions, yet possess weak solutions living in Sobolev spaces. The surprising fact: by enlarging the class of admissible objects, one gains — existence and well-posedness become routine, and classical smoothness is then recovered afterward via regularity theorems.

The Core Insight

Differentiation is a chore for functions but trivial for distributions. Classical calculus forbids differentiating non-smooth functions; distribution theory dodges this by defining the derivative through the identity

⟨ u^{'}, φ ⟩ = - ⟨ u, φ^{'} ⟩

, which always makes sense because the test function

φ

is smooth. Sobolev spaces then quantify how many such weak derivatives an object has and how integrable they are — turning "regularity" into a measurable, topological property and converting PDE problems into geometry in Banach/Hilbert spaces.

Formal Statement

Test Functions and Distributions

Let $Ω \subseteq R^{n}$ be open. The space of test functions is

D (Ω) = C_{c}^{\infty} (Ω) = {φ \in C^{\infty} (Ω) : supp φ is compact, supp φ \subset Ω} .

⟳A function is in

C_{c}^{\infty}

iff it is infinitely differentiable and vanishes outside a compact set; the standard bump

φ (x) = e^{- 1/ (1 - ∣ x ∣^{2})} 1_{∣ x ∣ < 1}

exhibits non-emptiness.

A sequence $φ_{j} \to φ$ in $D (Ω)$ iff there is a single compact $K$ containing all $supp φ_{j}$ and $\partial^{α} φ_{j} \to \partial^{α} φ$ uniformly for every multi-index $α$ .

A distribution $u$ is a continuous linear functional $u : D (Ω) \to R$ (or $C$ ); the space of all distributions is $D^{'} (Ω)$ . Continuity is equivalent to: for each compact $K \subset Ω$ there exist $C > 0$ and $N \in N$ with

∣ ⟨ u, φ ⟩ ∣ \leq C ∣ α ∣ \leq N \sum x \in K sup ∣ \partial^{α} φ (x) ∣ \forall φ \in D (Ω), supp φ \subset K .

⟳This seminorm bound is the explicit form of sequential continuity on the LF-space

D (Ω)

;

N

is the local order of

u

Distributional (Weak) Derivative

For $u \in D^{'} (Ω)$ and multi-index $α = (α_{1}, \dots, α_{n})$ , the derivative $\partial^{α} u \in D^{'} (Ω)$ is defined by

⟨ \partial^{α} u, φ ⟩ = (- 1)^{∣ α ∣} ⟨ u, \partial^{α} φ ⟩, ∣ α ∣ = α_{1} + \dots + α_{n} .

⟳Repeated integration by parts on a smooth

u

, with boundary terms killed by compact support of

φ

, forces the sign

(- 1)^{∣ α ∣}

; one then promotes this identity to a definition for arbitrary

u

A locally integrable $u \in L_{loc}^{1} (Ω)$ is identified with the distribution $φ \mapsto \int_{Ω} u φ d x$ . When $\partial^{α} u$ (distributional) is again represented by an $L_{loc}^{1}$ function $g$ , we call $g$ the weak derivative:

\int_{Ω} g φ d x = (- 1)^{∣ α ∣} \int_{Ω} u \partial^{α} φ d x \forall φ \in C_{c}^{\infty} (Ω) .

⟳Specialize [3] to a regular distribution and demand that the result be represented by an

L_{loc}^{1}

kernel.

Sobolev Spaces

Fix $k \in N_{0}$ and $1 \leq p \leq \infty$ . The Sobolev space is

W^{k, p} (Ω) = {u \in L^{p} (Ω) : \partial^{α} u \in L^{p} (Ω) for all ∣ α ∣ \leq k},

with norm

∥ u ∥_{W^{k, p} (Ω)} = (∣ α ∣ \leq k \sum ∥ \partial^{α} u ∥_{L^{p} (Ω)}^{p})^{1/ p} (1 \leq p < \infty), ∥ u ∥_{W^{k, \infty}} = ∣ α ∣ \leq k max ∥ \partial^{α} u ∥_{L^{\infty}} .

⟳Each

\partial^{α} u

is the weak derivative from [4]; the norm is the

ℓ^{p}

-aggregate of the component

L^{p}

norms, making

W^{k, p}

a closed subspace-style construction.

When $p = 2$ we write $H^{k} (Ω) := W^{k, 2} (Ω)$ , a Hilbert space with inner product

⟨ u, v ⟩_{H^{k}} = ∣ α ∣ \leq k \sum \int_{Ω} \partial^{α} u \overline{\partial^{α} v} d x .

⟳Polarize the

p = 2

norm; completeness follows from completeness of

L^{2}

on each derivative component.

Define $W_{0}^{k, p} (Ω)$ as the closure of $C_{c}^{\infty} (Ω)$ in the $W^{k, p}$ norm (zero boundary trace), and $H^{- k} (Ω) = (H_{0}^{k} (Ω))^{'}$ for negative-order spaces.

Completeness, Embedding, Density

(Banach/Hilbert) $W^{k, p} (Ω)$ is a Banach space for $1 \leq p \leq \infty$ ; $H^{k} (Ω)$ is a separable Hilbert space for $1 \leq p < \infty$ .
(Meyers–Serrin, " $H = W$ ") For $1 \leq p < \infty$ , $C^{\infty} (Ω) \cap W^{k, p} (Ω)$ is dense in $W^{k, p} (Ω)$ .
(Sobolev Embedding, $Ω = R^{n}$ , $k p < n$ ) With $p^{*} = \frac{n p}{n - k p}$ (the Sobolev conjugate),

W^{k, p} (R^{n}) ↪ L^{p^{*}} (R^{n}), ∥ u ∥_{L^{p^{*}}} \leq C (n, k, p) ∥ u ∥_{W^{k, p}} .

⟳For

k = 1

the Gagliardo–Nirenberg–Sobolev inequality follows from the fundamental theorem of calculus along each axis plus the generalized Hölder inequality; iterate for higher

k

(Morrey, $k p > n$ ) $W^{k, p} (R^{n}) ↪ C^{0, γ} (R^{n})$ with Hölder exponent $γ = k - n / p$ (mod 1), so functions become continuous (even Hölder).

(Rellich–Kondrachov) If $Ω$ is bounded with Lipschitz boundary and $k p < n$ , then $W^{k, p} (Ω) ↪↪ L^{q} (Ω)$ compactly for all $1 \leq q < p^{*}$ .

Why It Holds

Sobolev Inequality for $W^{1,1}(\mathbb R^n)$ — the prototype

We prove the sharp scaling case $p = 1$ , $k = 1$ : $∥ u ∥_{L^{n / (n - 1)}} \leq C ∥\nabla u ∥_{L^{1}}$ for $u \in C_{c}^{\infty} (R^{n})$ .

Step 1 — One-dimensional control. For each coordinate

i

, the fundamental theorem of calculus gives

∣ u (x) ∣ \leq \int_{- \infty}^{\infty} ∣ \partial_{i} u (x_{1}, \dots, t, \dots, x_{n}) ∣ d t =: g_{i} (\overset{x}{^}_{i}),

⟳

u (x) = \int_{- \infty}^{x_{i}} \partial_{i} u d t

and

u

has compact support, so bound the integral over all of

R

;

g_{i}

does not depend on

x_{i}

Step 2 — Multiply the $n$ estimates. Raise to the power

1/ (n - 1)

and multiply over

i

∣ u (x) ∣^{\frac{n}{n - 1}} \leq i = 1 \prod n (\int_{R} ∣ \partial_{i} u ∣ d t)^{\frac{1}{n - 1}} .

⟳Each factor

∣ u ∣^{1/ (n - 1)} \leq g_{i}^{1/ (n - 1)}

; multiplying all

n

of them produces

∣ u ∣^{n / (n - 1)}

on the left.

Step 3 — Integrate iteratively with generalized Hölder. Integrate in

x_{1}

, pulling out the factor independent of

x_{1}

, and apply Hölder with

n - 1

exponents each equal to

n - 1

. Repeating in

x_{2}, \dots, x_{n}

yields

\int_{R^{n}} ∣ u ∣^{\frac{n}{n - 1}} d x \leq i = 1 \prod n (\int_{R^{n}} ∣ \partial_{i} u ∣ d x)^{\frac{1}{n - 1}} .

⟳At each integration the term lacking the integration variable comes out; the remaining

n - 1

terms are paired against each other via Hölder’s inequality with conjugate exponents

(n - 1, \dots, n - 1)

Step 4 — AM–GM finish. Take the

(n - 1) / n

power and apply the arithmetic–geometric mean inequality to the product:

∥ u ∥_{L^{n / (n - 1)}} \leq i = 1 \prod n ∥ \partial_{i} u ∥_{L^{1}}^{1/ n} \leq \frac{1}{n} i = 1 \sum n ∥ \partial_{i} u ∥_{L^{1}} \leq ∥\nabla u ∥_{L^{1}} .

⟳Geometric mean

\leq

arithmetic mean of the

n

quantities

∥ \partial_{i} u ∥_{L^{1}}

; absorb constants.

Density of $C_{c}^{\infty}$ extends the inequality to all of $W^{1, 1}$ . The general $W^{1, p}$ case applies this to $∣ u ∣^{γ}$ with $γ = \frac{( n - 1 ) p}{n - p}$ and uses Hölder once more. $■$

Laurent Schwartz | 1915–2002

Schwartz unified decades of "improper" objects — Heaviside’s operational calculus, Dirac’s delta, Hadamard’s finite parts — into the rigorous theory of distributions in Théorie des distributions (1950–51). His decisive idea was duality: define generalized functions as continuous linear functionals on a carefully topologized space of test functions, so that operations on the singular object are dual to operations on smooth functions. For this he received the Fields Medal (1950). He emphasized that the "trick" was less in the objects than in choosing the right test-function topology, the inductive-limit space $D$ , whose dual is large enough to host delta and its derivatives.

Building Intuition

Think of a distribution as a measuring device, not a pointwise object. You never ask "what is $u (x)$ ?"; you ask "what is the weighted average $⟨ u, φ ⟩$ against this smooth probe $φ$ ?" The Dirac delta is the device that returns $φ (0)$ — it reads a value, not an average, which no ordinary function can do.

Differentiation works by passing the derivative to the probe. Since you can’t differentiate $u$ directly, you let the smooth $φ$ absorb the derivative, paying a sign $(- 1)^{∣ α ∣}$ as the price of integration by parts. Because $φ$ is infinitely smooth, this never fails — hence every distribution is infinitely differentiable.

Sobolev spaces add a bookkeeping ledger: not just "does the derivative exist as a distribution," but "is it integrable to power $p$ ?" The number $k$ counts derivatives; $p$ counts integrability. The Sobolev embedding theorem is the exchange rate: spend enough derivatives (large $k p$ relative to dimension $n$ ) and you buy back classical continuity. The competition is always against dimension $n$ — high dimension makes regularity expensive.

Duality of differentiation: the weak derivative passes $d/dx$ onto the smooth test function

Worked Examples

example[Weak derivative of the Heaviside step] Given: the Heaviside function $H (x) = 1$ for $x > 0$ , $0$ for $x < 0$ on $R$ . Find $H^{'}$ in $D^{'} (R)$ .

Step 1. Apply [3] with

∣ α ∣ = 1

⟨ H^{'}, φ ⟩ = - ⟨ H, φ^{'} ⟩ = - \int_{R} H (x) φ^{'} (x) d x = - \int_{0}^{\infty} φ^{'} (x) d x .

Step 2. Compute the elementary integral;

φ

has compact support so

φ (\infty) = 0

- \int_{0}^{\infty} φ^{'} (x) d x = - [φ (\infty) - φ (0)] = φ (0) .

Step 3. The functional

φ \mapsto φ (0)

is by definition the Dirac delta

δ_{0}

H^{'} = δ_{0} in D^{'} (R) .

Consequence (Sobolev):

H \in W_{loc}^{1, p}

? No —

δ_{0} \in / L^{p}

for any

p

. So

H \in / W^{1, p}

. But

H \in W^{0, p} = L^{p}

locally. The jump costs all regularity beyond

L^{\infty}

Is $u(x)=|x|^{-s}$ in $H^1$ of the unit ball in $\mathbb R^3$?

Given:

Ω = B (0, 1) \subset R^{3}

(

n = 3

u (x) = ∣ x ∣^{- s}

s > 0

. Determine the threshold

s

for which

u \in H^{1} (Ω) = W^{1, 2}

. Step 1 — $L^{2}$ condition. Using spherical coordinates

d x = 4 π r^{2} d r

\int_{Ω} ∣ u ∣^{2} d x = 4 π \int_{0}^{1} r^{- 2 s} r^{2} d r = 4 π \int_{0}^{1} r^{2 - 2 s} d r,

finite iff

2 - 2 s > - 1

, i.e.

s < 3/2

. Step 2 — Gradient.

∣\nabla u ∣ = s ∣ x ∣^{- s - 1}

, so

\int_{Ω} ∣\nabla u ∣^{2} d x = 4 π s^{2} \int_{0}^{1} r^{- 2 s - 2} r^{2} d r = 4 π s^{2} \int_{0}^{1} r^{- 2 s} d r,

finite iff

- 2 s > - 1

, i.e.

s < 1/2

. Step 3 — Intersect. The binding constraint is the gradient:

u = ∣ x ∣^{- s} \in H^{1} (B_{1} \subset R^{3}) ⟺ 0 < s < \frac{1}{2} .

Cross-check with embedding. Here

k p = 1 \cdot 2 = 2 < n = 3

, so

H^{1}

does not embed into

L^{\infty}

in 3D — unbounded functions like

∣ x ∣^{- s}

s < 1/2

, are legitimately in

H^{1}

. The Sobolev conjugate is

p^{*} = \frac{3 \cdot 2}{3 - 2} = 6

; indeed

∣ x ∣^{- s} \in L^{6}

iff

6 s < 3

iff

s < 1/2

, consistent.

Common Misconceptions

"Weak derivative = classical derivative wherever the function is differentiable." False in general. The weak derivative is determined by behavior on sets of positive measure plus singular structure. The Cantor–Lebesgue (devil's staircase) function

c (x)

has classical derivative

0

almost everywhere, yet its distributional derivative is the singular Cantor measure, not the zero function. So

c \in / W^{1, 1}

even though

c^{'} = 0

a.e. — the pointwise-a.e. derivative ignores the measure concentrated on a null set. "Sobolev functions have pointwise values." Generally false. Elements of

W^{k, p}

are equivalence classes modulo a.e. equality. A "value at a point" makes sense only after the trace theorem (on lower-dimensional sets) or Morrey embedding (when

k p > n

gives a continuous representative). Asking

u (0)

for

u \in H^{1} (R^{3})

is meaningless without further structure.

Multiplication of distributions is impossible in general

There is no associative product on $D^{'} (R)$ extending pointwise multiplication and the usual calculus (Schwartz's impossibility). Naively, $δ \cdot H$ would need to satisfy contradictory rules. Consider, with $H^{2} = H$ and $H^{'} = δ$ :

δ = (H)^{'} = (H^{2})^{'} = 2 H H^{'} = 2 H δ ⟹ δ = 2 H δ .

Yet also

H^{3} = H

gives

δ = (H^{3})^{'} = 3 H^{2} H^{'} = 3 H δ

, so

2 H δ = 3 H δ

, forcing

H δ = 0

and hence

δ = 0

— absurd. Moral: products like

δ^{2}

H δ

are undefined; one needs Colombeau algebras or microlocal wavefront-set conditions to multiply selectively.

Connections

Distributions and Sobolev spaces sit at the crossroads of analysis:

Fourier Transform — On $R^{n}$ , $H^{s}$ is characterized by $\int (1 + ∣ ξ ∣^{2})^{s} ∣ \overset{u}{^} (ξ) ∣^{2} d ξ < \infty$ , defining fractional and negative order spaces $H^{s}$ for all real $s$ . The transform turns differentiation into multiplication by $i ξ$ , making smoothness a decay condition on $\overset{u}{^}$ .
Lp Spaces — Sobolev spaces are built on top of $L^{p}$ ; the embedding theorems trade derivative count $k$ against integrability $p$ and dimension $n$ .
Weak Solutions of PDEs — The Lax–Milgram theorem on $H_{0}^{1}$ produces unique weak solutions of elliptic equations; this is the payoff of the whole machinery.
Hilbert Spaces — $H^{k}$ inherits orthogonality, projections, and the Riesz representation theorem, enabling variational methods.
Tempered Distributions — The dual $S^{'}$ of the Schwartz space $S$ is the natural home for the Fourier transform, intermediate between $D^{'}$ and finite-energy spaces.
Measure Theory — Radon measures are exactly the distributions of order $0$ , linking the abstract dual to classical integration.

Real-World Applications

Finite Element Method (FEM). Engineering simulation (structural mechanics, ANSYS/Abaqus) discretizes the weak form in $H_{0}^{1}$ ; piecewise-polynomial subspaces approximate Sobolev functions, with convergence rates set by Sobolev embedding and Céa's lemma.
Quantum Mechanics. Wavefunctions live in $L^{2}$ ; observables like momentum $- i ℏ\nabla$ require $H^{1}$ , and the kinetic energy form $∥\nabla ψ ∥_{L^{2}}^{2}$ is the Sobolev seminorm. Self-adjointness of the Schrödinger Hamiltonian is a Sobolev-space statement.
Image Processing — Total Variation denoising. The ROF model minimizes $\int ∣\nabla u ∣ + λ ∥ u - f ∥^{2}$ over $B V$ , the space of bounded-variation functions whose distributional gradient is a finite measure — allowing sharp edges (jumps) that $H^{1}$ forbids.
Fluid Dynamics — Navier–Stokes. Leray's weak solutions reside in $L_{t}^{\infty} L_{x}^{2} \cap L_{t}^{2} H_{x}^{1}$ ; global regularity (a Clay Millennium Problem) is precisely the question of whether these weak solutions stay smooth.
Seismology / Wave Propagation. Point-source earthquakes are modeled by $δ$ forcing terms; the response (Green's function) is a distribution, integrated against receiver sensitivities.
Electromagnetism. Point and surface charges are delta and surface-measure distributions; $\nabla \cdot E = ρ / ε_{0}$ is read distributionally, giving $Δ (1/4 π ∣ x ∣) = - δ$ as the fundamental solution.
Machine Learning — Sobolev training & PINNs. Physics-informed neural networks penalize Sobolev norms of the residual; generalization bounds and kernel methods invoke RKHS that are Sobolev spaces $H^{s}$ with $s > n /2$ (Morrey regime guaranteeing continuity).

Historical Development

1822 — Fourier introduces what becomes the delta as a limit of his integral kernel, used freely but without rigor.
1894–1920s — Heaviside develops operational calculus; physicists differentiate discontinuities pragmatically, drawing criticism from rigorists.
1908 — Hadamard introduces finite parts (parties finies) of divergent integrals — an early regularization that foreshadows distributions.
1926–1927 — Dirac formalizes the delta "function" $δ$ in quantum mechanics; mathematically illegitimate but operationally indispensable.
1935–1938 — Sergei Sobolev (Leningrad/Moscow) defines generalized functions and the spaces $W^{l, p}$ (now bearing his name) to solve the wave/Cauchy problem; proves the embedding inequality. This is the genuine first rigorous theory.
1936 — Bochner & Friedrichs develop related notions of generalized derivatives and mollification ("Friedrichs mollifiers").
1944–1948 — Laurent Schwartz crystallizes the duality framework, choosing the topology on $D$ and identifying $D^{'}$ .
1950–1951 — Schwartz publishes Théorie des distributions (2 vols); receives the Fields Medal (1950). Tempered distributions $S^{'}$ and the distributional Fourier transform follow.
1955 — Lax / Sobolev-Kondrachov / Rellich mature compact embeddings; Meyers–Serrin (1964) prove " $H = W$ ", unifying two historical definitions of Sobolev spaces.
1960s–70s — Hörmander founds microlocal analysis and the wavefront set, refining where distributions can be multiplied and how singularities propagate.

Subtleties & Limits

Multiplication and nonlinearity. Distribution theory is linear by design; nonlinear PDEs (Navier–Stokes, nonlinear wave/Schrödinger, stochastic PDEs) require multiplying rough objects, which is generically forbidden (see the impossibility counterexample). Partial resolutions — Colombeau algebras, paraproducts (Bony), rough paths (Lyons), and regularity structures (Hairer, Fields Medal 2014) — provide self-consistent calculi for specific equations, but no universal nonlinear theory of distributions exists. Global regularity of Navier–Stokes (Clay Millennium) is the headline open question framed entirely in Sobolev language: do 3D weak (Leray–Hopf) solutions in

L_{t}^{2} H_{x}^{1}

remain smooth for all time, or can the

H^{1}

norm blow up? Optimal constants and rigidity. The sharp constant in the Sobolev inequality [8] and the classification of extremizers (Aubin–Talenti functions) connect to the Yamabe problem in conformal geometry; sharp traces, fractional analogues (

H^{s}

W^{s, p}

via Gagliardo seminorms), and stability of these inequalities remain active.

Distributions and Sobolev Spaces Introduction

The Core Insight

Formal Statement

Test Functions and Distributions

Distributional (Weak) Derivative

Sobolev Spaces

Completeness, Embedding, Density

Why It Holds

Sobolev Inequality for $W^{1,1}(\mathbb R^n)$ — the prototype

Laurent Schwartz | 1915–2002

Building Intuition

Duality of differentiation: the weak derivative passes $d/dx$ onto the smooth test function

Worked Examples

Is $u(x)=|x|^{-s}$ in $H^1$ of the unit ball in $\mathbb R^3$?

Common Misconceptions

Multiplication of distributions is impossible in general

Connections

Real-World Applications

Historical Development

Subtleties & Limits

See Also

Before this

Where this leads