Sobolev Duality and Structure

Duality, reflexivity, and the Sobolev scale¶

One of the main rewards of working with Sobolev spaces instead of the full distributional framework is that Sobolev spaces are Banach spaces, and for $1 < p < \infty$ , they are reflexive. This means that the entire duality machinery from the duality chapter (weak topology, weak- $*$ topology, Banach–Alaoglu, weak compactness) applies directly, with no need for the inductive limit subtleties of $\mathcal{D}'(\Omega)$ .

The dual of $W_0^{k,p}(\Omega)$ ¶

Definition 1 ( $W_0^{k,p}(\Omega)$ and negative Sobolev spaces)

Let $W_0^{k,p}(\Omega)$ denote the closure of $C_c^\infty(\Omega)$ in $W^{k,p}(\Omega)$ . For $1 < p < \infty$ with conjugate exponent $p' = p/(p-1)$ , the negative Sobolev space is defined as the dual:

W^{-k,p'}(\Omega) := \big(W_0^{k,p}(\Omega)\big)^*.

(1)

The duality pairing extends the distributional pairing: for $T \in W^{-k,p'}(\Omega)$ and $\varphi \in C_c^\infty(\Omega) \subset W_0^{k,p}(\Omega)$ ,

\langle T, \varphi \rangle_{W^{-k,p'} \times W_0^{k,p}} = \langle T, \varphi \rangle_{\mathcal{D}' \times \mathcal{D}}.

(2)

So every element of $W^{-k,p'}$ is a distribution, but a tame one: it extends continuously from test functions to all of $W_0^{k,p}$ .

Remark 1

Concrete characterization¶

The dual characterization makes negative Sobolev spaces concrete. An element $T \in W^{-k,p'}(\Omega)$ can always be written as

\langle T, u \rangle = \sum_{|\alpha| \leq k} \int_\Omega f_\alpha \, D^\alpha u \, dx

(3)

for some $f_\alpha \in L^{p'}(\Omega)$ . Conversely, every such expression defines an element of $W^{-k,p'}$ . In other words:

W^{-k,p'}(\Omega) = \left\{ \sum_{|\alpha| \leq k} D^\alpha f_\alpha : f_\alpha \in L^{p'}(\Omega) \right\}

(4)

where the derivatives are taken in the distributional sense. This is a representation theorem analogous to the Riesz representation (Theorem 1): every continuous linear functional on $W_0^{k,p}$ comes from $L^{p'}$ data, with the pairing mediated by distributional derivatives.

The Sobolev scale¶

Sobolev spaces fit into the distributional framework as a graded scale of increasingly regular subspaces of $\mathcal{D}'(\Omega)$ :

\cdots \subset H^{-2}(\Omega) \subset H^{-1}(\Omega) \subset L^2(\Omega) \subset H^1(\Omega) \subset H^2(\Omega) \subset \cdots

(5)

The index $s$ measures regularity: positive $s$ means the function has $s$ derivatives in $L^2$ ; negative $s$ means the distribution is “mildly singular,” living $|s|$ levels below $L^2$ .

Example 1 (The Dirac delta in the Sobolev scale)

The Dirac delta $\delta \in \mathcal{D}'(\mathbb{R}^d)$ is not in $L^p$ for any $p$ , but it belongs to negative Sobolev spaces. In dimension $d$ with $p = 2$ , $\delta \in H^{-s}(\mathbb{R}^d)$ for $s > d/2$ , because the Sobolev embedding theorem guarantees that pointwise evaluation is bounded on $H^s$ when $s > d/2$ :

|\langle \delta, \varphi \rangle| = |\varphi(0)| \leq C \|\varphi\|_{H^s}.

(6)

In one dimension, $\delta \in H^{-s}(\mathbb{R})$ for $s > 1/2$ . In three dimensions, $\delta \in H^{-s}(\mathbb{R}^3)$ for $s > 3/2$ . The more singular the object, the further down the scale it lives.

Example 2 (Derivatives shift the scale)

Differentiation moves a distribution down the Sobolev scale: if $u \in H^s(\Omega)$ , then $D^\alpha u \in H^{s - |\alpha|}(\Omega)$ . For instance:

If $u \in H^1$ , then $u' \in L^2 = H^0$ .
If $u \in L^2$ , then $u' \in H^{-1}$ . The derivative exists as a distribution but is not a function.
The Heaviside function $H \in L^2(0,1)$ has $H' = \delta_{1/2} \in H^{-1}(0,1)$ .

The Sobolev index precisely tracks how many derivatives a function can “afford” before leaving $L^2$ .

The limits of the Sobolev scale¶

As $k$ increases, $H^{-k}$ grows larger, containing increasingly singular distributions. It is natural to ask whether the union exhausts all of $\mathcal{D}'(\Omega)$ .

Proposition 1

The inclusion

\bigcup_{k=0}^\infty H^{-k}(\Omega) \subsetneq \mathcal{D}'(\Omega)

(7)

is strict. The Sobolev scale captures exactly the finite-order distributions. Distributions of infinite order (where the order of the continuity estimate grows without bound as the compact set $K$ expands) escape every level of the scale.

Example 3 (A distribution of infinite order)

On $\mathbb{R}$ , define

T = \sum_{n=1}^\infty \delta^{(n)}(x - n).

(8)

On a compact set containing $x = n$ , the term $\delta^{(n)}(x - n)$ requires $k \geq n$ derivatives of the test function. No single $k$ works for all compact sets, so $T \notin H^{-k}(\mathbb{R})$ for any $k$ . But $T \in \mathcal{D}'(\mathbb{R})$ : for any test function $\varphi$ (which has compact support), only finitely many terms in the sum are nonzero, so $\langle T, \varphi \rangle$ is well-defined.

This is the precise sense in which $\mathcal{D}'(\Omega)$ is larger than the Sobolev scale: it accommodates distributions whose singularity worsens without bound across the domain. For such objects, the inductive limit topology of $\mathcal{D}$ is genuinely necessary; no single Banach space in the scale can hold them.

In practice, distributions arising from PDE theory almost always have finite order (often order $\leq 2$ ), so the Sobolev scale is sufficient for most applications.

Reflexivity: weak and weak- $*$ for free¶

Theorem 1 (Reflexivity of Sobolev spaces)

For $1 < p < \infty$ and $k \geq 0$ , the spaces $W^{k,p}(\Omega)$ and $W_0^{k,p}(\Omega)$ are reflexive Banach spaces. In particular, $H^k(\Omega) = W^{k,2}(\Omega)$ is a reflexive Hilbert space.

For $p = 1$ or $p = \infty$ , the Sobolev spaces are not reflexive.

Proof 1

Reflexivity means the canonical embedding $J : W^{k,p} \to (W^{k,p})^{**}$ is surjective, so the double dual adds nothing new. This has immediate consequences from the duality chapter:

Corollary 1 (Weak compactness in Sobolev spaces)

For $1 < p < \infty$ :

Banach–Alaoglu applies directly. The closed unit ball of $W^{k,p}(\Omega)$ is weak- $*$ compact in $(W^{-k,p'})^* = W^{k,p}$ . But because $W^{k,p}$ is reflexive, the weak and weak- $*$ topologies coincide, so the ball is weakly compact.
Every bounded sequence has a weakly convergent subsequence. If $(u_n)$ is bounded in $W^{k,p}(\Omega)$ , there exists a subsequence $u_{n_j} \rightharpoonup u$ in $W^{k,p}(\Omega)$ .
Weak sequential completeness. Every weakly Cauchy sequence in $W^{k,p}(\Omega)$ converges weakly.

This is the payoff: in $\mathcal{D}'(\Omega)$ , we needed the Montel property and the non-trivial inductive limit machinery to extract convergent subsequences. In $W^{k,p}$ for $1 < p < \infty$ , it is a direct consequence of reflexivity and Banach–Alaoglu, results we already proved in the duality chapter.

Remark 2 (The duality chapter toolkit, applied)

Here is how the duality chapter results specialize to Sobolev spaces:

Duality chapter result	Sobolev space consequence
Weak topology $\sigma(X, X^*)$ is Hausdorff	Weak limits in $W^{k,p}$ are unique
Banach–Alaoglu: unit ball is weak- $*$ compact	Bounded sets in $W^{k,p}$ are weakly precompact ( $1 < p < \infty$ )
Reflexive $\Rightarrow$ weak = weak- $*$	No need to distinguish weak and weak- $*$ on $W^{k,p}$
Weak convergence + compact operator $\Rightarrow$ strong convergence	$u_n \rightharpoonup u$ in $H^1$ $\Rightarrow$ $u_n \to u$ in $L^2$ (Rellich–Kondrachov)
Direct method: minimize over weakly compact sets	Existence of PDE solutions by energy minimization in $H^1$

The last row is the reason Sobolev spaces exist: they are the Banach spaces where the direct method works. Reflexivity gives weak compactness (subsequences exist), Rellich–Kondrachov gives compact embedding (the subsequences converge strongly in a weaker norm), and weak lower semicontinuity of the energy functional closes the argument.

Why $p = 1$ and $p = \infty$ are different¶

For $p = 1$ , the Sobolev space $W^{k,1}(\Omega)$ is not reflexive (just as $L^1$ is not reflexive). Its dual $(W_0^{k,1})^*$ contains measures and other singular objects. Bounded sequences in $W^{k,1}$ need not have weakly convergent subsequences. The direct method fails, and minimization problems over $W^{1,1}$ require the theory of functions of bounded variation ( $BV$ ) instead.

For $p = \infty$ , the space $W^{k,\infty}(\Omega)$ is essentially the Lipschitz functions (for $k = 1$ ), and its dual is again non-reflexive. The weak- $*$ topology on $(W^{k,\infty})^*$ does not coincide with the weak topology on $W^{k,\infty}$ .

The range $1 < p < \infty$ is the sweet spot where duality theory works cleanly.