Sobolev embedding theorem

If $s>d/2$, then $H^s({\mathbb{R}}^d)$ embeds continuously into bounded continuous functions.

For $d<p<\infty$, every Sobolev space element $u\in W^{1,p}({\mathbb{R}}^d)$ has a Hölder continuous function representative of exponent $1-d/p$, with the displayed bound. It also satisfies $\Vert u{\Vert }_{\infty }\le C_{d,p}(\Vert u{\Vert }_p+\Vert \nabla u{\Vert }_p)$. Averaging the fundamental theorem of calculus along a line segment over a ball givesThe Holder inequality bounds this by $C_{d,p}{r}^{1-d/p}\Vert \nabla u{\Vert }_p$, because $(d-1)p^{\prime }<d$. To compare two ball averages, translate a ball along the segment between their centres and use the same fundamental theorem of calculus along a line segment. These bounds give the displayed estimate. Density of smooth functions in a Sobolev space supplies the representative for nonsmooth $u$. For $p=\infty$, the corresponding conclusion is Lipschitz continuity.

For $p>n$, the Hölder continuous function representative $u^{\ast }$ of $u\in W^{1,p}({\mathbb{R}}^n)$ has a classical Frechet derivative at almost every point, equal to its weak derivative. At a point $x$ where the $p$-mean oscillation of $Du$ tends to zero, apply the Morrey inequality on a cube to $u^{\ast }(y)-u^{\ast }(x)-Du(x)\cdot (y-x)$. On a cube of side comparable to $|y-x|$, its oscillation is at mostThe required points have full measure by the Lebesgue differentiation theorem.

For a cube $Q$ of side length $r$ in ${\mathbb{R}}^n$, $p>n$, and a continuously differentiable function $v$, its average $v_Q=|Q{|}^{-1}{\int }_{Q}v$ satisfiesAveraging the fundamental theorem of calculus along a line segment from $y$ to $z\in Q$ and changing variables gives $|v(y)-v_Q|\le C_n{\int }_Q|Dv(w)||w-y{|}^{1-n}dw$. The Holder inequality applies because $(n-1)p^{\prime }<n$, and the kernel norm is $O(r^{1-n/p})$. Approximation extends the estimate to the continuous representative of a Sobolev space function. On ${\mathbb{R}}^n$ it yields a Hölder continuous function of exponent $1-n/p$ representing every $W^{1,p}$ class.

On a two-dimensional domain containing the origin, a logarithmic singularity such as $u(r)=(\log (1/r){)}^{\alpha }$ with $0<\alpha <1/2$ belongs locally to $H^1$ but is unbounded. Thus the critical first-order Sobolev space does not embed into $L^{\infty }$.

A Sobolev inequality bounds a function in one $L^q$ norm by its weak derivatives in another $L^p$ norm. In three dimensions, the first-order case $\Vert u{\Vert }_{L^6}\le C\Vert \nabla u{\Vert }_{L^2}$ is the endpoint estimate behind the continuous embedding $H^1\hookrightarrow L^6$.

For a continuously differentiable function $F$ on an interval $I=[t-\delta /2,t+\delta /2]$,To prove it, average $|F(t){|}^2-|F(u){|}^2$ over $u\in I$ using the fundamental theorem of calculus. On either half of $I$, the resulting derivative kernel has absolute value at most $1/2$. Since $|(|F{|}^2)\prime |\le 2|FF\prime |$, the displayed bound follows.

On a bounded three-dimensional domain, interpolation between $L^2$ and the Sobolev inequality $H^1\hookrightarrow L^6$ gives

For $d\ge 3$ and $u\in H^1({\mathbb{R}}^d)$,

For $1\le p<n$, the Sobolev conjugate exponent $p^{\ast }$ is defined byThe first-order Sobolev inequality maps one derivative in $L^p({\mathbb{R}}^n)$ to the function in $L^{p^{\ast }}({\mathbb{R}}^n)$.

The Gagliardo-Nirenberg interpolation inequalities bound an intermediate derivative norm by powers of a higher derivative norm and a lower norm. In two dimensions, one useful case is

On a bounded Lipschitz domain $\Omega \subset {\mathbb{R}}^d$, the Sobolev embedding $W^{1,p}(\Omega )\hookrightarrow L^q(\Omega )$ is compact whenever $q$ lies strictly below the critical Sobolev exponent.

Translations of one nonzero compactly supported smooth function have equal $W^{1,p}$ norms. On an unbounded domain they can have pairwise disjoint supports, so their mutual $L^p$ distances stay fixed and no subsequence converges in $L^p$.

For every bounded open $\Omega \subset {\mathbb{R}}^d$, the inclusion $H_0^1(\Omega )\hookrightarrow L^2(\Omega )$ is compact.

By the definition of $H_0^1(\Omega )$ as the closure of compactly supported smooth functions, extension by zero maps it continuously into $H^1({\mathbb{R}}^d)$ without creating a boundary distribution.

For $u\in H^1({\mathbb{R}}^d)$, Plancherel gives

Weak convergence on a bounded domain gives pointwise convergence of Fourier transforms and dominated convergence on bounded frequency balls. A uniform derivative bound controls the complementary high-frequency tails.

On bounded $\Omega$ with $V\in L^{\infty }(\Omega )$, weak convergence in $H_0^1(\Omega )$ makes the Dirichlet term lower semicontinuous and, by Rellich compactness, makes the potential term $\int V{u}^2$ continuous. Thusis weakly lower semicontinuous.

The infimum of $E(u)$ over $\Vert u{\Vert }_2=1$ is attained. A minimizing sequence is bounded in $H_0^1$ because $V$ is bounded below; weak compactness, Rellich strong $L^2$ convergence, and weak lower semicontinuity complete the direct-method argument.

Strong convergence on every bounded region combines with a tail estimate uniform in the sequence to give global strong convergence.

If local compactness makes every subsequential local limit agree with the global weak limit and the $L^2$ mass is uniformly tight, weak convergence upgrades to strong convergence.

Translations of one compactly supported bump have equal Sobolev norms. Widely separated translates have disjoint supports and remain a fixed positive distance apart in $L^2$, obstructing compactness on an unbounded domain.

A bounded sequence can fail to have a strongly convergent subsequence because its mass escapes to infinity even when its local regularity is uniformly controlled.

For $\varphi \in H^1(\mathbb{R})$ with $\Vert \varphi {\Vert }_2=1$, the dilationpreserves the $L^2$ norm and satisfies $\Vert u_R^{\prime }{\Vert }_2^2=R^{-2}\Vert {\varphi }^{\prime }{\Vert }_2^2$. Hence normalized functions on the line have Dirichlet-energy infimum zero, which no nonzero $L^2$ function attains.

A bounded function belongs to $C^{0,\alpha }({\mathbb{R}}^n)$ when

Write $\alpha =r+\beta$ with an integer $r\ge 0$ and $0<\beta \le 1$, using $r=\alpha -1$ at integer $\alpha$. For $f\in C^{r,\beta }$, subtract its degree-$r$ Taylor polynomial at $x_0$. For $r\ge 1$, the integral Taylor remainder with $f^{(r)}(t)-f^{(r)}(x_0)$ is bounded using the Hölder seminorm; for $r=0$ this is exactly Hölder continuity. The resulting bound is the displayed estimate. At integer order, the convention is $C^{r,1}$; a $C^{r+1}$ function on a compact closed interval also obeys this estimate.

Localized wavelets of support diameter $O(2^{-j})$ and $L^1$ norm $O(2^{-j/2})$, with enough vanishing moments, satisfySubtract a Taylor polynomial annihilated by the wavelet, apply the Hölder-Taylor remainder bound on its support of a function, and multiply by its $L^1$ norm. The same argument applies to localized boundary wavelets with the same cancellation; finitely many coarse scaling functions are handled separately.

The full norm is $|u{|}_{k,\alpha ;U}=|u{|}_{k;U}+[D^{k}u{]}_{\alpha ;U}$, combining the derivative supremum norm with the top-derivative Hölder seminorm. At order zero it is the sum of the supremum norm and the Hölder seminorm.

The Hölder seminorm of exponent $0<\alpha \le 1$ isIt measures Hölder continuity without including the supremum norm of the function itself.

If $f_j\to f$ uniformly and the Hölder seminorms are uniformly bounded, each fixed difference quotient converges. Taking the supremum over point pairs gives $[f{]}_{\mu }\le {\mathrm{liminf}}_j[f_j{]}_{\mu }$. The supremum norms converge, so the full Hölder norm has the same lower semicontinuity of a Hölder seminorm.

A Hölder class bounds derivatives through order $\lceil \beta \rceil -1$ and requires the highest derivative to be Hölder continuous with exponent $\beta -\lceil \beta \rceil +1$ and constant $L$.

A function is Hölder continuous with exponent $0<\alpha \le 1$ when $|f(x)-f(y)|\le C|x-y{|}^{\alpha }$ for some constant $C$.

If a real function on an interval obeys $|f(x)-f(y)|\le C|x-y{|}^{\alpha }$ for finite $C$ and $\alpha >1$, its difference quotients tend to zero everywhere. The mean value theorem makes it constant. Equivalently, divide a segment into $N$ pieces to obtain $|f(x)-f(y)|\le C|x-y{|}^{\alpha }{N}^{1-\alpha }\to 0$.

For $f\in L^2(0,1)$ and $0<\alpha <1/2$, the integral $g(x)={\int }_0^{1}f(y)|x-y{|}^{-\alpha }dy$ is well defined for every real $x$. Scaling the squared kernel difference over the real line and using the Cauchy-Schwarz inequality give $|g(x+h)-g(x)|\le C_{\alpha }\Vert f{\Vert }_2|h{|}^{1/2-\alpha }$.

If a holomorphic function on a half-rectangle satisfies $|f^{\prime }(x+iy)|\le C{y}^{-1+\alpha }$ with $0<\alpha \le 1$, then it extends to an $\alpha$-Hölder continuous function on the closed half-rectangle. Integrate vertically to height $r=|z-w|$, horizontally at that height, and vertically back; each contribution is $O(r^{\alpha })$.

On a ball of radius $R$, lower derivative norms of $u\in C^{l,\alpha }$ are bounded by an arbitrarily small multiple of the top Hölder seminorm plus a multiple of $\Vert u{\Vert }_{C^0}$. One endpoint form is

On a bounded regular domain, a bounded subset of $C^{k,\alpha }$ is relatively compact in $C^{l,\beta }$ whenever $k+\alpha >l+\beta$. This is an Arzela-Ascoli theorem argument applied successively to the derivatives.

If $s>n/2+\alpha$ with $0<\alpha \le 1$, then Fourier inversion, $|e^{it}-1|\le C_{\alpha }|t{|}^{\alpha }$, and Cauchy-Schwarz giveThe weighted frequency integral converges precisely because $s-\alpha >n/2$.