I have often encountered Hölder continuity in books on analysis, but the books I've read tend to pass over Hölder functions quickly, without developing applications. While the definition seems natural enough, it's not clear to me what we actually gain from knowing that a function is $\alpha$-Hölder continuous, for some $\alpha<1$.

I have some guesses, but they are just guesses: do $\alpha$-Hölder conditions give rise to useful weak solution concepts in PDEs? Are there important results that apply only to $\alpha$-Hölder functions, for some fixed $\alpha$? For $\alpha=1$ (Lipschitz continuity) the answer to both of these questions seems to be yes, but I know nothing for lower values of $\alpha$.

I'd be interested in answers that describe specific applications, as well as answers that give more of a ''big picture''.

  • 5
    $\begingroup$ We have a great regularity theory for elliptic equations with Holder continuous coefficients: The schauder estimates $\endgroup$ Commented Oct 8, 2021 at 15:48

2 Answers 2


Hölder continuous functions do not give rise to useful weak solutions in any context I am aware of: there are notions of weak solutions that are continuous, but the Hölder modulus is not relevant for the definition.

While there may be some rare results that require specific Hölder moduli with $\alpha < 1$, I can not think of any that I use in my research.

So why care about Hölder continuity at all? Here are a few reasons. I will say that this is coming from a purely PDE perspective, and that Hölder spaces are at their most useful when dealing with elliptic, parabolic, and some first-order PDE. For dispersive and wave equations, the fact that Hölder norms do not interact well with the Fourier transform is a strike against them. There are other (non-PDE) areas of analysis and geometry that find Hölder spaces useful for other reasons, but that would be for another answer.


Hölder spaces have very elementary and favorable compactness properties. A sequence of functions with bounded Hölder norms will have a uniformly convergent subsequence, and the Hölder norm is lower semicontinuous under uniform convergence. Uniform convergence is extremely, surprisingly, useful when studying some types of PDE, and is often enough to pass the entire PDE to the limit. This is the case with distributional solutions of linear equations, and more strikingly with viscosity solutions.

Easy to use and understand

The theory of Hölder spaces is not very deep. Unlike Sobolev spaces, which interact in subtle ways with the geometry of a domain's boundary, contain functions that generally don't make sense pointwise, require dealing with distributional derivatives, etc., Hölder spaces are just spaces of equicontinuous functions with little else going on.

It is easy to prove that a function is Hölder continuous, and common ways of doing so line up well with the way we approach PDE. One way to do this is to prove that $$ \max_{B_r(x)} u - \min_{B_r(x)} u \leq (1 - \theta)[ \max_{B_{2r}(x)} u - \min_{B_{2r}(x)}u] $$ for some $\theta > 0$, a decay of oscillation. Iterating this gives that $u$ has some Hölder modulus at $x$. This kind of statement is one we are happy to try and prove for solutions $u$: bounding the maximum of $u$ at a given scale in terms of $u$ on a larger ball is something we actually have the tools to do, at least for elliptic equations. There are also good approaches to showing $u$ is Hölder based on Sobolev or $L^p$ bounds at every scale (Morrey/Companato inequalities), and sometimes Sobolev spaces embed into Hölder spaces directly.

Another good aspect of Hölder spaces is that they let us talk about a fractional-power increase in smoothness without having to take (fractional?) derivatives, or any derivatives at all, and without needing the Fourier transform. Not having to take derivatives is a great technical convenience (see how the improvement of oscillation above is a statement about just the solution pointwise; this is great if differentiating the equation is problematic); not having to deal with anything fractional makes everything much more explicit; not needing the Fourier transform is good news for equations that interact poorly with it.

Our best theorems are true when $\alpha \in (0, 1)$

Sure, Hölder spaces might be nice, but why not just use $\alpha = 1$? It turns out that it is much, much harder to prove something is Lipschitz, and that often it is just not true. Consider the equation $$ \Delta u = f. $$ Heuristically we expect that $u$ is two derivatives smoother than $f$, because, well, that's what the equation seems to say: some second derivatives of $u$ equal $f$. The actual positive results in this direction are that if $f \in C^{0, \alpha}$, then $u \in C^{2, \alpha}$ (Schauder), that if $f \in L^p$ then $u \in W^{2, p}$ when $p \in (1, \infty)$ (Calderon-Zygmund), some similar theorems that are $k$ derivatives up from this, and much more complicated classifications of what happens at the endpoints $\alpha = 0, \alpha = 1, p = 1, p = \infty$. In particular, none of the endpoint versions are true, they all require modifications, different spaces, etc. This fact, that harmonic analysis theorems have more complicated endpoint versions, is a running theme in the field, and means that while we would love to work with $\alpha = 1$, often we just are not allowed to.

There are other types of theorems where we can prove that there exists an $\alpha > 0$ such that solutions (or their derivatives, or something related to them) are in $C^{0, \alpha}$. Here we may not be expecting the functions to be anywhere near Lipschitz. The most famous example of this is the De Giorgi-Nash-Moser theorem.

  • 1
    $\begingroup$ +1; another example for your last point: the 3D incompressible Euler equations is locally well-posed in $\alpha$-Holder spaces, but ill-posed in $C^k$ spaces; the ill-posedness is a theorem of D. Li and the late J. Bourgain $\endgroup$ Commented Oct 9, 2021 at 5:00
  • 1
    $\begingroup$ Would you mind telling me what the $W^{m,p}$ class is, I haven't come across that notation before and am interested in the result. Also Welcome to MSE :) $\endgroup$ Commented Oct 9, 2021 at 10:27
  • 1
    $\begingroup$ $W^{m,p}$ are Sobolev spaces en.wikipedia.org/wiki/Sobolev_space @user2628206 $\endgroup$
    – JackT
    Commented Oct 9, 2021 at 13:14
  • 1
    $\begingroup$ The Brownian motion as well, isn't it? The regularity of Brownian paths is expressed quite precisely by it being $\alpha$-Holder continuous for all $\alpha< \frac 12$ but not $\alpha$-Holder continuous for $\alpha \geq \frac 12$. In fact, this generalizes to a lot of solutions of various SPDEs, where the solution ends up having Holder-regular sample paths. $\endgroup$ Commented Oct 10, 2021 at 6:17
  • 1
    $\begingroup$ Regarding dispersive PDE, there's a cool chain rule by Monica Visan in Appendix A of "The defocusing energy-critical nonlinear Schrödinger equation in higher dimensions" arxiv.org/abs/math/0508298 $\endgroup$
    – dls
    Commented Oct 10, 2021 at 17:36

Another use of Hölder continuity is in hyperbolic dynamics. As a beginning, consider a compact $C^\infty$ manifold $M$ endowed with a $C^2$ Riemannian metric $\mathfrak{g}$ and a $C^1$ diffeomorphism $f:M\to M$. $f$ is called (uniformly) hyperbolic (or Anosov) if there is an invariant splitting $TM=S(f)\oplus U(f)$, where vectors in the stable bundle $S(f)$ contracts under $Tf$ exponentially fast and vectors in the unstable bundle $U(f)$ contracts under $Tf^{-1}$ exponentially fast when measured with respect to the norms induced by the metric (hence, by compactness, any metric) (see What is the constant of hyperbolicity? for further details). Heuristically $S_x(f)$ is the set of initial conditions $y$ infinitesimally close to $x$ whose forward orbits under $f$ converge to the forward orbit of $x$ exponentially fast in the time step, and a similar interpretation holds for $U_x(f)$.

Anosov showed in "Tangent Fields of Transversal Foliations in $\Upsilon$-Systems" that if one considers the assignments $x\mapsto S_x(f)$ and $x\mapsto U_x(f)$ as sections of the Grassmannian bundle $\operatorname{Gr}(TM)\to M$, $\operatorname{Gr}(TM)_x=\operatorname{Gr}(T_xM)=\{E\,|\, E \text{ is a linear subspace of } T_xM\}$, then they are locally Hölder continuous, provided that $f$ is $C^2$ (or $C^{1,\theta}$ in the sense of Definition of Hölder Space on Manifold). Here local Hölder continuity is to be interpreted as follows: let $x\in M$ and let $r_x(\mathfrak{g})\in\mathbb{R}_{>0}$ be the injectivity radius of at $x$, so that $\exp_x^\mathfrak{g}: T_xM(0; r_x(\mathfrak{g}))\stackrel{\cong_{C^1}}{\hookrightarrow} M$ is a $C^1$ embedding of the open ball in $T_xM$ centered at $0$ with radius $r_x(\mathfrak{g})$; denote by $N_x$ the image. Then if $y\in N_x$, there is a unique geodesic from $y$ to $x$ which defines a parallel transport $\Pi_{x\leftarrow y}: T_yM\stackrel{\cong}{\to} T_xM$ which induces an (isometric) isomorphism (of homogeneous spaces) $\operatorname{Gr}(\Pi)_{x\leftarrow y}=\operatorname{Gr}(\Pi_{x\leftarrow y}):\operatorname{Gr}(T_yM)\stackrel{\cong}{\to} \operatorname{Gr}(T_xM)$. Then for $\theta\in ]0,1]$, a section $E:M\to \operatorname{Gr}(TM)$ is locally $\theta$-Hölder continuous if

$$\exists C\in\mathbb{R}_{>0},\forall x\in M,\forall y\in N_x: \Vert \operatorname{proj}(E_x)-\operatorname{proj}(\operatorname{Gr}(\Pi)_{x\leftarrow y}(E_y)) \Vert_x\leq C d(x,y)^\theta,$$

where $\operatorname{proj}(V)$ is the orthogonal projection associated to the subspace $V$.

The importance of this is this: as long as the diffeomorphism $f$ is $C^1$, so that the tangent map $Tf$ is defined, $S(f)$ and $U(f)$ are well defined, and depend continuously on the basepoint. Even when $f$ is real analytic, $S(f)$ and $U(f)$ may fail to be differentiable; which causes problems in integrating them. However, since $x\mapsto S_x(f)$ and $x\mapsto U_x(f)$ are $C^{(0,\theta)}$ for some $\theta\in]0,1]$, they uniquely integrate to transverse foliations $\mathcal{S}(f)$ and $\mathcal{U}(f)$. The leaves of $\mathcal{S}(f)$ and $\mathcal{U}(f)$ are as regular as $f$, however their transverse regularity is governed by the regularity of $x\mapsto S_x(f)$ and $x\mapsto U_x(f)$; in particular they are locally Hölder. This in turn is the crucial property that connects ergodic theory to differentiable dynamics: the stable and unstable foliations being locally Hölder guarantees the so-called absolute continuity property of these foliations; which means that the holonomies between unstable leaves and holonomies between stable leaves do not collapse the leaf volumes induced by the Riemannian metric, so that Fubini theorems along the stable and unstable foliations are available (even though these foliations are not continuously differentiable in general as mentioned above). This in turn is used to establish ergodicity of Anosov diffeomorphisms preserving a Borel probability measure of Lebesgue class.

The notion of hyperbolicity we defined above has been generalized in various directions (partial hyperbolicity, nonuniform hyperbolicity, ...); and adaptations of the above argument is still very relevant in these generalized situations.


You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .