---
title: Pointwise Regularity in Analysis
url: https://www.emergentmind.com/topics/pointwise-regularity
type: topic
---

# Pointwise Regularity in Analysis

Pointwise regularity is the local study of how a function, a PDE solution, or a stochastic sample path is approximated near a fixed point by a simpler model, typically a polynomial, a homogeneous profile, or a multiscale scaling law. In the PDE literature this is commonly expressed by estimates of the form
\[
|u(x)-P(x)|\le K|x-x_0|^{k+\alpha},
\]
while in stochastic and fractal settings it is encoded by pointwise Hölder exponents, Besov-type local oscillation, or wavelet-coefficient decay. Recent work develops this perspective for fully nonlinear elliptic and parabolic equations, divergence-form equations, conical and rough domains, free boundaries, self-affine curves, and non-Gaussian processes, with an emphasis on identifying the precise local model rather than only a global modulus of continuity [2012.00324][2205.14291][1512.00948][2203.08487].

## 1. Definitions and local models

In the pointwise Schauder framework, a bounded function \(f\) is \(C^{k,\alpha}\) at \(x_0\) if there exists a polynomial \(P\) of degree \(k\) such that
\[
|f(x)-P(x)|\le K|x-x_0|^{k+\alpha}
\]
near \(x_0\). This definition is used systematically for interior and boundary regularity in fully nonlinear elliptic equations, including the boundary notions \(u\in C^{1,\alpha}(x_0)\) and \(u\in C^{2,\alpha}(x_0)\), and it is extended in general-form equations to endpoint classes \(C^k\) and \(C^{k,\mathrm{lnL}}\) corresponding to \(\alpha=0\) and \(\alpha=1\) [1901.06060][2012.00324].

A boundary variant replaces approximation of the function alone by simultaneous approximation of the boundary geometry. In one standard formulation, \(\partial\Omega\in C^{k,\alpha}(x_0)\) if, after a rigid motion, the boundary lies between two parallel tubes of thickness \(K|x'|^{k+\alpha}\) around the graph of a polynomial \(P(x')\). A further variant appears on cones: for viscosity solutions of \(F(D^2u)=0\) in \(C_\omega\), one says \(u\in C^{\alpha_\omega,\alpha}(0)\) if there exist a homogeneous profile \(\Psi\) and a scalar \(a\) such that
\[
|u(x)-a\Psi(x)|\le C|x|^\alpha \Psi(x)
\]
near the vertex. In the half-space case \(\Psi(x)=x_n\), this reduces to the usual \(C^{1,\alpha}\) boundary estimate [2205.14291][2204.09304].

In stochastic analysis, the local model is usually encoded by the pointwise Hölder exponent. For a process \(X\), one says that \(X\) is pointwise Hölder continuous of order \(\alpha>0\) at \(t\) if there exists a polynomial \(P_t\) of degree at most \(\lfloor\alpha\rfloor\) such that
\[
|X(t+h)-P_t(h)|=O(|h|^\alpha)\qquad(h\to 0),
\]
and the exponent is
\[
h_X(t)=\sup\Bigl\{\alpha>0:\;|X(t+h)-P_t(h)|=O(|h|^\alpha)\text{ as }h\to0\Bigr\}.
\]
For parameterized affine zipper curves one also distinguishes the liminf exponent
\[
\alpha(x)=\liminf_{y\to x}\frac{\log\|v(x)-v(y)\|}{\log|x-y|}
\]
from the regular exponent
\[
\alpha_r(x)=\lim_{y\to x}\frac{\log\|v(x)-v(y)\|}{\log|x-y|},
\]
when the limit exists [2203.08487][1608.04558].

A function-space formulation identifies pointwise regularity with local Besov decay. For self-affine lattice tilings, the pointwise Besov space \(B_{\infty,\infty}^s(x_0)\) is defined by the boundedness of
\[
\sup_{\ell\ge0}|M|^{\ell s}\,\mathrm{osc}_\infty^k(f;x_0,\ell),
\]
and one has the equivalence
\[
f\in C^s(x_0)\iff f\in B_{\infty,\infty}^s(x_0).
\]
Consequently,
\[
\alpha(x_0):=\sup\{s>0:f\in B_{\infty,\infty}^s(x_0)\}
\]
is exactly the pointwise Hölder exponent [1512.00948].

## 2. Interior pointwise regularity in elliptic and parabolic PDE

A broad interior theory is developed for fully nonlinear elliptic equations in general forms,
\[
F(D^2u,Du,u,x)=f(x),
\]
under a structure condition (SC2) that allows quadratic growth in the gradient. The model class includes equations such as
\[
a^{ij}(x)u_{ij}+\mu^{ij}(x)u_i u_j+b^i(x)u_i+c(x)h(u)=f(x)
\]
and
\[
\mathcal M^+(D^2u)+\mu|Du|^2+b(x)|Du|+c(x)h(u)=f.
\]
Within this framework, interior pointwise \(C^{1,\alpha}\), \(C^{2,\alpha}\), and \(C^{k,\alpha}\) regularity are obtained, together with endpoint \(C^k\) and \(C^{k,\mathrm{lnL}}\) results [2012.00324].

For divergence-form equations, an interior pointwise theory is formulated in Campanato form. In the parabolic setting, \(u\in C_2^{k,\alpha}(X_0)\) means that there exists a parabolic polynomial \(P\) of degree \(\le k\) such that
\[
\|u-P\|^*_{L^2(Q_r(X_0))}\le K r^{k+\alpha}.
\]
Under small-BMO assumptions on the leading coefficients and appropriate pointwise regularity of lower-order terms and data, weak solutions satisfy pointwise \(C^\alpha\) and \(C^{k,\alpha}\) estimates; if \(u\) vanishes to order \(k\), the required smoothness of the coefficients can be weakened by \(k\) or \(k+1\), depending on the term. The same framework yields a characterization of nodal sets:
\[
\mathcal L_k(u)=\bigcup_j \mathcal L_k^j,
\]
with each stratum contained in a finite union of \(j\)-dimensional \(C^{l,\alpha}\) manifolds [2405.07214].

For elliptic and parabolic equations with divergence-free drifts,
\[
-\Delta u+b(x)\cdot \nabla u=0,\qquad
u_t-\Delta u+b(x,t)\cdot \nabla u=0,
\]
interior pointwise \(C^\alpha\) regularity for any \(0<\alpha<1\) is proved under one of three scale-invariant smallness conditions on the drift: a Morrey-type \(C_p^{-1}\) smallness, an \(L^2\)-energy smallness, or a \(BMO^{-1}\) smallness. The proof is based on the energy inequality and the perturbation technique, combined with compactness and Campanato iteration [2402.18161].

A distinct line treats locally uniformly elliptic equations, where uniform ellipticity holds only on bounded regions of \((D^2u,Du,u)\)-space. For equations
\[
F(D^2u,Du,u,x)=f(x),
\]
with \(F\) only \(\rho\)-uniformly elliptic, pointwise \(C^{1,\alpha}\), \(C^{2,\alpha}\), and \(C^{k,\alpha}\) regularity are obtained under smallness assumptions on \(\|u\|_{L^\infty}\), on \(\|f\|_{C^{k-2+\alpha}(0)}\), or on the deviation from a model operator \(F_0\). The applications include the prescribed mean curvature equation, the Monge–Ampère equation, the \(k\)-Hessian equations, the \(k\)-Hessian quotient equations, and the Lagrangian mean curvature equation; the accompanying remarks state that the smallness assumptions are necessary in most cases [2405.07199].

## 3. Boundary, singular geometry, and free boundaries

Boundary pointwise regularity sharpens classical boundary Schauder theory by requiring only pointwise geometric control of the boundary at a single point. For fully nonlinear elliptic equations,
\[
F(D^2u,x)=f(x)\quad \text{in }\Omega,\qquad u=g\quad\text{on }\partial\Omega,
\]
if \(\partial\Omega\in C^{1,\alpha}(x_0)\), \(g\in C^{1,\alpha}(x_0)\), and \(f\) satisfies a Morrey-type bound, then \(u\in C^{1,\alpha}(x_0)\); if \(\partial\Omega\in C^{2,\alpha}(x_0)\), \(g\in C^{2,\alpha}(x_0)\), and \(f\in C^\alpha(x_0)\), then \(u\in C^{2,\alpha}(x_0)\) [1901.06060]. A parabolic analogue for
\[
u_t-F(D^2u,Du,u,x,t)=f \quad \text{in }\Omega\cap Q_1,\qquad u=g\quad\text{on }\partial\Omega\cap Q_1
\]
establishes boundary pointwise \(C^{k,\alpha}\) regularity for any \(k\ge1\), using flat-boundary model problems, compactness, and scaling [2208.01194].

For Dirichlet and oblique derivative problems with pointwise regular data, a higher-order boundary theory proves that if \(u\in C^{k,\alpha}(0)\), \(f\in C^{k+l-2,\alpha}(0)\), \(g\in C^{k+l,\alpha}(0)\), and \(\partial\Omega\in C^{l,\alpha}(0)\), then \(u\in C^{k+l,\alpha}(0)\). In the oblique case, if
\[
\Delta u=f \quad \text{in }\Omega\cap B_1,\qquad \beta(x)\cdot Du=g \quad\text{on }\partial\Omega\cap B_1,\qquad \beta_n\ge a_0>0,
\]
then \(f\in C^{-1,\alpha}(0)\), \(g\in C^\alpha(0)\), and \(\beta\in C^\alpha(0)\) imply \(u\in C^{1,\alpha}(0)\). These results are then applied to obstacle-type and one-phase problems to obtain \(C^\infty\) regularity of free boundaries from \(C^{1,\alpha}\) regularity [2204.09304].

The conical setting replaces Taylor polynomials by the homogeneous profile dictated by the geometry. Let \(\omega\subset S^{n-1}\) be a \(C^2\)-smooth open cap and \(C_\omega\) the cone based on \(\omega\). For bounded viscosity solutions of
\[
F(D^2u)=0 \quad \text{in } C_\omega^1,\qquad u=0\quad\text{on }(\partial C_\omega)^1,
\]
with \(F\) uniformly elliptic and positively \(1\)-homogeneous, there exist \(a\in\mathbb R\), \(\alpha\in(0,1)\), and \(C>0\) such that
\[
|u(x)-a\Psi(x)|\le C |x|^\alpha \Psi(x)\,\|u\|_{L^\infty(C_\omega^1)},\qquad x\in C_\omega^{1/2},
\]
where \(\Psi\) is the known homogeneous solution vanishing on \(\partial C_\omega\). This estimate leads directly to Liouville theorems on cones: if \(u(x)=O(\Psi(x))\) at infinity, then \(u\) is a multiple of \(\Psi\); if \(u=o(\Psi)\), then \(u\equiv0\) [2205.14291].

Boundary pointwise regularity has also been extended beyond smooth domains. On \(L\)-uniform domains, a notion of weak solution with nonzero boundary data is introduced through the auxiliary functions
\[
w^+(x)=\bigl(u(x)-\sup_{B_r(x_0)\cap \partial\Omega} g\bigr)^+,\qquad
w^-(x)=\bigl(u(x)-\inf_{B_r(x_0)\cap \partial\Omega} g\bigr)^-,
\]
together with subsolution inequalities. Under an admissibility condition and pointwise regularity assumptions \(f\in C^{-2,\alpha}\), \(g\in C^\alpha\), boundary pointwise \(C^\alpha\) regularity follows, and linearity with respect to harmonic functions yields \(C^{1,\alpha}\) and \(C^{2,\alpha}\) under stronger assumptions on the boundary and data [2509.17690]. For divergence-form equations with distributional coefficients, if a boundary point satisfies a measure condition, the multiplier \(V\) is form-bounded with sufficiently small constant, and the nonhomogeneous terms satisfy Dini decay, then the weak solution is continuous there in the \(L\)-sense, with an \(L^2\)-mean oscillation estimate implying pointwise Hölder regularity when the Dini remainder is integrable [2408.01073].

Free-boundary regularity is another setting where pointwise expansions are decisive. In the parabolic obstacle problem
\[
\Delta u-u_t=f\chi_{\{u>0\}},\qquad u\ge0,
\]
Lindgren and Monneau derive a Weiss-type monotonicity formula, a singular-point monotonicity formula, and a second-order Taylor expansion at singular free boundary points under pointwise Dini and double-Dini conditions. Under Dini continuity of \(f\), the regular set is locally a parabolic \(C^1\)-surface, while the singular set is locally contained in a union of parabolic \(C^1\) manifolds [1305.7349].

## 4. Wavelet, Besov, and multifractal formulations

In self-affine and multiresolution settings, pointwise regularity is encoded by decay of local oscillation or wavelet coefficients. For self-affine lattice tilings associated with an integer dilation matrix \(M\), local oscillation is measured by
\[
\operatorname{osc}_p^k(f;x,\ell)=|Q_\ell(x)|^{-1/p}\,\|f-P_{Q_\ell(x)}f\|_{L^p(Q_\ell(x))},
\]
and the corresponding Besov norm is
\[
\|f\|_{B_{p,q}^s(M)}
= \|f\|_{L^p} + \Bigl(\sum_{\ell=0}^\infty [|M|^{\ell s}\|\operatorname{osc}_p^k(f;\cdot,\ell)\|_{L^p}]^q\Bigr)^{1/q}.
\]
Under multiresolution analysis, one also has a wavelet characterization, and pointwise regularity at \(x_0\) is equivalent to decay of nearby wavelet coefficients:
\[
|b_{\ell,\varepsilon,j}(v)|\le C |M|^{-\ell(s+n/2)}
\]
whenever \(|M^\ell x_0-v|\lesssim 1\) [1512.00948].

For parameterized affine zipper fractal curves, dominated splitting of index-\(1\) provides a dynamical mechanism for well-defined local anisotropy. The pressure function \(P(t)\), defined as the unique real root of
\[
0=\lim_{n\to\infty}\frac1n\log\sum_{i_1\cdots i_n=0}^{N-1}\|A_{i_1\cdots i_n}\|^t(\lambda_{i_1}\cdots\lambda_{i_n})^{-P(t)},
\]
is continuous, strictly increasing, concave, and \(C^1\). Under SOSC, non-degeneracy, and dominated splitting, the Hausdorff dimensions of level sets of the pointwise Hölder exponent are given, on the stated ranges, by the Legendre-transform formula
\[
\dim_H E(\beta)=\inf_{t\in\mathbb R}\{t\beta-P(t)\}.
\]
Under Assumption A, the same formula extends to the full spectrum for the regular exponent \(\alpha_r\), and Assumption A is equivalent to the existence of \(\alpha_r\) for Lebesgue-a.e. point. The de Rham curve is a concrete application: for \(\omega\ne 1/3\), Assumption A holds, the curve is differentiable Lebesgue-a.e. with derivative zero, and the nondifferentiability set has dimension \(\tau-P(\tau)>0\), where \(P'(\tau)=1\) [1608.04558].

Wavelet methods can also be used algorithmically. The Iterated Amplitude Adjusted Wavelet Transform preserves the data histogram and the wavelet-coefficient magnitudes at each scale and location. Because the estimate
\[
|Wf(a,x_0)|\le K a^{h(x_0)+1/2}
\]
links the pointwise Hölder exponent to the modulus of wavelet coefficients, preserving those magnitudes preserves the pointwise Hölder regularity structure. The method is designed to preserve the multifractal properties of the data while randomizing the spatial arrangement of singularities, and it is used for testing oscillating singularities and velocity–intermittency coupling in turbulence [1701.00579].

## 5. Stochastic processes and refined pointwise oscillation

For stochastic processes, pointwise regularity is often finer than the sole value of the Hölder exponent. In the generalized Rosenblatt process \(R_{H_1,H_2}\), with self-similarity exponent \(\alpha_0=H_1+H_2-1\), wavelet methods identify three types of points on any interval. Ordinary points satisfy a modulus involving \(|t-s|^{\alpha_0}\log\log|t-s|^{-1}\), rapid points a stronger modulus involving \(|t-s|^{\alpha_0}\log|t-s|^{-1}\), and slow points satisfy a modulus with no logarithm. The wavelet-leader characterization is
\[
h_X(t)=\liminf_{j\to\infty}\frac{\log\bigl(\sup_{k:\,2^{-j}k\approx t}|d_{j,k}|\bigr)}{\log(2^{-j})},
\]
and the proof uses a Meyer-wavelet expansion, kernel cancellation for \(|j_1-j_2|>1\), and tail bounds in the second Wiener chaos. Compared with fractional Brownian motion, the logarithmic corrections differ by square-roots, reflecting the replacement of Gaussian first-chaos tails by exponential second-chaos tails [2203.08487].

For the multifractional Brownian motion \(B_H\), the pointwise exponent is governed exactly by the Hurst function. Under Condition 1.3, namely that for each \(t\) there exist \(\gamma\ge H(t)\) and constants \(c_t,R_t>0\) such that
\[
|H(s)-H(t)|\le c_t |s-t|^\gamma \qquad (|s-t|\le R_t),
\]
one has almost surely
\[
\alpha_{B_H}(t)=H(t)\qquad \forall t\in\mathbb R.
\]
Moreover, every non-empty open interval contains at least one slow point \(t\), meaning that
\[
|B_H(s)-B_H(t)|=O(|s-t|^{H(t)}).
\]
The same paper shows that this property persists for a non self-similar process and that the assumption on \(H\) can be weakened to the single-point logarithmic modulus
\[
|H(s)-H(t)|\le \frac{c_t}{\ln(|s-t|^{-1})}.
\]
Under this weaker condition, multifractional wavelet series and locally-deforming wavelet expansions still satisfy \(\alpha(t)=H(t)\) almost surely and exhibit the same slow/ordinary/rapid trichotomy [2302.06422].

A common misconception is that the pointwise exponent alone fully describes local behavior. The Rosenblatt and multifractional Brownian results show that this is false: points with the same exponent may still differ by logarithmic corrections in their modulus of continuity. The distinction between ordinary, rapid, and slow points is precisely a distinction at fixed exponent [2203.08487][2302.06422].

## 6. Methods, structural consequences, and neighboring notions

Across these works, the main proofs are scale-by-scale. In PDE, the dominant pattern is compactness plus perturbation: one subtracts an approximating polynomial or profile, rescales, proves precompactness, passes to a model limit equation, and iterates. This appears in fully nonlinear elliptic and parabolic settings, in divergence form, and in boundary problems, often with Campanato decay as the output norm [2405.07214][2208.01194][2012.00324]. On cones, the same philosophy is combined with a Hopf lemma, a boundary Lipschitz estimate, an improvement-of-oscillation lemma in rings, and dyadic iteration [2205.14291]. For fully fractional parabolic equations \((\partial_t-\Delta)^s u=f\), the proof instead uses an integral representation, a decomposition \(u=v_r+w_r\), a directional-average perturbation estimate for the fractional heat kernel, and new equivalent definitions of pointwise function spaces, yielding pointwise \(C^{k+\alpha+2s}\) or logarithmic regularity according to whether \(\alpha+2s\notin\mathbb Z\) or \(\alpha+2s\in\mathbb Z\) [2603.07511].

Pointwise estimates can also enforce geometric structure. In generated Jacobian equations, Guillen and Kitagawa prove an Aleksandrov estimate and a Sharp-Growth estimate under the analogue of the Ma–Trudinger–Wang condition, denoted \((G3w)\). These estimates imply strict \(G\)-convexity and, when \(G\) is \(C^{1,\alpha}\) in the \(x\)-variable, interior
\[
u\in C^{1,\beta}_{\mathrm{loc}}(\Omega),\qquad \beta\in(0,\alpha).
\]
The theory applies to the near-field reflector problem, where the reciprocal radial function \(u=1/\rho\) is represented through a generating function built from ellipsoids, and the resulting reflector surface is \(C^{1,\alpha}\) [1501.07332].

Several papers explicitly show that pointwise regularity is stronger than coarse Hölder control. In the conical fully nonlinear theory, the estimate \(u\in C^{\alpha_\omega,\alpha}(0)\) is described as strictly stronger than global \(C^\alpha\) or even local \(C^{1,\alpha}\) bounds because it identifies the exact homogeneous blow-up limit at the vertex [2205.14291]. Likewise, smallness assumptions in locally uniformly elliptic problems are not treated as purely technical: the stated remarks assert that they are necessary in most cases, since otherwise the blow-up operators may leave the uniformly elliptic regime [2405.07199].

A neighboring but distinct theme is pointwise convergence rather than pointwise regularity. For dispersive equations on compact symmetric spaces \(U/K\) of rank \(1\) or \(2\), the Sobolev threshold \(\alpha>1/2\) is sufficient for almost-everywhere convergence of
\[
u(x,t)=e^{-it\psi(\sqrt{-\Delta})}f(x)
\]
to \(f(x)\) as \(t\to0^+\). In the \(K\)-biinvariant rank-\(1\) case, the sufficiency improves to \(\alpha>1/3\) for the Schrödinger, Boussinesq, and Beam equations, while no maximal-function estimate can hold for \(\alpha<1/4\) [2512.09689]. This does not define pointwise regularity in the usual Hölder or polynomial-approximation sense, but it shows how local convergence questions interact with Sobolev smoothness thresholds.

Taken together, these results suggest that pointwise regularity is less a single theorem than a family of local asymptotic principles. Depending on the problem, the asymptotic model may be a Taylor polynomial, a homogeneous boundary profile, a wavelet-decay law, a Besov oscillation rate, or a modulus corrected by logarithms. What remains constant is the objective: to identify the exact local scale at which a function begins to resemble its canonical model.

Source: https://www.emergentmind.com/topics/pointwise-regularity