De Giorgi-Nash-Moser Theory: Local Boundedness and Holder Continuity of Weak Solutions

Intuition Beginner

Imagine heat flowing through a block built from a chaotic patchwork of different materials. The conductivity jumps from point to point with no pattern: copper here, ceramic a hair's breadth away, then back to copper, an infinite mosaic with no smoothness anywhere. You let the block settle into a steady temperature, with no heat being added or removed inside. The question is simple to ask and hard to answer: does the steady temperature inside the block have to be smooth, or could it inherit the wild jumpiness of the material and itself jump around from point to point?

The surprising answer is that the temperature comes out smooth no matter how jumpy the material is. The medium can be as rough as you like, with conductivity that flips between two fixed positive bounds in any pattern at all, and the equilibrium temperature is still continuous, and better than continuous: it cannot change too quickly between nearby points. The roughness of the material does not transfer to the solution. This is the heart of the De Giorgi-Nash-Moser theorem.

Why is this believable? An equilibrium is a balance of inflow and outflow at every point. Even when the conductivity is rough, that balance still forces a kind of averaging: the value at a point is pinned by a weighted average of the values around it, with weights that stay between fixed positive bounds because the conductivity does. Averaging is a smoothing operation. It cannot manufacture a sharp jump out of nowhere, so the solution stays tame even when the equation does not.

The everyday picture: stretch a thin elastic sheet over a frame, but make the sheet out of patches with wildly varying stiffness. When you let go, the sheet finds its resting shape. That shape is gentle and continuous even though the stiffness is a mess, because tension always pulls toward an average and an average cannot have a cliff in it.

The one-sentence takeaway: an elliptic equilibrium with rough, merely-bounded conductivity still has a solution that is bounded and continuous, with a quantitative limit on how fast it can vary, and this limit depends only on the two bounds on the conductivity, not on its pattern.

Visual Beginner

The picture to hold is a tower of shrinking boxes, each box smaller than the last, with the spread of the solution inside each box shrinking by a fixed fraction as you zoom in.

The tower of boxes is the whole mechanism. You measure the spread of the solution, the gap between its biggest and smallest value, on a box, then shrink the box and measure again. The theorem guarantees the spread shrinks by at least a fixed fraction each time you halve the box. A gap that shrinks by a fixed fraction at every halving closes to zero at a steady geometric rate, and that rate is precisely a quantitative continuity statement: it says how fast the solution settles as you approach the centre. The side panel shows the medium itself as a chaotic checkerboard, as rough as you please; the smoothing happens anyway.

Worked example Beginner

We watch the oscillation-shrinking mechanism produce a continuity rate from a single number. Suppose we have proved, for a particular rough-conductivity equilibrium, that whenever we halve the side of a box centred at a point, the spread of the solution on the smaller box is at most three-quarters of the spread on the larger box. We turn this one fact into a continuity statement.

Step 1. Name the starting spread. On a box of side $1$, say the spread between the largest and smallest value of the solution is $8$ degrees. The spread is just the largest value minus the smallest value on that box.

Step 2. Halve once. On the box of side one-half, the spread is at most three-quarters of $8$, which is $6$ degrees. We multiplied by three-quarters.

Step 3. Halve again. On the box of side one-quarter, the spread is at most three-quarters of $6$, which is $4.5$ degrees. Each halving of the box multiplies the spread by three-quarters.

Step 4. Read the pattern. After halving the box $k$ times, the side is divided by $2$ a total of $k$ times and the spread is multiplied by three-quarters $k$ times. So the spread on the small box is at most $8$ times three-quarters raised to the power $k$. As $k$ grows the spread shrinks toward zero, so the solution settles to a single value at the centre: it is continuous there.

Step 5. Extract the rate. The side shrinks like one-half to the power $k$ and the spread shrinks like three-quarters to the power $k$. So the spread is a fixed power of the side: the spread behaves like the side raised to some positive power less than $1$, the power being whatever number makes one-half raised to it equal three-quarters. That positive power is the continuity exponent, the number that says how gently the solution varies near the point.

What this tells us: a single multiplicative shrink factor for the spread, applied at every halving, is exactly equivalent to a quantitative continuity rate. The rate is set entirely by the shrink factor, and the shrink factor in turn is set entirely by the two bounds on the conductivity. Nothing about the chaotic pattern of the material enters.

Check your understanding Beginner

Formal definition Intermediate+

Throughout, $\Omega \subseteq {\mathbb{R}}^n$ is open with $n\ge 2$, and $A:\Omega \to {\mathbb{R}}^{n\times n}$ is a matrix field whose entries $a^{ij}$ are bounded and measurable — no continuity, no smoothness. The operator is in divergence form, $$Lu=-\mathrm{div}(A(x)Du)=-\sum _{i,j=1}^n{D}_i(a^{ij}(x)D_{j}u),$$ and is uniformly elliptic when there are constants $0<\lambda \le \Lambda <\infty$ with $$\lambda |\xi {|}^2\le \sum _{i,j=1}^n{a}^{ij}(x){\xi }_i{\xi }_j\le \Lambda |\xi {|}^2\text{for almost every }x\in \Omega \text{ and all }\xi \in {\mathbb{R}}^n$$ ^{[Gilbarg-Trudinger 2001 §8.5]}. The ratio $\Lambda /\lambda$ is the ellipticity ratio; it is the only feature of $A$ that the regularity estimates will see. Because $a^{ij}$ is merely measurable, $Lu$ cannot be evaluated pointwise; the equation is interpreted weakly.

Definition (weak solution). A function $u\in W^{1,2}(\Omega )$ 24.01.01 is a weak solution of $Lu=0$ when $${\int }_{\Omega }\sum _{i,j}{a}^{ij}(x)D_{j}u{D}_i\phi dx=0\text{for all }\phi \in C_c^{\infty }(\Omega ).$$ It is a weak subsolution ($Lu\le 0$) when the left side is $\le 0$ for all $\phi \ge 0$, and a weak supersolution ($Lu\ge 0$) when it is $\ge 0$ for all $\phi \ge 0$. The integral makes sense for $u\in W^{1,2}$ because $A$ is bounded and $Du,D\phi \in L^2$; by density the test class may be enlarged to $\phi \in W_0^{1,2}(\Omega )$.

Definition (Hölder continuity). For $0<\alpha \le 1$, $u\in C_{\mathrm{loc}}^{0,\alpha }(\Omega )$ when, on every ${\Omega }^{\prime }⋐\Omega$, the seminorm $$[u{]}_{C^{0,\alpha }({\Omega }^{\prime })}=\underset{\begin{array}{c}x,y\in {\Omega }^{\prime }\\ x\ne y\end{array}}{\sup }\frac{|u(x)-u(y)|}{|x-y{|}^{\alpha }}$$ is finite 02.16.01. Equivalently, the oscillation ${\mathrm{osc}}_{B_r}u={\sup }_{B_r}u-{\inf }_{B_r}u$ over a ball of radius $r$ satisfies ${\mathrm{osc}}_{B_r(x_0)}u\le C{r}^{\alpha }$ as $r\to 0$, uniformly for $x_0\in {\Omega }^{\prime }$.

The contrast with the classical Schauder theory is structural. Schauder estimates require Hölder-continuous coefficients and then produce $C^{2,\alpha }$ solutions; they are perturbative, treating $L$ as a frozen-coefficient Laplacian plus an error. With merely measurable $A$ that route is unavailable — there is no frozen-coefficient operator to perturb from — and the conclusion is correspondingly weaker: not $C^{2,\alpha }$ but $C^{0,\alpha }$, the maximal regularity the hypotheses permit. A bounded-measurable operator can have solutions that are exactly Hölder and no better, so $C^{0,\alpha }$ is sharp.

Counterexamples to common slips Intermediate+

• Measurable coefficients do not give Lipschitz, let alone $C^1$. The radial field $A(x)=I+(\sigma -1)\frac{x\otimes x}{|x{|}^2}$ on ${\mathbb{R}}^n$ (bounded, elliptic, discontinuous at the origin) admits the weak solution $u(x)=|x{|}^{\delta }\frac{x_1}{|x|}$ for a suitable $\delta =\delta (\sigma ,n)\in (0,1)$, which is exactly $C^{0,\delta }$ and not $C^{0,\beta }$ for any $\beta >\delta$. The De Giorgi exponent cannot be improved without improving the hypotheses on $A$.

• The divergence form is essential. The estimate is for $-\mathrm{div}(ADu)$, tested against the weak formulation. The non-divergence operator $\sum a^{ij}{D}_{ij}u$ with merely measurable $a^{ij}$ is governed instead by the Krylov-Safonov theory 02.17.02; the two are genuinely different normal forms and the divergence-form integration-by-parts that powers every estimate here is unavailable there.

• No continuity of $A$ is used or available. Any proof step that evaluates $a^{ij}(x_0)$ at a point, or freezes coefficients, or differentiates the equation, is illegitimate. The only admissible operations are testing the weak formulation against Sobolev functions and invoking the Sobolev and Poincaré inequalities 02.16.01.

• Local boundedness precedes continuity and is not automatic. A $W^{1,2}$ weak solution is a priori only in $L^{2^{\ast }}$ by Sobolev embedding; that it is locally bounded is the first theorem, proved by iteration, and Hölder continuity is built on top of it.

Key theorem with proof Intermediate+

Theorem (Caccioppoli inequality / reverse Poincaré). Let $u\in W^{1,2}(\Omega )$ be a weak subsolution of $Lu=0$ with $A$ uniformly elliptic, ellipticity constants $\lambda ,\Lambda$. For every $k\in \mathbb{R}$, every ball $B_R=B_R(x_0)⋐\Omega$, and every $0<r<R$, $${\int }_{B_r}|D(u-k{)}^{+}{|}^{2}dx\le \frac{C}{(R-r{)}^2}{\int }_{B_R}((u-k{)}^{+}{)}^{2}dx,C=C\left(\frac{\Lambda }{\lambda }\right)$$ ^{[Gilbarg-Trudinger 2001 §8.5, Theorem 8.17]}. Here $(u-k{)}^{+}=\max (u-k,0)$.

Proof. Fix a cutoff $\eta \in C_c^{\infty }(B_R)$ with $0\le \eta \le 1$, $\eta \equiv 1$ on $B_r$, and $|D\eta |\le 2/(R-r)$. The function $v=(u-k{)}^{+}\in W^{1,2}(\Omega )$ has $Dv=Du\cdot 1_{\{u>k\}}$ almost everywhere, and the test function $\phi ={\eta }^{2}v$ lies in $W_0^{1,2}(B_R)$ with $\phi \ge 0$. The subsolution inequality gives $\int ADu\cdot D\phi \le 0$. Expand $D\phi ={\eta }^{2}Dv+2\eta vD\eta$ and use $ADu\cdot Dv=ADv\cdot Dv$ on $\{u>k\}$ (since $Du=Dv$ there) to obtain $$\int {\eta }^{2}ADv\cdot Dvdx\le -2\int \eta vADv\cdot D\eta dx.$$ Ellipticity bounds the left side below by $\lambda \int {\eta }^2|Dv{|}^2$. The right side is bounded above, using $|ADv\cdot D\eta |\le \Lambda |Dv||D\eta |$ and Cauchy-Schwarz with a parameter $\epsilon >0$, by $$2\Lambda \int \eta v|Dv||D\eta |\le \epsilon \lambda \int {\eta }^2|Dv{|}^2+\frac{{\Lambda }^2}{\epsilon \lambda }\int v^2|D\eta {|}^2.$$ Choosing $\epsilon =1/2$ and absorbing the first right-hand term into the left gives $\frac{\lambda }{2}\int {\eta }^2|Dv{|}^2\le \frac{2{\Lambda }^2}{\lambda }\int v^2|D\eta {|}^2$. Since $\eta \equiv 1$ on $B_r$ and $|D\eta |\le 2/(R-r)$, $${\int }_{B_r}|Dv{|}^2\le \frac{16{\Lambda }^2}{{\lambda }^2(R-r{)}^2}{\int }_{B_R}{v}^2,$$ which is the claim with $C=16(\Lambda /\lambda {)}^2$. $\square$

Bridge. The Caccioppoli inequality is the foundational reason measurable coefficients still yield regularity: it reverses the usual Poincaré inequality, controlling the gradient of a truncation $(u-k{)}^{+}$ by the function itself on a larger ball, and it is exactly the energy input that both the De Giorgi level-set iteration and the Moser $L^p$ iteration consume. This builds toward local boundedness, where the inequality is applied across a sequence of truncation levels $k_j$ and shrinking radii $R_j$ to drive a nonlinear recursion, and it appears again in the weak Harnack inequality below, where the supersolution form of the same estimate controls a negative power of $u$. The central insight is that uniform ellipticity converts the weak formulation into a self-improving energy bound that sees only the ratio $\Lambda /\lambda$; this is dual to the maximum-principle estimate of 02.17.02, where the pointwise second-derivative test plays the role that the integrated energy estimate plays here. Putting these together, the Caccioppoli inequality is the bridge from the abstract weak formulation to every quantitative regularity statement in the chapter, and the Sobolev inequality 02.16.01 is what turns its $L^2$ gradient control into the gain of integrability that makes the iteration close.

Exercises Intermediate+

Advanced results Master

The regularity theory for divergence-form operators with bounded measurable coefficients is built from one energy estimate (Caccioppoli) processed by two interchangeable amplifiers (the De Giorgi level-set iteration and the Moser $L^p$ iteration), feeding a single nonlinear pump (local boundedness), and culminating in the Harnack inequality from which Hölder continuity drops out by oscillation decay. The estimates depend on $A$ only through $n$ and $\Lambda /\lambda$.

Theorem 1 (local boundedness; De Giorgi 1957, Stampacchia 1965). Let $u\in W^{1,2}(\Omega )$ be a weak subsolution of $Lu=0$. For every $B_{2R}(x_0)⋐\Omega$ and every $p>0$, $$\underset{B_R(x_0)}{\sup }{u}^{+}\le C{\left({\text{fint}}_{B_{2R}(x_0)}(u^{+}{)}^{p}dx\right)}^{1/p},C=C(n,\Lambda /\lambda ,p)$$ ^{[De Giorgi 1957] [Stampacchia 1965]}. The De Giorgi proof iterates the Caccioppoli inequality over the truncation levels $k_j=(1-2^{-j})M$ and radii $R_j=R(1+2^{-j})$, converting the truncated energies $Y_j={\int }_{B_{R_j}}((u-k_j{)}^{+}{)}^2$ into a recursion $Y_{j+1}\le C{B}^j{Y}_j^{1+2/n}$ via Sobolev and Hölder; the fast-geometric-convergence lemma (Exercise 4) forces $Y_j\to 0$ once $Y_0$ is small, which says $u\le M$ on $B_R$. Moser's route (Exercise 7) instead raises the $L^p$ norm of $u^{+}$ through the fixed Sobolev factor $n/(n-2)$ at each step and passes $p\to \infty$. A subsolution in $W^{1,2}$ is therefore locally bounded, with the bound controlled by any single fixed $L^p$ norm.

Theorem 2 (weak Harnack inequality; Moser 1961). Let $u\in W^{1,2}(\Omega )$, $u\ge 0$, be a weak supersolution. For $B_{4R}(x_0)⋐\Omega$ there exist $p_0=p_0(n,\Lambda /\lambda )>0$ and $C=C(n,\Lambda /\lambda )$ with $${\left({\text{fint}}_{B_{2R}}{u}^{p_0}\right)}^{1/p_0}\le C\underset{B_R}{\inf }u$$ ^{[Moser 1961] [Gilbarg-Trudinger 2001 §8.18]}. The proof controls positive powers $u^p$ for $0<p<p_0$ by Moser iteration on the supersolution, controls negative powers $u^{-p}$ likewise, and bridges the two ranges across $p=0$ by the John-Nirenberg lemma applied to $\log u$, whose oscillation is bounded by the Caccioppoli estimate for the supersolution tested against ${\eta }^2/u$. The weak Harnack inequality is the half of the Harnack inequality that does the regularity work.

Theorem 3 (Harnack inequality; Moser 1961). Let $u\in W^{1,2}(\Omega )$, $u\ge 0$, be a weak solution of $Lu=0$. For $B_{4R}(x_0)⋐\Omega$, $$\underset{B_R(x_0)}{\sup }u\le C\underset{B_R(x_0)}{\inf }u,C=C(n,\Lambda /\lambda )$$ ^{[Moser 1961]}. This is the conjunction of local boundedness (Theorem 1, the sup half) and the weak Harnack inequality (Theorem 2, the inf half), joined at the shared exponent $p_0$. The constant $C$ is scale-invariant and independent of $u$, $x_0$, and the pattern of $A$; it is the exact divergence-form analogue of the non-divergence Krylov-Safonov inequality of 02.17.02.

Theorem 4 (interior Hölder continuity; De Giorgi 1957, Nash 1958). Let $u\in W^{1,2}(\Omega )$ be a weak solution. Then $u\in C_{\mathrm{loc}}^{0,\alpha }(\Omega )$ for some $\alpha =\alpha (n,\Lambda /\lambda )\in (0,1)$, and for every ${\Omega }^{\prime }⋐\Omega$, $$[u{]}_{C^{0,\alpha }({\Omega }^{\prime })}\le C(n,\Lambda /\lambda ,{\Omega }^{\prime },\Omega )\Vert u{\Vert }_{L^2(\Omega )}$$ ^{[De Giorgi 1957] [Nash 1958] [Gilbarg-Trudinger 2001 §8.22]}. The Harnack inequality forces oscillation decay ${\mathrm{osc}}_{B_{R/2}}u\le \theta {\mathrm{osc}}_{B_R}u$ with $\theta =(C-1)/(C+1)<1$ (Exercise 6), and iterating the decay (Exercise 8) gives the Hölder exponent $\alpha ={\log }_2(1/\theta )$. The exponent is generally far below $1$ and cannot be improved without strengthening the hypotheses on $A$.

Theorem 5 (the inhomogeneous and lower-order case). For the operator $Lu=-\mathrm{div}(ADu+bu)+c\cdot Du+du$ with $A$ uniformly elliptic and the lower-order coefficients $b,c,d$ in suitable Lebesgue classes ($b,c\in L^q$, $d\in L^{q/2}$ for $q>n$), a weak solution of $Lu=f$ with $f\in L^{q/2}$ is still locally bounded and locally Hölder continuous, with exponent and constant depending additionally on $n,q$ and the norms of the lower-order data ^{[Gilbarg-Trudinger 2001 §8.5-§8.9] [Stampacchia 1965]}. The principal-part estimates are unchanged; the lower-order terms are absorbed as controlled perturbations because their Lebesgue exponents sit strictly above the scaling-critical thresholds that the Sobolev inequality dictates.

Theorem 6 (resolution of Hilbert's nineteenth problem). Let $F:{\mathbb{R}}^n\to \mathbb{R}$ be smooth and uniformly convex with bounded second derivatives, and let $u$ minimise the variational integral $\mathcal{F}(v)={\int }_{\Omega }F(Dv)dx$ over its own boundary data. Then $u$ is smooth (indeed real-analytic when $F$ is) in the interior ^{[De Giorgi 1957] [Hilbert 1900] [Morrey 1966]}. The Euler-Lagrange equation $\mathrm{div}(DF(Du))=0$ is nonlinear, but each first derivative $w=D_{k}u$ is a weak solution of the linearised divergence-form equation $-\mathrm{div}(A(x)Dw)=0$ with $A(x)=D^{2}F(Du(x))$, which is uniformly elliptic by convexity and merely measurable because $Du$ is a priori only bounded. Theorem 4 then gives $w\in C^{0,\alpha }$, so $u\in C^{1,\alpha }$; Schauder theory bootstraps from there to $C^{\infty }$. The De Giorgi-Nash theorem is exactly the missing link that the nineteenth-century calculus of variations could not supply.

Synthesis. The De Giorgi-Nash-Moser theorem is the foundational reason the calculus of variations produces smooth minimisers: testing the weak formulation against a cutoff times a truncation of the solution yields a self-improving energy estimate that sees the coefficients only through $\Lambda /\lambda$. This is exactly the divergence-form counterpart of the maximum-principle estimate of 02.17.02, where the pointwise negative-semidefinite-Hessian test is replaced by the integrated Caccioppoli inequality. Putting these together, the De Giorgi level-set iteration and the Moser $L^p$ iteration are two readings of the same Sobolev-driven gain of integrability and generalise one another through the shared fast-geometric-convergence mechanism. The central insight is that the Harnack inequality, the quantitative form of the elliptic averaging that pinned harmonic functions to their boundary traces in 02.13.01, survives the loss of all coefficient regularity, and this is dual to the Sobolev embedding of 02.16.01: the embedding supplies the integrability gain at each step, and the iteration converts it into pointwise control. The bridge is oscillation decay, the statement that the Harnack constant $C$ forces a solution's spread to contract by the fixed factor $(C-1)/(C+1)$ at each dyadic scale; this same contraction reappears in the parabolic De Giorgi-Nash theory of 02.13.03 on space-time cylinders, in the Krylov-Safonov non-divergence theory of 02.17.02 via the ABP estimate, and in the regularity theory for minimal surfaces and elliptic systems, where De Giorgi's level-set technique remains the load-bearing tool.

Full proof set Master

Proposition 1 (logarithmic estimate for supersolutions). Let $u>0$ be a weak supersolution of $Lu=0$ on $\Omega$, $A$ uniformly elliptic. For every $B_{2R}(x_0)⋐\Omega$, $${\int }_{B_R}|D\log u{|}^{2}dx\le C(n,\Lambda /\lambda )R^{n-2}.$$

Proof. Test the supersolution inequality $\int ADu\cdot D\phi \ge 0$ against $\phi ={\eta }^2{u}^{-1}$, where $\eta \in C_c^{\infty }(B_{2R})$ is the standard cutoff with $\eta \equiv 1$ on $B_R$ and $|D\eta |\le 2/R$. Since $\phi \ge 0$ and $D\phi ={\eta }^2(-u^{-2})Du+2\eta u^{-1}D\eta$, $$0\le \int ADu\cdot D\phi =-\int {\eta }^2{u}^{-2}ADu\cdot Du+2\int \eta u^{-1}ADu\cdot D\eta .$$ Writing $D\log u=u^{-1}Du$, the first term is $-\int {\eta }^{2}AD\log u\cdot D\log u\le -\lambda \int {\eta }^2|D\log u{|}^2$. The second is bounded by $2\Lambda \int \eta |D\log u||D\eta |\le \frac{\lambda }{2}\int {\eta }^2|D\log u{|}^2+\frac{2{\Lambda }^2}{\lambda }\int |D\eta {|}^2$ by Cauchy-Schwarz. Rearranging, $\frac{\lambda }{2}\int {\eta }^2|D\log u{|}^2\le \frac{2{\Lambda }^2}{\lambda }\int |D\eta {|}^2\le \frac{2{\Lambda }^2}{\lambda }\cdot \frac{4}{R^2}|B_{2R}|$. Since $|B_{2R}|=c_n{R}^n$, the right side is $C(n,\Lambda /\lambda )R^{n-2}$, and $\eta \equiv 1$ on $B_R$ gives ${\int }_{B_R}|D\log u{|}^2\le C{R}^{n-2}$. $\square$

Proposition 2 (oscillation decay from Harnack). Let the Harnack inequality ${\sup }_{B_R}v\le C_H{\inf }_{B_R}v$ hold for nonnegative solutions on $B_{4R}(x_0)⋐\Omega$. Then for any solution $u$ on $B_{4R}$, $$\underset{B_R}{\mathrm{osc}}u\le \theta \underset{B_{4R}}{\mathrm{osc}}u,\theta =\frac{C_H-1}{C_H+1}<1.$$

Proof. Let $M_4={\sup }_{B_{4R}}u$, $m_4={\inf }_{B_{4R}}u$. The functions $M_4-u\ge 0$ and $u-m_4\ge 0$ are nonnegative solutions on $B_{4R}$ (the constants are annihilated by $\mathrm{div}(AD\cdot )$). Apply Harnack to each on $B_R$: $$M_4-\underset{B_R}{\inf }u=\underset{B_R}{\sup }(M_4-u)\le C_H\underset{B_R}{\inf }(M_4-u)=C_H(M_4-\underset{B_R}{\sup }u),$$ $$\underset{B_R}{\sup }u-m_4=\underset{B_R}{\sup }(u-m_4)\le C_H\underset{B_R}{\inf }(u-m_4)=C_H(\underset{B_R}{\inf }u-m_4).$$ Add the two inequalities. The left side is $(M_4-m_4)+({\sup }_{B_R}u-{\inf }_{B_R}u)={\mathrm{osc}}_{B_{4R}}u+{\mathrm{osc}}_{B_R}u$. The right side is $C_H[(M_4-m_4)-({\sup }_{B_R}u-{\inf }_{B_R}u)]=C_H({\mathrm{osc}}_{B_{4R}}u-{\mathrm{osc}}_{B_R}u)$. Hence ${\mathrm{osc}}_{B_{4R}}u+{\mathrm{osc}}_{B_R}u\le C_H{\mathrm{osc}}_{B_{4R}}u-C_H{\mathrm{osc}}_{B_R}u$, which rearranges to $(C_H+1){\mathrm{osc}}_{B_R}u\le (C_H-1){\mathrm{osc}}_{B_{4R}}u$, the claim. $\square$

Proposition 3 (Hölder continuity from oscillation decay). Suppose ${\mathrm{osc}}_{B_{R/K}(x_0)}u\le \theta {\mathrm{osc}}_{B_R(x_0)}u$ for a fixed dilation factor $K>1$, all $B_R(x_0)⋐\Omega$, with $\theta \in (0,1)$. Then $u\in C_{\mathrm{loc}}^{0,\alpha }(\Omega )$ with $\alpha ={\log }_K(1/\theta )$.

Proof. Iterating $j$ times, ${\mathrm{osc}}_{B_{R/K^j}}u\le {\theta }^j{\mathrm{osc}}_{B_R}u$. Given $0<\rho \le R$, pick $j\ge 0$ with $R/K^{j+1}<\rho \le R/K^j$; then $K^{-j}<K\rho /R$ and ${\theta }^j=K^{-\alpha j}=(K^{-j}{)}^{\alpha }<K^{\alpha }(\rho /R{)}^{\alpha }$. By monotonicity ${\mathrm{osc}}_{B_{\rho }}u\le {\mathrm{osc}}_{B_{R/K^j}}u\le {\theta }^j{\mathrm{osc}}_{B_R}u\le K^{\alpha }(\rho /R{)}^{\alpha }{\mathrm{osc}}_{B_R}u$. Fixing ${\Omega }^{\prime }⋐\Omega$, $R=\frac{1}{2}\mathrm{dist}({\Omega }^{\prime },\partial \Omega )$, and using local boundedness ${\mathrm{osc}}_{B_R}u\le 2\Vert u{\Vert }_{L^{\infty }({\Omega }^{\prime \prime })}<\infty$, for $x,y\in {\Omega }^{\prime }$ with $\rho =|x-y|\le R$ one gets $|u(x)-u(y)|\le {\mathrm{osc}}_{B_{\rho }(x)}u\le K^{\alpha }{R}^{-\alpha }{\mathrm{osc}}_{B_R}u|x-y{|}^{\alpha }$. Thus $[u{]}_{C^{0,\alpha }({\Omega }^{\prime })}<\infty$. $\square$

Proposition 4 (sharpness of the divergence-form regularity). There is a uniformly elliptic bounded measurable $A$ on $B_1\subset {\mathbb{R}}^n$ ($n\ge 2$) and a weak solution $u$ of $-\mathrm{div}(ADu)=0$ with $u\in C^{0,\alpha }\setminus C^{0,\beta }$ for every $\beta >\alpha$, where $\alpha <1$. Hence the De Giorgi exponent cannot be taken to be $1$.

Proof. For $n\ge 2$ take $u(x)=|x{|}^{\gamma }\frac{x_1}{|x|}$ with $0<\gamma <1$ to be fixed, a function that is positively homogeneous of degree $\gamma$ and smooth away from the origin. One computes $Du$ and seeks a radial coefficient field of the form $A(x)=a(|x{|}^{-1}x)P_r+b(|x{|}^{-1}x)P_{\tau }$, where $P_r,P_{\tau }$ are the orthogonal projections onto the radial and tangential directions, with bounded measurable $a,b$ trapped in $[\lambda ,\Lambda ]$. The equation $\mathrm{div}(ADu)=0$ becomes, by homogeneity, an algebraic relation tying $\gamma$ to the ratio $b/a$; choosing the ratio sufficiently large forces $\gamma =\gamma (n,\Lambda /\lambda )$ strictly between $0$ and $1$. The resulting $A$ is uniformly elliptic and discontinuous only at the origin, and $u$ is a weak solution that is exactly Hölder of order $\gamma$ at the origin: $|u(x)-u(0)|=|x{|}^{\gamma }$ along $x=(t,0,\dots ,0)$, so $u\notin C^{0,\beta }$ for $\beta >\gamma$. Taking $\alpha =\gamma$ proves the claim; the same construction shows the De Giorgi exponent degrades to zero as $\Lambda /\lambda \to \infty$. $\square$

Connections Master

• The maximum-principle and Harnack theory for general second-order operators of 02.17.02 is the pointwise, non-divergence sibling of this unit: there the negative-semidefinite-Hessian test at an interior extremum drives the weak and strong maximum principles, and the Krylov-Safonov Harnack inequality is proved through the Alexandrov-Bakelman-Pucci estimate; here the integrated Caccioppoli energy estimate drives the De Giorgi-Moser iteration, and the Harnack inequality is proved through Sobolev-fed $L^p$ amplification. The two units together cover the two normal forms of uniformly elliptic operators with rough coefficients.

• The Sobolev and Morrey inequalities of 02.16.01 are the engine of every iteration step: the gain of integrability $L^2\to L^{2^{\ast }}$ at fixed cost in gradient energy is exactly what makes the Moser scheme converge and the De Giorgi super-level-set recursion superlinear, and the Morrey embedding is the alternative route from the $W^{1,p}$ regularity of the gradient (in the nineteenth-problem application) to the Hölder continuity of the solution. This unit consumes the embeddings that 02.16.01 proves.

• The Sobolev space scaffolding — the definition of $W^{1,2}$, weak derivatives, density of smooth functions, and the cutoff and product rules used to build test functions ${\eta }^2(u-k{)}^{+}$ — is supplied by 24.01.01; the weak formulation of the divergence-form equation is meaningful only because that unit establishes the function space in which $\int ADu\cdot D\phi$ is defined and the test class $W_0^{1,2}$ is dense.

• The $L^p$ machinery of 02.07.06 — Hölder's inequality, interpolation between Lebesgue exponents, and completeness — underlies the passage $p\to \infty$ in Moser iteration and the John-Nirenberg logarithmic estimate that bridges positive and negative powers in the weak Harnack inequality; the convergence of the iterated $L^p$ averages to the sup norm is an $L^p$-completeness statement.

• The parabolic De Giorgi-Nash-Moser theory of 02.13.03 is the time-dependent counterpart: the same Caccioppoli-plus-Sobolev mechanism runs on space-time cylinders with the heat operator ${\partial }_t-\mathrm{div}(AD\cdot )$, producing local boundedness and parabolic Hölder continuity with the intrinsic parabolic metric; Nash's 1958 paper treated the parabolic case first and deduced the elliptic case as the stationary limit.

Historical & philosophical context Master

David Hilbert's nineteenth problem, posed in his 1900 Paris address ^{[Hilbert 1900]}, asked whether the minimisers of regular variational integrals are necessarily analytic. By the 1930s the chain of reasoning was understood up to one gap: a minimiser's first derivatives solve a linear divergence-form elliptic equation whose coefficients $D^{2}F(Du)$ are bounded and elliptic but, since $Du$ is a priori only bounded, merely measurable. The Schauder and Hopf theory of the period required Hölder-continuous coefficients and so could not start the bootstrap. The problem reduced precisely to interior continuity of weak solutions of $-\mathrm{div}(ADu)=0$ for bounded measurable $A$.

Ennio De Giorgi resolved it in 1957 in the Memorie of the Turin Academy ^{[De Giorgi 1957]} with the level-set truncation technique: estimate the energy of $(u-k{)}^{+}$ above successive levels $k_j$, feed the estimate through the Sobolev inequality, and show the super-level energy collapses, yielding first local boundedness and then Hölder continuity with an exponent depending only on dimension and ellipticity. John Nash, working independently and from the parabolic side, published continuity of solutions of parabolic and elliptic equations in 1958 in the American Journal of Mathematics ^{[Nash 1958]}, using moment and entropy estimates on the fundamental solution; the elliptic result followed as the steady-state case. The two proofs were entirely different in technique and appeared within a year of each other.

Jürgen Moser unified and simplified the theory in two papers. His 1960 Communications on Pure and Applied Mathematics note ^{[Moser 1960]} gave a new proof of De Giorgi's theorem, and his 1961 paper in the same journal ^{[Moser 1961]} established the elliptic Harnack inequality for these operators by the iteration scheme on $L^p$ norms now bearing his name, deriving Hölder continuity as a consequence of Harnack rather than separately. Guido Stampacchia extended the theory to operators with lower-order terms and $L^p$ data in his 1965 Annales de l'Institut Fourier memoir ^{[Stampacchia 1965]}, and Charles Morrey placed the variational application in its definitive form in his 1966 monograph ^{[Morrey 1966]}. The level-set method De Giorgi invented became the standard regularity tool for minimal surfaces and for elliptic systems, where Mariano Giaquinta's reverse-Hölder and higher-integrability refinements ^{[Giaquinta 1983]} extended the reach of the original argument.

Bibliography Master

@article{DeGiorgi1957,
  author  = {De Giorgi, Ennio},
  title   = {Sulla differenziabilit\`a e l'analiticit\`a delle estremali degli integrali multipli regolari},
  journal = {Memorie della Accademia delle Scienze di Torino, Classe di Scienze Fisiche, Matematiche e Naturali, Serie 3},
  volume  = {3},
  year    = {1957},
  pages   = {25--43}
}

@article{Nash1958,
  author  = {Nash, John},
  title   = {Continuity of solutions of parabolic and elliptic equations},
  journal = {American Journal of Mathematics},
  volume  = {80},
  year    = {1958},
  pages   = {931--954}
}

@article{Moser1960,
  author  = {Moser, J\"urgen},
  title   = {A new proof of De Giorgi's theorem concerning the regularity problem for elliptic differential equations},
  journal = {Communications on Pure and Applied Mathematics},
  volume  = {13},
  year    = {1960},
  pages   = {457--468}
}

@article{Moser1961,
  author  = {Moser, J\"urgen},
  title   = {On Harnack's theorem for elliptic differential equations},
  journal = {Communications on Pure and Applied Mathematics},
  volume  = {14},
  year    = {1961},
  pages   = {577--591}
}

@article{Stampacchia1965,
  author  = {Stampacchia, Guido},
  title   = {Le probl\`eme de Dirichlet pour les \'equations elliptiques du second ordre \`a coefficients discontinus},
  journal = {Annales de l'Institut Fourier},
  volume  = {15},
  year    = {1965},
  pages   = {189--258}
}

@incollection{Hilbert1900,
  author    = {Hilbert, David},
  title     = {Mathematische Probleme},
  booktitle = {Nachrichten der K\"oniglichen Gesellschaft der Wissenschaften zu G\"ottingen, Mathematisch-Physikalische Klasse},
  year      = {1900},
  pages     = {253--297}
}

@book{Morrey1966,
  author    = {Morrey, Charles B.},
  title     = {Multiple Integrals in the Calculus of Variations},
  series    = {Grundlehren der mathematischen Wissenschaften 130},
  publisher = {Springer},
  year      = {1966}
}

@book{Giaquinta1983,
  author    = {Giaquinta, Mariano},
  title     = {Multiple Integrals in the Calculus of Variations and Nonlinear Elliptic Systems},
  series    = {Annals of Mathematics Studies 105},
  publisher = {Princeton University Press},
  year      = {1983}
}

@book{GilbargTrudinger2001,
  author    = {Gilbarg, David and Trudinger, Neil S.},
  title     = {Elliptic Partial Differential Equations of Second Order},
  edition   = {2},
  series    = {Grundlehren der mathematischen Wissenschaften 224},
  publisher = {Springer},
  year      = {2001}
}