Notes 295A PDF
Notes 295A PDF
Connor Mooney
2018 - 2019
Contents
1 Introduction 3
1.1 Variational Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Nonvariational Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Importance of Linear Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2 Harmonic Functions 5
2.1 Structure and Meaning of ∆ . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.3 Mean Value Property and Consequences . . . . . . . . . . . . . . . . . . . . 7
2.4 Maximum Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.4.1 Weak Maximum Principle . . . . . . . . . . . . . . . . . . . . . . . . 8
2.4.2 Strong Maximum Principle and Hopf Lemma . . . . . . . . . . . . . 10
2.4.3 Harnack Inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.5 Boundary Harnack Inequality . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.6 Energy Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.7 Representation Formulae . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.8 The Dirichlet Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3 Schauder Estimates 18
3.1 Pointwise Hölder Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.2 Laplace Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.3 Boundary Regularity for Laplace . . . . . . . . . . . . . . . . . . . . . . . . 21
3.4 Variable Coefficients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.5 Some Functional Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.6 Dirichlet Problem for Variable Coefficient Equations . . . . . . . . . . . . . . 26
4 Function Spaces 28
4.1 Sobolev Spaces and Approximation . . . . . . . . . . . . . . . . . . . . . . . 28
4.2 Sobolev Inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.3 Morrey Inequality and Morrey Spaces . . . . . . . . . . . . . . . . . . . . . . 28
4.4 BMO and John-Nirenberg Inequality (???) . . . . . . . . . . . . . . . . . . . 28
4.5 BV Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.6 Convex Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
1
5 Divergence-Form Equations: Weak Solutions 29
5.1 Maximum Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
5.2 Dirichlet Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
5.3 Regularity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
5.4 De Giorgi-Nash-Moser Harnack Inequality . . . . . . . . . . . . . . . . . . . 29
7 Applications 31
7.1 Minimal Surface Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
7.2 Fully Nonlinear Equations in 2D . . . . . . . . . . . . . . . . . . . . . . . . . 31
7.3 Concave Fully Nonlinear Equations . . . . . . . . . . . . . . . . . . . . . . . 31
2
1 Introduction
In the first part of the course we will discuss elliptic PDE. Broadly speaking, such equations
arise in two contexts: variational problems, and non-variational problems.
over all functions u : B1 ⊂ Rn → R with fixed boundary data ϕ. If u ∈ C 2 (B1 ) solves the
minimal surface equation
div(∇G(∇u)) = 0 in B1 , u|∂B1 = ϕ
then it is the unique minimizer of A(u). Notice that the equation has “divergence form.”
Roughly, we get global information about solutions by making various perturbations and
noticing that the area integral A(u) doesn’t change to leading order.
Notice that when ∇u = 0, the equation reduces to ∆u = div(∇u) = 0. Since the area
functional is invariant under rotations of the graph in Rn+1 , the minimal surface equation is
the Laplace equation over each tangent plane to the graph. Related is the observation that
if u = p · x + ψ, then ∆ψ = 0 to leading order in ; small perturbations from first-order
objects (planes) solve the Laplace equation.
of the graph of u. To solve problems like this, we use L∞ (rather than integral) techniques,
based on the maximum principle: if det D2 w ≥ det D2 u and they have the same boundary
data, then w ≤ u. The idea is then to start with a function that solves det D2 w > 1, and
then to perturb it upwards more and more until we get a solution (Perron’s method).
Notice that if u = 12 |x|2 + ψ, then to leading order in we have ∆ψ = 0; small pertur-
bations from second-order objects (quadratic polynomials) solve the Laplace equation.
3
We saw the Laplace equation
δij uij = 0
arise when solutions are small perturbations from simple objects. If we differentiate the
nonlinear equations in the ek direction we get
In particular, notice that if the gradient is continuous in the variational case, or the Hessian
is continuous in the non-variational case, we are dealing with linear equations in either
divergence or non-divergence form:
with continuous coefficients. Finally, if we have no information on the gradients resp. Hes-
sians, we are dealing with equations where the coefficients have no regularity, and all we
know is that the matrix aij is positive at every point. Understanding such problems is the
key to a satisfactory nonlinear theory.
The goal of this class is to understand each of these cases (aij = δij first, then continuous
coefficients, then rough coefficients) very well, giving a good foundation to deal with nonlinear
problems.
4
2 Harmonic Functions
2.1 Structure and Meaning of ∆
The Laplace equation
∆u = div(∇u) = δij uij = 0
for a C 2 function u on Rn is the fundamental example of an elliptic PDE. We have seen
it arise by taking small perturbations of “simple” solutions to nonlinear problems (planes
in the variational case, and quadratic polynomials in the non-variational case). It is also
natural from a symmetry viewpoint:
Proposition 1. Up to multiplication by constants, ∆ is the only linear, translation-invariant,
rotation-invariant second-order differential operator.
Proof. Linear, translation-invariant second-order operators can be written Lu = tr(AD2 u)
for some A ∈ Symn×n . Note that D2 (u(Bx)) = B T D2 u(Bx)B. Thus, rotation invariance
means that
tr(AM ) = tr(AOT M O) = tr(OAOT M )
for all orthogonal matrices O and M ∈ Symn×n . In particular, we may assume A is diagonal.
The result then follows by permuting coordinates.
Thus, ∆u measures how much, on average, u lies above its tangent linear function nearby a
point.
Proof. Using Taylor expansion, it is not hard to see that the quantity on the right is well-
defined and depends only on D2 u(x). In addition, it is linear, translation- and rotation-
invariant. By the previous proposition, it is a multiple of ∆. Testing it on |x|2 gives the
factor 2n.
We say that u ∈ C 0 (Rn ) satisfies the mean value property MVP if
Z
1
u(x) = u(y) dy
|∂Br | ∂Br (x)
for all r and x. ItR is a good exercise to show that this is equivalent to averaging in solid
1
balls: u(x) = |Br | Br (x) u(y) dy. By Proposition 2 we have:
5
Corollary 1. If u ∈ C 2 (Rn ) satisfies the MVP, then it is harmonic.
Actually, if u ∈ C 0 (Rn ) satisfies the MVP, then for a radially symmetric mollifier ρ = ρ(r)
supported in B1 with mass 1 we have
Z Z 1 Z 1 Z
u(x) = u(x) ρ dy = ρ(r)u(x)|∂Br | dr = ρ(r) u(x + y) dy dr = (u ∗ ρ)(x),
B1 0 0 |y|=r
2.2 Examples
It is important to have some examples to test out our theorems on harmonic functions. A
useful and rich family comes from looking at the real or imaginary parts of holomorphic
functions. (This reveals another invariance of ∆, conformal invariance).
1. Harmonic polynomials: re(z k ) = rk cos(kθ). As k → ∞, these illustrate the relation-
ship between growth rates of harmonic functions and the frequency of oscillation (an
important theme in “unique continuation”). As k crosses 1, we see how positive har-
monic functions in “acute corners” grow slowly, and in “obtuse corners” grow quickly.
2. Separated variables: re(ez ) = ex cos(y). These are similar to the harmonic polynomials,
written in rectangular coordinates. They grow exponentially in strips, and again, their
frequency of oscillation corresponds to growth rate.
3. Singular examples: log z gives rise to log r (defined away from 0) and θ, defined off of
a ray from the origin. The log is the “fundamental solution,” and its Laplace at the
origin is (as we will see) a Dirac mass. The function θ is interesting in that its graph is
also a minimal surface and an infinity-harmonic function (uij ui uj = 0), and it suggests
that in 2D, the Dirichlet problem may be well posed even on smooth domains with
line segments attached.
4. The function z log z gives rise to a harmonic function with Lipschitz data on {y = 0}
but logarithmically blowing up gradient in {y > 0}, showing that certain types of
boundary regularity don’t propagate to the interior.
There are analogues of these examples in higher dimensions. The most important example
is the fundamental solution
Γ = |x|2−n
in dimension n ≥ 3, which is (up to adding and multiplying by constants) the only rotation-
ally symmetric symmetric example. To see this, it is a good exercise to show that in polar
coordinates the Hessian of a radial function f (|x|) has the form
D2 f = diag(f 00 (r), f 0 /r, ..., f 0 /r).
6
Another way to explain the growth rate is to note that that the flux of ∇Γ through any
sphere ∂Br has to be the same, by the equation div(∇Γ) = 0. Thus, Γ0 (r) has to decay like
|∂Br | grows, i.e. Γ0 = cr1−n .
by the divergence theorem. The quantity we differentiated is the average of u on ∂Bλ (up
to multiplying by a constant), completing the proof.
Exercise: One can obtain the solid MVP directly by integrating by parts against a natural
test function. Take φr to be the fundamental solution outside Br , and a quadratic of the form
a + b|x|2 inside Br , with a,R b chosen so that φr has continuous derivative. Let ψ = φr − φ1
and integrate the relation B1 ∆uψ = 0 by parts twice to obtain the result.
It is important to note also that if ∆u ≥ (≤)0, we get the inequalities u ≤ (≥) its average.
Thus, for postive sub-harmonic functions (∆u ≥ 0), the average controls the value, and for
positive super-harmonic functions, the value controls the average.
The MVP says roughly that harmonic functions are their own convolutions. Thus, we
should expect good regularity properties. Very quickly we get:
Proof. Let ρ be smooth radial mollifier supported in B1/2 , and note that u = u ∗ ρ near 0.
Differentiating gives the result.
Using that the derivatives of a harmonic function are harmonic, we also get derivative
estimates of higher order:
|Dk u(0)| < C(n, k)kukL1 (B1 ) .
As a consequence we have the Liouville theorem:
Theorem 3. If u is harmonic in Rn and |u| < C(1 + |x|k ) for some C, k, then u is a
polynomial of degree at most k.
Proof. The functions R−k u(Rx) are harmonic and uniformly bounded in B1 as R → ∞. The
derivative estimates imply that Dk+1 u(0) = 0. Applying to the translations of u we see that
Dk+1 u ≡ 0.
7
Without assuming something about the boundary data, the analogue in a half-space is
false; think e−x cos(y) in the right half-plane, or even r−k cos(kθ) in {x > 1}. Liouville
theorems are useful when we have enough compactness properties to “zoom in on solutions”
and obtain global ones in the limit. If we classify the global ones, we gain information about
the local behavior of the original solution.
Derivative estimates give useful compactness properties:
Theorem 4. Any sequence of harmonic functions that is bounded in L1 (B1 ) contains a
locally uniformly convergent subsequence, and the limit is harmonic.
PDE arguments often rely on compactness arguments to reduce a problem to a simple case;
we will see an example (Harnack inequality) below.
Finally, note that rescaling the interior derivative estimate we get |∇u(0)| < C(n)r−1 supBr |u|.
By iterating this finitely many times we get the more precise information
|u − Pk |(x) ≤ (|Dk+1 u(ξ)|/(k + 1)!)|x|k+1 ( for some ξ ∈ B|x| ) ≤ sup |u|(C0 (n)|x|)k+1 .
B1
wt (x) = t(|x|2 − R2 )
8
with t large and decreasing, for R so large that Ω ⊂ BR . It is easy to see that for some t > 0,
wt touches (u − v) from below in Ω, contradicting the equation at that point.
As a consequence, the maximum (minimum) of a harmonic function on an increasing family of
domains is increasing (decreasing), and occurs at the boundaries. A topological consequence
is that the connected components of {u > l} all “go to the boundary,” for any linear function
l. This is powerful information in two dimensions. To see an example why, assume that
∆u = 0 in B1 and that u has boundary data with the “three point condition:” any plane
through the graph of the boundary data that crosses the data at three points has slope
bounded by K. (C 2 boundary data are good enough for this). Then |∇u| ≤ K in B1 .
Roughly, near any point point with nondegenerate Hessian, u is above and below its tangent
at the point in 4 regions divided by curves crossing orthogonally. These regions have to reach
the boundary (maximum principle), so the tangent plane crosses at four points. Thus, the
|∇u| at this point is bounded by K.
Another consequence is uniqueness for the Dirichlet problem: in a bounded domain, there
is at most one harmonic function that achieves the boundary data continuously. We must
be careful to work in bounded domains. Consider e.g. the functions x2 in {x2 ≥ 0}, or the
fundamental solution outside of a ball, or ex cos(y) in a horizontal strip; they vanish on the
boundary, but not inside the domain.
In unbounded domains, there are two things at play in the uniqueness problem. First,
the growth of the solution at ∞; and second, the geometry of the boundary at ∞. To see
how these affect the problem, consider first any domain in R2 that doesn’t contain a ball.
Any bounded harmonic function that vanishes on the boundary has to vanish (compare with
multiples of log). Here growth conditions were good enough for uniqueness. However, the
analogue is false in dimension n ≥ 3 (exercise). If we ask that u is bounded in for example
{xn ≥ 0} with zero boundary data instead, we do get uniqueness; argue by reflection and
Liouville theorem. Heuristically, the domain is “acting at infinity” enough to affect the
solution.
A closely related topic is removable singularities. If a harmonic function in a domain mi-
nus a point grows more slowly than the fundamental solution, we can remove the singularity.
This is proven by comparison with multiples of the fundamental solution. Connections to the
uniqueness problem in unbounded domains can be made explicit via the Kelvin transform:
if u is harmonic in B1 , then |x|2−n u(x/|x|2 ) is harmonic in Rn \B1 (see exercises).
One can also prove the interior derivative estimates using the maximum principle. Indeed,
it is a good exercise using the Cauchy-Schwarz inequality to prove that
∆(|∇u|2 ϕ2 + M u2 ) ≥ 0
in B1 for a cutoff function ϕ and a universal constant M (n). Conclude the derivative
estimate. Conclude that there are MVP-free proofs of Liouville.
Remark: The “movie” we make depends on the structure of the equation. Note that there
is non-uniqueness for the DP for ∆u + C(x)u = 0 when c is not less than 0 (eigenfunctions
of ∆). This is related to the fact that constants are not supersolutions. However, on small
domains (thin domains) we can make the movie (exercise). Also, we can now prove the
exponential growth in strips. Moving planes: EXERCISE.
9
2.4.2 Strong Maximum Principle and Hopf Lemma
The strong maximum principle is a rigidity result for the equality case in the weak maximum
principle:
Proof. The same barrier as in the proof of the strong maximum principle works.
The interior tangent ball is necessary; consider rα cos(αθ) for α > 1. (Or x1 x2 in the
positive quadrant, i.e. α = 2).
Consequence: Uniqueness for Neumann, mixed Dirichlet-Neumann problems in bounded
domains. The minimum is achieved at a boundary point, and the normal derivative vanishes.
Remark on Alexandrov’s theorem.
Consequence: Subharmonic plus bounded above in 2D implies bounded (comparison with
log); false in 3D (EXERCISE).
sup u ≤ C(n)u(0).
B1/2
Like the maximum principle, this property is shared by equations with rough coefficients.
Remarks: why constant should blow up in B1−δ at least like δ 1−n with fundamental
solution. Why polynomial blowup (iterate it). Gives oscillation decay (nonlocal version of
gradient estimate). Gives Liouville theorem.
Proof. Proof 1: By derivative estimates.
Proof of Oscillation Decay by compactness (only uses MP).
Proof 2 (Cheng - Yau): Gradient estimate for w := |∇ log(u)|2 depending only on C(n)
using the inequality ∆w + 2w1/2 |∇w| ≥ n2 w2 and usual maximum principle, comparison with
A
the family of supersolutions vA = (1−|x| 2 )2 . (EXERCISE).
10
2.5 Boundary Harnack Inequality
The Harnack inequality implied oscillation decay of harmonic functions, which can be viewed
as a quantitative version of the maximum principle. In this section we obtain a quantita-
tive version of the Hopf lemma (due to Krylov) implying “oscillation decay of boundary
gradients.” Below, Br+ = Br ∩ {xn ≥ 0}.
Lemma 1. Assume that ∆u = 0 in B1+ , with axn ≤ u ≤ bxn . There exist δ(n) > 0 and
a ≤ a0 ≤ b0 ≤ b such that
a0 x n ≤ u ≤ b 0 x n
+
in B1/2 , and
b0 − a0 ≤ (1 − δ)(b − a).
This lemma gives “oscillation decay of un (·, 0).”
Proof. By considering (u−axn )/(b−a) we may assume that a = 0 and b = 1. By considering
xn − u we may assume that u(en /2) ≥ 1/4.
The Harnack inequality says that u > c1 (n, µ) in {x : dist.(x, ∂B1+ ) > µ}. Fix µ < 1/10
small. Barriers centered at the points (x0 , 2µ) with |x0 | ≤ 1/2 give u ≥ c2 (n, µ) in {|x0 | <
+
1/2} ∩ {xn ≤ µ}. Since u > c1 in B1/2 otherwise, we have
ak x n ≤ u ≤ b k x n
in B2−k , with (bk −ak ) ≤ (1−δ)k C(n) = C(n)2−kα for α < 1 appropriate. Let l = limk→∞ ak .
It is then clear that
so the desired estimate holds at dyadic radii. It is a simple exercise to show it for all r by
considering the estimate in B2−k , with 2−k−1 ≤ r ≤ 2−k .
11
EXERCISE: Prove using the boundary Harnack inequality that if ∆u = 0 in Rn+ with
u|{xn =0} = 0 and u = O(|x|), then u is a multiple of xn .
EXERCISE: Assume that u satisfies the hypotheses of the boundary Harnack inequality,
and that ∇u is continuous up to {xn = 0}. Prove that un (·, 0) is C α , with an estimate, in
{|x0 | < 1/2}.
Prove using convexity of integrand. As a consequence it is not hard to show the maximum
principle (let f (s) = s for s < sup∂B1 u and 0 < f 0 < 1 for s > sup∂B1 u and consider the
competitor f (u).)
We also immediately get uniqueness from the variational principle: if u = 0 on ∂B1 , then
it is clear that any nonzero competitor has positive energy. This proof avoids the maximum
principle, and thus works e.g. for vectorial variational problems where the maximum principle
is not true.
A quantitative version of the variational principle is the Caccioppoli inequality:
Proof. Multiply the equation by uϕ2 and integrate by parts. EXERCISE: Show this is the
same as using the competitor (1 − ϕ2 )u in the variational principle and taking → 0.
By iterating we get estimates we get the L2 version of the gradient estimate:
for all k ≥ 1.
12
In the case n = 2, the energy estimate is particularly powerful. This is related to the fact
that the Dirichlet energy is invariant under the rescalings u → u(λx) for λ > 0 in 2D; we
will pursue it further when we discuss Sobolev spaces. For now we note that in R2 we “pay
nothing to cut off:” The Dirichlet energy of the function ψR (x) = min{1, 1−log(|x|)/ log(R)}
has energy Z
2π
|∇ψR |2 dx = →0
BR log R
as R → ∞. This gives a variational proof of the Liouville theorem in 2D, by choosing
ϕ = ψR in the energy estimate.
EXERCISE: Prove that sub-harmonic plus bounded above implies constant in 2D. Hints:
you may use that (u − c)+ has the energy estimate. Use the logarithmic cutoff trick.
A related result is that in 2D, the Dirichlet energy controls the oscillation for functions
with the maximum principle:
Proposition 9. If v is a function on B1 ⊂ R2 such that osc∂Br v increases with r, then
Z
2 2π
(osc∂Bδ v) ≤ |∇v|2 dx
| log δ| B1
Integrating this from r = δ to 1 and using that osc∂Br is increasing, we get the result.
This result gives a new proof of oscillation decay and the Harnack inequality for harmonic
functions in 2D based only on energy estimates. Indeed, if u is harmonic on B2 ⊂ R2 ,
then since u satisfies the maximum principle, we may apply the proposition and the energy
estimate to get
8π 2
Z Z
2 2π 2 2π
(oscBδ u) ≤ |∇u| dx ≤ u2 dx ≤ (oscB2 u)2 .
| log δ| B1 | log δ| B2 | log δ|
For δ < δ0 a fixed explicit constant, we see that the oscillation in Bδ is at most half the
oscillation in B2 . We may rescale and iterate to obtain a C α estimate as before.
For the Harnack inequality, one can apply the proposition to v := log u; see the HW.
13
Theorem 6. If ∆u = 0 in B1 ⊂ Rn and u(0) = 0, then the quantity
R
r Br |∇u|2 dx
f (r) := R
∂Br
u2 ds
is increasing with r, and is constant if and only if u is homogeneous.
For homogeneous functions, f (r) is constant, and exactly equals the homogeneity. For
more general functions, f (r) captures the dominant frequency in u at scale r. It is called
“Almgren’s frequency function.” A simpler version of the frequency formula arises in mini-
mal surfaces: if Σ is an area-minimizing hypersurface surface in Rn containing 0, then the
quantity |Σ ∩ Br |/rn−1 is increasing, and constant if and only if Σ is a cone. (There is “more
area than the equatorial slices”). The proof is easy by comparing with the cones over the
boundary traces (short computation). For harmonic functions, the situation is more com-
plicated because there is a two-parameter family of invariances: dilation, and multiplication
by constants, u → µu(λx).
Proof. Remark on proof by expansion into harmonic polynomials (EXERCISE).
We will give a variational proof. It is easy to check that under rescalings u → u(ax),
we have f → f (ar). It thus suffices to check that f 0 (1) ≥ 0. Computing, we see this is
equivalent to showing
Z Z
2
(|∇u| − 2αuuν ) ds − (n − 2) |∇u|2 dx ≥ 0,
∂B1 B1
where α := f (1). Let w be the alpha-homogeneous function that agrees with u on ∂B1 . Note
that the corresponding quantity for w vanishes, and that on ∂B1 we have
|∇u|2 = |∇w|2 + u2ν − α2 u2 , 2αwwν = 2α2 u2 .
By the variational principle, the above inequality is thus equivalent to
Z
(uν − αu)2 ds ≥ 0
∂B1
which is true.
The monotonicity of Rfrequency implies a quantitative version of unique continuation.
Indeed, let h(r) = rn−11
∂Br
u2 ds. The monotonicity says that r(log h)0 is positive and
increasing. In particular, r(log h)0 ≤ (log h)0 (1) := α. Integrating this we obtain
h(r) ≥ rα h(1),
i.e. the average of u on ∂Br grows with homogeneity at least that detected at r = 1.
This technique was successfully adapted to equations of the form aij (x)uij = 0 when aij
are Lipschitz. This is interesting because solutions to such problems are generally C 2, α for
any α < 1, but no better. (This is the ”Schauder perturbation theory,” our next topic).
There are counterexamples to unique continuation when n ≥ 3 when aij (x) are Hölder
continuous, with any exponent < 1; it is true in 2D for merely bounded coefficients using
techniques from complex analysis.
EXERCISE: Assume that u is harmonic on Rn . Show that ∂B1 u(λx)u(λ−1 x) ds is inde-
R
pendent of λ. Use the local version of this statement to prove weak unique continuation: if
u vanishes on a set with non-empty interior in a connected domain Ω, then u ≡ 0.
14
2.7 Representation Formulae
The classical approach to the Laplace equation is to start with the observation that for
Γ = c(n)|x|2−n , we have ∆Γ = δ0 :
The proof is by removing a small ball around x, integrating by parts, noting that Γν goes
like the spherical surface area, and taking a limit. Remark on why the scaling invariance
λn−2 Γ(λx) = Γ(x) is only possibility (preserves mass of ∆).
As a consequence we get mean-value type representation formulae: If u is harmonic on
Rn and ϕ is a cutoff with ϕ = 1 in B1 , we have
Z
u(x) = u(y)[2∇Γ(x − y) · ∇ϕ(y) − Γ(x − y)∆ϕ(y)] dy
Rn
for x ∈ B1 . Note that the last term in the integral is supported outside of B1 . In particular,
for x ∈ B1/2 , the weight is analytic and bounded with bounded derivatives giving new proofs
that u is analytic and of the interior gradient estimate.
We can use this to get more sophisticated representation formulae. First, Green’s formula
(integration by parts twice) is
Z Z
(u∆v − v∆u) dy = (u vν − v uν ) ds.
Ω ∂Ω
Choosing v(y) = Γ(x−y)−wx , where wx (y) is any function with the same values and normal
derivative as Γ on ∂Ω, we get Z
u(x) = u∆wx dy.
Ω
By taking Ω to be a ball around x and wx to be quadratic, we recover the MVP.
If we instead choose vx (y) = Γ(x − y) − hx (y) where h is harmonic and has the same
boundary values as Γ, we obtain the boundary representation formula
Z
u(x) = u vν ds.
∂Ω
In this case, we call vx (y) := G(x, y) the Green’s function of Ω. We note that G < 0, is
symmetric in x and y (exercise), and looks like a fundamental solution centered at x when
we let y vary (and vice versa).
In some cases we can compute G explicitly using symmetries. The easiest case is the
half-space Ω = {xn > 0}. Note that for x̃ the reflection over {xn = 0} over x, we have
Γ(x̃ − y) is harmonic in Ω and has the same values as Γ on the boundary. Thus, the Green’s
function is
G(x, y) = Γ(x − y) − Γ(x̃ − y),
15
xn
and Gν = C(n) |x−y| n . It is a good exercise to check that this weight Gν has mass 1 indepen-
16
point in 2D (or ball minus a line segment in 3D (exercise)). Notion of capacity: small sets
cannot affect the Laplace, in measure. Probabilistic interpretation.
Remarks on why we can’t carry out a variational approach yet (need some compactness:
Sobolev space theory)
17
3 Schauder Estimates
Discuss Newtonian potential for f when f ∈ C02 (Rn ), solvability of Dirichlet problem. Mo-
tivate the need for estimates with less regularity (C α ) for applications, e.g. fully nonlinear
and 2D Euler.
Motivate the 0 < α < 1. Counterexample to C 2 regularity of potential when f continuous
(removable singularity → no C 2 solution). Motivate further with convergence of second
derivatives of Newtonian potential. So Laplace of C 2 functions are not everything. Remark
on Dini continuity; vs. 1/(r log r) ∼ (log log r)0 from example). Remark on relevance for 2D
Euler equations.
for all r > 0. Likewise, we say that u is pointwise C k, α at x0 if there exist a polynomial P
of degree k and a constant K such that
18
Remark 2. We can replace B1 with any E ⊂ B1 and get the same result and estimate,
using the same proof as above. (The important thing is that the separation from tangent
holds globally).
Remark 3. Likewise, we can replace B1 with B1 and get the same result, now up to the
boundary. (Do e.g. the C 1, α case, and do the boundary C 1, α Harnack as a by-product).
Proof. Let v be the harmonic function with the same boundary data as u. By the maximum
principle, |v| ≤ 1. Let Px be the k th order Taylor approximation to v at x. By the interior
derivative estimates for v we have kPx kC k (B1 ) ≤ 1 and kv − Px kL∞ (Br (x)) ≤ C0 rk+1 for r > 0.
By the maximum principle (comparison with 0 (1 − |x|2 )) we have that |v − u| ≤ 0 in
B1 . Taking r0 so small that C0 r0k+1 ≤ r0k+α /2 and 0 < r0k+α /2, we are done.
In the case k = 1, this can be viewed as an “improvement of flatness” lemma: if u is
almost harmonic and in a cylinder with height to length ratio (flatness) 1, then at scale r0
is lies in a cylinder with flatness r0α < 1.
If we assume that f = ∆u has certain regularity properties, we can produce Taylor
expansions of u by iterating the approximation lemma. The important cases for us are k = 1
and k = 2:
Lemma 3. Fix 0 < α < 1 and assume that u ∈ C ∞ (B1 ) ∩ C(B1 ), with ∆u = f . Then
for some 0 , C0 (n, α), if |f | ≤ 0 and |u| ≤ 1 in B1 , for each x ∈ B1/2 there exists a linear
function Lx with kLx kC 1 (B1 ) ≤ C0 and
19
|∆w| = r01−α |f | ≤ r01−α 0 ≤ 0 , so we may apply the approximation lemma again to obtain a
linear function l1 with kl1 kC 1 (B1 ) ≤ C and
2(1+α)
ku − l0 − r01+α l1 (·/r0 )kL∞ (Br2 ) ≤ r0 .
0
Proceeding in the same way we produce linear functions lk with klk kC 1 (B1 ) ≤ C and
m
X k(1+α) (m+1)(1+α)
ku − r0 lk (·/r0k )kL∞ (Brm+1 ) ≤ r0 .
0
k=0
20
Thus, the C α behavior of ∆u controls the same behavior for the full Hessian. Again, the
same estimate
kwkC 2, α (B1/2 ) ≤ C(n, α)kf kC α (Rn )
holds for w = Γ ∗ f when f ∈ C0α (B2 ).
Proof. We produce the desired tangent quadratic at the origin; the other points in B1/2 can be
taken care of by translation. After subtracting f (0)|x|2 /2n we may assume that f (0) = 0.
After multiplying by 0 /(kukL∞ (B1 ) + kf kC α (B1 ) ) we can apply Proposition 4. Taking the
estimates back to u and applying our characterization of Hölder spaces Proposition 11, we
are done.
Using our regularity results, we obtain:
Proposition 14. Given any ϕ ∈ C(∂B1 ) and f ∈ C α (B1 ) with 0 < α < 1, there exists a
unique solution in C 2, α (B1 ) ∩ C(B1 ) to
∆u = f in B1 , u|∂B1 = ϕ
Proof. Extend f to be compactly supported in B2 without changing the C α norm too much.
Let w = Γ ∗ f . Then by approximation with smooth functions, we have w ∈ C 2,α (B10 ) with
the interior estimate. Let v be the harmonic function with boundary data ϕ − w. Taking
u = v + w and applying the interior derivative estimates for harmonic functions, we are
done.
∆u = f in B1 , u|∂B1 = ϕ
Proof. Following the proof of Proposition 15, we reduce to estimating the C 2, α norm of the
harmonic function with boundary data ϕ − w. This follows directly from analyzing the
representation formula.
Alternatively, we can approach the boundary estimates the same way we did the interior
case (perturb from harmonic function case). After subtracting ϕ we may assume that u = 0
on ∂B1 . In e.g. B99/100 we have the estimate by the interior case. Noting that kukL∞ (B1 ) ≤
C(n)kf kL∞ (B1 ) (maximum principle) and using the Kelvin transform centered at a boundary
point, we reduce the problem to showing:
21
Proposition 16. Assume that u ∈ C 2, α (B1+ ) ∩ C(B1+ ) satisfies ∆u = f in B1+ with u|xn =0 =
0 and f ∈ C α (B1+ ). Then u ∈ C 2, α (B1/2
+
) and
The starting point (and key step) for Proposition 16 is a boundary approximation /
“improvement of flatness” lemma:
Lemma 5. Fix 0 < α < 1 and assume that u ∈ C 2, α (B1+ ) ∩ C(B1+ ) with u = 0 on the flat
boundary. Then there exist 0 , r0 , C0 (n, α) such that if |∆u| < 0 in B1+ and |u| ≤ 1, then
+
for each x ∈ B1/2 there is a harmonic quadratic Qx with kQx kC 2 (B1 ) < C0 and
The proof is again to compare with the harmonic function with the same boundary values
as u. We may “ignore” the flat boundary by using that the odd reflection of this harmonic
is harmonic in B1 . By iterating this lemma exactly as in the interior case we obtain
Lemma 6. Fix 0 < α < 1 and assume that u ∈ C 2, α (B1+ ) ∩ C(B1+ ) with u = 0 on the flat
+
boundary and ∆u = f . If for some x0 ∈ B1/2 we have |f (x)| ≤ 0 |x − x0 |α for all x ∈ B1+ ,
and |u| ≤ 1, then there exists a harmonic quadratic Qx0 with kQx0 kC 2 (B1 ) ≤ C1 (n, α) and
If the estimates were in the same balls, we would be done after taking small. Since the ball
on the right side is bigger, we need to work a little harder. To that end we state a rescaled
Schauder estimate for the Laplace equation, and a useful interpolation inequality.
For g ∈ C α (Ω) we define [g]C α (Ω) = supy6=x |g(y)−g(x)|
|y−x|α
= kgkC α − kgkL∞ .
[D2 u]C α (B1−δ ) ≤ C(n, α)(δ −2−α kukL∞ (B1 ) + δ −α kf kL∞ (B1 ) + [f ]C α (B1 ) ).
22
Proof. For any x ∈ B1−δ , translate so that x = 0 and apply the interior estimate to u(δx)/δ 2
to get
kD2 ukL∞ (B1−δ ) ≤ C(n, α)(δ −2 kukL∞ (B1 ) + kf kL∞ (B1 ) + δ α [f ]C α (B1 ) ).
Thus, for any x, y ∈ B1−δ with |y−x| > δ/2 the desired inequality for |D2 u(y)−D2 u(x)|/|y−
x|α follows.
If y ∈ Bδ/2 (x), translate x to 0 and apply the interior Schauder estimate for u(δx)/δ 2 to
complete the result.
Next we prove a useful interpolation lemma.
kD2 vkL∞ (B1 ) ≤ K(n, α)(δ −2 kvkL∞ (B1 ) + δ α [D2 v]C α (B1 ) ).
Indeed, we may assume each quantity on the right is ≤ 1 after dividing by the sum, and if
a second derivative were very large at some point in BR then it remains large in a ball of
radius 1 around this point, contradicting the L∞ bound.
Then, apply this inequality to v(δx)/δ 2 in B1/δ .
Combining the previous two lemmas gives:
Lemma 9. Assume that u ∈ C 2, α (B1 ), and that aij (x)uij = f in B1 with kaij − δij kC α (B1 ) ≤
0 . Then
[D2 u]C α (B1−δ ) ≤ C(n, α)(δ −2−α (kukL∞ (B1 ) + kf kC α (B1 ) ) + 0 [D2 u]C α (B1 ) ).
[D2 u]C α (B1−δ ) ≤ C(n, α)(δ −2−α kukL∞ (B1 ) + 0 δ −α kD2 ukL∞ (B1 ) + 0 [D2 u]C α (B1 )
+ δ −α kf kL∞ (B1 ) + [f ]C α (B1 ) )
≤ C̃(n, α)(δ −2−α (kukL∞ (B1 ) + kf kC α (B1 ) ) + 0 δ −α kD2 ukL∞ (B1 )
+ 0 [D2 u]C α (B1 ) ).
Applying Lemma 8 to the term kD2 ukL∞ (B1 ) yields the desired result.
We iterate the previous lemma to obtain the interior Schauder estimate.
Proposition 17. There exists 0 (n, α) small such that if kaij − δij kC α (B1 ) ≤ 0 , and u ∈
C 2, α (B1 ) solves aij (x)uij = f in B1 , then
23
Proof. WLOG kukL∞ (B1 ) + kf kC α (B1 ) ≤ 1 after dividing by it. We will apply the previous
lemma iteratively in the balls B1−2−k with k ≥ 1. Let γ = 22+α and let
Lemma 9 gives
ak ≤ C(n, α)(γ k + 0 ak+1 ),
or after re-arrangement,
1 1 1 k
ak+1 ≥ ak + ak − γ .
2C0 2C0 0
1
Take 0 so small that 2C0
≥ γ. Then the previous inequality gives
If ak ≥ −1
0 γ
k−1
then we conclude that ak+1 ≥ −1 k −1
0 γ , so if a1 ≥ 0 we would have by
induction that ak → ∞, contradicting that u ∈ C (B1 ). Thus, a1 = [D2 u]C α (B1/2 ) ≤ −1
2, α
0 ,
and the result follows by interpolation.
Following the exact same steps one can prove the corresponding boundary estimate:
Proposition 18. There exists 0 (n, α) small such that if kaij − δij kC α (B1+ ) ≤ 0 , and u ∈
C 2, α (B1+ ) solves aij (x)uij = f in B1+ and u|xn =0 = 0, then
kukC 2, α (B + )
≤ C(n, α)(kukL∞ (B1+ ) + kf kC α (B1+ ) ).
1/2
[D2 u]C α (B + ≤ C(n, α)(δ −2−α kukL∞ (B1+ ) + δ −α kf kL∞ (B1+ ) + [f ]C α (B1+ ) )
1−δ )
for the equation ∆u = f. We also have the exact same interpolation estimate Lemma 8
in B1+ . Combining the two for the equation aij uij = f with 0 boundary data on the flat
boundary gives
[D2 u]C α (B + ≤ C(n, α)(δ −2−α (kukL∞ (B1+ ) + kf kC α (B1+ ) ) + 0 [D2 u]C α (B1+ ) ).
1−δ )
Then iterate with δ = 2−k in the same way as above to complete the proof.
Remark 4. The same estimates from Propositions 17 and 18 hold if we assume that
with kaij − δij kC α (B1 ) + kbkC α (B1 ) + kckC α (B1 ) ≤ 0 . It is a good exercise to show this.
24
Remark 5. The assumption that we are in a perturbative setting is reasonable, by scaling
considerations: if aij (x)uij + bi (x)ui + c(x)u = f (x), with C α coefficients and right side, then
ũ = u(rx) solves
aij (rx)ũij + rbi (rx)ũi + r2 c(rx)ũ = r2 f (rx).
Thus, for r small (“zooming in” a lot), we may assume that ka − δk, kbk, kck are very small
in C α , and we enter the perturbative setting.
By combining the interior and boundary Schauder estimates for variable coefficients, it
is not hard to show:
Theorem 7. Assume that u, ϕ ∈ C 2, α (B1 ) and aij (x)uij = f for some aij , f ∈ C α (B1 )
with u = ϕ on ∂B1 . Assume further that the eigenvalues of (aij ) are in (λ, Λ) for some
0 < λ ≤ Λ < ∞. Then
Sketch of proof: After subtracting ϕ we may assume that ϕ = 0. By the maximum princi-
ple, kukL∞ (B1 ) ≤ C(n, λ, Λ)kf kL∞ (B1 ) . (Compare with the function kf kL∞ (1−|x|2 )/(2nλ))).
For any x0 ∈ B1/2 , consider ũ(x) = u(x0 + r0 Bx) with r0 small, B B T = (aij )(x0 ),
and x ∈ B1 . Then ũ solves an equation with coefficients ãij satisfying kãij − δij kC α ≤
C(λ, Λ)r0α kaij kC α (B1 ) . By choosing r0 small enough depending on λ, Λ, kaij kC α we may
assume this quantity is smaller than 0 . Applying Proposition 17 (and recalling the bound
on kukL∞ (B1 ) ) and then scaling back and applying a simple covering argument proves the
interior estimate.
At the boundary we may perform a diffeomorphism that flattens ∂B1 , and apply the
same steps as above. The only difference is that after changing coordinates, the new equation
contains lower-order coefficients. However, under the above rescaling these coefficients get
very small in C α (see Remark 5), so we are in the perturbative setting and can apply the
version of Proposition 18 with lower order terms. Scaling back gives the result.
Remark 6. Assume the same as in Theorem 7, except with an equation of the form
with C α coefficients and right side. Provided c ≤ 0, the same estimate holds, with
The argument is the same. We use that c ≤ 0 to obtain the L∞ estimate for u in terms of
kf kL∞ (B1 ) by the maximum principle. For general c, the estimate is true if we add CkukL∞ (B1 )
to the right side, but false otherwise (e.g. eigenfunctions of ∆).
25
3.5 Some Functional Analysis
A priori estimates of the type given in Theorem 7 are useful for obtaining existence results.
Here we indicate one useful technique, based on the idea that if two equations can be con-
nected through a path of equations with “uniformly good” a priori estimates, and the first
equation is solvable, then so is the second. This is known as the “method of continuity.”
Theorem 8. Let V and W be Banach spaces with norms k.kV , k.kW , and assume that L0 , L1
are bounded linear maps from V to W . If the operators Lt := tL1 + (1 − t)L0 satisfy
kxkV ≤ KkLt (x)kW (1)
for some K fixed independent of t ∈ [0, 1], and L0 is surjective, then L1 is surjective.
Proof. Take w ∈ W . The equation Lδ (v) = w is equivalent to
L0 (v) = w + (L0 − Lδ )(v) = w + δ(L0 − L1 )(v).
Since L0 is surjective and by (1) it is injective, it is invertible. We may thus rewrite the
equation as
v = L−1 −1
0 w + δL0 (L0 − L1 )(v) := T (v)
26
Remark 7. The same result is true with lower-order terms bi (x)ui + c(x)u with C α co-
efficients, provided c ≤ 0, by the same argument and Remark 6. In other scenarios, we
may still get existence using an argument known as the “Fredholm alternative” (see e.g.
Gilbarg-Trudinger, Chapter 5). For general c, we don’t always have existence (e.g. the ODE
f 00 + f = 1; solutions are 2π-periodic, so arbitrary boundary conditions cannot be imposed).
Remark 8. When the coefficients of the equation have more (say k) derivatives that are
C α , we have a priori C k+2, α estimates by differentiating the equation and applying the usual
Schauder estimate. As a consequence, we get existence and uniqueness of solutions that are
C k+2, α in these cases. In particular, with smooth coefficients we get smooth solutions, and
derivative estimates of all orders depending on the derivatives of the coefficients.
Remark 9. As an application of the Schauder theory, we can reduce the study of nonlinear
problems to low-regularity situations. For example, if we obtain a priori C 1, α estimates for
equations of the form
Fij (∇u)uij = 0
where F is smooth and uniformly convex, then we obtain derivative estimates of all orders.
Equations of this structure arise naturally in variational problems; thus for these problems,
gradient continuity is the name of the game.
27
4 Function Spaces
4.1 Sobolev Spaces and Approximation
4.2 Sobolev Inequality
4.3 Morrey Inequality and Morrey Spaces
4.4 BMO and John-Nirenberg Inequality (???)
4.5 BV Functions
Isoperimetric inequality, sets of finite perimeter
28
5 Divergence-Form Equations: Weak Solutions
5.1 Maximum Principle
5.2 Dirichlet Problem
(Some functional analysis)
5.3 Regularity
Also Schauder theory for C 0 , C α coefficients.
29
6 Non-divergence Equations: Strong Solutions
6.1 Calderon-Zygmund Estimate for Laplace Operator
6.2 Continuous Coefficients
6.3 ABP Estimate
6.4 Krylov-Safonov Harnack Inequality
30
7 Applications
Using the theory we developed so far for linear divergence and non-divergence equations, we
can solve the Dirichlet problem for some interesting nonlinear problems.
31