You are on page 1of 11

Partial Differential Equations In other words, we can aspire to actually solve

the PDE more-or-less explicitly, to compute so-


Lawrence C. Evans lutions, or else to indirectly deduce properties of
Department of Mathematics, University of the solutions (without relying upon formulas or
California, Berkeley numerics). This article surveys these viewpoints,
with particular emphasis upon the last.
1 Overview Terminology. A partial differential equation
(PDE) is an equation involving an unknown func-
This article is an extremely rapid survey of the tion u of more than one variable and certain of
modern theory of partial differential equations its partial derivatives. The order of a PDE is the
(PDEs). Sources of PDEs are legion: mathemat- order of the highest order partial derivative of the
ical physics, geometry, probability theory, contin- unknown appearing within it.
uum mechanics, optimization theory, etc. Indeed, A system of PDEs comprises several equations
most of the fundamental laws of the physical sci- involving an unknown vector-valued function u
ences are partial differential equations and most and its partial derivatives.
papers published in applied math concern PDEs. A PDE is linear if it corresponds to a lin-
The following discussion is consequently very ear operator acting on the unknown and its par-
broad, but also very shallow, and will certainly tial derivatives; otherwise the partial differential
be inadequate for any given PDE the reader may equation is nonlinear.
care about. The goal is rather to highlight some
Notation. Hereafter u usually denotes the
of the many key insights and unifying principles
real-valued solution of a given PDE, and is usu-
across the entire subject.
ally a function of points x = (x1 , . . . , xn ) ∈ Rn ,
typically denoting a position in space, and some-
1.1 Confronting PDEs times also a function of t ∈ R, denoting time. We
∂u
Among the greatest accomplishments of the phys- write uxk = ∂x k
to denote the partial derivative
2
ical and other sciences are the discoveries of fun- of u with respect to xk , ut = ∂u ∂ u
∂t , uxk xl = ∂xk ∂xl ,
damental laws, usually partial differential equa- etc for higher partial derivatives. The gradient of
tions. The great problems for mathematicians, u in the variable x is
both pure and applied, are then to understand
∇u = (ux1 , . . . , uxn ).
the solutions of these equations, using theoreti-
cal analysis, numerical simulations, perturbation (In this article, ∇u always as above denotes the
theory, and whatever other tools we can find. gradient in the variables x, even if u also depends
But this very success in physics, that some on t.) We write the divergence Pnof a vector field
fairly simple looking PDEs, for example the Eu- F = (F 1 , . . . , F n ) as div F = i=1 Fxi i .
ler equations for fluid mechanics (see (11) below), The Laplacian of u is the divergence of its gra-
model very complicated and diverse physical phe- dient: Pn
nomena, causes all sorts of mathematical difficul- ∆u = ∇2 u = k=1 uxk xk . (1)
ties. Whatever general assertion we try to show Let us also write u = (u1 , . . . , um ) to display
mathematically must apply to all sorts of solu- the components of a vector-valued function. We
tions with extremely disparate behavior. always use boldface for vector-valued mappings.
It is therefore a really major undertaking to The solid n-dimensional ball with center x and
understand solutions of partial differential equa- radius r is denoted B(x, r), and ∂B(x, r) is its
tions, and for this there are at least three primary boundary, a sphere. More generally, ∂U means
mathematical approaches: the boundary of a set U ⊂ Rn ; and we denote by
• discovering analytical formulas for solutions, Z
either exact or approximate, f dS
• devising accurate and fast numerical methods, ∂U
and the integral of a function f over the boundary,
• developing rigorous theory. with respect to (n − 1)-dimensional surface area.

1
2

1.2 Some important partial 1.2.3 Systems of PDEs


differential equations In a system of conservation laws
Following is a listing of some of the most com-
monly studied PDEs. To streamline and clarify ut + div F(u) = 0, (8)
the presentation, we have mostly set various phys-
each component of u = (u1 , . . . , um ) typically
ical parameters to unity in these equations.
represents a mass, momentum or energy density.
1.2.1 First-order PDEs A reaction-diffusion system of partial differen-
tial equations has the form
First-order PDEs appear in many physical the-
ories, mostly in dynamics, continuum mechanics ut − ∆u = f (u). (9)
and optics. For example, in the scalar conserva-
tion law Here the components of u typically represent den-
ut + div F(u) = 0 (2) sities of, say, different chemicals, whose interac-
tions are modeled by the nonlinear term f .
the unknown u is the density of some physically
interesting quantity and the vector field F(u), its The simplest form of Maxwell’s equations reads
flux, depends nonlinearly upon u. 
Another important first-order PDE, the Et = curl B

Hamilton-Jacobi equation Bt = −curl E (10)

div E = div B = 0,

ut + H(∇u, x) = 0, (3)
appears in classical mechanics and in optimal con- in which E is the electric field and B the magnetic
trol theory. In these contexts, H is called the field.
Hamiltonian.
Fluid mechanics provides some of the most
complicated and fascinating systems of PDEs
1.2.2 Second-order PDEs in applied mathematics. The most important
Second-order PDEs model a significantly wider are Euler’s equations for incompressible, inviscid
variety of physical phenomena than do first-order fluid flow:
equations. For example, among its many other (
interpretations, Laplace’s equation ut + u · ∇u = −∇p
(11)
div u = 0,
∆u = 0 (4)
records diffusion effects in equilibrium. Its time and the Navier–Stokes equations for incompress-
dependent analog is the heat equation ible, viscous flow:
(
ut − ∆u = 0, (5) ut + u · ∇u − ∆u = −∇p
(12)
also known as the diffusion equation. div u = 0.
The wave equation
In these systems u denotes the fluid velocity and
utt − c2 ∆u = 0 (6) p the pressure.
superficially somewhat resembles the heat equa-
tion, but as the name suggests supports solutions 1.2.4 Higher order PDEs
with utterly different behavior. Equations of order greater than two are much
Schrödinger’s equation less common. Generally speaking, such higher or-
der PDEs do not represent fundamental physical
iut + ∆u = 0, (7)
laws, but are rather derived from such.
for which solutions u are complex-valued, is the For instance, we can sometimes rewrite a sys-
quantum mechanics analog of the wave equation. tem of two second–order equations as a single
3

fourth–order PDE. In this way the biharmonic 2 Understanding PDEs


equation
∆2 u = 0 (13) We explore in this section several general proce-
dures for understanding partial differential equa-
comes up in linear elasticity theory. tions and their solutions.
The Korteweg–deVries (KdV) equation
2.1 Exact solutions
ut + auux + buxxx = 0, (14)
The most effective approach is of course just to
a model of shallow water waves, similarly ap- solve the PDE outright, if we can. For instance
pears upon our combining a complicated system the boundary–value problem
of lower order equations appearing in appropriate (
asymptotic expansions. ∆u = 0 in B(0, 1)
u = g on ∂B(0, 1)
1.3 Boundary and initial conditions is solved by Poisson’s formula
Partial differential equations very rarely appear 1 − |x|2
Z
g(y)
alone: most problems require us to solve the u(x) = dS,
nα(n) ∂B(0,1) |x − y|n
PDEs subject to appropriate boundary and/or
initial conditions. If for instance we are to study α(n) denoting the volume of the unit ball in Rn .
a solution u = u(x), defined for points x lying in The solution of the initial-value problem for the
some region U ⊂ Rn , we usually prescribe also wave equation in one space dimension,
something about how u behaves on the bound- (
ary ∂U . Most common are Dirichlet’s boundary utt − c2 uxx = 0 in R × (0, ∞)
condition u = g, ut = h on R × {t = 0},
u = 0 on ∂U (15)
is provided by d’Alembert’s formula
and Neuman’s boundary condition Z x+ct
g(x + ct) + g(x − ct) 1
u(x, t) = + h(y) dy.
∂u 2 2c x−ct
=0 on ∂U , (16) (19)
∂ν
The wave equation can also be solved in higher
where ν denotes the outward-pointing unit nor- dimensions, but the formulas become increasingly
mal to the boundary and ∂u ∂ν := ∇u·ν is the outer complicated. For example, Kirchhoff ’s formula
normal derivative. If, say, u represents a temper-
Z
ature, then (15) specifies that the temperature is 1
u(x, t) = h dS
held constant on the boundary, and (16) that the 4πc2 t ∂B(x,ct)
heat flux through the boundary is zero. ( Z )
∂ 1
Imposing initial conditions is usually appropri- + g dS (20)
∂t 4πc2 t ∂B(x,ct)
ate for time-dependent PDEs, for which we re-
quire for the solution u = u(x, t) that satisfies this initial-value problem for wave equa-
tion in 3 space dimensions:
u(·, 0) = g, (17) (
utt − c2 ∆u = 0 in R3 × (0, ∞)
(21)
where g = g(x) is a given function, comprising u = g, ut = h on R3 × {t = 0}.
the initial data. For partial differential equations
that are second order in time, such as the wave The initial-value problem for the heat equation
equation (6), it is usually appropriate to specify (
also ut − ∆u = 0 in Rn × (0, ∞)
(22)
ut (·, 0) = h. (18) u = g on Rn × {t = 0}
4

has for all dimensions the explicit solution Asymptotic matching. Solution of PDEs of-
Z ten display differing properties in different subre-
1 |x−y|2
gions. In this circumstance we can try to fashion
u(x, t) = n/2
e− 4t g(y) dy. (23)
(4πt) Rn an approximate solution by first (a) constructing
simpler approximate solutions in each subregion
Certain nonlinear PDEs, including the KdV and then (b) appropriately matching these solu-
equation (14), are also exactly solvable; and dis- tions across areas of overlap.
covering these so-called integrable partial differ- A common such application is to boundary lay-
ential equations is a very important undertaking. ers. The outer expansion for the solution within
some region often has a form like
It is however a fundamental truth that we can- ∞
X
not solve most partial differential equations, if by u (x) ∼ k uk (x). (24)
“solve” we mean coming up with a more-or-less k=0
explicit formula for the answer.
Suppose we expect different behavior near the
boundary, which we take for simplicity to be the
2.2 Approximate solutions, plane {xn = 0}. We can then introduce the
perturbation methods stretched variables yn = xn /α , yi = xi (i =
1, . . . , n − 1) and put ū (y) = u (x). We look
It is consequently important to realize that we then for an inner expansion
can often deduce properties of solutions without

actually solving the partial differential equation, 
X
either explicitly or numerically. ū (y) ∼ k ūk (y). (25)
k=0
One such approach develops systematic pertur- The idea now is to match terms in the outer ex-
bation schemes to build small “corrections” to a pansion (24) in the limit xn → 0 with terms in the
known solution. There is a vast repertoire of such inner expansion (25) in the limit yn → ∞. Work-
techniques. Given a PDE depending on a small ing this out determines for instance the value of
parameter , the idea is to posit some form for α in the scaling.
the corrections and to plug this guess into the
differential equation, trying then to fine tune the
2.3 Numerical analysis of PDEs
form of the perturbations to make the error as
small as possible. These procedures do not usu- Devising effective computer algorithms for PDEs
ally amount to proofs, but rather construct self- is a vast enterprise, far beyond the scope of this
consistent guesses. article; and great ingenuity has gone into the de-
sign and implementation of such methods.
Multiple scales. Homogenization problems Among the most popular are the finite differ-
entail PDEs with effects occurring on differing ence methods (which approximate functions by
spatial or temporal scales, say of respective orders values at grid points), the method of lines (which
1 and . Often a goal is to derive simpler effective discretizes all but the time variable), the finite
PDE that yield good approximations. We guess element method and spectral methods (which rep-
the form of the effective equations by supposing resents functions using carefully designed basis
an asymptotic expansion of the form functions), multigrid methods (which employ dis-

X cretizations across different spatial scales), the
u (x) ∼ k uk (x, x/) level set method (which represents free bound-
k=0 aries as a level set of a function), and many other
schemes.
and showing that the leading term u0 is a function The design and analysis of such useful numer-
of x alone, solving some kind of simpler equation. ical methods [IV.29], especially for nonlinear
This example illustrates the insight that sim- equations, depends upon a good theoretical un-
pler behavior often appears in asymptotic limits. derstanding of the underlying PDE.
5

2.4 Theory and the importance of for some speed σ. More generally, a solution u of
estimates a PDE in more space variables having the form
The fully rigorous theory of PDEs focuses largely u(x, t) = v(y · x − σt)
upon the foundational issues of the existence,
is a plane wave. An extremely useful first step
smoothness, and, where appropriate, uniqueness
for studying a PDE is to look for solutions with
of solutions. Once these issues are resolved, at
these special structures.
least provisionally, theorists turn attention to un-
derstanding the behavior of solutions. Dispersion. It is often informative to look for
A key point is availability, or not, of strong an- plane wave solutions of the complex form
alytic estimates. Many physically relevant PDEs
u(x, t) = ei(y·x−σt) (27)
predict that various quantities are conserved, but
these identities are usually not strong enough to where σ ∈ C and y ∈ Rn . We plug the guess (27)
be useful, especially in three dimensions. For non- into some given linear PDE, thereby to discover
linear PDEs the higher derivatives solve increas- the so-called dispersion relationship between y
ingly complicated, and thus intractable, equa- and σ = σ(y) forced by the algebraic structure.
tions. And so a major dynamic in modern theory For example, inserting (27) into the Klein–
is the interplay between (a) deriving “hard” ana- Gordon equation
lytic estimates for PDEs and (b) devising “soft”
mathematical tools to exploit these estimates. In utt − ∆u + m2 u = 0 (28)
the remainder of this article, we present for many 1
gives σ = ±(|y|2 + m2 ) 2 . Hence the speed |y| σ
important PDEs the key estimates upon which
of propagation depends nonlinearly upon the fre-
rigorous mathematical theory is built.
quency of the initial data eiy·x . So waves of dif-
ferent frequencies propagate at different speeds;
3 Behavior of solutions hence the dispersion.
Since PDEs model so vast a range of physical and Solitons. As a nonlinear example, putting
other phenomena, their solutions display an even (26) into the KdV equation (14) with a = 6, b = 1
vaster range of behaviors. But some of these are leads to the ODE
more prevalent than others. −σv 0 + 6vv 0 + v 000 = 0,

3.1 Waves a solution of which is the explicit profile


√ 
Many partial differential equations of interest in σ σ
v(s) = sech2 s
applied mathematics support at least some solu- 2 2
tions displaying “wavelike” behavior. for each speed σ. The corresponding traveling
The wave equation. The wave equation is wave u(x, t) = v(x − σt) is called a soliton.
of course an example, as is most easily seen in
one space dimension from d’Alembert’s formula
3.2 Diffusion and smoothing
(19). This dictates that the solution has the gen-
eral form u(x, t) = F (x + ct) + G(x − ct) and From the explicit formula (23) we can read off a
consequently is the sum of right- and left-moving lot of interesting quantitative information about
waves with speed c. The wave-like behavior en- the solution u of the initial-value problem (22) for
coded within Kirchhoff’s formula (20) in 3 space the heat equation.
dimensions is somewhat less obvious. In particular, notice from (23) that if the initial
Traveling waves. A solution u of a PDE in- data function g is merely integrable, the solution
volving time t and the single space variable x ∈ R u is infinitely differentiable in both the variables x
is a traveling wave if it has the form and t at later times. So the heat equation instantly
smooths its initial data; this observation makes
u(x, t) = v(x − σt) (26) sense as the PDE models diffusive effects.
6

3.3 Propagation speeds the normal velocity of Γ(t) equals its curvature.
This is a geometric law of motion for the evolving
It is also easy to deduce from (23) that if u solves
black/white patterns emerging in the asymptotic
the heat equation, then values of the initial data
limit  → 0.
g(y) at all points y ∈ Rn contribute to determin-
ing the solution at (x, t) for times t > 0. We can Much more complex pattern formation effects
interpret this as an “infinite propagation speed” can be modeled by systems of reaction-diffusion
phenomenon. PDEs of the general form (9): see pattern for-
By contrast, for many time-dependent PDEs mation [X.YY],
we have “finite propagation speed”: there is no
influence of some of initial data upon the solution
until enough time passes. This is so for first-order 3.5 Blow-up
PDE in general, for the wave equation, and re-
Solutions of time-dependent PDEs may or may
markably also for some nonlinear diffusion PDEs,
not exist for all future times, even if their initial
such as the porous medium equation
conditions at time t = 0 are well behaved. Note
ut − ∆(uγ ) = 0 (29) for example that among solutions of the nonlinear
heat equation
with γ > 1. The particular explicit solution
ut − ∆u = u2 , (31)
 1
γ − 1 |x|2 γ−1

1
u(x, t) = α b− β 2β (30) subject to Neumann boundary conditions (16),
t 2γ t +
are those solutions u = u(t) that do not de-
n
for α = n(γ−1)+2 1
, β = n(γ−1)+2 and x+ = pend on x and consequently that solve the ODE
ut = u2 . It is not hard to show that solutions of
max{x, 0} shows clearly that the region of pos-
this equation go to infinity (“blow up”) at a finite
itivity moves outward at finite speed.
positive time, if u(0) > 0.
For more general initial data, there is an inter-
3.4 Pattern formation esting competition between the diffusive, and so
Nonlinear terms can interplay with diffusion and stabilizing, term ∆u and the destabilizing term
create interesting effects. For example let Φ(z) = u2 .
1 2 2
4 (z − 1) denote a “two well” potential, hav-
ing minima at z = ±1. Look now at this scalar
3.6 Shocks
reaction-diffusion problem in which  > 0 is a
small parameter: Solutions of PDEs can fail to exist for large times
( for other reasons than blow up in the sup-norm.
ut − ∆u = 12 Φ0 (u ) in R2 × (0, ∞) Another possibility is that the gradient of a solu-
u = g  on R2 × {t = 0}. tion becomes singular at some finite time.
This effect occurs for conservation laws (2).
For suitable designed initial data functions g  , it Consider for example this initial-value problem
turns out that that for Burgers’ equation:
lim u (x, t) = ±1; (
→0 ut + 12 (u2 )x = 0 in R × (0, ∞)
(32)
so that the solution asymptotically goes to one u = g on R × {t = 0}.
or the other of the two minima of Φ. We can
informally think of these regions as colored black Assume we have a smooth solution u and define
and white. the characteristic curve x(t) to solve the ODE
For each time t ≥ 0, denote by Γ(t) the curve (
between the regions {u (·, t) → 1} and {u (·, t) → ẋ(t) = u(x(t), t) (t ≥ 0)
−1}. Asymptotic matching methods reveal that x(0) = x0 .
7

Then Much more complicated free boundary prob-


lems occur in fluid mechanics, in which the un-
d known velocity u satisfies differing sorts of PDE
u(x(t), t) = ux (x(t), t)ẋ(t) + ut (x(t), t)
dt within the sonic and subsonic regions. We say
= ux (x(t), t)u(x(t), t) + ut (x(t), t) = 0, that the equations change type across the free
boundary.
according to the PDE (32). Consequently
u(x(t), t) ≡ g(x0 ) and also the characteristic 4 Some technical methods
“curve” x(t) is in fact a straight line.
So far, so good; and yet the foregoing often So vast is the field of partial differential equa-
implies that the PDE does not in fact possess a tions that no small handful of procedures can pos-
smooth solution, existing for all times. For we sibly handle them all. Rather, mathematicians
can easily build initial data g for which the char- have discovered over the years, and continue to
acteristic lines emanating from two distinct initial discover, all sorts of useful technical devices and
points cross at a later time. tricks. This section provides a selection of some
A major task for the rigorous analysis of Burg- of the most important.
ers’ equation and related conservation laws is
characterizing surfaces of discontinuity (called 4.1 Transform methods
shocks) for appropriately defined generalized so-
lutions. A panoply of integral transforms is available to
convert linear, constant coefficient PDEs into al-
gebraic equations. The most important is the
3.7 Free boundaries Fourier transform [I.19]
Some very difficult problems require not only Z
1
finding the solution of some PDE, but also the re- û(y) := e−ix·y u(x) dx.
(2π)n/2 Rn
gion within which it holds. Consider for example
the Stefan problem, which asks us to determine Consider, as an example, the equation
the temperature within some body of water sur-
rounded by ice. The temperature solves the heat −∆u + u = f in Rn . (33)
equation inside a region whose shape changes in
time as the ice melts and/or the water freezes. We apply the Fourier transform and learn that
The unknowns are therefore both the tempera- (1 + |y|2 )û = fˆ. This algebraic equation lets us
ture profile and the so-called free boundary of the easily find û, after which a somewhat tricky in-
water. version yields the formula
There are in general two sorts of such free |x−y|2

e−s− 4s
Z Z
boundary problems that occur in PDE theory: 1
u(x) = f (y) dyds.
those for which the free boundary is explicit, such (4π)n/2 0 Rn sn/2
as the Stefan problem, and those for which it is
implicit. An example of the latter is the obstacle Strongly related are Fourier series methods,
problem: which represent solutions of certain PDEs on
min{u, −∆u − f } = 0. bounded domains as infinite sums entailing sines
and cosines. Another favorite is the Laplace
The free boundary is transform, which for PDEs is mostly useful as a
transform in the time variable.
Γ = ∂{u > 0},
4.2 Energy methods; functional
along which the solution satisfies the overdeter-
analytic framework
mined boundary conditions u = 0, ∂u∂ν = 0. Many
important physical and engineering free bound- For many PDEs various sorts of “energy esti-
ary problems can be cast as obstacle problems. mates” are valid, where we use this term loosely
8

to mean integral expressions involving squared Then, assuming u goes to zero as |x| → ∞ fast
quantities. enough, we have
Z
Integration by parts. Important for what
ut utt + c2 ∇u · ∇ut dx

ė(t) =
follows is the integration by parts formula: Rn
Z
= ut (utt − c2 ∆u) dx = 0;
Z Z Z
uxi v dx = − uvxi dx + uvν i dS Rn
U U ∂U
This demonstrates conservation of energy.
for each i = 1, . . . , n. Here ν denotes the outward- For the nonlinear wave equation
pointing unit normal to the boundary. This is a
form of the Divergence Theorem [x.yy] from utt − ∆u + f (u) = 0 (35)
multivariable calculus.
a similar calculation works for the modified en-
ergy
4.2.1 Energy estimates Z  
1 2 1
e(t) = ut + |∇u|2 + F (u) dx,
Assume that u solves Poisson’s equation Rn 2 2

where f = F 0 .
−∆u = f in Rn . (34)

Then, assuming u goes to zero as |x| → ∞ fast 4.3 Variational problems


enough to justify the integration by parts, we By far the most successful of the nonlinear the-
compute that ories is the calculus of variations; and indeed a
fundamental question to ask of any given PDE is
n
whether or not it is variational, meaning that it
Z Z X
2
f dx = uxi xi uxj xj dx appears as follows.
Rn Rn i,j=1
Given the Lagrangian density function L =
Z n Z n
X X L(v, z, x), we introduce the functional
=− uxi xi xj uxj dx = (uxi xj )2 dx.
Rn i,j=1 Rn i,j=1 Z
I[u] := L(∇u, u, x) dx,
U
This identity implies something remarkable: if
the Laplacian ∆u (which is the sum of the defined for functions u : U → R, subject to given
pure second derivatives uxi xi for i = 1, . . . , n) boundary conditions that are not specified here.
is square-integrable, then each individual sec- Suppose hereafter that u is a minimizer of I[·].
ond derivative uxi xj for i, j = 1, . . . , n is square- We will show that u automatically solves an
integrable, even those mixed second derivatives appropriate partial differential equation. To see
that do not even appear in the equation (34). this, put i(τ ) := I[u + τ v], where v vanishes near
This is an example of regularity theory, which ∂U . Since i has a minimum at τ = 0, we can use
aims to deduce the higher integrability and/or the chain rule to compute
smoothness properties of solutions. Z
0 = i0 (0) = (∇v L · ∇v + Lz v) dx;
U
4.2.2 Time dependent energy estimates and so
As a next example suppose u = u(x, t) solves the
Z
wave equation (6) and define the energy at time 0= (−div(∇v L) + Lz ) v dx,
U
t:
in which L is evaluated at (∇u, u, x). Here we
Z
1
u2 + c2 |∇u|2 dx.

e(t) := write ∇v L = (Lv1 , . . . , Lvn ).
2 Rn t
9

This integral identity is valid for all functions where


v vanishing on ∂U ; whence follows the Euler– n n
Lagrange equation X
ij
X
Lu = − a (x)uxi xj + bi (x)uxi + c(x)u.
−div(∇v L(∇u, u, x)) + Lz (∇u, u, x) = 0. (36) i,j=1 i=1

We say L is elliptic provided the symmetric ma-


Nonlinear Poisson equation. For example
trix ((aij (x))) is positive definite. In usual appli-
the Euler–Lagrange equation for
cations u represents the density
Pn ofijsome quantity.
Z
1 The second-order term a u records
I[u] = |∇u|2 − F (u) dx, Pn xi xj
i,j=1
U 2 diffusion, the first-order term i=1 bi uxi repre-
sents transport, and the zeroth-order term cu de-
is the nonlinear Poisson equation
scribes the local increase or depletion.
−∆u = f (u) (37)
We use the maximum principle to show for in-
where f = F . 0 stance that if c > 0, then u cannot attain a pos-
itive maximum at an interior point. Indeed if u
Minimal surfaces. The surface area of the took on a positive maximum at some point x0 ,
graph of a function u is then the first term of Lu at x0 would be nonneg-
ative (according to (40)), the next term would
Z
1
I[u] = 1 + |∇u|2 2 dx, be zero (according to (39)) and the last would
U
be positive. But this is a contradiction, since
and the corresponding Euler–Lagrange equation Lu(x0 ) = 0.
is the minimal surface equation
  Nonlinear elliptic equations. Maximum
∇u principle techniques apply also to many highly
div 1 = 0. (38)
(1 + |∇u|2 ) 2 nonlinear equations, such as the Hamilton–
Jacobi–Bellman equation
The expression on the left is (n times) the mean
curvature of the surface; and consequently a min-
max {Lk u} = 0. (42)
imal surface has zero mean curvature. k=1,...,m

4.4 Maximum principles This is an important equation in stochastic op-


timization theory, in which each elliptic opera-
The integral energy methods just discussed can tor Lk is the infinitesimal generator of a different
for certain PDEs be augmented with pointwise stochastic process. We leave it to the reader to
maximum principle techniques. These are pred- use the maximum principle to show that a solu-
icated upon the elementary observation that if tion of (42) cannot attain an interior maximum
the function u attains its maximum at an interior or minimum.
point x0 , then
Related, but much more sophisticated maxi-
uxk (x0 ) = 0, k = 1, . . . , n (39) mum principle arguments can reveal many of the
subtle properties of solutions to the linear elliptic
and
equation (41) and the nonlinear equation (42).
n
X
uxk xl (x0 )ξk ξl ≤ 0, ξ ∈ Rn . (40)
k,l=1 4.5 Differential inequalities

Linear elliptic equations. Such insights are Since solutions of PDEs depend on many vari-
essential for understanding the general second- ables, another useful trick is to design appropriate
order linear elliptic equation integral expressions over all but one of these vari-
ables, so that these expressions satisfy interesting
Lu = 0, (41) differential inequalities in the remaining variable.
10

4.5.1 Dissipation estimates, gradient 4.5.3 Monotonicity formulas


flows
For monotonicity formulas we try to find inter-
For example, let u = u(x, t) solve the nonlinear esting expressions to integrate over balls B(0, r),
gradient flow equation with center say 0 and radius r. The hope is that
these integral quantities will solve useful differen-
ut − div(∇L(∇u)) = 0 (43) tial inequalities as functions of r.
As an example, consider the system
in Rn × (0, ∞). Put

1
Z −∆u = |Du|2 u, |u|2 = 1. (45)
e(t) := L(∇u) dx.
2 Rn for the unknown = (u1 , . . . , um ), where we
Pn u P m
Then, assuming u goes to zero rapidly as |x| → write |Du| = i=1 j=1 (ujxi )2 .
2

∞, we have A solution u is called a harmonic map into the


unit sphere. It is a challenging exercise to derive
Z from (45) the differential inequality
ė(t) = ∇L(∇u) · ∇ut dx
Rn Z !
d 1
Z Z
2 2
=− (div∇L(∇u))ut dx = − (ut ) dx ≤ 0. |Du| dx
Rn Rn
dr rn−2 B(0,r)
Z
This is a dynamic dissipation inequality. 2 P k k
= i,j,k uxi xi uxj xj dS ≥ 0,
rn ∂B(0,r)

4.5.2 Entropy estimates


from which we deduce that
Related are entropy estimates for conservation Z Z
1 2 1
laws. For this assume that u = u (x, t) solves |Du| dx ≤ n−2 |Du|2 dx
rn−2 B(0,r) R B(0,R)
the viscous conservation law

ut + F (u )x = uxx (44) if 0 < r < R. This inequality is often useful, as
it lets us deduce fine information at small scales
for  > 0. Suppose Φ is a convex function and r from that at larger scales R.
put Z
e(t) := Φ(u ) dx. 5 Theory and application
R
Then The foregoing listing of mathematical viewpoints
Z Z and technical tricks provides at best a glimpse
ė(t) = Φ0 ut dx = Φ0 (−Fx + uxx ) dx into the immensity of modern PDE theory, both
Z R R pure and applied.
Ψ(u )x + Φ00 (ux )2 dx

=−
ZR 5.1 Well-posed problems
= − Φ00 (ux )2 dx ≤ 0, A common goal of most of these procedures is
R
to understand a given PDE (plus appropriate
where Ψ satisfies Ψ0 = Φ0 F 0 . What is important boundary and/or initial conditions) as a well-
is that we have found not just one, but rather posed problem, meaning that (a) the solution ex-
a large collection of dissipation inequalities, cor- ists, (b) is unique, and (c) depends continuously
responding to each pair of entropy/entropy flux on the given data for the problem. This is usually
functions (Φ, Ψ). the beginning of wisdom, as well-posed problems
Finding and utilizing entropy/entropy flux provide the starting point for further theoretical
pairs for systems of conservation laws of the form inquiry, for numerical analysis and for construc-
(8) is a major challenge. tion of approximate solutions.
11

5.2 Generalized solutions 4. P. Garabedian, Partial Differential Equations,


Wiley, 1964.
A central theoretical problem therefore is fashion- 5. F. John, Partial Differential Equations, 4th ed,
ing for any given PDE problem an appropriate Springer, 1991.
notion of solution for which the problem is well- 6. S. Klainerman, “Partial Differential Equa-
posed. For linear PDEs the concept of “distribu- tions”, in Princeton Companion to Mathematics,
tional solutions” is usually the best, but for non- Princeton U Press, 2008.
linear problems there are many, including “viscos- 7. P. A. Markowich, Applied Partial Differential
Equations: A Visual Approach, Springer, 2007.
ity solutions”, “entropy solutions”, “renormalized
8. J. Rauch, Partial Differential Equations,
solutions”, etc. Springer, 1991.
For example the unique entropy solution of the 9. M. Renardy and R. Rogers, An Introduction to
initial–value problem (2) for a scalar conservation Partial Differential Equations, 2nd ed, Springer,
law exists for all positive times, but may sup- 2004.
port lines of discontinuities across so-called shock 10. W. Strauss, Partial Differential Equations: An
waves. Similarly, the unique viscosity solution of Introduction, 2nd ed, Wiley, 2008.
the initial–value problem for the Hamilton–Jacobi 11. M. Taylor, Partial Differential Equations I–III,
2nd ed, Springer, 2001.
equation (3) generally supports surfaces of dis-
continuity for its gradient. The explicit solution
(30) for the porous medium equation is likewise Biography of contributor
not smooth everywhere and so needs suitable in-
terpretation as a valid generalized solution. Professor of Mathematics, UC Berkeley, 1989–
The research literature teems with many such
notions and some of the deepest insights in the
field are uniqueness theorems for appropriate gen-
eralized solutions.

5.3 Learning more


This article, as promised, is a vast survey that
actually explains precious little in any detail.
To learn more, interested students should defi-
nitely consult other articles in this book, as well
as the following suggested reading. Markowich
[7] is a nice introduction, with lots of pictures,
and Strauss [10] is a very good undergraduate
text, containing derivations of the various formu-
las cited here. Klainerman’s survey article [6] is
extensive and provides some different viewpoints.
My graduate level textbook [3] carefully builds
up much of the modern theory of PDEs, but is
aimed at mathematically advanced students.

Further Reading
1. E. DiBenedetto, Partial Differential Equations,
2nd ed, Birkhäuser, 2009.
2. H. Brezis, Functional Analysis, Sobolev Spaces
and Partial Differential Equations, Springer,
2012
3. L. C. Evans, Partial Differential Equations, 2nd
ed, American Math Society, 2010