You are on page 1of 105

*************************************

MTH 849 – Partial Differential Equations
(based on L.C. Evans’s textbook)
by
Baisheng Yan
Department of Mathematics
Michigan State University
yan@math.msu.edu

Contents

Chapter 1.

Single First Order Equations

1

§1.1.

Transport Equation

1

§1.2.

Linear first order equation

2

§1.3.

The Cauchy problem for general first order equations

3

§1.4.

Introduction to Hamilton-Jacobi equations

Suggested exercises
Chapter 2.

Laplace’s Equation

13
21
24

§2.1.

Green’s identities

24

§2.2.

Fundamental solutions and Green’s function

25

§2.3.

Mean-value property

35

§2.4.

Maximum principles

37

§2.5.

Estimates of higher-order derivatives and Liouville’s theorem

38

§2.6.

Perron’s method for existence using subharmonic functions

41

§2.7.

Maximum principles for second-order linear elliptic equations

44

§2.8.

Weak solutions and existence in Sobolev spaces

48

Suggested exercises
Chapter 3.

Heat Equation

53
55

§3.1.

Fundamental solution of heat equation

55

§3.2.

Weak maximum principle and the uniqueness

61

§3.3.

Regularity of solutions

65

§3.4.

Nonnegative solutions

67

§3.5.

Mean value property and the strong maximum principle

69

§3.6.

Maximum principles for second-order linear parabolic equations

71

Suggested exercises
Chapter 4.
§4.1.

Wave Equation

Derivation of the wave equation

78
80
80
iii

2. Energy methods and the uniqueness 94 §4. One-dimensional wave equations and d’Alembert’s formula 81 §4.3.Contents iv §4.5. Higher dimensional wave equation 85 §4. Finite propagation speed for general second-order hyperbolic equations 96 Suggested exercises Index 99 100 .4.

. 0) = g(x). bn ) is a constant vector in Rn . The directional derivative along the direction (c. But if g is not in C 1 . t) = f (d) = f (x − ct). hence the function along the straight line x = ct + d is constant. t + s) = Du · B + ut = 0 ds and hence u is constant on the line. . then ux = ux1 + ux2 c. u(x. ∞). t) = f (x.Chapter 1 Single First Order Equations 1. i. Here x − ct = d is called a characteristic line. where B = (b1 . u is not a classical solution. the line through (x. t) with the direction (B. then u is a classical solution. 0) = g(x − Bt). (c) Non-homogeneous problem ( ut + B · Du(x.e. we have u(x. Geometric method: (ux . in Rn × (0. t) = u(x − Bt. t) = 0. t) u(x. u(x.1. t + s). Coordinate method: Change variable. ∞) 1 . Since d u(x + Bs. ut ) is perpendicular to (c. ut = ux1 (−c) + ux2 . hence ut + cux = ux2 (1 + c2 ) = 0. t) = f (x1 ) = f (x − ct). i. 0) = g(x). 1) is represented parametrically by (x + Bs.. (b) (general n) Let us consider the initial value problem ( ut + B · Du(x. (a) (n = 1) Find all solutions to ut + cux = 0. s ∈ R. If g ∈ C 1 . . As in (a). t) = 0 in Rn × (0. . x2 = cx + t. 0). This line hits the plane t = 0 when s = −t at (x − Bt. but it is a weak solution as we will see later. 1). . given a point (x. Transport Equation ut + B · Du(x.e. t). u(x. x1 = x − ct. 1) is zero.

s) in terms of (x(τ. s). then we would have d z(s) = c(x.1. s)ds. Solution: The characteristic ODEs with the initial data are given by   x˙ = x. ∞) The idea is to find a curve (x(s). Note that d ˙ z(s) = ux x˙ + ut t. t(τ. t)ux (x. 0) = x2 . . s) in terms of (x. s) and obtain the solution defined by u(x. u(x. we see that s = t. −t 0 1. t)z(s) + d(x. s) and t(τ. t)ut (x. s)). t) = z(0) = z(−t) + z 0 (s)ds = g(x − Bt) + f (x + (s − t)B. ds We see that if x(s).1) which is called the characteristic ODEs for the PDE. t). Example 1. Solve the initial value problem ut + xux = u. z(τ. t) = z(τ. t + s). From the expression. t) from x(τ.2. u(x. s) and z(τ. t). ds This linear ODE in z along with the ODE (1.2) Finally. then we plug into z(τ. 0) = g(x). t(s)) on this curve can be calculated easily. s). Let x(τ. t).1. with τ ∈ R being a parameter. t(τ. t) u(x. s) be the solutions. s) = es τ. t(0) = 0. in R × (0.   z˙ = z. s) = τ 2 es = e−t x2 . s) = s. s).1) can be solved easily if x(0) = τ. t)u + d(x. (1. Linear first order equation ( a(x. x(0) = τ. s) = es τ 2 . Solve (τ. if we can solve (τ. Then z˙ = f (x + Bs. Hence Z 0 Z t u(x. t˙ = b(x. t˙ = 1. s). t(τ. t(s) satisfy ( x˙ = a(x. t(s)) so that the values of z(s) = u(x(s).2. t + s). Therefore. Hence x(τ. Let z(s) = u(x + Bs. t(0) = 0 and z(0) = g(τ ) are given as initial data at s = 0. t) = c(x. Linear first order equation 2 This problem can be solved as in part (b). t) + b(x. (1. τ = xe−t . z(0) = τ 2 . t) = z(τ.

u. n. However. . . . . z. We are then concerned with finding a solution u in some domain Ω in Rn containing Γ with a given Cauchy data: u(x) = g(x) for x ∈ Γ. where F = F (p.) This expression is not very promising since it involves the second derivatives of u.6) n X ∂F ∂F ∂F uxi xj (x(s)) + uxi (x(s)) + = 0.5) n X ∂F ∂F ∂F (Du. x) = 0. x)uxi xj + (Du. . . ∂pj ∂z ∂xi j=1 where Fpj . x = (x1 . we differentiate pi (s) = uxi (x(s)) with respect to s to get (1. we can differentiate the equation F (Du. (i = 1. u. z(s). x(s)).3) F (p(s). The Cauchy problem for general first order equations Consider a general first-order PDE F (Du. Let Γ be parameterized by a parameter y ∈ D ⊂ Rn−1 with x = f (y). First. The Cauchy problem is to find a solution u(x) such that the hypersurface z = u(x) in the xz-space passes a prescribed (n − 1)-dimensional manifold S in xz-space. . u. .3. x) = 0 with respect to xi to obtain (1. · · · .1. xn ) ∈ Ω ⊂ Rn . . The idea is the following: To calculate u(x) for some fixed point x ∈ Ω. x ∈ Rn .7) x˙ j (s) = ∂F (p(s). We now attempt to choose the function x(s) so that we can compute z(s) and p(s). . · · · . . Fpn ). Fxn ). along which we can compute u easily. .. 1. Fz and Fxi are evaluated at (p(s). . z ∈ R. . .4) p˙i (s) = n X uxi xj (x(s))x˙ j (s). Dz F = Fz . Derivation of characteristic ODE. How do we choose the curve so that all this will work? Suppose that u is a smooth (C 2 ) solution and x(s) is our curve with x(0) ∈ Γ defined on an interval I containing 0 as interior point. j=1 d (Here and below. we need to find a curve connecting this x with a point x0 ∈ Γ. Then the Cauchy data is given by u(f (y)) = g(f (y)) = h(y) y ∈ D. ∂pj ∂z ∂xi j=1 Evaluating this identity along the curve x = x(s) we obtain (1. x) is smooth in p ∈ Rn . u. x)uxi + (Du. The method we will use is motivated from the transport equation and linear equation. u. Let Dx F = (Fx1 . n).1. namely pi (s) = uxi (x(s)) for i = 1.3. Assume the projection of S onto the x-space is a smooth (n − 1)-dimensional surface Γ. 2. Suppose x(s) satisfies (1. the “·” means “ ds ”. Let z(s) = u(x(s)) and p(s) = Du(x(s)). z(s). x) = 0. Dp F = (Fp1 . ∂pj . n. Hence (1. z(s). x(s)) ∀ j = 1. x(s)) = 0. 2. and is called the method of characteristics. . The Cauchy problem for general first order equations 3 1.3.

x(0) = f (y). This system of 2n+1 first-order ODEs together with F (p(s). z(s). z. The solution (p(s).7) to get x(s) if u is known in C 2 . x > 0.3.7). Solve ( xuy − yux = u. The Cauchy problem for general first order equations 4 Of course. p(s) ˙ = −Dx F (p(s). z(s). Since z(s) = u(x(s)). this is just an ODE for x(τ ). Combining (1. Example 1.10) z(s) ˙ = Dp F (p(s). u. we therefore need to find the initial data for p(0). So (1. If we know the initial data (p(0). x(s)) for at least small s. In this case. x(s)) is called the full characteristics and its projection x(s) is called the projected characteristics. we will know the value of u(x) at x = x(s). We will accomplish this after some examples. for those s where x(s) ∈ Ω. z(s) = u(x(s)). x(s)) = 0 are called the characteristic equations for F (Du. x(s)). z(s). (a) Linear equations B(x)Du(x) + c(x)u = d(x). z(s).4)-(1. x(s))p(s). x) = 0. and hence Dp F = B(x). 0) = g(x). z. u. x > 0. becomes x(s) ˙ = Dp F (p. F (p. z(s). x(0)) according to the Cauchy data and thus z(0) = h(y). when the solution u(x) is known. z(s). together with F (p. z. z(s). then solve z. 1. Let u ∈ C 2 (Ω) solve F (Du.7)-(1.1. Assume x(s) solves (1. x(s)) as 2n + 1 unknown functions. Theorem 1.3. Remark: The characteristic equations are useful because they form a closed system of ODEs for (p(s). z(s). x(s))pi (s) − ∂z ∂xi Finally we differentiate z(s) = u(x(s)) to get (1. x) = 0 in Ω.9) z(s) ˙ = n X j=1 pj (s)x˙ j (s) = n X pj (s) j=1 ∂F (p(s).1. z(s). Then p(s) and z(s) solve the equations in (1. z(0).8) p˙i (s) = − (1. One can solve the first set of ODEs. x) = 0. . x) = B(x)p + c(x)z − d(x). As we will choose the initial data (z(0). this will be an important part of the method of solving the equation using the characteristics. x(s)). then we can find the values of (p(s).9) into a vector form: x(s) ˙ = Dp F (p(s). x(s)) · p(s). Definition 1. (1. where p(s) = Du(x(s)). z(s) ˙ = B(x)p = −c(x)z(s).1. p is not needed.10). So we can solve (1. What we have just demonstrated is the following theorem. ∂pj We can rewrite equations (1. y > 0. x(s)) − Dz F (p(s). x(0)) when s = 0. for p(s) = Du(x(s)) and z(s) = u(x(s)). Special cases.2. In this case.10). u(x. z(s). x) = B(x). Dz F = c(x). z(s).2. (p(s).7) yields ∂F ∂F (p(s). x(s)).

z(0) = y 2 . y) = y 2 . s) ds = 2y(es − 1).3. 0) = g(x). z).) (b) Quasilinear equations B(x. F (p. s) = yes . The initial data with parameter y are given by x1 (0) = 0. u(0. s). Since u(0. x(0)) = 0. x1 > 0. u)Du + c(x. Example 1. y) = g((x2 + y 2 )1/2 )earctan(y/x) . This solution makes sense only if 1 − yg(x − y) 6= 0. p2 (0) = q2 (y). y > 0. and p1 (0) = q1 (y). we have 1 so p1 (0) = q1 (y) = y. z). z. p2 = p2 (y. Dp F = B(x. 2 Z s x1 = x1 (y. s)) = Z(y. together with F (p. u(x. 2 Using these initial data we can solve the characteristic ODEs to obtain 1 z = Z(y. s) = y + p1 (y. s) = y 2 e2s . from F (p(0). becomes ( x˙ = B(x. z). z. y) = g(x−y) 1−yg(x−y) . s)ds = y(es + 1).4. x2 (y. p1 = p1 (y. x˙2 = Fp2 = p1 . The Cauchy problem for general first order equations (Answer: u(x. Once again. z. z(0). p˙1 = −Fx1 − Fz p1 = p1 . x) = p1 p2 − z. F (p. x2 (0) = y. q1 (y)q2 (y) = y 2 . s) = 2yes . p2 (0) = ux2 (y. In this case.) (c) Fully nonlinear problem. x2 ) = x22 . x) = 0. which are autonomous system for x. z)p = −c(x.10). z˙ = B(x. x) = B(x. 0) = 2y = q2 (y).1. 5 . differentiating with respect to y. z˙ = p1 Fp1 + p2 Fp2 = p1 p2 + p2 p1 = 2p1 p2 = 2z. z. Solve ( ux + uy = u2 . z). p is not needed. 2 0 This implies u(x1 (y. p˙2 = −Fx2 − Fz p2 = p2 . Solution: In this case. 0 Z s 1 x2 = x2 (y. s) = p2 (y. ( ux1 ux2 = u. Example 1. where. Hence (1.3. u) = 0. s) = y 2 e2s . z)p + c(x. (Answer: u(x. The characteristic equations are x˙1 = Fp1 = p2 .

we say that a vector p0 is admissible at y0 (or that (p0 .. there exists a smooth function q(y) for y near y0 such that q(y0 ) = p0 and (1. (p. The Cauchy problem for general first order equations 6 Given x = (x1 . y) = F (p. y) = n X fyij (y)pi − hyj (y) (j = 1.12) can be written in terms of a map F(p. h(y0 ). z0 .   ∂F . x) = 0 in Ω. x2 ). ∂y1 ∂y1   . · · · . . We say that an admissible (p0 . f 2 . y0 ) is non-characteristic if   ∂f 1 (y0 ) ∂f n (y0 ) .  6= 0. y ∈ D. s) = x1 and x2 (y. The Cauchy problem.11) F (p0 . y0 ) is admissible) if (1. z0 = h(y0 ) = u(x0 ). In what follows. . · · · .13) in a neighborhood J of y0 in D ⊂ Rn−1 . where Fj (p. x2 ) = Z(y. and hence we must have u(x1 . by the implicit function theorem. ∂yn−1 ∂yn−1 Fp1 (p0 . s) = y 2 e2s = 1 (x1 + 4x2 )2 . we solve x1 (y. z0 .12) are satisfied.11) and (1. s) and then obtain the value u(x1 . f (y)). n − 1). n − 1. i=1 Fn (p. · · · . y0 ) = 0. f n ) = f. . 2. First of all. . y0 ) is admissible and non-characteristic. 2. y ∈ D ⊂ Rn−1 .. Given y0 ∈ D. x0 ) = F (p0 .. p = q(y) is admissible at y.3. that is. . y) = 0.3. we assume (p0 . x0 ) In this case. x2 ). x0 ) . . Note that an admissible p0 may or may not exist.13) F(q(y). Note that x1 + 4x2 = 4yes . z0 . p ∈ Rn . i=1 where (f 1 . . h(y). y) from Rn × D to defined by F = (F1 . . we assume that Γ is parameterized by x = f (y) with parameter y ∈ D ⊂ Rn−1 . Note that p0 is admissible at y0 if and only if F(p0 . · · · . pn0 ) may equal Du(x0 ). it may not be unique. Fix y0 ∈ D. · · · . .11) and (1. Let x0 = f (y0 ). Fpn (p0 . Let u(f (y)) = g(f (y)) := h(y). s) = x2 for (y. Fn ).3. . f (y0 )) = 0. We now discuss the method of characteristics to solve the Cauchy problem: F (Du. We assume that p0 ∈ Rn is given such that (1. In order that p0 = (p10 .1. Rn Conditions (1. u. 16 (Check this is a true solution!) 1. Let q(y) be determined by (1. y)|(p0 . it is necessary that (1.y0 ) = det  det 1 n   ∂f (y0 ) ∂f (y0 ) ∂p   .12) hyj (y0 ) = n X fyij (y0 )pi0 ∀ j = 1. where D is a domain and f is a smooth function on D. u(x) = g(x) on Γ. even when it exists. .

By continuous dependence of ODE. X(y.  Finally we define u(x) = Z(y(x).10) with initial data p(0) = q(y). s) = n X j=1 pj (y. x(0) = f (y) ∀ y ∈ J. f (y. In ∂f ∂X fact. . Then. Since these solutions depend on parameter y ∈ J. P. there exists an open interval I containing 0. Z(y. s). p2 (y. s(x)) for x ∈ V. s). s). Note that Zs (y.  ∂X(y. y0 ) be admissible and non-characteristic. Lemma 1.1. . z0 . So s(x) = 0. From the definition of u(x) and Step (2).3. Proof. s). s)x˙ j (y. 0) = f (y0 ) = x0 . y = y(x) ∈ J such that x = X(y. Let (p0 . (1) If x ∈ Γ ∩ V then x = f (y) = X(y. y(x) are C 2 in x ∈ V. using ∂X ∂y |s=0 = ∂y and ∂s |s=0 = Fp (p0 . In fact. Let x0 = f (y0 ). X are C 2 in (y. z(s). . pn (y. s) = Fp · Ps + Fz Zs + Fx · Xs = Fp · (−Fz P − Fx ) + Fz Fp · P + Fx · Fp = 0 from the characteristic ODEs. z(0) = h(y).5. s) = F (P (y. det |y=y0 .. s(x) are C 2 . s) and x = X(y. · · · . Then. Proof. s) n n ∂f (y0 ) ∂f (y0 ) . u(x). . where P (y. s)). xn (y. s). x(s)) be the solution to Equation (1. s) = (p1 (y. y) is admissible. s(x). s) . u. X(y. F (p . The inverse function theorem gives the answer. a neighborhood J of y0 . (2) The function f (y. z0 . to display the dependence on y. s). z = Z(y. and fs (y. . we denote them by p = P (y. s) = (x1 (y. it remains to show that p(x) = Du(x). s). we have F (p(x).. s) is C 2 we also have y(x). and a neighborhood V of x0 such that for each x ∈ V there exist unique s = s(x) ∈ I. Z. . x0 ) ∂y1 ∂yn−1  . s). s(x)) ∀ x ∈ V. x) = 0 on V with the Cauchy data u(x) = g(x) on Γ ∩ V . 0) = 0 since (q(y). y0 ) be admissible and non-characteristic. So Cauchy data follow. s). We have X(y0 . . Moreover.  .  ∂(y. 0) for some y ∈ J. Then the function u defined above solves the equation F (Du. we have the following: Theorem 1. The Cauchy problem for general first order equations 7 Let (p(s). s(x)) = h(y(x)) = h(y) = g(f (y)) = g(x). s)). Note that since X(y. . x2 (y.  6= 0 . s)) = 0 for all s ∈ I and y ∈ J. s).2. s). x0 ). Let (p0 . z .s=0 = det  . . To finish the proof. ∂f Fp1 (p0 . x ) pn 0 0 0 ∂y1 ∂yn−1 from the noncharacteristic condition. x) = 0 ∀ x ∈ V. (3) Let p(x) = P (y(x).. y(x) = y and hence u(x) = Z(y(x). we find that the Jacobian determinant  ∂f 1 (y )  1 (y ) 0 0 . · · · .

s). The Cauchy problem for general first order equations 8 and. pj (y. we have. s))r(s). n − 1. s) ∂yi ∂yi j=1 Then r(0) = hyi (y) − ∂f j j=1 qj (y) ∂xi (y) Pn = 0 by the choice of q(y) = P (y. n − 1.1. pj (y.  n  X ∂F ∂xj ∂F ∂pj ∂F ∂Z + =− . s). and n n 2 j X X ∂pj ∂xj ∂2Z j ∂ x − p − r(s) ˙ = ∂yi ∂s ∂yi ∂s ∂s ∂yi j=1 j=1  n  X ∂pj ∂xj ∂pj ∂xj − = . Z(y. 0). s) − (y. s)) = 0 and differentiating with respect to yi . .3. for i = 1. Z(y. n − 1. · · · . 2. ∂yi ∂s ∂yi ∂s ∂yi ∂s Finally from F (P (y. s) = (y. ∂xj ∂yi ∂pj ∂yi ∂z ∂yi j=1 Hence  n  X ∂pj ∂xj ∂pj ∂xj r(s) ˙ = − ∂yi ∂s ∂s ∂yi j=1     n X ∂pj ∂F ∂F ∂F j ∂xj = + + p ∂yi ∂pj ∂xj ∂z ∂yi j=1 (1. X(y. Fixed i = 1. for r(s) solves a linear ODE with r(0) = 0. 2.14) = − ∂F (P (y. . . we claim n X ∂Z ∂xj (y. . . s). s). for each i = 1. ∂z Consequently r(s) ≡ 0. s). ∂yi ∂s ∂s ∂yi j=1 In the last equality we have used the fact n X ∂xj ∂Z = pj . ∂s ∂s j=1 and hence n n j=1 j=1 X ∂ 2 xj X ∂pj ∂xj ∂2Z = pj + . s) ∂yi ∂yi j=1 The proof comes from the following argument. X(y. · · · . let n r(s) = X ∂Z ∂xj (y. s).

along the line x = y.  Example 1. 0) = ux (x. (a) ( xux + yuy = u. the problem does not have a solution.3. τ )} is a projected characteristic curve.6. τ ) = τ ∀ τ. if u were a solution. . In fact. x) + uy (x. 0)} is a projected characteristic curve. the problem can not have a solution unless g(x) = kx for some constant k. along the line y = 0. We can see easily that u = αx + (1 − α)y is a solution for any α ∈ R. 0) = g(x) ∀ x ∈ R. n−1 uxj (x) = Zs X ∂s ∂y i + Z yi ∂xj ∂xj i=1 n X ! n−1 n k X X ∂y i ∂s ∂x = + pk x˙ k pk ∂xj ∂yi ∂xj i=1 k=1 k=1 ! n n−1 X X ∂xk ∂y i k k ∂s = + p x˙ ∂xj ∂yi ∂xj i=1 k=1 = (1. using u(x) = Z(y(x). we would have g 0 (x) = u(x. Γ = {(τ. In the case g(x) = kx the solution u = kx. In this case. What happens if Γ is a projected characteristic surface? Let us look at two examples in this case. x) 1 u(x. u(τ.1. τ )} is a projected characteristic curve.15) n X k=1 pk ∂xk = ∂xj n X pk δkj = pj (x). then. In this case. x) = ux (x. In this case. dx x x Hence g(x) = kx. we would have 0= d u(x. 0) = = . s(x)) for all x ∈ V. s(x)) and x = X(y(x). Γ = {(τ. u(τ. k=1 So Du(x) = p(x) on V . 0) g(x) d u(x. dx x x (c) ( xux + yuy = u. Γ = {(τ. The Cauchy problem for general first order equations 9 (4) Finally. u(x. obtained by eliminating parameters in the characteristic solutions. if u is a solution. however. In fact. (b) ( xux + yuy = u. τ ) = 1 ∀ τ. x) = = 6= 0.

. xn xn If α < 0. . The projected characteristic curve Cy in the xt-space passing through the point (y.. xn (0) = 1 and z(0) = h(y1 . xn−1 . xn−1 ). 0) = g(x). hence the characteristic solution u becomes singular (undefined).1. Du = (ux1 .16) where u = u(x. ). t) becomes singular for some t > 0 if and only if there exist y1 6= y2 such that (1. j = 1. u(x. .. z(y. . s) = g(y). we have es = xn . n. When n = 1. Therefore. t) becomes singular for some t > 0 unless B(g(y)) is a nondecreasing function of y. u(x.7. uxn ). B : R → Rn and g : Rn → R are given smooth functions. xn (y. . . . 1) = h(x1 . z. . . x(y. s) in terms of x. yj = xj /xn . t) if and only if (1. and so the characteristic ODEs are   x˙ = B(z). s) = y + sB(g(y)). . . t). let q = (p. u(x. s) = eαs h(y) = xαn h( Example 1. two distinct projected characteristics Cy1 and Cy2 intersect at a point (x. n − 1. z(0) = g(y). t) = g(x − tB(u(x. n − 1.. t(0) = 0   z˙ = 0. x1 xn−1 . ..17) holds. x) = pn+1 + B(z) · p. s) = yj es . if B(g(y)) is not nondecreasing in y. . the only nice solution in a neighborhood of zero is identically zero since u(tx0 ) = tα u(x0 ). then there exist y1 < y2 satisfying . . s) = s. . g(y1 ) 6= g(y2 ). . t))). Furthermore. 2. . where y ∈ Rn is the parameter. · · · . z˙ = αz with the initial data xj (0) = yj for j = 1. At the intersection point (x. j = 1.17) y1 − y2 = t(B(g(y2 )) − B(g(y1 ))). t(y.3. . u(x) = z(y. . So xj (y. . s) = es . x(0) = y t˙ = 1. The Cauchy problem for general first order equations 10 Example 1. 0) is the line (x. Therefore. Hence u satisfies the implicit equation u(x. . Solution: The characteristic ODEs are x˙ j = xj . Solve ( ut + B(u) · Du = 0.8. t ∈ R). The solution is given by z(y. . Solve (P n j=1 xj uxj = αu u(x1 . . Solve (y. s) = eαs h(y). t) (x ∈ Rn . . pn+1 ) and F (q. t) = (y + tB(h(y)). In this case. t) along which u is a constant g(y). yn ). (1. In fact.

0) dx −∞ 0 −∞ holds for all smooth functions v : R × [0.18) u(x.21) Now select a test function v with compact support in V . 0) = g(x). which gives a contradiction. y > 0.3. by integration by parts and (1. Suppose R. t). Suppose V is a subdomain of R × (0. B(g(y2 )) − B(g(y1 )) If u were regular solution at (x. For certain first-order PDEs.19) becomes Z ∞Z ∞ 0= [R(u)vy + S(u)vx ] dxdy 0 −∞ ZZ = ZZ [R(u)vy + S(u)vx ] dxdy + [R(u)vy + S(u)vx ] dxdy. Let u be a weak solution to (1. g(y1 ) 6= g(y2 ).1. t) defined on whole R2 unless B(g(y)) is a constant function. Vl Vl This implies (1. ∞) → R with compact support.20). The Cauchy problem for general first order equations 11 B(g(y1 )) > B(g(y2 )) and hence two characteristics Cy1 and Cy2 intersect at a point (x. The same argument also shows that the problem cannot have a regular solution u(x. that are not necessarily in C 1 or even continuous. 0)v(x.4. t) with y2 − y1 t=− > 0. we say that u is a weak solution (or an integral solution) if Z ∞Z ∞ Z ∞ (1. Weak solutions.20) (R(u))y + (S(u))x = 0 in Vl . (1.19) (R(u)vy + S(u)vx ) dxdy = g(x. For example. t) would be equal to g(y1 ) and g(y2 ) by the analysis above. Similarly. (1.3. for the divergence form equation ( ∂R(u) ∂S(u) x ∈ R. (R(u))y + (S(u))x = 0 in Vr . ∂y + ∂x = 0. Then (1. S are smooth functions. Again (1. ZZ ZZ [R(u)vy + S(u)vx ] dxdy = − [(R(u))y + (S(u))x ]v dxdy Vl Z + C Vl [R(u− )νy− + S(u− )νx− ]v dl = Z C [R(u− )νy− + S(u− )νx− ]v dl. ∞) and C is a smooth curve x = ζ(y) which separates V into two regions Vl = V ∩ {x < ζ(y)} and Vr = V ∩ {x > ζ(y)}.18) such that u is smoothly extended onto both V¯l and V¯r . However. . it is possible to define weak solutions.19) becomes ZZ ZZ 0= [R(u)vy + S(u)vx ] dxdy = − [(R(u))y + (S(u))x ] v dxdy. 1. we have. since B(g(y1 )) > B(g(y2 )). then u(x. First we choose a test function v with compact support in Vl . Vl Vl Now since v has compact support in V .

h(x) = 1 − x.20). dy This condition sometimes is written as ζ 0 [[R(u)]] = [[S(u)]]. then u is a weak solution to (1.   0 if x ≥ 1. we deduce that Z [(R(u+ ) − R(u− ))νy− + (S(u+ ) − S(u− ))νx− ]v dl = 0. y) = 0 if x > ζ(y).19) for all test functions v with compact support in V ). and u− is the limit of u from Vl to the curve C.   0. Likewise. (1. uy + 2 x (Different R. and u+ is the limit of u from Vr to the curve C. S lead to different results.21) and (1. 0 ≤ y < 1.3.1. where   if x ≤ 0. Therefore. if x ≥ 1. . νy− ) = p . y) = 1−y if y ≤ x ≤ 1. 0 ≤ y < 1.22) dζ [R(u+ ) − R(u− )] = [S(u+ ) − S(u− )]. the unit out-normal ν − on C is given by (1. y ≥ 1. νy− ) is the unit out-normal of domain Vl . u(x. Vr ν+ C (νx+ . The Cauchy problem for general first order equations 12 where ν − = (νx− .9. Since the domain Vl is inside x < ζ(y). Note that ν − = −ν + .22). and define ( 1 if x < ζ(y). νy+ ) where = is the unit out-normal of domain Vr . 0) = h(x). C This equation holds for all test functions v as above. −ζ 0 (y)) ν − = (νx− . and so (R(u+ ) − R(u− ))νy− + (S(u+ ) − S(u− ))νx− = 0 on C. ZZ Z [R(u)vy + S(u)vx ] dxdy = [R(u+ )νy+ + S(u+ )νx+ ]v dl. if 0 ≤ x ≤ 1. with [[A]] standing for the jump of A across curve C. we obtain the Rankine-Hugoniot condition: (1. S(u) = u2 /2:  2 u = 0. The classical solution u(x. 0 ≤ y < 1.18) in V (defined by (1. y ≥ 1. 1  1−x u(x. In fact. y) (by the characteristics method) is well-defined for 0 < y < 1 and is given by  if x ≤ y. if u is smooth up to V¯l and V¯r and satisfies (1. 1 + (ζ 0 (y))2 Therefore. How should we define u for y ≥ 1? We set ζ(y) = (1 + y)/2. Consider the Burgers’ equation uy + uux = 0 as a divergence-form equation with R(u) = u. Example 1.) We study the weak solution of this equation with intial data u(x. 1.

t).4. x). x). x). p(0) = h(y). pn+1 ) and y = (x. (Note that u2 is continuous across both rays x = 0 (y > 0) and x = y (y > 0) and is a smooth solution on each of the three regions between the two rays. Can we find a further condition which ensures that the weak solution is unique and physical? The answer is yes. x). z. u(x.1.) However. y) = 1 if x > y/2. The initial data for the characteristic ODEs at s = 0 are given by x(0) = y.) So there are two different weak solutions for this Cauchy problem. y) = pn+1 + H(p. z(0) = g(y). [[S(u1 )]] = 12 and ζ 0 = 12 . x˙ = Hp (p. the equations ( p˙ = −Hx (p. we can define another such solution as follows:   if x > y 1 u2 (x. Then the method of characteristics this time does not lead to any ambiguity in defining u for all y > 0. 1). Define ( 0 if x < y/2 u1 (x. Introduction to Hamilton-Jacobi equations We study the initial value problem for the Hamilton-Jacobi equation ( ut + H(Du. x).) We now consider the initial data ( 0 if x ≤ 0. (Note that [[R(u1 )]] = 1. t) = ut + H(Du. it is usually called the entropy condition. Fz = 0. therefore. h(x) = 1 if x > 0. 0) = g(x). 0). We omit the details in this course. It is easy to check that the shock condition (Rankine-Hugoniot) is satisfied. x) = 0 with F (q. but does fail to define solution in the domain 0 < x < y. The equation can be written as F (Du. the sense of weak solutions so defined does not give the uniqueness of solution. (Note that [[R(u)]] = −1. Introduction to Hamilton-Jacobi equations 13 It can be checked that this u is a weak solution containing a shock (where u is discontinuous). x) · p − H(p. 1. . where q = (p. x). The function u2 is also a weak solution. In particular. [[S(u)]] = − 21 and ζ 0 = 12 . x) = 0. x. x). ut . y) = x/y if 0 < x < y   0 if x < 0. called a rarefaction wave. z˙ = Hp (p. So Fq = (Hp (p.4. t(0) = 0. x˙ = Hp (p. The characteristic equations are p˙ = −Hx (p. and u is a weak solution. x) are the so-called Hamilton’s equations in ODEs. Fy = (Hx (p. u. t˙ = 1.

s) = s and hence x = x(y. ˙ x). I[x] ≤ I[w] ∀ w ∈ A. ˙ x(s))ds with x(0) = y and x(t) = x. ˙ v = x˙ and p = p(y. Introduction to Hamilton-Jacobi equations 14 where h(y) is some function satisfying the admissible conditions. t). namely. Assume L(v. The minimizer x satisfies the Euler-Lagrange equation: − d (Lv (x(s). x). (1. t) = z(y. Theorem 1.1. ˙ x) + Lx (x. x ∈ Rn . We assume such a minimizer x exists in A. ˙ x + τ y) · y) ds. Define i(τ ) = I[x + τ y].4. v. w(t) = x}. ˙ w(s)) ds ∀ w ∈ A. called the Lagrangian. ˙ x) · y) ds = 0. 0 We use integration by parts in the first integral to obtain  Z t d ˙ x) + Lx (x. t) and z = z(y. · · · . d − Lv (x. x). ds  . x) · p − H(p. t) = g(y) + 0 This is the action functional in classical mechanics. Then z˙ = L(x. ds 0 This identity holds for all smooth functions y(s) satisfying y(0) = y(t) = 0. Then i(τ ) is smooth in τ and has minimum at τ = 0. x). Proof. t) are related by (1. p = p(y.23) L = Hp (p.24) u(x. We can explicitly compute i0 (τ ) to obtain Z t 0 i (τ ) = (Lv (x˙ + τ y.4. ˙ x) · y˙ + Lx (x. v = Hp (p. x). t). 0 A basic problem in the calculus of variations is to find a curve x(s) ∈ A minimizing the functional I[w]. − Lv (x. y ∈ Rn and a time A := {w : [0. y n (s)) be a smooth function satisfying y(0) = y(t) = 0. Let y(s) = (y 1 (s). 2.3. The calculus of variations. ˙ x(s)) = 0 ds (0 ≤ s ≤ t). ˙ x) = 0. ˙ x + τ y) · y˙ + Lx (x˙ + τ y. is a given smooth function. The quantities L = z. Fix two points x. 0 So i0 (0) = 0 implies Z t (Lv (x.23) we can eliminate p to have the relationship L = L(v. So t = t(y. 1. Suppose from (1. Define the action functional Z I[w] = t L(w(s). t] → Rn | w(0) = y. So the solution u is obtained from x(s) ˙ and x(s) by Z t L(x(s).1. and so for all 0 ≤ s ≤ t. ˙ x(s))) + Lx (x(s). So i0 (0) = 0. ˙ x) · y ds = 0. 1.

and Hpi (p. x) can be uniquely solved for v to have v = m p. Then Lv (v. x) − Lxi (v(p. x) associated with Lagrangian L(v. x). x(s)) = Hp · p(s) ˙ + Hx · x(s) ˙ = −Hp Hx + Hx Hp = 0. for each i = 1. ˙ x(s)). We make the following assumption: For all p. x) = p · vxi (p. x) · vxi (p. x ∈ Rn ). x(s)). Let L(v. x(s)).10. x). x(s)) = Hp (p(s). x). Finally. x) − Lv (v(p. x) = p · m¨ x(s) = F (x(s)). x). Thus x(s) ˙ = v(p(s). x) = 12 m|v|2 − φ(x). L is total action and H is total energy density. the equation p = Lv (v.  . x) = p · v(p. x) can be uniquely solved for v as a smooth function of (p. Furthermore H(p(s). x) = v i (p. we have x(s) ˙ = v(p(s). x(s)) (0 ≤ s ≤ t). p(s) ˙ = −Hx (p(s). x(s)) = −Hx (p(s). We compute by the definition of H(p.4. By definition of p(s) and the Euler-Lagrange equation. By assumption. p = Lv (v(p. x) + p · vpi (p.1. The minimizer x(s) and momentum p(s) satisfy Hamilton’s equations: x(s) ˙ = Hp (p(s). 2. m 2 m 2m In this example. x) = −Lxi (v(p. x). x). x). Theorem 1. The corresponding EulerLagrange equation is the Newton’s law: H(p. p(s) ˙ = Lx (x(s). ds So H(p(s). x). Given p ∈ Rn . · · · . Hence the Hamiltonian 1 1 1 1 p − m| p|2 + φ(x) = |p|2 + φ(x). Since p(s) = Lv (x(s). n. Introduction to Hamilton-Jacobi equations 15 We now set the generalized momentum to be p(s) := Lv (x(s). Proof. x) − L(v(p. x(s)) is constant on 0 ≤ s ≤ t. x) (p. x) = v i (v(p. v = v(p.4. x(s)). x) = mv. x). x). x(s)) for 0 ≤ s ≤ t. x ∈ Rn . ˙ x(s)) = Lx (v(p(s). x). x ∈ Rn ). We then define the Hamiltonian H(p. x(s)). x) (p. x(s)) = constant. x) − Lv (v(p. observe d H(p(s). equation 1 p = Lv (v. x) by H(p. F (x) = −Dφ(x). ˙ x(s)) (0 ≤ s ≤ t). Hxi (p. Example 1. x). x) · vpi (p.

Legendre-Fenchel transform and the Hopf-Lax formula.24) suggests us to study the following function: Z t  . if H is smooth then. Moreover.4. if H = L∗ is differentiable at p then. Introduction to Hamilton-Jacobi equations 16 1. q = Hp (p). if L ∈ C(Rn ) and H = L∗ then L = H ∗ . Theorem 1.4. For each L ∈ C(Rn ). we define the Legender-Fenchel transform of L to be L∗ (p) = sup [q · p − L(q)] q∈Rn (p ∈ Rn ). Given any function L : Rn → R.) Therefore. L∗ ∈ C(Rn ). p = Lq (q).1. we can eliminate p to have L = L(q) = H ∗ (q). In this case we say H. (L∗ )∗ = L. (For nonsmooth convex functions.2. from q = Hp (p). for any q ∈ Rn . if L is differentiable at q then. for any p ∈ Rn . the following conditions are equivalent: p · q = L(q) + L∗ (p). We also assume H : Rn → R is convex and satisfies the super-linear growth condition: H(p) = ∞.5. (1. Thus. Likewise. L are dual convex functions. L = Hp (p) · p − H(p). In order to solve the Hamilton-Jacobi equation. similar duality properties also hold. Moreover. the following conditions are equivalent: p · q = L(q) + H(p). see Problem 11. |p|→∞ |p| lim The set of all such convex functions on Rn is denoted by C(Rn ). We now study the simple case where Hamiltonian H = H(p) depends only on p ∈ Rn .

.

t) := inf L(w(s)) ˙ ds + g(w(0)) . 1 n u(x.

the equation ut (x. 0 where we assume that g : Rn → R is a Lipschitz continuous function. t > 0. . Lip(g) = sup x. t > 0. ∞) satisfying u(x. y∈R t and u is Lipschitz continuous on Rn × [0. t) = minn tL + g(y) . (Hopf-Lax formula) For x ∈ Rn . then at this point (x. 0) = g(x) for all x ∈ Rn . t). w ∈ C ([0. that is. Furthermore. R ). w(t) = x . x6=y |g(x) − g(y)| < ∞.6. Theorem 1. t). t)) = 0 is satisfied.     x−y u(x. t) + H(Du(x.y∈Rn . t]. if u is differentiable at some point (x. |x − y| The following Hopf-Lax formula is useful because it transforms this infinite-dimensional minimization problem to a finite-dimensional one and provides a good weak solution to the Hamilton-Jacobi equation.

4. then by Jensen’s inequality. which will also be used for checking the Hamilton-Jacobi equation Lemma 1. s) as initial data (at initial time s) to solve u(x. t) = min[tL( ) + g(z)] − tL( ) − g(y) z t t (choosing z = x ˆ − x + y) ≤ g(ˆ x − x + y) − g(y) ≤ Lip(g)|x − x ˆ|. t) ≥ infn tL + g(y) . t]. u(x. t)| ≤ Lip(g)|x − x ˆ|. Since L is convex. This proves |u(x. y∈R t−s In other words. t) − u(x. x−z x−y y−z tL( ) ≤ (t − s)L( ) + sL( ). For each x ∈ Rn and 0 ≤ s < t. t) on [s. t) − u(ˆ x. s) = sL( y−z s ) + g(z). t−s . 1. Proof. t) ≤ L(w(s))ds ˙ + g(w(0)) = tL t 0 Hence     u(x. Use this w we have   Z t x−y + g(y). we first prove the folllowing result. We prove that u is Lipschitz on Rn × [0. t) ≤ infn tL y∈R x−y t   + g(y) . t) = minn (t − s)L + u(y. L(w(s))ds ˙ + g(w(0)) ≥ tL( ) + g(y) ≥ infn tL y∈R t t 0 This proves   x−y u(x. given y ∈ Rn .11. t) = tL( ) + g(y). fix t > 0 and x. s) . Choose z ∈ Rn such that u(y. Choose y ∈ Rn such that x−y u(x. t 0 t 0 t where y = w(0). x ˆ ∈ Rn . Rn ) with w(t) = x. let w(s) = y + st (x − y) for s ∈ [0. ∞). to compute u(x. Step 1.1. y∈R t On the other hand. Introduction to Hamilton-Jacobi equations 17 Proof. Fix y ∈ Rn . Step 2. t) = infn tL + g(y) y∈R t and that the infimum is a minimum follows by the growth condition of L. s). To establish the Lipschitz condition on t.  Z t  Z 1 t 1 x−y L(w(s))ds ˙ ≥L w(s)ds ˙ = L( ). First. t) ≤ tL( ) + g(z) ≤ (t − s)L( ) + sL( ) + g(z) t t−s s x−y = (t − s)L( ) + u(y. Hence     Z t x−y x−y + g(y) . we can compute u at an earlier time s and then use u(y. If w ∈ C 1 ([0.     x−y u(x. t). We first show that     x−y u(x. t]. t Then x ˆ−z x−y u(ˆ x. t t−s s Hence x−z x−y y−z u(x. t].

z = st x + (1 − st )y. t0 ) − u(x0 − q. Let y ∈ Rn be such that u(x. t) − u(x. max |L∗ (w)|}. u(z. t0 ). y∈R t−s Finally the infimum is a minimum as seen from the above. Choose z such that namely. in this case. given any q ∈ Rn . s) ≤ L(0)|t − s|. Let  > 0 and. Finally. t) ≥ infn (t − s)L( ) + u(y. s) u(x. s) = minn (t − s)L y∈R t−s     x−y ≥ minn (t − s)L − Lip(g)|x − y| y∈R t−s = (t − s) minn {L(z) − Lip(g)|z|} (choosing y = x − (t − s)z)   z∈R = −(t − s) maxn max {w · z − L(z)} z∈R w∈BR (0) (R = Lip(g)) = −(t − s) max maxn {w · z − L(z)} = −(t − s) max L∗ (w).1. So. tˆ)| ≤ C(|x − x ˆ| + |t − tˆ|). = tL( t This proves   x−y u(x. t)| + |u(ˆ x. s)| ≤ C|t − s| ∀ x ∈ Rn . t0 − ). u(x. s) − u(x. From the lemma. w∈BR (0) z∈R w∈BR (0) Therefore. maxw∈BR (0) |L∗ (w)|}. t) − u(x. s) + (t − s)L( Continuation of Step 2. t−s 2. w∈BR (0) Step 3. x ∈ Rn . t) − u(ˆ x. Furthermore. t) ≤ (t − s)L(0) + u(x. tˆ)| ≤ |u(x. ∀ x ∈ Rn . t0 ) + H(Du(x0 . We show that ut (x0 .   x−y + u(y. t) = tL( x−y t ) + g(y). given tˆ > 0. t0 ). x−z x−y z−y = = .4. where C = max{|L(0)|. t) ≤ infn y∈R  x−y (t − s)L( ) + u(y. where C = max{Lip(g). w∈BR (0) Hence |u(x. x−z t−s = x−y t . we have |u(x. t0 > 0. Introduction to Hamilton-Jacobi equations 18 This proves  u(x. t−s t s Hence z−y x−z x−z ) ≤ sL( ) + g(y) + (t − s)L( ) t−s s t−s x−y ) + g(y) = u(x. t) − u(x. s) . t) − u(ˆ x. −|t − s| max L∗ (w) ≤ u(x. s) . |L(0)|. t > s ≥ 0. t) − u(ˆ x. s) t > s ≥ 0. let φ() = u(x0 . Let p0 = Du(x0 .  . t > 0 and x ˆ. t0 )) = 0. Assume u is differentiable at a point (x0 . t).

1.4. Introduction to Hamilton-Jacobi equations

19

Then φ0 (0) = ut (x0 , t0 ) + q · p0 . On the other hand, by the lemma above
u(x0 , t0 ) ≤ L(q) + u(x0 − q, t0 − )
and hence

φ() 

≤ L(q). This shows that φ0 (0) ≤ L(q), or
q · p0 − L(q) ≤ −ut (x0 , t0 ) ∀ q ∈ Rn .

So, H(p0 ) ≤ −ut (x0 , t0 ) and this proves that ut + H(Du) ≤ 0 at (x0 , t0 ).
) + g(y). For
To prove the other direction, let y ∈ Rn be such that u(x0 , t0 ) = t0 L( x0t−y
0
x0 −y
z−y 

0 <  < t0 , let z = z be such that t0 = t0 − ; that is, z = y + (1 − t0 )(x0 − y). Then  

x0 − y
z − y
x0 − y
u(x0 , t0 ) − u(z , t0 − ) ≥ t0 L(
) + g(y) − (t0 − )L(
).
) + g(y) = L(
t0
t0 − 
t0
Define   

φ() = u(x0 , t0 ) − u(z , t0 − ) = u(x0 , t0 ) − u y + (1 − )(x0 − y), t0 −  .
t0
Then

φ() 

≥ L( x0t−y
) for all  > 0. So φ0 (0) ≥ L( x0t−y
0
0
x0 − y
x0 − y
· Du(x0 , t0 ) + ut (x0 , t0 ) ≥ L(
).
t0
t0

Hence

x0 − y
x0 − y
· Du(x0 , t0 ) − L(
) ≥ −ut (x0 , t0 ).
t0
t0
We have finally proved that ut + H(Du) = 0 at (x0 , t0 ).
H(Du(x0 , t0 )) ≥ 

Example 1.12. (a) Find the Hopf-Lax solution to the problem
1
ut + |Du|2 = 0, u(x, 0) = |x|.
2
Solution: In this case, H(p) = 12 |p|2 and hence L(q) = 12 |q|2 . (Exercise!) So  

|x − y|2
u(x, t) = minn
+ |y| .
y∈R
2t
2

Given x ∈ Rn , t > 0, the function f (y) = |x−y|
+ |y| has a critical point at y = 0 and
2t
y
possible critical points at y0 6= 0 if f 0 (y0 ) = 0. Solving f 0 (y) = 0 leads to x = y + |y|
t, which
has no nonzero solution y if x = 0; in this case, u(0, t) = f (0) = 0. If x 6= 0 then f 0 (y) = 0
has nonzero solution
x
y0 = (|x| − t) .
|x|
Comparing f (y0 ) and f (0), we have u(x, t) = f (y0 ) if |x| > t and u(x, t) = f (0) if |x| ≤ t.
Hence
(
|x| − t if |x| ≥ t,
u(x, t) = |x|2 2
if |x| ≤ t.
2t
(b) Find the Hopf-Lax solution to the problem
1
ut + |Du|2 = 0, u(x, 0) = −|x|.
2
Solution: In this case,  

|x − y|2
u(x, t) = minn
− |y| .
y∈R
2t

1.4. Introduction to Hamilton-Jacobi equations

20

2

Given x ∈ Rn , t > 0, the function f (y) = |x−y|
− |y| has a critical point at y = 0 and
2t
y
t, which
possible critical points at y0 6= 0 if f 0 (y0 ) = 0. Solving f 0 (y) = 0 leads to x = y − |y|
0
has nonzero solutions y0 with |y0 | = t if x = 0. If x 6= 0 then f (y) = 0 has nonzero solution
x
y0 = (|x| + t) .
|x|
Comparing f (y0 ) and f (0) in all cases, we have u(x, t) = f (y0 ) for all x, t. Hence
t
u(x, t) = −|x| −
(x ∈ Rn , t > 0).
2
Example 1.13. (Problem 6 in the Text.)
Given a smooth vector field b(x) = (b1 (x), · · · , bn (x)) on Rn , let x(s) = X(s, x, t) solve
(
x˙ = b(x) (s ∈ R),
x(t) = x.
(a) Define the Jacobian
J(s, x, t) := det Dx X(s, x, t).
Show the Euler formula:
Js = (div b)(X)J.
(b) Demonstrate that
u(x, t) = g(X(0, x, t))J(0, x, t)
solves

(
ut + div(ub) = 0 in Rn × (0, ∞),
u=g
on Rn × {t = 0}.

Proof. (a) For any smooth matrix function M (s) of s ∈ R, it follows that
d
(det M (s)) = tr((adjM (s))M˙ (s)),
ds
where for any n × n matrix M , adjM is the adjugate matrix of M , satisfying
(1.25)

(Exercise: Prove (1.25)!)
Let M (s) = (Xxi j (s, x, t)) and B(x) = (bixj (x)). Then, by the equation,
M˙ (s) = B(X)M (s)

s ∈ R.

Therefore
= tr(B(X)M (s)(adjM (s))) = tr(B(X)J) = (div b)(X)J.
(b) The uniqueness of X(s, x, t) implies that
X(s, x, t) = X(s − t, x, 0) := X0 (s − t, x)

(x ∈ Rn , s, t ∈ R).

So X0 (s, x) satisfies the group property:
X0 (0, x) = x,

X0 (s, X0 (t, x)) = X0 (s + t, x).

Hence J(s, x, t) = J(s − t, x, 0) := J0 (s − t, x). By the Euler formula, J0 is given by
Rs

J0 (y, s) = e

0

(div b)(X0 (τ,y))dτ

.

Suggested exercises

21

From the group property of X0 , it follows that J0 (t + s, x) = J0 (s, X0 (t, x))J0 (t, x), which
implies
J(s, X0 (−s, x))J0 (−s, x) = J0 (0, x) = 1

(x ∈ Rn , s ∈ R).

Note that
u(x, t) = g(X0 (−t, x))J0 (−t, x).
Therefore, u(x, 0) = g(x) (x ∈ Rn ). We want to show that the classical solution built from
the characteristics method agrees with this function u(x, t). The characteristic ODEs are
x˙ = b(x),

t˙ = 1,

z˙ = −(div b)(x)z.

The initial data are given by
x(0) = y,

t(0) = 0,

z(0) = g(y)

(y ∈ Rn ).

Hence
x = x(y, s) = X0 (y, s),

t = t(y, s) = s

and
z = z(y, s) = g(y)e−

Rs
0

(div b)(X0 (τ,y))dτ

=

g(y)
.
J0 (y, s)

Solving (y, s) in terms of x = x(y, s) and t = t(y, s), we have
s = t,

y = X0 (−t, x).

Hence our classical solution w(x, t) from the characteristics method is given by
w(x, t) = z(y, s) =

g(X0 (−t, x))
= g(X0 (−t, x))J0 (−t, x),
J0 (t, X0 (−t, x))

which is exactly the function u(x, t). 

Suggested exercises
Materials covered are from Chapter 3 of the textbook. So complete the arguments that are
left in lectures. Also try working on the following problems related to the covered materials.
Chapter 3: Problems 4, 5, 6, 10, 11, 13, 16.
Homework # 1.
(1) (7 points) Write down an explicit formula for a function u solving the initial value
problem
(
ut + b · Du + cu = 0 in Rn × (0, ∞),
u=g
on Rn × {t = 0}.
Here c ∈ R, b ∈ Rn are constants, and g is a given function.

bn ) ∈ Rn .) Homework # 2. (c) Suppose u is a C 2 solution in each of two domains separated by a smooth curve x = ξ(y). b. |p|→∞ |p| lim min H(p) = H(0) = 0. (Hint: You may use the conclusion of Problem (3)(b) above. (b) (5 points) n n X 1 1 X L(q) = q · Aq + b · q = aij qi qj + bi qi 2 2 i.. Consider the quasilinear PDE uy + a(u) ux = 0 with initial data u(x. where h is also a given smooth function. show that u ≡ 0 on Ω. unless a(h(z)) is a nondecreasing function of z ∈ R.e. i=1 where A = (aij ) is a n × n symmetric.j=1 (q ∈ Rn ). p∈Rn (a) (5 points) Show that min H ∗ (q) = H ∗ (0) = 0. (4) (10 points) Consider the equation uy = (ux )3 . · · · . (3) (15 points) Let a(z) be a given smooth function of z ∈ R. and b = (b1 . (a) Use the characteristics method to show that the solution is given implicitly by u = h(x − a(u)y). r where 1 < r < ∞. i. Assume u is continuous but ux has a jump discontinuity across the curve. (1) Find the Legendre transform of L for (a) (5 points) 1 L(q) = |q|r (q ∈ Rn ). positive definite n × n-matrix. 0) = h(x). (b) Show that every C ∞ solution to the equation on whole R2 must be of the form u(x. If B(x) · x > 0 for all x ∈ ∂Ω. and also assume u restricted to the curve is smooth. (2) Let H : Rn → R be a convex function satisfying H(p) = +∞. y) = ax + by + c for some constants a. dξ/dy = a(u) along the curve. (Hint: Show that maxΩ u ≤ 0 and minΩ u ≥ 0). c ∈ R. q∈Rn . 0) = 2x3/2 . Prove the curve x = ξ(y) must be a characteristics. (a) Find the solution with initial data u(x.Suggested exercises 22 (2) (8 points) Let u be a C 1 solution of linear PDE B(x) · Du(x) = −u on the closed unit ball Ω in Rn . (b) Show that the solution becomes singular (undefined) for some y > 0.

∞) u(x. ui (x. for each t > 0. (i = 1. 2). t) − u2 (x. (3) (5 points) Let H(p) be a convex function satisfying the superlinear growth at ∞. t) has compact support in x ∈ Rn . Prove the L∞ -contraction inequality: sup |u1 (x. g2 be given Lipschitz functions. u2 are the Hopf-Lax solutions of the initial value problem ( uit + H(Dui ) = 0 in Rn × (0. and let g1 . Assume u1 . show that. u(x. t) is the Hopf-Lax solution to the Hamilton-Jacobi problem ( ut + H(Du) = 0 in Rn × (0. 0) = g(x). . If u(x. t)| ≤ sup |g 1 (x) − g 2 (x)| x∈Rn x∈Rn ∀ t > 0. 0) = g i (x). ∞).Suggested exercises 23 (b) (10 points) Let g be Lipschitz continuous and have compact support in Rn .

. The Laplace operator on Rn is defined by (2. k=1 A C 2 function u satisfying ∆u = 0 on a domain Ω ⊆ Rn is called a harmonic function on Ω. .5) is called Green’s second identity.3) we obtain Z Z Z (2. ∂ν ∂u ∂v − u ) dS.5) (v∆u − u∆v)dx = Ω ∂Ω (v u ∂v dS. f n ).3) is called Green’s first identity and Equation (2.4) yields Z Z (2. f n ) on Ω Z Z (2. ∂ν ∂ν Equation (2.Chapter 2 Laplace’s Equation 2.3)-(2. . ¯ let F~ = vDu = v∇u. Ω ∂Ω where ν is the outer unit normal to the boundary ∂Ω. . . ¯ states that The divergence theorem for C 1 -vector fields F~ = (f 1 . hence For any u. j=1 and for a smooth function u(x) we define the gradient of u by Du = ∇u = (ux1 .2) div F~ dx = F~ · ν dS.1.1) ∆u = n X uxk xk = div(∇u).4) u∆vdx = − ∇u · ∇vdx + Ω Ω ∂Ω Combining (2. Green’s identities For a smooth vector field F~ = (f 1 . Ω Ω ∂Ω Ω ∂Ω ∂ν Exchanging u and v in (2. we have div F~ = ∇u · ∇v + v∆u. 24 . we define the divergence of F~ by div F~ = n X ∂f j j=1 ∂xj = n X Dj f j = j=1 n X fxjj . v ∈ C 2 (Ω). uxn ) = div(∇u). · · · . . .3) v∆udx = − ∇u · ∇vdx + v∇u · νdS = − ∇u · ∇vdx + v dS. . Z Z Z Z Z ∂u (2.

n = 1. if Ω is connected. consequently. any two C 2 -solutions of the Neumann problem differ by a constant. v(r) = r is a solution. by (2.6) u∆udx = − |∇u|2 dx + Ω Ω ∂Ω u ∂u dS. ¯ be solutions to the Dirichlet problem above and let u = u1 − u2 . ∇u = 0. the constant must be zero. u2 ∈ C 2 (Ω) Then ∆u = 0 in Ω and ∂u ∂ν = 0 on ∂Ω. for such domains. Let u1 . Note that v(r) is well-defined for r > 0. r For n = 1. but is singular at r = 0 when n ≥ 2. ∂ν which implies that if there is a C 2 (Ω) solution u to the Neumann (boundary) problem  then R ∂Ω g(x)dS ∆u = 0 in Ω. This is a first-order linear ODE for s(r) and solving it yields that s(r) = cr1−n . v(r) = C ln r.. let s(r) = v 0 (r). r = |x| (radial function). Fundamental solutions and Green’s function We try to seek a harmonic function u(x) that depends only on the radius r = |x|. choose u = v. Therefore. ∂ν ¯ be solutions to the Neumann problem above and let u = u1 − u2 .3). ∇u = 0. u(x) = v(r). Therefore. 2.2. we have s0 (r) = − n−1 r s(r). For n ≥ 2. we have Z Z ∆udx = ∂Ω Ω 25 ∂u dS.6).2. however.3). . Fundamental solutions and Green’s function Taking v = 1 in (2. the C 2 (Ω)-solution of the Dirichlet problem is unique. ∂u ∂ν |∂Ω = g(x). hence. The Dirichlet (boundary) problem for Laplace’s equation is:  ∆u = 0 in Ω.6). hence.e. consequently u ≡ constant on each connected component of Ω. Computing ∆u for such a function leads to an ODE for v: n−1 0 ∆u = v 00 (r) + v (r) = 0. consequently u ≡ constant. = 0. But later we will prove such a uniqueness result for ¯ solutions in C 2 (Ω) ∩ C(Ω).   2−n Cr . and we have Z Z Z (2. n = 2. by (2. In (2. Let u1 . u|∂Ω = f. ¯ So u1 = u2 on Ω (we don’t need Ω is connected in this case). we obtain a solution for v(r):   Cr. since u = 0 on ∂Ω. u2 ∈ C 2 (Ω) Then ∆u = 0 in Ω and u = 0 on ∂Ω.2. i. n ≥ 3.

) Z (n = 2) (n = 3).1. Z ∆u(x) = Since B is arbitrary.R) 2. ) to have Z J = Φ(y)∆y f (x − y) dy Ω  Z Z  ∂f (x − y) ∂Φ(y) = f (x − y)∆Φ(y) dy + Φ(y) − f (x − y) dSy ∂νy ∂νy Ω ∂Ω  Z  ∂f (x − y) ∂Φ(y) = Φ(y) − f (x − y) dSy ∂νy ∂νy ∂Ω   Z ∂f (x − y) ∂Φ(y) = Φ(y) − f (x − y) dSy .7) φ(r) = − 2π ln r. Write Z Z ∆u(x) = Φ(y)∆y f (x − y) dy + B(0. Fundamental solutions. Moreover Z Φ(y)∆x f (x − y) dy = Φ(y)∆y f (x − y) dy. αn is the volume of the unit ball in Rn . 1. where  1 n = 1.1. For J we use Green’s second identity (2.R)\B(0.1.  − 2 r.2. We only prove this for n ≥ 2 and leave the case of n = 1 (where Φ(x) = − 12 |x|) as an exercise. Proof. Fix 0 <  < R. Definition 2.   1 2−n . Let f ∈ Cc2 (Rn ). it proves u ∈ C 2 (Rn ). R/2). Rn Then u ∈ C 2 (Rn ) and solves the Poisson’s equation −∆u(x) = f (x) (2. Therefore. We call function Φ(x) = φ(|x|) a fundamental solution of Laplace’s equation in Rn . Theorem 2.R) C 2 (B).) Now 2 |I | ≤ CkD f kL∞ ( C2 | ln | |Φ(y)| dy ≤ C2 B(0. B(0.2.2. For any f ∈ Cc2 (Rn ). Then Z Z Z u(x) = Φ(x − y)f (y) dy = Φ(y)f (x − y) dy = Φ(y)f (x − y) dy.R) This proves that u ∈ B(0. 1 n = 2. B(0.) Φ(y)∆y f (x − y) dy =: I + J . B(0. define Z u(x) = Φ(x − y)f (y) dy. (2.R) B(0. for n ≥ 3. n ≥ 3.5) with Ω = B(0. Hence I → 0 as  → 0+ . n(n−2)αn r Here. Fix any bounded ball B ⊂ Rn and x ∈ B.R) Rn where B(0. R) \ B(0.) . ∂νy ∂νy ∂B(0. by differentiation under the integral. R) is a large ball in such that f (x − y) = 0 for all x ∈ B and y ∈ / B(0.8) for all x ∈ Rn . uxi xj (x) = Φ(y)fxi xj (x − y) dy. Rn Rn B(0. Z Z uxi (x) = Φ(y)fxi (x − y) dy. Fundamental solutions and Green’s function 26 2.

2.2. ). Fundamental solutions and Green’s function 27 where νy = − y is the outer unit normal of ∂Ω on the sphere ∂B(0. Now .

.

Z Z .

∂f (x − y) .

.

.

∞ dS ≤ Cn−1 |φ()| → 0 Φ(y) .

≤ C|φ()|kDf kL .

.

.

) as  → 0+ . ∂B(0. Φ(x−y)f (y) dy. ∂Ω. Apply Green’s second identity   Z Z ∂v ∂u (u(y)∆v(y) − v(y)∆u(y)) dy = u(y) (y) − v(y) (y) dS ∂ν ∂ν Ω ∂Ω to functions u(y) and v(y) = Γ(x. Suppose Ω ⊂ Rn is a bounded domain with smooth boundary ¯ be any harmonic function in Ω. Green’s function. nαn  That is.) Z Z ∂Φ(y) dSy = − f (x − y) dSy → f (x).) Z ∂u − (Φ(y − x) − h(y)) (y)dS. y)dS − Γ(x. Furthermore. we finally prove that −∆u(x) = f (x) ∀ x ∈ Rn .) y as  → 0+ . Given any function u ∈ C 2 (Ω).9) ∂h ∂Φ + u(y) (y − x) − (y) dS ∂ν ∂ν ∂B(x. we have Z Z Z ∂Γ ∂u − (x. we can formally compute that (in terms of distributions) Z Z −∆u(x) = ∆x Φ(x − y)f (y) dy = − ∆y Φ(x − y)f (y) dy Rn Rn Z =− ∆y Φ(y)f (x − y) dy = hδ0 . Combining all the above. ). = R 1 dS and hence ∂B(0. f (x − y) ∂νy ∂B(0. hence ( 1 −1  (n = 2) ∂Φ(y) = ∇Φ(y) · νy = −φ0 () = 2π1 1−n ∂νy (n ≥ 3). ). where Φ(y) = φ(|y|) is the fundamental solution above. f i = f (0) If u(x) = R Rn ∀ f ∈ Cc∞ (Rn ).) ∂B(0. where δ0 is the Dirac measure concentrated at 0: hδ0 . Let Ω = Ω \ B(x. y)∆u(y) dy = u(y) ∂νy ∂νy Ω ∂Ω ∂Ω Z Z ∂Γ ∂u = u(y) (x.) ∂νy ∂B(0. ∂Ω). f (x − ·)i = f (x) Rn 2. y)dS − Γ(x. ∇Φ(y) = φ0 (|y|) |y| . The function Φ(x) formally satisfies −∆x Φ(x) = δ0 on x ∈ Rn .2. ∂ν ∂B(x.  The reason the function Φ(x) is called a fundamental solution of Laplace’s equation is as follows. y) (y)dS ∂ν ∂ν y y ∂Ω ∂Ω   Z (2. ∂Φ(y) ∂νy for all y ∈ ∂B(0.) . and since ∆v(y) = 0 on Ω . then. y) = Φ(y − x) − h(y) on Ω . y) (y)dS Γ(x.2. Let h ∈ C 2 (Ω) ¯ fix x ∈ Ω and 0 <  < dist(x.

However.9). However.2. ∂νy ∂νy ∂Ω Ω This formula permits us to solve for u if we know the values of ∆u in Ω and both u and on ∂Ω. in (2.  Z  Z ∂u ∂Γ (2. Then G(x. ∂u/∂ν is not known (and cannot be prescribed arbitrarily). y)∆u(y) dy.10). x) − hx (y) ¯ x 6= y). general bounded domains Ω. y) = Φ(y.) Z Z ∂h = − u(y)dSy − u(y) (y)dSy → u(x) ∂ν y ∂B(x. ∂νy Ω ∂Ω The function ∂G (x. y) (y) − u(y) (x. for bounded domains Ω with smooth boundary. y)∆u(y) dy (x ∈ Ω). ∂νy we deduce ¯ Theorem 2.)  → 0+ and noting that   ∂Φ ∂h u(y) (y − x) − (y) dSy ∂νy ∂νy ∂B(x.11) x h (y) = Φ(y − x) (y ∈ ∂Ω).) ∂u (y)dS → 0. We must modify this formula to remove the boundary integral term involving ∂u/∂ν. y ∈ ∂Ω) ∂νy is called Poisson’s kernel for domain Ω. Remark 2. Hence ∂νy (y − x) = −φ () = for y ∈ ∂B(x. (2. So. ∂u ∂ν Definition 2. y)dS − G(x. if exists for bounded domain Ω. ). y) = Φ(y − x) − h(y). ¯ is guaranteed by the general existence and regularity theory and existence of hx in C 2 (Ω) consequently for such domains the Green’s function always exists and is unique. Then. must be unique. . y ∈ Ω. y) = 0 for y ∈ ∂Ω and x ∈ Ω. the function Z K[g](x) = K(x. Fundamental solutions and Green’s function 28 where ν = νy is the outer unit normal at y ∈ ∂Ω = ∂Ω ∪ ∂B(x. ¯ solving Given x ∈ Ω. letting dS ∂B(x. ).) Z and Z (Φ(y − x) − h(y)) ∂B(x. We define Green’s function for domain Ω to be G(x. ). Z Z ∂G (2. for Poisson’s equation with Dirichlet boundary condition.) ∂B(x. A corrector function hx . we assume that there exists a corrector function h = hx ∈ C 2 (Ω) the special Dirichlet problem: ( ∆y hx (y) = 0 (y ∈ Ω). for all u ∈ C (Ω).2 (Representation formula).10) u(x) = Γ(x. we do not discuss these issues in this course. where h ∈ C 2 (Ω) 2 ¯ is harmonic in Ω. y)g(y)dSy (x ∈ Ω) K(x. Let Γ(x. hence. ¯ which may not be possible for Here we require the corrector function hx exist in C 2 (Ω). (x ∈ Ω. y) (x ∈ Ω. Given a function g on ∂Ω. y) = − ∂Ω is called the Poisson integral of g with kernel K.2. by (2. y) dS − Γ(x.2. Note that νy = − y−x  at ∂Φ 1 0 R y ∈ ∂B(x.3.12) u(x) = − u(y) (x.

We compute the limits of two terms on both sides of (2. (z ∈ U ).2. ). Proof.3 (Representation by Green’s function). then Z (2. u(x) = g(x) (x ∈ ∂Ω). Let v(z) = G(x. x) for all x.)     Z ∂v ∂v ∂w ∂w (z) − w(z) (z) dS = w(z) (z) − v(z) (z) dS. . w(z) = G(y. For sufficiently small  > 0. For the term on LHS. y ∈ Ω. z). we apply Green’s second identity on Ω = Ω \ (B(x. Fix x. ) ∪ ∂B(x.2.14) as  → 0+ . Fundamental solutions and Green’s function 29 ¯ solves the Dirichlet Theorem 2. ∂ν ∂ν ∂Ω This implies Z (2.13) u(x) = Z K(x. y ∈ Ω. x 6= y.4 (Symmetry of Green’s function). since w(z) is smooth near z = x. )) for functions v(z) and w(z) to obtain   Z ∂w ∂v v(z) (z) − w(z) (z) dS = 0. Ω Theorem 2. ) ∪ B(y. y)g(y)dS + ∂Ω G(x.) where ν denotes the inward unit normal vector on ∂B(x. x 6= y. If u ∈ C 2 (Ω) problem ( −∆u(x) = f (x) (x ∈ Ω). G(x.14) ∂B(x. y)f (y) dy (x ∈ Ω). z) Then ∆v(z) = 0 for z 6= x and ∆w(z) for z 6= y and v|∂Ω = w|∂Ω = 0. y) = G(y. v(z) ∂ν ∂ν ∂ν ∂ν ∂B(y.

Z .

.

.

∂w .

.

v(z) (z)dS .

.

≤ Cn−1 sup |v(z)| = o(1). .

) . ∂B(x.

y) is harmonic in x ∈ Ω for all y ∈ ∂Ω.14) = −w(x). y) < 0 for all x ∈ Ω and (2) Since G(x. y) = 0 for y ∈ ∂Ω. (Homework!) Since G(x. x) = G(x. y) = − ∂ν (x. v(z) = Φ(x − z) − hx (z) = Φ(z − x) − hx (z). hence Poisson’s kernel K(x. y).14) = −v(y). (1) Strong Maximum Principle below implies that G(x.) So lim LHS of (2. In particular. y . y) > 0 for all x. where the corrector hx is smooth in Ω. y) ≤ 0. where y νy is outer unit normal of Ω at y ∈ ∂Ω. x 6= y. y ∈ ∂G Ω. we know that G(x. y) is also harmonic in x ∈ Ω \ {y}. by the symmetry property. it follows that ∂ν (x. which is exactly G(y. y) is harmonic in x ∈ Ω for all ∂G y ∈ ∂Ω. we have y ∈ ∂Ω. →0+ proving w(x) = v(y).4.) ∂G ∂νy (x.) ∂B(x.  Remark 2. (In fact.) Also. G(x. Hence Z Z ∂v ∂Φ w(z) (z − x)dS = w(x). →0+ Likewise lim RHS of (2. y) is harmonic in y ∈ Ω \ {x}. ∂ν z∈∂B(x. lim w(z) (z)dS = lim + + ∂ν ∂ν →0 →0 ∂B(x.

for all x0 ∈ ∂Ω and δ > 0 Z (2. Let M = kgkL∞ .2.5 (Poisson integrals as solutions). y)dSy = 1 (x ∈ Ω). for each g ∈ C(∂Ω) its Poisson integral u(x) = K[g](x) is harmonic in Ω and satisfies lim x→x0 .2. Z K(x.δ) Then. For ε > 0. x∈Ω u(x) = g(x0 ) (x0 ∈ ∂Ω). y)dSy = 0. by Green’s representation theorem (Theorem 2. |y − x0 | < δ.15) lim K(x. ¯ That is. g = 1 and u ≡ 1. Proof. u=g on ∂Ω. u = K[g] can be extended to be a C 2 (Ω) ∩ C(Ω)-solution to the Dirichlet problem ( ∆u = 0 in Ω. y) ≥ 0 for all x ∈ Ω and y ∈ ∂Ω and that. Then . Fundamental solutions and Green’s function 30 (3) We always have that K(x.3). let δ > 0 be such that |g(y) − g(x0 )| <  ∀ y ∈ ∂Ω. x→x0 . with f = 0. x∈Ω ∂Ω\B(x0 . ∂Ω Theorem 2. Assume.

Z .

Z .

.

0 .

|u(x) − g(x )| = .

K(x. y)(g(y) − g(x )) dSy .

.

y)|g(y) − g(x0 )| dSy 0 B(x0 . Green’s functions for a half-space and for a ball.δ)  2.δ)∩∂Ω K(x. x∈Ω Z K(x. the representation formula may not be valid for such domains or valid only in certain extended sense. y)dSy + 2M B(x0 . For such a function to be the corrector function hx (y) we need Φ(y − x) = Φ(b(x)(y − a(x)) (y ∈ ∂Ω). however. ≤ K(x. y) dSy = ε. y) dSy ∂Ω\B(x0 . y) dSy .15).δ) Hence.δ) Z ≤ ε + 2M K(x.2. y ∈ ∂Ω). Then function Φ(b(x)(y − a(x)) is harmonic in Ω. y)|g(y) − g(x0 )| dSy ∂Ω ∂Ω Z Z ≤ K(x. it can be similarly defined for unbounded domains or domains with nonsmooth boundaries.δ)∩∂Ω ∂Ω\B(x0 . Such a construction is possible when Ω is a half-space or a ball.3. (Although a Green’s function is defined above for a bounded domain with smooth boundary.) .δ) Z Z ≤ε K(x. y)|g(y) − g(x0 )| dSy + K(x. lim sup |u(x) − g(x0 )| ≤ ε + 2M lim sup x→x0 . Fix x ∈ Ω and let b(x) ∈ R and a(x) ∈ Rn \ Ω. by assumption (2. ∂Ω\B(x0 . Green’s functions for certain special domains can be explicitly found from the fundamental solution Φ(x). x→x0 . which requires that |b(x)||y − a(x)| = |y − x| (x ∈ Ω. x∈Ω ∂Ω\B(x0 .

y) = 2xn 2xn = := H(x. Then (i) u ∈ C ∞ (Rn+ ) ∩ L∞ (Rn+ ). Fundamental solutions and Green’s function 31 Case 1. y 0 ) is harmonic in x ∈ Rn+ for each y 0 ∈ Rn−1 and (very complicated) Z H(x. we can easily see that hx (y) = Φ(y − x ˆ) solves (2. · · · . Rn−1 .2. y) = (x. · · · . Green’s function for a half-space. n ˆ = x (x ∈ R ). Theorem 2. the proof is similar. · · · . y ∈ ∂Rn+ ). Proof. lim x→x0 . y) = (y − x) − (y − x ˆ) ∂yn ∂yn ∂yn   −1 yn − xn yn + xn = − . where we assumed the domain Ω is bounded. 1. xn ) ∈ Rn | xn > 0}. Let Ω = Rn+ = {x = (x1 . −xn ). x2 . y 0 )dy 0 = 1 (x ∈ Rn+ ). y 0 )g(y 0 ) dy 0 = nαn Rn−1 (|x0 − y 0 |2 + x2n )n/2 Rn−1 (x ∈ Rn+ ). This theorem differs from the general theorem (Theorem 2. nαn |y − x|n |y − x ˆ|n The corresponding Poisson’s kernel of Rn+ is given by K(x. x2 .2. its reflection with respect to the hyper-plane xn = 0 is the point x ˆ = (x1 . n 0 nαn |x − y| nαn (|x − y 0 |2 + x2n )n/2 Hence the Poisson integral u = K[g] can be written as Z Z 2xn g(y 0 ) dy 0 (2. x Clearly x ˆ ˆ = x (x ∈ ∂Rn ) and Φ(x) = Φ(ˆ x) (x ∈ Rn ). + In this case. If x = (x1 . we give a full proof below just to emphasize the main idea.6 (Poisson’s formula for half-space). xn ) ∈ Rn . xn−1 . y) = − ∂G ∂G 2xn (x. Assume g ∈ C(Rn−1 ) ∩ L∞ (Rn−1 ) and define u by (2. If we write y = (y 0 . 0) with y 0 ∈ Rn−1 .16). y) = Φ(y − x) − Φ(y − x ˆ) (x. It can be verified that H(x. So a Green’s function for Rn+ is given by G(x. y ∈ Rn+ ).11).16) u(x) = H(x. then K(x. x2 . This formula is called Poisson’s formula for Rn+ . We show that this formula provides a solution to the Dirichlet problem. y 0 ). However. (iii) for all x0 ∈ ∂Rn+ .5). (ii) ∆u(x) = 0 in Rn+ . y) = ∂ν ∂yn nαn |x − y|n (x ∈ Rn+ . x∈Rn + u(x) = g(x0 ). Then ∂G ∂Φ ∂Φ (x.

Fundamental solutions and Green’s function 32 2. This also shows that Z ∆x H(x. we have Z α D u(x) = Dxα H(x. Since g is bounded. y 0 )g(y 0 )dy 0 = 0 (x ∈ Rn+ ).  > 0. Since H(x. y 0 ) is smooth in x ∈ Rn+ .2.2. y)g(y 0 )dy 0 Rn−1 is continuous in x ∈ Rn+ for all multi-indexes. Then if |x − x0 | < δ/2 and x ∈ Rn+ . . Now fix x0 ∈ ∂Rn+ . u defined above is also bounded. Choose δ > 0 so small that |g(y 0 ) − g(x0 )| <  if |y 0 − x0 | < δ. ∆u(x) = Rn−1 3.

Z

|u(x) − g(x )| =

.

.

H(x. y )(g(y ) − g(x )) dy .

n−1 0 0 0 0.

Combining these estimates.δ) Rn−1 \B(x0 .  .δ) = I + J.δ) 2n+2 kgkL∞ xn ≤ nαn 2n+2 kgkL∞ xn = nαn Z |y 0 − x0 |−n dy 0 Rn−1 \B(x0 . we deduce |u(x) − g(x0 )| ≤ 2 provided that |x − x0 | is sufficiently small. y 0 )dy 0 J ≤ 2kgkL∞ Rn−1 \B(x0 . nαn δ as xn → 0+ if x → x0 . 0 R Z ≤ 0 0 0 0 Z H(x. Thus Z H(x. Now that Z H(x. we have 1 |y 0 − x0 | ≤ |y 0 − x| + δ/2 ≤ |y − x| + |y 0 − x0 | 2 and so |y 0 − x| ≥ 12 |y 0 − x0 |. y )|g(y ) − g(x )|dy + Rn−1 ∩B(x0 . y 0 )dy 0 = .δ)  Z (n − 1)αn−1 = 2n+2 kgkL∞ ∞ r −n n−2 r  dr δ (n − 1)αn−1 xn → 0. I≤ Rn−1 Furthermore. y 0 )|g(y 0 ) − g(x0 )|dy 0 H(x. if |x − x0 | ≤ δ/2 and |y 0 − x0 | ≥ δ.

if x 6= 0. .5. y) = G(y. the corrector function hx is given by ¯ 1)). x 6= y. Given g ∈ C(∂B(0. but it is found from the symmetry of G: G(0. 1) = {x ∈ Rn | |x| < 1}. we deduce. 1). The Green’s function for ball B(0. 2y · x 1 |y − x|2 + = . |x|2 |x|2 |x|2 So b(x) = |x| will do the job. 1). y)yi = − ∂νy nαn |y − x|n |y − x|n i=1 i=1 |x|2 1 −1 (x ∈ B(0. y) cannot be given by first formula since ˜0 is undefined. We need to have |b(x)||y − x ˜| = |y − x| If y ∈ ∂B(0.2. Let Ω = B(0.1) |y − x| (x ∈ B(0. y) = Gyi (x. Green’s function for a ball. the point x |x|2 is called the inversion point to x with respect to the unit sphere ∂B(0. x 6= 0. y) = Φyi (y − x) − Φyi (|x|(y − x ˜))|x|   2 x i − yi |x| ((˜ x)i − yi 1 − = nαn |y − x|n (|x||y − x ˜|)n   1 xi − yi xi − |x|2 yi = − . y = 6 0). 1). Fundamental solutions and Green’s function 33 Case 2. y) = G(y.) yi Since Φyi (y) = φ0 (|y|) |y| = −yi nαn |y|n (y 6= 0 and n ≥ 2). nαn |y − x|n The same formula is also valid if x = 0 and y ∈ ∂B(0. y) dSy = n nαn ∂B(0. y ∈ ∂B(0.17) u(x) = g(y)K(x. if y ∈ ∂B(0. x 6= y). y) = ∂νy nαn |y − x|n (x ∈ B(0. The mapping x 7→ x ˜ is called inversion with respect to ∂B(0. y) = − ∂G 1 1 − |x|2 (x. 1) is given by ( Φ(y − x) − Φ(|x|(y − x ˜)) (x 6= 0. the Poisson’s kernel for B(0. nαn |y − x|n (|x||y − x ˜|)n So. 1).2. G(x. x ˜= Given x ∈ B(0. Gyi (x. since |x||y − x ˜| = |y − x| and νy = y. Consequently. 1) is given by K(x. we try to find the corrector function hx in the form of hx (y) = Φ(b(x)(y − x ˜)). its Poisson integral u = K[g] is given by Z Z g(y) dSy 1 − |x|2 (2. hx (y) = Φ(|x|(y − x ˜)) (y ∈ B(0. If x ∈ Rn \ {0}. 1)).1) ∂B(0. Definition 2. 1). = Therefore. 1)). 1)). 1) \ {0}). we have  n n  X ∂G 1 X xi yi − yi2 xi yi − |x|2 yi2 (x. for x 6= 0. (Note that G(0. 0) for y 6= 0. 1) then |y| = 1 and |y − x ˜|2 = 1 − 2y · x ˜ + |˜ x|2 = 1 − (y ∈ ∂B(0. 1)). 1). 0) = Φ(−y) − Φ(−|y|˜ y ) = Φ(y) − φ(1) (x = 0.

17). we deduce 2 Z 1 − | x−a g(a + ry) dSy x−a r | u(x) = u ˜( )= x−a n r nαn ∂B(0. Suppose u is a C 2 solution to the following Dirichlet problem on a closed ball B(a. changing z back to y. r).20) u(a) = − u(y)dSy . ∂B(a. u=g on ∂B(a. −n |z − x|n nαn r2 ∂B(a. r) = (x ∈ B(a.1) |y − r | Z g(z)r1−n dSz r2 − |x − a|2 = (z = a + ry). ¯ r). u(a) = nαn ∂B(a. = ∂B(a. a. K(x. (2.r) r Hence. if u is harmonic in Ω and B(a. then Z (2.r) |y|n nαn rn−1 ∂B(a. r)). y. (ii) ∆u = 0 in B(a. r). r)) The formula (2.r) ∂B(a. In this way. r).2.19) indeed gives a smooth solution in B(a.7 (Poisson’s formula for ball). a. r)).2. ( ∆u = 0 in B(0. This formula has a special consequence that at the center x = a it implies Z Z Z g(y) r 1 dSy = g(y) dSy = − g(y) dSy . r) ⊂⊂ Ω (this means closed ball B(a. 1). we show that Poisson’s formula (2. 1). . To conclude. Then function u ˜(x) = u(a + rx) solves (2.r) |y − x| Z g(y)K(x.r) This is the mean-value property for harmonic functions that we will give another proof of and study further in the section. (2.19) Z g(y) dSy r2 − |x − a|2 u(x) = n nαn r ∂B(a. r). r). r) with boundary condition g.18) u=g on ∂B(0. r) dSy (x ∈ B(a. r)) and define u by (2. the C 2 (B(0.19). y.r) where 1 r2 − |x − a|2 nαn r |y − x|n is the Poisson’s kernel for ball B(a. Fundamental solutions and Green’s function 34 ¯ 1))-solution u of the Dirichlet problem By Green’s representation formula. Theorem 2.18) with g˜(x) = g(a + rx) replacing g. n where a ∈ R is the center and r > 0 is the radius. is given by the formula (2. y ∈ ∂B(a. r). Assume g ∈ C(∂B(a. ( ∆u = 0 in B(a.19) is called Poisson’s formula on ball B(a. Therefore.r) ¯ r) ⊂ Ω). Then (i) u ∈ C ∞ (B(a.

15).  > 0. Proof. 1). 1). Then if |x − x0 | < δ/2 and x ∈ B(0. .5 above) by verifying condition (2. Fix x0 ∈ ∂B(0. Choose δ > 0 so small that |g(y) − g(x0 )| <  if y ∈ ∂B(0. lim x→x0 .3. We prove the boundary condition (iii) only. Mean-value property 35 (iii) for each x0 ∈ ∂B(a.r) u(x) = g(x0 ). x∈B(a. r).2. This theorem follows from the general theorem (Theorem 2. 1) and |y − x0 | < δ.

.

Z .

.

.

.

0 0 K(x. y)(g(y) − g(x )) dSy .

|u(x) − g(x )| = .

.

.

we deduce |u(x) − g(x0 )| ≤ 2 provided that |x − x0 | is sufficiently small. Then Z Z u(x) = − u(y)dS = − u(y) dy ∂B(x.1) u(x + ρz)dSz . We now give another proof. we write V ⊂⊂ Ω if the closure V¯ of V is a compact subset of Ω.δ) 2kgkL∞ (1 − |x|2 ) n −n 2 δ nαn = Cδ −n (1 − |x|2 ).r) for each ball B(x.8 (Mean-value property for harmonic functions). y)dSy = . Now that Z I≤ Furthermore.1))\B(x0 .  ≤ 2. r) ⊂⊂ Ω. Combining these estimates. Mean-value property For two sets V and Ω. ∂B(0. Let u ∈ C 2 (Ω) be harmonic.3.20).r) B(x.δ) = I + J.δ)∩∂B(0. nαn which goes to 0 as x → x0 . r) ⊂⊂ Ω. Thus Z J ≤ 2kgkL∞ K(x. y)|g(y) − g(x0 )|dSy + (∂B(0.1) Z K(x.1))\B(x0 . ∂B(0.1) Z K(x. The first equality (the spherical mean-value property) has been proved above by Poisson’s formula in (2. r). Let B(x. since |x| → |x0 | = 1. we have |y − x| ≥ |y − x0 | − |x − x0 | ≥ δ/2. y)dSy (∂B(0.1) |y − x0 | ≥ δ. . Theorem 2. if |x − x0 | ≤ δ/2 and K(x. y)|g(y) − g(x0 )|dSy ≤ B(x0 . For any ρ ∈ (0.ρ) u(y)dSy = − ∂B(0. Proof. let Z Z h(ρ) = − ∂B(x.

r).22) holds for all balls B(x.r) 0 ∂B(x. Z Z u(y) dy = B(x. where 0 < rx ≤ dist(x. . r) ⊂⊂ Ω with 0 < r < rx .3.ρ) ∂νy ∂B(x.ρ) For example.3) with v = 1. Hence Z + h(r) = h(0 ) = lim − u(x + ρz)dSz = u(x). Theorem 2.ρ) 0 Z ! Z r |∂B(x. ρ)| dρ = u(x)|B(x. ρ→0+ ∂B(0. 1)| = n|B(0.r) So.ρ) ∂B(x. Differentiating with respect to r yields (2. 1)|. r) with 0 < r < rx . Then u is harmonic in Ω. ∂Ω) is a number depending x. where 0 < rx ≤ dist(x. r) with 0 < r < rx . Mean-value property 36 Then. if and only if the spherical mean-value property (2. Let u ∈ C 2 (Ω) satisfy Z u(x) = − u(y)dS ∂B(x.r) r u(y)dSy dρ ∂B(x. r) with 0 < r < rx .) This proves that h is constant on (0.ρ) y−x dSy ρ Z ∇u(y) · νy dSy = − Z ρ ∂u(y) dSy = − ∆u(y) dy = 0.21) u(y)dSy = u(x).6.r)  Remark 2.2. ρ)| dρ 0 ∂B(x. using Green’s first identity (2.22) u(x) = − u(y) dy. B(x.21) holds for all spheres ∂B(x. by the polar coordinate formula. This follows from the polar coordinate formula as above: ! Z Z r Z u(y) dy = u(y)dSy dρ.r) for all ∂B(x.ρ) (Note that |∂B(0.ρ) 0 for all 0 < r < rx .1) Hence Z − (2. The ball mean-value property (2. r)|. ∂B(x. Z Z 0 h (ρ) = − ∇u(x + ρz) · zdSz = − ∂B(0.22) holds for all balls B(x. if (2. = u(x) 0 This proves the ball mean-value property Z (2. ∂Ω) is a number depending x.21) for all spheres ∂B(x. n B(x.9 (Converse to mean-value property). r) ⊂⊂ Ω with 0 < r < rx then ! Z r Z Z r u(y)dSy dρ = u(x) |∂B(x. B(x.1) Z = − ∇u(y) · ∂B(x.

r) with 0 < r < rx0 such that ∆u(y) > 0 on B(x0 . Indeed. Theorem 2. This proves that u is in C ∞ (Ω) and. n B(x0 . Then u ∈ C ∞ (Ω ). see Appendix C. r).r) a contradiction. Therefore. Then u ∈ C ∞ (Ω) is harmonic in Ω. ∂B(x0 . ¯ Ω ∂Ω .) Let x ∈ Ω .10 (Mean-value property and C ∞ -regularity).5. Maximum principles 37 Proof.11 (Maximum principle). ∂Ω).  The following result shows that if u is only continuous and satisfies mean-value property then u is in fact C ∞ and thus harmonic. Let Ω be a bounded open set in Rn . where η (x) = −n η(|x|/) is the standard radial mollifier. say.20 and 2. let Ω = {x ∈ Ω | dist(x. Assume u ∈ ¯ is a harmonic function in Ω. rx0 ). Let u ∈ C(Ω) satisfy Z u(x) = − u(y)dS ∂B(x. by the computation as above Z r 0 h (r) = − ∆u(y) dy > 0. If ∆u(x0 ) 6= 0. r) with 0 < r < dist(x. however.10 still holds if the mean-value property is satisfied for all sufficiently small spheres ∂B(x. Z Z −n η(|x − y|/)u(y) dy u (x) = η (x − y)u(y) dy =  B(x. (For more on smooth mollification. Proof. Theorem 2. every harmonic function on Ω is in C ∞ (Ω).) Ω −n Z =  ! Z u(y)dSy η(r/) 0 −n dr =  Z u(x) ∂B(x.21 below. B(0. ∂Ω) > }.4. then B(x. u is harmonic in Ω.ρ) The assumption says that h(ρ) is constant on ρ ∈ (0.)  Remark 2. ) ⊂⊂ Ω. Consider function Z h(ρ) = − u(y)dS (0 < ρ < rx0 ). Let u = u ∗ η be the smooth mollification of u. then there exists a ball B(x0 . 2.4. We show u (x) = u(x). Maximum principles Theorem 2.r)  η(r/)nαn rn−1 dr 0 Z = u(x) η (y) dy = u(x). by the previous theorem. See Lemmas 2.r) for all spheres ∂B(x. C 2 (Ω) ∩ C(Ω) (1) Then the maximum principle holds: max u = max u. For each sufficiently small  > 0.7. hence u = u on Ω and so u ∈ C ∞ (Ω ).2. ∆u(x0 ) > 0. r).

B(x. then (x ∈ Ω). 2 ¯ Then u ∈ C (Ω) ∩ C(Ω) is harmonic in Ω and u = 0 on ∂Ω.r)) rn+k for each ball B(x.r) B(x.2.13 (Local estimates on derivatives). by the maximum and minimum principles. To this end. words. 2. max u = max u = 0. applying the maximum principles to −u. Explain why (2) implies (1). If u ∈ C 2 (Ω) ∩ C(Ω) ¯ If u(x0 ) = 0 at some x0 ∈ Ω. the Dirichlet problem for Poisson’s equation ( −∆u = f in Ω.r) So equality must hold. Z Z u(x) = − u(y) dy ≤ − u(x) dy = u(x). hence. we have the following positivity result for all bounded open connected sets (domains) Ω: ¯ is harmonic and u|∂Ω ≥ 0 but 6≡ 0. Let u1 . Ω is connected and there exists an ¯ In other interior point x0 ∈ Ω such that u(x0 ) = maxΩ¯ u. which implies u ≡ 0 on ∂Ω. if u is a non-constant harmonic function in a bounded domain Ω. ¯ Let u = u1 − u2 . u2 be any two solutions to the problem in C 2 (Ω) ∩ C(Ω). then u(x) > 0 for all x ∈ Ω. u(x) < max u ∂Ω Proof. Estimates of higher-order derivatives and Liouville’s theorem Theorem 2. set S = {x ∈ Ω | u(x) = u(x0 )}.  Theorem 2. Estimates of higher-order derivatives and Liouville’s theorem 38 (2) (Strong maximum principle) If. Then Ck kukL1 (B(x. By the minimum principle. It is relatively closed in Ω since u is continuous. To show S is open. Hence S = Ω because Ω is connected. So we only prove (2). one also has the minimum principles: all previous results are valid if maximum is changed to minimum. u(x) = u(x0 ).12 (Uniqueness for Dirichlet problem). This proves the conclusion (2). by the mean-value property. so. We show S is an open set. then u(x) ≡ u(x0 ) for all x ∈ Ω. This set is nonempty since x0 ∈ S. which implies u(y) = u(x) for all y ∈ B(x.5. r) ⊂⊂ Ω. Proof. Since u(x) = u(x0 ) = maxΩ¯ u. ¯ strong minimum principle u ≡ 0 on Ω. In particular. r) ⊂ S and hence x ∈ S is an interior point of S. in addition.  If u is harmonic then −u is also harmonic. Hence max∂Ω = min∂Ω = 0.23) |Dα u(x)| ≤  . ¯ which implies u ≡ 0 on Ω. Assume B(x. Assume u is harmonic in Ω. proving that S is open. B(x. r). Note that the only nonempty subset that is both open and relatively closed in a connected set is the set itself. that is. Given f and g. ¯ Ω ∂Ω min u = min u = 0. so.5. u ≥ 0 on Ω. ¯ Ω ∂Ω ¯ Hence u1 ≡ u2 on Ω. take any x ∈ S. r) ⊂⊂ Ω and each multi-index α of order |α| = k. (2. u=g on ∂Ω ¯ can have at most one solution u ∈ C 2 (Ω) ∩ C(Ω). a contradiction. then by the Proof.

r/2) 1 B(x.r/2) 2n max |u(y)|. Z Z 2n ux1 (y)dy = ux1 (x) = − ux (y)dy αn rn B(x. say α = (1. 1. one has B(y.5. r) ⊂⊂ Ω. hence.r/2) Z 2n = ν1 u(y) dS.r/2) However. 0. r) ⊂⊂ Ω and hence .2. · · · .r/2) So 2n |ux1 (x)| ≤ αn rn Z |u(y)| dS ≤ ∂B(x. We use induction on k. ux1 is also harmonic. r/2) ⊂⊂ B(x. Estimates of higher-order derivatives and Liouville’s theorem 39 Proof. αn rn ∂B(x. Since u ∈ C ∞ is harmonic. for each y ∈ ∂B(x. if B(x. 0) then Dα u = ux1 . If k = 1. r/2). r y∈∂B(x.

Z .

.

.

2n .

.

u(z)dz .

≤ |u(y)| = .

.r)) . − kukL1 (B(x.

r/2) . B(y.

· · · . r) ⊂⊂ Ω. 2. as above.r/k)) . 1. Let |u(y)| ≤ M for all y ∈ Rn . Hence u is constant. Assume now k ≥ 2. 2. αn k = 1. hence. r)) n+k−1 n+k−1 k r ( k−1 r) k Combining the previous two inequalities. ≤ k−1 1 L (B(y. with C1 = 2n+1 n αn . n+1 r r r This inequality holds for all r since B(x. 0. α = β + (0. |Dα u(x)| ≤ nk β kD ukL∞ (B(x. C1 C1 M C1 kukL1 (B(x. By (2.r)) . r) ⊂⊂ Rn . n. by induction assumption with Dβ u at y. So Dα u = (Dβ u)xi and.23) with k = 1. Taking r → ∞ we have uxi (x) = 0 for each i = 1.r)) .  |uxi (x)| ≤ . k−1 k r) ⊂⊂ B(x. r If y ∈ B(x. 0). n. · · · . For example. where |β| = k − 1. αn rn Combining the inequalities above. we have |ux1 (x)| ≤ 2n+1 n kukL1 (B(x. for some i. · · · . we have |Dα (x)| ≤ Ck kukL1 (B(x.23) with k = 1. Proof. Then u is constant. rn+k k n+k−1 where Ck ≥ Ck−1 nk( k−1 ) .14 (Liouville’s Theorem). 2. for each i = 1. Let |α| = k. αn rn+1 This proves (2.r)) ≤ n+1 M αn rn = . Then. · · · .  Theorem 2. r/k) then B(y. |Dβ u(y)| ≤ k n+k−1 ) Ck−1 ( k−1 Ck−1 kuk kukL1 (B(x.r)) . · · · . Suppose u is a bounded harmonic function on whole Rn . we can choose Ck = (2n+1 nk)k . 2.

5. 2. a constant. Then u(x) ≥ 1 2n(N +1) u(y) ∀ x. y ∈ V. Let n ≥ 3 and f ∈ Cc∞ (Rn ). Then B(y.r) B(y. Ω) such that inequality sup u ≤ C inf u V V holds for all nonnegative harmonic functions u on Ω. r By Arzela-Ascoli’s theorem. y ∈ V .  |Duj (x)| ≤ Theorem 2. u − u ˜ = C. ∂Ω). we have CM (x ∈ V¯ .2r) B(y. 2r) ⊂⊂ Ω and hence Z Z Z 1 1 1 u(x) = − u(z)dz ≥ u(z)dz = n − u(z)dz = n u(y). Applying the local estimate above. Then any bounded solution of −∆u = f on Rn has the form Z Φ(x − y)f (y) dy + C u(x) = Rn for some constant C. · · · ).2. Let r = 14 dist(V. r) ⊂⊂ Ω for all x ∈ V¯ . r) ⊂ B(x.r) Thus 2n u(y) ≥ u(x) ≥ 1 2n u(y) if x. ∂Ω).17 (Harnack’s inequality). j = 1.15 (Representation formula). 2. By Liouville’s Theorem. we can cover V¯ by a chain of finitely many balls {Bi }N i=1 . Then B(x. |x − y| ≤ r. there exists a subsequence of {uj } which converges uniformly in V to a function u ¯ in V¯ . n rn α 2 2 2 n B(x. Proof. For each subdomain V ⊂⊂ Ω. Proof. This uniform limit function u ¯ certainly satisfies the mean-value property in V since each uj does and hence is harmonic in V . Let V ⊂⊂ Ω. ¯ Since V is connected and V is compact in Ω. Suppose {uj } be a sequence of harmonic functions in Ω and |uj (x)| ≤ M (x ∈ Ω. we can prove that the (Newton’s potential) function Z Φ(x − y)f (y) dy u ˜(x) = Rn is bounded on Rn . 2. Estimates of higher-order derivatives and Liouville’s theorem 40 Theorem 2. · · · ). N. y ∈ V and |x − y| ≤ r. Then there exists a subsequence {ujk } and a harmonic function u ¯ in V such that lim kujk − u ¯kL∞ (V ) = 0. j = 1. This function solves Poisson’s equation −∆˜ u = f . each of which has radius r/2 and Bi ∩ Bi−1 6= ∅ for i = 1.  Theorem 2. k→∞ Proof.  .16 (Compactness of sequence of harmonic functions). there exists a constant C = C(V. Let x. Since Φ(y) → 0 as |y| → ∞ if n ≥ 3. · · · . Let 0 < r < dist(V. Hence u − u ˜ is a bounded harmonic function on Rn .

Lemma 2.ρ (y) dSy = Muξ.ρ (x) = u(x) ≤ Mu (x.24).ρ (ξ) = Muξ.ρ ∈ σ(Ω) and u(x) ≤ uξ.ρ is simply Poisson’s integral of u|∂B(ξ.  Lemma 2. by the previous lemma. ρ). We need to show (2. ρ).24) holds if x ∈ / B(ξ. 0 all y ∈ B(x.2. ρ).24) is proved. By definition. Hence (2.r) for all sufficiently small r > 0. .ρ is subharmonic in Ω.ρ is subharmonic in B(ξ. ¯ ∩ σ(Ω). r0 ) ⊂ B(ξ. ρ). r0 ) ⊂⊂ Ω such ¯ ρ) or B(x.ρ (x) = ρ2 −|x−ξ|2 R u(y) nαn ρ ∂B(ξ. we have uξ. ρ) ⊂⊂ Ω. Then u(ξ) ≤ uξ. r0 ) ⊂ Ω \ B(ξ.ρ (y) is harmonic in B(x.25) holds if 0 < r < r0 is sufficiently small since u is subharmonic. u ≤ uξ. ρ). r ) or uξ. by (2. ρ).ρ (x.ρ (y) = u(y) for that either B(x.24) ∀ x ∈ Ω. (2. r) ∂B(x. while in second case. Let B(ξ.ρ (x) (2. (2. Perron’s method for existence using subharmonic functions 41 2. we have to show that for any x ∈ Ω Z (2.25) holds for all 0 < r < r0 since uξ. r) ≤ Muξ. uξ.ρ (x) ≤ − uξ. we define uξ. If x ∈ / ∂B(ξ.ρ is harmonic. In this case. then there exists a ball B(x. ρ) ∂B(ξ. Lemma 2. ¯ Ω ∂Ω For u ∈ C(Ω) and B(ξ.6.19.ρ) |y−x|n dSy if x ∈ B(ξ. (2. ρ).18. u(ξ) ≤ Mu (ξ. Proof.ρ on B(ξ.ρ) and hence is harmonic and takes the same boundary on ∂B(ξ. ρ) = Mu (ξ. r) for all sufficiently small r > 0.ρ (x) by ( u(x) if x ∈ Ω \ B(ξ.25) uξ. Proof. ρ) and hence either uξ. We denote by σ(Ω) the set of all subharmonic functions in Ω. uξ. Note that u − uξ. u is harmonic in Ω if and only if both u and −u belong to σ(Ω).ρ) holds for all sufficiently small ρ > 0.21. For u ∈ C(Ω) max = max u. We call u : Ω → R subharmonic in Ω if u ∈ C(Ω) and if for every ξ ∈ Ω the inequality Z u(ξ) ≤ − u(x)dS := Mu (ξ.25) if x ∈ ∂B(ξ.  Lemma 2. For u ∈ σ(Ω) and B(ξ.20. ρ) ⊂⊂ Ω. ρ). r0 ). Perron’s method for existence using subharmonic functions Definition 2.ρ (x.6. For u ∈ σ(Ω).8. ρ) the function uξ.ρ is in C(Ω). Note that in B(ξ. ρ) whenever B(ξ. Therefore uξ.ρ (ξ. ρ) ⊂⊂ Ω. hence. ρ) ⊂⊂ Ω. ρ) and equals zero on ∂B(ξ. ρ). To show that uξ. In first case.

by (2.16). m ≤ v j (x) ≤ M ¯ (x ∈ Ω). ¯ u(x) ≤ max u = M for u ∈ σg (Ω). · · · ). by the compactness theorem (Theorem 2. k→∞ k→∞ . Of course W depends on the choice of {xk } and the subsequence of {v j }.ρ in Ω. where 0 < ρ0 < ρ. in particular. there exists a subsequence uniformly convergent to a harmonic function W on B(ξ. v2 . then for all B(ξ. m ≤ uj (x) ≤ M (x ∈ Ω).ρ . 2.ρ .6. · · · ). Lemma 2. Hence u ≡ uξ.  Given g ∈ C(∂Ω). and lim v j (xk ) = wg (xk ) j→∞ (k = 1. v2 .26) W (xk ) = wg (xk ) (k = 1. · · · . and m ∈ σg (Ω). by the maximum principle. j→∞ Let v j = ujξ. Also. ρ). u ≤ uξ.22. Let v1 . Since {v j } is a bounded sequence of harmonic functions in B(ξ.2. be a sequence of points in B(ξ. ∂Ω Then m. The following has been left as a homework problem. · · · ). wg (x0 ) = W (x0 ) = lim W (xk ) = lim wg (xk ). Lemma 2. Let B(ξ. Proof. Define uj (x) = max{m. vk ∈ σg (Ω) and v = max{v1 . So σg (Ω) is non-empty. (k = 1. 2. ρ0 ). −u ≤ −uξ. · · · . Then v ∈ σg (Ω). We can find ujk ∈ σg (Ω) such that wg (xk ) = lim ujk (xk ) j→∞ (k = 1. · · · . Let x0 ∈ B(ξ. ρ).ρ . 2. Hence (2. ujj (x)} ¯ and Then uj ∈ σg (Ω). 2.26) and the continuity of W . ∂Ω M = max g. M are finite. ρ0 ). uj2 (x). ρ) ⊂⊂ Ω. Perron’s method for existence using subharmonic functions 42 Proof. ∂Ω Hence the function wg is well-defined in Ω. σg (Ω) = {u ∈ C(Ω) and wg (x) = sup u(x) (x ∈ Ω). ρ). ρ0 ). x ∈ Ω. If u. then v j is harmonic in B(ξ. · · · . define ¯ ∩ σ(Ω) | u ≤ g on ∂Ω}. x2 . −u ∈ σ(Ω). However the function wg is independent of all these choices. lim uj (xk ) = wg (xk ) ¯ (x ∈ Ω). With this choice of {xk }. u is harmonic in B(ξ. ρ) ⊂⊂ Ω. ρ0 ) such that x1 = x0 and xk → x0 . vk }.23. wg is harmonic in Ω. · · · ). We first choose {xk } in B(ξ. u∈σg (Ω) Set m = min g. uj1 (x). Let x1 .

so is wg in B(ξ. Hence u(x) ≤ g(x) (x ∈ ∂Ω ∩ B(η. δ)). Perron’s method for existence using subharmonic functions 43 Since {xk } is arbitrary. −U ∈ C(Ω) −U ≤ −g on ∂Ω. Since Qη (x) < 0 on compact set ∂Ω \ B(η. Since W is harmonic in B(ξ. ρ0 ).26) implies that wg ≡ W in B(ξ.6. x∈Ω We consider function −w−g (x) which is defined by −w−g (x) = − sup u(x) = inf U (x). (2. We first prove lim inf wg (x) ≥ g(η). there exists a δ > 0 such that g(x) > g(η) −  for x ∈ ∂Ω ∩ B(η.27). We now prove (2. Since g is continuous. it follows that Qη (x) ≤ −γ on ∂Ω \ B(η. To study the behavior of wg (x) as x approaches a boundary point η ∈ ∂Ω. x→η. ρ0 ). u(η) = g(η) − . Let K = M −m γ u(x) = g(η) −  + KQη (x) ≤ M − Kγ = m ≤ g(x) (x ∈ ∂Ω \ B(η.2.24. This proves that wg is harmonic in Ω. and Then u ∈ C(Ω) u(x) ≤ g(η) −  (x ∈ ∂Ω). x→η. We now choose {xk } to be a dense sequence in B(ξ. with v ∈ σ−g (Ω). So wg is continuous in Ω. satisfies ¯ ∩ σ(Ω). ρ0 ). u ∈ σg (Ω). δ). If η ∈ ∂Ω is regular. Thus u ≤ g on ∂Ω and. x∈Ω wg (x) = g(η). ρ0 ). we have proved that the function wg is harmonic in Ω. ≥ 0. we need some property of the boundary ∂Ω at point η. Then (2. ¯ A point η ∈ ∂Ω is called regular if there exists a barrier function Qη (x) ∈ C(Ω)∩σ(Ω) such that Qη (η) = 0. K > 0 be constants and u(x) = g(η) −  + KQη (x). Proof. 2. Then g(η) −  = lim x→η.x∈Ω This proves (2. ¯ ∩ σ(Ω).28) lim sup wg (x) ≤ g(η). by definition. δ). δ)). Then where γ > 0 is a constant. x∈Ω Let  > 0. δ). . v∈σ−g (Ω) where U = −v.  Therefore. then lim x→η.27) x→η. Lemma 2. ρ0 ). So u(x) ≤ wg (x) (x ∈ Ω). 1. Qη (x) < 0 (x ∈ ∂Ω \ {η}). this proves the continuity of wg on B(ξ.x∈Ω u(x) ≤ lim inf wg (x).

x∈Ω x→η. the function wg (x) defined above is harmonic in Ω and has boundary limit g.28) as follows: lim sup wg (x) ≤ lim sup (−w−g (x)) = − lim inf w−g (x) ≤ g(η).) The maximum principle for harmonic functions depends heavily on the mean-value property of harmonic functions and so does the Perron’s method for the solvability of Dirichlet problem of Laplace’s equation. Maximum principles for second-order linear elliptic equations 44 ¯ and For each u ∈ σg (Ω) and each U = −v as above. On the other hand. u=g on ∂Ω ¯ for all continuous boundary data g ∈ C(∂Ω) if and only if has a solution u ∈ C 2 (Ω) ∩ C(Ω) every point η ∈ ∂Ω is regular. (Explain why). 2. x∈Ω  Theorem 2. u − U ≤ 0 in Ω. x→η.  Remark 2. for any u and U above and hence wg (x) ≤ −w−g (x) ∀ x ∈ Ω. where Φ(x) = φ(|x|) is the fundamental solution above. it follows that u − U ∈ σ(Ω) ∩ C(Ω) u − U ≤ g + (−g) = 0 on ∂Ω. If every point η ∈ ∂Ω is regular. ρ) in the exterior domain Rn \ Ω such that point η ∈ ∂Ω if there is a closed ball B = B(x B ∩ ∂Ω = {η}. then for each g ∈ C(∂Ω). we deduce (2. Proof. Existence of the so-called weak solutions for some of these equations is based on energy method in Sobolev spaces and other modern methods of PDEs and will be briefly introduced later. A domain Ω is said to have the exterior ball property at boundary ¯ 0 .25.2. if you are interested in more of the modern theory of PDEs. the proof of maximum principles and the technique of Perron’s method are strongly limited to Laplace’s equation and can not be used to other more general important PDEs in geometry. We say that Ω has the exterior ball property if it has this property at every point ¯ η ∈ ∂Ω.7. u(x) = Qη (x). physics and applications. x∈Ω x→η.7. if Ω is strictly convex. hence η is regular.4 of the textbook.9. so. . the Dirichlet problem is always uniquely solvable in C 2 (Ω)∩C(Ω). In this case. g(x) = −|x−η| in C 2 (Ω)∩C(Ω). Maximum principles for second-order linear elliptic equations (This material is from Section 6. given η ∈ ∂Ω. η is regular point of ∂Ω since the barrier function at η can be chosen as Qη (x) = Φ(x − x0 ) − φ(ρ). Applying (2. by the maximum principle.27) above to w−g . extending wg to ∂Ω by g gives ¯ a solution u to the problem in C 2 (Ω) ∩ C(Ω). Therefore. then Ω has the exterior ball property. For such domains. ¯ This proves that u(x) ≤ U (x) (x ∈ Ω) Hence. We study the maximum principles for second-order linear elliptic PDEs in this section. In particular. Let Ω ⊂ Rn be bounded. you should register in the sequence courses MTH 940-941. The Dirichlet problem ( ∆u = 0 in Ω. if the Dirichlet problem is solvable for boundary data ¯ then the solution u is a barrier function at η.

if Lu is elliptic.7. j = 1. t=0 dt2 i. (Warning: I am using a different notation for c− (x) from the Textbook. we assume that aij (x) = aji (x).26. i=1 Here Di u = uxi and Dij u = uxi xj . 2. Maximum principles for second-order linear elliptic equations 45 2. the matrix (aij (x)) is positive definite. u(x0 ) is a maximum of u in Ω. for some x0 ∈ Ω. Without the loss of generality. i=1 aij (x)Dij u(x) + i. with smallest eigenvalue not less than λ(x) > 0. We consider the second-order linear differential operator Lu(x) = − n X ij a (x)Dij u(x) + i. Lemma 2.j=1 for any ξ = (ξ1 . Let Lu be elliptic in Ω and u ∈ C 2 (Ω) satisfy Lu < 0 in Ω. where c− (x) = − min{0. then u can not attain a nonnegative maximum in Ω.10. Second-order linear elliptic PDEs. . Weak maximum principle and the uniqueness. for each x ∈ Ω. . and (by the second-derivative test) n X d2 u(x0 + tξ) | = Dij u(x0 )ξi ξj ≤ 0.j=1 k=1 If λ(x) ≥ λ0 > 0 for all x ∈ Ω. ξ ∈ Rn . bi and c are well-defined everywhere in Ω. If c(x) ≥ 0. c(x)}. We argue by contradiction.e. i. ξn ) ∈ Rn . Definition 2. then u can not attain a maximum in Ω.j=1 n X bi (x)Di u(x) + c(x)u(x). aij = n X (i. n). Lemma 2. then there is an invertible matrix B = (Bij ) such that A = B T B. bki bkj k=1 2.2. Then Dj u(x0 ) = 0. i. Throughout this section we assume that all functions aij . (This is to say the Hessian matrix (Dij u(x0 )) ≤ 0. . 2.j=1 n X bi (x)Di u(x) + c+ (x)u(x). c(x)}. Suppose. The operator Lu is called elliptic in Ω if there exists λ(x) > 0 (x ∈ Ω) such that n n X X ij a (x)ξi ξj ≥ λ(x) ξk2 .1. n). If A = (aij ) is a positive definite matrix.27.7. .) We write aij (x0 ) = n X k=1 bki bkj (i. then. · · · . j = 1.2.j=1 ˜ Lu(x) =− n X n X bi (x)Di u(x). c(x)}. we say Lu is uniformly elliptic in Ω. If c ≡ 0. i=1 where c+ (x) = max{0. Proof. ∀ x ∈ Ω. So.7. . We also denote c− (x) = min{0. · · · . We use the following notations: L0 u(x) = − n X aij (x)Dij u(x) + i.) First we recall a fact in linear algebra..

28 (Weak Maximum Principle). then the proof is still valid if the maximum is changed to the supremum. noticing that v + = −u− = − min{0. n) ¯ satisfies Lu ≤ 0 (in this case for some constant M > 0. 1. |bk | ≤ M for some k. Assume Lu(x) = 0. i = 1. λ(x) ≥ 0. ¯ Ω ¯ Ω ∂Ω ∂Ω . 2. ¯ Ω ∂Ω ¯ Ω ∂Ω where u+ (x) = max{0. Then   b1 (x) eαx1 < 0 Lv(x) = (−a11 (x)α2 + b1 (x)α)eαx1 ≤ αa11 (x) −α + 11 a (x) if α > M + 1.j=1 which implies that Lu(x0 ) ≥ c(x0 )u(x0 ) ≥ 0 either when c ≥ 0 and u(x0 ) ≥ 0 or when c ≡ 0. It follows from Theorem 3.j=1 n X n X Dij u(x0 )bki bkj ≤ 0. If Ω+ is empty.7.8 we can easily see the following Theorem 2. Assume Ω+ is not empty. Let c(x) ≥ 0 and u ∈ C 2 (Ω) ∩ ¯ Then C 0 (Ω).29). then L0 u(x) = Lu(x) − c(x)u(x) ≤ 0 in Ω+ . u}. ¯ Ω Ω+ ∂(Ω+ ) ∂Ω 2.2. ∂Ω Setting ε → 0+ ∂Ω ∂Ω proves the theorem.30) max u ≤ max u+ . 1. Let Lu ≤ 0. max |u| = max |u| if Lu = 0. 2. Maximum principles for second-order linear elliptic equations 46 where B = (bik ) is a matrix.  Remark 2.6 that for any x ∈ Ω.29) max u = max u. the result is trivial. k=1 i. This gives a contradiction to Lu(x0 ) < 0 in both cases.  Theorem 2..11. then (2. It ¯ follows from Lemma 3. as |b1 (x)| a11 (x) ≤ |b1 (x)| λ(x) ≤ M. Let v(x) = eαx1 .29. Hence n X aij (x0 )Dij u(x0 ) = i. ¯ Ω ∂Ω − min u = max v ≤ max v + = − min u− . Consider w(x) = u(x) + εv(x) for any ε > 0. Let Lu be elliptic in Ω and satisfy (2. u(x) + εv(x) ≤ max(u + εv) ≤ max u + ε max v. Note that max∂(Ω+ ) u = max∂Ω u+ . but satisfies akk (x) (b) If Ω is unbounded. We apply the inequality in Step 1 to u and v = −u.) i. Let Ω+ = {x ∈ Ω | u(x) > 0}. Let Lu be elliptic in Ω and |bi (x)|/λ(x) ≤ M (x ∈ Ω. Proof. · · · . Then Lw = Lu + εLv < 0 in Ω. (a) The weak maximum principle still holds if (aij ) is nonnegative definite. ¯ Ω ∂Ω Proof.8 that max u = max u = max u = max u+ . From the proof of Theorem 3.e. (Then use v = eαxk . If c(x) = 0 and u ∈ C 2 (Ω) ∩ C 0 (Ω) we say u is a sub-solution of L) in a bounded domain Ω. max u ≤ max u+ if Lu ≤ 0. u(x)}. but is bounded in a slab |x1 | ≤ N . to deduce (2.

π). then the same conclusion holds provided u(x0 ) ≥ 0. there is a ε > 0 such that wε (x) < 0 for all |x| = R2 . then solution to the Dirichlet problem Lu = f in Ω. where c > 0 is a constant. u(x) < u(x0 ) for all |x| = R2 . u|∂Ω = g ¯ is unique in C 2 (Ω) ∩ C(Ω).29). (a) The maximum principle and the uniqueness fail if c(x) < 0. the same conclusion holds no matter what sign of c(x) is.2. π)n in dimensions n ≥ 2. But u 6≡ 0. · · · . In addition. (b) Nontrivial solutions can be constructed for equation −∆u−cu = 0 with zero Dirichlet boundary data in the cube Ω = (0. we assume that B = B(0. since v|∂B = 0. ∂ν (b) If c(x) ≥ 0. If c(x) ≥ 0 and Ω is bounded. consider wε (x) = u(x) − u(x0 ) + εv(x).30. bn (x))) ≤ |x| ≤ R. By assumption. u(0) = u(π) = 0.7. Let Lu be uniformly elliptic with bounded coefficients in a ¯ satisfy Lu ≤ 0 in B. Remark 2.7.12. the weak maximum principle applies to L. while the book uses c = − min{0. (a) If c(x) = 0 in B. so. Assume that x0 ∈ ∂B such that u(x) < u(x0 ) ball B and u ∈ C 2 (B) for every x ∈ B. Strong maximum principle. provided α > 0 is fixed and sufficiently large. Lemma 2. hence. 3. ˜ = Lu − c(x)u + c+ (x)u = Lu − c− (x)u is an elliptic operator with nonnegative Recall Lu ˜ coefficient for the zero-th order term.j (2. where ν is outer unit normal to B at x0 .  Theorem 2. Without the loss of generality. Let Lu be elliptic in Ω and satisfy (2. − − (Warning: Again. (c) If u(x0 ) = 0.31 (Hopf’s Lemma). Proof. Consider 2 2 v(x) = e−α|x| − e−αR .) It is easy to check that h i X X 2 ˜ Lv(x) = −4 aij (x)α2 xi xj + 2α (aii (x) − bi (x)xi ) e−α|x| + c+ (x)v(x) i. R). ¯ then ∂u (x0 ) > 0. c}. For any ε > 0. The following lemma is needed to prove a version of the strong maximum principle for elliptic equations. c = min{0. if n = 1.3. Then ˜ ε (x) = εLv(x) ˜ Lw + Lu(x) − c− (x)u(x) − c+ (x)u(x0 ) < −c− (x)u(x) − c+ (x)u(x0 ) ≤ 0 on R 2 ≤ |x| ≤ R in all cases of (a). c} here.31) ≤ h < 0 on R 2 i i 2 − 4λ0 α2 |x|2 + 2α tr(A(x)) + 2α|b(x)||x| + c+ (x)) e−α|x| (here A(x) = (aij (x)). b(x) = (b1 (x). Maximum principles for second-order linear elliptic equations 47 This yields maxΩ¯ |u| = max∂Ω |u|. For example. 2. (b) and (c). 1. the function u(x) = sin x satisfies −u00 − u = 0 in Ω = (0. 2. we have wε (x) = u(x) − u(x0 ) ≤ 0 on .

(The proof for some special cases can be found in Textbook. Let u ∈ C 2 (Ω) ∩ C(Ω).34 (Harnack’s inequality). ∂ν as required. Weak solutions and existence in Sobolev spaces 48 |x| = R. Finally. R) be the largest ball contained in Ω− and centered at y. Suppose that u is not a constant but there exists a x ¯ ∈ Ω such that u(¯ x) ≥ u(x) for all x ∈ Ω.8. Let B = B(y. (b) If c = 0. Let V ⊂⊂ Ω be connected and Lu be uniformly elliptic in Ω with bounded coeffients. (a) If c(x) ≥ 0. C) < dist(y. ∂Ω). c = 0 and u is not a constant. Let M = u(¯ x).2.8. Weak solutions and existence in Sobolev spaces (This material is from parts of Chapter 5 and Sections 6.) Theorem 2.) We study the second-order linear elliptic operator in divergence form: n n X X ij Lu(x) = − Dj (a (x)Di u(x)) + bi (x)Di u(x) + c(x)u(x). Ω. Then the sets Ω− = {x ∈ Ω | u(x) < M }. Hence u(x) < u(x0 ) for all 0 x ∈ B.  Theorem 2. then min u(y) < u(x) < max u(y) y∈∂Ω y∈∂Ω ∀ x ∈ Ω. Then Hopf’s Lemma would imply that ∂u ∂ν (x ) > 0. then Theorem 2. Ω0 = {x ∈ Ω | u(x) = M } are both nonempty (since u 6≡ M ). i.33.1 and 6. c(x) ≥ 0 and u is not a constant. Then. We assume M ≥ 0 if c(x) ≥ 0. We proceed by contradiction. then u can not attain a nonnegative maximum in Ω unless u is a constant. which actually follows from the proof of the similar result for parabolic equations later. and thus ∅ = 6 ∂Ω0 ∩ Ω ⊂ C = ∂Ω− ∩ Ω.j=1 i=1 . 2. |u(x)| < max |u(y)| ∀ x ∈ Ω. Let y ∈ Ω− be such that dist(y. we state without proof the following Harnack’s inequality for nonnegative solutions of elliptic equations. Hence ∂wε 0 ∂u 0 ∂v ∂u 0 2 (x ) = (x ) + ε (x0 ) = (x ) − 2εRαe−αR . this contradicts to the fact that Dj u(x ) = 0 since u has maximum at x0 ∈ Ω.32 (Strong Maximum Principle). where ν is the outnormal of B 0 0 at x .  ¯ If Lu = 0. Then there exists a point x0 ∈ C. y∈∂Ω If Lu = 0. Let L be uniformly elliptic in a bounded domain Ω with bounded coefficients and u ∈ C 2 (Ω) such that Lu ≤ 0 in Ω. L) > 0 such that sup u ≤ C inf u V V for each nonnegative C 2 function u of Lu = 0 in Ω. with x0 ∈ ∂B.2 of the Texttbook. The weak maximum principle implies that wε (x) ≤ 0 for all x0 is a maximum point of wε on R2 ≤ |x| ≤ R and thus 0≤ That is R 2 ≤ |x| ≤ R. then u can not attain a maximum in Ω unless u is a constant. ∂ν ∂ν ∂ν ∂ν ∂u 0 2 (x ) ≥ 2εRαe−αR > 0. there exists C = C(V. Proof.

8. Example 2. In fact.1. v) = (uv + Du · Dv) dx. Moreover. Ω i. · · · . n) Ω Ω Cc∞ (Ω). Sobolev spaces H 1 (Ω) and H01 (Ω). then for each function v ∈ Cc∞ (Ω). we have Z Z n n X X  ij i a (x)uxi (x)vxj (x) + b (x)v(x)uxi (x) + c(x)u(x)v(x) dx = f v dx. We study the following Dirichlet problem: ( Lu = f in Ω. Show that u = |x|−α ∈ H 1 (Ω) if and only if α < (b) Show that u = 1 − |x| ∈ H01 (Ω). v] := Ω ij a (x)uxi (x)vxj (x) + i. · · · . (2. n}.j=1 n X  bi (x)v(x)uxi (x) + c(x)u(x)v(x) dx. for all test functions φ ∈ In this case we call wi the weak derivative of u in Ω with respect to xi and denote wi = uxi = Di u. Ω Theorem 2. ∀ x ∈ Ω. the definition of Du can be even extended as long as it is in L2 (Ω).13.j=1 Ω i=1 Define n X Z (2. bi and c are all bounded functions in Ω. v ∈ H 1 (Ω)). Let Ω = B(0. Let Ω be a bounded domain in Rn . wn ) to be the weak gradient vector of u in Ω.36. (a) Let α > 0. Definition 2. k=1 If λ(x) ≥ λ0 > 0 for all x ∈ Ω. A function u ∈ L2 (Ω) is called weakly differentiable in Ω if there exist functions wi (i = 1. 1) be the unit ball in Rn . · · · . · · · .j=1 n X ξk2 . we say Lu is uniformly elliptic in Ω.35. 2. Then H 1 (Ω) is a linear subspace of L2 (Ω). v] we do not need u to be C 2 but we only need Du exists and is in L2 (Ω) in order to define B[u. Weak solutions and existence in Sobolev spaces 49 Here we assume that aij (x) = aji (x) and aij . v) defines an inner product on H 1 (Ω) that makes H 1 (Ω) a Hilbert space. v]. 2.32) u=0 on ∂Ω. vLu = 0 on Ω. 2. denote Du = ∇u = (w1 . The operator Lu is called elliptic if there exists λ(x) > 0 such that n X aij (x)ξi ξj ≥ λ(x) i. i=1 In B[u.2. n−2 2 . If aij are smooth and u is a RC 2 solution to the Dirichlet problem. w2 . Define H 1 (Ω) = {u ∈ L2 (Ω) | u is weakly differentiable with Di u ∈ L2 (Ω) for i = 1. ξ ∈ Rn .33) B[u.8. In H 1 (Ω) we define Z (u. (u. kuk = (u. u)1/2 (u. We define H01 (Ω) to be the closure of Cc∞ (Ω) in the Hilbert space H 1 (Ω). Using integration by parts. n) in L1loc (Ω) such that Z Z uφxi dx = − φwi dx (i = 1. So Ω vLu dx = 0. .

write y = (y1 . we have deduced (2. a)n . y 0 ) dx1 (y1 ∈ (−a. y 0 ) = ux1 (x1 . Q Q Therefore. {uj } is also a Cauchy sequence in H 1 (Ω).34) (u ∈ H01 (Ω)). ¯ by zero To prove this. y )| ≤ (y1 + a) 0 2 a 0 2 |ux1 (x1 .8. Ω Proof. Fix any u ∈ Cc∞ (Ω). Hence there exists a subsequence {ujk } converging to some u ¯ ∈ H 1 (Ω) under the H 1 -norm. a) and y ∈ Q ≡ (−a. we deduce Z Z 2 2 |u(y)| dy ≤ 4a |ux1 (y)|2 dy. v ∈ H01 (Ω)). −a 2 Z a |u(y1 . y )| dx1 ≤ 2a −a and consequently. 0 Then. 1. assume Ω ⊂⊂ Q ≡ (−a. Z (2. Extend u to Q 0 0 0 outside Ω. a). Z 0 2 y1 |u(y1 . Then Z y1 u(y1 .14.  Remark 2. u)H 1 . We first prove the Poincar´ e’s inequality kukL2 (Ω) ≤ CkDukL2 (Ω) (2. (a) For a bounded domain Ω with Lipschitz boundary. Z Z |ux1 (x1 . −a By H¨older’s inequality. that is. y 0 )|2 dx1 (y 0 ∈ Q0 ). y ) with y1 ∈ (−a. y 0 )|2 dx1 . This proves that H0 (Ω) is complete under the H01 -norm. the SobolevRellich-Kondrachov compactness theorem asserts that H 1 (Ω) is compactly embedded in L2 (Ω). Let Ω be bounded. Certainly u ¯ ∈ H01 (Ω) and {ujk } 1 1 also converges to u ¯ in the H0 -norm. y 0 ∈ Q0 ). and hence H01 (Ω) is a Hilbert space. Then H01 (Ω) is a Hilbert space under the inner product Z (u.34). a)n−1 . v)H01 = Du · Dv dx (u. we have another version of the Poincar´ e’s inequality: For bounded domains Ω with Lipschitz boundary. y )| dy1 ≤ 4a −a a |ux1 (x1 . 1/2 2. Ω . a)n−1 . kukL2 (Ω) ≤ 2akDukL2 (Ω) (u ∈ Cc∞ (Ω)). every bounded sequence in H 1 (Ω) has a subsequence convergent strongly in L2 (Ω). by (2. For each y ∈ Q.35) ku − − ukL2 (Ω) ≤ CkDukL2 (Ω) (u ∈ H 1 (Ω)). −a Integrating over y 0 ∈ Q0 = (−a.34) from the above inequality. Suppose {uj } is a Cauchy sequence in H01 (Ω) under the H01 -norm: kukH01 = (u.37. (b) By this compactness theorem.2. Weak solutions and existence in Sobolev spaces 50 Theorem 2. Since H01 (Ω) is the closure of Cc∞ (Ω) in H 1 (Ω).

v) ∀ v ∈ H. We compute that for all u ∈ H k(I − tA)uk2 = kuk2 + t2 kAuk2 − 2t(Au. Theorem 2. We show that. for some β > 0. l : H → R is a bounded linear functional on H). βkuk2 ≤ B[u. w ∈ H and a. Assume (i) B is bounded. Proof.15. u] = (Au. for some α > 0. |B[u. w] B[w. we have a unique w0 ∈ H such that l(v) = (w0 . For each fixed u ∈ H. v]| ≤ αkukkvk ∀ u. A is bounded. there exists a unique element w = Au ∈ H such that B[u.2. there exists a unique u ∈ H such that (2.. (See the textbook for a different proof.8. i. the existence of u will be proved by using the contraction mapping theorem. b ∈ R. A weak solution to Dirichlet problem (2. From (i). and hence by the Riesz representation theorem.32) is a function u ∈ H01 (Ω) such that Z f v dx ∀ v ∈ H01 (Ω). v] = (w. kAuk2 = B[u. By the Riesz representation theorem again. the functional v 7→ B[u. map T is a strict contraction. v. This defines a map A : H → H and it is easy to see that A is linear. klkH ∗ . u) ≤ kuk2 (1 + t2 α2 − 2βt) ≤ γkuk2 . v) = (Au. v] is defined by (2. that is. by (ii). and (ii) B is strongly positive (also called coercive). vxi taken as weak derivatives. B[u. u] + bB[w.e. (Lax-Milgram Theorem) Let H be a Hilbert space and B : H × H → R be a bilinear form. Definition 2. v] = Ω where B[u.. w] + bB[v. u] ≥ βkuk2 ∀ u ∈ H.36) B[u. A function B[·. w] = aB[u.2. Au] ≤ αkukkAuk. Furthermore. and hence kAuk ≤ αkuk for all u ∈ H. ·) and norm k · k. v) for all v ∈ H and klkH ∗ = kw0 k.e. ·] : H × H → R is called a bilinear form if B[au + bv. v ∈ H. w ∈ H we have kT (v) − T (w)k = k(I − tA)(v − w)k.38. v] = l(v) Moreover.) Note that solution u to Au = w0 is equivalent to a fixed-point of map T : H → H defined by T (v) = v − tAv + tw0 (v ∈ H) for any fixed t > 0. the solution u satisfies kuk ≤ 1 β (v ∈ H). Weak solutions and existence in Sobolev spaces 51 2. B[u. Let H denote a real Hilbert space with inner product (·. v] is in H ∗ . u) ≤ kAukkuk and hence kAuk ≥ βkuk for all u ∈ H. We will show that the equation Au = w0 has a unique solution u ∈ H. . for t > 0 small enough. i. The existence of weak solution falls into the general framework of Lax-Milgram Theorem in functional analysis. v ∈ H. Then.8. for each l ∈ H ∗ (that is. Note that for all v. au + bv] = aB[w.33) with uxi . v] for all u. Weak solution to Dirichlet problem. The uniqueness of u follows easily from the property kAu − Avk ≥ βku − vk for all u.

 2.32) has a unique weak solution u ∈ H01 (Ω). Let Z l(v) = f (x)v(x) dx (v ∈ H01 (Ω)).8. This u solves Au = w0 . Moreover.3. By the uniform ellipticity with constant λ0 > 0. Then there exists a constant M depending on the ellipticity constant and the L∞ -bound of coefficients bi such that. Weak solutions and existence in Sobolev spaces 52 for some 0 < γ < 1 if we choose t such that 0 < t < α2β2 . v ∈ H01 (Ω)). Using the boundedness of the coefficients. v]| ≤ αkDukL2 (Ω) kDvkL2 (Ω) = αkukH01 kvkH01 (u. Therefore.2.j=1 By the boundedness of bi and Cauchy’s inequality with ε. We state the following special existence theorem. Existence of weak solutions.34). Let Lu be uniformly elliptic in Ω with bounded coefficients. if c(x) ≥ M . by Cauchy’s inequality and Poincar´e’s inequality (2. Ω Then l is a bounded linear functional on H01 (Ω). the Dirichlet problem (2. we have . (This also uses Poincar´e’s inequality (2. Z X n aij (x)uxi uxj dx ≥ λ0 kDuk2L2 (Ω) = λ0 kuk2H 1 . Theorem 2.34).) 2. 0 Ω i. The proof is complete.34). 1. Proof.39 (Existence of weak solutions). map T : H → H is √ a contraction (with constant γ < 1) on H and thus has a fixed point u by contraction mapping theorem. then.) 3. from kf kH ∗ = kw0 k = kAuk ≥ βkuk. for each f ∈ L2 (Ω). we deduce that |B[u. we have kuk ≤ β1 klkH ∗ .8. (This uses Poincar´e’s inequality (2.

Z n .

.

X .

λ λ0 .

.

0 i kDuk2L2 (Ω) + M kuk2L2 (Ω) = kuk2H 1 + M kuk2L2 (Ω) . b (x)uxi u dx.

≤ .

0 .

Ω .

u] ≥ βkuk2H 1 (Ω) . which will not be studied in this course. β = 0 λ0 >0 2 (u ∈ H01 (Ω)). v] on Hilbert space H = H01 (Ω) satisfies all conditions of the Lax-Milgram Theorem. Z λ0 2 kukH 1 + (c(x) − M )u2 dx (u ∈ H01 (Ω)). · · · . the bilinear form B[u. u] ≥ 0 2 Ω This is called the energy estimate. what about if c(x) is not sufficiently large? This is related to the eigenvalue problem of Lu and will not be studied in this course. n). we deduce that if c(x) ≥ M then B[u. Also.  Remark 2.. 4. we have the unique existence of weak solution. either. . therefore. By the energy estimate above. Consequently. Note that M = 0 if all bi ’s are zero (i.e. Can we show the weak solution is smooth enough to be a classical solution? This is the regularity problem. B[u. Therefore. 2 2 i=1 where M depends only on λ0 and the L∞ -bound of the coefficients bi (i = 1.16. Lu has no first-order terms). 2.

uk } ∈ σ(Ω). Homework # 3. show that u = max{u1 . ¯ ∩ σ(Ω). 10. y) ≥ 0 (x ∈ Ω. (5) (20 points) Recall that u : Ω → R is subharmonic in Ω if u ∈ C(Ω) and if for every ξ ∈ Ω the inequality Z u(ξ) ≤ − u(x)dS := Mu (ξ. So complete the arguments that are left in lectures.ρ) holds for all sufficiently small ρ > 0. y)∆u(y)dy (x ∈ Ω). x 6= y). uk ∈ σ(Ω). Prove that u(x) ≡ 0 on Ω \ {0}. u2 . and g(x) = |x| for x ∈ ∂Rn+ . y)dSy = 1 (x ∈ Ω). y)u(y)dSy − G(x.Suggested exercises 53 Suggested exercises Materials covered are from Chapters 2 and 6 of the textbook with some new materials added. ∂Ω x2 x2 x2 (2) (15 points) Let Ω be the ellipsoid 21 + 32 + 43 < 1 in R3 . · · · . Let ∂G G(x. 12. for all u ∈ C (Ω). (4) (15 points) Let u be the solution of ( ∆u = 0 in Rn+ u=g on ∂Rn+ given by Poisson’s formula for the half-space. 8. y) be Green’s function for Ω and K(x. Then. Chapter 6: Problems 1. and limx→0 |x|n−2 u(x) = 0. 4. We denote by σ(Ω) the set of all subharmonic functions in Ω. Also try working on the following problems related to the covered materials. ρ) ∂B(ξ. Chapter 2: Problems 2–11. 6. we have Z Z u(x) = K(x. y ∈ Ω. Use (without proof) the ¯ fact that Ω has Green’s function G(x. y) (x ∈ Ω. y ∈ ∂Ω) be y 2 ¯ Poisson’s kernel for Ω. Suppose that u ¯ is continuous on Ω\{0}. (a) For u ∈ C(Ω) Ω (b) If u1 . show that max ¯ = max∂Ω u. K(x. 13 Ω¯ (3) (15 points) Let n ≥ 3 and Ω ⊂ Rn be a bounded domain and 0 ∈ Ω. 9. . u = 0 on ∂Ω. K(x. Show that u ∈ σ(Ω) if and only if −∆u ≤ 0 in Ω. y ∈ ∂Ω). y) to prove that for all u ∈ C 2 (Ω) max |u| ≤ max |u| + ¯ Ω ∂Ω 6 max |∆u|. |x| ≤ 1. where g is continuous. ∂Ω Ω Prove the following statements: Z G(x. y) = − ∂ν (x. bounded on ∂Rn+ . y) > 0 (x. (c) Let u ∈ C 2 (Ω). Show that Du is unbounded near x = 0. (1) (15 points) Let Ω be a bounded domain in Rn with smooth boundary ∂Ω. harmonic in Ω\{0}. 5. · · · .

(3) (10 points) Given f ∈ L2 (Ω). operator in a bounded domain Ω. ∂ν = 0 (What is the necessary condition of f for existence of smooth solutions u? Show the condition is necessary and sufficient for the existence of weak solution. P P (1) (10 points) Let Lu = − ni. Show that Lv ≤ 0 in Ω if λ is large enough. Let u ∈ C 3 (Ω) 2 2 Set v = |Du| + λu . ∂u on ∂Ω. the Dirichlet problem ( Lu + σ(u) = f in Ω u=g on ∂Ω ¯ can have at most one solution u ∈ C 2 (Ω) ∩ C(Ω). Deduce kDukL∞ (Ω) ≤ C (kDukL∞ (∂Ω) + kukL∞ (∂Ω) ) for some constant C.) . Formulate and prove the existence of weak solutions to the Neumann boundary problem in H 1 (Ω): ( −∆u = f in Ω.j=1 a (x)uxi xj be a uniformly elliptic ¯ be a solution of Lu = 0 in Ω.j=1 aij (x)uxi xj + ni=1 bi (x)uxi be a uniformly elliptic operator in a bounded domain Ω. show that.Suggested exercises 54 Homework # 4. If σ is a C 1 function on R with σ 0 (s) ≥ 0 for all s ∈ R. given any f and g. ij ¯ and Lu = − Pn (2) (10 points) Let aij ∈ C 1 (Ω) i.

3. and the problem in the book gives another way. in which there is no source or sink. V Since V is arbitrary. but points in opposite direction (flux is from regions of higher concentration to lower concentration): F = −a∇u (a > 0). which is called the heat equation when a = 1. the rate of change of the total quantity within V equals the negative of the net flux F through ∂V : Z Z d udx = − F · νdS. Therefore we obtain the equation ut = a div ∇u = a∆u. we obtain the following nonhomogeneous heat equation ut − ∆u = f (x. For many applications F is proportional to the gradient of u. t) x ∈ Ω. ∞). 55 .1. Here we discuss yet another way of finding a special solution to the heat equation. we should have ut = − div F.Chapter 3 Heat Equation The heat equation. Fundamental solution of heat equation As in Laplace’s equation case. The textbook gives one way to find such a specific solution. If there is a source in Ω. we would like to find some special solutions to the heat equation. also known as the diffusion equation. chemical concentration. t ∈ (0. describes in typical physical applications the evolution in time of the density u of some quality such as heat. etc. Let V be any smooth subdomain. dt V ∂V The divergence theorem tells us d dt Z Z udx = − V div Fdx.

t) is a solution to the one-dimensional heat equation ut = uxx . function u(x) = u1 (x1 . . We start with the following observations: (1) If uj (x. t) = w0 (t)v( xt ) − w(t)v 0 ( xt ) xt2 . 2 2 2 + w(t)v 0 ( xt ) 2t . then u(x1 . t)u1 (x2 . n. (2) If u(x. A direct computation yields 2 2 2 ut (x.1. xn ) ∈ Rn and t > 0. . Fundamental solution and the heat kernel. Therefore x2 1 u = u1 (x. t) = 2 w(t)v( xt ). 2 s this equation is satisfied if we choose λ = −1/2 and 4v 0 + v = 0 and hence v(s) = e− 4 . t) = u1 (x1 . j = 1.1. we need w0 (t)v − x2 w 00 4x2 [v + 2v 0 + v 0 ] = 0. 2 ux (x. t t t Separation of variables yields that w0 (t)t 4sv 00 (s) + 2v 0 (s) + sv 0 (s) = w(t) v(s) with s = x2 t . t) = w(t)v 0 ( xt ) 2x t . Fundamental solution of heat equation 56 3.1. t<0 is called the fundamental solution of the heat equation ut = ∆u. This simple fact is left as an exercise. t) · · · un (xn . uxx (x. t > 0) is a solution of the heat equation ut = ∆u for t > 0 and x ∈ Rn .3. . t) = w(t)v 00 ( xt ) 4x t2 For ut = uxx . then so is w(λ)u(λx. . both sides of this equality must be constant. t) = 1 tn/2 e− |x|2 4t (x ∈ Rn . . So 1 s(4v 00 + v 0 ) + (4v 0 − 2λv) = 0. t) = √ e− 4t t is a solution of ut = uxx . · · · . In this case. say. . t) are solution to the one-dimensional heat equation ut = uxx for x ∈ R and t > 0. λ2 t) for any real λ. xn . The constant 1 (4π)n/2 in the fundamental solution Φ(x. By observation (1) above. λ. t) · · · u1 (xn . Definition 3. there should be a solution of the form u(x. Therefore. . The function Φ(x. w(t) 2 1 from which we have w(t) = t− 2 . t) is to make valid the following . t) =   1 e− (4πt)n/2 0 |x|2 4t t > 0. w0 (t)t 1 =− . Especially. . t) is a solution to the heat equation ut = ∆u for x = (x1 .1.

Initial-value problem. y. (iii) for each x0 ∈ Rn . t)dx = 1. K(x. 1. t)g(y)dy Rn Z |x−y|2 1 − 4t e g(y)dy (x ∈ Rn . R Proof. t) > 0 and Rn K(x. t) = |x−y|2 1 − 4t e (4πt)n/2 (x. 2. t) = Φ(x − y. lim x→x0 . t) dy = kgkL∞ (x ∈ Rn . ∞)) ∩ L∞ (Rn × (0. y. choose a δ > 0 such that  |g(x0 ) − g(y)| < if |x0 − y| < δ. t > 0) the heat kernel in Rn . Rn Proof. ∞)) The fact u ∈ and solves the heat equation follows from the integrability of α n D Φ in R .2. unlike the fundamental solution of Laplace’s equation. t > 0). t)g(y)dy = K(x. Furthermore. y. For each t > 0.2. .1. Rn C ∞ (Rn × (0. all space-time derivatives Dα Φ are integrable on Rn for each t > 0. 2 Assume |x − x0 | < δ/2. Then (i) u ∈ C ∞ (Rn × (0. and the fact Kt = ∆x K in Rn × Rn × (0. ∞). In the following. (3. differentiation under the integral.1. Clearly. Define Z Z Φ(x − y.2). −∞  Note that. ∞)). This is a straight forward computation based on the fact (verify!) Z ∞ √ 2 e−x dx = π. Also notice that as t → 0+ it follows that Φ(x. y. x ∈ Rn . Then. Hence Z |u(x. t) → ∞. y ∈ Rn . the fundamental solution Φ(x. t) → 0 (x 6= 0) and Φ(0. 3. t)| ≤ kgkL∞ K(x. Z Φ(x. t) → δ0 in distribution as t → 0+ . t > 0). t) of the heat equation is C ∞ in x ∈ Rn for each t > 0. t) = Rn (3. we will see that Φ(·. t→0+ u(x. 0) = g(x).3.1. t) dy = 1 for all x ∈ Rn and t > 0. Fundamental solution of heat equation 57 Lemma 3. We now study the initial-value problem or Cauchy problem of the heat equation ( ut = ∆u. y.2) Theorem 3. For any fixed x0 ∈ Rn and  > 0. ∞).1) u(x. and define u by (3. t ∈ (0. x ∈ Rn . t) = g(x0 ). t > 0). (ii) ut = ∆u (x ∈ Rn . Assume g ∈ C(Rn ) ∩ L∞ (Rn ). We call the function K(x. = (4πt)n/2 Rn u(x.

Z .

.

.

t) − g(x )| = . 0 0 |u(x.

.

Φ(x − y. t)(g(y) − g(x ) dy .

.

n R .

δ) √ Rn \B(0. ∞) (in fact. we have |u(x. t e −|z|2 16 dz. 0) = 0 (x ∈ Rn ). Now I < /2. Moreover. ∞) (3. 2 Then for all |x − x0 | < t0 . t) defined by (3. note that if |y − x0 | ≥ δ. Thus |y − x| ≥ 21 |y − x0 |. ∞): ( u ˜t − ∆˜ u = 0 in Rn × (s. t) dy J ≤ 2kgkL∞ Rn \B(x0 .δ) by the change of variable z = 0 y−x √ . its domain of influence is still all of x ∈ Rn . (a) If g is bounded. Given s > 0.δ/ t) Rn \B(x0 . 3. g ≥ 0 and 6≡ 0.δ) B(x0 .δ) := I + J. t > 0) u(x.1. We select a 0 < t0 < δ/2 such that Z C dy Rn \B(x0 . t) (x ∈ Rn . t)|g(y) − g(x )| dy + ≤ Rn \B(x0 . Duhamel’s principle and the nonhomogeneous problem. the function u defined through the heat kernel K above is C ∞ in Rn × (0.3) u ˜=f on Rn × {t = s}. (b) Even when g is not continuous but only bounded. t) has the smoothing effect in the sense that function u becomes infinitely smooth in whole space-time as soon as time t > 0 no matter how rough the initial data g is at t = 0. continuous. then function u(x. Now we solve the initial-value problem of nonhomogeneous heat equation: ( ut − ∆u = f (x.δ) ≤ ≤ C tn/2 C tn/2 Z − e Z e− |y−x|2 4t |y−x0 |2 16t Z dy = C √ Rn \B(0. Fundamental solution of heat equation Z 58 Z 0 Φ(x − y. since |x − x0 | < δ/2 ≤ |y − x0 |/2.2) is in fact positive for all x ∈ Rn and t > 0. then |y − x0 | ≤ |y − x| + |x − x0 | ≤ |y − x| + |y − x0 |/2. 0 < t < t0 . We interpret these observations by saying the heat equation has infinite propagation speed for disturbances of initial data.δ/ t0 ) e −|z|2 16  dz < . . suppose we solve the following homogeneous problem on Rn × (s.1. Therefore. 2 2  Remark 3. analytic). u(x.3. There is a general method of solving nonhomogeneous problem using the solutions of homogeneous problem with variable initial data.3. t) − g(x0 )| ≤ I + J <   + = . Consequently Z Φ(x − y. t) depends on the values of g(y) at all points y ∈ Rn no matter how far y and x are away. Even the initial g is compactly supported. the heat kernel K(x. To estimate J. t)|g(y) − g(x0 )| dy Φ(x − y. such a method is usually known as the Duhamel’s principle.2. y.

we have Z tZ u(x. t − s)f (y. t − s) dyds (i. y. we use C12 (Ω × I) to denote the space of functions u(x. s) dyds 0 Rn (3. t − s) dyds (x ∈ Rn . t) = U (x.1. s) dy v(x. t) = U (x.4) Z t Z |x−y|2 1 − 4(t−s) = f (y. ut are continuous in Ω × I. t. ∞). s) on Rn × {t = 0}. t) = Rn (x ∈ Rn . s) dy ds (x ∈ Rn . Then. . ∞)) satisfies (i) ut (x. e n/2 0 (4π(t − s)) Rn In the following. t) = f (x. In this way. (ii) for each x0 ∈ Rn . Rewriting. t > 0). t→0+ u(x. ∞)). 0) dy 0 Rn Rn and Z tZ uxi xj (x. lim x→x0 .3. By change of variables. Proof. t) = Φ(y. t > 0). j = 1. Theorem 3. t > 0). · · · . s) = Rn (x ∈ Rn .3) as Z K(x.3. y. t. t) = 0 Rn Φ(y. s)fxi xj (x − y.4). ∞)) has compact support in Rn × [0. t)f (x − y. t) = Φ(x − y. t) by (3. t − s)f (x. t) − ∆u(x. t − s) dyds + Φ(y. Du. t > s). s) dy u ˜(x. t) = 0. t)f (x. t − s). s)ds (x ∈ Rn . t > 0). Z tZ u(x. Then u ∈ C12 (Rn × (0. As f has compact support and Φ(y. t) will solve the Cauchy problem ( vt − ∆v = 0 in Rn × (0. Define u(x. Rn 0 1. Assume f ∈ C12 (Rn × [0. s)f (x − y. Fundamental solution of heat equation 59 Let u ˜(x. ∞) v(x. t > 0) 0 solves the nonhomogeneous problem above. Duhamel’s principle asserts that the function Z t u(x. t) = Φ(y. 0) = f (x. we obtain a solution to (3. n). t) (x ∈ Rn . t) such that u. 2. s) is smooth near s = t > 0. s)f (x − y. D2 u. we have Z tZ Z ut (x. t) = v(x. One solution of v to this problem is given above by Z K(x. This proves u ∈ C12 (Rn × (0. then v(x.

Later. t) u(x. 0) dy + Rn Z tZ = Φ(y. t) = Φ(x − y.4.2. RεR Now |Jε | ≤ C 0 Rn Φ(y. t)f (x − y. There are infinitely many solutions to Problem (3. t − ε) − f (x − y. Nonuniqueness for Cauchy problems. t) = lim (Iε + Jε + N ) ε→0+ Z Φ(y.  Corollary 3. t − s)] dyds Rn 0 Z Φ(y. note ku(·. t)kL∞ ≤ tkf kL∞ → 0 as t → 0+ . 0) = g(x) (x ∈ Rn . t)dy = f (x. The following result shows that Cauchy problems for the heat equation do not have unique solutions if we allow all kinds of solutions.1. t) − ∆u(x. we have Z tZ Iε = [(∂s − ∆y )Φ(y. t) = Φ(y. t)dy + Φ(y. s)dyds ≤ εC. s)[(−∂s − ∆y )f (x − y. s)[(∂t − ∆x )f (x − y. 3. ut (x. Therefore. t)f (x − y. ε)[f (x − y.5. Fundamental solution of heat equation 60 2.4. 0) dy Rn :=Iε + Jε + N. ε)f (x − y. t > 0). t − ε) − f (x − y. Z tZ ut (x. t > 0) (x ∈ Rn ). t)f (x − y. t)]dy − N. t − s)f (y.3. Finally. 3. t) − ∆u(x. t)dy + = lim ε→0+  Z Rn Φ(y. by (iii) of Theorem 3. t)]dy Rn Z Φ(y. ε)f (y. ε)[f (x − y. ε)f (x − y. Rn Rn since (∂s − ∆y )Φ(y. . ε)f (x − y. t − ε)dy − Φ(y. the function Z Z tZ u(x. 0) dy n Rn Z R Z = Φ(y. s)dyds Rn 0 Rn is a solution of ( ut − ∆u = f (x. t − s)] dyds + n Z0 R + Φ(y. Theorem 3. t) = lim ε→0+ Rn (x ∈ Rn .1). Under the previous hypotheses on f and g. ε)f (x − y. t)dy = lim ε→0+ Rn Z Φ(x − y.1. t − s) dyds n ε Z R Z + Φ(y. s)]f (x − y. s) = 0. t − s)] dyds ε Rn Z εZ Φ(y. s)[(−∂s − ∆y )f (x − y. we will show that the problem will have unique solution if the solution satisfies some growth condition. Integration by parts. t)g(y)dy + Φ(x − y.

First we have the weak maximum principle for sub-solutions. but is not analytic in t. (θt)k (b) For this function h. . 2. . t ∈ R.  Exercise 3. · · · ). T ]}. . Weak maximum principle and the uniqueness 61 Proof. We formally solve the following Cauchy problem  x ∈ R. 2. t) = h(t) for all t > 0. t ∈ (0. We define the parabolic boundary of ΩT to be ΓT = ∂ 0 ΩT := ΩT \ ΩT . . j=0 A formal computation from ut = uxx gives us h0 (t) = h(t).2. h1 (t) = 0. 0) = 0. . We have the Taylor expansion in terms of x. ΓT = ∂ 0 ΩT := (∂Ω × [0.  ut = uxx . prove that the function u given by (3. (a) Let α > 1 and h be defined by (3. . 3. T ] := {(x. ∞ X u(x. ∞) with u(x. is not identically zero since u(0. t ∈ (−∞. 3. that is. t > 0. T ]) ∪ (Ω × {t = 0}). Show that there is a constant θ > 0 such that k! − 1 t−α |h(k) (t)| ≤ e 2 (t > 0.1.7) above. We only need to construct a nonzero solution of the one-dimensional heat equation with 0 initial data.6) is a solution to the heat equation in R × (0. x2k . 1. Therefore h2k (t) = (3. Parabolic cylinders and the weak maximum principle. Actually. t) = 0. u(x. (3. t) = h(t). t ≤ 0. h0j (t) = (j + 2)(j + 1)hj+2 . t) | x ∈ Ω. . Consider the parabolic cylinder: ΩT = Ω × (0. u(0. 0) = 0.3 (not required). We choose for some α > 1 and define the function h(t) by ( 0. T > 0. k = 1. (2k)! u(x. called a Tychonoff solution. t) is an entire function of x for any real t.2. but this solution.6) 1 (k) h (t).7) h(t) = −α −t e . 1. . Weak maximum principle and the uniqueness More practical question is the existence and uniqueness of the mixed-value problems in a bounded open set. t ∈ R (3. t) = h2k+1 (t) = 0. We assume Ω is a bounded open set in Rn .3. j = 0. ∞). .5)  ux (0. t) = hj (t)xj .2. ∞ X h(k) (t) k=0 (2k)! k = 0. Then u given above is a honest solution of the heat equation and u(x.

t0 ) ∈ ΩT .) Then max u = max u. t).2. 0 0 ∂ ΩT ΩT ∂ ΩT So v + T ≤ max u + T. Weak maximum principle and the uniqueness 62 Theorem 3. 0) = g(x). this would limply vt − ∆v ≥ 0 at (x0 . 3. Let u ∈ C12 (Rn × (0. x ∈ Rn . The proof is standard and is left as an exercise. T ]. t). u(x. t0 ) ∈ ∂ 0 ΩT and so max v = max v ≤ max u. x ∈ ∂Ω. t) ∈ ΩT ). ut − ∆u ≤ 0. Theorem 3.6 (Weak maximum principle).2. a contradiction since vt − ∆v < 0. 2 u(x. Hence (x0 .2.2. consequently. t0 ) ≥ 0. a > 0. T ]) and satisfy   for 0 < t < T. Uniqueness of mixed-value problems. Then sup x∈Rn . Let u ∈ C12 (ΩT )∩C(ΩT ) and satisfy ut −∆u ≤ 0 in ΩT . Let v(x0 . t) ∈ ΩT .T ] u(x.  u(x. t0 ). t0 ) we have ∆v(x0 . max u ≤ max(v + t) ≤ max v + T = max 0 0 ΩT ΩT ∂ ΩT ΩT ∂ ΩT Setting  → 0+ .6 immediately implies the following uniqueness result for the initial-boundary value problem (or the mixed-value problem) of heat equation in ΩT . t) = h(x.7.  Similarly. The mixed-value problem   ut − ∆u = f (x. max u ≤ max 0 ∂ ΩT ΩT The opposite inequality is obvious and hence the equality follows. t) ≤ Aea|x| for 0 < t < T. at maximum point (x0 . t ∈ [0.3. t0 ) ≤ 0 and vt (x0 . 0 ∂ ΩT ΩT Proof. Then x0 ∈ Ω and t0 ∈ (0. (x. t∈[0. we say u is a subsolution of heat equation.8 (Weak maximum principle on Rn ). t0 ) = maxΩT v. T ]) ∩ C(Rn × [0. t) = sup g(z). Theorem 3. 0) = g(x) for x ∈ Rn for constants A. T ] can have at most one solution u in C12 (ΩT ) ∩ C(ΩT ). (In this case. We now extend the maximum principle to the region Rn × (0. 3. Maximum principle for the Cauchy problems. x ∈ Ω. z∈Rn . we deduce u. x ∈ Rn .3. Theorem 3. the weak minimum principle holds for the super-solutions satisfying ut − ∆u ≥ 0 in ΩT . Consider v = u − t for any  > 0. T ]. Then vt − ∆v = ut − ∆u −  ≤ − < 0 ((x.   u(x. Supose (x0 .

. i2 = −1 and Φ(x. t) − r2 µ 4(T +−t) e (T +  − t)n/2 r2 µ 4(T +−t) e (T +  − t)n/2 r2 µ 2 4(T +) e ≤ Aea(r+|y|) − (T + ))n/2 2 ≤ Aea|x| − 2 2 ≤ Aea(r+|y|) − µ(4(a + γ))n/2 e(a+γ)r . where a + γ = 4(T1+) for a positive number γ > 0. T ). Weak maximum principle and the uniqueness 63 Proof. r). t) = µ e 4(T +−t) satisfies wt = ∆w (T +−t)n/2 C is a constant. µ u(y. and consider v(x. Let ε > 0 be such that 4a(T + ) < 1. Comparing the coefficients of r2 inside both exponentials. x ∈ Rn 2 satisfying the growth condition |u(x. until we reach to time T . say. Rn Therefore. µ v(y. T +  − t). µ > 0. where is the fundamental solution of heat equation). 0) < u(x. t ∈ (0. t) = u(x. t) ≤ supRn g for all (x. then v(x. (T +  − t)n/2 Rn This being valid for all µ > 0 implies that u(y. t) = CΦ(i(x − y). ∂ 0 ΩT (T +  − t)n/2 Rn So.2. t) (or by noting w(x. t) − ≤ max v ≤ sup g (0 ≤ t ≤ T ). and hence in Rn × (0. We first assume that 4aT < 1. We have from the maximum principle that (3. The Tychonoff solution constructed by power series above cannot satisfy 2 the growth condition |u(x.4. If x ∈ ∂Ω. vt − ∆v ≤ 0 Now fix r > 0 and let Ω = B(y. Remark 3. we can choose r > 0 sufficiently large (depending on µ and y) so that 2 2 Aea(r+|y|) − µ(4(a + γ))n/2 e(a+γ)r ≤ sup g. by (3. t)| ≤ Aea|x| near t = 0. 0) = g(x) ≤ supRn g.  Theorem 3. t) ∈ ∂ 0 ΩT . 0 ≤ t ≤ T . t). t) = u(y.9. 3. a > 0. for some constants A. If x ∈ Rn . t ∈ (0. 2. T ). · · · . x ∈ Rn . u(x.3. 1. T ]) ∩ C(Rn × [0. by letting µ → 0+ . Finally. T1 = 8a . |x−y|2 A direct calculation shows that the function w(x. Hence. t) − |x−y|2 µ 4(T +−t) e (T +  − t)n/2 (x ∈ Rn . we can repeatedly apply the result above on intervals 1 [0. if 4aT ≥ 1. t)| ≤ Aea|x| for all x ∈ Rn . Fix y ∈ Rn . t) ≤ max v 0 ∂ ΩT (0 ≤ t ≤ T ). then v(x. 0) = g(x). where. T ]) to the Cauchy problem ut − ∆u = f (x. Now consider the circular cylinder ΩT .8) v(y. we have v(x. t) = u(x. T ]. T1 ].8). t) ≤ supRn g. There exists at most one solution u ∈ C12 (Rn × (0. [T1 . 0 ≤ t ≤ T ). 2T1 ]. t) ≤ sup g + .

12 (Backward uniqueness). t) = 0 for x ∈ ∂Ω. which is not easy itself at all. Theorem 3. Z Z Z e0 (t) = 2 uut (x. Energy method for uniqueness. Consider the mixed-value problem ( ut − ∆u = f in ΩT . t) in ΩT . Let e(t) be the energy associated with u. If u(x. 0) = g(x) for x ∈ Ω. T ]. then u ≡ 0 in ΩT . Ω Ω Ω which implies that e(t) ≤ e(0) = 0 for all t ∈ (0. Then u = u2 − u1 ∈ C12 (ΩT ) solves ( ut − ∆u = 0 in ΩT .11.2. t ∈ [0. Let u1 . 1.10. u2 be the solutions. Proof. u(x. t ∈ [0. u=g on ∂ 0 ΩT . T ). There exists at most one solution u ∈ C12 (ΩT ) of the following Neumann problem   ut − ∆u = f (x. We have already proved the uniqueness for the mixed-value problems of heat equation from the weak maximum principle. T ].3. t) Proof. Let Z e(t) = u2 (x. Ω Here we abused the name of “energy” since the L2 norm of the temperature has no physical meaning. t) dx (0 ≤ t ≤ T ). Theorem 3.   ∂u for x ∈ ∂Ω.  We can also use the energy method to prove the following backward uniqueness ressult. Hence u ≡ 0 in ΩT . t)dx = 2 u∆udx = −2 |∇u|2 dx ≤ 0. The proof is standard and is left as exercise. u(x. We set the “energy” at t associated with u to be Z (3. We have by Green’s identity. Proof. T ) = 0 for all x ∈ Ω. Suppose u ∈ C12 (ΩT ) solves ( ut − ∆u = 0 in ΩT . u=0 on ∂ 0 ΩT .2. Now we introduce another method to show the uniqueness concerning more regular solutions and domains. Weak maximum principle and the uniqueness 64 3. There exists at most one solution in C12 (ΩT ) of the mixed-value problem.9) e(t) = u2 (x. t) = h(x. t)dx.4. Assume Ω is a bounded smooth domain in Rn . Ω As before e0 (t) = −2 Z Ω |∇u|2 dx . ∂ν (x.  Theorem 3.

We use contradiction.  Remark 3. Ω Now Z Z Z 2 |∇u| dx = − Ω u∆u dx ≤ Ω 2 2 Z 2 1/2 (∆u) dx . The backward uniqueness theorem for the heat equation asserts that if two temperature distributions on Ω agree at some time T > 0 and have the same boundary values for times 0 ≤ t ≤ T . Then f 00 (t) = e00 (t)e(t) − e0 (t)2 ≥0 e(t)2 (t1 ≤ t < t2 ).5. 1. Suppose u ∈ C12 (ΩT ) solves the heat equation in ΩT . t0 . e(t2 ) = 0. t0 . r) = {(y. Fix (x0 . Regularity of solutions In order to establish the regularity of a solution of the heat equation in a bounded domain we use Green’s identity and the fundamental solution as we did for Laplace’s equation. there must exist an interval [t1 . t) ∈ Cc (R × R) such that ζ ≡ 0 outside C. t1 < t < t2 ). 3. 4 2 ∞ n Let ζ(x. Consider the closed circular cylinder C(x. Regularity of solutions 65 and hence Z 00 Z e (t) = −4 ∇u · ∇ut dx = 4 Ω (∆u)2 dx. t. t0 . t1 < t < t2 ) and so e((1 − τ )t1 + τ t) ≤ e(t1 )1−τ e(t)τ Letting t → t− 2. This is a contradiction since e(t) > 0 on [t1 . t2 ] ⊂ [0. ζ ≡ 1 on C 0 and 0 ≤ ζ ≤ 1. Suppose otherwise that e(t) 6≡ 0. r). T ]. Then u ∈ C ∞ (ΩT ). (0 < τ < 1.3.3. t2 ).3. . 0 ≤ e((1 − τ )t1 + τ t2 ) ≤ e(t1 )1−τ e(t2 )τ = 0 (0 < τ < 1). Consequently f ((1 − τ )t1 + τ t) ≤ (1 − τ )f (t1 ) + τ f (t) (0 < τ < 1. with e(t) > 0 on t ∈ [t1 . Ω 2. Set f (t) = ln e(t) for t ∈ [t1 . r) is contained in ΩT . t2 ). then these temperature distributions must be identical within Ω at all earlier times. They will remain the same until a time when the boundary data become different.13 (Smoothness). We show that e(t) = 0 for all 0 ≤ t ≤ T and then we are done. C 00 := C(x0 . r). since e(t2 ) = 0. s) | |y − x| ≤ r. Theorem 3. Ω 2 |∇u| dx (e (t)) = 4 2 u dx Ω Thus 0 1/2 Z ≤ e(t)e00 (t). t2 ). t0 ) ∈ ΩT and choose r > 0 so small that C := C(x0 . Consider two small circular cylinders 3 1 C 0 := C(x0 . Hence f is convex on [t1 . Proof. t − r2 ≤ s ≤ t}. Since e(T ) = 0. t2 ).

Then u is a C ∞ -solution to the heat equation in (ΩT ) and hence the formula (3. 0 ≤ t ≤ t0 ). r/2) ⊂ C(x.10) in the previous proof. let u = η *u be the mollification of u. then we have ZZ (3. s) is C ∞ in (x. we may assume the point is (0. Proof.l for k. t − s) · Dζ]u(y. t0 ]) and v = 0 on Rn × {t = 0}. t) ∈ C 00 ). Then we let  → 0 and deduce that (3. by uniqueness and Theorem 3. s) − 2Dζ · Du] dyds Z ZC = [Φ(x − y. We have derived the formula (3.10) proves u is C ∞ (C 00 ). t − s) · Dζ(y. y. s) dyds (x ∈ Rn . There exists a constant Ck. 0). t. s) dyds ((x. t. the formula (3. Theorem 3. ZZ u(x. Then. t − s)(ζs + ∆ζ) + 2Dy Φ(x − y.10) also holds for u in C 00 .3. t) (x ∈ Rn . t) = Γ(x. ∞). vt − ∆v = ζt u − 2Dζ · Du − u∆ζ := f˜ in Rn × (0. we have Z tZ Φ(x − y. r) ⊂ ΩT and all solutions u ∈ C12 (ΩT ) of the heat equation in ΩT . 1. t) ∈ ΩT .3. l = 0. t) ∈ C 00 . T ]. y.10) for C ∞ -solution u of the heat equation in ΩT . t) = Φ(x − y. t ∈ (0. Hence. v(x. Then ZZ u(x. s) = Φ(x − y.3.10) holds for u . C\C 0 3. s) dyds. but the following argument aims to establishing an identity that is valid for C12 -solutions once valid for C ∞ -solutions of the heat equation. Note that f˜ is C ∞ and has compact support in Rn × [0.t. t) ∈ C(1/2)). t0 ). 3/4) and C(1/2) = C(0. 2.  Let us now record some estimates on the derivatives of solutions to the heat equation. t0 )). y. t)u(x. s). t − s)(ζs + ∆ζ)(y. 0. 0.10) u(x. 1) ⊂ ΩT and let C(3/4) = C(0.r/2) for all cylinders C(x. Since Γ(x. s) dyds C ZZ = [Φ(x − y. C(1)\C(3/4) . Suppose first that the cylinder C(1) = C(0. If u is a C12 -solution. Regularity of solutions 66 2. s) ∈ C \ C 0 . by (3. moreover v is bounded on Rn × [0. t) = ζ(x. s) + 2Dy Φ(x − y. t − s)[(ζs − ∆ζ)u(y.t. C\C 0 Let Γ(x. y. 0. Furthermore. Let v(x. s)u(y. t − s)(ζs + ∆ζ) + 2Dy Φ(x − y.r)) (|α| = k) r C(x. We temporarily assume that u ∈ C ∞ (C).14 (Estimates on derivatives). t. t. y) = Γ(x. s) dyds ((x. t) = 0 Rn Let (x. Fix some point (x. t.l max |Dxα Dtl u| ≤ k+2l+n+2 kukL1 (C(x. t) ∈ C 00 and (y. t − s)f˜(y. Then v ∈ C ∞ (Rn × (0. s)u(y. s) dyds C ZZ = Φ(x − y. Upon shifting the coordinates. t − s)f˜(y. t − s) · Dζ]u(y. This seems contradicting to what we needed to prove. 1. 1/2). · · · such that Ck. t.

|x| ≤ ρ. This section we discuss yet another important uniqueness result due to D. l = 0. rn+2 |Dyα Dsl u(y. t) = u(rx. 0) = ζ (x)g(x). 0 ≤ t < T . ζ a (x) = a − |x| for a − 1 < |x| < a. t) = r2l+k Dyα Dsl u(rx. in terms of the growth condition. 0) ≤ g(x) ≤  + u(x. and let ut . t) ≥ 0. s)| ≤ Then. Now suppose C(r) = C(0. Then Let  > 0 and let ρ > a + 2M 2πe ( v a (x. R Note that ζ a (y)g(y) ∈ Cc (Rn ). t)g(y)dy. v a (x. t)dy ≤ √ .15 (Widder’s Theorem). t)ζ a (y)g(y)dy. s) ∈ C(r/2)). r) ⊂ ΩT and let C(r/2) = C(0. for each pair of k.3. 0 < t < T .l kukL1 (C(1)) 2l+k+n+2 r ((y. 0) = g(x). Dxα Dtl v(x. 1. a v (x. u(x. 0.11) applied to v. Nonnegative solutions From the example of Tychonoff solution above of heat equation. u(x. 0 < t < T. t) ∈ C(1)).  3. r2 t) and kvkL1 (C(1)) = 1 kukL1 (C(r)) . 0. t) ≤  ≤  + u(x. y. Using K(x. 0). R Proof.l . Widder for nonnegative solutions. s)|u(y. Some additional conditions are needed for uniqueness. This completes the proof. Assume that ut = uxx . t)| ≤ C(1)\C(3/4) for some constant Ck. ζ a (x) = 0 for |x| ≥ a. we know that the initial data can not determine the solution uniquely. as we discussed in the previous section. we have Ck. · · · and all |α| = k. For a > 1 define the cut-off function ζ a (x) by ζ a (x) = 1 for |x| ≤ a − 1. 2. |x| = ρ. 0 < t < T . r/2). s) dyds ≤ Ck. y. we know that a vta − vxx =0 a for x ∈ R. t) = K(x. Consider the expression Z v a (x. Let u be a solution to the heat equation in ΩT . t). we have for . ux and uxx exist and be continuous for x ∈ R. 2πe |x| − a −a √ a a . is real analytic and is represented by Z u(x. 2πe|x−y| a 2Ma a 1 K(x. t.11) |Dx Dt u(x. y. V. Note that for all |α| = k. t) = K(x. t) ≤ |x| > a. 2.4. ZZ α l |Dxα Dtl Γ(x. Then vt − ∆v = 0 in the cylinder C(1). y. Nonnegative solutions 67 Consequently. r2 t) ((x. for example. with (3. Theorem 3. Let Ma be the maximum of g(x) for |x| ≤ a.4. Then u is determined uniquely for x ∈ R. We rescale by defining v(x.l kukL1 (C(1)) (3. Let u be defined and continuous for x ∈ R. a 0 ≤ v (x) ≤ Ma Z √ 1 . y. 0 < t < T . 1.

consider W (x. t) is convex in x. s) ≤ e W (0. t) ≥ 0. W (x. hence. we have. s) = 0 for s ∈ [0. t) = v(x. 2. That is u(x. W (x. t). 0 0 0 From Step 1. Let w = u − v. wt = wxx . s) for x ∈ R. t). Therefore. T ). we reduced the problem to the case g = 0. Regularity of v(x. T − ).4. t) ≤  + u(x. Let ρ → ∞ and we find the same inequality for all x ∈ R. s). t) for |x| ≤ ρ.) We introduce the new function Z t W (x. is a non-decreasing bounded functions of a. s)dy ≥ p 2xW (x. v a (x. with some constants a.12) above with t = T − . t) ≥ 0 and w(x. Nonnegative solutions 68 By the maximum principle. x2 3. it follows that W (x. It remains to show that w ≡ 0. 0 ≤ t < T.12) W (x. t) ≥ K(0. x > 0. Z Z W (0. 0 ≤ t < T. s) ≤ ex 2 / W (0. s) = 0 for s ∈ [0. t) = Wt (x. Then w is continuous for x ∈ R. we deduce that r π(t − s) x2 /(t−s) (3. t) ≤ u(x. t)dH + W (x − H. Since  > 0 is arbitrary. s)ds. t) ≤ W (x + H. t) = lim v (x. Since ζa x ∈ R. Since W (x. t) ≤ W (x + H. A > 0. t)dH = W (y. t). 4π(t − s) The similar argument can also apply to the case x < 0. t)dy. t > s ≥ 0). y. Hence W satisfies the assumption of Theorem 3.8 for all x ∈ R and 0 ≤ s ≤ T − 2. 0 ≤ t < T and 0 ≤ v(x. t − s)W (y. t) = w(x. (That is. t) = K(x. for all t > s. s)dy ≥ R 2 /(t−s) −x e ≥p 4π(t − s) Z 0 2x K(0. 0 Then Wxx (x. we find that Z a v(x. Then W (x. 0) = 0. 2W (x. T − ). t) ≤ u(x. So W (x. s)dy 0 2 /(t−s) −x e 2x W (y. s) is bounded √ for |x| ≤ πT and 0 ≤ s ≤ T − 2. t) (x ∈ R. 0 ≤ t < T . Given x > 0. 0 ≤ s ≤ T − 2.3. t) = w(x. Setting  → 0 yields that v a (x. y. y.  . t)g(y)dy a→∞ R exists for x ∈ R. Using (3. t). for |x| ≥ πT and 0 ≤ s ≤ T − 2. t) + W (x − H. t − s)W (y. 0 ≤ t < T . Now for T > √  > 0. 0) = 0. w(x. t) can be obtained from the analyticity of v a (x. integrating the inequality with respect to H from 0 to x we find Z x Z x Z 2x 2xW (x. t) for any H > 0.

Mean-value property on the heat ball. whose boundary is a level set of Φ(x − y. s) in E(1).5. s) E(r) u(ry.5. u is a subsolution to the heat equation). s) = 1 for (y.3. t. 1. t − s). s) ∈ E(1)). Mean value property and the strong maximum principle 69 3. n |y|2 ln H(y. rn This is a bounded region in space-time. r) ⊂ ΩT . Theorem 3. Mean value property and the strong maximum principle In this section we derive for the heat equation some kind of analogue of the mean value property for harmonic functions.r) |x − y|2 dyds (t − s)2 for all E(x. s) ≤ ∆y v(y. Then vs (y.5. −s) = |y|2 1 4s e (−4πs)n/2 (y ∈ Rn . s) = u(ry. Definition 3. Fix x ∈ Rn . Set 1 φ(r) = rn ZZ = ZZ u(y.16 (Mean-value property for the heat equation).13) 4(ln H)s + 2n |y|2 =− . s) | Φ(x − y. Then 1 u(x. t = 0 and write E(r) = E(0.6. r2 s) ((y. 2s . s) ∈ ∂E(1). t) ≤ n 4r ZZ u(y. r) = {(y. Then H(y. 0. and (3. 2 s s ∇(ln H) = y .t. r > 0. 3. Proof. r). Let H(y. r2 s) E(1) |y|2 dyds s2 |y|2 dyds. s2 Let v(y.1. s) = − ln(−4πs) + ≥0 2 4s ((y. Let u ∈ C12 (ΩT ) satisfy ut − ∆u ≤ 0 in ΩT (that is. s) ∈ E(1)). s) E(x. s < 0). We define the heat ball E(x. s) = Φ(y. We may assume that x = 0. t − s) ≥ 1 }. t ∈ R. t.

ZZ Z 0 Z |y|2 dyds = 4π|y|2 e−τ dydτ 2 n s 2 τ E(1) −∞ |y| ≤− 2π τ e Z 0 Z √− n τ eτ 2π = 4πnαn rn+1 e−τ drdτ −∞ 0 Z n+2 n 4πnαn n n+2 0 2 = ( ) (−τ ) 2 e 2 τ dτ n + 2 2π −∞ Z 4πnαn n n+2 2 n+4 ∞ n+2 −t n = ( ) 2 ( ) 2 t 2 e dt (t = − τ ) n + 2 2π n 2 0 4nαn 2 1 n/2 n = ( ) Γ( + 2) (here Γ(s) is the Gamma-function) n+2n π 2 1 n/2 n n/2 = 2nαn ( ) Γ( ) = 4 (noting that nαn = 2π n ). =− r E(1) s Consequently. ut (ry. |y|2 ≤ 2ns ln(−4πs)}. r2 s) = r12 vs (y. |y|2 ≤ n − 2π τ eτ }. s2 1 4π Note that E(1) = {(y. s) → (y. s) | − ≤ s < 0. s)) 2 dyds = r s E(1)   ZZ 2 1 |y| = y · ∇v 2 + 4vs y · ∇(ln H) dyds r s E(1) "Z Z # ZZ 1 |y|2 = y · ∇v 2 dyds − 4 (nvs + y · (∇v)s ) ln Hdyds r s E(1) E(1) "Z Z # ZZ 1 |y|2 = y · ∇v 2 dyds − 4 (nvs ln H − y · ∇v(ln H)s ) dyds r s E(1) E(1) # "Z Z ZZ 2n 1 vs ln Hdyds (using (3. Therefore.13b)). ZZ  |y|2 0 φ (r) = ∇u(ry. s). r2 s) = 1r ∇v(y. It remains to show that ZZ E(1) |y|2 dyds = 4. 2 E(1) s 3. Mean value property and the strong maximum principle 70 2. the set E(1) is mapped ˜ ˜ one-to-one and onto a set E(1) in (y. ln H ≥ 0 in E(1)) r s E(1) E(1) ZZ y  2n · ∇v − 2∇v · ∇ ln H dyds = 0 (using (3.3. 0). φ is an increasing function of r. and hence ZZ |y|2 + φ(r) ≥ φ(0 ) = u(0. Γ( ) 2 π 2 This completes the proof. r2 s) · y + 2rsut (ry.13a)) =− y · ∇v dyds + 4n r s E(1) E(1) "Z Z # ZZ 1 2n ≥− y · ∇v dyds + 4n ∆v ln Hdyds (as vs ≤ ∆v. τ ). 0) dyds. where τ = ln(−4πs) and hence s = − 4π . s). We calculate φ0 (r) as follows. τ ) | τ ∈ (−∞. Under the change of eτ variables (y. s) + 2svs (y.  .5. Note that ∇u(ry. r2 s) dyds s2 E(1) ZZ |y|2 1 (y · ∇v(y. τ ) space given by E(1) = {(y.

r0 ) = M for some point (z0 .  Remark 3. Select times t0 > t1 > · · · > tm = t such that the line segments Li connecting (xi−1 . xm = x} such that the line segments connecting xi−1 to xi lie in Ω for all i = 1. u ≡ M on Li and hence u(x. Since u is continuous. this is a contradiction to r0 being the infimum.3. 2. 2. t0 ) ≤ n u(y. Let Ω be a bounded domain in Rn . We prove r0 = s0 .) 3. ti ) lie in ΩT . Maximum principles for second-order linear parabolic equations (This material is from Section 7. t) = M for all points (x. t) = M for all (x. Assume that u ∈ C12 (ΩT ) ∩ C 0 (ΩT ) satisfies ut − ∆u ≤ 0 in ΩT (that is. r)). (The following argument is not needed if Ω is convex. ti−1 ) to (xi . s) dyds ≤ dyds = M. Strong maximum principle. · · · . r). Let M = u(x0 . we see that u(y. m. t) ∈ L. 1. r0 .6. Now fix any x ∈ Ω and 0 ≤ t < t0 . r) for all sufficiently small r. Hence r0 = s0 and so u ≡ M on L.2. u is a subsolution to the heat equation) and there exists a point (x0 . s) ∈ E(x0 . s) ≡ M on E(x0 . Then for all sufficiently small r > 0. Note that E(z0 . t0 . According to Step 2. t0 ] | u(x. 2 4r (t0 − s)2 4rn E(x0 . (A positive initial local source of heat will generate the positive temperature immediately afterwards everywhere away from the boundary. then u is positive everywhere within ΩT .1 of the textbook.) There exist points {x0 . Then u(z0 . t0 ) ∈ ΩT such that u(x0 . s) ≡ M ((y. Draw any line segment L in ΩT connecting (x0 . r0 . (b) If a solution u to the heat equation attains its maximum (or minimum) value at an interior point (x0 .6. From the previous argument. t) ∈ ΩT . t0 ) to another point (y0 .t0 . r) contains L ∩ {r0 − σ < t ≤ r0 } for some σ > 0. E(x0 . and we employ the mean value theorem to get ZZ ZZ 1 |x0 − y|2 |x0 − y|2 M M = u(x0 .r) E(x0 .r) (t0 − s) As equality holds only if u(y.5. Consider r0 = inf{s ∈ [s0 . (a) The weak maximum principle follows from the strong maximum principle above. ΩT Then u is constant in Ωt0 . T ]. · · · . t0 .) . t0 ) then u is constant at all earlier times t ≤ t0 . If the initial data u(x. 0) = g(x) is nonnegative and is positive somewhere. u(x. the solution may change at later times t > t0 if the boundary conditions alter after t0 . s0 ) in ΩT with s0 < t0 . The solution will not respond to changes of the boundary data until these changes happen. (c) Suppose u solves the heat equation in ΩT and equals zero on ∂Ω×[0.17 (Strong Maximum Principle). This is another illustration of infinite propagation speed of the disturbances of initial data for the heat equation. t0 ) = max u. Suppose for the contrary that r0 > s0 . Maximum principles for second-order linear parabolic equations 71 3. However. t0 .7.t0 . s ≤ t ≤ t0 }. the infimum r0 is attained. r) ⊂ ΩT . Theorem 3. x1 . t0 ) = maxΩT u. Proof. t) = M on E(z0 . r0 ) ∈ L ∩ ΩT .

t0 ) = maxΩT v > 0 for some (x0 . Then (a) max u = max u 0 (b) if c(x.j=1 n X ξk2 .j=1 which is a contradiction. t) ∈ ΩT .j=1 aij vxi xj + cv ≥ cv ≥ 0 at (x0 . this proves (3. T ]. consequently at maximum point (x0 . (vxi xj (x0 . 3. t0 )) ≤ 0 (in matrix sense) and vt (x0 . i. t0 ) ∈ ΩT then x0 ∈ Ω and t0 ∈ (0. let v(x0 . t)u.14)(b) is obviously valid if maxΩT v ≤ 0. To prove (3. t) ∈ ΩT . t)t] ≤ − < 0 ((x.3. t) = 0 in ΩT . If (x0 . k=1 If λ(x. T ]. The operator ∂t + L is called parabolic in ΩT if there exists λ(x. t)ξi ξj ≥ λ(x. t) i. Weak maximum principle.18 (Weak maximum principle). t) ≥ 0 in ΩT . We show that (a) (3. we assume that aij (x. t0 ) we have Dv(x0 . ξ ∈ Rn . Therefore. t). t0 ). t0 ). Proof. Let v(x0 . t) = aji (x. t0 ) ≥ 0. t0 ) ∈ ΩT .14) max v = max v 0 ∂ ΩT ΩT (b) u+ max v ≤ max 0 ∂ ΩT ΩT if c(x. t)uxi + c(x. vt + Lv = ut + Lu − [ + c(x. i=1 where aij . t0 ) = 0. t) ∈ ΩT ). (vxi xj (x0 . t) > 0 such that n X aij (x. t) = 0 in ΩT .14)(a). t0 ) = maxΩT v for some (x0 . we must have (x0 . Assume ∂t + L is parabolic in ΩT . t0 ) ∈ ΩT then x0 ∈ Ω and t0 ∈ (0. we say the operator ∂t + L is uniformly parabolic in ΩT .14)(a). Note that in both cases of (a) and (b) above. . Theorem 3. So we assume maxΩT v > 0. ∂ ΩT ΩT max u ≤ max u+ 0 ∂ ΩT ΩT if c(x.1. ∀ (x. t0 ) ≥ 0. we say that u is a subsolution of ut + Lu = 0). t0 ) ∈ ∂ 0 ΩT . 1. Let u ∈ C12 (ΩT ) ∩ C(ΩT ) satisfy ut + Lu ≤ 0 in ΩT (in this case. where Ω is a bounded open set in Rn . bi and c are all bounded functions in ΩT = Ω × (0. as vt + Lv ≤ − on ΩT .6. Fix  > 0 and let v = u − t. T ]. consequently at maximum point (x0 . Note that (3.6. t0 )) ≤ 0 (in matrix sense) and vt (x0 . if c(x. t0 ) ∈ ΩT . t) ≥ 0 in ΩT . t) ≥ λ0 > 0 for all (x. Maximum principles for second-order linear parabolic equations 72 We consider the second-order linear differential operator ∂t + L defined by ut + Lu = ut − n X aij (x. and hence vt + Lv = vt − n X aij vxi xj ≥ 0 at (x0 . t0 ) = 0.j=1 n X bi (x. and hence vt + Lv = vt − n X i. t)uxi xj + i. Without the loss of generality. If (x0 . t0 ) we have Dv(x0 .

Theorem 3.19 (Weak maximum principle for general c). Then max u ≤ eCT max u+ . t) = 0 in ΩT . hence max v = max v ≤ max u ≤ max u+ . we must have (x0 . 0 ∂ ΩT ΩT where C is any constant such that c(x.14). t) + C ≥ 0 in ΩT . For example. t) ∈ ΩT . if c(x. ˜ where Lw ˜ = Lw + Cw. t) = 0 in ΩT . max u = max(eCt w) ≤ eCT max w ≤ eCT max w+ ≤ eCT max u+ . t) = c(x. t) ≥ 0 in ΩT . Therefore. u(x. by the previous theorem. max w ≤ max w+ . t) ≥ 0 in ΩT . Clearly the equality holds in the case of (a) above.   u(x. t0 ) ∈ ∂ 0 ΩT .6. t) = h(x. 0 0 ΩT ΩT ∂ ΩT ΩT ∂ ΩT  The weak maximum principle for general c(x. Assume ∂t + L is parabolic in ΩT . Theorem 3. Maximum principles for second-order linear parabolic equations 73 which is also a contradiction.3. C = − inf ΩT c− (x. t ∈ [0. t) implies the uniqueness of mixed-value problem for parabolic equations regardless of the sign of c(x. as vt + Lv ≤ − on ΩT . Setting  → 0+ in both cases. x ∈ Ω. Let u ∈ C12 (ΩT ) ∩ C(ΩT ) satisfy ut + Lu ≤ 0 in ΩT (that is. Consider w(x. So we assume maxΩT u > 0. 0 ∂ ΩT ΩT Consequently. 0 0 0 ∂ ΩT ΩT ∂ ΩT ∂ ΩT 2. Proof. (x. (a) max u ≤ max 0 ∂ ΩT ΩT (b) max u ≤ max u+ 0 ∂ ΩT ΩT if c(x. t). t). T ] can have at most one solution u in C12 (ΩT ) ∩ C(ΩT ). Hence. since maxΩT u > 0. t) = e−Ct u(x. we deduce that max u = max(v + t) ≤ max v + T = max v + T ≤ max u + T 0 0 ΩT ΩT ∂ ΩT ΩT ∂ ΩT max u = max(v + t) ≤ max v + T ≤ max u+ + T 0 ΩT ΩT ∂ ΩT ΩT if c(x. t). c˜(x. is parabolic and has the zeroth order coefficient The operator ∂t + L. we deduce u if c(x. u is a subsolution of ut + Lu = 0).20.  We have the following result for general c(x. The inequality is obviously valid if maxΩT u ≤ 0. By (3. t). t). x ∈ ∂Ω. t). . 0) = g(x). Then wt + Lw = e−Ct (ut + Lu) − Ce−Ct u = e−Ct (ut + Lu) − Cw and hence wt + (Lw + Cw) = e−Ct (ut + Lu) ≤ 0 in ΩT . t) + C ≥ 0 in ΩT . The mixed-value problem   ut + Lu = f (x.

x2 ∈ V. By symmetry (or applying the maximum principle to −u). Then we have the following Theorem 3. t2 ]. t2 ) − v(x1 . st2 + (1 − s)t1 )ds ds 1 [(x2 − x1 ) · Dv + (t2 − t1 )vt ]ds = 0 Z ≥ 1 [(t2 − t1 )vt − |Dv|2 ] ds − C(). t1 ) = 0 Z d v(sx2 + (1 − s)x1 . we may assume that u > 0 (otherwise consider u +  and let  → 0) and assume that V is a ball since in general V is covered by a finite number of balls. 0 Therefore to show (3. Proof. t). t2 ).e. 0 < t1 < t2 ≤ T ).3. t) = ln u(x. u2 be two solutions of the problem in C12 (ΩT ) ∩ C(ΩT ). t1 . 0 ΩT ∂ ΩT Hence u ≤ 0 in ΩT . We consider the special case when the operator ∂t + L is given simply by n X ut + Lu = ut − aij (x. Let v(x.15). Assume V ⊂⊂ Ω is connected. t2 and the coefficients of L such that sup u(. t)ξi ξj ≤ Λ|ξ|2 ((x. Dij v = − = − Di vDj v. (x1 . A direct computation shows that Dj u Dij u Di uDj u Dij u Dj v = . Define u = u1 − u2 . u u ij .. The idea is to show that for some γ > 0 v(x2 .. proving the theorem. there is a constant C depending only on V. 2 u u u u P X ut ij aij Dij u vt = = = (aij Dij v + aij Di vDj v) := α + β. t)Dij u ij=1 and is uniformly parabolic in ΩT . 2. u1 ≤ u2 in ΩT . for 0 < t1 < t2 ≤ T . ij We also assume aij is smooth on ΩT .16) vt ≥ ν|Dv|2 − γ in V × [t1 .15) This will imply u(x1 . t) ∈ ΩT . there are positive constants λ and Λ such that X λ|ξ|2 ≤ aij (x. Hence max u ≤ eCT max u+ = 0. Let u1 . t2 ) − v(x1 .21. t2 ) V V whenever u ≥ 0 is a smooth solution of ut + Lu = 0 in ΩT . Then u ∈ C12 (ΩT ) ∩ C(ΩT ) solves ut + Lu = 0 in ΩT and satisfies u = 0 on ∂ 0 ΩT .  3.6. that is..2. Then. t1 ) ≥ −γ (3. we have u2 ≤ u1 in ΩT . Note that Z 1 v(x2 . ξ ∈ Rn ).6. This proves u1 = u2 in ΩT . 1. t1 ) ≤ eγ u(x 2 . it suffices to show that (3. Let Ω be a smooth and bounded domain of Rn . Without loss of generality. Maximum principles for second-order linear parabolic equations 74 Proof. i. Harnack’s inequality and the strong maximum principle. t1 ) ≤ C inf u(.

with suitable choice of  > 0 in (3.17).19) αt − X ij aij Dij α + n X bi Di α ≥ λ2 |D2 v|2 − C|Dv|2 − C. Note that. kl Dk β = 2 kl X aij Dik vDj v + X ij Dkl β = 2 X Dk aij Di vDj v. to show vt = α + β ≥ ν|Dv|2 − γ.17) ( > 0). we only need to show that w := α + κβ ≥ −γ for some κ ∈ (0. 3. Notice that Dij vt = Dij α + Dij β and hence X αt = [aij Dij vt + aij. ijkl where R is also a term satisfying (3.17). β(x. Maximum principles for second-order linear parabolic equations 75 where X α(x.18) akl Dkl β = 2 kl X akl aij Dikl vDj v + 2 ijkl =2 X X aij akl Dik vDjl v + R ijkl aij Di αDj v + 2 ij X aij akl Dik vDjl v + R. there exists a constant C > 0 such that (3. . t) = X ij aij Di vDj v. t) = aij Dij v. ij Note that β ≥ λ|Dv|2 . To this end.6. ijkl where R is also a term satisfying (3.17). Hence X (3.t Dij v] ij = X = X aij Dij α + ij X aij Dij β + ij aij Dij α + 2 aij. by the uniform parabolicity and linear algebra. we calculate X X Di α = akl Dikl v + Dkl vDi akl . ijkl Hence. ij aij Dikl vDj v + 2 X ij aij Dik vDjl v + R. 1/2) to be determined later. i=1 where bi = −2 n X j=1 aij Dj v. Hence.3. ij where R is a term satisfying |R| ≤ |D2 v|2 + C()|Dv|2 + C (3. X aij akl Dik vDjl v ≥ λ2 |D2 v|2 .t Dij v ij X ij X aij Dj vDi α + 2 X ij aij akl Dik vDjl v + R.

· · · . t0 ). (ζ 4 ) kl −4wζ 2 D where |R1 | ≤ Note also that Dk wDl = k ζDl ζ and |bk | ≤ C|Dv|. kl k Note that Ht = µ + ζ 4 wt + w(ζ 4 )t .21).t Di vDj v + 2 aij Di vt Dj v − akl Dkl β ij kl ij =2 X =2 X kl aij Dj v(Di vt − Di α) − 2 ij X aij akl Dik vDjl v + R ijkl aij Dj vDi β − 2 ij X aij akl Dik vDjl v + R. i kl By (3. t2 ] and ζ = 0 on ∂ 0 ΩT . t0 ) < 0 and hence w = α + κβ < 0. by (3. |w(x0 . t0 ) ∈ ΩT such that H(x0 . 2 k=1 4. 1/2) sufficiently small such that n X λ2 2 2 (3. Maximum principles for second-order linear parabolic equations 76 To derive a similar inequality for β(x. t0 ). t0 )| ≤ C|D2 v|(x0 .21) wt + Lw + bk Dk w ≥ |D v| − C|Dv|2 − C. 2. at (x0 . kl k k Cζ 2 |w|.23) 0≥µ+ζ 4  λ2 2 2 |D v| − C 4  + R2 . t). (3. which implies that w + µt ≥ 0 in V × [t1 . and so by (3. t0 ). We now prove this claim.22) (3. t0 ) ≤ C|D2 v|(x0 . we can choose κ ∈ (0.20) βt − akl Dkl β + bi Di β ≥ −C(|D2 v|2 + |Dv|2 + 1).20). Dk H = ζ 4 Dk w + wDk (ζ 4 ). t) = ζ 4 w + µt ≥ 0 in ΩT .22) 0≥µ+ζ 2 where |R2 | ≤ C(ζ 2 |w| + ζ 3 |w||Dv|). Claim: H(x. ijkl where R is a term satisfying (3. Hence. at (x0 . Therefore. This proves the theorem.18) X X X X βt − akl Dkl β = aij. w ≥ −µt2 > −γ.6. Dkl H = ζ 4 Dkl w + Dl (ζ 4 )Dk w + Dk (ζ 4 )Dl w + wDkl (ζ 4 ). we compute by (3. then there is a point (x0 .17). t0 ). with the same bi defined above. Hence. we have |Dv|2 (x0 . . Choose a cutoff function ζ ∈ C ∞ (ΩT ) such that 0 ≤ ζ ≤ 1 and ζ ≡ 1 on V × [t1 . t2 ]. 5. So. X X (3. Let µ be a large positive constant. Suppose not.  2  4 λ 2 2 2 |D v| − C|Dv| − C + R2 . At this minimum point. ζ 6= 0. ζDk w = −4wDk ζ (k = 1.3. Since H(x0 .19) and (3. n) and ! X X X X 0 ≥ µ + ζ 4 wt − akl Dkl w + bk Dk w + bk wDk (ζ 4 ) − 2 akl Dk wDl (ζ 4 ) + R1 . for the function w = α+κβ. X X Dk H = ζ 3 (4wDk ζ + ζDk w) = 0. Ht − akl Dkl H + bk Dk H ≤ 0. t0 ) < 0 is the minimum of H on ΩT .

we can derive the Harnack’s inequality for elliptic equations. V. this implies w ≡ 0 in Wt0 . with x0 ∈ W. we have |R2 | ≤ Cζ 2 |D2 v| + Cζ 3 |D2 v|3/2 ≤ ζ 4 |D2 v|2 + C().6. Since V is arbitrary. t) ≥ 0 in Wt0 . 3. that aij (x. for any connected subdomain V ⊂⊂ Ω. (ii) Similarly if ut + Lu ≥ 0 in ΩT and u attains its minimum over ΩT at a point (x0 . Maximum principles for second-order linear parabolic equations 77 where |R2 | ≤ C(ζ 2 |w| + ζ 3 |w||Dv|). Then wt + Lw = 0. w(x. we have the following theorem. Then any solution u(x) to Lu = 0 is a time-independent solution of parabolic equation ut +Lu = 0. V ) inf u(x) x∈V x∈V whenever u ≥ 0 is a smooth solution to Lu = 0 in Ω. t0 ) = 0. Now let w = M − v(x. 2. then u must be a constant on Ωt0 . t0 ) = u(x0 . we have u(x. Let 0 < t < t0 . Then. t. And we can also deduce the following stronger version of maximum principle.23). We prove (i) as (ii) can be shown by replacing u with −u. Choose any connected V ⊂⊂ W with x0 ∈ V. Therefore. Let v be the solution of vt + Lv = 0 in WT and v|∂ 0 WT = u|∂ 0 WT . t0 ) in Wt0 . Corollary 3. (We assume the existence of such a solution. t) ≤ C inf w(y.j=1 be uniformly elliptic in Ω with smooth aij . Proof. we have u = M on ∂ 0 Wt0 . t0 ) ∈ ΩT . t).3. Theorem 3. t0 ) ∈ ΩT . 1. The above conclusion holds for arbitrary open sets W ⊂⊂ Ω. where M := maxΩT u = u(x0 . Select any smooth. in the operator L above. t) ≤ M for all (x. (i) If ut + Lu ≤ 0 in ΩT and u attains its maximum over ΩT at a point (x0 . then u must be a constant on Ωt0 . and therefore u ≡ M on Ωt0 . V ) > 0 such that sup u(x) ≤ C(L. there exists a constant C(L. t0 ) such that 0 ≤ w(x. which is a contradiction to (3. t) ∈ WT . t0 ). t0 ) = Cw(x0 .23 (Strong maximum principle). y∈V This implies that w ≡ 0 in Vt0 . open set W ⊂⊂ Ω. t) = aij (x) is independent of t. Suppose. Assume that u ∈ C12 (ΩT ) ∩ C 0 (ΩT ).  From this Harnack’s inequality. t0 ). Let Lu = − n X aij (x)Dij u i. Since v = u on ∂ 0 WT . But therefore v ≡ M = u(x0 . t) ≤ v(x.22 (Harnack’s inequality for elliptic equations). At (x0 . From this we deduce that v(x0 . provided µ > 0 is sufficiently large.  . using Young’s inequality with . Using Harnack’s inequality we have a constant C = C(L.) Then by the weak maximum principle. t0 ) = M is maximum of v.

So complete the arguments that are left in lectures. t)uxx + 2b(x. t)u ((x. t) | 0 < x < 1. z ∈ Rn . Z K(x. y ∈ Rn . s > 0). t > 0). (1) (10 points) Write down an explicit formula (without proof) for a solution of the Cauchy problem ( ut − ∆u + cu = f in Rn × (0. t) = (4πt)−n/2 e−|x−y| 2 /4t (x. (5) (10 points) Let ΩT = {(x. 0 < t ≤ T }. 0) = g(x) (x ∈ Rn ) has a unique solution u satisfying |u(x. u(x. t) = √ e h(s) ds 4π 0 (t − s)3/2 for a solution of the initial/boundary-value problem of the heat equation ut −uxx = 0 in the quadrant x > 0. Homework # 5. where a. 7. t + s) (x. Show directly without quoting maximum principles that |u(x. y. Show that the Cauchy problem ut − ∆u = 0 in Rn × (0. Chapter 7: Problems 1. t) K(y. ∞). t)}. where C is a constant. 0 ∂ ΩT where C = max{0. t) = h(t). ∞) → R with h(0) = 0. t > 0. c are given bounded functions in ΩT and a > 0 in ΩT . t) ≤ (n/2eπ)n/2 |x − y|−n (x. ∞) u=g on Rn × {t = 0}. t)ux + c(x. derive the formula Z t x2 1 x − 4(t−s) u(x. z. Also try working on the following problems related to the covered materials. Chapter 2: Problems 12–17. t > 0). (3) (10 points) Let K(x. t) be the heat kernel defined by K(x. Let u ∈ C12 (ΩT ) ∩ C(ΩT ) be a solution of ut = a(x. b. y ∈ Rn .Suggested exercises 78 Suggested exercises Materials covered are from Chapters 2 and 7 of the textbook with some new materials added. 0) = 0 (x > 0). 8. y. z. (2) (10 points) Let g ∈ C(Rn ) be bounded and satisfy g ∈ L1 (Rn ). t)| ≤ eCT max |u|. where c ∈ R is a constant. t) ∈ ΩT ). Rn (4) (10 points) Given a continuous function h : [0. y. and let ∂ 0 ΩT be the parabolic boundary of ΩT . s) dy = K(x. y. Prove (a) (b) K(x. supΩT c(x. t ≥ 0). u(0. . t > 0 satisfying the conditions: u(x. t)| ≤ C(1 + t)−n/2 (x ∈ Rn .

t). T ] can have at most one solution u in C12 (ΩT ) ∩ C(ΩT ). the maximum principles and uniqueness theorem for linear parabolic equations. t). given any functions f. (Hint: You may use. u(x. t ∈ [0. t) = h(x. Assume B : Rn → R and σ : R → R are smooth functions. x ∈ Ω. without proof. the following initial/boundary value problem   (x.) . Show that.   u(x. x ∈ ∂Ω.Suggested exercises 79 (6) (10 points) Let Ω be a bounded open set in Rn and 0 < T < ∞. 0) = g(x). g. h. t) ∈ ΩT . ut − ∆u + B(∇u) + σ(u) = f (x.

For example. V ∂V This identity is true for any region.2) utt − ∆u = f (x. Derivation of the wave equation The wave equation is a simplified model for a vibrating string (n = 1). Let V represent any smooth subregion of Ω.1) and the nonhomogeneous wave equation (4.1. Newton’s law says (assume the mass is 1) that Z Z utt = F · νdS. 4. the unknown function u : Ω × [0.Chapter 4 Wave Equation In this part we investigate the wave equation utt − ∆u = 0 (4. dt2 V V and the net force is Z F · νdS. Here x ∈ Ω ⊂ Rn . t) subject to appropriate initial and boundary conditions. ∞) → R. ν is the unit outnormal on ∂V . We shall discover that solutions to the wave equation behave quite differently from solutions of Laplace’s equation or the heat equation. or elastic solid (n = 3). hence the divergence theorem tells that utt = div F. t) represents the displacement in some direction of the point at time t ≥ 0. t > 0. The acceleration within V is then Z Z d2 udx = utt dx. 80 . these solutions are generally not C ∞ and exhibit finite speed of propagation. ∂V where F denoting the force acting on V through ∂V . In this physical interpretation u(x. membrane (n = 2).

for a classical solution u. (i) From d’Alembert’s formula (4.e. h. ut (x. 2 2 x−t Remark 4. 0) = g(x). The physical interpretation strongly suggests it will be mathematically appropriate to specify two initial conditions: initial displacement u(x. We derive the solution u in terms of g. 0) = ut (x. 0) = h(x) on R × {t = 0}. F = F (Du). that is. So. Note that utt − uxx = (ut − ux )t + (ut − ux )x = vt + vx . this is the case when g ∈ C 2 and h ∈ C 1 . the resulting equation is the wave equation. and so we obtain utt − a∆u = 0. t) = a(x − t).4. 0).4). One-dimensional wave equations and d’Alembert’s formula 81 For elastic bodies.4) u(x.2. 0) = g(x). where v = ut − ux .3) u(x. function v solves the initial value problem for linear transport equation: vt + vx = 0. t) = g(x + t) + v(x − (s − t). v(x. (4. i. One-dimensional wave equations and d’Alembert’s formula We first focus on the following Cauchy problem for 1-D wave equation: ( utt − uxx = 0 in R × (0. we have v(x.1. t) = a(x − t). if F and G are C 2 . the function u above is indeed a C 2 solution to the wave equation. Again this is an initial value problem for a nonhomogeneous linear transport equation of u.. F is a function of Du. 4. .2. s) ds 0 Z t = g(x + t) + a(x − s + t − s) ds 0 Z 1 x+t a(y) dy = g(x + t) + 2 x−t Z 1 x+t = g(x + t) + [h(y) − g 0 (y)] dy. 0) and initial velocity ut (x. Solving this problem. ut − ux = v(x. t) = F (x + t) + G(x − t) for appropriate functions F and G. 2 x−t from which we deduce the d’Alembert’s formula: Z g(x + t) + g(x − t) 1 x+t + (4. we use the linearization F (Du) ≈ aDu to approximate F. t) = h(y) dy. 0) − ux (x. Conversely. where g. any solution u to the wave equation utt − uxx = 0 has the form u(x. h are given. For small u and small Du. ∞). 0) = h(x) − g 0 (x) := a(x). When a = 1. u(x. with speed 1 and travelling to right and left. Solving this problem gives Z t u(x. Note that F (x + t) and G(x − t) are called traveling waves.

1 (Parallelogram Property).  utt − uxx = 0 u(x. One-dimensional wave equations and d’Alembert’s formula 82 (ii) Even when initial data g and h are not smooth. where functions g. which is called the domain of influence of x0 . ∞) of slope 1 or −1. Therefore. solutions to the onedimensional wave equation do not lose the regularity from initial data. which. Lemma 4. t) = k(t) t > 0. with A and C being two opposite points. We easily have the following property.4. k satisfy certain smoothness and compatibility conditions. h. t > 0}. Solve the initial-boundary value problem  x > 0.5) u(A) + u(C) = u(B) + u(D).1. t) | x0 − t < x < x0 + t. Note that. if initial data g ∈ C k and h ∈ C k−1 .4) defines a (weak) solution to the Cauchy problem. t0 ). 0) = g(x). (See Figure 4. where ABCD is any parallelogram in R × [0. t0 ) depends only on the initial data from x0 −t0 to x0 +t0 . t > 0. Thus the wave equation does not have the smoothing effect as does the heat equation.2. . we still consider the formula (4.  u(0. ut (x. Parallelogram Property Example 4. However. 0) = h(x) x > 0. (iii) The value u(x0 . The Cauchy data at (x0 . [x0 − t0 .) t D C A B x 0 Figure 4. Any solution u of the 1-D wave equation satisfies (4. as we shall see later.1. 0) influence the value of u in the region I(x0 ) = {(x. then u ∈ C k but is not in general smoother. x0 + t0 ] is called the domain of dependence for the point (x0 . is not the case for the higher dimensional wave equation.2.

) − u( . we need certain conditions in order that this formula does define a smooth solution in the domain x > 0. t) = g(x + t) − g(t − x) 1 R t+x  k(t − x) + + 2 t−x h(y)dy 2 (x ≥ t ≥ 0). t) is given by (4.) If (x.2. t) := A is in the region (II). t − x) + u( . (0 ≤ x < t). Of course. x) = + 2 2 Z 2x h(y)dy. then we can use the parallelogram property to obtain u(x. t) D Region (I) B C E := (x. k.4.4): Z g(x + t) + g(x − t) 1 t+x u(x. g(2x) + g(0) 1 u(x. on x = t. Mixed-value problem on x > 0. t > 0. 2 2 t−x Therefore. then u(x. the solution u to the problem is given by  R   g(x + t) + g(x − t) + 1 t+x h(y)dy 2 x−t 2 (4. t) := E is in the region (I).)  . that is. (See Figure 4.6) u(x.2. t > 0 Solution. t) x 0 Figure 4. that is x ≥ t. (Exercise: derive such conditions on g. 2 2 x−t In particular. ) 2 2 2 2   Z Z g(x + t) + g(0) 1 x+t g(t − x) + g(0) 1 t−x = k(t − x) + + h(y)dy − + h(y)dy 2 2 0 2 2 0 Z t+x g(x + t) − g(t − x) 1 = k(t − x) + + h(y)dy. t) = u(B) + u(D) − u(C) x+t x+t t−x t−x = u(0. t) = + h(y)dy.2. h. 0 ≤ x < t. 0 If (x. One-dimensional wave equations and d’Alembert’s formula 83 t Region (II) A := (x.

we consider (4. u(x.4. . Solve the initial-boundary value problem  (x. Tj (t) = aj cos(jt) + bj sin(jt). we find that λ = −j 2 for all j = 1. π).3. t > 0. 2.3. π) × (0. Another way to solve this problem is using the method of separation of variables. From X 00 (x) = λX(x). and Xj (x) = sin(jx). t) = 0.7) u(x. (See Figure 4. . Mixed-value problem on 0 < x < π. In all other regions we use the formula (4. t > 0 Example 4. t) ∈ (0.) We can solve u in the region (I) by formula (4. .4). ∞) satisfying the boundary condition that are in the form of u(x. ut (x. t) = u(π. t) = X(x)T (t). t) = ∞ X j=1 [aj cos(jt) + bj sin(jt)] sin(jx). X(0) = X(π) = 0. 0) = h(x). X(0) = X(π) = 0. 0) = g(x). π) × R+ . To make sure that u satisfies the initial condition..  utt − uxx = 0. One-dimensional wave equations and d’Alembert’s formula 84 t (IV) (II) (III) A D (I) B C E π 0 x Figure 4. First try find the solutions of wave equation in (0. x ∈ (0.  u(0. T (t) X(x) where λ is a constant. which implies that X 00 (x) T 00 (t) = = λ. Solution. Then we should have X 00 (x)T (t) = T 00 (t)X(x).2. .5).3.

for any fixed x ∈ Rn . Let u ∈ C m (Rn × [0. ∞)) with m ≥ 2 solve ¯+ ×R ¯ + ) and solves the Cauchy problem (4. that is. ut (x.3. (4. (4. ∂B(x. n B(x. t) = − ∆u(y. after some computations. 0) = g(x).r) Z H(x.r) Lemma 4. we compute.r) From this we deduce that limr→0+ Ur (x. for r > 0. and we have u(x.8). and this proves that U ∈ C m (R r t . r). x ∈ Rn . t) = 0 and. ∞) × [0. t > 0. Then. r) = − h(y) dSy = Mh (x. r > 0. t) dSy = Mu(·. To show the regularity of U with respect to (r. (x. n ∂B(x.4 (Euler-Poisson-Darboux equation).r) Z G(x. r. t)dy. π]. Spherical means and the Euler-Poisson-Darboux equation. that  Z Z 1 Urr (x. ut (x. Higher dimensional wave equation 85 Set t = 0 for u and ut . 0) = h(x). t) ∈ Rn × (0. t) ∈ C m (R r t of the Euler-Poisson-Darboux equation ( Utt − Urr − n−1 r Ur = 0 in (0.t) (x. ∂B(x. ∞).3. r). 4. We can further compute Urrr and other higher-order ¯+ × R ¯ + ). t) ∈ [0. Ut = H on (0. t) = − u(y.  4.1.3. r. This reduction requires the spherical mean method. r) = − g(y) dSy = Mg (x. ∂B(x.9) U = G. r. r. 0) = g(x) = ∞ X aj sin(jx). Let x ∈ Rn . The issue of convergence will not be discussed here.r) B(x. Higher dimensional wave equation In this section we solve the Cauchy problem for n-dimensional wave equation (n ≥ 2) ( utt − ∆u = 0. r. bj = h(x) sin(jx)dx. ∞) × (0. 1. U (x. derivatives of U up to the m-th order. aj = π 0 jπ 0 Substitute these coefficients into (4. ∞) × {t = 0}. r).8) u(x. Note that the regularity of U with respect to t follows easily. Proof. 0) = h(x) = j=1 ∞ X jbj sin(jx).r) Thus limr→0+ Urr (x. Z Z π 2 2 π g(x) sin(jx)dx.7) and we obtain the solution u in terms of trigonometric series. j=1 So aj and jbj are the Fourier coefficients of functions g(x) and h(x) on [0. Define the spherical means: Z U (x. Z r Ur (x. The idea is to reduce the problem to a certain one spatial-dimensional problem that we are able to solve. t) = − ∆u dS + −1 − ∆u dy.4. t). ∞). ∞). r. t) = n1 ∆u(x.

t)dy n B(x. by formula (4.3.ρ) and so (4. t) = lim U (x. U ˜ (r.11) (r n−1 1 Ur ) r = nαn Z utt (y. t) = lim r→0+ " = lim r→0+ r→0+ ˜ (r. ∂B(x.4. r.t) and so ∂ ∂t Z − ∂B(x. t) = [G(r + t) − G(t − r)] + 2 2 −r+t We recover u(x.r) which expands into the Euler-Poisson-Darboux equation. t) U r # Z r+t ˜ + t) − G(t ˜ − r) G(r 1 ˜ ˜ 0 (t) + H(t). ∂B(x.  4. t) as follows: u(x. this completes the proof of the theorem. function U that ˜ (0. Higher dimensional wave equation 86 2.r) Thus r n−1 1 Ur = nαn Z 1 utt (y.t) ∂B(x.1) ! g(x + tz)dSz .t) (4. ˜ + H(y) dy = G 2r 2r −r+t This gives the Kirchhoff ’s formula for 3-D wave equation: ! Z Z ∂ u(x.1) ∂ = ∂t Z − ∂B(0. Solution in R3 and Kirchhoff ’s formula. t) dSy = rn−1 Utt . U (r. r ˜ (r. t)dy = nαn B(x.t) ! g(y)dSy g(x + tz)dSz ∂B(0. t) satisfies the 1-D wave equation in r > 0. one can see easily from the Euler-Poisson-Darboux equation that 1 (4. 0) = rG := G.3.10) = = Z r − ∆u(y.13) Z = − (th(y) + g(y) + Dg(y) · (y − x)) dSy (x ∈ R3 .6) with k = 0. t > 0). Ur = (4.12) (rU )rr = rUrr + 2Ur = (r2 Ur )r = rUtt = (rU )tt . we have for 0 ≤ r ≤ t Z 1 r+t ˜ 1 ˜ ˜ ˜ H(y) dy. Using the formula above. nαn rn−1 B(x.t) where we have used Z − Z g(y)dSy = − ∂B(x. ∂B(x.r) Z 1 utt (y. t) = t− g(y)dSy + t − h(y)dSy ∂t ∂B(x.r) Z 1 ∆u(y. ˜ Ut (r. t) dSy 0 dρ. ˜ U Hence. t)dy nαn rn−1 B(x. t) = 0.2. For the most important case n = 3. t)dy. t > 0. t) = rU (x. Note That is.r) Z r ! Z utt (y. 0) = rH := H. r. The initial condition is satisfied easily.

t) with parameter z ∈ B(x. We now discuss it in more details. x2 ). ∞). where g¯(x1 . 1 − |w|2 B(0.3. t) ∈ R3 × (0. x2 . ∂t ˜ x. t) solves ( u ¯tt − ∆¯ u = 0. x2 . 0) = (x. Define u ¯(x1 .t) Z 1 g(z) p = dz 2 2πt B(x. The trick is to consider u as a solution of a 3-D wave equation and then apply Kirchhoff’s formula to find u. x2 .t) t − |z − x|2 Z t g(z) p = − dz. u ¯t (x.1) We will give some remarks about Kirchhoff’s formula later. x2 .3.t) t2 − |z − x|2 Z t2 h(z) p + − dz (x ∈ R2 . and hence p z−x t Dγ(z) = − p 1 + |Dγ(z)|2 = p . y3 = ±γ(z). t) = u ¯(¯ x. Then u ¯(x. 0) ∈ R3 . This is called the Hadamard’s method of descent.t) ∂B(0. t).1) .8) with n = 2. Assume u ∈ C 2 (R2 ×[0. t) = ∂t Z t − 2 g(z) p t2 − |z − x|2 B(x. t ∂B(x. t) = u(x1 . x3 ) = h(x1 . t > 0). (x. x3 ) = g(x1 . x3 .t) ∂ B(¯ ∂ B(¯ ˜ x. t) is the ball in R3 centered at x where B(¯ ¯ of radius t.t) t − |z − x|2 Therefore. ∞) solves (4. Solution in R2 by Hadamard’s method of descent. x2 . y3 ). x ∈ R3 . We would like to deduce a formula for u in terms of g.3. ¯ 1 . . then ! Z Z ∂ ¯ g¯(y)dSy + t − h(y)dS t− u(x. (4.t) ˜ x. ˜ We parameterize y ∈ ∂ B(¯ x.4. t) = y. t) in R2 by y = (z. 2 2 B(x. h.14) Since ! Z t2 g(z) p − dz 2 B(x.t) Z p 2 = g(z) 1 + |Dγ(z)|2 dz 4πt2 B(x. Higher dimensional wave equation 87 Z = − Z 1 Dg(x + tz) · zdSz = − Dg(y) · (y − x)dSy . p where γ(z) = t2 − |z − x|2 .t) Z dz = t − g(x + tw) p dw. 4. 2 2 2 t − |z − x| t − |z − x|2 We can then compute that Z − g¯(y)dSy = ˜ x. 2 2 B(x. 0) = h(x). 0) = g¯(x). x2 ) and h(x If we write x = (x1 . x2 ) ∈ R2 and x ¯ = (x1 .t) ∂ B(¯ Z 1 g¯(y)dSy 4πt2 ∂ B(¯ ˜ x. ¯ u ¯(x.t) t − |z − x|2 ∂ u(x.

for all n ≥ 2).t) B(x. k − 1) are independent Now assume that n = 2k + 1 (k ≥ 1).4. h. t ≥ 0). Let f : R → R be C k+1 .4. t) to compute u(x.2. 2.t) So we can rewrite (4. t) ∈ [0.. 3 (in fact. t) = (r U (x. r dr dr j=0 where of f . ∞). β0k = (2k − 1)!! = (2k − 1)(2k − 3) · · · 3 · 1 and βjk (j = 1. the domain of influence of the point (x0 . t > 0). then you only feel the disturbance (hear the sound) once exactly at time t = |x − x0 |. ∞) × [0.1) Z Z Dg(z) · (z − x) g(z) p p dz + t − dz. we need the information of initial data f and g in the whole disk B(x.14) to deduce the so-called Poisson’s formula for 2-D wave equation: Z 1 tg(z) + t2 h(z) + tDg(z) · (z − x) p (4. (a) There is a fundamental difference for wave equations between n = 1 and n = 2. (b) There are also fundamental differences between n = 3 and n = 2 for wave equations. you will feel the disturbance (feel the wave) all the time t ≥ |x − x0 |. ut ∈ C m−2 and thus there is a loss of regularity for wave equation when n = 2. Higher dimensional wave equation 88 we obtain ! Z Z Z ∂ g(z) g(x + tw) Dg(x + tw) · w 2 p p p t − dz = − dw + t − dw 2 2 2 ∂t t − |z − x| 1 − |w| 1 − |w|2 B(x.15) u(x. . Then. Solution for general odd dimensions n. dr2 r dr r dr   k−1 X 1 d k−1 2k−1 dj f (ii) (r f (r)) = βjk rj+1 j (r). r. For general odd dimensions n. for k = 1. t) to compute u(x. in this case.3. ut are at least as smooth as g. r ∂r .) 4. That is. we use the following result whose proof is left as an exercise. In R2 . a disturbance at x0 will affect the values of the solution in the region |x − x0 | ≤ t. where u. t). Set   1 ∂ k−1 2k−1 V (r. 2. if g ∈ C m . = t− 2 2 t − |z − x| t2 − |z − x|2 B(x. if n = 2. 2 2 2 B(x.1) B(0. t) = − dz (x ∈ R2 . t) and. although the effect on you will die out as t → ∞.t) t − |z − x| Remark 4. 0) = g(x). a “disturbance” originating at x0 propagates along a sharp wavefront |x − x0 | = t. Then the function U (x. (Imagine you are at position x in Rn and there is a sharp initial disturbance at position x0 away from you. t) defined above is C k+1 in (r. h ∈ C m−1 then u ∈ C m−1 . . the solution u depends on the derivative Dg of the initial data u(x. While in R3 we only need the information of initial data f and g on the sphere ∂B(x. r. . Therefore. and is the solid disc |x − x0 | ≤ t for n = 2.3.  2     d 1 d k−1 2k−1 1 d k 2k 0 (i) (r f (r)) = (r f (r)). · · · . But for n = 2. This is called the Strong Huygens’ principle in R3 . 3. If n = 3. Let u ∈ C k+1 (Rn × [0. t)) (r > 0. ∞)) be a solution to wave equation utt − ∆u = 0 in Rn × (0. however.t) B(0. 0) is the surface |x − x0 | = t for n = 3. This is different from the 1-D case.5 (Some useful identities). Lemma 4. ∞). In both Kirchhoff’s formula and Poisson’s formula.

11): (r2k Ur )r = r2k Utt . That is. and V (0+ . t). t−r where 1 ∂ k−1 2k−1 ˜ ) (r U (x.  Then. t)) = βjk rj+1 j U (x. by (4. t) = [G(r + t) − G(t 2 2 Z r+t ˜ H(y)dy. we have ∂ 2 1 ∂ k−1 2k−1 1 ∂ k 2k )( ) (r U) = ( ) (r Ur ) 2 ∂r r ∂r r ∂r 1 ∂ k−1 2k−1 1 ∂ k−1 1 2k ) ( (r Ur )r ) = ( ) (r Utt ) = Vtt . t) = U (x. t) = t − gdS (n − 2)!! ∂t t ∂t ∂B(x. (4. Proof. (x. t) u(x. Vtt = Vrr for r > 0.19) belongs to C (R × .5. t > 0. t) = ( ) (r U (x. r. r. t > 0).5. Higher dimensional wave equation 89 Lemma 4. G(r) =( r ∂r 1 ∂ k−1 2k−1 ˜ H(r) =( ) (r Ut (x.6. we derived the following formula for the solution of the Cauchy problem ( utt − ∆u = 0. 0+ . r ∂r (4.18) u(x. we have for 0 ≤ r ≤ t (4.6). t ∂t ∂B(x.16) 1 ˜ ˜ − r)] + 1 V (r.t) When n = 3 this formula agrees with the Kirchhoff’s formula we derived earlier for wave equation in R3 . r. r. 0) = f (x).t) (4. by (ii) of Lemma 4.4. then u defined by (4. t) = U (x.5. t) = 0 follows from part (ii) of Lemma 4.17) But recall u(x. 0)). k−1 X 1 ∂ k−1 2k−1 ∂j V (r. = ( r ∂r r r ∂r Vrr = ( where we have used the Euler-Poisson-Darboux equation (4. That V (0+ .7 (Solution of wave equation in odd dimensions). x ∈ Rn when n = 2k + 1.19) Z  1 ∂  n−3  i 2 + tn−2 − hdS (x ∈ Rn . t) = 0. t) = lim r→0 β k r 0 " # Z r+t ˜ + t) − G(t ˜ − r) G(r 1 1 ˜ lim + H(y)dy = 2r 2r t−r β0k r→0 = 1 ˜0 ˜ [G (t) + H(t)]. t) ∈ Rn × (0. Using part (i) of Lemma 4. r ∂r ∂r j=0 and so V (r. If n ≥ 3 is odd. g ∈ 2 n C m+1 (Rn ) and h ∈ C m (Rn ) for m ≥ n+1 2 . β0k Therefore.3. t) and. Theorem 4. ∞). 0) = g(x). Z h ∂  1 ∂  n−3   1 2 n−2 u(x. ut (x. 0)). for odd n ≥ 3. 0+ .

then in order that solution u defined by (4. 1.t)→(x0 . . t) = h(x0 ). t) = t Gt . t) for odd n ≥ 3. t) γn ∂t t ∂t and Z ∆G(x.t) This proves utt − ∆u = 0 in Rn × (0.5(ii). . Suppose h ≡ 0. ∞) and satisfies the Cauchy condition in the sense that. utt (x. in order for u to be of C 2 we only need g ∈ C 2 and h ∈ C 1 .t) 1 = nαn Z n−1 ∆g dS = t Z − ∂B(x. Let γn = (n − 2)!! and k = n−1 2 . 0).19) becomes Z  1  ∂  1 ∂ k−1  2k−1 u(x. .t)  Remark 4.t) nαn tn−1 B(x. 2. D 2 g and of h. n−1 n−3 Dg. Higher dimensional wave equation 90 (0. t>0 u(x.3. Therefore there is a possible loss of smoothness for higher-dimensional wave equations.t)→(x0 . γn ∂t t ∂t ∂B(x. Proof.5(i). So we give a similar proof for case (b). we need g ∈ C m+1 (Rn ) and h ∈ C m (Rn ) for some m = n+1 2 .3. by Lemma 4. note that Z Z 1 t ∆g dy = ∆g dy Gt = − n B(x. solves wave equation utt − ∆u = 0 in Rn × (0. t − (t Gt )t = γn ∂t t ∂t t γn ∂t t ∂t ∂B(x. . for each x0 ∈ Rn .19) is of C 2 (Rn × (0.4. However. t) = t G(x.20) ut (x.t) and so ∂ (t2k Gt )t = ∂t 1 nαn ! Z ∆g dy B(x. For example. ∞)). The proof in case (a) is given in the text. (b) If n = 1. t) → (x0 . t) → 0 as (x. 0). . 1  1 ∂ k  2k  1  ∂  1 ∂ k  2k  (4. (a) To compute u(x. t) → (x0 . t) = βj (j + 1)t γn ∂tj ∂tj+1 γn j=0 as (x. 0) = g(x0 ) → u(x. if n = 3 then we need 3 2 g ∈ C and h ∈ C .t) On the other hand. In the second identity of (4.0). t) = t Gt . lim (x. ∞)).t) Z ∆g(x + z) dSz = − ∆g dS. G(x.0).t) Hence utt = Z   1  ∂  1 ∂ k−1  1 2k 1  ∂  1 ∂ k−1  2k−1 ∆g dS . t) =  1  ∂  1 ∂ k−1  2k−1 t ∆G(x. γn t ∂t γn ∂t t ∂t The first identity implies ut (x. We may separate the proof in two cases: (a) g = 0. if n ≥ 3 is odd. t > 0. . t) = ∆x − Z g(x + z) dSz = − ∂B(0. . t) = − gdS.20). (b) h = 0. t) . Then 1 ≤ k ≤ m − 1 and the function u defined by (4. ∞). Also.t) ∆g dS. ∆u(x. ∂B(x.   k−1 j+1 G j β0k 1 X k j+1 ∂ j∂ G + t G(x0 . t>0 ut (x. Dh. . t) = g(x0 ). ∂B(x. t). lim (x.t) By Lemma 4. we only need the information of g. ∂B(0. D 2 h on the boundary ∂B(x. . t > 0.

vt (x.5.t)→(x0 . (n + 1)αn+1 (n − 1)!! n!! 2 We deduce that. The trick is to use Hadamard’s method of descent. t) = tn − dy n!! ∂t t ∂t t2 − |y − x|2 B(x. 2 (n + 1)αn+1 B(x.8 (Solution of wave equation in even dimensions). Assume that n is even in this section.t) (4. Higher dimensional wave equation 91 4.21)  1 ∂  n−2  Z i h(y) 2 n p t − + dy .19) to obtain the formula for v(x. it is apparently true that v(x. yn+1 = ±γ(z).0).t) Z 2 g(z) p dz = n−1 2 (n + 1)αn+1 t t − |z − x|2 B(x. lim (x. Then Z h ∂  1 ∂  n−2   1 2 n−1 u(x. lim (x. Note that if u(x. Goal is to derive a representation formula similar to (4. to compute that Z − Z 1 g¯(y)dSy = g¯(y)dSy (n + 1)αn+1 tn ∂ B(¯ ˜ x.t) Z 2tαn g(z) p = − dz. xn+1 . ∞). we parameterize y ∈ ∂ B(¯ p y = (z.21) belongs to C (R × n (0. xn+1 . t) with parameter z ∈ B(x. xn+1 ) ≡ g(x).19).4. t ∂t t2 − |y − x|2 B(x.t)→(x0 . ˜ x. t>0 This follows from the theorem in odd dimensions. Like before. ∞) and satisfies the Cauchy condition in the sense that.t) ˜ x.t) ∂ B(¯ Z p 2 = g(z) 1 + |Dγ(z)|2 dz (n + 1)αn+1 tn B(x. t) = u(x. 0) = g¯(x. ut (x. we may use (4.3. solves wave equation utt −∆u = 0 in R ×(0.t) t − |z − x|2 Note that 2αn 1 1 n/2 = (using αn = Γ(πn+2 ) ).t) When n = 2 this formula agrees with the Poisson’s formula we derived earlier for wave equation in R2 . If n ≥ 2 is even. Theorem 4. + t ∂t ˜ x. 0) and B(¯ ¯ of radius r in Rn+1 . used before for n = 2. ∞) with initial data ¯ xn+1 ) ≡ h(x). t). t) = t − g¯dS (n − 1)!! ∂t t ∂t ˜ x.3. t) is a solution of the wave equation in Rn × (0. t) = h(x0 ). g ∈ 2 n C m+1 (Rn ) and h ∈ C m (Rn ) for m ≥ n+2 2 . yn+1 ). t>0 u(x. ∞)). xn+1 . Solution for even n by Hadamard’s method of descent. Since n + 1 is odd. then u defined by (4. t) is a solution to the wave equation in Rn+1 × (0.t) ∂ B(¯ Z i  1 ∂  n−2  2 ¯ tn−1 − hdS . for even n. r) to be ball centered x set x ¯ = (x. t) in Rn by As above. xn+1 . . γ(z) = t2 − |z − x|2 .0). v(x.  Z  1 h ∂  1 ∂  n−2 g(y) 2 p u(x. 0) = h(x. t) = g(x0 ). t) = v(¯ x.t) ∂ B(¯ ˜ x. for each x0 ∈ Rn .

0) = f (x. (4. 0 0 Hence utt − ∆u = f (x. t − s) x ∈ Rn . t) = 0 in Rn × (0. s)ds. Z t u(x. In either case. (4. t − s) ds x ∈ Rn . Nonhomogeneous problems and Duhamel’s principle. s)ds.19) or (4. That is. 0 Z 0 t Z utt (x.22). 0 is the solution to the non-homogeneous problem (4. t. Assume n ≥ 2 and f ∈ C [n/2]+1 (Rn × [0.23) U (x. v(x. 4. t) = U (x. 0) is the surface |x − x0 | = t for odd n. t. t. t. s)ds. ∞). then U (x. 1. t) + Utt (x. While.23) is given by formula (4. we let U (x.6.23). t. t) = Ut (x.4. Therefore. s) for x ∈ Rn .  . Proof. s)ds = Ut (x. we have u ∈ C 2 (Rn × [0.4. t) = v(x. t) in Rn × (0. Then Z t (4. The regularity of f shows that the solution U to (4. t) + 0 Z t ∆u(x. t > s. s) − ∆U (x. t) − ∆v(x. 0) = ut (x.24) u(x. t) + Ut (x. t.3. for even dimensions. 2. ∞)). ∞)). s) = 0 in Rn × (s. Higher dimensional wave equation 92 Remark 4. For even n we need information of f and g (and their derivatives) in the whole solid ball B(x. (b) For odd n. Ut (x. ∞)) and is a solution to (4. 0) = 0. the domain of influence of the point (x0 . We now turn to the initial value problem for the non-homogeneous equation ( utt − ∆u = f (x. 0) = 0. t > 0. a disturbance at x0 will affect the values of the solution in the region |x − x0 | ≤ t. t). t) = ∆U (x. s) = v(x. vt (x. s) = 0. t. s)ds 0 is in C 2 (Rn × [0. ∞). s. ut (x. A direct computation shows that Z t Z t ut (x. (a) There is a fundamental difference between odd n and even n. s. s) for x ∈ Rn . is the solution to (4. 0) = 0. Clearly u(x. t. t Utt (x. t) to compute u(x.22). t. a “disturbance” originating at x0 propagates along a sharp wavefront |x − x0 | = t.22) u(x. Motivated by the Duhamel’s principle used above for the heat equation. s)ds = f (x. t) = U (x. ∞). s) be the solution to the initial value problem ( Utt (x. t. s) = f (x. Theorem 4. t. t) is the solution to the initial value problem ( vtt (x. This is called the Strong Huygens’ principle. and is the solid ball |x − x0 | ≤ t for even n. Note that if v(x. t). t. 0) = 0 for x ∈ Rn .3.9.21) above.

Solution. is given by Z v(x. t − r) dSy dr r f (y. s).t) Hence the solution U to (4. t − s) = (t − s) − f (y. t) = 2 So Z x+t s sey dy = (ex+t − ex−t ).t) ! f (y. Find a solution of the following problem ( utt − ∆u = f (x. ut (x. ∞). 0) = 0.23) is given by Z U (x. ∞). 0) = 0. 0) = sex . Using Kirchhoff’s formula.11. (x. Solution. x ∈ R3 . vt (x. t) = U (x. t) ∈ R × (0. t) = Z U (x. s) dSy ds ∂B(x.t−s) ∂B(x. t. t − |y − x|) dy. vt (x. Find a solution of the following problem  utt − uxx = tex . t.t−s) f (y.t−s) so that Z t u(x. 0) = 0. Higher dimensional wave equation 93 Example 4. t). 0) = f (x. u(x. ∂B(x. t) = t − f (y. s) dSy . (x.  . (x. s) = v(x. s) ds = 0 = = = t Z (t − s) − 0 1 4π Z tZ 1 4π Z tZ 1 4π Z 0 0 ∂B(x. |y − x| Note that the domain of dependence is a solid ball now. s) dSy ds t−s f (y. 0) = 0. v(x. t) ∈ R3 × (0. ut (x. t. the solution v to the following Cauchy problem ( vtt − ∆v = 0. t − s) = (ex+t−s − ex−t+s ). 2 x−t s U (x. 2  Example 4. t) ∈ R × (0. u(x. ∂B(x.r) B(x. v(x. 0) = 0. (x. t.10. x ∈ R3 . From d’Alembert’s formula.4.3. s)ds 0 1 = [ex+t (−te−t − e−t + 1) − ex−t (tet − et + 1)] 2 1 = (−2tex + ex+t − ex−t ). 2 Hence Z t u(x. s) dSy . has solution 1 v(x. 0) = 0. ∞). t) ∈ R3 × (0. s) = v(x. we know that the problem  vtt − vxx = 0. ∞).

19) and (4. then v = u − u ˜ solves   vtt − ∆v = 0 in ΩT . h as dimension n gets larger and larger.   ut = h on Ω × {t = 0}. Therefore v ≡ 0 from v = 0 on ΓT . t)|2 )dx (0 ≤ t ≤ T ).21).2. Proof. Fix x0 ∈ Rn .12 (Uniqueness for wave equation). Energy methods and the uniqueness There are some subtle issues about the uniqueness of wave equation using the formulas (4.g. T ] and let ΓT = ∂ 0 Ω = ΩT \ ΩT . t) | 0 ≤ t ≤ t0 . (4.25) u=g on ΓT .4. set ΩT = Ω×(0. There exists at most one solution u ∈ C 2 (ΩT ) of (4.25). Instead we will need the fact that certain quantities of the wave behave nicely (e. t0 ): K(x0 . however. Uniqueness of mixed-value problem. v = 0 on ΓT .13 (Finite propagation speed).4. If u ˜ is another solution.  4. We are interested in the uniqueness of initial-boundary value problem or mixed-value problem:   utt − ∆u = f in ΩT . we cannot use such formula to claim the uniqueness of the C 2 solutions. Let u ∈ C 2 be a solution of utt − ∆u = 0 in Rn × (0. t0 > 0 and consider the backwards wave cone with apex (x0 . Ω Divergence theorem yields (note that it seems v ∈ C 3 is needed in using the divergence theorem below. 4. Energy methods and the uniqueness 94 4. We can use the energy method to give another proof of the domain of dependence for wave equation in whole space.1. t0 ) × {t = 0}. . Theorem 4. t) + |Dv(x.  vt = 0 on Ω × {t = 0}. For initial data that are only C 2 . If u = ut = 0 on B(x0 . As above. ∞).4. t0 ) := {(x. Define the energy 1 e(t) = 2 Z (vt2 (x.4.4.. t0 ). |x − x0 | ≤ t0 − t}. then u ≡ 0 within the cone K(x0 . Let Ω be a bounded open set with a smooth boundary ∂Ω. the end result holds for only v ∈ C 2 by a smoothing argument) Z 0 e (t) = (vt vtt + Dv · Dvt )dx Z ZΩ ∂v = vt (vtt − ∆v)dx + vt dS = 0 ∂ν Ω ∂Ω since v = 0 on ∂Ω implies that vt = 0 on ∂Ω. Theorem 4. conserved). These formulas only hold under more and more smoothness assumptions on the initial data g. Domain of dependence.

0) = 1+|x| 2.t0 −t) t Then e(t) = 1 2 Z 0 t0 −t Z |ξ|=1 (0 ≤ t ≤ t0 ).14. The point here is that energy methods provide a much simpler proof. t0 ) has no effect on the solution within the cone K(x0 . t0 ). from | ∂u ∂ν | = |Du · ν|| ≤ |Du||ν| = |Du|. t)|2 ]dS 2 ∂B(x0 . we can try to find a solution in the form of u(x. we can use Kirchhoff’s formula to find the solution. which becomes exactly the Euler-Poisson-Darboux equation. t) − |Du(x. t)|2 ](t0 − t)n−1 dSξ 2 |ξ|=1 t Z Z 1 = [ut utt + Du · Dut ]dx − [u2t (x. it follows that ut ∂u 1 2 1 1 1 − ut (x. Since solution is unique. t) = v(|x|. 1 u(x. We already know this from the representation formula if u(x.4. t0 ) and consequently has finite propagation speed. t) + Du · Dut (x0 + rξ. ut (x. Energy methods and the uniqueness 95 Remark 4.3. t) + |Du(x. 0) = 0. t) + |Du(x0 + rξ.t0 −t) Z 1 − [u2 (x. Z t0 −t Z 0 [ut utt (x0 + rξ. t)|2 ∂ν 2 2 2 2 1 = − (|ut | − |Du|)2 ≤ 0. Solve the Cauchy problem of the wave equation ( utt − ∆u = 0. t) − |Du(x.t0 −t) ∂B(x0 . by Divergence theorem. However. in R3 × (0. 0) are sufficiently smooth. t) + |Du(x. Thus ut = Du = 0. t)|2 )dx 2 B(x0 . t)| dS ≤ 0 = ut ∂ν 2 2 ∂B(x0 .5.26) − ut (x. In theory. t)|2 )rn−1 drdSξ . t)]rn−1 drdSξ e (t) = 0 |ξ|=1 Z 1 − [u2 (x0 + (t0 − t)ξ. Define the local energy Z 1 e(t) := (u2 (x. Instead. t)|2 ]dS 2 B(x . In particular. the computation would be too complicated. 2 Now e0 (t) ≤ 0 implies that e(t) ≤ e(0) = 0 for all 0 ≤ t ≤ t0 . Examples of other initial and boundary value problems. Details are left as exercise.t −t) ∂B(x0 . Proof. and consequently u ≡ 0 in K(x0 . Uniqueness of wave equation can be used to find the solutions to some mixed-value problems. any solution found in special forms will be the unique solution. .t0 −t) Z 0 0 Z ∂u = ut [utt − ∆u]dx + ut dS ∂ν B(x0 .t0 −t) since. we see that any “disturbance” originating outside B(x0 .t0 −t) t   Z 1 ∂u 1 2 2 (4.4. and so. t) by solving an equation for v. t)|2 ≤ |ut ||Du| − u2t (x.4. ∞). t) + |Du(x0 + (t0 − t)ξ.  4. (u2t (x0 + rξ. 0) and ut (x. Example 4. t) + |Du(x. t) − |Du(x.

Hence u = u ˜|xn >0 is the solution to the original problem. T ] where T > 0.27) utt + Lu = 0 (x. xn ) ∈ Rn−1 × (0. Finite propagation speed for general second-order hyperbolic equations Example 4. t ∈ (0. v satisfies the wave equation in n + 1 dimension. t) = u(x. ut (x. 4. In any case.) We study a class of special second-order equations of the form (4. Example 4. the answer is yes. Then solve the wave equation to get a solution u ˜ on whole space Rn × (0. t ∈ [0. 0. we call equation (4. h to odd functions F and G in xn . ∞). in both cases.16. 0) = h(x) 96 in Rn × (0.15. 2 and {uj (x)}∞ j=1 is an orthonormal basis in L (Ω). If λ = −µ2 < 0. t) cos(µxn+1 ). uj |∂Ω = 0. 0) = g(x). t) = 0 = Ω × (0. let v(x. let v(x. T ]. g(x) and h(x) with respect to uj (x) respectively. x ∈ Ω. We need some machinary for Laplace’s equation to answer this question. Tj0 (0) = bj . t)eµxn+1 . ∞) utt − ∆u = 0. ∞). t). ut (x. xn+1 . 0) = g(x).   u(˜ x.2.4 in the textbook. Solve ( utt − ∆u + λu = 0 u(x. xn+1 . ut (x. this gives the solution u.4. h. Extend the functions g. Solve in ΩT . ∞). Finite propagation speed for general second-order hyperbolic equations (This is from §7. aj . Tj (0) = aj . where L has a simple form Lu = − n X aij (x)Dij u. j=1 where Tj00 (t) −∆uj = λj uj .j=1 with smooth symmetric coefficients (aij ) satisfying uniform ellipticity on Rn . Show solution v must be of separate form: v(x.5. t)ρ(xn+1 ). but we will not study this in this course. Let λ ∈ R and λ 6= 0. − λj T (t) = fj (t). and fj (t). x = (˜ u(x. In this case. Let Ω be a bounded domain and ΩT   utt − ∆u = f (x. t) = u(x.   u(x. x ∈ ∂Ω. . Use the method of separation of variables to find a solution of the form ∞ X u= uj (x)Tj (t). Then. t) u(x. 0) = g(x). Solve   x. Example 4. 0) = h(x). t) ∈ Rn × (0. t) = u(x. The question is whether there is such an orthonormal basis. bj are the Fourier coefficients of f (x. t) = 0. If λ = µ2 > 0.5. 0) = h(x).27) uniformly hyperbolic. xn > 0. (x ∈ Rn ). The initial data of v at t = 0 are also known from g. xn+1 . i. ∞).17.

i. (n − 1)-dimensional surface for each t ∈ [0. Lemma 4. ∞) be given. t) ∈ Rn × (0. Then Cr is the level set {φ(x) = r}. t0 ) let Ct = {x ∈ Rn | q(x) < t0 − t}. t) dx (0 < t < t0 ). t) dS |Dφ(x)| β(x.  u2t + n X i. t) dS. we compute Z  n  X 0 e (t) = ut utt + aij uxi uxj t dx Ct 1 − 2 i. |Dq(x)|  Apply this lemma.j=1 In order to compute e0 (t). Assume ∂Ct = {x ∈ Rn | q(x) = t0 − t} be a smooth. {φ(x)=r} |Dφ(x)| t Ct From this. We assume C is given in terms of a function q(x) smooth positive on Rn \ {x0 } and q(x0 ) = 0 by C = {(x. We would like to find some sort of a curved backward cone C. Let φ(x) = t0 − q(x). we need the following result. t) β(x. t) dS. t0 ). Finite propagation speed for general second-order hyperbolic equations 97 Let u be a smooth solution to (4. Ct Then α0 (t) = Z Z βt (x. For each t ∈ [0. t0 ) | q(x) < t0 − t}. We define the energy 1 e(t) = 2 Z  u2t + Ct n X  aij uxi uxj dx (0 ≤ t ≤ t0 ). Let β(x. t) dS |Dφ(x)| Z ! dr − {φ(x)=t} Z βt (x. we have 0 Z t0 Z α (t) = {φ(x)=r} t βt (x. t)dx − ∂Ct Ct β(x. t)dx − = Ct Z ∂Ct β(x. t0 ) ∈ Rn × (0.j=1 aij uxi uxj  1 dS |Dq| .4. |Dq(x)| Proof. By Coarea Formula. t) be a smooth function and define Z α(t) = β(x. t0 ) such that u ≡ 0 within C if u = ut = 0 on C0 = C ∩ {t = 0}. ! Z t0 Z Z β(x. t)dx = α(t) = dS dr.5.27).18.j=1 Z ∂Ct =:A − B. with vertex (x0 . Let (x0 .

ν2 . In fact. ν= Since hv. νn ) is the usual outer unit normal to ∂Ct . integrating by parts. Finite propagation speed for general second-order hyperbolic equations 98 Note that aij uxi uxj t = (aij uxi ut )xj − (aij uxi )xj ut .j=1 n X i.5. ∂Ct i. by Cauchy-Schwarz’s inequal- n n n .j=1 i.j=1 n X Z =− ut Ct Z (aij )xj uxi dx + aij uxi νj ut dS.4. defines an inner product on Rn .j=1 aij vi wj Dq |Dq| on ∂Ct = {q(x) = t0 − t}. Then. wi = ity Pn i.j=1 where ν = (ν1 . we deduce that Z Z n n   X X A= ut utt − (aij uxi )xj dx + aij uxi νj ut dS Ct ∂Ct i. · · · .

X .

X 1/2  X 1/2 .

.

a u ν ≤ a u u a ν ν . .

ij xi j .

ij xi xj ij i j i.j=1 aij νi νj 1/2 dS.j=1 Therefore.j=1 i.j=1 1 . i.30) n X aij qxi qxj = 1. smooth and satisfying (4. Note that condition (4.j=1 i. this condition is satisfied if q solves the generalized Eikonal equation: (4. This proves that e ≡ 0 if e(0) = 0. .j=1 In particular. Let C and Ct be defined as above.19 (Finite propagation speed).27) and satisfies u = ut = 0 on C0 . i.30) for all x 6= x0 and q(x0 ) = 0. Assume q is a function positive.j=1  u2t + ∂Ct i. we have proved the following theorem Theorem 4. |Dq| then we obtain that |A| ≤ Ce(t) + B and so e0 (t) ≤ Ce(t). If u is a smooth solution to (4.j=1 n X aij uxi uxj n  X i. then u ≡ 0 within the cone C.28) is equivalent to (4.j=1 If we have (4. n X Z |A| ≤ Ce(t) + ∂Ct 1 ≤ Ce(t) + 2 Z aij uxi uxj 1/2 n X 1/2 |ut | aij νi νj dS i.28) n X aij νi νj 1/2 ≤ i.j=1 Therefore. i.29) n X aij qxi qxj ≤ 1.

t0 ) when t = 0. Find the explicit solution to the given problem in the form of u(x. t > 0. u(x. 0) = 0. Also try working on the following problems related to the covered materials. 6.   ut (0. Chapter 2: Problems 21-24.t0 −t) t Compute e0 (t) and show e0 (t) ≤ Ce(t). where βjk are constants. u(x. 1 (b) (10 points) Assume h(x) = 1+|x| 2 . 3. utt = uxx . Assume u(x. . m. t) | 0 ≤ t ≤ t0 . t0 ) if u = ut = 0 in B(x0 . (2) (15 points) Solve the mixed value problem for 1-D wave equation   x > 0. t) = U (|x|. t) t > 0. (1) (10 points) Let f : R → R be C m+1 . Show that u ≡ 0 on K(x0 . ∞). x > 0. ∞). 2 )( r dr ) P k−1 k j+1 dj f 1 d k−1 2k−1 (r). show that if α = −1 the problem has no solution unless g ≡ 0. Moreover. t) ≤ t for all x ∈ R3 and t ≥ 0. (a) (5 points) If 0 ≤ h(x) ≤ 1. 0) = h(x). 0) = 0. t)? (4) (10 points) Show that there exists a constant K such that K |u(x. 0) = g(x). t). (5) (10 points) Let f (p. 0) = 0 R and U (0) = R3 (|g| + |Dg| + |D2 g|) dy < ∞. u) be a smooth function on (p. 2. ut (x. (b) ( r dr ) (r f (r)) = j=0 βj r drj (c) β0k = (2k − 1)!! = (2k − 1)(2k − 3) · · · 3 · 1. What is u(0. Let Z 1 e(t) = (u2 + |Du|2 + u2 ) dx (0 ≤ t ≤ t0 ). let K(x0 . m. Prove that for k = 1. . u) = 0 on Rn × [0. ut (x. u(x. So complete the arguments that are left in lectures. 0) = 0. 7. Chapter 12: Problems 2. . some problems being from Chapter 12 of the textbook. t)| ≤ U (0) ∀ t > 0 t whenever u is a smooth solution to the Cauchy problem of 3-D wave equation utt − ∆u = 0 in R3 × (0. show that 0 ≤ u(x. ∞). t) solve the wave equation in R3 × R utt = ∆u. . Homework # 6. ut .Suggested exercises 99 Suggested exercises Materials covered are from Chapter 2 of the textbook with some new material (from Chapter 7) added. 2 B(x0 . d2 1 d k−1 2k−1 d k 2k 0 (a) ( dr (r f (r)) = ( 1r dr ) (r f (r)). 0. Given any (x0 . (3) Let h ∈ C 2 (R3 ) and let u(x. m. t) = αux (0. t0 ) = {(x. t0 ) ∈ Rn × (0. t) is a smooth solution to the nonlinear wave equation utt − ∆u + f (Du. u) ∈ Rn × R × R and f (0. ut (x. 0) = g(x). . where α 6= −1 is a constant and g ∈ C 2 (R+ ) vanishes near x = 0. |x − x0 | ≤ t0 − t}. Hint: Use the energy method.

51 compactly embedded. 85 exterior ball property.Index backward uniqueness theorem. 24 Green’s function. 56 generalized Eikonal equation. 73 mixed-value problems. 51 maximum principle. 48 heat ball. 50 contraction mapping theorem. 26 Poisson’s formula. 51 strongly positive. 45. 24 domain of dependence. 33 Kirchhoff’s formula for 3-D wave equation. 57 infinite propagation speed. 43 bilinear form. 36 barrier function. 31 regular. 51 bounded. 52 Riesz representation theorem. 80 fundamental solution. 55 heat kernel. 37 mean-value property. 51 Cauchy problem. 40 outer unit normal. 65 ball mean-value property. 25 divergence. 24 divergence form. 98 gradient. 52 energy method. 91 Harnack’s inequality. 26 fundamental solution of the heat equation. 72 parabolic boundary. 49 energy estimate. 24 parabolic. 88 Poisson’s formula on ball. 44 Euler-Poisson-Darboux equation. 33 inversion point. 43 regularity. 51 d’Alembert’s formula. 44 exterior ball property at boundary point. 33 Hadamard’s method of descent. 31 radial function. 57 coercive. 51 100 . 52 elliptic. 82 Duhamel’s principle. 31 Poisson’s formula for 2-D wave equation. 28 Green’s function for Rn + . 24 Lax-Milgram Theorem. 62 initial-value problem. 34 minimum principles. 87. 25 reflection. 71 initial-boundary value problem. 48 divergence theorem. 82 domain of influence. 51 Lax-Milgram Theorem Theorem. 86 Laplace operator. 51 bounded linear functional. 33 Poisson’s kernel for ball. 69 heat equation. 27 Dirichlet (boundary) problem. 58. 81 Dirac measure. 34 Poisson’s kernel of Rn + . 28 Poisson’s equation. 61 Poincar´ e’s inequality. 38 mixed-value problem. 31 Green’s function for ball. 57 inversion. 1). 44 finite speed of propagation. 92 eigenvalue problem. 61 Neumann (boundary) problem. 50 Poisson integral of g with kernel K. 25 Newton’s potential. 62. 34 Poisson’s kernel. 28 Poisson’s kernel for B(0.

81 weak derivative. 49 101 . 50 spherical mean method. 85 spherical mean-value property. 72 wave equation. 46 subharmonic in Ω. 84. 51 weakly differentiable. 63 uniformly elliptic. 81 Tychonoff solution. 96 Sobolev-Rellich-Kondrachov compactness theorem. 49 weak gradient vector. 92 Strong maximum principle. 61. 49 weak maximum principle. 88. 35 Strong Huygens’ principle.Index separation of variables. 38 sub-solution of L. 62 traveling waves. 96 uniformly parabolic. 62 super-solutions. 49 uniformly hyperbolic. 72 subsolution of heat equation. 45. 62 weak solution to Dirichlet problem. 41 subsolution of ut + Lu = 0. 61 weak minimum principle.