MAT-INF4300 Appendices A.3 Notation Notation for derivatives. Assume u : U 7→ R, x ∈ U. Provided that the limit exists, u(x + hei ) − u(x) ∂u (x) = lim . h→0 ∂xi h The usual shorthand version. ∂u ∂2u ∂3u u xi = , u xi xj = , u xi xj xk = ∂xi ∂xi xj ∂xi xj xk The multiindex notation consists of a vector of the form α = (α1 , . . . , αn ), where each component αi is a nonnegative integer. This vector is called the multiindex, and has order |α| = α1 + · · · + αn . Given a multiindex α we define D α u(x) := ∂ |α| u(x) = ∂xα11 . . . ∂xαnn u α α n n ∂xn . . . ∂xn If k is a nonnegative integer, D k u(x) := D α u(x) | |α| = k is the set of all partial derivatives of order k. By ordering the partial derivatives, we can also regard D k u(x) as a point in Rn . X 1/2 . |D k u| = |D α u|2 |α|=k If in the special case k = 1 we have the elements of Du being a vector. Du = ux1 , . . . , uxn = gradient vector. If k = 2 we regard the elements of D 2 u as being arranged in a matrix: ∂2u 2u . . . ∂x∂1 ∂x ∂x21 n 2 . . D u= = Hessian matrix . ∂2u ∂xn x1 ∆u = n X i=1 ... ∂2u ∂x2n uxixi = tr D 2 u = Laplacian of u. 1 B.2. Elementary Inequalities *a. Cauchy’s inequality ab ≤ a2 b2 + 2 2 (a, b ∈ R) Proof 0 ≤ (a − b)2 = a2 − 2ab + b2 a2 b2 2ab ≤ a + b =⇒ ab ≤ + 2 2 2 2 *b. Cauchy’s inequality with ǫ ab ≤ ǫa2 + b2 4ǫ (a, b > 0, ǫ > 0). Proof. Using the regular Cauchy’s inequality. √ b 2ǫa2 b2 b2 √ 2ǫa ab = + = ǫa2 + ≤ 2 2(2ǫ) 4ǫ 2ǫ c. Young’s inequality. Let 1 < p, q < ∞ and ap bq + ab ≤ p q 1 p + 1 q = 1. Then (a, b > 0). Proof. The mapping x 7→ ex is convex, and consequently 1 p+ 1 q ab = elog a+log b = e p log a log bq 1 ap bq 1 p q + ≤ elog a + elog b = p q p q *d. Young’s inequality with ǫ. ab ≤ ǫap + C(ǫ)bq (a, b > 0, ǫ > 0) for C(ǫ) = (ǫp)−q/p q −1 . Proof. Using Young’s inequality. ab = (ǫp)1/p a ǫpap + = p bq (ǫp)q/p q b (ǫp)1/p ≤ p (ǫp)1/p a p + b (ǫp)1/p q q = ǫap + bq (ǫp)−q/p q −1 = ǫap + C(ǫ)bq 2 e. Hölder’s inequality. Assuming 1 < p, q < ∞ and u ∈ Lp (U) and v ∈ Lq (U), we have Z |uv|dx ≤ kukLp (U ) kvkLq (U ) 1 p + 1 q = 1. Then, if U Proof. By homogeneity, we may assume kukLp = kvkLq = 1. Then by Young’s inequality, for 1 < p, q < ∞, Z Z Z 1 1 p |u| dx + |v|q dx = 1 = kukLp kvkLq |uv|dx ≤ p q U U U f. Minowski’s inequality. Assume 1 ≤ p ≤ ∞ and u, v ∈ Lp (U). Then, ku + vkLp (U ) ≤ kukLp (U ) + kvkLp (U ) Proof. ku + Z U |u+v| So, p−1 vkpLp (U ) = (|u|+|v|)dx ≤ Z p U |u + v| dx = Z Z U |u + v|p−1 |u + v|dx ≤ Z 1/p Z 1/p p−1 p p p |u| dx + |v| dx |u+v| dx p U U p p = ku + vkp−1 Lp (U ) kukL (U ) + kvkL (U ) p p ku + vkpLp (U ) ≤ ku + vkp−1 Lp (U ) kukL (U ) + kvkL (U ) ku + vkLP (U ) ≤ kukLp (U ) + kvkLp (U ) U =⇒ *g. General Hölder inequality. Let 1 ≤ p1 , . . . , pm ≤ ∞, with and assume uk ∈ Lpk (U) for k = 1, . . . , m. Then Z U |u1 . . . um |dx ≤ m Y k=1 P 1 i pi =1 kuk kLpk (U) Proof. Induction using Hölder’s inequality. 3 h. Interpolation inequality for Lp -norms. Assume 1 ≤ s ≤ r ≤ t ≤ ∞ and 1 θ (1 − θ) = + . r s t Suppose also u ∈ Ls (U) ∩ Lt (U). Then u ∈ Lr (U), and kukLr (U ) ≤ kukθLs (U ) kuk1−θ Lt (U ) . = 1. Proof. We can use Hölder’s inequality since θrs + (1−θ)r t Z Z r |u| dx = |u|θr |u|(1−θ)r dx U ≤ Z U |u| U s θr θr θrs Z (1−θ)r t t dx |u|(1−θ)r (1−θ)r . U i. Cauchy-Schwarz inequality. (x, y ∈ Rn ). |xy| ≤ |x||y| Proof. Let ǫ > 0 and note 0 ≤ |x ± ǫy|2 = |x|2 ± 2ǫxy + ǫ2 |y|2 =⇒ ±xy ≤ Provided y 6= 0 we set ǫ = ±xy ≤ |x| . |y| So 1 2 ǫ 2 |x| + |y| 2ǫ 2 1 2ǫ = |y| 2|x| and ǫ 2 = |x| . 2|y| We get, |y| 2 |x| 2 |x||y| |x||y| |x| + |y| = + = |x||y| 2|x| 2|y| 2 2 |xy| ≤ |x||y| 4 C.2 Gauss-Green Theorem In this section we assume U is a bounded, open subset of Rn , and ∂U is C 1 . Gauss-Green Theorem. Suppose u ∈ C 1 Ū . Then Z Z uxi dx = uν i dS (i = 1, . . . , n). U ∂U The integral of the partial derivative of u over xi over the vector x in U, is equal to the surface integral of the antiderivative u times the unit vector ν i on the boundary ∂U. This result is not proven, but just accepted as a fact. Integration by parts formula. Let u, v ∈ C 1 (Ū ). Then Z Z Z uxi vdx = − uvxi dx + uvν i dS (i = 1, . . . , n) U U ∂U Proof First we apply the Gauss-Green with uv instead of u Z Z (uv)xi dx = uvν i dS. U ∂U Then we observe that (uv)xi = uxi v + uvxi by the product rule of differentiation, so we get: Z Z Z Z (uv)xi dx = uxi v + uvxi dx = uxi vdx + uvxi dx U U U Setting both these equal each other, we get the identity. Green’s formulas For u, v ∈ C 2 (U ). (i) Z U △udx = Z ∂U ∂u dS ∂ν Proof We use Gauss-Green’s theorem with uxi instead of u and arrive at the equality: Z Z uxixi dx = uxi ν i dS. U ∂U Summing over i = 1, . . . , n, we get n Z n Z X X uxixi dx = i=1 U i=1 5 uxi ν i dS ∂U From elementary calculus; Z X n uxixi dx = U i=1 Z n X uxi ν i dS. ∂U i=1 In the integral on the left side, we recognise the expression of the Laplacian. On the right side we have the definition of the expression for the dot product of Du · ν which is the definition of ∂u . Hence, we have ∂ν Z Z ∂u dS. △udx = U ∂U ∂ν (ii) Z U Dv · Dudx = − Z Z u△vdx + U ∂U ∂v udS ∂ν Proof Using the integration by parts formula with vxi instead of v. Z Z Z uxi vxi dx = − uvxi xi dx + uvxi ν i dS U U ∂U Summing from i = 1 to n and taking the integral over the entire sum, and factor out u from the sums on the right side. Z X n U i=1 uxi vxi dx = − Z U u n X vxi xi dx + i=1 Z u ∂U n X vxi ν i dS i=1 On the left side, we see we have the dot product of the gradients. In the first term on the right side we have the Laplacian of v and in the last term we get ∂v the defining expression for ∂ν . Finally: Z Z Z ∂v Du · Dvdx = − u△vdx + u dS U U ∂U ∂ν (iii) Z U u△v − v△udx = 6 Z u ∂U ∂v ∂u −v ∂ν ∂v Proof Using identity (ii) and shifting the terms: Z Z Z Du · Dvdx = − u△vdx + U Z U u△vdx = U Z u ∂U u ∂U ∂v dS − ∂ν Z U ∂v dS ∂ν Du · Dvdx And by using (ii) again with u and v interchanged and rewriting we get Z Z Z ∂u v△udx = v dS − Dv · Dudx U ∂U ∂ν U Subtracting the left and right sides yields Z Z Z Z Z Z ∂v ∂u u△vdx − v△udx = u dS − Du v dS + Dv Dudx · Dvdx − · U U U U ∂U ∂ν ∂U ∂ν We collect the terms in the integrals, and end up with Z Z ∂v ∂u u△v − v△udx = u − v dS ∂ν U ∂U ∂ν Polar Coordinates (i) Let f : Rn 7→ R be continuous and summable. Then Z Z ∞Z f dx = dS dr Rn 0 B(x0 ,r) for each point x0 ∈ Rn . The integral of the function f over the entire space is the same as taking the surface integral over the ball with radius r from 0 to ∞. (ii) In particular d dr Z B(x0 ,r) Z f dx = f dS ∂B(x0 ,r) for each r > 0. Differentiating the integral over the ball with radius r gives us the surface integral over the boundary for the same ball. 7 C.7. Uniform convergence Suppose that {fk }∞ k=1 is a sequence of real-valued functions in R such that |fk (x)| ≤ M k = 1, . . . , x ∈ Rn for some constant M. Every function in the sequence is bounded by the constant M over the entire domain. We also need the sequence of fk ’s to be uniformly equicontinuous. Then there exists a subsequence {fkj }∞ j=1 ⊂ {fk }∞ and a continuous function f such that k=1 fkj → f uniformly on compact subsets of Rn . To say {fk } is uniformly equicontinuous means that for each ε > 0, there exists a δ > 0 such that |x − y| < δ implies fk (x) − fk (y)| < ε for x, y ∈ Rn for k = 1, . . .. (All the fk ’s are continuous). D.1 Banach Spaces We let X denote a real linear space. Definition A mapping k.k : X 7→ [0, ∞) is called a norm if: (i) ku + vk ≤ k|uk + kvk for all u, v ∈ X. (Triangle inequality). (ii) kλuk = |λ|kuk for all u ∈ X, λ ∈ R. (iii) kuk = 0 ⇐⇒ u = 0. Hereafter we assume X is a normed, linear space. Definition We say a sequence {uk }∞ k=1 ⊂ X converges to u ∈ X, written uk → u, if lim kuk − uk = 0. k→∞ Definitions (i) A sequence {uk }∞ k=1 ⊂ X is called a Cauchy sequence if for each ε > 0 there exists N > 0 such that kuk − ul k < ε for all k, l ≥ N. (ii) X is complete if each Cauchy sequence in X converges; that is, ∞ whenever {uk }∞ k=1 is a Cauchy sequence, there exists u ∈ X such that {uk }k=1 converges to u. (iii) A Banach space X is a complete, normed linear space. 8 Definition We say X is separable if X contains a countable, dense subset. Examples (i) Lp spaces. Assume U is an open subset of Rn , and 1 ≤ p ≤ ∞. If f : U 7→ R is measurable, we define R 1/p p |f | dx if 1 ≤ p < ∞ U kf kLp (U ) := ess supU |f | if p = ∞. We define Lp (U) to be the linear space of all measurable functions f : U 7→ R for which kf kLp (U ) < ∞. Then Lp (U) is a Banach space, provided we identify two functions which agree a.e. Hölder spaces. (iii) Sobolev spaces. D.2 Hilbert Spaces Let H be a real linear space. Definition A mapping (·, ·) : H × H 7→ R is called an inner product if (i) (u, v) = (v, u) for all u, v ∈ H (ii) the mapping u 7→ (u, v) is linear for each v ∈ H. (iii) (u, u) ≥ 0 for all u ∈ H. (iv) (u, u) = 0 if, and only if, u = 0. Notation If (·, ·) is an inner product, the associated norm is kuk := (u, u)1/2 (u ∈ H) (1) The Cauchy-Schwarz inequality states |(u, v)| ≤ kukkvk (u, v ∈ H). (2) This inequality was proved in appendix B.2. Using (2) we easily verify (1) defines a norm on H. Definition A Hilbert space H is a Banach space endowed with an inner product which generates the norm. 9 Examples (a) The space L2 (U) is a Hilbert space, with Z (f, g) = f gdx. U 1 (b) The Sobolev space H (U) is a Hilbert space, with Z (f, g) = f g + Df · Dgdx. U Definitions (i) Two elements u, v ∈ H are orthogonal if (u, v) = 0. (ii) A countable basis {wk }∞ k=1 ⊂ H is called orthonormal if (wk , wl ) = 0 (k, l = 1, . . . ; k 6= l) kwk k = 1 (k = 1, . . .). If u ∈ H and {wk }∞ k=1 ⊂ H is an orthonormal basis, we can write u= ∞ X (u, wk )wk , k=1 where the series converges in H. In addition, 2 kuk = ∞ X (u, wk )2 . k=1 Definition If S is a subspace of H, S ⊥ = {u ∈ H | (u, v) = 0} for all v ∈ S is the subspace orthogonal to S. D.3 Bounded Linear Operators Definition (i) A bounded linear operator u∗ : X 7→ R is called a bounded linear functional on X. (ii) We write X ∗ to denote the collection of all bounded linear functionals on X; X ∗ is the dual space of X. Theorem (Riesz Representation Theorem) H ∗ can be canonically identified with H; more precisely, for each u∗ ∈ H ∗ there exists a unique element u ∈ H such that hu∗ , vi = (u, v) for all v ∈ H. (This mapping is a linear isomorphism of H ∗ onto H). 10 E.1 Lebesgue Measure Lebesgue measure provides a way of describing the ”size” or ”volume” of certain subsets of Rn . Definition A collection of M of subsets of Rn is called a σ-algebra if (i) ∅, Rn ∈ M (ii) A ∈ M =⇒ Ac ∈ M. ∞ ∞ (iii) If {Ak }∞ k=1 ⊂ M, then ∪k=1 Ak , ∩k=1 Ak ∈ M Theorem (Existence of Lebesgue -measures and -sets There exists a σ-algebra M of subsets of Rn and a mapping | · | : M 7→ [0, +∞] with the following properties: (i) Every open subset of Rn , and thus every closed subset of Rn , belong to M. B. (ii) If B is a ball in Rn , then |B| equals the n-dimensional volume of (iii) If {Ak }∞ k=1 ⊂ M and the sets are pairwise disjoint, then ∞ ∞ X [ |Ak | (countable additivity) Ak = k=1 k=1 (iv) If A ⊆ B, where B ∈ M and |B| = 0, then A ∈ M and |A| = 0. Notation The sets in M are called Lebesgue measurable sets and | · | is the ndimensional Lebesgue measure. If some property holds everywhere on Rn except for a measurable set with Lebesgue measure zero, we say the property holds Almost everywhere, abbreviated ”a.e”. Remarks From the previous information we see that |A| equals the volume of any set A with piecewise smooth boundary. We see that |∅| = 0 and that if the sets in (iii) are not disjoint, we have a ≤ instead of equality (subadditivity). 11 E.2 Measurable functions and integration Definition Let f : Rn 7→ R. We say f is a measurable function if f −1 (U) ∈ M for each open subset U ∈ R. Note in particular that if f is continuous, then f is measurable. The sum and product of two measurable functions are measurable and for a sequence {fk }of measurable functions, lim sup fk and lim inf fk , are also measurable. Theorem (Ergoroff’s Theorem) let {fk }, f be measurable functions and fk → f a.e on A, where A ⊂ Rn is measurable, |A| < ∞. Then for each ε > 0, there exists a measurable subset E ⊂ A such that (i) |A − E| ≤ ε. (ii) fk → f uniformly on E. Now if f is a nonnegative, measurable function, it is possible, by an approximation of f with simple functions, to define the Lebesgue integral Z f dx. Rn This agrees with the usual integral if f is continuous or Riemann integrable. If f is measurable, but not necessarily nonnegative, we define Z Z Z + f dx = f dx − f − dx, Rn Rn Rn provided at least one of the terms on the right hand side is finite. In this case we say f is integrable. Definition A measurable function f is summable if Z |f |dx < ∞. Rn 12 Remark Note carefully the terminology: a measurable function is integrable if it has an integral (which may be infinite) and is summable if this integral is finite. Notation If the real-valued function f is measurable, we define the essential supremum of f to be ess sup f := inf µ ∈ R |{f > µ} = 0 . E.3 Convergence theorems for integrals The Lebesgue theory of integration is especially useful since it provides the following powerful convergence theorems. Theorem (Fatou’s Lemma) Assume the functions {fk } are nonnegative and measurable. Then, Z Z fk dx. lim inf fk dx ≤ lim inf Rn k→∞ k→∞ Rn Theorem (Monotone Convergence Theorem) Assume the functions {fk } are measurable, with 0 ≤ f1 ≤ f2 ≤ · · · ≤ fk ≤ fk+1 ≤ . . . , then Z lim fk dx = lim Rn k→∞ k→∞ Z fk dx. Rn Theorem (Dominated Convergence Theorem) Assume the functions {fk } are integrable and fk → f a.e. Suppose also |fk | ≤ g a.e for some summable function g. Then Z Z fk dx → f dx. Rn Rn 13 —————————————————————- 24/08-2010 Normal ODE’s are of the form ′ y (t) = f t, y(t) . y(0) = c PDE’s have the form Υ t, x1 , . . . , xn , u, ut, ux1 , . . . = 0 | {z } ind. var. where u = u(t, x1 , . . . , xn ), a function of several variables. Lipschitz continuity is whenever |f (x)−f (y)| ≤ L|x−y| for some constant L. We write f ∈ C k (Ω) for some function f : Ω 7→ R which means f, f ′ , . . . , f (k) exist and f (k) is continuous. As opposed to ODE, there are usually no general solutions for PDE’s. Recall Taylor’s theorem for a series expansion. (x − a)n (n) f (a). n! f (x) = f (a) + (x − a)f ′ (a)0 . . . + Proving using the fundamental theorem of calculus. Z x Z ′ f (t) = f (x) − f (a) =⇒ f (x) = f (a) + a x f ′ (t) a and by integration-by-parts: ′ ′ = f (a) + xf (x) − af (a) − = f (a) + Z x tf ′′ (t)dt a x ′′ a where Z Z ′ ′ f (t)dt + xf (a) − af (a) − x a Z f ′′ (t)dt = x f ′ (x) − f ′ (a) ′ = f (a) + (x − a)f (a) + Z a tf ′′ (t)dt a x (x − t)f ′′ (t)dt and by expanding forever, we eventually arrive at (*). 14 x (*) The n-dimensional version. f (x) = f (a) + Z |α|≤K 1 D α f (a)(x − a)α α! α Where the one dimensional notation would be D α f = ∂∂xαf and the multi|α| f dimensional notation means D α = ∂xα1∂,...,∂x αn which can also be written as n 1 α1 αn ∂x1 , . . . , ∂xn f . Here we have the vectors |α = (α1 , . . . , αn ) ∈ Rn , x = (x1 , . . . , xn ) ∈ Rn the scalar |α| = α1 + · · · + αn and the cross-faculty α! = α1 ! . . . αn !. When we set |α| = K for some positive, integer constant K this means all combinations α1 + · · · + αn = K. Notation wise the three vectors (x − a)α = (x1 − a1 )α1 . . . (xn − an )αn . Proof Fix some x ∈ Rn . Join a and x by u(t) = a + t(x − a), reducing the multidimensional to the 1-dimensional case, so we can use Taylor’s expansion. We have performed a linear interpolation, so if this was for n = 2, we would have for two points in the plane: The function f u(t) : R 7→ R with f u(1) = f (x) and f u(0) = f (a). Now we use the 1-dimensional Taylor’s form. n X 1 dk f u(t) f (x) = f u(1) = f (a) + k! dtk t=0 k=1 X K · D α f u(t) t=0 (x − a)α = f (a) + α |α|≤K X K D α f (a)(x − a)α = f (a) + α |α|≤K Gauss-Green formula For some open set U ⊆ Rn and some function u : U 7→ R, where u ∈ C 1 Ū ∂u where we have the multi-dimensional u = u(x1 , . . . , xn ) and uxi = ∂x i Z Z uxi dx = uν i ds U ∂U 15 where ∂U denotes the boundary of U. ν = (ν 1 , . . . , ν n ) is the outward norm from the boundary: We use ds because the integral over ∂U has a different measure than the first integral. Green’s Formula (Important) P For a Laplacian: △u = ni=1 uxixi (with no cross derivatives) we have, for u, v ∈ C 2 (Ū Z Z ∂u ds (i) △udx = U ∂U ∂ν . = Du · ν (dot product) and Du = u , . . . , u where ∂u x x n 1 ∂ν This can be proved with a 1-dimensional Gauss-Green theorem: Z Z ux dx = u · νds. U ∂U Verifying Green’s formula (i) using Gauss-Green’s theorem with u instead of ux Z Z ux ux = (ux · ν)ds. U (ii) Z U ∂U DuDvdx = − since (uv)xi = uvxi + uxi v. (iii) Z U Z u△vdx + U (u△v − v△u)dx = Z ∂U Z ∂U u ∂v ∂u ds −v ∂ν ∂ν Convolution Let f, g be measurable functions. (f ∗ g)(x) = Z Rn f (x − y)g(y)dy assuming that the integrals exist. 16 ∂v ds ∂ν Properties (i) f ∗ g = g ∗ f which follows from change of variables. (ii) f ∗ (g ∗ h) = (f ∗ g) ∗ h (iii) supp(f ∗ g) ⊂ supp(f ) ∗ supp(g) where supp = support. So supp(f ) = {x ∈ Rn : f (x) 6= 0} and x 6∈ supp(f ) ⇒ f (x) = 0: the region of the function that is not 0. Proof of (iii) Assume x ∈ supp(f ∗ g). Assume for contradiction x 6∈ supp(f ) + supp(g). This implies ∀y ∈ supp(g), x − y 6∈ supp(f ) But (f ∗ g)(x) = Z Rn f (x − y)g(y)dy = Z supp(g) f (x − y)g(y)dy or else it would be zero. So x − y 6∈ supp(f ) = 0 which means f (x − y) = 0 and the integral is 0. Contradiction. Thus, x ∈ supp(f ∗ g) =⇒ x ∈ supp(f ) + supp(g). We consider the function η(x) = C exp 1 |x|2 −1 0, , |x| ≤ 1 |x| > 1 R so supp(y) = B(0, 1): the closed unit ball. We choose C such that R⋉ ηdx = R 1, η ε (x) = ε1n η xε . For this function supp(η ε ) = B(0, ε). We have Rn η ε dx = 1. The support is decreasing as ε → 0; the function is higher and narrower, but the area remains 1. 17 ε↓0 We have the Dirac delta: η ε (x) R= δ(x). This is a measure where δ(x) = 0 except x = 0 where δ(0) = ∞ and δ(x)dx = 1. We introduce the regularization of f . Z ε f (x − y)ηε (y)dy f = f ∗ nε = | {z } Rn ∈C ∞ where f ε : is a smooth, well behaved function. Some properties: (i) f ε ∈ C ∞ (U) (ii) f ε → f a.e as ε ↓ 0. If f is “not smooth” the derivative does not exist, so we use the convolution to obtain a more manageable function. In fact this is the standard way of dealing with non-smooth functions while working with PDE’s. —————————————————————- 27/08-2010 We consider an elliptic PDE, the Laplace equation △u = 0. We have x ∈ Rn , P n u : Rn 7→ R, △u = i=1 uxi xi . Another important elliptic PDE is the Poisson equation △u = f . Definition of distribution D(Ω) = φ : φ ∈ Cc∞ (Ω) the space of test functions. ∂k u Cc∞ (Ω) = u : ; exists ∀k supp(u) is compact k ∂x |{z} ∞ deriv. The distribution T : D(Ω) 7→ R and satisfies T (φn ) → 0 if φn → 0. If T1 = T2 ⇔ T1 (φ) = T2 (φ) ∀φ ∈ D(Ω). If (the first is differentiated) T1′ = T2 ⇒ (T2 , φ) = −(T1 , φ′ ) ∀φ ∈ D(Ω), in the sense of distribution. (1) measure B(0, r) = α(n)r n : the volume of unit ball in Rn . (2) Z — r↓0 ∂B(0,r) g(x)ds(x) −→ g(0) 18 where the special integral sign is the average of the integral: Z Z 1 —f (x)dx = f (x)dx. α(n)r n B(0,r) We will now look closer at the Laplacian equation, △u = 0 in Rn . What u satisfies this equation? This equation has a radial equation, which is a strong property and benefitial.pIt means u(x) = u(|x|), so if x = (x1 , . . . , xn ), we can instead use |x| = x21 + . . . + x2n . We denote u(x) = u(|x|) = v(r) where r = |x|, the radius. We can then consider △v = 0 By the chain rule uxi = v ′ (r) xri : u(x1 , . . . , xn ) = v(r) =⇒ ∂r ∂ ∂ ∂u = v(r) = v(r) u xi = ∂xi ∂xi ∂r ∂xi The double derivative (product rule): u xi xi x2i 1 x2i ′ = v (r) · + v (r) − 3 r r r ′′ (n − 1) ′ v (r) = 0 r We hve not reduced the PDE to a second degree ODE. △u = 0 =⇒ v ′′ (r) + Solution: c log(v ′ ) = (1 − n) log r + log c =⇒ v ′ = n−1 r b log r + c n = 2 v(r) = b +c n≥3 r n−2 for some constants b and c. Claim Choose constants b, c such that −△x u = δ0 (Dirac distr.) in the sense of distribution. Then Z Z − △x ugdx = g(x)δ0 = g(0). The Laplacian equation △u(0) = 0 since |x| = r → 0 means the solutions tend to infinity. 19 Fundamental solution for the Laplacian. △x u = δ0 . This is the case in the solution we found v(r). The fundamental solution is: 1 − 2π log |x| n=2 φ(x) = 1 1 · n ≥3 n(n−2)α(n) |x|n−2 We will prove the claim −△x φ = δ0 where φ is the fundamental solution. We set △φ = 0 in Rn \{0} so we can find solutions. Z △x φ = δ0 =⇒ − △x φg(x)dx = g(0) ∀g ∈ D(Ω) (⋆) Rn For φ we associate a distribution Fφ . Z (Fφ , g) = φ(x)g(x)dx Rn ∀g ∈ D(Ω) = D(Rn ). The g is a smooth, test function. We are interested in (F△φ , g) = and using integration by parts: Z Z Z ′′ ′ ′ f gdx = − f g dx = f g ′′ dx so we get = We continue. Z Rn φ△gdx = Fφ , △g F△φ , g = Z R⋉ △φgdx φ(x)△g(x) Rn We split up this integral in two parts. Z Z = φ(x)△g(x) + B(0,δ) R φ(x)△g(x) := I + J Rn \B(0,δ) Solving I: There are two situations, n = 2 and n ≥ 3. First we calculate for n = 2. Z Z 1 1 log |x| △g(x)dx ≤ k△gkL∞ log |x|dx 2π {z } B(0,δ) 2π B(0,δ) | φ(x) Going to polar coordinates. Z 2π Z δ Z δ ≤ C log |r|rdrdθ ≤ C log |r|rdr ≤ C log |δ|δ 2 0 0 0 20 where C log |δ|δ 2 → 0 as δ → 0. Now we consider the case n ≥ 3 Z 1 1 · n−2 △g(x)dx B(0,δ) n(n − 2)α(n) |x| Z δ Z ≤ k△gkL∞ 1 ds(y)dr n−2 0 B(0,δ) |y| Z Z δ Z δ 1 1 ds(y)dr ≤ C nα(n)r n−1 dr ≤C n−2 n−2 r r B(0,δ) 0 0 Z δ Cnα(n) 2 δ → 0 as δ → 0 = Cnα(n) rdr = 2 0 Thus, I → 0 as δ → 0. We direct our attention to the other integral, J. Z φ(x)△g(x)dx Rn \B(0,δ) Z ∂g ∂φ g(x)ds(x)+ φ(x) ds(x) = ∂ν ∂(Rn \B(0,δ) Rn \B(0,δ) ∂(Rn \B(0,δ) ∂ν R R ∂v = 0 − J1 + J2 where Ω (u△v − v△u)dx = Ω (u ∂ν − v ∂u ). Useful facts are ∂ν ∂φ x x 1 that ∇x φ(x) = − nα(n)|x|n , ν = − |x| and ∂ν = △φ · ν = nα(n)|x| n−1 (scalar product). Z 1 J1 = − g(x)ds(x) n−1 ∂(B(0,δ) nα(n)|x| Z Z 1 g(x)ds(x) = − — g(x)ds(x) → −g(0) as δ ↓ 0 =− nα(n)δ n−1 ∂(B(0,δ) ∂B(0,δ) Z Z △φ(x)g(x)dx− Z J2 = Z ∂g ∂g φ(x) ds(x) ≤ k kL∞ |φ(x)|dx ∂ν ∂ν ∂B(0,δ) ∂B(0,δ) Again we must consider the two cases n = 2 and n ≥ 3. First n = 2. Z log |x|ds(x) ≤ C log |δ|(2πδ) ≤ Cδ log |δ| → 0 as δ ↓ 0 ∂B(0,δ) Similar for n ≥ 3 which means J2 → 0 as δ ↓ 0. So, finally, I + J = g(0). We let δ → 0 to cover the whole space. −△x φ = δ0 21 and remember △φ 6= 0 ∀x ∈ Rn . Z v(x) = φ(x − y)f (y)dy = φ ∗ f Rn △x v(x) = Z Rn △x φ(x − y)f (y)dy = Applying −△x φ = δ0 , we get =− Z Z Rn △yφ(x − y)f (y)dy δ{x} f (y)dy = −f (x) so −△u = f which is the Poisson equation. If you know φ, the fundamental solution, you know v(x) which satisfies −△u = f . Theorem Assume f ∈ C 2 (Rn ), u(x) = where −△x φ = δ0 , then Z Rn φ(x − y)f (y)dy (i) u ∈ C 2 (Rn ) (ii) −△u = f where (i) follows from f ∈ C 2 , since for φ ∗ f , vx = φx ∗ f = φ ∗ fx . —————————————————————- 31/08-2010 Recap Z — Z — 1 u(x)dx = α(n)r n B(x,r) Z 1 u(y)ds(y) = nα(n)r n ∂B(x,r) u(x)dx. B(x,r) Z u(y)ds(y) ∂B(x,r) Harmonic Functions We say that u : Rn 7→ R is harmonic if △u = 0. Mean value theorem Let Ω ⊂ Rn . If u ∈ C 2 (Ω) is harmonic, then Z Z u(x) = — u(y)ds(y) = — u(y)dy ∀B(x, r) ⊂ Ω. ∂B(x,r) B(x,r) 22 Proof (First equality) For r > 0 Z φ(r) = — u(y)ds(y). ∂B(x,r) For r = 0, φ(r) = u(x), since limr→0+ φ(r) = u(x). Claim: φ is constant, which means φ′ (r) = 0. Z φ(r) = — u(y)ds(y); y = x + rz ∂B(x,r) Z =— u(x + rz)ds(z) ∂B(0,1) Z φ (r) = — ′ Z =— By definition. Since ∂u ∂ν ∂B(0,1) ∂B(x,r) ∇u(x + rz)zds(z) ∇u(y) y−x ds(y) r = ∇u · ν Z =— ∂u ds(y) ∂B(x,r) ∂ν Z ∂u 1 ds(y) = n−1 nα(n)r ∂B(x,r) ∂ν Integration by parts. 1 = nα(n)r n−1 Z B(x,r) ∇(∇u)dy We recognise this as the Laplacian. Z 1 = △u(y)dy = 0 nα(n)r n−1 B(x,r) The equality follows since u is harmonic, so △u = 0. Thus φ(r) = φ(0) = u(x). Proving the other equality. Z u(x) = — u(y)dy. B(x,r) 23 Begin by using a very common step. Z Z r Z u(y)dy = B(x,r) = Z 0 0 ∂B(x,s) hZ r n−1 nα(n)s — ∂B(x,s) u(y)ds(y) dr i u(y)ds(y) dr The part inside the brackets is u(x) by the first part. Z r = nα(n)sn−1 u(x)dr = α(n)u(x)r n 0 Thus, Z — u(y)dy = u(x) B(x,r) This theorem has a converse. Theorem If u ∈ C 2 (Ω) satisfies Z u(x) = — u(y)ds(y), ∂B(x,r) ∀B(x, r) ⊂ Ω, then u is harmonic, i.e △u = 0. Proof Assume Z φ(r) = — u(y)ds(y) = u(x) ∂B(x,r) ′ is a constant so φ (r) = 0 and Z r φ (r) = —△u(y)dy. n ′ Let △u 6= 0 for contradiction. Then △u > 0 ot △u < 0 in B(x, r), then φ′ (r) > 0 or φ′ (r) < 0. But φ′ (r) = 0 so △u = 0. Contradiction. Maximum Principle Let Ω ⊂ Rn be open and bounded and assume u ∈ C 2 (Ω)∩C(Ω̄) is harmonic. (1) max u(x) = max u(x) Ω̄ ∂Ω (2) If Ω is connected and ∃x0 ∈ Ω such that u(x0 ) = max u(x) Ω̄ If both these conditions are satisfied, u is a constant in Ω. This theorem has applications to harmonic functions. 24 Proof We note that condition 2 implies condition 1, so we only have to prove condition 2. We assume that x0 ∈ Ω such that M = u(x0 ) = max u(x) for some Ω̄ constant M. This satisfies the mean value theorem, so Z M =— u(y)dy = u(x0 ) = max u(x) ≤ M. Ω̄ B(x,r) The equality holds if u(y) = M, ∀y ∈ B(x0 , r). This means we have u(y) included in B(x0 , r), which means it is constant there. The set A = {x : u(x) = M} is, by connectedness both open and closed, so A = Ω. By using −u instead of u the maximum principle will become the minimum principle with the same theorem except we use ’min’ instead of ’max’. We have shown how to solve (the boundary value problem) −△u = f in Ω u=g on ∂Ω Can we say if this solution is unique? We assume u, v are two solutions and we claim u = v. We set w = u−v w e = v−u which means w, w e both satisfy △w = 0 in Ω w = 0 on ∂Ω We can then apply the maximum principle max |u − v| = max |u − v| = 0 ⇒ u = v ∂Ω Ω We use the maximum principle to prove unique solutions for e.g parabolic equations. Theorem Smoothness of Harmonic Functions Let Ω ⊆ Rn be open and bounded. If u ∈ C(Ω) satisfies Z u(x) = — u(y)ds(y), ∀B(x, r) ⊂ Ω ∂B(x,r) 25 then, u ∈ C ∞ (Ω), we have a smooth function. In particular, u harmonic implies it is smooth. A recap. For 1/(|x|2 −1) Ce |x| < 1 η(x) = 0 |x| ≥ 1 R we choose the constant C such that Rn η(x)dx = 1 and supp(η) ⊂ B(0, 1). Now 1 x ηε (x) = n η ε ε R with Rn ηε (x)dx = 1 (the approximation function) and supp(ηε ) ⊂ B(0, ε). Proof We introduce uε (x) = Z Ω Then, ηε (x − y)u(y)dy = ηε ∗ f. (i) uε ∈ C ∞ (Ω) (ii) uε (x) = u(x). We prove these to prove the theorem. Claim (i) is a smooth C ∞ by the convolution property. Z (uε )x = (ηε )x (x − y)u(y)d Ω (uε )xx = Z Ω (ηε )xx (x − y)u(y)d so uε ∈ C ∞ because ηε ∈ C ∞ . Because of this we only need to show (ii) to complete the proof. Z uε (x) = ηε (x − y)u(y)dy B(x,ε) By def. of ηε . Z |x − y| 1 u(y)dy = n η ε B(x,ε) ε Z Z r 1 ε u(y)ds(y) dr η = n ε 0 ε ∂B(x,r) Z Z 1 ε r = n u(y)ds(y)dr η ε 0 ε ∂B(x,r) 26 1 = n ε Z 0 ε Z r n−1 nα(n)r — u(y)ds(u)dr η ε ∂B(x,r) Now we can use the smoothness property. Z 1 ε r = n nα(n)r n−1 u(x)dr η ε 0 ε Z Z 1 ε r = u(x) n ds(y)dr η ε 0 ε ∂B(0,r) Combining the two integrals back to one. Z |y| 1 = u(x) n dy η ε B(0,ε) ε Z = u(x) ηε (y)dy = u(x) B(0,ε) | {z } =1 So u(x) = uε (x) and the smoothness is proved: u(x) ∈ C ∞ (Ω). —————————————————————- 03/09-2010 We remember that u is harmonic implies that u satisfies MVP, the mean value property, and the reverse: if u ∈ C 2 and u satisfies MVP that implies that u is harmonic. The Max principle states that max u = max u ∂U U which is used for proof of a unique solution. We also have the result that u harmonic ⇒ u ∈ C ∞ (U). Liouville’s Theorem Let u : Rn 7→ R be harmonic and bounded. Then u is constant. 27 Proof We fix some x0 ∈ Rn , so Z u(x0 ) = — u(y)dy. B(x0 ,r) From a previous result, u harmonic ⇒ u ∈ C ∞ (R⋉). Since △u = 0, then △uxi = 0 for all i = 1, . . . , n, and furthermore uxi is itself harmonic. This means, by the MVP, Z Z 1 uxi (x0 ) = — uxi (y)dy = ux (y)dy α(n)r n B(x0 ,r) i B(x0 ,r) Integration by parts, 1 = α(n)r n Z uνi dS(y). ∂B(x0 ,r) 1 |uxi (x0 )| ≤ kνkL∞ α(n)r n Using that u is bounded, ≤ CkukL∞ (Rn Z ∂B(x0 ,r) |u|dS(y) 1 n · α(n)nr n−1 = C ∀r n α(n)r r R since ∂B(x0 ,r) dS(y) = α(n)nr n−1 . If we take the limit, r → ∞, then uxi (x0 ) ≡ 0 ∀x0 ∈ Rn , which is true only when u is constant. Representation formula let f ∈ Cc2 (Rn ), and let n ≥ 3. Then every bounded solution of −△u = f , x ∈ Rn , has the form Z u(x) = φ(x − y)f (y)dy + C Rn where C is some constant and φ is the fundamental solution of the Laplacian. Proof First: why is this not true when n = 2? Because for n = 2, φ(x) = − 1 log |x| → ∞ as |x| → ∞ 2π so the integral can be unbounded. However, for n ≤ 3, we have φ(x) = K |x|n−2 K= 28 1 n(n − 2)α(n) which is not bounded (but that is not obvious). we have already proved that Z u(x) = φ(x − y)f (y)dy Rn is a solution for the Poisson equation. We must now show that this solution is bounded, i.e show that: Z 1 |u(x)| = K f (y)dy n−2 Rn |x − y| Z Z 1 1 ≤ K f (y)dy+|K f (y)dy ≤ |I1 +I2 | ≤ C n−2 n−2 B(x,ε) |x − y| Rn \B(x,ε) |x − y| for some upper bound C. First we see that I2 is bounded. This is because f ∈ Cc2 (Rn ), that is, f has compact support. That means that outside some specified region f will be 0. Hence the integral is not taken over the entire set. Other parts blow up when x → y, but we have removed the ε-ball around x, so y cannot be arbitrarily close to x. This means it is finite. From these observations, I2 is bounded. I1 has the form Z rZ 1 I1 = dsdr n−2 0 ∂B(x,r) r which we have already solved. From this we know that I1 is bounded. Since both I1 and I2 are bounded, then |I1 + I2 | ≤ C, so Z u(x) = φ(x − y)f (y)dy Rn is a bounded solution of −△u = f . Let u and u e be two bounded solutions. For w = u − u e, we have −△w = 0 which means w is constant (Liouville’s theorem), and this in turn means that u=u e0C. Any two solutions are just shifted by a constant. Harnack’s Inequality (For V ⊂⊂ U we mean V ⊂ V ⊂ U, where V is a compact set). For each connected, open set V ⊂⊂ U, ∃C, C > 0 depending on V , such that sup u ≤ C inf u v V for all non-negative, harmonic functions u in U. 29 Proof We set 1 r = dist(V, ∂U) = inf{dist(V, ∂U)} 4 and choose x, y such that |x − y| ≤ r. Z Z 1 1 udz = n u(y) u(x) = — udz ≥ n — 2 2 B(y,r) B(x,2r) This is true for any x, y, so =⇒ 2n u(y) ≥ u(x) ≥ 1 u(y) 2n V connected and V compact means that ∃{Bi }N i=1 ⇒ u(x) ≥ 1 2nN u(y) And u(x) ≥ 21n u(y) means u(y) ≥ 21n u(z) so, naturally, u(x) ≥ 21n u(y) ≥ 2n 12 u(z). Since we have compactness, we have a finite number of “balls” that cover the entire set: u(x) ≥ 2nN1u(y) . This theorem says we have a sense of uniformity. The different values don’t differ more than a constat factor C. Green’s Function We need Green’s function to classify all solutions of −△u = f in Ω u=g on ∂Ω Let us first consider the one-dimensional version. 30 −u′′ (x) = f (x) in Ω u(0) = u(1) = 0 on ∂Ω What is u? We set the “fundamental theorem”: Z x u(x) = C1 + u′ (y)dy 0 ′ u (y) = C2 + Z y ′′ 0 u (z)dz = C2 − Z =⇒ u(x) = C1 + C2 x − We define F (y) := Ry 0 0 f (z)dz, so Z tZ 0 y f (z)dz = 0 Z Use integration by parts, Z x Z x ′ yF (y)dy = xF (x) − yF (y) 0 − 0 Z x y Z y f (z)dz 0 f (z)dzdy. 0 x F (y)dy 0 x yf (y)dy = 0 Z 0 x (x − y)f (y)dy. Using the boundary condition u(0) = u(1) = 0, Z 1 C1 = 0, C2 = (1 − y)f (y)dy 0 so u(x) = x Green’s function Z 1 0 (1 − y)f (y)dy − Z 0 x (x − y)f (y)dy. y(1 − x) 0 ≤ y ≤ x G(x, y) = x(1 − y) x ≤ y ≤ 1 So, u(x) = x Z 1 0 (1 − y)f (y)dy − Z x 0 (x − y)f (y)dy = Some basic properties are • G is continuous. • G(x, y) = G(y, x) (symmetric). 31 Z 0 1 G(x, y)f (y)dy. • Writing the solution in a more convenient way. Z u(x) = φ(x − y)f (y)dy + C Rn solves −△u = f . Green’s function is the “fundamental solution” to −△u = f in Ω u=g on ∂Ω For later usage: Green’s formula: Z Z u△v − v△udy = Ω u ∂Ω ∂u ∂u − v dS(y). ∂ν ∂ν Multi-dimensional case Let U ⊆ Rn be open and bounded. −△u = f in U u=g on ∂U Let u ∈ C 2 (U ) be any function. Fix x ∈ U such that B(x, ε) ⊂ U. We apply Green’s formula: Vε = U\B(e, ε) to u(y) and φ(y − x). Z Z ∂φ ∂u u(y) (y−x)−φ(y−x) (y)dS(y) u(y)△φ(y−x)−φ(y−x)△u(y)dy = ∂ν ∂ν ∂Vε Vε For the boundary ∂Vε we have to consider both the inner and outer boundary as shown in the picture: We have seen before that: Z ∂u φ(y − x) dS(y) → 0 as ε ↓ 0 ∂ν B(x,ε) In addition, we have also already seen that Z ∂φ u(y) (y − x)d? → u(x) as ε ↓ 0 ∂ν ∂B(x,ε) 32 (These are the inner boundaries in the two terms we get after Green’s function). We know that φ(x) = 0 if x 6= 0. The terms u(y)△φ(y − x) = 0 since △φ(y − x) = 0 when y 6= x and this is the case since we have removed the ball B(x, ε) around x. As ε ↓ 0 Z Z ∂u u(x) = φ(y − x) dS(y) − φ(y − x)△u(y)dy. ∂ν ∂U U and Vε → U when ε → 0. From this expression we have a problem with the term not know this. ∂U ∂ν since we do Correction function Fix x, φx − φx (y) such that △φx = 0 in U, φx = φ(y − x) on ∂U. . (We need this to find a way around the problem with ∂U ∂ν Apply Green’s formula to φx and u (using △φx = 0 on U). Z Z ∂u ∂φx x dS(y) − φx (y) − φ (y)△u(y)dy = u(y) ∂ν ∂ν U ∂U We use that φx = φ(y − x) on ∂U. Z ∂φx ∂u = u(y) dS(y) − φ(y − x) ∂ν ∂ν ∂U Implementing the corrector function: G(x, y) := φ(y − x) − φx (y) for x 6= y Z Z ∂G u(x) = − G(x, y)△u(y)dy − u(y) (x, y)dS(y) ∂ν U ∂U = △yG(x, y) · ν. We have another expression that does not include where ∂G ∂ν ∂u (which is unknown on the boundary). ∂ν Finally we get, Z Z ∂G dS(y) u(x) = f (y)G(x, y)dy − g(y) ∂ν U ∂U 33 —————————————————————- 07/09-2010 For u ∈ C 2 (U) −△u = f in U u=g on ∂U the solution has the form Z ∂G + u(x) − g(y) ∂ν ∂U Z f (y)G(x, y)dy. U We have G(x, y) = φ(x − y) − φx (y) for △y φx (y) = 0 U x φ (y) = φ(y − x) y ∈ ∂U Theorem Green’s function is symmetric: G(x, y) = G(y, x). Proof We set v(z) = G(x, z), and w(z) = G(y, z). We want to check if G(x, y) = v(y) = w(x) = G(y, x). We have: △v = △w = 0 since x 6= z and y 6= z for the first and second equality, respectively. We consider the set U\ B(x, ε) ∪ B(y, ε) = V depicted in the following image. 0= Z 0= 0= Z Z V v△w − w△vdx v ∂w ∂v − w dS ∂ν ∂ν v ∂v ∂w − w dS ∂ν ∂ν ∂B(y,ε) ∂B(x,ε) 34 Z = V Z DvDw − DwDvdx w B(x,ε) Z ∂v ∂w dS = − v ∂ν ∂ν | {z } Vanishes v ∂B(y,ε) ∂w ∂v − w dS ∂ν ∂ν We look closer at the problematic part: Z ∂w 1 ≤ C n−2 εn−1 → 0 when ε → 0. v ∂ν ε As we have shown before: lim ε→0 Similarly, Z w Z v ∂B(x,ε) ∂B(y,ε) ∂φ dS = w(x). ∂ν ∂w → v(y). ∂ν We now know that y 7→ G(x, y) is harmonic, and x 7→ G(y, x) is harmonic, as well as G itself. Green’s function for a half space We will end up with a general solution for the Laplacian problem. Green’s function for R+ n (like the upper plane for n = 2 or the positive octant in n = 3. Formally we write Rn+ = {x ∈ Rn |xn > 0}. For the half-space we get the boundary problem: −△u = f x ∈ Rn+ u=g x ∈ ∂Rn+ for which we can find the general solution. To construct this solution we solve the auxiliary problem: −△y φx = 0 Rn+ x φ (y) = φ(x − y) Rn+ For the vector x = (x1 , . . . , xn ) we define x e = (x1 , . . . , −xn ), which is a reflection of x as shown in the following image. 35 Claim: φx (y) = φ(y − x e). This is harmonic. Consider y ∈ ∂Rn+ . φ is constant on the surface: φ(x − y) = φ(|x − y|): Since φ is a function of the magnitude, and not the point, φ(x−y) = φ(y −e x). A candidate for the solution is Z ∂G (y)g(y)dS(y) u(x) = − ∂ν Rn + and we have and G(x, y) = φ(x − y) − φ(y − x e) ∂G ∂G =− ∂ν ∂yn ∂G xn − yn 1 −xn − yn n =− − | ∂ν nα(n) |x − y|n |e x−y and using that |e x − y|n = |x − y|n on the boundary, == 1 ∂xn · . nα(n) |x − y|n The candidate is on the form ∂xn u(x) = nα(n) Z R⋉−1 g(y) dy |x − y|n and apparently we have convolution on the kernel, and we introduce the Poisson kernel: Z = K(x, y)g(y)dy Rn−1 which is a candidate, not a solution. 36 Theorem We assume g ∈ C 1 (Rn−1 ) ∩ L∞ (Rn−1 ). Then (i) u ∈ C ∞ (Rn−1 ) ∩ L∞ (Rn+ ) (ii) △u = 0 in R+ n (iii) For x ∈ Rn+ , we have limx→xo u(x) = g(x0 ) for x0 ∈ Rn−1 . (When you approach a point on the boundary from inside, you get the right path). Proof For x ∈ Rn+ , y 7→ K(x, y) is C ∞ (∂Rn+ ). K is a smooth function on the boundary as long as x is not on the boundary. Thus, u ∈ C ∞ (Rn+ . The integral of K, the kernel, is exactly 1. Z Z Z ∞ 1 2xn 2xn (n − 1)α(n − 1)r n−2 K(x, y)dy = dy = dr n nα(n) Rn−1 |x − y|n nα(n) 0 (x2n + r 2 ) 2 Rn−1 ! Z 2xn (n − 1)α(n − 1) ∞ 1 r n−2 dr = nα(n) xn 1 + ( xrn )2 0 Substitution, s = r , xn dr = xn ds x x 2 n n (n − 1)α(n − 1) = n nα(n)xn 1 , cos2 θ Second substitution, s = tan θ, ds = 2(n − 1)α(n − 1) ⇒ nα(n) Z π 2 ∞ sn−2 xn−2 n 0 (1 + s2 ) 2 1 + s2 = 1 . cos2 θ sinn−2 θ 1 n cos θ dθ cosn−2 θ cos2 θ 0 2(n − 1)α(n − 1) = nα(n) Z Z π 2 sinn−2 θdθ = 1 0 The step showing that this equals 1 is a proof by induction. It is quite easily verified for n = 2, as α(n) is the volume/area/length of the unit circle in Rn : 2α(1) 2α(2) Z π 2 1dθ = 0 2(2) π = 1. 2π 2 Since g is bounded by assumption and the integral of the kernel is 1, we have Z u(x) = K(x, y)g(y)dy ≤ kgkL∞ . Rn−1 37 By definition y 7→ K(x, y) is harmonic, and K= ∂G ∂φx − . ∂yn ∂yn Since K(x, y) = K(y, x), it follows that x 7→ K(y, x) is harmonic as well. Using that g is continuous |y − x0 | ≤ δ ⇒ |g(y) − g(x0 )| < ε. δ 0 We choose some x ∈ R+ n such that |x − x | ≤ 2 . We then verify (iii) in the proof by considering Z 0 |u(x) − g(x )| = K(x, y) g(y) − g(x0 ) dy Splitting the integral in two parts. Z Z 0 ≤ I +J = K(x, y) g(y) −g(x ) dy + |x0 −y|≤ε |x0 −y|>ε K(x, y)g(y) −g(x0)dy The integral of the kernel is 1, and by continuity, |g(y) − g(x0 )| ≤ ε, so I ≤ ε, so it is arbitrarily small. xn What happens when y is far away from x0 ? Since K(x, y) = |x−y| n , when |y − x| is large, K becomes smaller. Furthermore, |y −x0 | = |y −x+x−x0 | ≤ |y −x|+|x−x0 | < |y −x|+ δ 1 ≤ |y −x0 |+|y −x|. 2 2 In short |y − x0 | ≤ 21 |y − x0 | + |y − x| so 21 |y − x0 | ≤ |y − x|, and we have removed the δ-dependence. Since g is bounded, we can assume the ’worst case scenario’ and set |g(y) − g(x0 )| ≤ 2kgkL∞ , so Z 1 2n+2 kgkL∞ xn dy → 0 J ≤ 2kgkL∞ K(x, y)dy ≤ nα(n) |y − x0 |n |x0 −y|>ε as xn → 0+ . Both parts of the integral tend to 0, so the difference between u(x) and g(x0 ) becomes arbitrarily small, so u(x) → g(x0 ) as x → x0 . 38 —————————————————————- 10/09-2010 Theorem (Poisson’s formula for a ball) Assume g ∈ C(∂B(0, r)) and define u by Z r 2 − |x|2 g(y) u(x) = dS(y) nα(n)r ∂B(0,r) |x − y|n then, (i) u ∈ C ∞ (B(0, r) (ii) △u = 0 in B(0, r) (iii) For x ∈ B(0, r), limx→x0 u(x) = g(x0 ) for all x0 ∈ ∂B(0, r). Proof If we can prove (ii), (i) follows since all harmonic functions are C ∞ . We see that ∀x the mapping y 7→ G(x, y) is harmonic when y 6= x, and by symmetry x 7→ G(x, y) is harmonic. Hence, x 7→ Since G is harmonic, P ∂G ∂yi △x X ∂G ∂yi yi = −K(x, y). is harmonic. X ∂G u(x) = ∂yi Z yi = X ∂ △x G yi ∂yi K(x, y)g(y)dS =⇒ ∂B(0,r) △x u(x) = Z ∂B(0,r) △x K(x, y)g(y)dS = 0 which proves that (ii) u is harmonic and it follows that (i) u ∈ C ∞ B(0, r) . From the expression for u(x), we have the Kernel, Z r 2 − |x|2 1 K(x, y) = . nα(n)r ∂B(0,r) |x − y|n From the maximum principle, we know that the boundary problem: −△u = f in B(0, r) u=g on ∂B(0, r) 39 has a unique solution. From assumption, g ∈ C ∂B(0, r) , which means that for y, x ∈ ∂B(0, r), |y − x0 | < δ =⇒ |g(y) − g(x0 )| < ε. Having established this, we examine Z 0 0 K(x, y) g(y) − g(x ) dS |u(x) − g(x )| = ∂B(0,r) We split the integral in two: Z Z 0 ≤ K(x, y) g(y)−g(x ) dS+ ∂B(x,δ)∩∂B(x0 ,r) ∂B(x,δ)\∂B(x0 ,r) K(x, y)g(y)−g(x0)dS which we call I1 + I2 , and they are the surface integrals in the image: By continuity we have |g(y) − g(x0 )| < ε, and the surface integral of K(x, y) is 1, so Z <ε+ K(x, y)g(y) − g(x0 )dS ∂B(x,δ)\∂B(0,r) Taking a closer look at the boundary: And, as on page 39. in the text book, we get |y − x| ≥ 21 |y − x0 |, and when we return to the integral, we get Z r 2 − |x|2 4 ∞ dS → 0 <ε+ · kgk K |y − x0 | ∂B(x,δ)\∂B(x0 ,r) nα(n)r as x → x0 , which implies |x| → r, and it follows that r 2 − |x|2 → 0. 40 Energy Methods Again we specify the Poisson boundary problem and denote it (P). −△u = f in B(0, r) (P) u=g on ∂B(0, r) We assume ∂U is C 1 and u ∈ C 2 U . To prove that the solution is unique: assume u, v are solutions and set w = v − u. −△w = 0 in U w=0 on ∂U Z Z Z ∂w Green’s formula 0= w△w == w dS − |Dw|2dx ∂ν U ∂U U = 0, the first term vanishes, and we are left with |Dw| = 0 in U. Since ∂w ∂ν There is no gradient so w must be flat: w(y) = w(x) for x ∈ U and y ∈ ∂U We know from this that there is at most one solution (there is of course the possibility that there is no solution). We have A which is a class of functions, and we define it as o n 2 A = u|u ∈ C (C), u = g on ∂U or in other words, any function taking the values we want. Z 1 I u := |Du|2 − f udu 2 U Theorem We have the equivalence u solves (P) ⇐⇒ I[u] ≤ I[w] ∀w ∈ A. In physics this is the “least energy” or “minimized energy property”. Proof Implications from left to right. We assume u solves (P) and w ∈ A. Z Z 0 = (−△u − f )(u − w)dx = Du · (Du − Dw) − f u + f wdx U U = Z U |Du|2 − DuDw − f u + f wdx We use a very important, elementary inequality. It is obvious that 12 (a − b)2 ≥ 0, which is true both for numbers and vectors. Multiplying (or dot 41 multiplying) and shifting terms yields 12 a2 + 12 b2 ≥ ab. using this, we get −DuDw ≥ − 21 (|Du|2 + |Dw|2): Z 1 1 ≥ |Du|2 − |Du|2 − |Dw|2 − f u + f wdx 2 2 U Z Z 1 1 2 |Du| − f udx − |Dw|2 − f wdx = I[u] − I[w]. = U 2 U 2 In summary 0 ≥ I[u] − I[w] which leaves us with I[u] ≤ I[w]. For the converse we assume we have a minimiser. For a small number τ and v ∈ C0∞ (U), then u + τ v ∈ A, we define i(τ ) := I[u + τ v], for i(0) ≤ i(τ ) for τ 6= 0 and i′′ (0) = 0. Z i 2 d h 1 ′ i (τ ) = Du + τ Dv − f (u + τ v) dx U dt 2 Z i τ2 d h 1 Du|2 + τ Du · Dv + |Dv|2 − f u − τ f v dx = 2 U dt 2 We now have a polynomial, which is easy to differentiate. Z = DuDv + τ |Dv|2 − f vdx U ′ 0 = i (0) = Z U Integration by parts. Z DuvdS + Z Z U ∂U DuDv − f vdx U −△u · v − f vdx Using that v = 0 on the boundary, Z = (−△u − f )vdx = 0. U This is quite strong, as the only assumption we made was that v ∈ C0∞ (U), so Z wvdx = 0∀v ∈ C0∞ (U) =⇒ w = 0. U (The trick is that v ≈ w). 42 —————————————————————- 14/10-2010 43 —————————————————————- 17/10-2010 44 —————————————————————- 21/10-2010 Weak Derivative We say v = D α u in the weak sense if for u, v ∈ L1loc (U) we have Z Z α |α| vφdx ∀φ ∈ Cc∞ (U). u · D φdx = (−1) U U Sobolev Space The definition of the Sobolev space is that if u ∈ W k,p(U), then D α u exists in the weak sense for all |α| ≤ k. We write kukW k,p(U ) < ∞, and we define X kukW k,p (U ) = |α|≤k kukW k,p(U ) = kD α ukpLp (U ) X |α|≤k p1 1 ≤ p < ∞. kD α ukL∞ (U ) p = ∞. As we have shown earlier, k · kW k,p (U ) is a norm. Example Let U = B(0, 1) ⊂ R and let u(x) = |x|−a . For what values of a > 0, n, p is u ∈ W 1,p (U)? We recall the definition of the weak derivative, which for k = 1 amounts to: u ∈ W k,p (U) which means D α u ∈ Lp (U) for all |α| ≤ 1. By the definition of Sobolev norms, kukpW 1,p (U ) = X |α|≤1 kD α ukpLp (U ) k = | {z } Z p B(0,1) |u| dx = Z B(0,1) We utilize the coarea formula. Z 1Z Z tZ −ap |x| dS dr = 0 + α=0 Calculating the first term: kukpLp (U ) = kukpLp (U ) ∂B(0,r) 0 45 n X |i=1 kuxi kpLp (U ) {z α=1 |x|−ap dx ∂B(0,r) r −ap dSdr } We use that |∂B(0, r)| = cr n−1 , so Z 1 =C r −ap+n−1dr. 0 This is finite if −ap + n − 1 > −1 which is equivalent to saying ap < n. We move on to the sum of the first derivatives, and start by calculating what it is. X a −ax − uxi = ∂xi |x|−a = ∂xxi x2i 2 = a+2 |x| which implies that |Du(x)| = |x|aa+1 . We now calculate the Lp -norm for this gradient. Z Z Coarea p |Du(x)| dx = a |x|−(a+1)p dx == B(0,1) a Z 1Z 0 B(0,1) r −(a+1)p dSdr = aC ∂B(0,r) Z 1 r −(a+1)p+n−1 dr. 0 This is finite if −(a + 1)p + n − 1 > −1 which is equivalent to (a + 1)p < n. Thus, u, Du are Lp (U) if (n + 1)p < n (as this is a stricter condition than the first). (If u, Du ∈ Lp (U), then by definition they are in W 1,p (U). (2) Let φ ∈ Cc∞ (U) and fix ε > 0. Define Vε = U\B(0, ε). <BILDE> Note that Vε → U as ε → 0. Using integration by parts; Z Z Z uφν i dSx . uxi φdx + uφxi dx = − ∂Vε Vε Vε Vε has two boundaries: an inner and an outer boundary: ∂Vε = ∂U ∪∂B(0, ε), but by compact support φ(∂U) = 0, so the surface integral of the outer boundary vanishes. Z Z uxi φdx + uφν idSx . =− ∂B(0,ε) Vε We estimate the second term. Z Z i ≤ uφν dS x ∂B(0,ε) ∂B(0,ε) 46 |x|−a |φ|dSx ≤ (since ∂B(0, ε) ⇔ |x| = ε) Z kφkL∞ ε−a dSx ≤ Cε−a+n−1 → 0 as ε → 0, if a + 1 ≤ n. ∂B(0,ε) Moreover, u, Du ∈ L1 (U) if a + 1 < n. Taking the limit ε → 0. Z Z Z uxi φdx + lim uφν i dSx uφxi dx = − lim lim ε→0 ε→0 Vε ε→0 Vε ∂B(0,ε= (we can put the limit inside the integral) Z Z Z Z α uφxi dx = − uxi φdx + 0 =⇒ uD φdx = − D α uφdx U U U U for (a + 1) < n. Thus, • Du exists in the weak sense if a + 1 < n. • u, Du ∈ Lp if (a + 1)p < n. and based on this we know that u ∈ W 1,p (U) if (a + 1)p < n or a < n−p . p Theorem: Properties of Weak Derivatives Assume u, v ∈ W k,p (U), |α| ≤ k. Then (i) D α u ∈ W k−|α|,p(U) and D β (D α u) = D α (D β u) = D α+β u. for all α, β where |α| + |β| ≤ k. (ii) For each λ, µ ∈ R, λu + µv ∈ W k,p (U). and D α (λu + µv) = λD u + µD α v for all |α| ≤ k. α (iii) If V ⊂ U, then u ∈ W k,p (V ). P β≤α (iv) ξ ∈ Cc∞ (U) then ξu βIf α−β α α! D ξD u where αβ = β!(α−β)! . β ∈ W k,p (U) and D α (ξu) Proof (i) Fix φ ∈ Cc∞ (U), then D β φ ∈ Cc∞ (U) is also a test function. Z Z Z α β |α| |α| α β uD α+β φdx. uD (D φdx) = (−1) D u D φ dx = (−1) |{z} U U U = Inf. diff. (The partial derivatives of φ is the normal derivative, not the weak kind). We know that u is k-times differentiable and |α + β| ≤ k. Z Z α+β β |α+β| α D α+β uφdx =⇒ D uφdx = (−1) = (−1) (−1) U U 47 Z α β |β| D D φdx = (−1) Z Z U U D α+β uφdx =⇒ U D β (D α u) = D α+β u in the weak sense. For the other claim. kD α ukpW k−|α|,p (U ) = X |β|≤k−|α| kD β (D α u)kpLp (U ) = X |α+β≤k kD α+β ukpLp (U ) = kukW k,p (U ) . (ii) Fix φ ∈ Cc∞ (U). Z Z Z α α (λu + µv)D φdx = λ uD φdx + µ vD αφdx = U U |α| (−1) (−1)|α| Z Z U α λD uφdx + (−1) |α| Z µD α vφdx = U U (λD α u + µD α v)φdx =⇒ D α (λu + µv) = λD α u + µD α v. U Clearly, kλu + µvkW k,p(U ) ≤ |λ|kukW k,p(U ) + |µ|kvkW k,p(U ) , where we used that k · kW k,p (U ) is a norm. (iii) Let φ ∈ Cc∞ (V ), then φ ∈ Cc∞ (U). Z Z α |α| D α uφdx uD φdx = (−1) U U φ: only defined for V (??). Z Z α |α| D α uφdx. uD φdx = (−1) V V Justifies that u has weak derivatives. Also, clearly, kukW k,p(V ) ≤ kukW k,p (U ) . (iv) We use an induction argument on |α|. First assume that |α| = 1, then for any φ ∈ Cc∞ (U): Z Z α (ξu)D φdx = u ξD α φ dx | {z } U U D α (ξφ)−D α (ξ)φ = Z α U uD (ξφ)dx − Z U α uD ξφdx = − 48 Z U α D u · ξφdx − Z U uD α ξ · φdx = − Z (D α ξu + ξDα u) φdx, U and by definition of the weak derivative: X α β Dξ D α−β u D (ξu) = D ξ · u + ξD u = β β≤α α α α (and we have assumed |α| = 1). We assume this holds for l < k and the formula is valid for all |α| ≤ l. Choose α with |α| = l + 1 which implies α = β + γ where |β| = l and γ = 1. Claim: the formula is valid for all such α. Z α ξuD φdx = U Z β γ |β| ξuD (D φ)dx = (−1) |{z} U ∈Cc∞ Z D β (ξu)D αφdx U and by the induction assumption, Z X β |β| D β ξDβ−σ uD γ φdx. = (−1) U σ≤β σ Our goal now is to try and transfer the α outside.. Z X β |β|+|γ| D γ (D σ ξDβ−σ u)φdx = (−1) σ U σ≤β Since |γ| = 1 we can use the formula from the induction step. Z X β γ+σ β−σ D ξD u + D β ξDβ+γ−σu φdx = (−1)|α| | {z } | {z } σ U σ≤β (I) X β σ≤β σ D γ+σ ξD I β−σ II X ρ=σ+γ u == γ≤ρ≤β+γ β D ρ ξDβ+γ−ρu ρ−γ Starts from order 1, so |γ| = 1, X β D ρ ξDβ+γ−ρ u + D β+γ ξu = IRes. ρ−γ γ≤ρ≤β+γ 49 (II) X β X β σ β+γ−σ β+γ−σ D σ ξDβ+γ−σ = IIRes D ξD u = ξD u+ σ σ γ≤σ≤β σ≤β Adding these together. X β β D σ ξDβ+γ−σ u + ξDβ+γ−σ u + D β+γ ξu. + IRes + IIRes = σ σ−γ γ≤σ≤β and using β σ−γ + β σ β+γ σ = , we get X β + γ D γ ξDβ+γ−σ u + ξDβ+γ u + D β+γ ξu = σ γ≤σ≤β X α X β + γ γ β+γ−σ D α ξDα−σ u D ξD u= = σ σ σ≤γ σ≤β+γ Therefore, Z ξDα φdx = (−1)|α| Z U U ! X α D α ξDα−γ uφ dx σ σ≤α which, finally, implies, X α D σ ξDα−σ u. D (ξu) = σ σ≤α α Theorem Sobolev Spaces as function spaces For each k = 1, . . . and 1 ≤ p ≤ ∞, the Sobolev space is a Banach space. Proof We know that k · kW k,p (U ) is a norm, so it only remains to prove that the Sobolev space is complete, i.e for every Cauchy sequence in W k,p (U) converges to some function in the same space W k,p (U). Assume that {um } is a Caucy sequence in W k,p(U). Then {D α um } is Cauchy in Lp (U). Since kum − un kW k,p (U ) → 0 as l, m → ∞ by definition we have kD α um − D α ul kLp (U ) → 0 as l, m → ∞. 50 By completeness of Lp (U), D α um → uα in Lp (U). In particular, take α = (0, . . . , 0), then um → u(0,...,0) in Lp (U). We define u = u(0,...,0) and claim that uα = D α u. If we can show this, we have shown that D α um → D α u ⇒ um → u in W k,p(U) for |α| ≤ k. To show this, we first fix a φ ∈ Cc∞ (U). Then Z Z α uD φdx = lim um D α φdx. R R m→∞ U R U (where u = lim um = lim um ). The functions un have weak derivatives up to order k, so Z Z α |α| |α| uα φdx =⇒ = (−1) lim D um φdx = (−1) n→∞ U U D α u = uα for all |α| ≤ k, hence D α um → D α u in Lp (U), which implies um → u in W k,p (U), since by defintion X kum − ukW k,p (U ) = kD α um − D α ukpLp (U ) . In conclusion: W k,p (U) is a Banach space. 51 —————————————————————- 24/10-2010 52 —————————————————————- 28/09-2010 Smooth Approximation Recall • • ∃ηε ∈ Cc∞ B(0, ε) , Z ηε dx = 1. Rn Uε = x ∈ U | dist(x, ∂U) > ε Theorem - Local Approximation by Smooth functions For some integer k ≥ 0 and 1 ≤ p < ∞, let u ∈ W k,p(U). Set uε = ηε ∗ u in Uε . Then, (i) uε ∈ Cc∞ (Uε ) for each ε > 0 (ii) k,p uε −→ u in Wloc (U). (We showed a related result for Lp spaces, but this theorem states it is also true for Sobolev spaces). Proof Proving (i). k=0 u ∈ W k,p (U) =⇒ u ∈ Lp (U) =⇒ uε ∈ C ∞ (Uε ) where the last implication follows from the properties of mollifiers. Proving (ii). We begin proving point (ii) with verifying the following claim. For all |α| ≤ k, α ε Dαu D | {zu} = ηε ∗ |{z} Norm. der Weak der. The left side derivatives are normal, partial derivatives since uε is a smooth function, and the derivatives on the right side are the weak derivatives as defined for Sobolev spaces. Proof of claim By the definition of mollifiers. ε u (x) = Z U ηε (x − y)u(y)dy, 53 so α ε D u (x) = Z U Dxα ηε (x − y)u(y)dy where we have, by the definition of weak derivatives (?), Dxα ηε (x − y) = (−1)|α| Dyα ηε (x − y) and plugging this in the first expression, Z α ε |α| Dyα ηε (x − y)u(y)dy. D u (x) = (−1) U But for some fixed x ∈ Uε , ηε (x − y) ∈ Cc∞ (U), so Z α |α| |α| ηε (x − y)D u(y)dy (−1) = (−1) U We get (−1)2|α| = 1, and we are left with the desired result: Z ηε (x − y)D αu(y)dy = ηε ∗ D α u(x). U By assumption and construction of Sobolev spaces, D α u(x) ∈ Lp (Uε ) hence, By definition, D α uε −→ D α u in Lp (V ), V ⊂⊂ U. keε − ukpW k,p(V ) = X |α|≤k kD α uε − D α ukpLp (V ) → 0 k,p as ε → 0, so we can conclude uε → u in Wloc (U). Theorem - Global approx. by smooth functions. Assume U is bouned and suppose u ∈ W k,p(U), 1 ≤ p < ∞. Then there exists functions um ∈ C ∞ (U) ∩ W k,p(U) such that um → u as m → ∞ in W k,p(U). Proof (Difficult) (1) Define 1 Ui = x ∈ U | dist(x, ∂U) > i for i = 1, 2, . . ., so Ui ⊂ Ui+1 . 54 We write Vi = Ui+3 − U i , and we get the sets as indicated by the image. S We choose V0 ⊂⊂ U such that U = ∞ i=0 Vi , e.g V0 = U3 . (2) Let {ξi }∞ i=0 be a smooth partition of unity subordinate to the open sets ∞ {Vi }i=0 , that is, suppose 0 ≤ ξi P ≤ 1, ξi ∈ Cc∞ (Vi ), i ∈ N ∞ i=0 ξi = 1 on U. (This is a theorem from Topology). Now, if u ∈ W k,p(U), then ξi u ∈ W k,p (U) (we have already proved this), and supp(ξi u)⊂ Vi . (3) Define U i = ηεi ∗ (ξi u) for each εi > 0, for i = 0, 1, . . .. Note the following: fro sufficiently small εi > 0, then supp(U i ) ⊂ Wi , i = 1, 2, . . . where Wi = U+4 − U . The support of the product ξi u is the support of of the sum of the sets each of them have support. Then, for ξi > 0, δ kui − ξi ukW k,p (U ) ≤ i+1 (*) 2 for some small δ > 0, i = 0, 1, . . .. By the local approximation theorem, kui − ξiuk converges to 0. (4) Define v= ∞ X ui i=0 for a fixed x ∈ U. For each V ⊂⊂ U, there are at most finitely many non-zero terms in the sum. For example, if x ∈ W1 , then ui (x) = 0 for i ≥ 4, and if x ∈ Wj then ui(x) for |i − j| ≥ 4. Because of this the sum on the set V will be finite, and it follows that the full sum is finite. Using the finite sum and that ui ∈ C ∞ (U), the function v is smooth. We will use the following: u=1·u= 55 ∞ X i=1 ξi u. We want to approximate v by y. For V ⊂⊂ U, kv − ukW k,p (V ) ≤ ∞ X i=0 i ku − ξi ukW k,p (V ) ∞ X δ ≤ =δ i+1 2 i=0 The first inequality is the triangle inequality, and the second one follows from (*). kv − ukW k,p (U ) = sup kv − ukW k,p (V ) ≤ δ. V ⊂⊂U We have proved theorem: v ∈ C ∞ (U) ∩ W k,p (U). Theorem - Global approx. by smooth functions up to the boundary. Assume U is bounded and ∂U ∈ C 1 . Suppose u ∈ W k,p (U) for 1 ≤ p < ∞, then ∃um ∈ C ∞ (U ) such that um → u in W k,p(U). Proof (1) Fix x0 ∈ ∂U. Since ∂U ∈ C 1 , ∃r > 0 and a C 1 function γ : Rn−1 7→ R such that U ∩ B(x0 , r) = x ∈ B(x0 , r) | xn > γ(x′ ) where x′ = (x1 , . . . , xn−1 ). (This follows since the boundary is C 1 , see appendix). We call V = U ∩ B(x0 , 2r ). See image. (2) Now define xε = x + λεen , (x ∈ V, ε > 0, λ > 0) and en is (0, . . . , 0, 1). We only add λ to the last component. 56 Claim B(xε , ε) ⊂ U ∩ B(x0 , r) ⇔ y ∈ B(xε , ε) ⇒ y ∈ B(x0 , r) for yn < γ(y ′ ) and y ′ = (y1 , . . . , yn−1). To verify this claim: we can write y ∈ B(xε , ε) as y = xε + εw for |w| = 1. |y − x0 | = |xε + εw − x0 | = |x + λεen + εw − x0 | ≤ |x − x0 | + ε|λ + 1| < r since |x − x0 | < 2r and ε|λ + 1| < within the ball. We have r 2 for sufficietly small ε, thus we know y is w1 .. ε y = x + εw = x + λεen + εw) = x + ε(λen + w = x + ε . wn − 1 λ + wn so, y1 x1 + εw1 .. y = ... = . yn xn + ε(λ + wn ) where we note that yn = xn + ε(λ + wn ). We can approximate γ(y ′ ) = γ(x′ + εw ′ ) ≤ γ(x′ ) + εM for some M > 0, since Z ε Z ε d ′ ′ ′ ′ ′ (∂(x +ετ )dτ = γ(x )+ ∇γ ·wdτ ≤ γ(x′ )+εM. γ(x +εw ) = γ(x )+ 0 0 dτ Since x ∈ V we also have γ(x′ ) < xn so we can conclude that γ(y ′) < xn +εM. From earlier we have yn = xn + ε(λ + wn ) ⇒ xn = yn − ε(λ + wn ) ⇒ γ(y ′) = yn − ε(λ + wn )0εM = yn − ε(λ + wn − M). If we choose λ > 0 so large that λ + wn > M, we get yn > γ(y ′ ). Claim shown. (3) We define uε (x) := u(xε ), for x ∈ V , and we write v ε = ηε ∗ uε . We do this to get away from the boundary since the mollification vanishes at the boundary due to the support. We mollify at the point xε instead, which is well within the ball. Claim: v ε ∈ C ∞ (V ). 57 We will verify this. By definition of mollifiers; Z ε v (x) = ηε ∗ uε (x) = ηε (x − y)uε(y)dy B(x,ε) and by our definition: uε (x) = u(xε ), we get Z Z ε = ηε (x − y)u(y )dy = ηε (x − y)u(y + λεen )dy B(x,ε) B(x,ε) Change of variable, z = y + λεen , Z Z = ηε (x + λεen − z)u(z)dz = B(x+λεen ,ε) B(xε ,ε) ηε (xε − y)u(y)dz = ηε ∗ u(xε ). So, we have shown v ε (x) = ηε ∗ u(xε ). Now x ∈ V implies xε ∈ W ⊂⊂ U which again implies v ε ∈ C ∞ (V ). Claim shown. (4) Claim: v ε → u in W k,p(V ). We look at all α, where |α| ≤ k. kD α v ε − D α ukLp (V ) ≤ kD α v ε − D α uε kLp (V ) + kD α uε − D α ukLp (V ) The first term tends to 0 as ε → 0 since uε is the mollification, and the second term tends to 0 as ε → 0 by the local approximation theorem. We need that Lp is continuous under translation, that is: Z |f (y) − f (x)|p dx −→ 0 as y → x. We have uε = eε (x) = u(xε ) = u(x + λεen ) which goes to u(x) as ε → 0, and since Lp is continuous under translation, kD α uε − D α ukLp (V ) → 0 as ε → 0. (5) Since ∂U is compact, we can find finitely many xi0 ∈ ∂U, ri > 0 and corresponding sets Vi = U ∩ B(xi0 , r2i ) and functions vi ∈ C ∞ (V i ) for i = 1, . . . , N. such that N [ ri ∂U ⊂ B(xi0 , ) 2 i=1 58 (that is, we can cover the entire boundary with a finite amount of balls Vi ), and for each such ball Vi the function vi satisfies: kvi − ukW k,p (Vi ) ≤ δ. (**) (For each Vi we aruge in the exact same way as we did for V , only repeating it N times). S ∞ We choose V0 ⊂⊂ U such that U ⊂ N i=0 Vi and select v0 ∈ C (V 0 ) satisfying kv0 − ukW k,p (V0 ) ≤ δ. (***) (which we can do by a previous theorem). (6) Let {ξi }N i=0 be a smooth partition of unity on U, subordinate to the sets ri N V0 , B(xi0 , ) i=1 . 2 Define v= We write u = PN i=0 i=0 ξi u. α N X ξi vi =⇒ v ∈ C ∞ (U ). For all |α| ≤ k α kD v − D ukLp (U ) ≤ N X i=0 kD α (ξi vi ) − D α (ξi u)kLp (U ) which follows by the triangle inequality. = N X i=0 kD α ξi (vi − u) kLp (U ) where we can write α D ξi (vi − u) = so we have X β β≤α α D β ξi − D α−β (vi − w) N X X β β α−β D ξ − D (v − u) = i i α i=0 β≤α Lp (U ) ≤ N X X β i=0 β≤α | α kDξαi kL∞ (U ) · kD α−β (vi − u)kLp (U ) | {z } {z } ≤kvi −uk k,p W ≤C∈R 59 (U ) ≤C N X i=0 kvi − ukW k,p(U ) ≤ CNδ. In the last inequality we used (**) and (***), and since δ can be arbitrarily small, this means v → u in W k,p(U). 60 —————————————————————- 01/10-2010 2.5.8 We have the boundary conditions △u = 0 in Rn+ u = g on ∂Rn+ Assume g ∈ L∞ (bounded) and g(x) = |x| for x ∈ ∂Rn+ and |x| ≤ 1. Show that the gradient Du is not bounded near x = 0. The solution u(x) si given by Poisson’s formula for a half space, so we have Z g(y) 2xn dS. (x ∈ Rn+ ) u(x) = nα(n) ∂Rn+ |x − y|n We are given a hint, which is to estimate (en denotes the n’th unit vector: en = (0, . . . , 0, 1) ∈ Rn ) u(λen ) − u(0) . λ Clearly u(0) = 0. Z g(y) 2λ dS (x ∈ Rn+ ). u(λen ) = nα(n) ∂Rn+ |λen − y|n We have: y ∈ ∂Rn+ ⇔ y = (y1 , . . . , yn−1, 0) ⇒ λen − y = (−y1 , . . . , −yn−1 , λ). 2 |λen − y| = y12 + y22 + . . . + yn−1 + λ2 n 21 = |y| + λ2 21 so |λen − y|n = (λ2 + |y|2) 2 . Returning to the function, we have Z 2λ g(y) u(λen ) = n dS. 2 nα(n) ∂Rn+ (λ + |y|2) 2 61 It follows that Z u(λen ) − u(0) g(y) 2 = u(λen ) = n dS. 2 λ nα(n) ∂Rn+ (λ + |y|2) 2 By a definition from Evans 2 = nα(n) Z Rn−1 (λ2 g(y) n−1 y. n d 2 + |y| ) 2 Now we split the integral in two parts, and use what we know about u. Z Z 2 |y| g(y) n−1 n−1 = y+ y n d n d 2 2 nα(n) |y|≤1 (λ2 + |y|2) 2 |y|>1 (λ + |y| ) 2 for y ∈ Rn−1 . However, now we see that, from the first integral, as λ → 0, (λ2 1 |y| . n → 2 2 |y|n−1 + |y| ) If we integrate this limit (which we can due to monotone convergence theorem), using the coarea formula, Z Z 1Z 1 1 n−1 d y = dS dt = n−1 n−1 y≤1 |y| 0 |y|=r r Z 1 Z 1 1 1 n−2 Cr dr = C dr = +∞. n−1 0 r 0 r What we have calculated is the derivative of xn , and uxn = ∞ =⇒ Du unbounded. 5.4 Extensions Extensions of W 1,p (U) to W 1,p (Rn ). Assume U is bounded and ∂U ∈ C 1 . Choose V ⊃⊃ U for a bounded V . Then there exists a bounded linear operator E : W 1,p (U) → W 1,p (Rn ) such that for each u ∈ W 1,p (U): (i) Eu = u a.e in U. (ii) supp(Eu) ⊂ V . (iii) kEukW 1,p (Rn ) ≤ CkukW 1,p (U ) for C ∈ R and C only depends on n,p and U. Definition: Eu is called the extension of u to Rn . Proof Fix x0 ∈ ∂U and assume first that ∂U is flat near x0 , lying in the plane {xn = 0}. 62 For r ≥ 0, let B + = B(x0 , r) ∩ {xn ≥ 0} B − = B(x0 , r) ∩ {xn ≤ 0} where B + ⊂ U , and B − ⊂ Rn \U. We make the temporary assumption u ∈ C 1 U, and then we define u(x) x ∈ B+ u(x) := −3u(x1 , . . . , xn−1 , −xn ) + 4u(x1 , . . . , xn−1 , − x2n ) x ∈ B − so for some x ∈ B − , we get u(x) ∈ B + (a higher order reflection). (1) Claim: u ∈ C 1 (B) (where B = B + ∪ B − ). To prove this claim, we write u+ = uB+ . u− = uB− , It suffices to show that D α u− {xn =0} = D α u+ {xn =0} , ∀|α| ≤ 1. If we can show this, the claim follows. To show it, we first observe that − u+ = u− on {xn = 0}. (−3u(0) + 4u(0) = u(0)), and u+ xi = uxi on {xn = 0} for i = 1, . . . , n − 1, and for i = n we verify this directly: xn u− ) xi (x) = u B − = 3uxn (x1 , . . . , xn−1 , −xn ) − 2uxn (x1 , . . . , xn−1 , − 2 so, we have + u− = u = u x n {xn =0} xn {xn =0} xn {xn =0} 63 and therefore the claim follows. (2) Claim: kukW 1,p (B) ≤ CkukW 1,p (B+ ) . By definition, kukW 1,p (B) = where, and Z Z B− X kalpha≤1 p B kD |u| dx = − p α ukpLp (B) Z |u (x)| dx = = Z p B |u| dx + p + | u (x) | dx + B + | {z } u(x) Z B− n Z X i=1 B |uxi (x)|p dx, |u− (x)|p dx Z − 3u(x′ , xn ) + 4u(x′ , − xn p dx. 2 B− Using the inequality |a + b|p ≤ C(|a|p + |b|p ), we get Z Z xn p ′ p ′ C |u(x , −xn )| dx + |u(x , − )| dx 2 B− B− and using a change of variable, ≤C Z B+ |u(x)|p dx. So, we have established that Z Z p |u(x)| dx ≤ C B+ |u(x)|p dx B+ |uxi (x)|p dx, B and by a similar argument, Z Z p |uxi (x)| dx ≤ C B and when we combine these, we get kukW 1,p (B) ≤ CkukW 1,p (B+ ) which proves the claim. (3) Consider the situation where ∂U is not flat near x0 . In this case we can find a C 1 mappnig Φ with inverse Ψ such that Φ straightens out the boundary ∂U near x0 . 64 yi = xi , i = 1, . . . , n − 1 y = Φ(x) yn = xn − γ(x1 , . . . , xn−1 ) xi = yi , i = 1, . . . , n − 1 x = Ψ(y). xn = yn + γ(y1 , . . . , yn−1 ) In the ”new” domain we inscribe a ball around y0 inside the transformed ball and call each part B + and B − . In addition we write u′ (y) = u (Ψ(y)) , x = Ψ(y), y = Ψ(x). With the same reasoning as we used with the assumption of a flat boundary, we can extend u′ to u′ defined on B such that u′ ∈ C 1 (B). Moreover, we can also use the inequality, ku′ kW 1,p (B) ≤ Cku′ kW 1,p (B+ ) . By definition, u(x) = u′ (Φ(x)) =⇒ uxi = u′yi (Φ(x)) + u′yn (Φ(x))γxi for i = 1, . . . , n − 1, and for i = n uxn (x) = u′xn (Φ(x)). Let W = Ψ(B) ⇔ B = Ψ(W ). Then, by change of variable and our previous definitions; kukW 1,p (W ) ≤ ku′ kW 1,p (B) and by (*) ≤ Cku′ kW 1,p (W + ) ≤ CkukW 1,p (W + ) ≤ CkukW 1,p (U ) 65 (*) so, kukW 1,p (W ) ≤ CkukW 1,p (U ) (**) (4) Since ∂U is compact ∃{xi0 }N i=1 ⊂ ∂U (a set of points on the boundary), N and open sets {Wi }i=1 and extensions ui of u to Wi (i = 1, . . . , N) such that N ∂U ⊂ ∪N i=1 Wi . We choose W0 ⊂⊂ U such that U ⊂ ∪i=0 Wi . P N N We let {ξ}i=0 be a partition of unity wrt {Wi }i=0 . We write u = N i=0 ξi ui , N where u0 = u. Note that supp(U ⊂ ∪i=0 Wi . Then kukW 1,p (Rn ) ≤ C N X i=0 (∗∗) kui kW 1,p (Wi ) ≤ C In short, we havae established that N X i=0 kukW 1,p (U ) = CNkukW 1,p (U ) . n kuk1,p W (R ) ≤ CkukW 1,p (U ) . (I) (5) Write Eu = u, with u 7→ Eu being linear (follows from the definition). The construction so far is for u ∈ C 1 (U ). Now assume u ∈ W 1,p (U). Then ∃um ∈ C ∞ (U ) such that kum − ukW 1,p (U ) → 0 as n → ∞. Thus, (I) kEum − Eun kW 1,p (Rn ) ≤ Ckum − ul kW 1,p (U ) → 0 as m, l → ∞. 1,p which means {Eum }∞ (Rn ), hence it converges m=1 is a Cauchy sequence in W (since Sobolev spaces are complete) to some u = Eu in W 1,p (Rn ). So, we know, since um ∈ C ∞ (smooth) kEum kW 1,p (Rn ) ≤ Ckum kW 1,p (W ) ↓ for all u ∈ W 1,p (U). ↓ as m → ∞ kEukW 1,p (Rn ) ≤ CkukW 1,p (U ) 5.5 Traces Theorem - Trace Theorem Assume U is bounded and ∂U ∈ C 1 . Then there exists a bounded, linear operator T : W 1,p (U) 7→ Lp (∂U) such that (i) T u = u if u ∈ C(U ) ∩ W 1,p (U). ∂W (ii) kT ukLp (∂U ) ≤ CkukW 1,p (U ) for all u ∈ W 1,p (U). Definition: T u is the trace of u on ∂U. 66 Proof (1) Assume u ∈ C 1 (U), and assume x0 ∈ ∂U and ∂U is flat near x0 . Let b = B(x0 , r ) (the smaller ball contained in B). We also B = B(x0 , r), B 2 b We define Γ = B b ∩ ∂U, and introduce ξ ∈ Cc∞ (B), ξ ≥ 0 and ξ = 1 on B. ′ n−1 x = (x1 , . . . , xn−1 ) ∈ R = {xn = 0}. We take the integral of u over Γ, and since ξ = 1 on Γ we can write: Z Z Z p ′ p ′ |u| dx = ξ|u| dx ≤ ξ|u|pdx′ Γ Γ {xn =0} (since {xn = 0} is a bigger domain than Γ). Using Green’s formula, we get Z Z Z p p−1 p pξ|u| sgn(u)uxn dx ξxn |u| dx + =− (ξ|u| )xn dx = − B+ B+ B+ ≤C Z B+ p |u| + p|u| p−1 |uxn |dx and we take a closer look at p|u|p−1|uxn |. Using Young’s inequality: ab ≤ q ap + bq for 1/p + 1/q = 1, we get p |u|p−1|uxn | ≤ |uxn |p |u|(p−1)q |uxn |p p − 1 p + = + |u| p q p p since 1/p + 1/q = 1 ⇒ 1/q = (p − 1)/p ⇒ q = p/(p − 1), so we get Z Z Z p p ′ p |u|p +|Du|pdx ≤ CkukpW 1,p (B+ ) |u| dx ≤ C |u| + |uxn | dx ≤ C Γ B+ B+ where C only depends on ξ and p. 67 —————————————————————- 05/10-2010 Theorem from last time Assume U is bounded with ∂U ∈ C 1 . Then there exists a bounded, linear operator T : W 1,p (U) 7→ K p (∂U) such that (i) T u = u∂U if u ∈ C(U) ∩ W 1,p (U) (ii) kT ukLp (∂U ) ≤ CkukW 1,p (U ) for all u ∈ W 1,p (U) for some constant C = C(n, p, U). Proof Assume first u ∈ C 1 (U), and x0 ∈ ∂U and assume ∂U is flat near x0 . As we had last time: Z Γ p ′ |u| dx ≤ C Z B+ |u|p + |Du|p dx (I) The second step is to assume ∂U is not flat near x0 , in which case we can simply straighten out ∂U near x0 with functions Φ and Ψ: Applying (I) and change of variables, we get Z Z p |u| dS ≤ |u|p + |Du|pdx Γ U for Γ ⊂ ∂U containing x0 . 68 (II) Since the boundary is compact we can find points {x0 }N i=1 ⊂ ∂U, Γi ⊂ ∂U where i = 1, . . . , N such that ∂U = ∪N Γ and by (II), which applies to each i=1 i i, Z Z p |u| dS ≤ |u|p + |Du|p dx =⇒ Γi Then U kukLp (Γi ) ≤ CkukW 1,p (U ) . kukLp (∂U ) ≤ N X i=1 i = (1, . . . , N) kukLp (Γi ) ≤ C · NkukW 1,p (U ) which is the desired estimate. We write T u = u∂U . kT ukLp (∂U ) ≤ CkukW 1,p (U ) , u ∈ C 1 (U). (III) Next, assume u ∈ W 1,p (U) and we show that the inequality is true for this as well. By the approximation theorem ∃um ∈ C ∞ (U) such that um → u in W 1,p (U) (IV). Then, since the functions are smooth: (III) kT um − T ul kLp (∂U ) ≤ Ckim − ul kW 1,p (U ) −→ 0 by (IV) as m, l → ∞. Thus, {T um } is Cauchy in Lp (∂U), so T um → u∗ := T u in Lp (∂U). We combine this with (IV). (III) kT um kLp (∂U ) ≤ Ckum kW 1,p (U ) ↓ (III) ↓ as m → ∞ kT ukLp (∂U ) ≤ CkukW 1,p (U ) Assume in addition that u ∈ C)U, then um → u uniformly on U , which again means T um → T u uniformly on U . T um = um ∂U =⇒ T u = u∂U Recall: W01,p (U) is the closure of Cc∞ (U) in W 1,p (U), that is u ∈ W01,p (U) ⇐⇒ ∃um ∈ Cc∞ (U) ⇒ um → u in W 1,p (U). Theorem Assume U is bounded and ∂U ∈ C 1 . Suppose u ∈ W 1,p (U). Then u ∈ W01,p (U) ⇐⇒ T u = 0 on ∂U. (where we call T u the trace of u). Proof : Evans page 261. 69 5.6 - Sobolev Inequalities Assume 1 ≤ p < n. Motivation: we want to establish an estimate of the form kukLq (Rn ) ≤ CkDukLp (Rn ) (I) for all u ∈ Cc∞ (Rn ), and some constant C > 0 which is independent of the function u and 1 ≤ q < ∞. Question: Assuming that the estimate (I) holds, what is the relation beween p,q and the dimension n? We choose u ∈ Cc∞ (U) where u 6= 0. We define uλ (x) = u(λx) for λ > 0, x ∈ Rn . Then by (I), kuλ kLq (Rn ) ≤ CkDuλ kLp (Rn ) But, Z Z 1 |uλ(x)| dx = |u(λx)| dx = n λ Rn Rn q q Z Rn (II) |u(y)|q dy. where we made the substitution y = λx, dy = λn dx, which yields dx = (1/λn )dy. Also, Z Z Z Z λp p p p p |Dux (x)| dx = |Du(λx)| dx = n |Du(y)|pdy. |D(u(λx))| dx = λ λ n n n n R R R R where we used the same substitution. Now, by (II) n λ− q kukLq (Rn ) ≤ Cλ p−n p kDukLp (Rn ) which is equivalent to, n n kukLq (Rn ) ≤ Cλ1− p + q kDukLp (Rn ) We note that if 1 − np + nq > 0, then by letting λ → 0 we get 0 on the right hand side, which means we get a contraction. If 1 − np + nq < 0 we get a contraction by letting λ → ∞. Both these scenarios must be untrue since we have assumed u 6= 0. So, for (I) to hold we must have 1− n n 1 1 1 np + = 0 ⇐⇒ = − ⇐⇒ q = . p q q p n n−p Definition: A q satisfying this is called the Sobolev conjugate of p and is denoted p∗ . 70 Theorem. The Gagliardo-Nirenberg-Sobolev Inequality np (the Sobolev conjugate), then there exists Assume 1 ≤ p < n and p∗ = n−p a constant C > 0 depending only on p and n such that kukLp∗ (Rn ) ≤ CkDukLp (Rn ) for all u ∈ Cc∞ (Rn ). Proof Case p = 1: Since supp(U) is compact, Z xi u(x) = uxi (x1 , . . . , yi , . . . , xn )dyi (i = 1, . . . , n) −∞ (u vanishes before infinity because it is compact), so Z ∞ |u(x)| ≤ uxi (x1 , . . . , yi, . . . , xn )dyi (i = 1, . . . , n) −∞ which implies, n |u(x)| ≤ n Z Y ∞ −∞ i=1 |uxi (x1 , . . . , yi, . . . , xn )| and taking the 1/(n − 1)’th power on both sides: |u(x)| n n−1 ≤ n Z Y i=1 ∞ −∞ |uxi (x1 , . . . , yi , . . . , xn )| 1 n−1 We use the shorthand notation: uxi (x) = uxi (x1 , . . . , yi, . . . , xn ), so we write n n−1 |u(x)| ≤ n Z Y i=1 ∞ −∞ |uxi (x)| 1 n−1 = Z ∞ −∞ |ux1 (x)| dyi 1 n−1 n Z Y i=2 ∞ −∞ |uxi (x)| 1 n−1 We take the integral on both sides wrt x1 . Z ∞ −∞ |u(x)| n n−1 dx1 ≤ Z ∞ −∞ |ux1 (x)| dyi 1 Z n−1 ∞ n Z Y −∞ i=2 ∞ −∞ |uxi (x)| 1 n−1 dx1 (*) We want to estimate the integral on the far right and do so using the general Hölder inequality. Z f1 . . . fn dx ≤ Z f1p1 71 p1 1 ... Z f1pn p1 n if Pn 1 i=1 pi = 1, or the shorthand version: Z Y n i=1 We set fi = Z n Z Y fi dx ≤ i=1 ∞ −∞ fipi |uxi (x)|dyi 1 n−1 p1 i . pi = n − 1 , Pn 1 1 since i=2 pi = n−1 (n − 1) = 1, which means we can use the Hölder inequality. We get 1 1 n−1 n−1 Z ∞Y n Z ∞ Z ∞ n Z ∞ Y dx1 ≤ |uxi (x)|dyidx1 |uxi (x)|dyi −∞ i=2 −∞ −∞ i=2 −∞ which implies, when we use the inequality on (*), Z ∞ −∞ |u(x)| n n−1 dx1 ≤ Z ∞ −∞ |ux1 (x)| dyi 1 n−1 n Z Y ∞ −∞ i=2 We now integrate both sides wrt to x2 , define k := ZZ ∞ k −∞ |u(x)| dx1 dx2 ≤ Z Z ∞ −∞ |uxi (x)dx1 dx2 k Z Z ∞ −∞ | This implies, I= where f1n−1 I≤ Z Z = Z ∞ n Y −∞ i=1, i6=2 "Z ∞ −∞ |ux1 (x)|dyi ∞ −∞ |ux1 (x)|dyi dx2 This implies, ZZ ∞ −∞ |u(x)|k dx1 dx2 ≤ Z Z ∞ −∞ |ux1 (x)dyi dx2 k −∞ Z n Z Z Z Y = fin−1 72 i=3 | −∞ |uxi (x)|dyidx1 {z } k dx2 fi , 3≤i≤n 1 n−1 −∞ |ux1 (x)|dyi. ∞ −∞ ∞ ∞ |uxi (x)|dx1 dx2 dyi ∞ −∞ k Y n ZZ ∞ #n−1 1 n−1 |uxi (x)|dyi dx1 1 n−1 1 n−1 f1 −∞ k Z Z −∞ |ux1 (x)| dyi {z } i=1, i6=2 i=3 ∞ ∞ Y Z fi dx2 ≤ Z |ux1 (x)|dx1 dx2 k k ! n ZZZ Y i=3 . ∞ −∞ |uxi (x)|dx1 dx2 dyi !k Continuing in this way wrt x3 , . . . , xn we will eventually arrive at the inequality: Z ... Z ∞ −∞ |u(x)| n n−1 n Z Y dx1 . . . dxn ≤ ... Z ∞ −∞ i=1 |uxi (x)|dx1 . . . dyi . . . dxn 1 n−1 But now we simply use |uxi (x)| ≤ |Du(x)|, which is true for i = 1, . . . , n. Using that the integral from −∞ to ∞ n times is the integral over Rn we get Z Rn |u(x)dx| = n n−1 n Z Y Z and if we take the n−1 n ∞ |Du(x)|dx1 . . . dyi . . . dxn −∞ i=1 Rn |u(x)dx| Rn ... Z 1 n−1 Z |Du(x)|dx = Rn i=1 So we have: ≤ n Z Y n n−1 ≤ Z Rn 1 n−1 n n−1 |Du(x)|dx n n−1 |Du(x)|dx power on each side, we get Z Rn |u(x)dx| n n−1 n−1 n ≤ Z Rn |Du(x)|d and this is equivalent to n ≤ CkDukL1(Rn ) . kukL n−1 (Rn ) np We have verified this for p = 1, and note that p∗ = n−p so when p = 1 we n ∗ have p = n−1 which is what we have on the left side. We now generalise this. For p = 1 we just established that Z Rn |u(x)| n−1 n n n−1 ≤C Z Rn |Du(x)|dx. (**) The case 1 < p < n. We apply the inequality (**) to v = |u|γ , where γ > 1 is to be chosen later. Z Rn |u(x)| γn n−1 n−1 n ≤C ≤C Z Rn Z |u| Rn γ |D(|u| )|dx = C (γ−1)p p−1 p−1 Z p Rn 73 Z Rn |Du|p γ|u|γ−1 |Du|dx p1 We have p−1 p + Z 1 p Rn = 1, so Hölder’s inequality is okay. The expression implies, γn n−1 |u(x)| n−1 n ≤C Z Rn |u| (γ−1)p p−1 p−1 p kDukLp (Rn ) where we have what we need on the right hand side. Now we choose γ such that (γ − 1)p n p −p γn = ⇐⇒ γ( − )= n−1 p−1 n−1 p−1 p−1 −p np − n − np + p p−n n−1 ⇐⇒ γ = ⇐⇒ γ = −p ⇐⇒ γ = p >1 (n − 1)(p − 1) p−1 n−1 n−p So, So, setting γn n−1 γn p (n−1) n np = · = = p∗ . n−1 n−p (n − 1) n−p = p∗ in the previous expression, means Z Rn But, n−1 − p−1 n p |u| dx ≤ CkDukLp (Rn ) p∗ np − p − np + n n−p 1 n−1 p−1 − = = = ∗ n p np np p which implies, and finally concludes the proof, kukLp∗ (Rn ) ≤ CkDukLp (Rn ) 74 ∀u ∈ Cc 1 (Rn ) —————————————————————- 12/10-2010 Theorem: Morrey’s Inequality For n < p ≤ ∞ we have kukC 0,γ (Rn ) ≤ CkukW 1,p (U ) ∀u ∈ Cc1 (R) where γ = 1 − np . Proof Z — B(x,r) |u(y) − u(x)|dy ≤ C Z B(x,r) |Du(y)| |x − y|n−1 kukL∞ (Rn ) ≤ CkukW 1,p (Rn ) Now let W = B(x, r) ∩ B(y, r) for any x, y ∈ Rn and |x − y| = r. 75 (1) (2)