HARMONIC ANALYSIS HUIQIANG JIANG 1. Test functions and distributions Let Ω be a nonempty open set in Rn . The test function space D (Ω) = C0∞ (Ω) consists of all infinitely differentiable functions with compact support. For every cpt K ⊂ Ω, we use DK (Ω) ⊂ D (Ω) to define the collection of all infinitely differentiable functions with compact support in K. We can introduce a metric in DK (Ω) such that for any f, g ∈ DK (Ω), dK (f, g) = max 0≤N <∞ where 1 kf − gkN 2N 1 + kf − gkN kf kN = max kDα f kL∞ , N ≥ 0. |α|≤N We use τK to denote the topology of DK (Ω) equipped with such metric. The topology of D (Ω) can be defined precisely. Let β be the collection of all convex balanced sets W ⊂ D (Ω) such that DK (Ω) ∩ W ∈ τK for every compact K ⊂ Ω. Let τ be the collection of all unions of sets of the form φ+W with φ ∈ D (Ω) and W ∈ β. Theorem 1. τ is a topology in D (Ω) and β is a local base for τ . The topology τ makes D (Ω) into a locally convex topological vector space. Many important properties of D (Ω) are included in the following theorem. Theorem 2. (a) A convex balanced subset V of D (Ω) is open iff V ∈ β. (b) The topology τK of any DK (Ω) coincides with the subspace topology that DK inherits from D (Ω). (c) If E is bounded subset of D (Ω), then E ⊂ DK (Ω) for some compact K ⊂ Ω and for each N ≥ 0, there exists MN < ∞ s.t., kφkN ≤ MN for any φ ∈ E. (d) D (Ω) has the Heine-Borel property. (e) If {φi } is a Cauchy sequence in D (Ω), then {φi } ⊂ DK (Ω) for some compact K ⊂ Ω and lim kφi − φj kN = 0 for any N ≥ 0. i,j→∞ (f ) If φi → 0, then {φi } ⊂ DK (Ω) for some compact K ⊂ Ω and Dα φi → 0 uniformly for every multi-index α. (g) In D (Ω), every Cauchy sequence converges. Theorem 3. Every differential operator Dα : D (Ω) → D (Ω) is continuous. Definition 1. A continuous linear functional on D (Ω) is called a distribution. The space of all distributions in Ω is denoted by D0 (Ω). 1 2 HUIQIANG JIANG Theorem 4. If Λ is a linear functional on D (Ω), the following two conditions are equivalent: (a) Λ ∈ D0 (Ω). (b) For every compact K ⊂ Ω, there exists N ≥ 0 and C > 0 s.t. |Λφ| ≤ C kφkN holds for every φ ∈ DK (Ω). Remark 1. If Λ is such that one N will do for all K, then the smallest such N is called the order of Λ. Example 1. δx is a distribution of order 0. Example 2. Every locally integrable complex function is a distribution of order 0. Example 3. Any locally finite Borel measure is a distribution of order 0. The space of distribution D (Ω) is a topological vector space with weak-* topology, and it has many other nice structures. Definition 2. If α is a multi-index and Λ ∈ D0 (Ω), Dα Λ ∈ D0 (Ω) is a linear functional defined by the formula |α| (Dα Λ) (φ) = (−1) Λ (Dα φ) , φ ∈ D (Ω) . Proposition 1. For any multi-indices α, β and for any Λ ∈ D0 (Ω), Dβ Dα Λ = Dα Dβ Λ = Dα+β Λ. Remark 2. If f has continuous partial derivatives of all orders up to N , then Dα Λf = ΛDα f for any |α| ≤ N. Definition 3. Suppose Λ ∈ D0 (Ω) and f ∈ C ∞ (Ω). We can define a distribution f Λ such that (f Λ) (φ) = Λ (f φ) , φ ∈ D (Ω) . Locally, distributions are derivatives of distributions induced by continuous functions: Theorem 5. Suppose Λ ∈ D0 (Ω) and K is a compact subset of Ω. Then there is a continuous function f in Ω and there is a multi-index α such that for every φ ∈ DK (Ω), Z |α| Λφ = (−1) f (x) Dα φ (x) dx. Ω i.e., Λ = Dα f on K. These results on distributions are proved in Analysis 3, please check Chapter 6 of Rudin’s Functional Analysis for details. HARMONIC ANALYSIS 3 2. Fourier Series We consider complex valued periodic functions on R with period 2π which can 1 be viewed as functions defined on the the unit circle ¡ S¢ . The Fourier Coefficients of the function f ∈ L1 S1 ≡ L1 [−π, π] are defined by Z π 1 an = f (x) e−inx dx. 2π −π Formally, the Fourier Series of f is ∞ X f (x) ∼ an einx . n=−∞ Since f ∈ L1 [−π, π], we have 1 |an | ≤ 2π Z π |f (x)| dx = −π 1 kf kL1 . 2π Moreover, we have the Riemann-Lebesgue lemma: Theorem 6. For every f ∈ L1 [−π, π], lim an = 0. n→±∞ ¡ ¢ If f ∈ C k S1 , i.e., f ∈ C k (R) is periodic with period 2π, then integration by parts yields for any n 6= 0, ° k ° ° k ° 1 ° d f° 1 ° d f° ° ° ° . |an | ≤ ≤ k° ° ° ° k k 2πn dx L1 n dxk °L∞ ¡ ¢ In particular, if f ∈ C 2 S1 , Weierstrass’s M-test implies that the Fourier Series of f ∞ X an einx n=−∞ ¡ ¢ converges uniformly to a continuous function g ∈ C S1 . We will show that g = f later. To study the convergence of Fourier series, we define the partial sums sN (f, x) = N X an einx , n=−N and the Fejer sums N 1 X sn (f, x) . N + 1 n=0 SN (f, x) = Direct calculations yield Z π sN (f, x) = f (y) kN (x − y) dy −π where ¡ ¢ 1 sin N + 12 z kN (z) = 2π sin z2 and Z π SN (f, x) = f (y) KN (x − y) dy −π 4 HUIQIANG JIANG where " ¡ ¢ #2 sin N + 12 z 1 KN (z) = . π (N + 1) sin z2 If we take f ≡ 1, then for each N ≥ 0, sN (f, x) ≡ SN (f, x) ≡ 1 and hence Z π Z π kN (x) dx = KN (x) dx = 1. −π −π We observe also that KN (x) ≥ 0 and for each 0 < δ < π, KN converges to 0 uniformly on δ ≤ |x| ≤ π as N → ∞ which implies Z KN (x) dx = 0. lim N →∞ δ≤|x|≤π ¡ ¢ Theorem 7. For any f ∈ C S1 , the Fejer sum SN (f, x) converge to f uniformly on S1 . We leave the proof as an exercise. Theorem 8. For any f ∈ Lp [−π, π], p ≥ 1, the Fejer sum SN (f, x) converges to f in Lp [−π, π]. Proof. Young’s inequality implies 1 kSN (f, x)kLp ≤ 2 p kf kLp . ¡ ¢ Since C S1 is dense in Lp [−π, π], the result follows from 7. ¤ ¡ ¢ Corollary 1. For any f ∈ C 2 S1 , the Fourier series of f converges to f uniformly. Proof. Since 1 kf 00 kL∞ , n2 sN (f, x) converges uniformly, since SN (f, x) converges uniformly to f , we conclude sN (f, x) converges uniformly to f . ¤ ¡ ¢ Theorem 9. If f ∈ C α S1 for some α > 0, then for any x, |an | ≤ lim sN (f, x) = f (x) . N →∞ Proof. 1 sN (f, x) − f (x) = 2π Z π −π µ ¶ 1 (f (x − y) − f (x)) sin N + ydy. sin y2 2 Since (f (x − y) − f (x)) sin y2 is integrable, the result follows from Riemann-Lebesgue lemma. ¤ ¡ ¢ Remark 3. Actually, the result can be strengthened to: If f ∈ C α S1 for some α > 0, then sN (f, x) converges to f uniformly. ¡ ¢ Remark 4. There exists f ∈ C S1 such that sN (f, x) doesn’t converge to f pointwisely. HARMONIC ANALYSIS 5 ¡ ¢ Theorem 10. If f ∈ BV S1 , then for any x, lim sN (f, x) = N →∞ Proof. We first consider 1 λ→∞ π Z f (x+) + f (x−) . 2 π lim f (−y) −π Let Z ∞ G (y) = y sin λy dy. y sin y dy. y Then we have Z 1 π sin λy f (−y) dy π 0 y Z 1 λπ ³ y ´ sin y = f − dy π 0 λ y Z 1 λπ ³ y ´ =− f − dG (y) π 0 λ ³ y ´¯λπ 1 Z λπ ³ y´ ¯ = −G (y) f − ¯ + G (y) df − λ 0 π 0 λ Z 1 π = −G (λπ) f (−π+) + G (0) f (0−) + G (λy) df (−y) , π 0 hence 1 lim λ→∞ π Similarly 1 λ→∞ π Z π f (−y) sin λy 1 dy = f (0−) . y 2 f (−y) sin λy 1 dy = f (0+) . y 2 0 Z 0 lim −π So we have Z sin λy f (0+) + f (0−) 1 π f (−y) dy = . lim λ→∞ π 0 y 2 Finally, we have from Riemann-Lebesgue lemma µ ¶ Z 1 π sin λy sin λy lim f (−y) − dy = 0. λ→∞ π 0 y 2 sin y2 ¤ ¡ ¢ ¡ ¢ Remark 5. If f ∈ BV S1 ∩ C S1 , then sN (f, x) converges to f uniformly. Now we consider f ∈ Lp [−π, π], p ≥ 1. When p = 2, we have the well-known result Theorem 11. Let f ∈ L2 [−π, π], then sN (f ) → f in L2 [−π, π] as N → ∞. Moreover, ∞ X 2 2 |an | . kf kL2 = n=−∞ 6 HUIQIANG JIANG On the other hand, we can’t expect sN (f ) → f in L1 [−π, π] as N → ∞ if f ∈ L1 [−π, π]. We define TN : f 7→ sN (f ) . If for any f ∈ L1 [−π, π], TN f → f in L1 [−π, π] as N → ∞, then uniform boundedness principle implies the existence of C > 0 such that kTN f kL1 ≤ C kf kL1 holds for any f ∈ L1 [−π, π] and for any N . By duality, we have kTN f kL∞ ≤ C kf kL∞ holds for any f ∈ L∞ [−π, π] and for any N which is impossible since Z π lim |kN (x)| dx = ∞. N →∞ −π p Theorem 12. Let f ∈ L [−π, π], 1 < p < ∞, then sN (f ) → f in Lp [−π, π] as N → ∞. We will prove this result later. 3. Fourier Transform ¡ d ¢ If f ∈ L R ; C , its Fourier transform F (f ) = fˆ is defined by Z 1 fˆ (y) = f (x) e−ix·y dx. d/2 Rd (2π) 1 Since (3.1) ° ° ° ˆ° °f ° 1 kf kL1 , d/2 (2π) ¡ ¢ ¡ ¢ the Fourier transform F : L1 Rd → L∞ Rd . Now let’s list some basic properties of Fourier transform: ¡ ¢ Theorem 13. (i) Translation: Let f ∈ L1 Rd ; C . If g (x) = f (x − a) and h (x) = eiax f (x) for some a ∈ Rd , then L∞ ≤ ĝ (y) = e−iay fˆ (y) and ĥ (y) = fˆ (y − a) . ¡ ¢ (ii) Scaling: Let f ∈ L1 Rd ; C and λ > 0. If h (x) = f (x/λ), then ĥ (y) = λd fˆ (λt) . ¡ ¢ ¡ ¢ (iii) Convolution: Let f, g ∈ L1 Rd ; C . Then f ∗ g ∈ L1 Rd ; C and ˆ d/2 (f ∗ g) = (2π) fˆ · ĝ. HARMONIC ANALYSIS Proof. (iii) ˆ (f ∗ g) (y) = = = Z 1 d/2 (2π) 1 (f ∗ g) (x) e−ix·y dx Rd Z d/2 (2π) 1 Rd Z d/2 (2π) = (2π) 7 d/2 µZ ¶ f (x − z) g (z) dz e−ix·y dx Rd µZ ¶ f (x − z) e Rd −i(x−z)·y dx e−iz·y g (z) dz Rd fˆ (y) ĝ (y) . ¤ ¡ ¢ A function f ∈ C ∞ Rd ; C is said to be rapidly decreasing if for any n ≥ 0, ³ ´n 2 (3.2) pn (f ) = sup sup 1 + |x| |(Dα f ) (x)| < ∞. |α|≤n x∈Rd i.e., P · Dα f is uniformly bounded for every polynomial P and for every multiindex α. The Schwartz space S = Sd is defined as the function space of rapidly decreasing functions on Rd . The countable collection of norms pn in (3.2) defines a locally convex topology for S. The topology of S is compatible with the metric d (f, g) = max n≥0 1 pn (f − g) . 2n 1 + pn (f − g) It is easy to check f ∈ S implies f ∈ L1 . Moreover, if f ∈ S, then for every multiindex α, Dα f , xα f ∈ S and we have the following property of Fourier transform. Proposition 2. If f ∈ S and α is a multi-index, then (Dα f ) ˆ (y) = (iy) fˆ (y) , ³ ´ α Dα fˆ (y) = ((−ix) f ) ˆ (y) . α Proof. Direct calculation. ¤ Theorem 14. (a) S is a Fréchet space. (b) If P is a polynomial, g ∈ S and α is a multi-index, then each of the three mappings f → P f, f → gf, f → Dα f is a continuous linear mapping of S into S. ¡ ¢ ¡ ¢ (c) D Rd is dense in S and the identity mapping of D Rd into S is continuous. Proof. (a) Let {fk } ⊂ S be a Cauchy sequence. Then for each n, {fk } is a Cauchy sequence w.r.t the norm pn . Hence, for any multi-indices α, β, there exists continuous function gαβ such that xα Dβ fk → gαβ uniformly as k → ∞. Moreover, gαβ (x) = xα Dβ g00 , so we have fi → g00 in S. Hence S is complete and it is a Fréchet space. (b) Since S is a Fréchet space, a mapping Λ : S → S is continuous iff fk → 0 implies Λfk → 0. We also recall that fk → 0 iff for each n, pn (fk ) → 0. The continuity of the three mappings follows easily. 8 HUIQIANG JIANG ¡ ¢ (c) Let ϕ ∈ D Rd be such that ϕ (x) ≡ 1 for any |x| < 1. For any f ∈ S, we define ³x´ fk (x) = f (x) ϕ . k Then for any multi-indices α, β, ¯ ¯ max ¯xα Dβ (f − fk )¯ x ¯ ¯ ¯X µ ¶ ¯ h ³ ´i ¯ ¡ ¢ x ¯¯ β α x Dβ−γ f (x) Dγ 1 − ϕ = max ¯¯ γ k ¯¯ |x|≥k ¯ γ≤β µ ¶ ¯³ ¯ ´ ¡ ¢ 1 X β ¯ ¯ 2 ≤ max kDγ ϕkL∞ ¯ 1 + |x| xα Dβ−γ f (x)¯ → 0. γ |x|≥k 1 + k 2 γ≤β ¡ ¢ Hence, fk → f in S. So D Rd ¡ is dense in S. ¢ d The identity mapping of ¡D R ¢ into S is continuous iff for each K compact, d the ¡identity mapping of D R into S is continuous. ¡Now K ¢ for K compact, since ¢ d DK Rd is a Fréchet space, the identity mapping of D R into S is continuous iff K ¡ ¢ fk → 0 in DK Rd implies fk → 0 in S. The latter statement is easy to verify. ¤ Theorem 15. The Fourier transform is a continuous linear mapping of S into S. Proof. For any f ∈ S and for any multi-index α, β, ³ ´ ³ ´ β y α Dβ fˆ (y) = y α (−ix) f ˆ (y) α+β ¡ α ¡ β ¢¢ = (−i) D x f ˆ (y) which is uniformly bounded, hence fˆ ∈ S. Let fk → 0 in S, then for any multi-index α, β, ° ³ ´ ° ° α ¡ β ¢° °¡ ° ¡ ¢¢ 1 ° α ° ° D x fk ° 1 °y Dβ fˆk (y)° ∞ = ° Dα xβ fk ˆ (y)°L∞ ≤ L d/2 L (2π) ° ° ° °³ ° ° ° ´d ° ¡ β ¢° ° 1 1 ° ° 2 α ° 1 + |x| ≤ D x fk ° ° → 0. ³ ´ ° ∞° d d/2 ° ° ° 2 (2π) L ° 1 + |x| ° L1 Hence fˆk → 0 in S. So the Fourier transform is continuous. ¤ ¡ d¢ We use C0 R to denote continuous functions converging to zero at infinity. ¡ ¢ ¡ ¢ Corollary 2. If f ∈ L1 Rd , then fˆ ∈ C0 Rd . ¡ ¢ ¡ ¢ Proof. For each f ∈ L1 Rd , there exists fn ∈ S such that fn → f in L1 Rd . Now ¡ ¢ ¡ ¢ (3.1) implies that fˆn → fˆ in L∞ Rd . Since fˆn ∈ S, we conclude fˆ ∈ C0 Rd . ¤ Before introducing the inverse transform, we need the following lemma on approximate delta function. ¡ ¢ Proposition 3. Let {kn } be a sequence of functions in L1 Rd satisfying (1) kn (x) ≥ 0 a.e. x ∈ Rd ; (2) Z Rd kn (x) dx = 1; HARMONIC ANALYSIS (3) For any δ > 0, 9 Z lim kn (x) dx = 0. n→∞ |x|>δ ¡ ¢ ¡ ¢ Then for any f ∈ C Rd ∩ L∞ Rd , Z lim f (y) kn (x − y) dx = f (x) . And for any f ∈ L p n→∞ ¡ d Rd ¢ R , 1 ≤ p < ∞, Z ¡ ¢ f (y) kn (x − y) dx converges to f in Lp Rd . fn (y) = Rd We also recall that Z 2 e−(x+it) dx = √ π R holds for any t ∈ R. Theorem 16. The Fourier transform F : S → S is invertible. And for any f ∈ S, Z 1 f (x) = fˆ (y) eix·y dy. d/2 Rd (2π) Proof. We compute Z 1 d/2 (2π) = lim ε→0 = lim ε→0 = lim fˆ (y) eix·y dy Rd Z 1 d/2 (2π) Z 1 d (2π) 1 ε→0 (2πε) = f (x) . Here we used the fact that 2 ε fˆ (y) eix·y e− 2 |y| dy Rd Z ε Rd Z d/2 2 f (z) e−iz·y eix·y e− 2 |y| dydz Rd f (z) e− |x−z|2 2ε dz Rd 2 |z| 1 e− 2ε (2πε)d/2 is a family of approximate delta functions. ¤ Hence, we can define inverse Fourier transform F −1 : S → S s.t. for any g ∈ S, Z 1 F −1 (g) (x) = ǧ (x) = g (y) eix·y dy. d/2 d R (2π) Apparently, F −1 : S → S is also continuous. Corollary 3. If f, g ∈ S, then f ∗ g ∈ S. Proof. f, g ∈ S imply fˆ, ĝ ∈ S, hence fˆĝ ∈ S and d/2 (f ∗ g) ˆ = (2π) fˆ · ĝ ∈ S. Hence, f ∗ g = F −1 ((f ∗ g) ˆ) ∈ S. ¤ 10 HUIQIANG JIANG ¡ ¢ Proposition 4. If f ∈ L1 Rd , then as ε → 0, fε defined by Z 2 ε 1 fε (x) = fˆ (y) eix·y e− 2 |y| dy d/2 Rd (2π) ¡ ¢ converges to f in L1 . If in addition fˆ ∈ L1 Rd , then Z 1 f (x) = fˆ (y) eix·y dy a.e. x ∈ Rd . d/2 Rd (2π) Proof. We compute Z 1 d/2 (2π) Since Z 1 2 ε fˆ (y) eix·y e− 2 |y| dy = d/2 (2πε) Rd f (z) e− |x−z|2 2ε dz Rd 2 |z| 1 e− 2ε (2πε)d/2 is a family of approximate delta functions, Z |x−z|2 1 − 2ε f (z) e dz → f (x) in L1 as ε → 0. d/2 Rd (2πε) ¡ ¢ If in addition fˆ ∈ L1 Rd , Lebesgue dominated convergence implies, Z 1 f (x) = fˆ (y) eix·y dy a.e. x ∈ Rd . d/2 Rd (2π) ¤ Remark 6. If f ∈ L ¡ 1 ¢ ¡ ¢ ¡ ¢ Rd and fˆ ∈ L1 Rd , then actually f, fˆ ∈ C0 Rd . Now we prove the well known Parseval formula Theorem 17. If f, g ∈ S, then Z Z fˆĝdy. f ḡdx = Rd Rd ¡ ¢ Using the inner product of L2 Rd , we can write ³ ´ (f, g) = fˆ, ĝ . In particular, when f = g, we have the Plancherel identity ° ° ° ˆ° °f ° 2 = kf kL2 . L Proof. Using Fubini theorem, we have Z Z Z f ḡdx = fˆ (y) eix·y ḡ (x) dxdy d d d R ZR R = fˆ (y) ĝ (y) dy. Rd ¤ Corollary 4. If f ∈ L ¡ 1 ¢ ¡ ¢ ¡ ¢ Rd ∩ L2 Rd , then fˆ ∈ L2 Rd , and ° ° ° ˆ° °f ° 2 = kf kL2 . L HARMONIC ANALYSIS 11 ¡ ¢ ¡ ¢ Proof. Choose, fn ∈ S, such that fn → f in L1 Rd ∩ L2 Rd . Since ° ° °ˆ ° °fn − fˆm ° 2 = kfn − fm kL2 , L n o ¡ ¢ fˆn is a Cauchy sequence in L2 Rd , hence there exists g ∈ L2 , s.t. fˆn → g ¡ ¢ ¡ ¢ ¡ ¢ in L2 Rd . On the other hand, fn → f in L1 Rd implies, fˆn → fˆ in L∞ Rd . ¡ ¢ Hence fˆ = g ∈ L2 Rd . ¤ ¡ ¢ Hence we can extend Fourier¡ transform as a linear isometry from L2 Rd onto ¢ ¡ ¢ ¡ ¢ itself. ¡Actually, for any f ∈ L2 Rd , if {fn } ⊂ L1 Rd ∩ L2 Rd satisfies fn → f ¢ 2 d in L R , then fˆ = lim fˆn in L2 -sense. n→∞ ¡ ¢ Theorem 18. Let f ∈ L Rd . Define Z 1 gn = f (x) e−ix·y dx, d/2 |x|<n (2π) 2 then fˆ = lim gn in L2 -sense. n→∞ Moreover, if hn = Z 1 (2π) d/2 fˆ (x) eix·y dx, |x|<n then f = lim hn in L2 -sense. n→∞ Proof. We leave the proof as an exercise. ¤ ¡ ¢ Remark 7. If f ∈ Lp Rd , 1 < p < 2, then we can write f = f1 + f2 where ¡ ¢ ¡ ¢ f1 ∈ L1 Rd and f2 ∈ L2 Rd , and hence we can define fˆ = fˆ1 + fˆ2 . To define Fourier transform on more general function, we need to introduce tempered distributions. Definition 4. A continuous linear functional on S is called a tempered distribution. And we use S 0 to denote the collection of all tempered distributions. ¡ ¢ ¡ d¢ Since D Rd is dense in S and the identity mapping I of D R into S is ¡ ¢ continuous, every Λ ∈ S 0 gives rise to Λ◦I ∈ D0 Rd . Moreover, the map Λ → Λ◦I is one-to-one. Hence, we can view tempered distributions as distributions which have continuous extension to S. Example 4. Every distribution with compact support is tempered. Example 5. Suppose µ is a positive Borel measure on Rd such that ³ ´−k 2 1 + |x| dµ (x) < ∞. Then µ is a tempered distribution. 12 HUIQIANG JIANG Example 6. Suppose 1 ≤ p < ∞, N > 0 and g is a measurable function on Rd such that Z ³ ´−k 2 p |g (x)| dx < ∞. 1 + |x| Rd Then g is a tempered distribution. In particular, every Lp function is a tempered distribution and so is every polynomial. 2 Example 7. e|x| is not a tempered distribution. ¡ ¢ Proof. Let φ ∈ D Rd be a radial symmetry function such that ³ φ (r)´ > 0 when 2 1 < r < 2 and φ (r) = 0 otherwise. One can check that e−|x| /2 φ |x| → 0 in S n while µ ¶ Z |x| |x|2 −|x|2 /2 e e φ dx = ∞. n Rd ¤ Theorem 19. If α is a multi-index, P is a polynomial, g ∈ S and Λ is a tempered distribution, then so are Dα Λ, P Λ and gΛ. Proof. For any f ∈ S, (Dα Λ) (f ) = (−1) |α| Λ (Dα f ) , since Dα : S → S is continuous, Dα Λ is continuous. Continuity of P Λ and gΛ follows similarly. ¤ Now we can define Fourier transform of tempered distributions. Definition 5. For any u ∈ S 0 , we define for any f ∈ S, ³ ´ û (f ) = u fˆ . Then û ∈ S 0 . If f ∈ L1 , then f is a tempered distribution which we denote by uf . And for any g ∈ S, we have Z ûf (g) = uf (ĝ) = f (x) ĝ (x) dx Rd Z Z Z = f (x) e−ix·y g (y) dydx = fˆ (y) g (y) = ufˆ (g) . Rd Rd Rd Hence, the definition of Fourier transform of tempered distributions is consistent with the Fourier transform of integrable functions. Similarly, one ¡ can ¢ show that it is also consistent with the Fourier transform of functions in L2 Rd . Theorem 20. The Fourier transform is a continuous, linear bijection from S 0 onto S 0 . Its inverse is also continuous. Proof. Let W be a nbhd of 0 in S 0 . Then there exists functions f1 , · · · , fn ∈ S s.t. {u ∈ S 0 : |u (fk )| < 1, 1 ≤ k ≤ n} ⊂ W. Define, ¯ ³ ´¯ n o ¯ ¯ V = u ∈ S 0 : ¯u fˆk ¯ < 1, 1 ≤ k ≤ n . Then V is a nbhd of 0 in S 0 , and F (V ) ⊂ W . Hence F : S 0 → S 0 is continuous. Since F 4 = I, F is a bijection and F −1 = F 3 is continuous. ¤ HARMONIC ANALYSIS 13 Theorem 21. If u ∈ S 0 , and α is a multi-index, then α (Dα u) ˆ = (iy) û, α Dα û = ((−ix) u) ˆ. Proof. By definition. ¤ d Example 8. 1̂ = (2π) 2 δ and δ̂ = 1 d (2π) 2 . Proof. For any f ∈ S, ³ ´ Z 1̂ (f ) = 1 fˆ = d fˆ (y) dy = (2π) 2 f (0) . Rd An for any f ∈ S, ³ ´ δ̂ (f ) = δ fˆ = fˆ (0) = Z 1 d (2π) 2 f (x) dx. Rd ¤ Example 9. Let α be a multi-index, then d (xα ) ˆ = (2π) 2 iα Dα δ. Since any distribution supported at 0 is a linear combination of distributions of the form Dα δ, the Fourier transform of f is supported at 0 iff f is a nonzero polynomial. Definition 6. Let u ∈ S 0 , and ϕ ∈ S. We define for any x ∈ Rd (u ∗ ϕ) (x) = u (ϕ (x − ·)) . ¡ ¢ Theorem 22. Let u ∈ S 0 , and ϕ ∈ S. Then u ∗ ϕ ∈ C ∞ Rd . Moreover, for any multi-index α, Dα (u ∗ ϕ) = Dα u ∗ ϕ = u ∗ Dα ϕ. Proof. Since the mapping x 7→ ϕ (x − ·) is continuous from Rd into S, u ∗ ϕ is continuous. Now µ ¶ (u ∗ ϕ) (x + tei ) − (u ∗ ϕ) (x) ϕ (x + tei − ·) − ϕ (x − ·) =u , t t since ϕ (x + tei − y) − ϕ (x − y) → ∂i ϕ (x − y) in S, t we have ∂i (u ∗ ϕ) (x) = lim t→0 (u ∗ ϕ) (x + tei ) − (u ∗ ϕ) (x) = u (∂i ϕ (x − ·)) , t ¡ ¢ hence u ∗ ϕ has a continuous i-th partial derivative. By induction, u ∗ ϕ ∈ C ∞ Rd . ¤ ¡ ¢ Lemma 1. Let u ∈ S 0 , and ϕ, φ ∈ D Rd , we have (u ∗ ϕ) ∗ φ = u ∗ (ϕ ∗ φ) . 14 HUIQIANG JIANG Proof. For any h > 0, we define the Riemann sum X ϕ (x − kh) φ (kh) fh (x) = hd k where for each h and x, the summation has finitely many nonzero terms. It is straight forward to verify that fh → ϕ ∗ φ in S. Hence, u ∗ (ϕ ∗ φ) = lim+ u ∗ fh h→0 X = lim+ hd (u ∗ ϕ) (x − kh) φ (kh) h→0 k = (u ∗ ϕ) ∗ φ. Here the last equality holds because (u ∗ ϕ) (x − y) φ (y) is Riemann integrable in y. ¤ ¡ d¢ Now let ϕ ∈ D R be the standard mollifier and for any ε > 0, we define ³x´ ϕε (x) = ε−n ϕ . ε Lemma 2. For any φ ∈ S, we have φ ∗ ϕε → φ in S. Proof. Direct calculation. ¤ Theorem 23. If u ∈ S 0 , then u ∗ ϕε → u in the weak-* topology of S 0 . Proof. For any φ ∈ S, we have ³ ´ u (φ) = u ∗ φ̃ (0) ¡ ¢ where φ̃ (x) = φ (−x). For any φ ∈ D Rd , (u ∗ ϕε ) (φ) ³ ´ = (u ∗ ϕε ) ∗ φ̃ (0) ³ ³ ´´ = u ∗ ϕε ∗ φ̃ (0) µ³ ´˜ ¶ = u ϕε ∗ φ̃ . µ³ Since the expression u ϕε ∗ φ̃ ´˜ ¶ is continuous for φ ∈ S, we conclude u∗ϕε ∈ S 0 and the above formula holds for φ ∈ S. Since ϕε ∗ φ̃ → φ̃ in S, we have lim (u ∗ ϕε ) (φ) = u (φ) . ε→0 ¤ ¢ Remark 8. Similar proof ¡ ¢ will yield that if u ∈ D R , then u ∗ ϕε → u in the weak-* topology of D0 Rd . 0 ¡ d Remark 9. The formulas for Fourier transforms of regular functions continue to hold for tempered distributions as long as they are well defined. For example, if u ∈ S 0 and ϕ ∈ S, then d (u ∗ ϕ) ˆ = (2π) 2 ϕ̂û. HARMONIC ANALYSIS 15 Definition 7. Let Ω ⊂ Rd be open and k be a nonnegative integer and 1 ≤ p ≤ ∞. Sobolev space W k,p (Ω) is the collection of functions f ∈ Lp and the distributional derivatives Dα f ∈ Lp for any |α| ≤ k. W k,p (Ω) is a Banach space with norm p1 X p kf kW k,p (Ω) = kDα f kLp (Ω) if 1 ≤ p < ∞, |α|≤k kf kW k,∞ (Ω) = max kDα f kL∞ (Ω) . |α|≤k When p = 2, W k,2 (Ω) = H k (Ω) is a Hilbert space. ¡ ¢ Let f ∈ H k Rd . Then we have for any |α| ≤ k, ¡ ¢ α f = (iy)α fˆ ∈ L2 Rd . [ D ³ ´k/2 ¡ ¢ ¡ ¢ 2 fˆ ∈ L2 Rd . Moreover, the norm Hence, f ∈ H k Rd iff 1 + |y| °³ ´k/2 ° ° ° 2 ° kf kH̃ k = ° 1 + |y| fˆ° ° L2 is equivalent to the ¡ d ¢original norm. This expression also allows us to extend ¡ d ¢the k k definition of H R to k ∈ R. One can identify the dual space of H R as ¡ d¢ −k H R . Theorem 24 (Sobolev). Let Ω ⊂ Rd be open and k > d2 . Any f ∈ H k ( ) can be represented by a continuous function. ¡ ¢ Proof. For any x ∈ Ω, let ϕ ∈ D Rd¡ has ¢ compact support in Ω and ϕ = 1 in Br (x) for some r > 0. Then ϕf ∈ H k Rd and¡ϕf ¢= f in Br (x). Since continuity is a local property, we assume f itself is in H k Rd . Now °³ ° Z ¯ ¯ ´k/2 ° ´−k/2 ° ° ° °³ ° ¯ ˆ¯ 2 2 ˆ ° ° ° ° < ∞. f ° ° 1 + |y| ¯f ¯ dx ≤ ° 1 + |y| ° Rd L2 So we have fˆ ∈ L ¡ 1 ³ ´∨ ¢ ¡ ¢ Rd and f = fˆ ∈ C0 Rd . L2 ¤ 4. The maximal function 4.1. Distribution function. We first recall the concept of distribution function. Definition 8. Let g be defined on Rn . For each t ≥ 0, we define λ (t) = |{x ∈ Rn : |g (x)| > t}| . The function λ is called the distribution function of |g|. The function λ is monotone decreasing on [0, ∞) and hence measurable. Any quantity dealing solely with the size of g can be expressed in terms of the distribution function λ (α). If g ∈ Lp (Rn ), then Z Z ∞ p |g (y)| = − tp dλ (t) . Rn And if g ∈ L ∞ 0 n (R ), then kgk∞ = inf {t : λ (t) = 0} . 16 HUIQIANG JIANG Proposition 5. If g ∈ Lp (Rn ), p ≥ 1, then for any t > 0, Z 1 p λ (t) ≤ p |g (y)| dy. t Rn Proof. Z Z p |g (y)| dy ≥ tp Rn dy = tp λ (t) . |g(x)|>t ¤ 4.2. A covering lemma. Lemma 3. Let E be a measurable subset of Rn which is covered by the union of a family of balls {Bα }α∈Λ of bounded radius. Then from this family, we can select a disjoint sequence, B1 , B2 , · · · , s.t. X |Bk | ≥ 5−n |E| . Proof. We choose B1 s.t. 1 sup r (Bα ) . 2 α∈Λ When B1 , B2 , · · · , Bk are chosen, we choose Bk+1 s.t., r (B1 ) ≥ 1 sup {r (Bα ) : α ∈ Λ, Bα ∩ Bi = ∅, 1 ≤ i ≤ k} . 2 This P procedure terminates if the set in the above expression is empty. P If |Bk | = ∞, we are done. So we now assume |Bk | < ∞. Since r (Bk ) ≥ 1 ∗ 2 r (Bj ) for any j ≥ k, either {Bk } is finite or limk→∞ r (Bk ) = 0. where Bk is the ball with the same center as Bk and with radius 5r (Bk ). Given any α ∈ Λ such that Bα is not one of {Bk }. If {Bk } is finite, then for some j, Bj ∩ Bα 6= ∅ and r (Bj ) ≥ 12 r (Bα ) which implies Bα ⊂ Bj∗ . If {Bk } is not finite, then there exists first k, s.t. r (Bk+1 ) < 12 r (Bα ). Hence, for some 1 ≤ j ≤ k, Bj ∩ Bα 6= ∅ and S r (Bj ) ≥ 21 r (Bα ) which implies Bα ⊂ Bj∗ . Hence, in any case, Bα ⊂ Bk∗ . So we r (Bk+1 ) ≥ k have E⊂ [ Bα ⊂ [ α∈Λ and |E| ≤ X |Bk∗ | = 5n Bk∗ k X |Bk | . ¤ 4.3. Hardy-Littlewood maximal function. We recall that if f ∈ L1loc (Rn ), then Z 1 f (y) dy f (x) = lim+ r→0 |Br (x)| Br (x) holds for a.e. x ∈ Rn . To study such limit, we introduce: Definition 9. Let f ∈ L1loc (Rn ). The maximal function of f is defined by Z 1 M (f ) (x) = sup |f (y)| dy, x ∈ Rn . r>0 |Br (x)| Br (x) HARMONIC ANALYSIS 17 Theorem 25. Let f be a given function defined on Rn (a) If f ∈ Lp (Rn ), 1 ≤ p ≤ ∞, then M f is finite a.e.. (b) If f ∈ L1 (Rn ), then for any t > 0, A m {x : (M f ) (x) > t} ≤ t Z |f | dx, Rn where we can take A = 5n . (c) If f ∈ Lp (Rn ), with 1 < p ≤ ∞, then M f ∈ Lp (Rn ) and kM f kp ≤ Ap kf kp where µ Ap = 2 5n p p−1 ¶ p1 , 1 < p < ∞ and A∞ = 1. Proof. (b) Let Et = {x : (M f ) (x) > t}. For any x ∈ Et , there exits a ball Bx centered at x, s.t., Z |f (y)| dy > t |Bx | . Bx The family {Bx } is a cover of set Et , hence the covering lemma implies the existence of disjoint sub-collection {Bxk } s.t. X |Bxk | ≥ 5−n |Et | . Hence, |Et | ≤ 5n X |Bxk | < Z Z 5n X 5n |f | dx ≤ |f | dx. t t Rn Bx k (c) The case p = ∞ is trivial with A∞ = 1. Now we assume 1 < p < ∞. For any t > 0, we define f1 (x) = f (x) if |f (x)| > 2t , f1 (x) = 0 otherwise. Then |f (x)| ≤ |f1 (x)| + 2t which yields t M f (x) ≤ M f1 (x) + . 2 Hence, ½ Et = {x : (M f ) (x) > t} ⊂ t x : (M f1 ) (x) > 2 ¾ . So we have ½ ¾ Z t 2A 2A λ (t) = |Et | ≤ m x : (M f1 ) (x) > ≤ kf1 kL1 = |f | . 2 t t |f (x)|> 2t 18 HUIQIANG JIANG Hence, Z p kM f kp ∞ Z ∞ t dλ (t) = p tp−1 λ (t) dt 0 0 ÃZ ! Z ∞ p−1 2A ≤p t |f | dx dt t 0 |f |> 2t ÃZ ! Z 2|f | p−2 t dt dx = 2pA |f (x)| =− p Rn 0 Z 2pA p−1 |f (x)| (2 |f |) dx p − 1 Rn 2p pA p p = kf kp = App kf kp . p−1 = ¤ Remark 10. Ap → ∞ as p → 1 which suggests that part (c) will not hold for p = 1. Actually, M f is never integrable unless f = 0 a.e.. The estimate in (c) is called type (p, p), while the estimate in (b) is called type weak (1, 1). Corollary 5. If f ∈ L1loc (Rn ), then f (x) = lim r→0+ 1 |Br (x)| Z f (y) dy Br (x) holds for a.e. x ∈ Rn . Proof. We assume f ∈ L1 (Rn ). For any x ∈ Rn , we define Z 1 fr (x) = f (y) dy |Br (x)| Br (x) and ¯ ¯ ¯ ¯ Ωf (x) = ¯¯lim sup fr (x) − lim sup fr (x)¯¯ . r→0+ r→0+ n For any ε > 0, we choose, for any δ > 0, h ∈ C (R ) and denote g = f − h s.t. kgk1 = kf − hk1 < δ. Then Ωf (x) = Ωg (x) ≤ 2M g. Now m {x : Ωf (x) > ε} = m {x : Ωg (x) > ε} n ε o 2 · 5n ≤m x : M g (x) > ≤ kgk1 ≤ 2 · 5n δ. 2 ε Since δ can be arbitrarily small, we have m {x : Ωf (x) > ε} = 0. Hence Ωf (x) = 0 a.e. x ∈ Rn . Hence limr→0+ fr (x) exists a.e.. On the other hand, since fr → f in L1 , there exists a subsequence rk s.t. frk → f a.e. x ∈ Rn . Hence, f (x) = lim+ fr (x) r→0 holds for a.e. x ∈ Rn . ¤ HARMONIC ANALYSIS 19 4.4. Generalized maximal function. Definition 10. Let F be a family of measurable subsets of Rn . We say F is regular, if there exists c > 0 s.t. if S ∈ F, then there exists a ball B centered at 0, s.t. S ⊂ B, m (S) ≥ cm (B). Here are some examples: Example 10. The family F = {δU }δ>0 is regular if U is bounded and measurable with m (U ) > 0. Example 11. The family F of all cubes with the property that their distance from the origin is bounded by a constant multiple of their diameter. Example 12. The family F of all balls containing the origin. Example 13. Any subfamily of a regular family is regular. Given a regular family F of measurable subsets of Rn . We define the maximal function Z 1 MF (f ) (x) = sup |f (x − y)| dy. S∈F m (S) S Let c > 0 be the number in the Definition of regular family, then we have 1 MF (f ) (x) ≤ M (f ) (x) . c Hence strong (p, p), 1 < p ≤ ∞ and weak (1, 1) estimates hold for MF . Corollary 6. Suppose F is a regular family of measurable subsets of Rn such that inf m (S) = 0. S∈F If f ∈ L1loc (Rn ), then f (x) = lim S∈F m(S)→0 1 m (S) Z f (x − y) dy S holds for a.e. x ∈ Rn . Definition 11. Let f ∈ L1loc (Rn ). A point x ∈ Rn is said to be a Lebesgue point of f if Z 1 lim+ |f (y) − f (x)| dy = 0. r→0 |Br (x)| Br (x) The Lebesgue set of f is the collection of all Lebesgue points of f . Theorem 26. Let f ∈ L1loc (Rn ). (a) The complement of Lebesgue set is of measure zero. (b) Suppose F is a regular family of measurable subsets of Rn such that inf m (S) = 0. S∈F Then f (x) = lim S∈F m(S)→0 n 1 m (S) holds for every Lebesgue point x ∈ R . Z f (x − y) dy S 20 HUIQIANG JIANG Proof. (a) For any s ∈ Q, 1 lim+ r→0 |Br (x)| Z |f (y) − s| dy = |f (x) − s| Br (x) holds for a.e. x ∈ Rn . We define the exceptional set Es , then |Es | = 0. Let E = ∪s∈Q Es , then |E| = 0. For any x 6∈ E, we have Z 1 lim+ |f (y) − s| dy = |f (x) − s| r→0 |Br (x)| Br (x) holds for every s ∈ Q. Since Q is dense in R, Z 1 lim |f (y) − f (x)| dy = 0, r→0+ |Br (x)| Br (x) i.e., x is a Lebesgue point. (b) Let x be a Lebesgue point. Then ¯ ¯ Z Z ¯ 1 ¯ ¯ ¯≤ 1 f (x − y) dy − f (x) |f (x − y) − f (x)| dy ¯ m (S) ¯ m (S) S Z S 1 ≤ |f (x − y) − f (x)| dy → 0 c |B| B as |B| → 0. ¤ 4.5. An interpolation theorem for Lp . Let 1 ≤ p, q ≤ ∞ and T be a mapping from Lp (Rn ) to measurable functions in Rn . Then T is of type (p, q) if for some A > 0, kT f kq ≤ A kf kp holds for every f ∈ Lp (Rn ) . Similarly, if in addition, q < ∞, T is of weak type (p, q) if for some A > 0, µ ¶q A kf kp m {x : |T f (x)| > t} ≤ holds for every f ∈ Lp (Rn ) . t When q = ∞, T is of weak type (p, q) if it is of type (p, q). If T is of type (p, q), then T is of weak type (p, q). Assume q < ∞ and T is of type (p, q). For every f ∈ Lp (Rn ), µ ¶q A kf kp 1 q m {x : |T f (x)| > t} ≤ q kT f kq ≤ . t t Hence, T is of weak type (p, q). The following is a special case of Marcinkiewicz interpolation theorem: Theorem 27 (Marcinkiewicz). Suppose 1 < r ≤ ∞ and T maps L1 (Rn ) + Lr (Rn ) into the space of measurable functions in Rn . Suppose that T is (i) subadditive: For any f, g ∈ L1 (Rn ) + Lr (Rn ), |T (f + g) (x)| ≤ |T f (x)| + |T g (x)| ; (ii) of weak type (1, 1): For any f ∈ L1 (Rn ), m {x : |T f (x)| > t} ≤ A1 kf k1 ; t HARMONIC ANALYSIS 21 (iii) of weak type (r, r): If r < ∞, for any f ∈ Lr (Rn ), µ ¶r Ar kf kr m {x : |T f (x)| > t} ≤ t and if r = ∞, kT f k∞ ≤ A∞ kf k∞ . Then for any 1 < p < r, kT f kp ≤ Ap kf kp holds for every f ∈ Lp (Rn ) where ·µ Ap = r¶ 2r A1 (2Ar ) + p−1 r−p ¸ p1 p if r < ∞. Proof. The case r = ∞ follows from the proof of (p, p) estimate for maximal function. Now we assume 1 < r < ∞. Fix f ∈ Lp . For any t > 0, we define f1 = f χ|f |>t and f2 = f χ|f |≤t . Then f1 ∈ L1 and f2 ∈ Lr . Since |T (f1 + f2 ) (x)| ≤ |T f1 (x)| + |T f2 (x)| , we have ½ ¾ ½ ¾ t t {x : |T f (x)| > t} ⊂ x : |T f1 (x)| > ∪ x : |T f2 (x)| > . 2 2 Hence, λ (t) = |{x : |T f (x)| > t}| ¯½ ¾¯ ¯½ ¾¯ ¯ t ¯¯ ¯¯ t ¯¯ ≤ ¯¯ x : |T f1 (x)| > + x : |T f (x)| > 2 2 ¯ ¯ 2 ¯ µ ¶r 2Ar kf2 kr 2A1 kf1 k1 + ≤ t t µ ¶r Z Z 2A1 2Ar r = |f | dx + |f | dx. t t |f |>t |f |≤t Recall that Z p kT f kp Now we have Z ∞ t p−1 0 and Z p−1 t 0 Hence, µ p kT f kp ≤ tp−1 λ (t) dt. 0 à Z ! 1 1 p |f | dx dt = kf kp t |f |>t p−1 à ∞ ∞ =p 1 tr ! Z r |f | dx dt = |f |≤t r¶ 2A1 (2Ar ) + p−1 r−p 1 p kf kp . r−p p p p kf kp = App kf kp . ¤ 22 HUIQIANG JIANG 4.6. Behavior near general points of measurable sets. Suppose E is a given measurable subset of Rn . We say x ∈ Rn is a point of density of E if lim r→0 |E ∩ Br (x)| = 1. |Br (x)| Proposition 6. Almost every point x ∈ E is a point of density of E, and almost every point x 6∈ E is not a point of density of E. Now let E be a closed set. For any x ∈ Rn , we denote δ (x) = dist (x, E). Proposition 7. Let E be a closed set. Then for almost every x ∈ E, δ (x + y) = o (|y|). This holds in particular if x is a point of density of E. Proof. Let x be a point of density of F and suppose ε > 0 is given. If δ (x + y) > ε |y|, then ¯ ¯ µ ¶n ¯E ∩ B(1+ε)|y| (x)¯ ε ¯ ¯ ≤1− ¯B(1+ε)|y| (x)¯ 1+ε which is impossible if |y| is sufficiently small. ¤ Now we consider the integral of marcinkiewicz: Z δ (x + y) I (x) = n+1 dy. |y|≤1 |y| We need the following lemma: Lemma 4. Let E be a closed set such that |E c | < ∞ and Z δ (x + y) I∗ (x) = n+1 dy. n |y| R Then I∗ (x) < ∞ for a.e. x ∈ E. Moreover, Z I∗ (x) dx ≤ nωn |E c | E where ωn is the volume of the unit ball in Rn . Proof. Z Z Z I∗ (x) dx = E δ (x + y) E Rn E Ec n+1 |y| δ (y) Z Z = Z = |x − y| ÃZ δ (y) Ec E ÃZ Z ≤ dydx n+1 dydx 1 = nωn Z Ec n+1 dx |x − y| δ (y) Ec ! |x|≥δ(y) 1 |x| n+1 dx dy ! dy dy = nωn |E c | . ¤ HARMONIC ANALYSIS 23 Theorem 28. Let E be a closed set. (a) When x 6∈ E, I (x) = ∞; (b) For almost every x ∈ F , I (x) < ∞. Remark 11. We can also define for λ > 0, Z δ λ (x + y) (λ) I (x) = dy. n+λ |y| |y|≤1 4.7. Decomposition in cubes of open sets in Rn . Theorem 29 (Calderón Zygmund decomposition). Let f be a nonnegative integrable function on Rn and α > 0. Then there exists a decomposition of Rn so that (i) Rn = F ∪ Ω, F ∩ Ω = ∅; (ii) f (x) ≤ α a.e. on F . (iii) Ω is the union of closed cubes, Ω = ∪k Qk , whose interiors are disjoint and for each k, Z 1 α< f (x) dx ≤ 2n α. |Qk | Qk Moreover, 1 kf k1 . α Proof. We decompose Rn into a mesh of equal cubes such that Z Z 1 1 f (x) dx ≤ f (x) dx ≤ α. |Q| Q |Q| Rn |Ω| ≤ Let Q0 be a fixed cube, we divide it into 2n congruent cubes. For each new cube Q00 , Case 1: Z 1 f (x) dx ≤ α. |Q00 | Q00 Then we continue the process. Case 2: 1 |Q00 | Z f (x) dx > α. Q00 We select Q00 as one of Qk . Let Ω = ∪k Qk , and F = Rn \Ω. Then f (x) ≤ α a.e. on F . ¤ Theorem 30. Let F be a nonempty closed set in Rn . Then its complement Ω = F c can be written as a union of cubes Ω= ∞ [ Qk , k=1 whose interiors are disjoint and for each k, diam Qk ≤ dist (Qk , F ) ≤ 4 diam Qk . Proof. Consider a mesh M0 of cubes with unit length whose vertices are integral points. Let Mk = 2−k M0 . We also define for some c > 0, ª © Ωk = x : c2−k < dist (x, F ) < c2−k+1 . 24 HUIQIANG JIANG Let F0 = √ Then with c = 2 n, we have [ {Q ∈ Mk : Q ∩ Ωk 6= ∅} . k diam Q ≤ dist (Q, F ) ≤ 4 diam Q S for each Q ∈ F0 . We have Ω = Q. We observe that if Q◦1 ∩ Q◦1 6= ∅, then one Q∈F0 cube contains the other. For each Q ∈ F0 , let Q0 be the maximal cube in F0 which contain Q. Such maximal cube is unique and all the interiors of maximal cubes are disjoint. Let {Qk } be the collection of all maximal cubes in Ω. We have the desired properties. ¤ With the help of Theorem 30, we can improve certain aspects of the Calderon Zygmund decomposition. Theorem 31. Let f be a nonnegative integrable function on Rn and α > 0. Let F = {x : M f (x) ≤ t} and Ω = F c = {x : M f (x) > t} . We can write Ω as the union of closed cubes, Ω = ∪k Qk , whose interiors are disjoint and for each k, diam Qk ≤ dist (Qk , F ) ≤ 4 diam Qk . Moreover, for each k, 1 |Qk | Z f (x) dx ≤ cα Qk where c is a number depending only on n. Proof. Since F is closed, its complement Ω = F c can be written as a union of cubes Ω= ∞ [ Qk , k=1 whose interiors are disjoint and for each k, diam Qk ≤ dist (Qk , F ) ≤ 4 diam Qk . For each cube Qk , let pk ∈ F be such that dist (Qk , F ) = dist (Qk , pk ). Let rk = max dist (x, pk ) ≤ 5 diam Qk , x∈Qk then Qk ⊂ Brk (pk ). So we have Z 1 |f (x)| dx |Brk (pk )| Brk (pk ) Z f (x) dx t ≥ M f (pk ) ≥ 1 |Brk (pk )| Qk Z 1 ≥ f (x) dx. c |Qk | Qk ≥ HARMONIC ANALYSIS Hence 1 |Qk | 25 Z f (x) dx ≤ ct. Qk ¤ Remark 12. We also have |Ω| ≤ 2n kf k1 . t 5. Singular Integral 5.1. Convolution of measures. The dual space of C0 (Rn ) is M (Rn ), the space of all finite Borel measures. For any µ1 , µ2 ∈ M (Rn ), µ = µ1 ∗ µ2 is defined by Z Z f (x + y) dµ1 (x) dµ2 (y) . µ (f ) = Rn Rn We have µ1 ∗ µ2 = µ2 ∗ µ1 and kµk ≤ kµ1 k kµ2 k. Let 1 ≤ p < ∞. If f ∈ Lp (Rn ), then f ∗ µ ∈ Lp (Rn ) and kf ∗ µkp ≤ kf kp kµk . The map T : f 7→ f ∗ µ is a bounded linear operator from Lp (Rn ) to Lp (Rn ) which commutes with translation. Such mappings can be classified. Theorem 32. Suppose T : Lp (Rn ) → Lq (Rn ) is linear, bounded and commutes with translations. Then there exists a unique tempered distribution u ∈ S 0 such that T (ϕ) = ϕ ∗ u for any ϕ ∈ S. Before presenting its proof, we consider special cases when p = q where we can say more on the tempered distribution u. Theorem 33. Suppose T : L1 (Rn ) → L1 (Rn ) is linear, bounded and commutes with translations. Then there exists a unique µ ∈ M (Rn ) such that T (f ) = f ∗ µ for any f ∈ L1 (Rn ). Moreover, kT k = kµk. Proof. Let u ∈ S 0 be such that T (ϕ) = ϕ ∗ u for any ϕ ∈ S. We have for any ϕ ∈ S, kϕ ∗ uk1 ≤ kT k kϕk1 . Suppose ϕε is the standard mollifier, we have kϕε ∗ uk1 ≤ kT k . Hence, ϕεk ∗ u * µ for a subsequence εk → 0. On the other hand, we also have ϕεk ∗ u → u in S 0 . So we have u = µ. kT k = kµk can be verified since lim kϕε ∗ µk = kµk . ε→0 ¤ Theorem 34. Suppose T : L2 (Rn ) → L2 (Rn ) is linear, bounded and commutes with translations. Then there exists a multiplier m ∈ L∞ (Rn ), such that Tcf = m · fˆ for any f ∈ L2 (Rn ). 26 HUIQIANG JIANG Proof. Let u ∈ S 0 be such that T (ϕ) = ϕ ∗ u for any ϕ ∈ S. Then for ϕ = e− x2 2 , n/2 Tcϕ = (2π) ϕ̂ · û. For any φ ∈ D (Rn ), since ϕ̂ = e− x2 2 we have φ ϕ̂ ∈ S and φc n/2 T ϕ = (2π) φ · û ∈ L2 . ϕ̂ So we have û ∈ L2loc (Rn ). Now for any ϕ ∈ S, we have ° ° ° ° n/2 (2π) kϕ̂ · ûk2 = °Tcϕ° = kT ϕk2 ≤ A kϕ̂k2 . 2 Hence û ∈ L∞ (Rn ). Let m = (2π) n/2 û ∈ L∞ (Rn ). We have for any ϕ ∈ S, Tcϕ = mϕ̂. Since S is dense in L2 (Rn ), the continuities of T and Fourier transform imply Tcf = m · fˆ for any f ∈ L2 (Rn ). ¤ Lemma 5. Suppose 1 < p < ∞ and T : Lp (Rn ) → Lp (Rn ) is linear, bounded and commutes with translations. Then T , restricted to Lp (Rn ) ∩ Lq (Rn ), extends to a map T : Lq (Rn ) → Lq (Rn ) which is linear, bounded and commutes with translations. Proof. Let u ∈ S 0 be such that T (ϕ) = ϕ ∗ u for any ϕ ∈ S. We observe that for any ψ ∈ C0∞ (Rn ), if u ∈ C0∞ (Rn ), we have Z (ψ ∗ u) (x) ϕ (x) dx Rn Z Z = ψ (y) u (x − y) ϕ (x) dxdy Rn Rn Z Z = ψ (y) u (z) ϕ (y + z) dxdy n n ZR R = ψ (y) (u ∗ ϕ̃) (−y) dxdy. Rn Using Lemma 1, we can show that for any ϕ, ψ ∈ C0∞ (Rn ), Z Z (ψ ∗ u) (x) ϕ (x) dx = ψ (y) (ϕ̃ ∗ u) (−y) dxdy, Rn i.e., Rn Z Z (T ψ) (x) ϕ (x) dx = Rn ψ (y) (T ϕ̃) (−y) dxdy. Rn For fixed ψ ∈ C0∞ (Rn ), ¯Z ¯ ¯ ¯ ¯ ¯ ¯ n (T ψ) (x) ϕ (x) dx¯ ≤ kψkq kT ϕ̃kp ≤ kT k kψkq kϕkp R holds for any ϕ, ψ ∈ C0∞ (Rn ), hence kT ψkq ≤ kT k kψkq . So T can be extended to a continuous linear functional T : Lq (Rn ) → Lq (Rn ). ¤ HARMONIC ANALYSIS 27 Theorem 35. Suppose 1 < p < ∞ and T : Lp (Rn ) → Lp (Rn ) is linear, bounded and commutes with translations. Then there exists a multiplier m ∈ L∞ (Rn ), such that Tcϕ = m · ϕ̂ for any ϕ ∈ S. Proof. T : Lp (Rn ) → Lp (Rn ) and T can be extended to T : Lq (Rn ) → Lq (Rn ) with 1 1 + = 1. p q Hence Marcinkiewicz interpolation theorem implies T : L2 (Rn ) → L2 (Rn ). The conclusion now follows from Theorem 34. ¤ Now we prepare to prove Theorem 32. Lemma 6. If f ∈ W n+1,p (Rn ), 1 ≤ p ≤ ∞, then f equals a.e. a continuous function g satisfying X kgk∞ ≤ C kDα f kp . |α|≤n+1 Proof. We first assume f ∈ D (Rn ). Then we have ¯ ¯ ³ ´− n+1 X 2 ¯ˆ ¯ 2 ¯f (x)¯ ≤ c 1 + |x| ¯ ¯ ¯ α¯ |x| ¯fˆ (x)¯ |α|≤n+1 ³ = c 1 + |x| 2 ´− n+1 2 2 ´− n+1 2 ¯ ¯ ¯[ ¯ ¯Dα f (x)¯ X |α|≤n+1 ³ ≤ c 1 + |x| X kDα f k1 . |α|≤n+1 Hence X kf k∞ ≤ c2 Z kDα f k1 |α|≤n+1 ³ 1 + |x| 2 ´− n+1 2 X dx ≤ C Rn kDα f k1 . |α|≤n+1 Hence the case p = 1 follows from the density of D (Rn ) in W n+1,1 (Rn ). For 1 < p ≤ ∞, f ∈ W n+1,p (Rn ) implies ϕf ∈ W n+1,1 (Rn ) where ϕ ∈ D (Rn ) is a cutoff function s.t. ϕ = 1 on B1 (0) and ϕ = 0 on Rn \B2 (0). Then ϕf is continuous, and X X |f (0)| = |ϕf (x)| ≤ C kDα (ϕf )k1 ≤ C2 kDα f kL1 (B2 (x)) ≤ C3 X |α|≤n+1 |α|≤n+1 kDα f kp . |α|≤n+1 Since the choice of origin is arbitrary, we have X kgk∞ ≤ C3 kDα f kp . |α|≤n+1 ¤ Now we prove Theorem 32: 28 HUIQIANG JIANG Proof of Theorem 32. Let ϕ ∈ S. Since τte1 ϕ − ϕ → ∂x1 ϕ in Lp as t → 0+ , t we have µ ¶ τte1 ϕ − ϕ (T ϕ) (x + te1 ) − (T ϕ) (x) =T (x) → T (∂x1 ϕ) t t in Lq as t → 0+ . Hence ∂x1 (T ϕ) = T (∂x1 ϕ) ∈ Lq . Similar calculation yields Dα (T ϕ) = T (Dα ϕ) ∈ Lq for any α. Hence T ϕ is continuous and X X |(T ϕ) (0)| ≤ C kDα (T ϕ)kq = C kT (Dα ϕ)kq |α|≤n+1 ≤ C kT k |α|≤n+1 X α kD ϕkp . |α|≤n+1 Hence the map ϕ → (T ϕ) (0) is a continuous map on S, and there exists a unique ũ ∈ S 0 s.t. (T ϕ) (0) = ũ (ϕ) . 0 Now we define u ∈ S s.t u (ϕ) = ũ (ϕ̃), then we have (T ϕ) (x) = T (τx ϕ) (0) = ũ (ϕ (x + ·)) = u (ϕ (x − ·)) = (ϕ ∗ u) (x) . ¤ 5.2. Singular integrals: the heart of the matter. In this section, we will study a typical singular integral. Theorem 36. Let K ∈ L2 (Rn ). We suppose: (a) ¯ ¯ ¯ ¯ ¯K̂ (x)¯ ≤ B. (b) K ∈ C 1 (Rn \ {0}) and |∇K (x)| ≤ B |x| 1 n+1 . p For f ∈ L ∩ L , 1 < p < ∞, define Z (T f ) (x) = K (x − y) f (y) dy. Rn Then kT f kp ≤ Ap kf kp where Ap = c (n, p) B. Proof. We assume B = 1. Step 1. T is of type (2, 2). Since T f = K ∗ f , n Tcf = (2π) 2 K̂ · fˆ, we have ° ° n ° ° kT f k2 = °Tcf ° ≤ (2π) 2 2 ° ° n ° ˆ° °f ° = (2π) 2 kf k2 . 2 Step 2. T is of weak type (1, 1). Fix α > 0, we define F = {x : M f (x) ≤ t} and Ω = F c = {x : M f (x) > t} . HARMONIC ANALYSIS 29 From Theorem 31, we can write Ω as the union of closed cubes, Ω = ∪k Qk , whose interiors are disjoint and for each k, diam Qk ≤ dist (Qk , F ) ≤ 4 diam Qk . Moreover, for each k, 1 |Qk | Z f (x) dx ≤ cα. Qk Moreover, c kf k1 . α Here we use c to denote different constants depending only on n. Now we define ½ for x ∈ F, R f (x) g (x) = 1 ◦ |Qk | Qk f (x) dx for x ∈ Qk . |Ω| ≤ Then g is defined a.e. Let b = f − g, then we have b (x) = 0 for x ∈ F and Z b (x) dx = 0 Qk for each cube Qk . Since Z Z Z g 2 (x) dx = Rn f 2 (x) dx + g 2 (x) dx F Ω Z 2 ≤α |f (x)| dx + (cα) |Ω| F ≤ cα kf k1 . We have ¯n c c 4 α o¯¯ ¯ 2 2 ¯ ≤ 2 kT gk2 ≤ 2 kgk2 ≤ kf k1 . ¯ x : |T g (x)| > 2 α α α Now we estimate T b. Write bk (x) = b (x) χQk . We have for any x ∈ F , ¯Z ¯ ¯Z ¯ ¯ ¯ ¯ ¯ ¯ ¯ ¯ |T bk (x)| = ¯ K (x − y) b (y) dy ¯ = ¯ [K (x − y) − K (x)] b (y) dy ¯¯ Qk Qk Z |b (y)| ≤ c diam Qk n+1 dy. Qk |x − y| Since Z Z |b (y)| dy ≤ Qk |f (y)| dy + cα |Qk | ≤ cα |Qk | , Qk and diam Qk ≤ dist (Qk , F ) ≤ 4 diam Qk . we have for any x ∈ F , Z |T bk (x)| ≤ cα Qk δ (y) n+1 dy |x − y| 30 HUIQIANG JIANG where δ (y) = dist (y, F ). Hence, Z δ (y) |T b (x)| ≤ cα Ω |x − y| n+1 dy. Applying Lemma 4, we have Z |T b (x)| dx ≤ cα |Ω| ≤ c kf k1 . F Hence, ¯n α o¯¯ ¯ ¯ x ∈ Rn : |T b (x)| > ¯ 2 o¯ ¯n α ¯ ¯ ≤ ¯ x ∈ F : |T b (x)| > ¯ + |Ω| 2 c ≤ kf k1 . α Hence |{x : |T f (x)| > α}| ¯n α o¯¯ ¯¯n α o¯¯ ¯ ≤ ¯ x : |T g (x)| > ¯ + ¯ x : |T b (x)| > ¯ 2 2 c ≤ kf k1 . α Step 3. Interpolation theorem implies the case 1 < p < 2. Step 4. The case 2 < p < ∞ comes from the standard duality argument since Z (K ∗ f ) (x) ϕ (x) dx Rn Z = (K ∗ ϕ̃) (−x) ϕ (x) dx. Rn ¤ 2 n Remark 13. The assumption K ∈ L (R ) can be dropped. Remark 14. We could use standard Calderón Zygmund decomposition instead of Theorem 31. 5.3. Singular integrals: Some extension and variants. Theorem 37. Let K ∈ L2 (Rn ). We suppose: (a) ¯ ¯ ¯ ¯ ¯K̂ (x)¯ ≤ B. (b) For any |y| > 0, Z |K (x − y) − K (x)| dx ≤ B. |x|≥2|y| For f ∈ L1 ∩ Lp , 1 < p < ∞, define Z (T f ) (x) = K (x − y) f (y) dy. Rn Then kT f kp ≤ Ap kf kp where Ap = c (n, p) B. HARMONIC ANALYSIS 31 Proof. We only need to establish weak (1, 1) estimate. Let F, Ω and √ Qk be defined in the Calderon Zygmund decomposition to |f |. We define Q∗k = 2 nQk , [ Ω∗ = Q∗k and F ∗ = Rn \Ω∗ . k Let yk be the center of Qk . We have for any x 6∈ Q∗k and y ∈ Qk , |x − yk | ≥ 2 |y − yk | . Since Z T bk (x) = Z K (x − y) b (y) dy = [K (x − y) − K (x − yk )] b (y) dy, Qk Qk we have for any x ∈ F ∗ Z XZ |T b (x)| ≤ F∗ = k x6∈Q∗ k k Qk XZ Z ≤ Ω Hence, Z |[K (x − y) − K (x − yk )] b (y)| dydx y∈Qk |b (y)| dy |b (y)| dy ≤ c kf k1 . ¯n α o¯¯ ¯ ¯ x ∈ Rn : |T b (x)| > ¯ 2 o¯ ¯n α ¯ ¯ ≤ ¯ x ∈ F ∗ : |T b (x)| > ¯ + |Ω∗ | 2 c c ≤ kf k1 + c |Ω| ≤ kf k1 . α α ¤ Remark 15. Theorem 36 is a special case of the above theorem. Theorem 38. Suppose the kernel K satisfies (a) For any |x| > 0, −n |K (x)| ≤ B |x| . (b) For any |y| > 0, Z |K (x − y) − K (x)| dx ≤ B. |x|≥2|y| (c) For any 0 < R1 < R2 < ∞, Z K (x) dx = 0 R1 <|x|<R2 For f ∈ Lp , 1 < p < ∞, define for ε > 0, Z (Tε f ) (x) = f (x − y) K (y) dy. |y|≥ε Then kTε f kp ≤ Ap kf kp where Ap = c (n, p) B. Also, lim Tε f = T f ε→0 32 HUIQIANG JIANG exists in Lp for each f ∈ Lp and kT f kp ≤ Ap kf kp . Lemma 7. Let ½ K (x) for |x| ≥ ε, 0 for |x| < ε. ¯ ¯ ¯ ¯c sup ¯K ε (y)¯ ≤ cB Kε (x) = Then y where c = c (n). Proof. We first assume ε = 1. We observe that K1 satisfies (a) For any |x| > 0, −n |K1 (x)| ≤ B |x| . (b) For any |y| > 0, Z |K1 (x − y) − K1 (x)| dx ≤ cB. |x|≥2|y| (c) For any 0 < R1 < R2 < ∞, Z K1 (x) dx = 0. R1 <|x|<R2 Z To see this, |K1 (x − y) − K1 (x)| dx |x|≥2|y| Z Z ≤ |K1 (x − y) − K1 (x)| dx + |x|≥2|y| Z ≤B+ 1≤|x−y|≤ 23 Z Z |x|≥2|y| |x|≤1 |K1 (x − y)| dx + |K1 (x − y)| dx + |x|≥2|y| |x−y|≤1 |K1 (x)| dx |K1 (x)| dx ≤ cB. 1≤|x|≤2 Now since K1 ∈ L2 (Rn ), for any |y| > 0, Z 1 K̂1 (y) = lim K1 (x) e−ix·y dx R→∞ (2π)n/2 |x|≤R Z Z 1 1 −ix·y = K (x) e dx + lim K1 (x) e−ix·y dx. 1 n/2 n/2 2π 2π R→∞ |x|≤ |y| ≤|x|≤R (2π) (2π) |y| We estimate ¯Z ¯ ¯ ¯ ¯ ¯ −ix·y K1 (x) e dx¯ ¯ ¯ ¯ |x|≤ 2π |y| ¯Z ¯ ¯ £ −ix·y ¤ ¯¯ ¯ =¯ K1 (x) e − 1 dx¯ ¯ |x|≤ 2π ¯ |y| Z |x| |K1 (x)| dx ≤ |y| 2π |x|≤ |y| Z 1 ≤ B |y| 2π |x|≤ |y| = 2nωn πB. |x| n−1 dx HARMONIC ANALYSIS πy . |y|2 On the other hand, let z = 33 We have e−iz·y = e−iπ = −1 and Z 2π ≤|x|≤R |y| K1 (x − z) e−ix·y dx Z =− 2π ≤|x+z|≤R |y| Hence, K1 (x) e−ix·y dx. Z lim R→∞ 2π ≤|x|≤R |y| K1 (x) e−ix·y dx Z 1 lim (K1 (x) − K1 (x − z)) e−ix·y 2π 2 R→∞ |y| ≤|x|≤R Z Z 1 1 −ix·y K (x) e K1 (x) e−ix·y dx − dx + 2π 2π 1 2 2π|x|≤ |y| 2 2π|x|≥ |y| = |y| ≤|x+z| |y| ≥|x+z| = I + II + III. We have |II| ≤ 1 2 1 |III| ≤ 2 and Z Z π 2π ≤|x|≤ |y| |y| 2π 3π ≤|x|≤ |y| |y| Z |K1 (x)| dx ≤ cB, |K1 (x)| dx ≤ cB Z |I| ≤ 2π ≤|x|≤∞ |y| |K1 (x) − K1 (x − z)| dx = |K1 (x) − K1 (x − z)| dx ≤ cB |x|≥2|z| Hence, we conclude ¯ ¯ ¯c ¯ sup ¯K 1 (y)¯ ≤ cB. y The general ε case can be obtained by a scaling. Now we are ready to prove Theorem 38: Proof of Theorem 38. Since ¯ ¯ ¯c ¯ sup ¯K ε (y)¯ ≤ cB, y we have kTε f kp ≤ Ap kf kp where Ap = c (n, p) B. Now assume f ∈ C0∞ (Rn ), then we have Z (Tε f ) (x) = f (x − y) K (y) dy |y|≥ε Z Z = f (x − y) K (y) dy + (f (x − y) − f (x)) K (y) dy |y|≥1 = I + II. ε≤|y|≤1 ¤ 34 HUIQIANG JIANG Since f ∈ L1 and K ∈ Lp (Rn \B1 (0)), I ∈ Lp . Now II is compactly supported and since |f (x − y) − f (x)| ≤ A |y| , II converges uniformly to Z (f (x − y) − f (x)) K (y) dy. |y|≤1 Hence, Z Z (f (x − y) − f (x)) K (y) dy f (x − y) K (y) dy + Tε f → |y|≤1 |y|≥1 in Lp . Now standard density argument implies Tε f → T f for each f ∈ Lp . ¤ Let 1 , x ∈ R1 . πx It is easy to check that K satisfies the assumptions in Theorem 38. Hence for any f ∈ Lp , the Hilbert transform Z f (x − y) 1 T f = lim dy + y ε→0 π |y|≥ε K (x) = exists in the Lp norm. 5.3.1. Singular integral operators which commute with dilations. Let Tf = K ∗ f p for f ∈ L . T commutes with dilation iff K is homogeneous of degree −n. i.e., K (x) = Ω (x) n , |x| where Ω is homogeneous of degree 0, i.e., Ω (εx) = Ω (x) for any ε > 0. When K is homogeneous of degree −n, the cancellation property in Theorem 38 is equivalent to Z Ω (x) dσ = 0. Sn−1 For the assumption b, we have Lemma 8. Assume that Ω satisfies the following ”Dini-type” condition Z 1 ω (δ) dδ < ∞ (5.1) δ 0 where the modulus of continuity ω (δ) = sup |Ω (x) − Ω (x0 )| . |x−x0 |≤δ, |x|=|x0 |=1 Then there exists B > 0, such that for any y > 0, Z |K (x − y) − K (x)| dx ≤ B. |x|≥2|y| HARMONIC ANALYSIS Proof. 35 · ¸ Ω (x − y) − Ω (x) 1 1 K (x − y) − K (x) = + Ω (x) n n − n . |x − y| |x − y| |x| Now ¯ ¯ ¯ 1 1 ¯¯ ¯ n − n dx ≤ c. ¯ |x| ¯ |x|≥2|y| |x − y| Z Since ¯ ¯ ¯y¯ ¯ x−y x ¯¯ ¯ ¯ ¯ ≤ c − ¯ ¯ ¯ |x − y| |x| ¯ x holds for any |x| ≥ 2 |y|, we have ¯ µ µ ¶ µ ¶¯ ¶ ¯ x−y |y| x ¯¯ |Ω (x − y) − Ω (x)| = ¯¯Ω −Ω ≤ ω c |x − y| |x| ¯ |x| and ¯ ¯ ¯ Ω (x − y) − Ω (x) ¯ ¯ ¯ dx n ¯ ¯ |x − y| |x|≥2|y| µ ¶ Z |y| 1 ≤c ω c n dx |x| |x| |x|≥2|y| Z 2c ω (δ) =c dδ < ∞. δ 0 Z Hence, Z Z c 2 |K (x − y) − K (x)| dx ≤ c |x|≥2|y| ω (δ) dδ + c kΩk∞ δ 0 µZ 1 ≤c 0 ¶ ω (δ) dδ + kΩk∞ . δ ¤ Remark 16. The Dini-type condition is satisfied if Ω is Hölder continuous. Theorem 39. Let Ω be homogeneous of degree 0. Suppose that Z Ω (x) dσ = 0 Sn−1 and Dini condition (5.1) holds. For 1 < p < ∞ and f ∈ Lp (Rn ), define Z Ω (y) Tε f = n f (x − y) dy. |y|≥ε |y| (a) There exists constant Ap independent of ε, s.t. kTε f kp ≤ Ap kf kp . (b) lim Tε f = T f ε→0 exists in Lp norm, and kT f kp ≤ Ap kf kp . 36 HUIQIANG JIANG (c) If f ∈ L2 (Rn ), then the Fourier transforms of f and T f are related by Tcf (y) = m (y) fˆ (y) where m is a homogeneous function of degree 0. Proof. (a) (b) follows from Theorem 38. (c) Let ( Ω(x) |x|n Kε,η (x) = 0 Then we have K̂ε,η (y) = if ε ≤ |x| ≤ η, otherwise. Z 1 n/2 (2π) ε≤|x|≤η Ω (x) −ix·y dx. n e |x| Writing r = |x|, R = |y|, x = |x| x0 and y = |y| y 0 for x, y 6= 0, we have for any y 6= 0, Z 1 Ω (x0 ) −iRrx0 ·y0 K̂ε,η (y) = e dx n/2 rn ε≤|x|≤η (2π) Z Z η 1 Ω (x0 ) −iRrx0 ·y0 = e drdσ (x0 ) n/2 r n−1 S ε (2π) Z Z η ´ 1 Ω (x0 ) ³ −iRrx0 ·y0 = e − cos (Rr) drdσ (x0 ) n/2 r Sn−1 ε (2π) Z 1 = Ω (x0 ) Iε,η (x0 , y) dσ (x0 ) n/2 Sn−1 (2π) where Z Iε,η (x0 , y) = ε η ´ 1 ³ −iRrx0 ·y0 e − cos (Rr) dr. r Now, Z η 1 sin (Rrx0 · y 0 ) dr ε r Z ∞ sin t π →− dt sign (x0 · y 0 ) = − sign (x0 · y 0 ) t 2 0 Im Iε,η = − as ε → 0 and η → ∞. Moreover, |Im Iε,η | is uniformly bounded for 0 < ε < η < ∞. On the other hand, Z η 1 Re Iε,η = [cos (Rrx0 · y 0 ) − cos (Rr)] dr ε r 1 → ln 0 0 |x · y | HARMONIC ANALYSIS 37 as ε → 0, η → ∞ and |Re Iε,η | ≤ 2 ln |x01·y0 | . To see this, we consider for λ, µ > 0, Z η cos (λr) − cos (µr) dr r ε Z λη Z µη cos r cos r = dr − dr r r λε µε Z λη Z µε cos r cos r = dr + dr r r µη λε µ ¶ Z λ Z µε Z µε 1 r 1 cos r − 1 = cos dr + dr + dr r η r r µ λε λε µ ¶ Z λ Z µε 1 µ cos r − 1 r = cos dr + ln + dr. r η λ r µ λε Since ¯Z ¯ ¯ ¯ µε λε we have Z lim ε→0, η→∞ Moreover ¯Z ¯ ¯ ¯ ε η ¯ cos r − 1 ¯¯ dr¯ ≤ |µ − λ| ε, r cos (λr) − cos (µr) µ dr = ln . r λ ¯ cos (λr) − cos (µr) ¯¯ dr¯ r ε ¯ ¯Z ¯ λη cos r ¯ ¯¯Z µε cos r ¯¯ ¯ µ¯ ¯ ¯ ¯ ¯ dr¯ + ¯¯ dr¯¯ ≤ 2 ¯ln ¯ . ≤¯ ¯ ¯ µη r r λ λε η Hence, we have lim Iε,η (x0 , y) = − ε→0, η→∞ πi 1 sign (x0 · y 0 ) + ln 0 0 2 |x · y | and for any 0 < ε < η < ∞, µ |Iε,η (x0 , y)| ≤ c ln 1 +1 0 |x · y 0 | ¶ for some universal constant c. Since for any |y 0 | = 1, Z 1 ln 0 0 dσ (x0 ) ≤ c, |x ·y | n−1 S we have for some c = c (n) , sup 0<ε<η<∞ ° ° ° ° °K̂ε,η (y)° ∞ ≤ c, and Lebesgue dominated convergence implies Z 1 lim K̂ε,η (y) = Ω (x0 ) I (x0 , y 0 ) dσ (x0 ) n/2 ε→0, n−1 S (2π) η→∞ where I (x0 , y 0 ) = − πi 1 sign (x0 · y 0 ) + ln 0 0 . 2 |x · y | 38 HUIQIANG JIANG Now for any f ∈ L2 , we have n ˆ ˆ 2 T[ ε,η f (y) = (2π) K̂ε,η f (y) → m (y) f (y) in L2 as ε → 0, η → ∞, where Z Ω (x0 ) I (x0 , y 0 ) dσ (x0 ) . m (y) = Sn−1 Hence, Tε,η f converges to some function g ∈ L2 . Now fix ε, it is easy to see Tε,η f → Tε f in L2 . So we conclude g = T f . Hence Tcf (y) = m (y) fˆ (y) . ¤ From the proof, we see Z Ω (x0 ) I (x0 , y 0 ) dσ (x0 ) m (y) = Sn−1 with I (x0 , y 0 ) = − πi 1 sign (x0 · y 0 ) + ln 0 0 . 2 |x · y | of Hilbert transform, we have 1 In particular, for the kernel πx Z i m (y) = − x0 sign (x0 · y 0 ) dσ (x0 ) = −i sign (y) . 2 n−1 S Theorem 40. Suppose Ω satisfies the conditions of the previous theorem. Let f ∈ Lp (Rn ), 1 ≤ p < ∞. (a) limε→0 Tε f (x) exists for a.e. x; (b) Let T ∗ f (x) = sup |Tε f (x)| . ε>0 Then the map f 7→ T ∗ f is of weak type (1, 1). (c) If 1 < p < ∞, then kT ∗ f kp ≤ Ap kf kp . Proof. Check the proof of (b) and (c) in Stein’s book if you are interested. We now prove (a) assuming that (b) and (c) hold. We define for any f ∈ Lp (Rn ), ¯ ¯ ¯ ¯ Λf (x) = ¯¯lim sup Tε f (x) − lim inf Tε f (x)¯¯ . ε→0 ε→0 Then Λf (x) ≤ 2T ∗ f (x). We also observe that if f ∈ D (Rn ), then Λf (x) ≡ 0. If 1 < p < ∞, writing f = f1 + f2 , where f1 ∈ D (Rn ), we have kΛf kp = kΛf2 kp ≤ 2 kT ∗ f2 kp ≤ 2Ap kf2 kp . Since we can make kf2 kp arbitrarily small, we have kΛf kp = 0. Hence, Λf (x) = 0 a.e. in Rn . If p = 1, we have for any t > 0, ¯½ ¾¯ ¯ ∗ t ¯¯ ¯ |{Λf > t}| = |{Λf2 > t}| ≤ ¯ T f2 > 2 ¯ 2c kf2 k1 . t Since we can make kf2 k1 arbitrarily small, we have |{Λf > t}| = 0 for every t > 0, i.e., Λf (x) = 0 a.e. in Rn . ¤ ≤ HARMONIC ANALYSIS 39 5.4. Hilbert transform on a circle and Fourier series. For any f ∈ Lp [−π, π], we can define the Hilbert transform Z cot (x/2) Sf (y) = lim f (x − y) dx. ε→0 ε≤|x|≤π 2π Comparing this with the modified Hilbert transform Z 1 f (x − y) dx. S̃f (y) = lim ε→0 ε≤|x|≤π πx We see that Sf (y) exists for a.e. y ∈ [−π, π] and S is of type (p, p) for any 1 < p < ∞. Lemma 9. For any n ∈ Z, Proof. Since 1 2π Rπ −π ¡ ¢ S einx = −i sign (n) einx . sin[(N + 12 )y ] dy sin y2 = 1 for N = 0, 1, 2, · · · , we have Z ¡ ¡ inx ¢¢ S e y = lim ε→0 = −ie ε≤|x|≤π Z π iny −π π Z = −ieiny −π cos (x/2) in(y−x) e dx 2π sin (x/2) cos (x/2) sin (nx) dx 2π sin (x/2) ¡¡ ¢ ¢ ¡¡ ¢ ¢ sin n + 12 x + sin n − 12 x dx 4π sin (x/2) = −i sign (n) einy . ¤ Hence, if f (x) = then iSf (y) = X P0 f = then Tf = 1 2π Z an einx , an einx − n≥1 We define X X an einx . n≤−1 π f (x) dx = a0 (f ) , −π X 1 (iS + I + P0 ) f = an einx . 2 n≥0 We also define for any integer k, Mk f = eikx f It is easy to verify that sN = M−N T MN − MN +1 T M−(N +1) . Now we apply the singular integral theory to prove the Lp convergence of Fourier series: Theorem 41. Let f ∈ Lp [−π, π], 1 < p < ∞, then sN (f ) → f in Lp [−π, π] as N → ∞. 40 HUIQIANG JIANG Proof. Since S is of type (p, p) and P0 is of type (p, p), we conclude T is of type (p, p). Obviously MN is of type (p, p), sN is of type (p, p). Since sN (f ) → f ¡ hence ¢ in Lp [−π, π] as N → ∞ for f ∈ C 2 S 1 , standard density argument implies that sN (f ) → f in Lp [−π, π] as N → ∞ for every f ∈ Lp [−π, π]. ¤ 6. Riesz transforms, Poisson Integrals and Spherical Harmonics 6.1. The Riesz transforms. Recall that for f ∈ Lp (R), 1 ≤ p < ∞, the Hilbert transform of f is defined by Z 1 f (x − y) Hf (y) = lim+ dy, y ∈ R. y ε→0 π |y|≥ε Since d = −i sign (y) fˆ, Hf Hilbert transform is unitary on L2 (R) and H 2 = −I. For any δ 6= 0, we define the dilation operator τδ (f ) (x) = f (δx). For any δ > 0, Hτδ = τδ H and for any δ < 0, Hτδ = −τδ H. Theorem 42. Suppose T is a bounded operator on L2 (R) which satisfies (a) T commutes with translations; (b) for any δ > 0, T τδ = τδ T and for any δ < 0, T τδ = −τδ T . Then T is a constant multiple of the Hilbert transform. Proof. There exists m ∈ L∞ (R), s.t. Tcf (y) = m (y) fˆ (y) for a.e. y ∈ R. Since −1 Fτδ = |δ| τδ−1 F. We have ¢ −1 −1 ¡ T[ τδ f (y) = m (y) τd τδ−1 fˆ (y) = m (y) |δ| fˆ δ −1 y , δ f (y) = m (y) |δ| and −1 T[ τδ f (y) = sign (δ) τ[ τδ−1 Tcf (y) δ T f (y) = sign (δ) |δ| ¡ ¢ ¡ ¢ ¡ ¢ −1 −1 = sign (δ) |δ| Tcf δ −1 y = sign (δ) |δ| m δ −1 y fˆ δ −1 y . Hence ¡ ¢ m (y) = sign (δ) m δ −1 y which implies that for some constant c, m (y) = −ci sign (y) a.e.. ¤ Now we generalize the Hilbert transform to multi-dimensions. For any f ∈ Lp (Rn ), 1 ≤ p < ∞, the Riesz transform of f is defined by Z yj Rj f (x) = lim cn n+1 f (x − y) dy, j = 1, 2, · · · , n ε→0 |y|≥ε |y| where ¡ ¢ Γ n+1 2 cn = (n+1)/2 . π HARMONIC ANALYSIS Ω (x) 41 x j For the Riesz kernel Kj (x) = |x| and Ωj (x) = cn |x|j , we have n Z m (y) = Ωj (x0 ) I (x0 , y 0 ) dσ (x0 ) Sn−1 · ¸ Z πi 1 = cn x0 − sign (x0 · y 0 ) + ln 0 0 dσ (x0 ) 2 |x · y | Sn−1 Z πi = − cn x0 sign (x0 · y 0 ) dσ (x0 ) . 2 Sn−1 y Since m commutes with rotations, we have m (y) = c |y| . Choose y = en , we have for n ≥ 2, Z πi c = − cn x0n sign (x0n ) dσ (x0 ) 2 Sn−1 Z πi = − cn |x0n | dσ (x0 ) 2 Sn−1 Z 1 ¡ ¢ n−2 dt = −πicn t 1 − t2 2 (n − 1) ωn−1 √ 1 − t2 0 ¡ n+1 ¢ Γ 2 = −πi (n+1)/2 ωn−1 π ¡ n+1 ¢ n−1 Γ 2 π 2 ¡ n+1 ¢ = −πi (n+1)/2 π Γ 2 = −i. Here we used ωn = π n/2 ¡ ¢. Γ 1 + n2 Hence, yj ˆ d R f (y) . j f (y) = −i |y| To understand the interaction of dilations and rotations with the Riesz transform, we consider multipliers which commute with dilations and rotations. Lemma 10. Suppose m : Rn → Rn is homogeneous of degree 0, and for any rotation ρ, m (ρx) = ρm (x) . Then x m (x) = c |x| for some constant c. Proof. Fix any unit vector e, we have m (ρe) = ρm (e) . Hence, ρe = e implies ρm (e) = m (e), so there exists constant c (e) s.t. m (e) = c (e) e. Now let e1 , e2 be two unit vectors, and ρ be a rotation such that ρe1 = e2 , then m (ρe1 ) = ρm (e1 ) = c (e1 ) e2 , m (ρe1 ) = m (e2 ) = c (e2 ) e2 , 42 HUIQIANG JIANG hence c (e1 ) = c (e2 ). Hence m (e) = ce. Since m is homogeneous of degree 0, m (x) = c x |x| for some constant c. ¤ To translate the above lemma in terms of Riesz transforms, we consider the interaction of dilations and rotations with Fourier transform. For any δ > 0, we have Fτδ = δ −n τδ−1 F. And for any rotation about the origin ρ, we define ρf (x) = f (ρx), then Z 1 (Fρf ) (y) = e−ix·y f (ρx) dx n/2 Rn (2π) Z −1 1 e−iρ x·y f (x) dx = n/2 n R (2π) Z 1 = e−ix·ρy f (x) dx n/2 Rn (2π) = (ρFf ) (y) . Proposition 8. Let T = (T1 , T2 , · · · , Tn ) be an n-tuple of bounded transformations on L2 (Rn ). Suppose (a) Each Tj commutes with the translation in Rn ; (b) Each Tj commutes with the dilations in Rn ; (c) For every rotation ρ of Rn , ρTj ρ−1 f = n X ρjk Tk f. k=1 Then there exists constant c such that Tj = cRj , j = 1, 2, · · · , n. We now consider a function u ∈ C 2 (Rn ) with compact support. We have u[ xi xj = −yi yj û = − yi yj 2 |y| û = −R\ i Rj ∆u. |y| |y| Hence, ∂2u = −Ri Rj ∆u. ∂xi ∂xj i.e., once we know ∆u, we know all second order partial derivatives of u. Theorem 43. Suppose u ∈ C 2 (Rn ) is compact supported. Then for any 1 < p < ∞, we have ° 2 ° ° ∂ u ° ° ° ° ∂xi ∂xj ° ≤ Ap kukp p where Ap is a constant depending only on n and p. HARMONIC ANALYSIS 43 6.2. Poisson Integral. We write Rn+1 = {(x, t) : x ∈ Rn , t > 0}. Given a func+ n tion f defined on R , we want to find harmonic function u on Rn+1 such that u = f + n n ˜ on R . We use ∆ to denote the Laplacian on R , and ∆ the Laplacian on Rn+1 . Then ˜ = ∆u + utt = 0. ∆u We write û = û (y, t) the Fourier transform of u in x-variable, then 2 − |y| û + ûtt = 0. We also impose boundary condition û (y, 0) = fˆ and û (y, ∞) = 0. Then formally û = e−|y|t fˆ. Hence, u (x, t) = Z 1 (2π) eix·y e−|y|t fˆ (y) dy. n/2 Rn If f ∈ L2 (Rn ), u defined above is a smooth harmonic function in Rn+1 and + u (·, t) → f in L2 (Rn ) norm as t → 0+ . Hence, indeed u is a solution to the Dirichlet problem in Rn+1 + . Now we write µZ ¶ Z 1 ix·y −|y|t −iyz u (x, t) = e e e f (z) dz dy n (2π) Rn Rn Z = P (x − z, t) f (z) dz Rn where P (x, t) = 1 n (2π) Z eix·y e−|y|t dy Rn is called the Poisson kernel in Rn+1 + . Lemma 11. For any γ > 0, e −γ 1 =√ π Z 0 ∞ e−u − γ 2 √ e 4u du. u Proof. e −γ 1 = π Z ∞ −∞ eiγs 1 ds = 2 1+s π Z µZ ∞ e −∞ iγs ∞ ¶ −(1+s2 )u e du ds. 0 ¤ Proposition 9. P (x, t) = ³ cn t 2 |x| + t2 with cn = ´ n+1 2 Γ ¡ n+1 ¢ 2 π n+1 2 . 44 HUIQIANG JIANG Proof. Z 1 eix·y e−|y|t dy n (2π) Rn µ ¶ Z Z ∞ −u |y|2 t2 1 e 1 ix·y − 4u √ √ e e du dy = n π 0 u (2π) Rn ¶ Z ∞ −u µZ |y|2 t2 1 1 e ix·y − 4u √ √ = e e dy du. n π 0 u (2π) Rn P (x, t) = Now Z eix·y e− Rn Z = e dy ¯ ¯2 √ ¯ ¯ √ |x|2 u −¯ 2ty −i tu x¯ − t2 u Rn Hence |y|2 t2 4u = e− |x|2 u t2 = e− |x|2 u t2 = e− |x|2 u t2 Z dy ¯ ¯ ¯ √ ¯2 −¯ 2ty ¯ u e dy Rn µ √ ¶n µZ ∞ ¶n 2 u −s2 e ds t −∞ µ √ ¶n 2 πu . t √ n Z ∞ −u |x|2 u 1 (2 π) e √ √ t−n un/2 e− t2 du n π u (2π) 0 Z ∞ n−1 t 1 = n+1 ³ e−s s 2 ds n+1 ´ 2 π 2 0 2 |x| + t2 ¡ ¢ Γ n+1 t 2 = . n+1 ³ ´ n+1 2 π 2 2 2 |x| + t P (x, t) = ¤ Proposition 10. The Poisson kernel P (x, t) satisfies: (i) P > 0 for any x ∈ Rn , t > 0; (ii) for any t > 0, Z P (x, t) dx = 1. Rn (iii) For any δ > 0, Z lim+ t→0 P (x, t) dx = 0. |x|≥δ Proof. Direct verification. ¤ Theorem 44. Suppose f ∈ Lp (Rn ), 1 ≤ p ≤ ∞, and let u (x, t) = P (·, t) ∗ f. (i) u is a smooth harmonic function in Rn+1 + ; (ii) for each x ∈ Rn , sup |u (x, t)| ≤ M f (x) ; t>0 HARMONIC ANALYSIS 45 (iii) If p < ∞, u (·, t) → f in Lp (Rn ) as t → 0+ ; (iv) For a.e. x ∈ Rn , lim+ u (x, t) = f (x) ; t→0 (v) If p = ∞ and f is continuous in Rn , then u (x, t) → f (x) uniformly on any compact subset of Rn as t → 0+ . Proof. (i) Trivial. (ii) Assume x = 0 and M f (0) < ∞. Define Z g (r) = |f (x)| dx ≤ ωn rn M f (0) , Br (0) then ¯Z ¯ Z ¯ ¯ ¯ P (−x, t) |f (x)| dx P (−x, t) f (x) dx¯¯ ≤ |u (0, t)| = ¯ Rn Rn µZ ¶ Z ∞ = P (r, t) |f (x)| dσ (x) dr 0 ∂Br Z ∞ Z ∞ ∞ 0 = P (r, t) g (r) dr = g (r) P (r, t)|0 − Pr (r, t) g (r) dr 0 Z ∞ Z ∞ 0 =− Pr (r, t) g (r) dr ≤ −M f (0) Pr (r, t) ωn rn dr 0 0 Z ∞ ∞ = − M f (0) ωn rn P (r, t)|0 + M f (0) P (r, t) nωn rn−1 dr 0 = M f (0) . Here we used the fact that P is radial and monotone decreasing in |x|. (iii) It follows from the fact that P (·.t) is an approximate delta function as t → 0+ . (iv) First assume 1 ≤ p < ∞. Define for any x, ¯ ¯ ¯ ¯ Ωf (x) = ¯¯lim sup u (x, t) − lim inf u (x, t)¯¯ . t→0+ t→0+ Then |Ωf (x)| ≤ 2M f (x) and hence for any δ > 0, ¯½ ¾¯ ¯ δ ¯¯ ¯ |{x : Ωf (x) > δ}| ≤ ¯ x : Ωf (x) > ≤ Ap kf kp . 2 ¯ It is easy to verify that Ωf (x) = 0 if f is continuous and compact supported. For any f ∈ Lp , writing f = f1 + f2 such that f1 is continuous and compact supported, and kf2 kp small. Then we have |{x : Ωf (x) > δ}| = |{x : Ωf2 (x) > δ}| ≤ Ap kf2 kp which can be made arbitrarily small. Hence |{x : Ωf (x) > δ}| = 0. 46 HUIQIANG JIANG i.e., limt→0+ u (x, t) exists a.e. x ∈ Rn . Because of (iii) lim u (x, t) = f (x) t→0+ a.e.. Now for f ∈ L∞ (Rn ), if suffices to show that for any R > 0, lim u (x, t) = f (x) t→0+ a.e. x ∈ BR (0) ⊂ Rn . We write f = f1 + f2 where f1 = f χB2R , then lim f2 ∗ P = 0 t→0 uniformly on BR (0). and the result follows from the L1 case for f1 . (v) Similar. ¤ 6.3. Spherical harmonics. Let Pk , k ≥ 0 be the set of all homogeneous polynomials in Rn of order k. Hence p ∈ Pk iff X p (x) = cα xα . |α|=k Pk is a vector space of dimension µ ¶ n+k−1 (n + k − 1)! dk = = . n−1 (n − 1)!k! We introduce an inner product on Pk by hp, qi = p (D) q̄. Definition 12. A solid spherical harmonic of degree k is a homogeneous harmonic polynomial in Rn of order k. A (surface) spherical harmonic of degree k is the restriction on Sn−1 of solid spherical harmonic of degree k in Rn . We use Hk to denote all spherical harmonics of degree k and we will not distinguish solid spherical harmonic and surface spherical harmonic. Lemma 12. For k ≥ 2, the linear map ϕ : Pk → Pk−2 defined by ϕ (p) = ∆p is onto. Proof. If not, there exists 0 6= q ∈ Pk−2 , q⊥∆p for any p ∈ Pk , hence D E 2 0 = hq, ∆pi = |x| q, p 2 which is a contradiction if we take p = |x| q. ¤ Remark 17. As a direct consequence of the lemmas, for k ≥ 2, the dimension of Hk is µ ¶ µ ¶ n+k−1 n+k−3 dk − dk−2 = − . n−1 n−1 Moreover, the dimension of H0 is 1 and the dimension of H1 is n. When n = 3, the dimension of Hk is 2k + 1 for every k ≥ 0. HARMONIC ANALYSIS 47 Remark 18. Let ∆S denote the Laplace-Beltrami operator on Sn−1 , i.e., the spherical Laplacian. Then we have for any f defined on Sn−1 , ∆S f = ∆g ³ where g = f x |x| ´ . In polar coordinates, we have for any f defined on Rn , ∆f = frr + 1 n−1 fr + 2 ∆S f. r r Hence, if p ∈ Hk , we have ∆S p = −k (k + n − 2) p. i.e., p is an eigenfunction of the spherical Laplacian. Theorem 45. If p ∈ Pk , then 2 2l p (x) = p0 (x) + |x| p1 (x) + · · · + |x| pl (x) where l = £k¤ 2 and pj (x) ∈ Hk−2j . Proof. Since any polynomial of order ≤ 1 is harmonic, we can assume k ≥ 2. We 2 claim Pk is the direct sum of |x| Pk−2 and Hk . For any q ∈ Pk s.t. D E 2 0 = |x| p, q holds for any p ∈ Pk−2 , we have 0 = hp, ∆qi , hence ∆q = 0 and q ∈ Hk . The theorem then follows from induction. ¤ Corollary 7. The restriction of any polynomial on Sn−1 is a sum of spherical harmonics. Corollary 8.¡ The ¢collection¡ of finite ¢ linear combinations of spherical harmonics is dense in C Sn−1 and L2 Sn−1 . ¡ ¢ ¡ ¢ Proof. Density in C Sn−1¡ follows from Weierstrass theorem. Density in L2 Sn−1 ¢ follows from density in C Sn−1 . ¤ ¢ ¡ The inner product in L2 Sn−1 is defined by Z (f, g) = f (x) ḡ (x) dσ (x) . Sn−1 ∞ Theorem ¡ ¢ 46. The finite dimensional spaces {Hk }k=0 are mutually orthogonal in L2 Sn−1 . Hence ∞ ¡ ¢ X Hk . L2 Sn−1 = k=0 48 HUIQIANG JIANG Proof. If k 6= j and p ∈ Hk , q ∈ Hj , then Z (k − j) (p, q) = (k − j) p (x) q̄ (x) dσ (x) Sn−1 Z ∂p ∂ q̄ = q̄ − pdσ (x) ∂ν n−1 ∂ν ZS = ∆pq̄ − ∆q̄pdσ (x) = 0. |x|≤1 Hence (p, q) = 0. ¤ When n = 2, the above decomposition gives rise to the Fourier series on S1 . ¡ ¢ Proposition 11. Suppose f ∈ L2 Sn−1 and f= ∞ X pk k=0 where pk ∈ Hk . Then f is smooth iff for any N > 0, Z ¡ ¢ 2 |pk | = O k −N S n−1 as k → ∞. Proof. If f is smooth,for any positive integer r, Z Z Z r 2 r r |pk | , pk ∆s f = ∆s pk f = [−k (k + n − 2)] hence for k large, Z 2 |pk | ≤ 1 r |k (k + n − 2)| So we have µZ 2 ¶ 21 µZ |pk | Z 2 |pk | ≤ |∆rs f | 2 ¶ 12 ≤ c ³R |pk | k 2r 2 ´ 12 . c . k 4r On the other hand, if for any N > 0, Z ¡ ¢ 2 |pk | = O k −N S n−1 as k → ∞, we can show that Z 2 |∆rs f | ≤ c. ¢ ¡ Hence, f ∈ H 2r Sn−1 . Sobolev emdedding implies f is smooth. ¤ Theorem 47 (Identity of Hecke). Suppose p is a solid spherical harmonic of degree k in Rn , then 2 −|x|2 /2 = (−i)k p (y) e−|y| /2 . pe\ HARMONIC ANALYSIS Proof. We have −|x|2 /2 (y) = pe\ = = Z 1 (2π) 1 (2π) 1 (2π) 49 n/2 p (x) e−|x| Rn 2 e−|y| n/2 2 /2 e dx p (x) e−|x+iy| Z 2 /2 −ix·y Z e−|y| n/2 = q (y) e−|y| 2 2 /2 dx Rn p (x − iy) e−|x| /2 2 /2 dx Rn /2 where q is a polynomial. Replacing y by iy, we also have Z 2 1 p (x + y) e−|x| /2 dx = p (y) q (iy) = n/2 n R (2π) since p is harmonic. Hence k q (y) = p (−iy) = (−i) p (y) . ¤ 6.4. Higher Riesz Transforms. We consider for k ≥ 1 the singular integral with p(x) the kernel |x| n+k where p ∈ Hk . When k = 1, we recover Riesz transform. Such singular integrals are hence called higher Riesz transforms. The main theorem we want to establish is: Theorem 48. Let p ∈ Hk , k ≥ 1. Then the multiplier corresponding to the transform Z p (x) T f (y) = lim f (y − x) dx ε→0 |x|≥ε |x|n+k is p (x) γk k |x| where ¡ ¢ n π 2 Γ k2 ¢. ¡ γk = (−i) Γ k+n 2 k Remark 19. When k = 1, we have ¡ ¢ n π 2 Γ 12 π (n+1)/2 γ1 = −i ¡ 1+n ¢ = −i ¡ n+1 ¢ Γ 2 Γ 2 which recovers the multiplier for Riesz transform. We first prove several lemmas. Lemma 13. Let p ∈ Hk , k ≥ 0. Then 2 |y| n −δ|x|2 /2 = δ −k− 2 (−i)k p (y) e− 2δ . pe\ Proof. Simple scaling. ¤ 50 HUIQIANG JIANG Lemma 14. Let p ∈ Hk , k ≥ 0. For any α ∈ (0, n) and for any ϕ ∈ S, we have Z Z p (x) p (x) ϕ̂ (x) dx = γ ϕ (x) dx k,α n+k−α k+α Rn |x| Rn |x| where k α− n 2 γk,α = (−i) 2 ¡ ¢ Γ k+α 2 ¢. ¡ Γ k+n−α 2 Proof. For any δ > 0, we have Z Z 2 n k p (x) e−δ|x| /2 ϕ̂ (x) dx = δ −k− 2 (−i) Rn p (x) e− |x|2 2δ ϕ (x) dx. Rn Multiplying both sides with δ β−1 with k+n−α , 2 β= and integrate δ from 0 to ∞, we have ¶ Z µZ ∞ Z 2 k δ β−1 e−δ|x| /2 dδ p (x) ϕ̂ (x) dx = (−i) Rn Rn 0 Now Z ∞ Ã δ µZ β−1 −δ|x|2 /2 e dδ = 0 and |x| 2 Z 0 2 à =2 where we used u = Hence, Z p (x) Rn |x|2 2δ n+k−α |x| ∞ δ− k+α 2 −1 |x| 2 2 k+α 2 e− |x|2 2δ e− |x|2 2δ ¶ dδ p (x) ϕ (x) dx. k+n−α 2 Γ k+n−α 2 ¶ 1 |x| k+n−α dδ ∞ k+α 2 −1 u µ k+α 2 −1 µ uβ−1 e−u dδ = 2 !− k+α Z 2 Γ δ− 0 0 ∞ = !−β Z ∞ e−u du 0 k+α 2 ¶ 1 |x| k+α 2 and dδ = − |x| 2u2 du. ϕ̂ (x) dx = = 2 k+α 2 Γ ¡ k+α ¢ Z p (x) (−i) k+n−α ¡ 2 ϕ (x) dx ¢ k+α k+n−α 2 2 Γ Rn |x| 2 ¡ ¢ Z Γ k+α p (x) k α− n 2 2 ¡ k+n−α ¢ (−i) 2 ϕ (x) dx. k+α Γ Rn |x| 2 k ¤ Lemma 15. Let p ∈ Hk , k ≥ 1. For any α ∈ (0, n) and for any ϕ ∈ S, we have Z Z p (x) p (x) lim+ ϕ̂ (x) dx = lim ϕ̂ (x) dx. n+k−α n+k + α→0 ε→0 Rn |x| |x|≥ε |x| , HARMONIC ANALYSIS Proof. Z p (x) Rn Z |x| n+k−α ϕ̂ (x) dx Z p (x) = Z = Hence, ϕ̂ (x) dx + n+k−α [ϕ̂ (x) − ϕ̂ (0)] dx + |x|≥1 |x| |x|≤1 Z Z n+k−α |x| p (x) Rn = |x| Z |x|≤1 = lim+ ε→0 ϕ̂ (x) dx n+k−α Z p (x) |x| n+k−α ϕ̂ (x) dx. ϕ̂ (x) dx Z n+k p (x) [ϕ̂ (x) − ϕ̂ (0)] dx + |x| |x|≥1 p (x) n+k |x| |x|≥ε |x| |x|≥1 p (x) lim α→0+ p (x) n+k−α |x| p (x) |x|≤1 51 n+k ϕ̂ (x) dx ϕ̂ (x) dx. ¤ Now we can prove the main theorem. Proof of Theorem 48. Let f ∈ S. Then for any fixed x, there exists ϕ such that f (x − y) = ϕ̂ (y) and we have Z 1 ϕ (y) = f (x − z) eiy·z dz n/2 (2π) 1 = Rn Z n/2 Rn (2π) = fˆ (y) eix·y . Hence Z lim+ ε→0 p (x) |x| |x|≥ε Z Rn = lim+ γk,α α→0 Z = γk,0 |x| Z n+k−α Rn |x| k ϕ̂ (x) dx p (x) p (x) Rn ϕ̂ (x) dx p (x) = lim α→0+ n+k f (u) eiy·(x−u) du |x| k+α ϕ (x) dx Z ϕ (x) dx = γk,0 where k γk,0 = (−i) 2 −n 2 So we have Z T f (y) = lim+ ε→0 |x|≥ε p (x) |x| Rn p (x) k |x| eix·y fˆ (x) dx ¡ ¢ Γ k2 ¡ ¢. Γ k+n 2 f (y − x) dx = γk,0 n+k Z Rn p (x) k |x| eix·y fˆ (x) dx. 52 HUIQIANG JIANG Hence, p (x) Tcf (x) = (2π) 2 γk,0 k fˆ (x) = m (x) fˆ (x) |x| n where n m (x) = (2π) 2 γk,0 p (x) = γk k |x| Since S is dense in L2 , the theorem is proved. p (x) |x| k . ¤ Theorem 49. Let c be a constant and Ω be a homogeneous function on Rn of degree 0, which is smooth on Sn−1 and Z Ωdσ = 0. Sn−1 Define Z T f (x) = cf (x) + lim ε→0 |y|≥ε Ω (y) n f (x − y) dy. |y| Then there exists a homogeneous function m of degree 0 which is smooth on Sn−1 such that Tcf = mfˆ. Moreover, the reverse is also true. Proof. Write Ω (x) = ∞ X pk (x) , k=1 then on Sn−1 m (x) = c + ∞ X γk pk (x) k=1 n−1 which is smooth whenever Ω is smooth on S . ¤ 7. Differentiability Properties in Terms of Function Spaces 7.1. Riesz Potential. Let f ∈ S. Recall 2 (−∆f ) ˆ (x) = |x| fˆ. We can define for any β > −n, ³ ´ β β (−∆) 2 f ˆ (x) = |x| fˆ. β β Since |x| fˆ ∈ L1 (Rn ) for any β > −n, (−∆) 2 f is well defined as inverse Fourier transform of integrable function. For any α ∈ (0, n), we define the Riesz potential of f ∈ S by −α 2 (Iα f ) (x) = (−∆) f. Hence, −α ˆ Id f α f (x) = |x| and (Iα f ) (x) = ³ 1 (2π) n/2 |x| Apparently, Iα f is well defined as long as |x| −α −α ´∨ ∗ f. fˆ ∈ L1 . HARMONIC ANALYSIS 53 Proposition 12. For any f ∈ S. (a) Iα (Iβ f ) = Iα+β f if α > 0, β > 0 and α + β < n. (b) If n ≥ 3, then ∆ (Iα f ) = Iα (∆f ) = −Iα−2 f holds for any α ∈ (2, n). And when α = 2, we have ∆ (I2 f ) = I2 (∆f ) = −f. Proof. (a) Since α + β < n −α−β ˆ −α \ I\ f = |x| Id α+β f = |x| β f = Iα (Iβ f ). (b) For any α ∈ (2, n), 2 2−α b ∆\ (Iα f ) = − |x| Id f = −I\ α f = − |x| α−2 f , Iα\ (∆f ) = |x| −α c = − |x|2−α fb = −I\ ∆f α−2 f . When α = 2, ∆ (I2 f ) = I2 (∆f ) = −f follows similarly. ¤ ³ −α Now we calculate |x| ´∨ : Lemma 16. Let α ∈ (0, n). Then for any ϕ ∈ S, Z Z −n+α −α |x| ϕ̂ (x) dx = cα |x| ϕ (x) dx Rn Rn where γ0,α = 2 Hence, ³ 1 n/2 (2π) where −α |x| ¡ ¢ Γ α2 ¢. ¡ Γ n−α 2 α− n 2 ´∨ = 1 −n+α |x| γα ¡ ¢ 2α π n/2 Γ α2 ¡ ¢ . γα = Γ n−α 2 Proof. Take k = 0 in Lemma 14, we have Z Z −n+α |x| ϕ̂ (x) dx = γ0,α Rn where γ0,α = 2 Hence, γ0,α Z |x| Rn α− n 2 ϕ (x) dx ¡ ¢ Γ α2 ¡ ¢. Γ n−α 2 Z −α −α |x| Rn Z ϕ (x) dx = |x| −n+α ϕ̂ (x) dx = Rn i.e., −n+α −α \ |x| = γ0,α |x| Rn −n+α \ |x| (y) ϕ (y) dy, 54 HUIQIANG JIANG and ³ 1 |x| n/2 (2π) −α ´∨ = 1 n/2 γ0,α (2π) |x| −n+α . ¤ Corollary 9. For any α ∈ (0, n) and f ∈ S, Z 1 1 (7.1) (Iα f ) (x) = f (y) dy γα Rn |x − y|n−α where ¡ ¢ 2α π n/2 Γ α2 ¡ ¢ . γα = Γ n−α 2 We now ask for what pairs p and q, we have kIα f kq ≤ A kf kp . A necessary condition follows from a scaling argument. Let (p, q) be such a pair. Since for any δ > 0, Iα (τδ f ) = δ −α τδ Iα f and kτδ f kp = δ −n/p kf kp , we have kIα (τδ f )kq = δ −α kτδ Iα f kq n n =δ −α− q kIα f kq ≤ δ −α− q A kf kp n n =δ −α− q + p A kτδ f kp . Hence, we must have 1 1 α = − . q p n p n Theorem 50. Let 0 < α < n, 1 ≤ p < q < ∞, 1q = p1 − α n and f ∈ L (R ). n (a) Iα f defined in (7.1) converges absolutely for a.e. x ∈ R . (b) If p > 1, then kIα f kq ≤ Ap,q kf kp where Ap,q is a constant independent of f . (c) If p = 1, then for any t > 0, µ |{x : Iα f (x) > t}| ≤ A kf k1 t ¶q where A is a constant independent of f . Proof. (a) Fix µ > 0. Let K = 1 K1 ∈ L and K2 ∈ L 0 p 1 , γα |x|n−α where Moreover, Z kK1 k1 = |x|≤µ K1 = Kχ{|x|≤µ} and K2 = K − K1 . Then 1 1 + = 1. 0 p p 1 γα |x| n−α dx = nωn α µ ≡ c1 µα , αγα HARMONIC ANALYSIS and ÃZ kK2 kp0 = µ 1 ¶p0 55 ! p10 dx n−α γα |x| µ ¶1 n 1 ωn q p0 − nq = µ ≡ c2 µ− q . 0 γα p p p Hence for any f ∈ L , K1 ∗ f ∈ L , K2 ∗ f ∈ L∞ and |x|>µ Iα f = K ∗ f = K1 ∗ f + K2 ∗ f. converges absolutely a.e. in Rn . (b) (c) Assume kf kp = 1. For each µ > 0, kK1 ∗ f kp ≤ kK1 k1 kf kp = c1 µα and n kK2 ∗ f k∞ ≤ kK2 kp0 kf kp = c2 µ− q . For any t > 0, we have |{x : |Iα f (x)| > t}| ¯½ ¾¯ ¯½ ¾¯ ¯ t ¯¯ ¯¯ t ¯¯ ¯ ≤ ¯ x : |(K1 ∗ f ) (x)| > + x : |(K2 ∗ f ) (x)| > 2 ¯ ¯ 2 ¯ p p kK1 ∗ f kp c 2p ≤ 1 p µαp ≤ p t (t/2) if we choose Hence, n t = c2 µ− q . 2 1 . tq i.e., Iα is of weak type (p, q). In particular, (c) holds. Now the following Marcinkiewicz interpolation theorem implies (b). ¤ |{x : |Iα f (x)| > t}| ≤ c Theorem 51 (Marcinkiewicz). Assume 1 ≤ pi ≤ qi ≤ ∞, i = 1, 2 and p1 < p2 , q1 6= q2 . Let T be a subadditive transformation defined on Lp1 (Rn ) + Lp2 (Rn ). Suppose that T is of weak type (pi , qi ), i = 1, 2. Then for any θ ∈ (0, 1) and 1 (1 − θ) θ 1 (1 − θ) θ = + , = + , p p1 p2 q q1 q2 T is of type (p, q). 7.2. The Sobolev spaces. For any nonnegative integer k, the Sobolev space W k,p (Rn ) = Lpk (Rn ) is defined as the space of functions f with Dα f ∈ Lp (Rn ) whenever |α| ≤ k. W k,p (Rn ) is a Banach space with the norm p1 X p kf kW k,p (Rn ) = kDα f kp . |α|≤k An equivalent norm can be given by X |α|≤k kDα f kp . 56 HUIQIANG JIANG When p = 2, H k (Rn ) = W k,p (Rn ) is a Hilbert space with inner product X Z (f, g) = Dα f Dα gdx. |α|≤k Rn Proposition 13. Let 1 ≤ p < ∞. Then f ∈ W k,p (Rn ) iff there exists a sequence {fm } ⊂ D (Rn ) s.t. lim kf − fm kp = 0 m→∞ and for each |α| ≤ k, {Dα fm } converges in Lp norm. In particular, D (Rn ) is dense in W k,p (Rn ). Proof. Let {ϕε } be the standard mollifier. If f ∈ W k,p (Rn ), then gε = ϕε ∗ f ∈ C ∞ (Rn ) and lim kf − gε kp = 0. ε→∞ Since Dα gε = ϕε ∗ Dα f , for each |α| ≤ k lim kDα f − Dα gε kp = 0. ε→∞ Let η ∈ D (Rn ) be a smooth cutoff function s.t η (x) = 1 on B1 (0). For each g ∈ C ∞ (Rn ) ∩ W k,p (Rn ), define hδ (x) = g (x) η (δx), then lim khδ − gkp = 0. ε→∞ The existence of {fm } follows from a Diagonal argument. The reverse is trivial. ¤ Lemma 17. Suppose f ∈ D (Rn ), then n Z 1 X ∂f yj f (x) = (x − y) n dy. nωn j=1 Rn ∂xj |y| Proof. For any ξ ∈ Sn−1 , we have Z f (x) = ∞ ξ · ∇f (x − ξt) dt. 0 Integrating over ξ ∈ Sn−1 , we have Z Z ∞ 1 f (x) = ξ · ∇f (x − ξt) dtdσ (ξ) nωn Sn−1 0 n Z yj 1 X ∂f (x − y) n dy. = nωn j=1 Rn ∂xj |y| ¤ Lemma 18. Suppose f ∈ W 1,1 (Rn ), then 1 ° n ° n ° n ° Y X ° ∂f ° ° ∂f ° 1 ° ° ° ° ≤ kf k n ≤ ° ∂xj ° ° ∂xj ° . n−1 n 1 1 j=1 j=1 HARMONIC ANALYSIS 57 Proof. We assume f ∈ D (Rn ). If n = 1, we have ¯Z x ¯ ¯ ¯ |f (x)| = ¯¯ f 0 (y) dy ¯¯ ≤ kf 0 k1 , −∞ hence kf k∞ ≤ kf 0 k1 . Suppose the inequality holds for n = k, then for n = k + 1, we write x = (x0 , xn ), and Z n n n−1 kf k n = |f (x)| n−1 dx n−1 Rn ! 1 ¯ µZ ¯ ¶ n−1 Z ÃZ ¯ ∂f ¯ ¯ ¯ dxn ≤ |f (x)| dx0 dxn ¯ ¯ R Rn−1 R ∂xn ! 1 ¯ ¶ n−1 µZ ¶ n−2 Z õZ Z ¯ n−1 ¯ ∂f ¯ n−1 0 0 ¯ dxn dx ¯ |f (x)| n−2 dx ≤ dxn ¯ ¯ R Rn−1 R ∂xn Rn−1 1 n−1 ° ° 1 Z ° n−1 Y° ° ∂f ° n−1 ° ∂f ° ° ° ° ≤° dxn ° ∂xn ° ° ∂xj ° 1 n−1 R 1 L (R ) j=1 ° ° 1 YZ ° ∂f ° n−1 n−1 ° =° ° ∂xn ° 1 j=1 R ° ° ° ∂f ° ° ° ° ∂xj ° 1 n−1 dxn L1 (Rn−1 ) 1 ° n−1 n ° Y ° ∂f ° ° ° = . ° ∂xj ° j=1 1 Since D (Rn ) is dense in W 1,1 (Rn ), standard density argument implies that the inequality holds for f ∈ W 1,1 (Rn ). ¤ The following theorem is essential for Sobolev spaces. Theorem 52. Suppose k is a positive integer and 1q = p1 − nk . (a) If 1 ≤ p < nk , then W k,p (Rn ) ⊂ Lq (Rn ) and the natural inclusion is continuous. (b) If p = nk ≥ 1, then W k,p (Rn ) ⊂ Lrloc (Rn ) for any 1 ≤ r < ∞. (c) If p > nk and p ≥ 1, then W k,p (Rn ) ⊂ C (Rn ). Proof. It suffices to prove the Theorem when k = 1 and the general case follows from an induction. (a) Let f ∈ D (Rn ), then µ ¶ n \ X ∂f xj d ∂f Rj = −i = |x| fˆ. ∂x |x| ∂x j j j=1 Hence, f = I1 n X j=1 µ Rj ¶ ∂f ∂xj 58 HUIQIANG JIANG if n ≥ 2. If 1 < p, q < ∞, then n ≥ 2, so part (i) when p > 1 follows from the properties of Riesz transform and Riesz potential. When p = 1, part (i) becomes 1 ° n n ° Y ° ∂f ° ° ° kf k n ≤ ° ∂xj ° n−1 1 j=1 which is the lemma above. (b) It follows from (a). (c) Assume f ∈ D (Rn ) with support in BR (0), then for any x ∈ BR (0) ¯ ¯ ¯ ¯ n Z ¯ 1 X ¯ y ∂f j ¯ (x − y) n dy ¯¯ |f (x)| = ¯ |y| ¯ nωn j=1 Rn ∂xj ¯ ¯ ¯ ¯ ¯ n Z ¯ 1 X ¯ y ∂f j ¯ =¯ (x − y) n dy ¯¯ |y| ¯ nωn j=1 |x−y|≤R ∂xj ¯ ° ° ° ° n ° ° ° 1 X° ° ∂f ° ° yjn ° ≤ ° ° ° nωn j=1 ∂xj p |y| °Lp0 (B2R (0)) ° n ° X ° ∂f ° ° ° ≤c ° ∂xj ° p j=1 since ° ° 0 ° yj °p ° n° ° |y| ° p0 Z 1 ≤ |y| B2R (0) L (B2R (0)) Z 2R = nωn (n−1)p0 dy 0 r−(n−1)p +n−1 dr 0 p−n nωn (p − 1) = (2R) p−1 , p−n here we used (n − 1) p p−n +n= > 0. p−1 p−1 Density argument yields the conclusion for f ∈ W 1,p (Rn ). − (n − 1) p0 + n = − 7.3. Bessel Potential and Bessel Spaces. Bessel potential Jα is defined by −α 2 Jα = (I − ∆) Hence, for any f ∈ S, . ³ ´− α2 2 Jd f = 1 + |x| fˆ α and Jα f = Gα ∗ f where Gα (x) = ·³ 1 n 1 + |x| 2 ´− α2 ¸∨ (2π) 2 Z ³ ´− α2 1 2 = 1 + |y| eix·y dy. n (2π) Rn ¤ HARMONIC ANALYSIS 59 ³ ´− α2 2 If α > n, then 1 + |x| ∈ L1 , and hence Gα (x) = Z 1 n (2π) ³ ´− α2 2 1 + |y| eix·y dy. Rn Applying the identity ³ 1 + |x| 2 ´− α2 = Γ Z 1 ¡α¢ ∞ 2 α e−t(1+|x| ) t 2 −1 dt, 0 2 we have 1 Gα (x) = n ¡α¢ (2π) Γ 2 = 1 n ¡α¢ (2π) Γ 2 n 2 µZ Z ∞ e Rn Z ∞ −t(1+|y|2 ) 0 ¶ dt eix·y dy µZ α e−t t 2 −1 e−t|y| 2 +ix·y ¶ dy dt Rn 0 Z t α 2 −1 ∞ α−n π t 2 −1 dt e n ¡ ¢ (2π) Γ α2 0 Z ∞ |x|2 α−n 1 = e−t− 4t t 2 −1 dt. ¡α¢ n (4π) 2 Γ 2 0 = |x|2 −t− 4t Now we claim that the above expression holds whenever α > 0. Proposition 14. For any α > 0, define Gα (x) = Z 1 n 2 (4π) Γ ¡α¢ ∞ e−t− |x|2 4t t α−n 2 −1 dt. 0 2 Then Gα ∈ L1 (Rn ) and cα (y) = G 1 (2π) n 2 ³ ´− α2 2 1 + |y| . Proof. Since Z e− |x|2 4t Z Rn n n 2 e−|u| (4t) 2 du = (4πt) 2 , dx = Rn we have Z Rn Gα (x) dx = 1 n 2 (4π) Γ 1 Z ¡α¢ 2 ∞ µZ ∞ µZ −t− e ¡ ¢ n (4π) 2 Γ α2 0 Rn Z ∞ α 1 = ¡α¢ e−t t 2 −1 dt = 1. Γ 2 0 = |x|2 4t Rn 0 Z e−t− |x|2 4t ¶ α−n dx t 2 −1 dt ¶ α−n dx t 2 −1 dt 60 HUIQIANG JIANG Now, Z n cα (y) = (2π) 2 G = = = = Gα (x) eix·y dx ¶ Z ∞ µZ |x|2 α−n 1 e− 4t eix·y dx e−t t 2 −1 dt ¡α¢ n (4π) 2 Γ 2 0 Rn ¶ Z ∞ µZ α−n 1 1 |x−2ity|2 −t|y|2 − 4t dx e−t t 2 −1 dt e ¡α¢ n 2 (4π) Γ 2 0 Rn Z ∞ n 2 α−n 1 (4πt) 2 e−t(1+|y| ) t 2 −1 dt ¡α¢ n (4π) 2 Γ 2 0 Z ∞ ³ ´− α2 2 α 1 2 ¡α¢ e−t(1+|y| ) t 2 −1 dt = 1 + |y| . Γ 2 0 Rn ¤ Proposition 15. (a) If 0 < α < n, ³ ´ |x| −n+α + o |x| as |x| → 0. γ (α) −n+α Gα (x) = (b) For any α > 0, ³ 1 ´ Gα (x) = O e− 2 |x| as |x| → ∞. Proof. (a). Let 2 2 |x| |x| , dt = − 2 du. 4t 4u u= Hence, Z 1 n 2 (4π) Γ = ¡α¢ 2 1 n 2 ∞ e− |x|2 4t t α−n 2 −1 dt 0 Z ¡α¢ ∞ à −u e 2 |x| 4u ! α−n 2 −1 (4π) Γ 2 0 ¯ x ¯α−n Z ∞ ¯ ¯ n−α 2 = e−u u 2 −1 du ¡α¢ n 2 (4π) Γ 2 0 −n+α = where |x| γ (α) ¡ ¢ n 2α π 2 Γ α2 ¡ ¢ . γ (α) = Γ n−α 2 2 |x| du 4u2 HARMONIC ANALYSIS 61 So we have 1 n−α − |x| Gα (x) γ (α) Z ∞ n−α ¡ ¢ |x|2 α−n |x| = 1 − e−t e− 4t t 2 −1 dt ¡α¢ n (4π) 2 Γ 2 0 à ! α−n ¶ Z ∞µ 2 −1 n−α 2 2 |x|2 |x| |x| |x| − 4u −u = e 1−e du ¡α¢ n 4u 4u2 (4π) 2 Γ 2 0 ¶ Z ∞µ |x|2 n−α 1 1 − e− 4u e−u u 2 −1 du → 0 as x → 0. = α n ¡α¢ 2 π2Γ 2 0 (b) When |x| ≥ 1 Gα (x) = ≤ ≤ 1 n 2 ¡α¢ n 2 ¡α¢ (4π) Γ 1 (4π) Γ 1 n 2 (4π) Γ ≤ ce− Z |x| 2 2 2 ¡α¢ 2 ∞ e−t− |x|2 4t t α−n 2 −1 dt 0 Z ∞ t 1 t t 1 |x| 2 e− 2 − 8t − 2 − |x|2 8t t α−n 2 −1 dt 0 Z ∞ e− 2 − 8t − t α−n 2 −1 dt 0 . ¤ For any α > 0, since Gα ∈ L1 (Rn ), for any f ∈ Lp (Rn ), 1 ≤ p ≤ ∞, the Bessel potential of f , Jα f = Gα ∗ f ∈ Lp (Rn ) is well defined. We also define J0 f = f for any f ∈ Lp (Rn ). Proposition 16. For any α, β ≥ 0, Jα Jβ = Jα+β . Lemma 19. Let α ≥ 0. (a) Jα is a bijection from S onto S. (b) Let 1 ≤ p ≤ ∞ and g1 , g2 ∈ Lp (Rn ). Then Jα (g1 ) = Jα (g2 ) implies g1 = g2 . Proof. (a) Since Fourier transform is a bijection from S onto S, it follows from the definition that for any f ∈ S, ³ ´− α2 2 Jd fˆ. α f = 1 + |x| (b) Let g = g1 − g2 . Jα (g) = 0 implies that for any ϕ ∈ S, Z Z 0= Jα (g) ϕ = gJα (ϕ) . Rn Rn Since Jα (ϕ) can be any function in S, g = 0. Definition 13. The potential space Lpα (Rn ) is defined as Lpα = Jα (Lp (Rn )) . ¤ 62 HUIQIANG JIANG The Lpα -norm of f is defined by kf kp,α = kgkp if f = Jα (g) . Remark 20. k·kp,α is indeed well defined since Jα (g1 ) = Jα (g2 ) implies g1 = g2 . Obviously Lpα is a Banach space. Lemma 20. Let α > 0. (i) There exists a finite measure µα on Rn s.t. its Fourier transform is given by α |x| µ cα (x) = ³ 1 + |x| 2 ´ α2 . (ii) There exists a finite measure να on Rn s.t. ³ ´α α 2 2 = νc 1 + |x| α (x) (1 + |x| ) . Proof. (a) ³ |x| à α 1 + |x| 2 ´ α2 = 1− =1+ ! α2 1 1 + |x| ∞ X 2 k µα¶ ³ 2 (−1) k k=1 " = (2π) n 2 δ0 + ∞ X k=1 1 + |x| 2 ´−k #ˆ µα¶ 2 (−1) G2k (x) k k =µ cα (x) . Here " µα = (2π) n 2 δ0 + ∞ X k=1 # µα¶ 2 G (−1) 2k (x) k k is a finite measure. (b) We need Weiner’s theorem: If ϕ ∈ L1 (Rn ) satisfies ϕ̂ + 1 is nowhere zero, then there exists ψ ∈ L1 (Rn ) s.t. −1 (ϕ̂ + 1) Let " ϕ = (2π) n 2 ∞ X k=1 = ψ̂ + 1. # µα¶ 2 (−1) G2k (x) + Gα (x) , k k then α ϕ̂ + 1 = ³ |x| + 1 ´α 2 2 1 + |x| never vanishes, hence, there exists ψ ∈ L1 (Rn ) s.t. ³ ´α 2 2 1 + |x| ψ̂ + 1 = , α |x| + 1 HARMONIC ANALYSIS i.e., ³ 1 + |x| 2 ´ α2 α = ψ̂ + 1 + |x| ³ 63 ´ ψ̂ + 1 . n Thus we can take να = ψ + (2π) 2 δ0 . ¤ Lemma 21. Suppose 1 < p < ∞, and α ≥ 1. Then f ∈ Lpα (Rn ) iff f ∈ Lpα−1 (Rn ) ∂f and for each j, ∂x ∈ Lpα−1 (Rn ). Moreover, the two norms kf kp,α and j ° n ° X ° ∂f ° ° ° kf kp,α−1 + ° ∂xj ° p,α−1 j=1 are equivalent. Proof. 1. If f ∈ Lpα (Rn ), then there exists g ∈ Lp (Rn ), f = Jα (g). Since f = Jα (g) = Jα−1 (J1 g) and J1 g ∈ Lp (Rn ), we have f ∈ Lpα−1 (Rn ) with kf kp,α−1 = kJ1 gkp ≤ kgkp = kf kp,α . Now for each j, ³ ´− α2 d ∂f 2 ĝ = −ixj fˆ = −ixj 1 + |x| ∂xj ³ ´− α−1 |x| −ixj 2 2 = 1 + |x| ³ ´ 12 |x| ĝ 2 1 + |x| ³ ´ ³ ´− α−1 2 2 d µ c1 · R = 1 + |x| jg ³ ´− α−1 n 2 −\ 2 (2π) 2 µ1 ∗ Rj g. = 1 + |x| Hence, ³ ´ ∂f −n = Jα−1 (2π) 2 µ1 ∗ Rj g ∈ Lpα−1 (Rn ) . ∂xj Moreover, ° ° ° ∂f ° ° ° ° ∂xj ° p,α−1 ° ° ° ° −n = °(2π) 2 µ1 ∗ Rj g ° ≤ c kRj gkp ≤ C kgkp . 2. If f ∈ Lpα−1 (Rn ) and for each j, gj in Lp (Rn ), such that p ∂f ∂xj ∈ Lpα−1 (Rn ). Then there exists g and f = Jα−1 (g) and for each j, ∂f = Jα−1 (gj ) . ∂xj ³ ´− α−1 2 2 Since fˆ = 1 + |x| ĝ, we have ³ ´− α−1 ³ ´− α−1 d d ∂f ∂g 2 2 2 2 = −ixj fˆ = 1 + |x| . (−ixj ) ĝ = 1 + |x| ∂xj ∂xj 64 HUIQIANG JIANG However, ³ ´− α−1 ³ ´− α−1 d d ∂f ∂g 2 2 2 2 = 1 + |x| . ĝj = 1 + |x| ∂xj ∂xj Hence, we have ∂g = gj . ∂xj Now, the following formal calculation can be justified, ³ ´− α−1 2 2 fˆ = 1 + |x| ĝ ³ ´− α2 ³ ´1 2 2 2 = 1 + |x| 1 + |x| ĝ α ³ ´− 2 2 = 1 + |x| νc α (x) (1 + |x|) ĝ µ ³ ´− α2 µ \ µ ∂g ¶¶¶ 2 −n = 1 + |x| (2π) 2 να ∗ g + Rj . ∂xj Hence, f ∈ Lpα (Rn ) and kf kp,α ° µ ¶¶° µ ° ° ∂g −n 2 ° =° ν ∗ g + R (2π) α j ° ° ∂xj p ° ¶° µ ° ° ∂g ° ≤ c° °g + Rj ∂xj ° p à ° ! X° ° ∂g ° ° ° ≤ c kgkp + ° ∂xj ° . p ¤ Theorem 53. Suppose k is a nonnegative integer and p ∈ (1, ∞). Then Lpk (Rn ) = W k,p (Rn ) . Proof. By induction. ¤ Suppose that f ∈ Lp (Rn ). We define the Lp modulus of continuity: ωp (t) = kf (x + t) − f (x)kp . Then lim ωp (t) = 0. |t|→0 Let’s first recall the Minkowski integral inequality. Theorem 54. Let (X, dµ) and (Y, dν) be two σ-finite measure spaces and 1 ≤ p < ∞. Then for any measurable function F (x, y) on X × Y , µZ µZ ¶p ¶ p1 ¶ p1 Z µZ p |F (x, y)| dµx dνy ≤ |F (x, y)| dνy dµx . Y X X Y If (X, dµ) is the counting measure for {1, 2} and (Y, dν) is the Lebesgue measure on Rn . We deduce kf1 + f2 kp ≤ kf1 kp + kf2 kp . HARMONIC ANALYSIS 65 Proposition 17. Suppose 1 < p < ∞. Then f ∈ Lp1 (Rn ) iff f ∈ Lp (Rn ) and ωp (t) = O (|t|) as |t| → 0. Proof. Suppose f ∈ D (Rn ). Let t = |t| t0 . Z |t| f (x + t) − f (x) = (∇f, t0 ) (x + st0 ) ds. 0 Hence, Minkowski’s inequality implies kf (x + t) − f (x)kp ≤ |t| ° n ° X ° ∂f ° ° ° ° ∂xj ° . p j=1 n Since D (R ) is dense in Lp1 n n (R ), for any f ∈ D (R ), we have ° n ° X ° ∂f ° ° ° kf (x + t) − f (x)kp ≤ |t| ° ∂xj ° p j=1 and hence ωp (t) = O (|t|) as |t| → 0. On the other hand, suppose f ∈ Lp (Rn ) and ωp (t) = O (|t|) as |t| → 0. Let ej be the unit vector in j-th direction. Since ° ° ° f (x + sej ) − f (x) ° ° ° = O (1) , ° ° s p there exists a subsequence, sk → 0, s.t., f (x + sk ej ) − f (x) * fj weakly in Lp (Rn ) as k → ∞. sk Since for any ϕ ∈ D (Rn ), Z Z f (x + sk ej ) − f (x) ϕ (x) − ϕ (x − sk ej ) ϕ (x) dx = − f (x) dx, sk sk Rn Rn we have, letting k → ∞, Z Z ∂ϕ fj ϕ (x) dx = − f (x) dx. ∂xj Rn Rn Hence, ∂f = fj ∈ Lp (Rn ) ∂xj and f ∈ Lp1 (Rn ). ¤ 8. Extensions and Restrictions 8.1. Decomposition of open sets into cubes. Let’s recall the following theorem we proved. Theorem 55. Let F be a nonempty closed set in Rn s.t. F 6= Rn . Then its complement Ω = F c can be written as a union of cubes ∞ [ Qk , Ω= k=1 whose interiors are disjoint and for each k, diam Qk ≤ dist (Qk , F ) ≤ 4 diam Qk . 66 HUIQIANG JIANG ∞ Let F be the collection {Qk }k=1 . We say that two cubes Q1 , Q2 in F touch if Q1 ∩ Q2 6= ∅. Proposition 18. Suppose two cubes Q1 , Q2 ∈ F touch. Then 1 diam Q1 ≤ diam (Q1 ) ≤ 4 diam Q2 . 4 Proof. Since Q1 ∩ Q2 6= ∅, diam Q2 ≤ dist (Q2 , F ) ≤ dist (Q1 , F ) + diam Q1 ≤ 5 diam Q1 . Since diam Q2 = 2k diam Q1 for some integer k, we deduce diam Q2 ≤ 4 diam Q1 . ¤ n Proposition 19. Suppose Q ∈ F . Then there are at most N = 12 cubes in F which touch Q. ¡ ¢ Let ε ∈ 0, 14 be fixed. For any Qk ∈ F with center xk , we define ¡ ¢ Q∗k = (1 + ε) Qk − xk + xk . Proposition 20. Each point of Ω is contained in at most N of the cubes Q∗k . Proof. Let Q and Qk be two cubes in F. We consider the union of Qk with all cubes in F which touches Qk , then this union contains Q∗k as its interior because of the proposition 18. Hence, Q∗k ∩ Q 6= ∅ iff Qk ∩ Q 6= ∅. ¤ Let Q0 be the unit cube centered at the origin. Fix ϕ ∈ D (Rn ), s.t. 0 ≤ ϕ ≤ 1, ϕ = 1 on Q0 and ϕ = 0 on Rn \Q∗0 . Let µ ¶ x − xk ϕk (x) = ϕ lk where xk is the center of Qk and lk is the side length of Qk . We now define ϕk (x) ϕ∗k (x) = P∞ , x ∈ Ω. k=1 ϕk (x) Then we have ∞ X ϕ∗k (x) ≡ 1 for x ∈ Ω = F c . k=1 8.2. Extension theorems of Whitney type. We first consider the regularized distance. Let F be a nonempty closed set in Rn s.t. F is not the whole space. Let δ (x) be the distance of x from F . Then δ is a Lipschitz function. Theorem 56. There exists a function ∆x = ∆ (x, F ) defined in F c s.t. (a) c1 δ (x) ≤ ∆x ≤ c2 δ (x), x ∈ F c ; (b) ∆x is smooth in F c and ¯ ¯ α ¯ ¯ ∂ 1−|α| ¯ ¯ ¯ ∂xα ∆ (x)¯ ≤ Bα (δ (x)) where Bα , c1 , c2 are constants independent of F . HARMONIC ANALYSIS 67 Proof. Let ∆x = ∞ X (diam Qk ) ϕk (x) . k=1 If x ∈ Qk , then δ (x) = dist (x, F ) ≤ dist (Qk , F ) + diam Qk ≤ 5 diam Qk , and hence ∆x = ∞ X (diam Qk ) ϕk (x) ≥ (diam Qk ) ϕk (x) = diam Qk ≥ k=1 Also if x ∈ Q∗k , 1 δ (x) . 5 then 1 3 diam Qk ≥ diam Qk . 4 4 c ∗ Since any x ∈ F lies in at most N of the Qk , we have X 4N ∆x ≤ diam Qk ≤ δ (x) . 3 ∗ δ (x) ≥ dist (Qk , F ) − x∈Qk For any α, |α| ≥ 1, there exists Aα such that, |Dα ϕk (x)| ≤ Aα (diam Qk ) −|α| ∆x is smooth in F c and for any x ∈ F c , X |Dα ∆ (x)| ≤ diam Qk |Dα ϕk (x)| x∈Q∗ k ≤ X x∈Q∗ k ≤ N Aα 1−|α| Aα (diam Qk ) µ ¶1−|α| 1 δ (x) 6 1−|α| = Bα (δ (x)) where we used the fact that if x ∈ Q∗k , then δ (x) = dist (x, F ) ≤ dist (Qk , F ) + 2 diam Qk ≤ 6 diam Qk . ¤ ∞ Given closed F and the cube decomposition {Qk }k=1 of F c . For each Qk , fix a point pk ∈ F s.t. dist (Qk , F ) = dist (Qk , pk ) . Let f be a continuous function on F . We define f (x) if x ∈ F, ∞ P E0 (f ) (x) = ∗ f (pk ) ϕk (x) if x ∈ F c . k=1 Proposition 21. E0 (f ) is continuous on Rn and smooth in F c . 68 HUIQIANG JIANG Proof. E0 (f ) is smooth in F c because for each x ∈ F c , there exists a nbhd of x such that the summation is finite. Now we need to show the continuity of E0 (f ) for each y ∈ F . If x ∈ Q∗k , then |y − pk | ≤ |y − x| + |x − pk | ≤ c |y − x| . For each ε > 0, there exists δ, s.t. |f (z) − f (y)| < ε whenever |z − y| < δ. Now for any x ∈ F c , s.t. |y − x| < δc , |E0 (f ) (x) − f (y)| ≤ ∞ X |f (pk ) − f (y)| ϕ∗k (x) k=1 = X |f (pk ) − f (y)| ϕ∗k (x) < ε x∈Q∗ k X ϕ∗k (x) ≤ ε. x∈Q∗ k Hence E0 (f ) is continuous at y. ¤ For any 0 < γ ≤ 1, we define γ Lip (γ, Rn ) = {f : |f (x)| ≤ M, |f (x) − f (y)| ≤ M |x − y| , x, y ∈ Rn } . Then Lip (γ, Rn ) is a Banach space with norm kf kC γ = kf k∞ + [f ]γ = kf k∞ + inf x,y∈R x6=y n |f (x) − f (y)| . γ |x − y| We remark here that if γ > 1, then Lip (γ, Rn ) defined as above contains only constant functions. If F ⊂ Rn is a nonempty closed set, we can define the Banach space Lip (γ, F ) in a similar way. Theorem 57. The linear extension operator E0 maps Lip (γ, F ) continuously into Lip (γ, Rn ) if 0 < γ ≤ 1. The norm of this mapping has a bound independent of the closed set F . Proof. First, we observe that for any x ∈ F c , −|α| |Dα ϕ∗k (x)| ≤ A0α (diam Qk ) . c Now we claim for any x ∈ F , ¯ ¯ ¯ ∂ ¯ γ−1 ¯ ¯ . ¯ ∂xi E0 (f ) (x)¯ ≤ c [f ]γ (δ (x)) In fact, for any x ∈ F c , choose y ∈ F s.t. δ (x) = |x − y|, we have ¯ ¯ ¯ ¯¯ ∞ ¯ ∗ ¯ ∂ ¯ ¯X (x) ∂ϕ ¯ k ¯ ¯=¯ E (f ) (x) (f (p ) − f (y)) ¯ 0 k ¯ ∂xi ¯ ¯ ∂xi ¯ k=1 ¯ ∗ ¯ ∞ X ¯ ∂ϕk (x) ¯ ¯ ¯ ≤ |f (pk ) − f (y)| ¯ ∂xi ¯ k=1 ≤ c [f ]γ (δ (x)) where c only depends on n. γ−1 HARMONIC ANALYSIS 69 Now for any x ∈ F c , y ∈ F , ¯ ¯∞ ¯X ¯ ¯ ¯ ∗ (f (pk ) − f (y)) ϕk (x)¯ |E0 (f ) (x) − E0 (f ) (y)| = ¯ ¯ ¯ k=1 X ≤c |f (pk ) − f (y)| x∈Q∗ k ≤ c [f ]γ X γ γ |pk − y| ≤ c [f ]γ |x − y| . x∈Q∗ k For any x, y ∈ F c , let L be the line segement from x to y. Case I: |x − y| ≤ dist (L, F ). Then |E0 (f ) (x) − E0 (f ) (y)| ≤ |x − y| sup |∇E0 (f ) (x0 )| x0 ∈L ≤ c [f ]γ |x − y| (δ (x)) γ−1 γ ≤ c [f ]γ |x − y| . Case II: |x − y| > dist (L, F ). Then there exists x0 ∈ L and y 0 ∈ F s.t. |x0 − y 0 | < |x − y| , hence |x − y 0 | < 2 |x − y| and |y − y 0 | < 2 |x − y| . So we have |E0 (f ) (x) − E0 (f ) (y)| ≤ |E0 (f ) (x) − E0 (f ) (y 0 )| − |E0 (f ) (y) − E0 (f ) (y 0 )| γ γ γ ≤ c [f ]γ |x − y 0 | + c [f ]γ |y − y 0 | ≤ c [f ]γ |x − y| . ¤ More generally, let ω (δ) be a modulus of continuity satisfying (i) ω(δ) is increasing as δ → 0; δ (ii) ω (2δ) ≤ cω (δ) for any δ > 0. We define Lip (ω, F ) to define the collection of continuous functions f such that kf kC ω = kf k∞ + [f ]ω = kf k∞ + inf x,y∈Rn x6=y |f (x) − f (y)| < ∞. ω (|x − y|) Then Lip (ω, F ) is a Banach space. Corollary 10. The linear extension operator E0 maps Lip (ω, F ) continuously into Lip (ω, Rn ). Let γ > 1 and k be an integer s.t. k < γ ≤ k + 1. We say that f ∈ Lip (γ, F ) if f is defined on F and there exists f (α) , |α| ≤ k, such that, for any 0 ≤ |α| ≤ k, if f (α) (x) = X |α+β|≤k then f (α+β) (y) β (x − y) + Rα (x, y) , β! ¯ ¯ ¯ (α) ¯ γ−|α| ¯f (x)¯ ≤ M and |Rα (x, y)| ≤ M |x − y| 70 HUIQIANG JIANG © ª holds for any x, y ∈ F . Since for f ∈ Lip (γ, F ), the collection f (α) may not be © (α) ª unique, we could view f as an element in Lip (γ, F ) and take the smallest |α|≤k M as its © norm. ª Let f (α) |α|≤k ∈ Lip (γ, F ). We define f (0) (x) if x ∈ F, ³n o´ ∞ (α) P Ek f = ∗ P (x, pj ) ϕj (x) if x ∈ F c j=1 where P (x, y) = X f α (y) α (x − y) , x ∈ F c , y ∈ F. α! |α|≤k Theorem 58. The linear extension operator Ek maps Lip (γ, F ) continuously into Lip (γ, Rn ) if k < γ ≤ k + 1. The norm of this mapping has a bound independent of the closed set F . 8.3. Extension theorem for Sobolev functions in Lipschitz domain. Let k ≥ 0 be an integer and 1 ≤ p ≤ ∞. Similar to W k,p (Rn ), we can define W k,p (Ω) for any open subset Ω ⊂ Rn . The goal of this section is to find an extension operator from W k,p (Ω) to W k,p (Rn ). We first consider special Lipschitz domains whose boundary are represented by uniform Lipschitz graphs. Let ϕ : Rn → R be a function satisfying |ϕ (x) − ϕ (x0 )| ≤ M |x − x0 | for any x, x0 ∈ Rn . Let © ª Ω = (x, y) ∈ Rn+1 : x ∈ Rn , y ∈ R, y > ϕ (x) . The smallest such M is called the bound of the specail Lipschitz domain Ω. Lemma 22. There exists a continuous function ψ defined on [1, ∞) which is rapidly decreasing at ∞ and satisfies Z ∞ ψ (t) dt = 1, 1 Z ∞ tk ψ (t) dt = 0 for k ∈ N. 1 Proof. Let ψ (t) = µ ¶ 1 e Im e−ω(t−1) 4 πt iπ where ω = e− 4 . ¤ Lemma 23. Let F = D and ∆ (x, y) be the regularized distance from F . Then for any (x, y) ∈ F c , ϕ (x) − y ≤ c∆ (x, y) where c is a constant depending only on M . Proof. For any (x, y) ∈ F c , µ ¶− 12 1 δ (x, y) ≥ 1 + 2 (ϕ (x) − y) . M HARMONIC ANALYSIS Since ∆ (x, y) ≥ the lemma holds with 71 1 δ (x, y) , 5 µ ¶1 1 2 c=5 1+ 2 M ¤ Let δ ∗ (x, y) = 2c∆ (x, y) . ¡ ¢ We define for any f ∈ W k,p (Ω) ∩ C ∞ Ω such that f and all its partial derivatives are bounded in Ω, ½ f (x, y) if y ≥ ϕ (x) , R∞ E (f ) (x, y) = ∗ ψ (t) f (x, y + tδ (x, y)) dt if y < ϕ (x) . 1 ¡ ¢ Proposition 22. E (f ) ∈ C ∞ Rn+1 and kE (f )kW k,p (Rn+1 ) ≤ Ak,n (M ) kf kW k,p (Ω) . Proof. First, ¤ ¡ ¢ Theorem 59. The extension operator E maps W k,p (Ω) continuously into W k,p Rn+1 . Proof. Let η be a nonnegative smooth function with compact support in the interior of the cone Γ− = {(x, y) : M |x| < |y| , y < 0} . s.t. Z η = 1. Rn+1 For any ε > 0, define ηε (u) = 1 εn+1 For any f ∈ W k,p (Ω), we define fε (u) = f ∗ ηε = ¡ ¢ Then fε ∈ C ∞ Ω and η ³u´ ε , u ∈ Rn+1 . Z Rn+1 f (u − v) ηε (v) dv. kfε kW k,p (Ω) ≤ kf kW k,p (Ω) . Now if 1 ≤ p < ∞, then lim kfε − f kW k,p (Ω) = 0. ε→0+ Let Eε (f ) = E (fε ) . ¡ ¢ Then Eε (f ) is a Cauchy sequence in W k,p Rn+1 and hence E (f ) = lim+ Eε (f ) ε→0 defines an extension operator satisfying kE (f )kW k,p (Rn+1 ) ≤ Ak,n (M ) kf kW k,p (Ω) . 72 HUIQIANG JIANG And if p = ∞, k ≥ 1, lim kfε − f kW k−1,∞ (Ω) = 0. ¡ ¢ So Eε (f ) is a Cauchy sequence in W k−1,∞ Rn+1 and hence ε→0+ E (f ) = lim Eε (f ) ε→0+ defines an extension operator satisfying kE (f )kW k,∞ (Rn+1 ) ≤ Ak,n (M ) kf kW k,∞ (Ω) . ¤