PIOTR HAJLASZ
1. Preliminaries
Let us first fix notation that will be used throughout the book. We will be working mostly in the Euclidean spaces and with the Lebesgue measure. The Lebesgue measure of a n measurable set E will be denoted by |E|. The volume of a ball of radius r is |B(0, r)| = ωnr . Then it is well known that the (n − 1)-dimensional measure of the sphere Sn−1(0, r) equals n−1 nωnr . The integral average over a set E of positive measure will be denoted by Z 1 Z f(x) dx = f(x) dx. E |E| E The Lebesgue measure on a surface will be denoted by dσ so the integration in the spherical coordinates will take the form Z Z ∞ Z f(x) dx = f(sθ) dσ(θ) ds. Rn 0 Sn−1 Unless we will explicitly state otherwise, all functions and all function spaces will be com- plex valued.
n ∞ A generic open set in R will usually be denoted by Ω and C0 (Ω) will stand for a space of smooth functions with compact support in Ω. The space of continuous functions n n vanishing at infinity will be denote by C0(R ), i.e. f ∈ C0(R ) if f is continuous and lim f(x) = 0. |x|→∞
n ∞ n C0(R ) is a Banach space with respect to the supremum norm k · k∞ and C0 (R ) is a n dense subset of C0(R ). We will use the multiindex notation ∂|α| Dαu = , xα = xα1 · ... · xαn α1 αn 1 n ∂x1 . . . ∂xn where α = (α1, . . . , αn), |α| = α1 + ... + αn. Then the product rule takes the form X α! Dα(fg) = DβfDγg, where α! = α ! · ... · α !. β!γ! 1 n β+γ=α
Date: February 27, 2017. 1 2 PIOTR HAJLASZ
The characteristic function of a set E will be denoted by χE. By δa we will denote the Dirac measure centered at a i.e., Z f(x) dδa(x) = f(a). Rn By an increasing function we will mean a non-decreasing function and increasing func- tions that are actually increasing, will be called strictly increasing. Similar terminology applies to decreasing and strictly decreasing functions. By sgn x we will denote the sign of the number x, i.e. sgn x = 1 for positive x, sgn x = −1 for negative x and sgn 0 = 0. Various constants will usually be denoted by C. By writing C(n, m) we will indicate the the constant C(n, m) depends on n and m only. We adopt the rule that a constant C may change its value within one string of estimates. This will allow us to minimize the use of indices. HARMONIC ANALYSIS 3
2. The Maximal function
2.1. The maximal theorem.
1 n Definition 2.1. For a locally integrable function f ∈ Lloc(R ) the Hardy-Littlewood max- imal function is defined by Z n Mf(x) = sup |f(y)| dy, x ∈ R . r>0 B(x,r)
The operator M is not linear but it is subadditive. We say that an operator T from a space of measurable functions into a space of measurable functions is subadditive if
|T (f1 + f2)(x)| ≤ |T f1(x)| + |T f2(x)| a.e. and |T (kf)(x)| = |k||T f(x)| for k ∈ C. The following integrability result, known also as the maximal theorem, plays a fundamental role in many areas of mathematical analysis.
Theorem 2.2 (Hardy-Littlewood-Wiener). If f ∈ Lp(Rn), 1 ≤ p ≤ ∞, then Mf < ∞ a.e. Moreover
(a) For f ∈ L1(Rn) 5n Z (2.1) |{x : Mf(x) > t}| ≤ |f| for all t > 0. t Rn (b) If f ∈ Lp(Rn), 1 < p ≤ ∞, then Mf ∈ Lp(Rn) and p 1/p kMfk ≤ 2 · 5n/p kfk for 1 < p < ∞, p p − 1 p
kMfk∞ ≤ kfk∞ . Remark 2.3. Note that if f ∈ L1(Rn) is a non-zero function, then Mf 6∈ L1(Rn). Indeed, R if λ = B(0,R) |f| > 0, then for |x| > R, B(0,R) ⊂ B(x, R + |x|) so Z λ Mf(x) ≥ |f| ≥ n , B(x,R+|x|) ωn(R + |x|) Since the function on the right hand side is not integrable on Rn, the statement (b) of the theorem is not true for p = 1. 4 PIOTR HAJLASZ
Remark 2.4. If g ∈ L1(Rn), then the Chebyschev inequality 1 Z (2.2) |{x : |g(x)| > t}| ≤ |g| for t > 0 t Rn 1 is easy to prove. Thus if an (not necessarily linear) operator T is bounded in L , kT fk1 ≤ Ckfk1, it immediately follows from Chebyschev’s inequality that C Z (2.3) |{x : |T f(x)| > t}| ≤ |f(x)| dx. t Rn The maximal function M is not bounded in L1, but it still satisfies the weaker estimate (2.3). For this reason estimate (2.1) is called the weak type estimate.
In the proof of Theorem 2.2 we will need the following two results. Theorem 2.5 (Cavalieri’s principle). If µ is a σ-finite measure on X and Φ : [0, ∞) → [0, ∞) is increasing, absolutely continuous and Φ(0) = 0, then Z Z ∞ Φ(|f|) dµ = Φ0(t)µ({|f| > t}) dt . X 0
Proof. The result follows immediately from the equality Z Z Z |f(x)| Φ(|f(x)|) dµ(x) = Φ0(t) dt dµ(x) X X 0 and the Fubini theorem. Corollary 2.6. If µ is a σ-finite measure on X and 0 < p < ∞, then Z Z ∞ |f|p dµ = p tp−1µ({|f| > t}) dt . X 0
The next result has many applications that go beyond the maximal theorem. Here and in what follows by 5B we denote the ball concentric with B and with five times the radius. Theorem 2.7 (5r-covering lemma). Let B be a family of balls in a metric space such that sup{diam B : B ∈ B} < ∞. Then there is a subfamily of pairwise disjoint balls B0 ⊂ B such that [ [ B ⊂ 5B. B∈B B∈B0 More precisely, for every B ∈ B there is B0 ∈ B0 such that B ∩ B0 6= ∅ and B ⊂ 5B0. If the metric space is separable, then the family B0 is countable (or finite) and we can 0 ∞ arrange it as a (possibly finite) sequence B = {Bi}i=1 so ∞ [ [ B ⊂ 5Bi . B∈B i=1 Remark 2.8. Here B can be either a family of open balls or closed balls. In both cases the proof is the same. HARMONIC ANALYSIS 5
Proof. Let sup{diam B : B ∈ B} = R < ∞. Divide the balls in the family B according to their diameters: n R R o F = B ∈ B : < diam B ≤ . j 2j 2j−1 S∞ Clearly B = j=1 Fj.
First we select balls as large as possible so we choose balls from the family F1 and we take as many as we can: we define B1 ⊂ F1 to be a maximal family of pairwise disjoint balls.
Suppose the families B1,..., Bj−1 are already defined.
In the construction of Bj we want to select balls form Fj, but we have to make sure that balls that we choose do not intersects with any of the previously selected balls. Thus we have to select balls from n j−1 o ˆ 0 0 [ Fj = B ∈ Fj : B ∩ B = ∅ for all B ∈ Bi . i=1 ˆ Then we define Bj to be the maximal family of pairwise disjoint balls in Fj 0 S∞ Finally we define B = j=1 Bj. It follows immediately from the definition of Bj that Sj 0 0 Sj every ball B ∈ Fj intersects with a ball in i=1 Bj. If B ∩ B 6= ∅, B ∈ i=1 Bi, then R R diam B ≤ = 2 · < 2 diam B0 2j−1 2j 0 and hence B ⊂ 5B .
1 n Proof of Theorem 2.2. (a) Let f ∈ L (R ) and let Et = {x : Mf(x) > t}. For x ∈ Et, there is rx > 0 such that Z Z −1 |f| > t so |B(x, rx)| < t |f| . B(x,rx) B(x,rx) 1 n Observe that supx∈Et rx < ∞, because f ∈ L (R ). The family of balls {B(x, rx)}x∈Et forms a covering of the set Et so applying the 5r-covering lemma there is a sequence of S∞ pairwise disjoint balls B(xi, rxi ), i = 1, 2,... such that Et ⊂ i=1 B(xi, 5rxi ) and hence ∞ n ∞ Z n Z n X 5 X 5 |Et| ≤ 5 |B(xi, rxi )| ≤ |f| ≤ |f| . t t n i=1 i=1 B(xi,rxi ) R The proof is complete.
p n (b) Let f ∈ L (R ), 1 < p ≤ ∞. Since kMfk∞ ≤ kfk∞ we can assume that 1 < p < ∞. Let f = f1 + f2, where
f1 = fχ{|f|>t/2}, f2 = fχ{|f|≤t/2} be a decomposition of |f| into its ‘lower’ and ‘upper’ parts. It is easy to check that f1 ∈ 1 n ∞ n L (R ) and clearly f2 ∈ L (R ).
The idea now if to apply the weak type estimate from (a) to f1 and the estimate kMfk∞ ≤ kfk∞ to f2. This will give us an estimate for the size of the level set |{Mf > t}| 6 PIOTR HAJLASZ and the result will follow from Corollary 2.6. The same idea is the core of the so called real interpolation method and we will see it again in the proof of the Marcinkiewcz Interpolation Theorem 8.9.
Since |f| ≤ |f1| + t/2 we have Mf ≤ Mf1 + t/2 and hence
{Mf > t} ⊂ {Mf1 > t/2} . Thus 2 · 5n Z (2.4) |Et| = |{Mf > t}| ≤ |f1(x)| dx t Rn 2 · 5n Z = |f(x)| dx . t {|f|>t/2} Cavalieri’s principle gives Z Z ∞ |Mf(x)|p dx = p tp−1|{Mf > t}| dt Rn 0 Z ∞ 2 · 5n Z ≤ p tp−1 |f(x)| dx dt 0 t {|f|>t/2} Z Z 2|f(x)| = 2 · 5np |f(x)| tp−2 dt dx Rn 0 p Z = 2p · 5n |f(x)|p dx p − 1 Rn and the results follows. Remark 2.9. Note that we proved in (2.4) the following inequality 2 · 5n Z (2.5) |{x : Mf(x) > t}| ≤ |f(x)| dx t {|f|>t/2} which is slightly stronger than (2.1). Later we will see that the measure of the set |{Mf > t}| has a similar bound from below, see Proposition 2.33.
Remark 2.10. For a positive measure µ on Rn we define the maximal function by µ(B(x, r)) Mµ(x) = sup . r>0 |B(x, r)| A minor modification of the proof of Theorem 2.2(a) leads to the following result.
Proposition 2.11. If µ is a finite positive Borel measure on Rn, then 5n |{x : Mµ(x) > t}| ≤ µ( n) for all t > 0. t R Remark 2.12. It is natural to inquire whether in the Euclidean case, there is a version of Theorem 2.7 for families of cubes. The answer is in the positive and it follows immediately from Theorem 2.7 and the fact that cubes with sides parallel to coordinate axes are balls n in R with respect to the metric d∞(x, y) = maxi |xi − yi|. In the following result 5Q will denote a cube concentric with Q and with five times the diameter. HARMONIC ANALYSIS 7
Corollary 2.13. Let F be a family of open or closed cubes in Rn with sides parallel to coordinate directions such that sup{diam Q : Q ∈ F} < ∞. Then there is a subfamily F 0 ⊂ F of pairwise disjoint cubes such that [ [ Q ⊂ 5Q. Q∈F Q∈F 0 More precisely for every Q ∈ F there is Q0 ∈ F 0 such that Q ∩ Q0 6= ∅ and Q ⊂ 5Q0.
Although we are interested mostly in the Euclidean setting, it was important to formulate Theorem 2.7 in a the setting of metric spaces as we could conclude Corollary 2.13 directly from Theorem 2.7.
In the next two sections we will show applications of the maximal theorem.
2.2. Fractional integration theorem. As a first application of the maximal theorem we will prove a result about integrability of the Riesz potentials. Definition 2.14. For 0 < α < n and n ≥ 2 we define the Riesz potential by n α α 1 Z f(y) π 2 2 Γ (I f)(x) = dy , where γ(α) = 2 . α n−α n−α γ(α) n |x − y| R Γ 2
At this moment the particular value of the constant γ(α) is not important to us. We could even replace this constant by 1.1 Theorem 2.15 (Hardy-Littlewood-Sobolev). Let α > 0, 1 < p < ∞ and αp < n. Then there is a constant C = C(n, p, α) such that p n kIαfkp∗ ≤ Ckfkp for all f ∈ L (R ) , where p∗ = np/(n − αp). 2 p q Remark 2.16. It is not difficult to show directly that if Iα : L → L is bounded, then q = p∗. The idea is to use the change of variables x 7→ tx, t > 0, and see how both sides of the inequality kIαfkq ≤ Ckfkp change. This is so called a scaling argument.
We precede the proof with a technical lemma. Lemma 2.17. If 0 < α < n, and δ > 0, then there is a constant C = C(n, α) such that Z |f(y)| α n n−α dy ≤ Cδ Mf(x) for all x ∈ R . B(x,δ) |x − y|
Proof. For x ∈ Rn and δ > 0 consider the annuli δ δ A(k) = B x, − B x, . 2k 2k+1
1 It will play an important role later: I2f is the convolution with the fundamentals solution of −∆, see Theorem 5.60. Also Riesz potentials will be studied in Section 7 in connection with fractional powers of the Laplace operator. 2A nice exercise. 8 PIOTR HAJLASZ
We have ∞ Z |f(y)| X Z |f(y)| dy = dy |x − y|n−α |x − y|n−α B(x,δ) k=0 A(k) ∞ α−n X δ Z ≤ |f(y)| dy 2k+1 k=0 A(k) ∞ α−n n X δ δ Z ≤ ωn k+1 k |f(y)| dy 2 2 k k=0 B(x,δ/2 ) α−n ∞ ! 1 X 1 ≤ ω δα Mf(x) . n 2 2kα k=0 The proof is complete.
Proof of Theorem 2.15. Fix δ > 0. H¨older’sinequality and integration in polar coordinates yield 0 Z |f(y)| Z dy 1/p n−α dy ≤ kfkp (n−α)p0 Rn\B(x,δ) |x − y| Rn\B(x,δ) |x − y| Z ∞ 1/p0 n−1−(n−α)p0 = kfkp nωn s ds δ α−(n/p) = C(n, p, α)δ kfkp , n−1 because nωn equals the (n − 1)-dimensional measure of the unit sphere S and n − (n − α)p0 < 0. This and the lemma give α α−(n/p) |Iαf(x)| ≤ C δ Mf(x) + δ kfkp . Taking3 Mf(x)−p/n δ = kfkp yields αp αp 1− n |Iαf(x)| ≤ C(Mf(x)) n kfkp which is equivalent to αp ∗ ∗ p p n p |Iαf(x)| ≤ C(Mf(x)) kfkp . Integrating both sides over Rn and applying boundedness of the maximal function in Lp gives the result.
2.3. The Lebesgue differentiation theorem. We will show now how to prove the Lebesgue differentiation theorem from the maximal theorem.
n For h ∈ R we define τhf(x) = f(x + h). p n Lemma 2.18. If f ∈ L (R ), 1 ≤ p < ∞, then kτhf − fkp → 0 as h → 0.
3 α α−(n/p) We apply here a standard trick: we take δ > 0 such that δ Mf(x) = δ kfkp. HARMONIC ANALYSIS 9
This result is obvious if f is a compactly supported smooth function, because in that ∞ p case τhf converges uniformly to f. The general case follows from the density of C0 in L . 1 n R 1 Lemma 2.19. Let f ∈ L (R ) and let fr(x) = B(x,r) f(y) dy. Then fr → f in L as r → 0. In particular, there is a sequence ri → 0 such that Z lim f(y) dy = f(x) a.e. i→∞ B(x,ri)
Proof. We have Z Z Z |fr(x) − f(x)| dx ≤ |f(y) − f(x)| dy dx Rn Rn B(x,r) Z Z = |f(x + y) − f(x)| dy dx Rn B(0,r) Z (2.6) = kτyf − fk1 dy . B(0,r) 1 Since τyf → f in L as y → 0, the right hand side of (2.6) converges to 0 as r → 0. The 1 existence of a sequence ri follows from the fact that from a sequence convergent in L we can extract a subsequence convergent a.e.
The Lebesgue differentiation theorem gives much more.
1 n Theorem 2.20 (Lebesgue differentiation theorem). If f ∈ Lloc(R ), then Z lim f(y) dy = f(x) a.e. r→0 B(x,r)
1 n Proof. Since the theorem is local in nature we can assume that f ∈ L (R ). Let fr(x) = R B(x,r) f(y) dy and define
Ωf(x) = lim sup fr(x) − lim inf fr(x) . r→0 r→0
It suffices to prove that Ωf = 0 a.e. Indeed, this property means that fr → g converges a.e. to a measurable function g. Since by Lemma 2.19, fri → f a.e. we get g = f a.e. Observe that Ωf ≤ 2Mf, hence for any ε > 0, Theorem 2.2(a) yields C Z |{x :Ωf(x) > ε}| ≤ |f| . ε Rn 2 Let h be a continuous function such that kf − hk1 < ε . Continuity of h implies Ωh = 0 everywhere and hence Ωf ≤ Ω(f − h) + Ωh = Ω(f − h) , so C Z |{Ωf > ε}| ≤ |{Ω(f − h) > ε}| ≤ |f − h| ≤ Cε . ε Rn Since ε > 0 can be arbitrarily small we conclude Ωf = 0 a.e. 10 PIOTR HAJLASZ
1 n If f ∈ Lloc(R ), then we can define f at every point by the formula Z (2.7) f(x) := lim sup f(y) dy . r→0 B(x,r) According to the Lebesgue differentiation theorem this is a representative of f in the class of functions that coincide with f a.e.
1 n n Definition 2.21. Let f ∈ Lloc(R ). We say that x ∈ R is a Lebesgue point of f if Z lim |f(y) − f(x)| dy = 0 , r→0 B(x,r) where f(x) is defined by (2.7).
1 n Theorem 2.22. If f ∈ Lloc(R ), then the set of points that are not Lebesgue points of f has measure zero.
Proof. For c ∈ Q let Ec be the set of points for which Z (2.8) lim |f(y) − c| dy = |f(x) − c| r→0 B(x,r) does not hold. Clearly |E | = 0 and hence the set E = S E has measure zero too. Thus c c∈Q c for x ∈ Rn \ E and all c ∈ Q, (2.8) is satisfied. If x ∈ Rn \ E and f(x) ∈ R, approximating f(x) by rational numbers one can easily check that Z lim |f(y) − f(x)| dy = |f(x) − f(x)| = 0 . r→0 B(x,r) Indeed, if ε > 0 is arbitrary and c ∈ Q is such that |f(x) − c| < ε/3, then there is δ > 0 such that for 0 < r < δ Z Z |f(y) − f(x)| dy ≤ |f(y) − c| dy + |c − f(x)| B(x,r) B(x,r) ε < |f(x) − c| + + |c − f(x)| < ε. 3 The proof is complete. n p Definition 2.23. We say that x ∈ R is a p-Lebesgue point of f ∈ Lloc, 1 ≤ p < ∞ if Z lim |f(y) − f(x)|p dy = 0 . r→0 B(x,r)
The same method as above leads to the following result that we leave as an exercise.
p n Theorem 2.24. If f ∈ Lloc(R ), 1 ≤ p < ∞, then the set of points that are not p-Lebesgue points of f has measure zero.
Definition 2.25. Let E ⊂ Rn be a measurable set. We say that x ∈ Rn is a density point of E if |B(x, r) ∩ E| lim = 1 . r→0 |B(x, r)| HARMONIC ANALYSIS 11
Applying the Lebesgue theorem to the characteristic function of the set E, i.e., f = χE we obtain Theorem 2.26. Almost every point of a measurable set E ⊂ Rn is its density point and a.e. point of Rn \ E is not a density point of E.
The Lebesgue theorem shows that for almost all x, the averages of f over balls centered as x converge to f(x). It is natural to inquire whether we can replace balls by other sets like cubes or even balls, but not centered at x. Definition 2.27. We say that a family F of measurable subsets of Rn is regular at x ∈ Rn if
(a) The sets are bounded and have positive measure. (b) There is a sequence {Si} ⊂ F with |Si| → 0 as i → ∞. (c) There is a constant C > 0 such that for all S ∈ F we have |S| ≥ C|BS|, where BS is the smallest closed ball centered at x that contains S.
Note that if |Si| → 0, then the radius of BSi approaches to zero so the sets Si approach to x. Examples of families regular at x include:
• Family of all balls that contain x. • Family of all cubes centered at x. • Family of all cubes that contain x. 1 n Theorem 2.28. If f ∈ Lloc(R ), x is a Lebesgue point of f, and if F is a family regular at x, then Z lim f(y) dy = f(x) . S∈F |S|→0 S
Proof. We have Z Z Z 1 f(y) dy − f(x) ≤ |f(y) − f(x)| dy ≤ |f(y) − f(x)| dy S S |S| BS |B | Z 1 Z ≤ S |f(y) − f(x)| dy ≤ |f(y) − f(x)| dy → 0 |S| BS C BS as |S| → 0.
Note that if F is a family regular at 0 and if we define the maximal function associated with F by Z MF f(x) = sup |f(x − y)| dy , S∈F S then a routine calculation shows that
(2.9) MF f(x) ≤ CMf(x) , so MF satisfies the claim of Theorem 2.2 (with different constants). In particular MF is a bounded operator in Lp, 1 < p ≤ ∞. 12 PIOTR HAJLASZ
Let F be the family of all rectangular boxes in Rn that contain the origin and have sides parallel to the coordinate axes. With the family we can associate the maximal function Z Mff(x) = sup |f(x − y)| dy . S∈F S
Note that the family F is not regular at 0 and hence the boundedness of Mf in Lp, 1 < p ≤ ∞ cannot be concluded from (2.9). However, we have Theorem 2.29 (Zygmund). For 1 < p < ∞ there is a constant C = C(n, p) > 0 such that
(2.10) kMffkp ≤ Ckfkp . p Moreover, if f ∈ Lloc, 1 < p < ∞, then Z (2.11) lim f(x − y) dy = f(x) a.e. diam S→0 S∈F S
Proof. First we will prove how to conclude (2.11) from (2.10). Note that since the family is not regular at 0, (2.11) is not a consequence of Theorem 2.28 (see also Theorem 2.30). However, Z Z 0 ≤ lim sup f(x − y) dy − lim inf f(x − y) dy ≤ 2Mff(x) diam S→0 diam S→0 S S S∈F S∈F and hence (2.11) follows from (2.10) by almost the same argument as the one used to deduce the Lebesgue Differentiation Theorem from Theorem 2.2. We leave details to the reader. Thus it remains to prove (2.10). In dimension one Z x+b Mff(x) = sup |f(y)| dy a,b>0 x−a is the so called non-centered Hardy-Littlewood maximal function. Clearly Mff ≤ CMf which gives (2.10). For simplicity of notation we will prove the higher dimensional result for n = 2 only. The proof in the case of general n > 1 is similar. We have
Z x2+b2 Z x1+b1 Mff(x1, x2) = sup |f(y1, y2)| dy1 dy2 a ,b >0 1 1 x2−a2 x1−a1 a2,b2>0 Z x2+b2 Z x1+b1 ≤ sup sup |f(y1, y2)| dy1 dy2 . a2,b2>0 x2−a2 a1,b1>0 x1−a1 On the right hand side we have iteration of one dimensional non-centered maximal func- tions. First we apply the maximal function to variable y1 and evaluate it at x1 and then we apply the maximal function to the variable y2 and evaluate it at x2. It is easy to see now that inequality (2.10) follows from the one dimensional case applied twice and the Fubini theorem. HARMONIC ANALYSIS 13
Surprisingly (2.11) does not hold for p = 1 and hence the maximal function Mff does not satisfy the weak type estimate (2.1).4 Namely one can prove the following result that we leave without a proof. Theorem 2.30 (Saks). Let F be the family of all rectangular boxes in Rn that contain the origin and have sides parallel to the coordinate axes. Then the set of functions f ∈ L1(Rn) such that Z lim sup f(x − y) dy = ∞ for all x ∈ Rn diam S→0 S S∈F 1 n is a dense Gδ subset of L (R ). In particular it is not empty.
2.4. The Calder´on-Zygmund decomposition. The following result will play an im- portant role in what follows. Theorem 2.31 (Calder´on-Zygmund decomposition). Suppose f ∈ L1(Rn), f ≥ 0 and α > 0. Then there is an open set Ω and a closed set F such that
(a) Rn = Ω ∪ F , Ω ∩ F = ∅; (b) f ≤ α a.e. in F ; S∞ (c) Ω can be decomposed into cubes Ω = k=1 Qk with pairwise disjoint interiors such that Z α < f(x) dx ≤ 2nα , k = 1, 2, 3,... Qk
Proof. Decompose Rn into a grid of identical cubes, large enough to have Z f(x) dx ≤ α Q for each cube in the grid. Take a cube Q from the grid and divide it into 2n identical cubes. Let Q0 be one of the cubes from this partition. We have two cases: Z Z f(x) dx > α or f(x) dx ≤ α . Q0 Q0 0 If the fist case holds we include the open cube Q to the family {Qk}. Note that Z Z Z α < f = 2n|Q|−1 f ≤ 2n f ≤ 2nα Q0 Q0 Q so the condition (c) is satisfied. If the second case holds we divide Q0 into 2n identical cubes and proceed as above. We continue this process infinitely many times or until it is S∞ terminated. We apply it to all the cubes of the original grid. Let Ω = k=1 Qk, where the cubes are defined by the first case of the process. It remains to prove that f ≤ α a.e. in the set F = Rn \ Ω. The set F consists of faces of the cubes (this set has measure zero) ˜ ˜ and points x such that there is a sequence of cubes Qi with the property that x ∈ Qi, ˜ R R diam Qi → 0, f ≤ α. According to Theorem 2.28 for a.e. such x we have f → f(x) Q˜i Q˜i and hence f ≤ α a.e. in F . −1 Corollary 2.32. Let f, α and Ω be as in Theorem 2.31. Then |Ω| < α kfk1. 4Such estimate would imply (2.11). 14 PIOTR HAJLASZ
Proof. We have ∞ ∞ Z X X −1 −1 |Ω| = |Qk| ≤ α |f| ≤ α kfk1 . k=1 k=1 Qk The proof is complete.
In Remark 2.9 we obtained an upper bound for the measure of the level set of the maximal function which improves the weak type estimate from Theorem 2.2. We will prove now a similar lower bound.
Proposition 2.33. If f ∈ L1(Rn), then 2−nC(n)−1 Z 2 · 5n Z |f(x)| dx ≤ |{x : Mf(x) > t}| ≤ |f(x)| dx. t {|f|>C(n)t} t {|f|>t/2} n/2 where we can take C(n) = ωnn .
Proof. For the proof of the right inequality see (2.4). In order to prove the left inequality we apply the Calder´on-Zygmnud decomposition to the function |f| and α = s > 0. If S Ω = k Qk is an open set as in Theorem 2.31, then Z s < |f(x)| dx ≤ 2ns for all k. Q k √ Let `k be the edge-length of the cube Qk. Then for any x ∈ Qk, Qk ⊂ Bx = B(x, n`k) so Z Z |Bx| n s < |f(y)| dy ≤ |f(y)| dy ≤ ωnn 2 Mf(x). Qk |Qk| Bx Hence −n Z −1 − n X 2 X |{x : Mf(x) > sω n 2 }| ≥ |Q | ≥ |f(x)| dx n k s k Qk 2−n Z ≥ |f(y)| dy s {|f|>s} where the last inequality follows from the fact that |f| ≤ s for almost all x 6∈ Ω which means S n/2 that the set {|f| > s} is contained in Ω = k Qk. Now it suffices to take s = tωnn .
2.5. Integrability of the maximal function. In Remark 2.3 we observed that if f ∈ L1(Rn), then Mf 6∈ L1(Rn). However, we showed that the function cannot be globally integrable. It turns out that, in general, the maximal function need not be even locally integrable. We will actually characterize all functions such that the maximal function is locally integrable. Definition 2.34. We say that a measurable function f belongs to the Zygmund space L log L if |f| log(e + |f|) ∈ L1.
It is easy to see that for a space with finite measure we have \ Lp ⊂ L log L ⊂ L1 , p>1 HARMONIC ANALYSIS 15 so the Zygmund space is an intermediate space between all Lp for p > 1 and L1.
Theorem 2.35 (Stein). Suppose that a measurable function f defined in Rn equals zero outside a ball B. Then Mf ∈ L1(B) if and only if f ∈ L log L(B).
Proof. Suppose that f ∈ L log L(B).5 Cavalieri’s principle (Theorem 2.5) and weak type estimates from Proposition 2.33 give Z Z ∞ Mf(x) dx = |{x ∈ B : Mf(x) > t}| dt B 0 Z ∞ ≤ |B| + |{x ∈ B : Mf(x) > t}| dt 1 Z ∞ 2 · 5n Z ≤ |B| + |f(x)| dx dt 1 t {|f|>t/2} Z Z max{2|f(x)|,1} dt = |B| + C(n) |f(x)| dx B 1 t Z = |B| + C(n) |f(x)| max{0, log(2|f(x)|)} dx B Z ≤ |B| + 2C(n) |f(x)| log(e + |f(x)|) dx < ∞. B
To prove the other implication we assume Mf ∈ L1(B) and we want to prove that f ∈ L log L(B). In the proof of the first implication we used the right inequality from Proposition 2.33, but now we want to use the left inequality. However, the following cal- culations seem to disprove integrability of Mf on B. Z Z ∞ Z ∞ C Z Mf(x) dx = |{x ∈ B : Mf(x) > t}| dx ≥ 1 |f(x)| dx dt B 0 0 t {|f(x)|>C2t} Z Z |f(x)|/C2 dt = C1 |f(x)| dx = +∞. B 0 t | {z } = +∞ if f(x) 6= 0. These calculations are erroneous. Although the function f is supported in the ball B only, the maximal function Mf is positive on all of Rn and in the left estimate in Proposition 2.33 we deal with the measure of the set n {x ∈ R : Mf(x) > t} and not of the set {x ∈ B : Mf(x) > t}. We will show not how to overcome this difficulty.
1 n Note that Mf ∈ Lloc(R ). Indeed, Mf is integrable in B and locally bounded outside the closed ball B, so we need to verify integrability of Mf in a neighborhood of the boundary of B.6 But it is easy to see that if x is near the boundary and outside the ball,
5The proof of integrability of Mf will be similar to the proof of Lp integrability of the maximal function when f ∈ Lp, p > 1. 6 Integrability on B and local boundedness outside B does not imply local integrability in Rn. The function f(x) = 0 for x ≤ 0 and f(x) = 1/x for x > 0 is integrable on (−∞, 0] and is locally bounded on (0, ∞). 16 PIOTR HAJLASZ we can estimate the value of Mf(x) by (constant times) the value of the maximal function in a point being the reflection of x across the boundary. Thus the integrability of Mf near the boundary follows from the integrability of Mf in B. Note also that the set {Mf > 1} is bounded, because Mf(x) decays to zero as x → ∞. Thus local integrability of Mf and boundedness of the set {Mf > 1} implies that the function Mf in integrable in {Mf > 1}. With these remarks we can complete the proof as follows. Z ∞ > Mf(x) dx {Mf>1} Z ∞ ≥ |{Mf > t}| dt 1 Z ∞ 2−nC−1 Z ≥ |f(x)| dx dt 1 t {|f|>Ct} ! Z Z max{|f(x)|/C,1} dt = 2−nC−1 |f(x)| dx B 1 t Z = 2−nC−1 |f(x)| max{0, log(|f(x)|/C)} dx . B This easily implies that f ∈ L log L. Remark 2.36. A more careful analysis leads to the following version of Stein’s theorem. In an open set Ω ⊂ Rn we define the local maximal function by nZ o MΩf(x) = sup |f| : x ∈ Q ⊂ Ω , Q where the supremum is taken over all cubes Q in Ω that contain x.
n 1 Theorem 2.37 (Stein). Let Q ⊂ R be a cube. Then MQf ∈ L (Q) if and only if f ∈ L log L(Q). Moreover Z Z Z −(n+1) |f| n+2 5 MQf ≤ |f| log e + ≤ 2 MQf , Q Q |f|Q Q where Z |f|Q = |f| . Q
p 2.6. The Minkowski integral inequality. Let Fk ∈ L (Y, ν) and denote the variable by y. If we write F (k, y) instead of Fk(y), then the classical Minkowski inequality k|F1| + ... + |Fm|kp ≤ kF1kp + ... + kFmkp can be rewritten as
!1/p 1/p Z X p X Z (2.12) |F (k, y)| dν(y) ≤ |F (k, y)|p dν(y) . Y k k Y Since the sum over k can be regarded as an integral with respect to the counting measure, the following result is a natural generalization of the Minkowski inequality. HARMONIC ANALYSIS 17
Theorem 2.38 (Minkowski’s integral inequality). If µ and ν are σ-finite measures on X and Y respectively and if F : X × Y → R is measurable, then for 1 ≤ p < ∞ we have Z Z p 1/p Z Z 1/p |F (x, y)| dµ(x) dν(y) ≤ |F (x, y)|p dν(y) dµ(x) . Y X X Y Remark 2.39. It would be natural to try approximate F by simple functions in x and apply the discrete version of the inequality (2.12). The proof given below uses however, a very different and much more elegant argument.
Proof. If p = 1 then we actually have equality by Fubini’s theorem. Thus assume that 1 < p R p < ∞. We will use the fact that the L (Y, ν) norm of the function y 7→ X |F (x, y)| dµ(x) equals to the norm of the corresponding functional on Lq(Y, ν), p−1 + q−1 = 1. Z Z p 1/p |F (x, y)| dµ(x) dν(y) Y X Z Z = sup h(y) |F (x, y)| dµ(x) dν(y) h∈Lq(ν) Y X khkq=1 Z Z = sup h(y)|F (x, y)| dν(y) dµ(x) h∈Lq(ν) X Y khkq=1 Z Z 1/q Z 1/p ≤ sup |h(y)|q dν(y) |F (x, y)|p dν(y) dµ(x) h∈Lq(ν) X Y Y khkq=1 | {z } 1 Z Z 1/p = |F (x, y)|p dν(y) dµ(x) . X Y The proof is complete.
2.7. Convergence of averages. In this section we will show that a large class of averages of f converge to f. Our approach will be based on the maximal function estimates. This will lead to far reaching generalizations of the Lebesgue Differentiation Theorem 2.20. Note that in the proofs of Theorem 2.20 and its generalizations discussed above (Theorems 2.28 and 2.29), maximal function also plays an important role. The results of this section will be used later in Section 3.3 Since many averages are constructed with the help of convolution, let us recall that the convolution of measurable functions f and g is defined as Z Z (f ∗ g)(x) = f(x − y)g(y) dy = f(y)g(x − y) dy. Rn Rn The equality of the integrals follows from a linear change of variables. This equality actually means that f ∗ g = g ∗ f. One can also easily check that (f ∗ g) ∗ h = f ∗ (g ∗ h).
Theorem 2.40. Let g ∈ L1(Rn). If f ∈ Lp(Rn), 1 ≤ p < ∞, then f ∗ g ∈ Lp(Rn) and
(2.13) kf ∗ gkp ≤ kfkpkgk1. 18 PIOTR HAJLASZ
n n If f ∈ C0(R ), then f ∗ g ∈ C0(R ) and
(2.14) kf ∗ gk∞ ≤ kfk∞kgk1.
Proof. Let f ∈ Lp(Rn), 1 ≤ p < ∞. The inequality can be obtained from the Minkowski Integral Inequality (Theorem 2.38) as follows Z Z p 1/p
kf ∗ gkp = f(x − y)g(y) dy dx Rn Rn Z Z p 1/p ≤ |f(x − y)| |g(y)| dy dx Rn Rn | {z } |{z} dµ dν Z Z 1/p ≤ |f(x − y)|p dx |g(y)| dy Rn Rn = kfkpkgk1 . n The proof in the case of f ∈ C0(R ) is left to the reader as an exercise. Remark 2.41. In the above proof one could use the classical Minkowski inequality instead of the integral one via the following estimate Z Z |f(x − y)| |g(y)| dy = |f(x − y)| |g(y)|1/p|g(y)|1/q dy Rn Rn Z 1/p Z 1/q ≤ |f(x − y)|p|g(y)| dy |g(y)| dy . Rn Rn We leave details to the reader.
Theorem 2.40 is a special case of a more general Young’s inequality. We leave the proof as an exercise. Theorem 2.42 (Young’s inequality). If 1 ≤ p, q, r ≤ ∞ and q−1 = p−1 + r−1 − 1, then
kf ∗ gkq ≤ kfkpkgkr .
For ϕ ∈ L1(Rn) and t > 0 we define x ϕ (x) = t−nϕ . t t R R Note that n ϕt = n ϕ. R R −1 n −1 Example 2.43. If ϕ = ωn χB(0,1), then ϕt = (ωnt ) χB(0,t) so 1 Z Z f ∗ ϕt(x) = n f(x − y) dy = f(y) dy. ωnt B(0,t) B(x,t) 7 1 1 Thus in that case f ∗ ϕt → f a.e. and in L , provided f ∈ L . Note also that
(2.15) sup(|f| ∗ ϕt)(x) = Mf(x). t>0
7See Lemma 2.19 and Theorem 2.20. HARMONIC ANALYSIS 19
8 Actually this identity plays a crucial role in the proof of the pointwise convergence f ∗ϕt → f a.e. In what follows we will study convergence of f ∗ ϕt for more general functions ϕ. Not surprisingly we will be looking for conditions that will guarantee an estimate similar to (2.15).9
1 n −n p n Theorem 2.44. Let ϕ ∈ L (R ) and let ϕt(x) = t ϕ(x/t) for t > 0. If f ∈ L (R ), n 1 ≤ p < ∞ or f ∈ C0(R ) and p = ∞, then Z (2.16) lim kf ∗ ϕt − afkp = 0, where a = ϕ(x) dx. t→0+ Rn Remark 2.45. The two most interesting cases are when a = 1 and when a = 0.
n n 10 Remark 2.46. If f ∈ C0(R ), then f ∗ ϕt ∈ C0(R ) and (2.16) with p = ∞ means that f ∗ ϕt converges to af uniformly. p Remark 2.47. If a = 1, then (2.16) means that f ∗ ϕt → f in L so the operators f 7→ f ∗ ϕt converge pointwise, i.e. for every f, to the identity operator f 7→ f. For that reason the family of functions {ϕt : t > 0} is often called an approximation of identity (provided a = 1).
Proof. We will prove the result when f ∈ Lp(Rn), 1 ≤ p < ∞, leaving the case of f ∈ n C0(R ) to the reader. We have Z
|f ∗ ϕt(x) − af(x)| = (f(x − y) − f(x)) ϕt(y) dy . Rn Given δ > 0, the Minkowski Integral Inequality yields Z Z 1/p p kf ∗ ϕt − afkp ≤ |f(x − y) − f(x)| dx |ϕt(y)| dy Rn Rn Z = kτ−yf − fkp|ϕt(y)| dy Rn Z = kτ−tyf − fkp|ϕ(y)| dy, Rn where the last equality follows from a simple linear change of variables.
1 n Note that kτ−tyf −fkp|ϕ(y)| ≤ 2kfkp|ϕ(y) ∈ L and for every y ∈ R , kτ−tyf −fkp → 0 as t → 0 (Lemma 2.18) so Z kf ∗ ϕt − afkp ≤ kτ−tyf − fkp|ϕ(y)| dy → 0 as t → 0 Rn by the Dominated Convergence Theorem.
Under assumptions of Theorem 2.44, if f ∈ Lp(Rn), 1 ≤ p < ∞, there is a sequence ti → 0 such that f ∗ ϕti → af a.e. as i → ∞, but a more challenging question is to find conditions that would guarantee f ∗ ϕt → af a.e. as t → 0. That would be a true generalization of the Lebesgue Differentiation Theorem.
8Along with weak type estimates for the maximal function, see Theorem 2.20. 9See Theorem 2.49. 10See Theorem 2.40. 20 PIOTR HAJLASZ
Definition 2.48. Let ϕ ∈ L1(Rn). We say that Ψ is a radially decreasing majorant of ϕ if (a) Ψ(x) = η(|x|) for some11 η : [0, ∞) → [0, ∞].12 (b) η is decreasing.13. (c) |ϕ(x)| ≤ Ψ(x) a.e.
Every function ϕ ∈ L1 has the least radially decreasing majorant14
Ψ0(x) = ess sup |ϕ(y)|. |y|≥|x| Thus the existence of an integrable radially decreasing majorant Ψ of ϕ is equivalent to 1 Ψ0 ∈ L . Theorem 2.49. Suppose that ϕ ∈ L1(Rn) has an integrable radially decreasing majorant 1 n Ψ(x) = η(|x|) ∈ L (R ) . 1 n n Then for f ∈ Lloc(R ) and all x ∈ R
(2.17) sup(f ∗ ϕt)(x) ≤ kΨk1Mf(x) . t>0 If in addition f ∈ Lp( n), 1 ≤ p < ∞, and R ϕ(x) dx = a, then R Rn
(2.18) f ∗ ϕt → af a.e. as t → 0.
Proof. In order to prove (2.17) it suffices to prove that for a radially decreasing and inte- grable function Ψ we have
(2.19) |f| ∗ Ψ(x) ≤ kΨk1Mf(x).
Indeed, since Ψt is also radially decreasing and integrable, applying (2.19) to Ψt will give
|f ∗ ϕt(x)| ≤ |f| ∗ Ψt(x) ≤ kΨtk1Mf(x) = kΨk1Mf(x). Thus it remains to prove (2.19). For a positive integer k let
k·2k k·2k X 1 X 1 s (x) = χ k (x) = χ (x) . k 2k {Ψ≥i/2 } 2k Bi i=1 i=1 Since Ψ is radially symmetric the sets n i o B = x : Ψ(x) ≥ i 2k are balls15 centered at zero.
11Functions of this form, i.e. functions constant on spheres Sn−1(0, r) are called radially symmetric. 12It may happen that η(t) = +∞ for all t. 13i.e. non-increasing 14 However, it may happen that Ψ0 = +∞ everywhere. 15Open or closed – find an example such that for some t the balls are open while for other values of t they are closed. HARMONIC ANALYSIS 21
The sequence {sk} is increasing and convergent to Ψ almost everywhere. Indeed, for 16 x 6= 0, Ψ(x) < ∞ so Ψ(x) < k0, for some k0 ∈ N. Then for k ≥ k0 ` ` + 1 ≤ Ψ(x) < , for some ` = 0, 1, 2, . . . k · 2k − 1. 2k 2k Since the inequality Ψ ≥ i/2k is satisfied for i = 1, 2, . . . , `, the definition of the function k sk gives that sk(x) = `/2 so 1 Ψ(x) − < s (x) ≤ Ψ(x). 2k k
That means sk(x) → Ψ(x) for all all x 6= 0. Observe also that the sequence {sk} is k k k+1 increasing, sk ≤ sk+1. Indeed, if sk(x) = `/2 , then Ψ(x) ≥ `/2 = (2`)/(2 ) so the definition of sk+1 gives 1 ` s (x) ≥ · 2` = = s (x). k+1 2k+1 2k k Thus if we can prove
(2.20) |f| ∗ sk(x) ≤ kΨk1Mf(x), inequality (2.19) will follow from the Monotone Convergence Theorem. We have k·2k X |Bi| χB |f| ∗ s (x) = |f| ∗ i (x). k 2k |B | i=1 i Since17 Z k·2k χB X |Bi| |f| ∗ i (x) = |f(x − y)| dy ≤ Mf(x) and = ks k ≤ kΨk , |B | 2k k 1 1 i Bi i=1 inequality (2.20) follows. The proof of (2.17) is complete. It remains now to prove convergence (2.18). If p = 1, then
(2.21) Ωf(x) := lim sup f ∗ ϕt − lim inf f ∗ ϕt ≤ 2kΨk1Mf(x) t→0+ t→0+ and almost the same argument as in the proof of Theorem 2.20 yields that18 Ωf = 0 a.e. 1 so f ∗ ϕt converges a.e. to a measurable function. Since f ∗ ϕt → af in L (Theorem 2.44), (2.18) follows.
The case 1 < p < ∞ uses a similar argument. If f ∈ Lp(Rn), 1 < p < ∞, then Chebyschev’s inequality (2.2) and Theorem 2.2(b) yield Z p p 1 p |{x : Mf(x) > t}| = |{x : |Mf(x)| > t }| ≤ p |Mf(x)| dx t Rn Z C(n, p) p ≤ p |f(x)| dx. t Rn 16Because Ψ ∈ L1. 17 Because the balls Bi are centered at the origin. 18We will actually repeat this argument in the case of 1 < p < ∞. 22 PIOTR HAJLASZ
Hence for ε > 0, inequality (2.21) yields Z C p |{x :Ωf(x) > ε}| ≤ p |f(x)| dx, ε Rn n 19 where the constant C depends on n, p and kΨk1. If h ∈ C0(R ), then Ωh=0. Taking n p p+1 h ∈ C0(R ) such that kf − hkp ≤ ε we conclude that |{x :Ωf(x) > ε}| = |{x : Ω(f − h)(x) > ε}| < Cε so Ωf = 0 a.e. and exactly as in the case of p = 1, (2.18) follows.
19 Because h ∗ ϕt converges uniformly (and hence pointwise) to h. HARMONIC ANALYSIS 23
3. The Fourier transform
3.1. Fourier transform.
Definition 3.1. For f ∈ L1(Rn) we define the Fourier transform as Z n ˆ −2πix·ξ X F(f)(x) = f(ξ) = f(x)e dx where x · ξ = xjξj . n R j=1 Theorem 3.2. The Fourier transform has the following properties
∧ 1 n n ˆ (a) : L (R ) → C0(R ) is a bounded linear operator with kfk∞ ≤ kfk1.
(b) (f ∗ g)ˆ = fˆgˆ for f, g ∈ L1(Rn).
(c) If f, g ∈ L1(Rn), then fg,ˆ fgˆ ∈ L1(Rn) and Z Z fˆ(x)g(x) dx = f(x)ˆg(x) dx. Rn Rn (d) If τhf(x) = f(x + h), then ˆ 2πih·ξ 2πih·x ˆ (τhf)ˆ(ξ) = f(ξ)e and (f(x)e )ˆ(ξ) = f(ξ − h). −n (e) If ft(x) = t f(x/t), then ˆ ˆ (ft)ˆ(ξ) = f(tξ) and (f(tx))ˆ(ξ) = (f)t(ξ) . (f) If ρ ∈ O(n) is an orthogonal transformation, then (f(ρ ·))ˆ(ξ) = fˆ(ρξ).
∧ 1 n ∞ n ˆ Proof. (a) Clearly, : L (R ) → L (R ) is a bounded linear mapping with kfk∞ ≤ kfk1. Indeed, Z ˆ −2πix·ξ |f(ξ)| ≤ f(x)e dx = kfk1. Rn The Dominated Convergence Theorem implies that the function fˆ is continuous. It remains to prove that fˆ(ξ) → 0 as |ξ| → ∞. Let ξ 6= 0. Since eπi = −1 we have Z Z fˆ(ξ) = f(x)e−2πix·ξ dx = − f(x)e−2πix·ξeπi dx Rn Rn Z ξ = − f(x) exp − 2πi x − 2 · ξ dx Rn 2|ξ| 24 PIOTR HAJLASZ Z ξ −2πix·ξ = − f x + 2 e dx . Rn 2|ξ| Hence Z ˆ 1 ξ −2πix·ξ f(ξ) = f(x) − f x + 2 e dx 2 Rn 2|ξ| and thus Lemma 2.18 yields 1 |fˆ(ξ)| ≤ f − τ ξ f → 0 as |ξ| → ∞. 2 2|ξ|2 1 (b) Z Z (f ∗ g)ˆ(ξ) = f(x − y)g(y) dy e−2πix·ξ dx Rn Rn Z Z = f(x − y)e−2πi(x−y)·ξ dx g(y)e−2πiy·ξ dy Rn Rn = fˆ(ξ)ˆg(ξ) . (c) fg,ˆ fgˆ ∈ L1, because the functions f,ˆ gˆ are bounded and the equality of the integrals easily follows from the Fubini theorem. (d) Z Z −2πix·ξ −2πi(x−h)·ξ (τhf)ˆ(ξ) = f(x + h)e dx = f(x)e dx Rn Rn Z = e2πih·ξ f(x)e−2πix·ξ dx = e2πih·ξfˆ(ξ) . Rn The second equality follows from a similar argument. (e) Z Z −n x −2πix·ξ −2πi(ty)·ξ (ft)ˆ(ξ) = t f e dx = f(y)e dy Rn t Rn Z = f(y)e−2πiy·(tξ) dy = fˆ(tξ) . Rn The second formula follows from the first one if we replace t by t−1. (f)20
Z Z (f(ρ ·))ˆ(ξ) = f(ρx)e−2πix·ξ dx = f(y)e−2πi(ρ−1y)·ξ dy Rn Rn Z = f(y)e−2πiy·(ρξ) dy = fˆ(ρξ) . Rn Equality (ρ−1y) · ξ = y · (ρξ) follows from the fact that the mapping x 7→ ρx is an isometry. 20This property is in some sense obvious: the Fourier transform does not depend on the choice of the coordinate system since it depends on the scalar product x · ξ. Thus the Fourier transform should stay the same if we rotate the coordinate system by ρ. HARMONIC ANALYSIS 25
1 n 1 n Theorem 3.3. Suppose f ∈ L (R ) and xkf(x) ∈ L (R ), where xk is the k-th coordinate ˆ function. Then f is differentiable with respect to ξk and ∂fˆ (−2πixkf(x))ˆ = (ξ) . ∂ξk
Proof. Let ek be the unit vector along the k-th coordinate. Then the second part of The- orem 3.2(d) gives ∧ fˆ(ξ + he ) − fˆ(ξ) e−2πi(hek)·x − 1 k = f(x) (ξ) → (−2πix f(x))ˆ(ξ) . h h k 1 The convergence follows from the continuity of the Fourier transform in L . 1 n p Definition 3.4. We say that f ∈ Lloc(R ) is L -differentiable with respect to xk if there is g ∈ Lp(Rn) such that Z p f(x + hek) − f(x) − g(x) dx → 0 as h → 0. Rn h p The function g is called the L -partial derivative of f with respect to xk and is denoted by g = ∂f/∂xk. 1 n 1 Theorem 3.5. If f ∈ L (R ) is L -differentiable with respect to xk, then ∧ ∂f ˆ = 2πiξkf(ξ) . ∂xk
Proof. The first part of Theorem 3.2(d) and the L1-differentiability of f give ∧ ∧ ∂f e2πi(hek)·ξ − 1 ∂f f(x + he ) − f(x) − fˆ(ξ) = − k → 0 ∂xk h ∂xk h as h → 0, so ∧ 2πi(hek)·ξ ∂f ˆ e − 1 ˆ (ξ) = lim f(ξ) = 2πiξkf(ξ) . ∂xk h→0 h The proof is complete. Remark 3.6. With each polynomial X α P (x) = aαx |α|≤m of variables x1, . . . , xn we associate a differential operator |α| X α X ∂ P (D) = aαD = aα α . ∂x 1 . . . ∂xαn |α|≤m |α|≤m 1 n Then under suitable assumptions Theorems 3.3 and 3.5 have the following higher order generalizations (3.1) P (D)fˆ(ξ) = (P (−2πix)f(x))ˆ(ξ), (P (D)f)ˆ(ξ) = P (2πiξ)fˆ(ξ).
The next result is our first (and actually one of the most important) example of the Fourier transform. 26 PIOTR HAJLASZ
Theorem 3.7. If f(x) = e−π|x|2 , then fˆ(ξ) = e−π|ξ|2 .
Proof. First observe that it suffices to prove the result in dimension 1. Indeed, assuming it for n = 1, the case of general n follows from Fubini’s theorem Z n Z ∞ n 2 Y 2 Y 2 2 ˆ −π|x| −2πix·ξ −πxk −2πixkξk −πξk −π|ξ| f(ξ) = e e dx = e e dxk = e = e . n R k=1 −∞ k=1 In the case n = 1 we will show two different proofs. The first one will use the contour integration while the second one will use differential equations. Thus in the remaining part of the proof we will assume that f(x) = e−πx2 is a function of one variable. Proof by contour integration. The function e−πz2 is holomorphic and hence its integral along the following curve equals zero.
Letting R → ∞ we obtain Z ∞ Z ∞ e−πx2 dx = e−π(x+iξ)2 dx . −∞ −∞ The left hand side equals 1, so Z ∞ 1 = e−πx2 e−2πixξeπξ2 dx . −∞ Hence Z ∞ e−πξ2 = e−πx2 e−2πixξ dx = fˆ(ξ) . −∞ The proof is complete. Proof by differential equations. We have f 0(x) = −2πxf(x), −i(−2πixf(x)) = f 0(x) so Theorems 3.3 and 3.5 yield ∧ 0 −i − 2πixf(x) (ξ) = fb0(ξ) , (fˆ) (ξ) = −2πξfˆ(ξ). | {z } | {z } (fˆ)0(ξ) 2πiξfˆ(ξ) HARMONIC ANALYSIS 27
Solving this differential equation for the unknown function ξ 7→ fˆ(ξ) yields Z ∞ fˆ(ξ) = fˆ(0)e−πξ2 = e−πξ2 , because fˆ(0) = e−πx2 dx = 1. −∞
3.2. Measures of finite total variation. Let us first recall basic facts regarding mea- sures of finite total variation.
If µ is a complex (Borel) measure on Rn, then there is a unique positive measure |µ| called the total variation of measure µ and a Borel function h : Rn → C, |h(x)| = 1 for all x ∈ Rn such that Z n µ(E) = h(x) d|µ|(x) for all Borel sets E ⊂ R . E The space B(Rn) of complex measures of finite total variation kµk := |µ|(Rn) is a Banach space with the norm k · k.
Note that f ∈ L1(Rn) defines a measure of finite total variation by Z µ(E) = f(x) dx. E In that case Z Z |µ|(E) = |f(x)| dx so kµk = |f(x)| dx = kfk1 < ∞. E Rn This proves that L1(Rn) is a closed subspace of B(Rn). In a special case when µ is a real valued measure, called signed measure, we have h : n + − R → {±1} so if we define h = hχ{h=1} and h = hχ{h=−1}, then Z µ±(E) = h±(x) d|µ|(x) E are positive measures such that (3.2) µ = µ+ − µ− and |µ| = µ+ + µ−. The representation (3.2) is called the Hahn decomposition of µ. Note that the measures µ+ and µ− are supported on disjoint sets.
n Theorem 3.8 (Riesz representation theorem). The dual space to C0(R ) is isometri- cally isomorphic to the space of measures of finite total variation. More precisely, if n ∗ Φ ∈ (C0(R )) , then there is a unique measure µ of finite total variation such that Z n Φ(f) = f dµ for f ∈ C0(R ). Rn Moreover kΦk = kµk = |µ|(Rn). 28 PIOTR HAJLASZ
This result allows us to define convolution of measures.
If f, g ∈ L1(Rn), then f ∗ g ∈ L1(Rn) ⊂ B(Rn) and hence it acts as a functional on n C0(R ) by the formula Z Z Z Φ(h) = h(x)(f ∗ g)(x) dx = h(x) f(x − y)g(y) dy dx Rn Rn Rn Z Z = h(x)f(x − y) dx g(y) dy Rn Rn Z Z = h(x + y)f(x)g(y) dx dy . Rn Rn This suggests how to define convolution of measures.
n If µ1, µ2 ∈ B(R ), then Z Z Φ(h) = h(x + y) dµ1(x) dµ2(y) Rn Rn n n defines a functional on C0(R ) and hence there is a unique measure µ ∈ B(R ) such that Z Z Z n h(x + y) dµ1(x) dµ2(y) = h(x) dµ(x) for all h ∈ C0(R ). Rn Rn Rn Definition 3.9. We denote the measure µ by
µ = µ1 ∗ µ2 and call it convolution of measures.
Clearly
µ1 ∗ µ2 = µ2 ∗ µ1 and kµ1 ∗ µ2k ≤ kµ1k kµ2k.
If dµ1 = f dx, dµ2 = g dx, then
d(µ1 ∗ µ2) = (f ∗ g) dx, so the convolution of measures extends the notion of convolution of functions. If dµ1 = f dx and µ ∈ B(Rn), then Z Z Z Z h(x + y) dµ1(x) dµ(y) = h(x + y)f(x) dx dµ(y) Rn Rn Rn Rn Z Z = h(x)f(x − y) dx dµ(y) Rn Rn Z Z = h(x) f(x − y) dµ(y) dx. Rn Rn Thus µ1 ∗ µ can be identified with a function Z 1 n x 7→ f(x − y) dµ(y) ∈ L (R ), Rn so we can write Z f ∗ µ = f(x − y) dµ(y), kf ∗ µk1 ≤ kfk1 kµk. Rn HARMONIC ANALYSIS 29
Theorem 3.10. If 1 ≤ p < ∞, f ∈ Lp(Rn) and µ ∈ B(Rn), then
kf ∗ µkp ≤ kfkp kµk.
Proof is almost the same as that for Theorem 2.40 as we leave it to the reader.
Definition 3.11. The Fourier transform of a measure of finite total variation µ ∈ B(Rn) is defined by Z µˆ(x) = e−2πix·ξ dµ. Rn Proposition 3.12. If µ ∈ B(Rn), then µˆ is a bounded and continuous function such that kµˆk∞ ≤ kµk.
Proof. Boundedness ofµ ˆ is obvious and continuity follows from the Dominated Conver- gence Theorem. n Proposition 3.13. If µ1, µ2 ∈ B(R ), then µ\1 ∗ µ2 =µ ˆ1 · µˆ2.
We leave the proof as an exercise. Remark 3.14. Convolution of measures and the Fourier transform of measures play a fundamental role in probability.
3.3. Summability methods. The answer to an important question of how to reconstruct a function f from its Fourier transform fˆ is provided by the inversion formula21 Z (3.3) f(x) = fˆ(ξ) e2πix·ξ dξ . Rn It seems that this formula requires both f and fˆ to be integrable. Integrability of f is needed for the existence of fˆ and integrability of fˆ is needed for the integral in (3.3) to make sense. We will actually prove in Corollary 3.26 that if f, fˆ ∈ L1, then (3.3) is true a.e. There is, however, a problem here. Equality (3.3) means that f(x) is the Fourier transform of fˆ ∈ L1 evaluated at −x. Since the Fourier transform of an integrable function fˆ ∈ L1 n n 1 ˆ belongs to C0(R ), it follows that f ∈ C0(R ). That means, if f ∈ L \ C0, then f cannot be integrable so the right hand side of (3.3) does not make much sense. Is there any chance to give meaning to the integral at (3.3), beyond the class of integrable Fourier transforms so that the equality at (3.3) would be true for all f ∈ L1? The answer is in the positive but we need to interpret the integral at (3.3) using suitable summability methods.22
21Compare it with the formula for Fourier series: Z 1 X fˆ(n) = f(x)e−2πixn dx, f(x) = fˆ(n)e2πixn. 0 n∈Z
22An example of a summability method for sequences is given by the so called Ces`aro method: with each sequence {an} we associate the limit of (a1 + ... + an)/n. This extends the notion of limit far beyond the class of convergent sequences. 30 PIOTR HAJLASZ
n If Φ ∈ C0(R ) is such that Φ(0) = 1, then for t > 0 we define the Φ-mean of a function f as Z MΦ,tf = Mtf = f(x)Φ(tx) dx. Rn 1 n R + If f ∈ L ( ), then MΦ,tf → n f(x) dx as t → 0 , but the limit of Mtf might also exist R R for some non-integrable functions f. R Definition 3.15. We say that n f is Φ-summable to ` ∈ if lim + MΦ,tf = `. R R t→0
In the case of Φ(x) = e−|x| and Φ(x) = e−|x|2 we obtain the so called the Abel and the Gauss summability methods. More precisely we have Definition 3.16. For each t > 0 the Abel mean and the Gauss mean of a function f are Z Z −t|x| −t|x|2 At(f) = f(x)e dx and Gt(f) = f(x) e dx . Rn Rn R 23 We say that n f(x) dx is Abel summable (Gauss summable) to ` ∈ if limt→0 At(f) = ` R R (limt→0 Gt(f) = `).
−|x|2 −t2|x|2 Remark 3.17. If Φ(x) = e , then Φ(tx) = e so MΦ,tf = Gt2 (f). This however, does not cause any problems when you compare notions of the Φ-summability and the Gauss summability because limt→0+ MΦ,tf = limt→o Gt2 (f). Remark 3.18. Clearly, in the case of integrable functions, the Abel and the Gauss methods give the usual integral. However, the means At(f) and Gt(f) are finite whenever f is bounded so for some bounded, but non-integrable f, the integral R f(x) dx might still Rn 1 n n be Abel or Gauss summable to a finite value. For example if f ∈ L (R ) \ C0(R ), the ˆ 2πix·ξ n the function ξ 7→ f(ξ)e is in C0(R ) but it is not integrable. However, as we will see (Theorems 3.23 and 3.30) for almost all x ∈ n, the integral R fˆ(ξ)e2πix·ξ dξ is both Abel R Rn and Gauss summable to f(x). This gives some meaning of the inversion formula (3.3) for generic f ∈ L1(Rn).
We want to find a reasonably large class of Φ-means (including the Abel and the Gauss means) such that Z fˆ(ξ)e2πix·ξ dξ Rn is Φ-summable to f(x) i.e. Z ˆ 2πix·ξ + Mt(x) = f(ξ)e Φ(tξ) dξ → f(x) as t → 0 Rn 1 and we can consider both, convergence Mt → f almost everywhere and in L (with respect to variable x). The main result reads as follows Theorem 3.19. Let Φ ∈ L1( n) be such that ϕ = Φˆ ∈ L1( n) and R ϕ(x) dx = 1. For R R Rn f ∈ L1(Rn) let Z ˆ 2πix·ξ Mt(x) = f(ξ)e Φ(tξ) dξ Rn 23 R a R ∞ −tx Exercise. Prove that if lima→∞ 0 f(x) dx = `, then At = 0 f(x)e dx converges to ` as t → 0. HARMONIC ANALYSIS 31 be the Φ-mean of the integral (3.3). Then 1 (3.4) Mt → f in L as t → 0. If in addition ϕ = Φˆ has an integrable radially decreasing majorant, then
(3.5) Mt → f a.e. as t → 0.
Proof. The lemma below is a link between the Φ-mean and the approximation by convo- lution discussed in Section 2.7. Lemma 3.20. If f, Φ ∈ L1(Rn) and ϕ = Φˆ, then Z Z ˆ 2πix·ξ f(ξ)e Φ(tξ) dξ = f(x − y)ϕt(−y) dy = f ∗ ϕ˜t(x), where ϕ˜(y) = ϕ(−y). Rn Rn
Proof. For h ∈ Rn define (h) 2πix·h 1 n g (x) = e Φ(tx) ∈ L (R ) . Recall that24 2πih·x ∧ ∧ (w(x)e ) (ξ) =w ˆ(ξ − h) and (w(tx)) (ξ) = (w ˆ)t(ξ). Hence (h) ∧ ˆ (g ) (ξ) = (Φ)t(ξ − h) = ϕt(ξ − h) so replacing h by x we get (x) ∧ (g ) (ξ) = ϕt(ξ − x). We have Z Z Z fˆ(ξ)e2πix·ξΦ(tξ) dξ = fˆ(ξ)g(x)(ξ) dξ = f(ξ)(g(x))∧(ξ) dξ Rn Rn Rn Z Z = f(ξ)ϕt(ξ − x) dξ = f(x − y)ϕt(−y) dx. Rn Rn We used here Theorem 3.2(c) and in the last equality the change of variables −y = ξ−x.
Now (3.4) follows directly from Theorem 2.44 and (3.5) follows from Theorem 2.49. Remark 3.21. Theorem 3.19 is not easy to apply since given Φ ∈ L1, we have to compute Φˆ and show its integrability. In the case of the Gauss mean Φ(x) = e−|x|2 it can be easily done with the help of Theorem 3.7, but the case of the Abel mean Φ(x) = e−|x| is much harder due to difficulty of computing the Fourier transform of Φ(x) = e−|x|.
In Theorems 3.23 and 3.30 we will investigate the case of the the Gauss and the Abel and methods. Theorem 3.22. Let f(x) = e−4π2t|x|2 , t > 0. Then
(a) W (x, t) := fˆ(x) = (4πt)−n/2e−|x|2/(4t). (b) The function W has the following scaling property with respect to t: if ϕ(x) = W (x, 1), then W (x, t) = ϕt1/2 (x).
24see Theorem 3.2. 32 PIOTR HAJLASZ
(c) Z W (x, t) dx = 1 for all t > 0. Rn Proof. We can write f(x) = e−4π2t|x|2 = u((4πt)1/2x), where u(x) = e−π|x|2 . ∧ ˆ −π|ξ|2 25 Since (ψ(tx)) (ξ) = (ψ)t(ξ) andu ˆ(ξ) = u(ξ) = e we obtain ˆ −n/2 −|ξ|2/(4t) W (ξ, t) = f(ξ) = (ˆu)(4πt)1/2 (ξ) = u(4πt)1/2 (ξ) = (4πt) e which is (a). In particular
ϕ(ξ) = W (ξ, 1) = u(4π)1/2 (ξ).
Since (ψt1 )t2 (x) = ψt1t2 (x) we get W (ξ, t) = u 1/2 (ξ) = u 1/2 (ξ) = ϕ 1/2 (ξ) (4πt) (4π) t1/2 t R R which is (b). Finally equality n ψt = n ψt , t1, t2 > 0 along with W (ξ, t) = ϕ 1/2 (ξ) R 1 R 2 t yields (c): Z Z 1 Z 2 W (ξ, t) dξ = W ξ, dξ = e−π|ξ| dξ = 1. Rn Rn 4π Rn
As an application we obtain
Theorem 3.23 (The Gauss-Weierstrass summability method). If f ∈ L1(Rn), then Z Z (3.6) fˆ(ξ)e2πix·ξe−4π2t|ξ|2 dξ = f(y)W (x − y, t) dy → f as t → 0+ Rn Rn both in L1(Rn) and almost everywhere.
Proof. First we will prove equality. Let Φ(x) = e−4π2|x|2 . Using notation from Theorem 3.22 ˆ Φ(x) = W (x, 1) = ϕ(x). Since ϕ(x) = ϕ(−x),ϕ ˜t1/2 (x) = ϕt1/2 (x) = W (x, t). Thus Lemma 3.20 yields Z Z fˆ(ξ)e2πix·ξe−4π2t|ξ|2 dξ = fˆ(ξ)e2πix·ξΦ(t1.2ξ) dξ Rn Rn Z = f ∗ ϕ˜t1/2 (x) = W (x − y, t)f(y) dy. Rn Theorem 3.22 shows that the function Φ satisfies the assumptions of Theorem 3.19 so both 1 convergences in L and almost everywhere follow.
Remark 3.24. The function W (x, t) = ϕt1/2 (x) satisfies the assumptions of Theorem 2.49 with a = 1 so for f ∈ Lp, 1 ≤ p < ∞ Z (3.7) f(y)W (x − y, t) dy → f a.e. and in Lp as t → 0+. Rn 25The function W is defined as a Fourier transform so it will be more convenient for us to prove (a)-(c) with the variable x replaced by ξ. HARMONIC ANALYSIS 33
Remark 3.25. W (x, t) is called the Gauss-Weierstrass kernel. Under suitable assumptions about f, differentiation under the sign of the integral shows that the function
Z 1 Z |x−y|2 − 4t w(x, t) = W (x − y, t)f(y) dy = n/2 e f(y) dy Rn (4πt) Rn ∂w n+1 satisfies the heat equation ∂t = ∆xw in the interior of R+ . As we showed, the function w(x, t) converges to f(x) as t → 0, so w is a solution to ∂w n+1 ∂t = ∆xw on R+ , w(x, 0) = f(x), x ∈ Rn. Corollary 3.26. If both f and fˆ are integrable26, then Z (3.8) f(x) = fˆ(ξ)e2πix·ξ dξ a.e. Rn
Proof. Since fˆ ∈ L1, the left hand side of (3.6) converges to R fˆ(ξ)e2πix·ξ dξ so the result Rn follows from Theorem 3.23.
The above inversion formula motivates the following definitions. Definition 3.27. The inverse Fourier transform is defined by Z fˇ(x) = F −1(f)(x) = f(ξ)e2πix·ξ dξ. Rn 1 n ˆ ˆ n Corollary 3.28. If f1, f2 ∈ L (R ) and f1 = f2 on R , then f1 = f2 a.e.
1 ˆ 1 Proof. Let f = f1 − f2. Then f ∈ L and f = 0 ∈ L so (3.8) yields that f = 0 a.e.
The following result provides another example of a function that satisfies the assumptions of Theorem 3.19. Theorem 3.29. Let f(x) = e−2π|x|t, t > 0. Then
(a) t P (x, t) := fˆ(x) = c , n (t2 + |x|2)(n+1)/2 where Γ n+1 c = 2 . n π(n+1)/2 (b) The function P has the following scaling property with respect to t: if ϕ(x) = P (x, 1), then P (x, t) = ϕt(x). (c) Z P (x, t) dx = 1 for all t > 0. Rn By the same arguments as before we obtain.
26The assumption about integrability of fˆ is very strong. As already observed, the equality (3.8) implies n that f equals a.e. to a function in C0(R ). 34 PIOTR HAJLASZ
Theorem 3.30 (The Abel summability method). If f ∈ L1(Rn), then Z Z fˆ(ξ)e2πix·ξe−2π|ξ|t dξ = f(y)P (x − y, t) dy → f as t → 0+ Rn Rn both in L1(Rn) and almost everywhere.
Remark 3.31. The function P (x, t) = ϕt(x) satisfies the assumptions of Theorem 2.49 with a = 1 so for f ∈ Lp, 1 ≤ p < ∞ Z f(y)P (x − y, t) dy → f a.e. and in Lp as t → 0+. Rn Remark 3.32. P (x, t) is called the Poisson kernel. Under suitable assumptions about f, differentiation under the sign of the integral shows that the function Z u(x, t) = P (x − y, t)f(y) dy Rn n+1 satisfies the Laplace equation ∆(x,t)u = 0 in the interior of R+ . As we showed u(x, t) converges to f as t → 0+ so u(x, t) is a solution to the Dirichlet problem in the half-space n+1 ∆(x,t)u = 0 on R+ , u(x, 0) = f(x), x ∈ Rn
The proof of Theorem 3.29 is substantially more difficult than that of Theorem 3.22. Since the formula for the Fourier transform involves the Γ function we need to recall its basic properties. Definition 3.33. For 0 < x < ∞ we define Z ∞ Γ(x) = tx−1e−t dt . 0 Theorem 3.34.
(a) Γ(x + 1) = xΓ(x) for all 0 < x < ∞. (b) Γ(n + 1) =√n! for all n = 0, 1, 2, 3,... (c) Γ(1/2) = π.
Proof. (a) follows from the integration by parts. Since Γ(1) = 1, (b) follows from (a) by induction. The substitution t = s2 gives Z ∞ Z ∞ Z ∞ Γ(x) = tx−1e−t dt = s2(x−1)e−s2 2s ds = 2 s2x−1e−s2 ds 0 0 0 and hence ∞ 1 Z 2 √ Γ = 2 e−s ds = π . 2 0
The formula Γ(x) = Γ(x + 1)/x allows us to define Γ(x) for all negative non-integer values of x. For example √ 3 Γ(−1/2) Γ(1/2) 4 π Γ − = = = . 2 −3/2 (−3/2)(−1/2) 3 HARMONIC ANALYSIS 35
Definition 3.35. For x ∈ (−∞, 0) \ {−1, −2, −3,...}, the function Γ(x) is defined by Γ(x + k) Γ(x) = , x(x + 1) · ... · (x + k − 1) where k is a positive integer such that x + k ∈ (0, 1).
Lemma 3.36. √ Z π/2 n+1 n πΓ 2 sin θ dθ = n for n = 1, 2, 3,... 0 nΓ 2
Proof. Denote the left hand side by an and the right hand side by bn. Easy one time integration by parts27 shows that n + 1 a = (n + 1)(a − a ), a = a . n+2 n n+2 n+2 n + 2 n Also elementary properties of the Γ function show that n + 1 b = b n+2 n + 2 n and now it is enough to observe that a1 = 1 = b1 = 1, a2 = π/4 = b2. Lemma 3.37.
(a) The volume of the unit ball in Rn equals 2πn/2 πn/2 (3.9) ω = = . n nΓ(n/2) n Γ 2 + 1 n (b) The (n − 1)-dimensional measure of the unit sphere in R equals nωn
Proof.
It follows from the picture and the Fubini theorem that the volume of the upper half of the ball equals Z 1 1 n−1 ωn = ωn−1r(h) dh . 2 0 The substitution h = 1 − cos θ, dh = sin θ dθ, r(h) = sin θ
27Use sinn+2 Θ = (− cos θ)0 sinn+1 Θ. 36 PIOTR HAJLASZ and Lemma 3.36 give 1 Z π/2 π1/2Γ n+1 ω = ω sinn θ dθ = ω 2 , 2 n n−1 n−1 n 0 nΓ 2 so 1/2 n+1 2π Γ 2 ωn = n ωn−1 . nΓ 2 If 2πn/2 an = n nΓ 2 then a direct computation shows that a1 = 2 = ω1 and that an satisfies the same recur- rence relationship as ωn, so ωn = an for all n. The second equality in (3.9) follows from Theorem 3.34(a). (b) follows from the fact that the (n−1)-dimensional measure of a sphere of radius r equals to the derivative with respect to r of the volume of an n-dimensional ball of radius r. 2 We will also need the following result. Lemma 3.38. For β > 0 we have ∞ −u 1 Z e 2 e−β = √ √ e−β /(4u) du . π 0 u
Proof. Applying the theory of residues to the function eiβz/(1 + z2) one can easily prove that Z ∞ −β 2 cos βx e = 2 dx . π 0 1 + x This and an obvious identity Z ∞ 1 −(1+x2)u 2 = e du 1 + x 0 yields Z ∞ −β 2 cos βx e = 2 dx π 0 1 + x ∞ ∞ 2 Z Z 2 = cos βx e−ue−ux du dx π 0 0 ∞ ∞ 2 Z Z 2 = e−u e−ux cos βx dx du π 0 0 ∞ ∞ 2 Z 1 Z 2 = e−u e−ux eiβx dx du π 0 2 −∞ ∞ ∞ (x=−2πy) 2 Z Z 2 2 = e−u π e−4π uy e−2πiβy dy du π 0 −∞ | {z } W (β,u) ∞ r 2 Z 1 π 2 = e−u e−β /(4u) du π 0 2 u HARMONIC ANALYSIS 37
∞ −u 1 Z e 2 = √ √ e−β /(4u) du . π 0 u The proof is complete.
Proof of Theorem 3.29. (a) By a change of variables formula it suffices to prove the formula for t = 1. We have ∞ −u Z Z 1 Z e 2 2 e−2π|x|e−2πix·ξ dx = √ √ e−4π |x| /(4u) du e−2πix·ξ dx Rn Rn π 0 u ∞ −u 1 Z e Z 2 2 = √ √ e−4π |x| /(4u)e−2πix·ξ dx du π 0 u Rn | {z } W (ξ,(4u)−1) ∞ −u r n 1 Z e u 2 = √ √ e−u|ξ| du π 0 u π Z ∞ 1 −u(1+|ξ|2) (n−1)/2 = (n+1)/2 e u du π 0 Z ∞ 1 1 −s (n−1)/2 = (n+1)/2 2 (n+1)/2 e s ds . π (1 + |ξ| ) 0 | {z } Γ[(n+1)/2] (b) is obvious. (c) Because of the scaling property (b) it suffices to consider t = 1. We have Z Z dx P (x, 1) dx = cn 2 (n+1)/2 Rn Rn (1 + |x| ) Z ∞ Z polar dσ n−1 = cn 2 (n+1)/2 r dr 0 Sn−1(0,1) (1 + r ) Z ∞ rn−1 = cn n ωn 2 (n+1)/2 dr 0 (1 + r ) Z π/2 r=tan θ n−1 = cn n ωn sin θ dθ 0 Γ n+1 2πn/2 π1/2Γ n = 2 n 2 π(n+1)/2 n n−1 n Γ 2 (n − 1)Γ 2 = 1 . The proof is complete.
Corollary 3.26 applied to f(x) = e−4π2t|x|2 and to f(x) = e−2πt|x| yields Corollary 3.39. Z Z W (ξ, t)e2πix·ξ dξ = e−4π2t|x|2 and P (ξ, t)e2πix·ξ dξ = e−2π|x|t Rn Rn for all x ∈ Rn. 38 PIOTR HAJLASZ
The Weierstrass and Poisson kernels have the following semigroup property.
Corollary 3.40. If Wt(x) = W (x, t) and Pt(x) = P (x, t), then for t1, t2 > 0
(Wt1 ∗ Wt2 )(x) = Wt1+t2 (x) and (Pt1 ∗ Pt2 )(x) = Pt1+t2 (x) for all x ∈ Rn.
Proof. It follows from Corollary 3.39 with x replaced by −x that ˆ −4π2t|x|2 ˆ −2π|x|t Wt(x) = e and Pt(x) = e . Hence Theorem 3.2(b) yields ˆ ˆ ˆ (Wt1 ∗ Wt2 )ˆ(x) = Wt1 (x)Wt2 (x) = Wt1+t2 (x) ˆ ˆ ˆ (Pt1 ∗ Pt2 )ˆ(x) = Pt1 (x)Pt2 (x) = Pt1+t2 (x) and the result follows from Corollary 3.28. 2
3.4. The Schwarz class and the Plancherel theorem.
n ∞ n Definition 3.41. We say that f belongs to the Schwarz class S (R ) = Sn if f ∈ C (R ) and for all multiindices α, β α β sup |x D f(x)| = pα,β(f) < ∞. x∈Rn That means all derivatives of f rapidly decrease to zero as |x| → ∞, faster than the inverse of any polynomial.
∞ n −|x|2 Clearly C0 (R ) ⊂ Sn, but also e ∈ Sn so functions in the Schwarz class need not be compactly supported.
{pα,β} is a countable family of norms in Sn and we can use it to define a topology in Sn.
Definition 3.42. We say that a sequence (fk) converges to f in Sn if
lim pα,β(fk − f) = 0 for all multiindices α, β. k→∞
This convergence is metrizable. Indeed, dα,β(f, g) = pα,β(f − g) is a metric and if we 0 0 arrange all these metrics in a sequence d1, d2,..., then ∞ X d0 (f, g) d(f, g) = 2−k k 1 + d0 (f, g) k=1 k defines a metric in Sn such that fn → f in Sn if and only if fn → f in the metric d.
Proposition 3.43. The space Sn has the following properties.
(a) Sn equipped with the metric d is a complete metric space. ∞ n (b) C0 (R ) is dense in Sn. (c) If ϕ ∈ Sn, then τhϕ → ϕ in Sn as h → 0, where τhϕ(x) = ϕ(x + h). HARMONIC ANALYSIS 39
(d) The mapping α β Sn 3 ϕ 7→ x D ϕ(x) ∈ Sn is continuous. (e) If ϕ ∈ Sn, then ϕ(x + he ) − ϕ(x) ∂ϕ k → (x) as h → 0. h ∂xk
in the topology of Sn. (f) If ϕ, ψ ∈ Sn, then ϕ ∗ ψ ∈ Sn and Dα(ϕ ∗ ψ) = (Dαϕ) ∗ ψ = ϕ ∗ (Dαψ) for any multiindex α.
We leave the proof as an exercise.
Theorem 3.44. The Fourier transform is a continuous, one-to-one mapping of Sn onto Sn such that (a) ∧ ∂f ˆ (ξ) = 2πiξjf(ξ), ∂xj (b) ∂fˆ (−2πixjf)ˆ(ξ) = (ξ) , ∂ξj (c) (f ∗ g)ˆ = fˆg,ˆ (d) Z Z f(x)ˆg(x) dx = fˆ(x)g(x) dx , Rn Rn (e) Z f(x) = fˆ(ξ)e2πix·ξ dξ . Rn
Proof. We already proved (a), (b), (c) and (d). Now we will prove that the Fourier transform is a continuous mapping from Sn into Sn. Formulas (a) and (b) imply that ξαDβfˆ(ξ) = C(Dα(xβf))ˆ(ξ).
Since kgˆk∞ ≤ kgk1 we get ˆ α β ˆ α β pα,β(f) = kξ D f(ξ)k∞ ≤ CkD (x f)k1 . An application of the Leibnitz rule28 implies that Dα(xβf) equals a finite sum of expressions of the form xβi Dαi f. Since Z βi αi 2 n βi αi 2 −n kx D fk1 = |(1 + |x| ) x D f(x)|(1 + |x| ) dx Rn 28Product rule. 40 PIOTR HAJLASZ
≤ C(n) sup |(1 + |x|2)nxβi Dαi f(x)| < ∞ x∈Rn ˆ it follows that f ∈ Sn. One can also easily deduce that the mapping ˆ : Sn → Sn is continuous. ˆ ˆ If f ∈ Sn than f ∈ Sn and hence both f and f are integrable, so (e) follows from the inversion formula. This formula also shows that the Fourier transform applied four times is an identity on Sn and hence the Fourier transform is a bijection on Sn. 2 Theorem 3.45 (Plancherel). The Fourier transform is an L2 isometry on a dense subset 2 Sn of L ˆ kfk2 = kfk2, f ∈ Sn, and hence it uniquely extends to an isometry of L2 ˆ 2 n kfk2 = kfk2, f ∈ L (R ) . Moreover for f ∈ L2(Rn) Z fˆ(ξ) = lim f(x)e−2πix·ξ dx R→∞ |x| ˆ¯ ¯ Proof. Given f ∈ Sn let g = f, sog ˆ = f. Indeed, Z ¯ Z gˆ(ξ) = fˆ(x)e−2πix·ξ dx = fˆ(x)e2πix·ξ dx = f¯(x) . Rn Rn Hence Theorem 3.44(d) gives Z Z Z Z ¯ ˆ ˆˆ¯ ˆ kfk2 = ff = fgˆ = fg = ff = kfk2 . Rn Rn Rn Rn 2 2 Thus the Fourier transform is an L isometry on Sn. Since Sn is a dense subset of L it uniquely extends to an isometry of L2. Now for f ∈ L2(Rn) we have 2 1 L L 3 fχB(0,R) −→ f as R → ∞ and hence Z 2 −2πix·ξ L ˆ (fχB(0,R))ˆ(ξ) = f(x)e dx −→ f(ξ) |x|≤R as R → ∞. Similarly Z 2 fˆ(ξ)e2πix·ξ dξ −→L f(x) as R → ∞. |ξ| Proposition 3.46. If f, g ∈ L2(Rn), then Z Z fˆ(x)g(x) dx = f(x)ˆg(x) dx . Rn Rn 2 Proof. Approximate f and g in L by functions in Sn, apply Theorem 3.44(d) and pass to the limit. 1 n 2 n Consider the class L (R ) + L (R ) consisting of functions of the form f = f1 + f2, 1 2 f1 ∈ L , f2 ∈ L . Then we define ˆ ˆ ˆ f = f1 + f2. In order to show that the Fourier transform is well defined in the class L1 + L2 we need to show that it does not depend on the particular choice of the representation f = f1 + f2. 1 2 1 2 Indeed, if we also have f = g1 + g2, g1 ∈ L , g2 ∈ L , then f1 − g1 = g2 − f2 ∈ L ∩ L and hence ˆ ˆ f1 − gˆ1 = (f1 − g1)ˆ = (g2 − f2)ˆ =g ˆ2 − f2, ˆ ˆ f1 + f2 =g ˆ1 +g ˆ2 . It is an easy exercise to show that p n 1 n 2 n L (R ) ⊂ L (R ) + L (R ), for 1 ≤ p ≤ 2, and hence the Fourier transform is well defined on Lp(Rn), 1 ≤ p ≤ 2 and p n 2 n n ˆ: L (R ) → L (R ) + C0(R ), for 1 ≤ p ≤ 2. Later we will prove the Hausdorff-Young Inequality which implies that p n p0 n ˆ: L (R ) → L (R ), for 1 ≤ p ≤ 2, where p0 is the H¨olderconjugate to p. 42 PIOTR HAJLASZ 4. Miscellaneous results about the Fourier transform In this chapter we will show some interesting results about the Fourier transform. Some of them, but not all, will be used later. 4.1. The Poisson summation formula. Theorem 4.1 (Poisson summation formula). If f ∈ C1(R) and C (4.1) |f(x)| + |f 0(x)| ≤ for x ∈ , 1 + x2 R then ∞ ∞ X X f(n) = fˆ(n). n=−∞ n=−∞ Remark 4.2. Here fˆ(n) stand for the Fourier transform of f evaluated at n – do not get confused with the Fourier series. P Note that it follows from the growth condition (4.1) that the series n f(n) converges and that any function f ∈ Sn satisfies (4.1). Proof. It follows from (4.1) that ∞ X (4.2) g(x) = f(x + k) k=−∞ is a periodic function of class C1 with period 1. Therefore g is can be represented as a Fourier series29 ∞ X g(x) = gˆ(n)e2πinx n=−∞ where ∞ ∞ Z 1 X Z 1 X Z k+1 gˆ(n) = g(x)e−2πinx dx = f(x + k)e−2πinx dx = f(x)e−2πinx dx 0 k=−∞ 0 k=−∞ k Z ∞ = f(x)e−2πinx dx = fˆ(n) −∞ 29Nowg ˆ(n) are Fourier series coefficients – do not get confused with the Fourier transform! HARMONIC ANALYSIS 43 so ∞ ∞ ∞ X X X f(k) = g(0) = gˆ(n) = fˆ(n). k=−∞ n=−∞ n=−∞ As an application we will prove the Jacobi identity. Definition 4.3. The Jacobi ‘theta’ function is defined by ∞ X 2 ϑ(t) = e−πn t, t > 0. n=−∞ Clearly ϑ ∈ C∞(R). − 1 Corollary 4.4 (Jacobi identity). For t > 0 we have ϑ(t) = t 2 ϑ(1/t). Proof. The function f(x) = t−1/2e−πx2/t, t > 0 can be written as f(x) = t−1/2u(xt−1/2) where u(x) = e−πx2 . Hence30 ˆ −1/2 −1/2 1/2 −πξ2t f(ξ) = t (ˆu)t−1/2 (ξ) = t ut−1/2 (ξ) = u(ξt ) = e . Thus the Poisson summation formula yields ∞ ∞ X 2 X 2 t−1/2 e−πn /t = e−πn t i.e., t−1/2ϑ(1/t) = ϑ(t). n=−∞ n=−∞ Corollary 4.5. For t > 0 we have ∞ X 1 π 1 + e−2πt = . t2 + n2 t 1 − e−2πt n=−∞ Remark 4.6. It is easy to show by taking the limit as t → 0+ in the above identity that ∞ X 1 π2 = . n2 6 n=1 Proof. In Theorem 3.29 we showed that the Fourier transform of f(x) = e−2π|x|t equals31 t fˆ(x) = π(t2 + x2) 30 ∧ −n We use here two facts: (ϕ(sx)) (ξ) = (ϕ ˆ)s(ξ) = s ϕˆ(ξ/s) with n = 1 andu ˆ(ξ) = u(ξ). 31 We are in dimension n = 1 and c1 = 1. 44 PIOTR HAJLASZ so the Poisson summation formula gives32 ∞ ∞ t X 1 X 1 + e−2πt = e−2π|n|t = π t2 + n2 1 − e−2πt n=−∞ n=−∞ from which the theorem follows. 4.2. The Heisenberg inequality. Celebrated Heisenberg’s uncertainty principle asserts that position and momentum of a particle cannot be measured at the same time and it can be written as an inequality h σ σ ≥ x p 4π where σx and σp are standard deviations of position and momentum and h is the Plank constant. Translating it into our language the above inequality can be formulated as Theorem 4.7 (Heisenberg’s inequality). For any f ∈ L2(R) and a, b ∈ R, s s Z ∞ Z ∞ kfk2 (x − a)2|f(x)|2 dx (ξ − b)2|fˆ(ξ)|2 dξ ≥ 2 . −∞ −∞ 4π 2 Moreover the equality holds if and only if f(x) = Ce2πibxe−k(x−a) for some C ∈ C and k > 0. Proof. We will prove the result under the assumption that f ∈ S1. The case of general f ∈ L2 is left to the reader as an exercise. Let f(x) = e2πibxg(x − a). Since fˆ(ξ) =g ˆ(ξ − b)e−2πa(ξ−b) we easily obtain that the Heisenberg inequality is equivalent to Z ∞ 1/2 Z ∞ 1/2 kgk2 (4.3) |x|2|g(x)|2 dx |ξ|2|gˆ(ξ)|2 dξ ≥ 2 . −∞ −∞ 4π Note that (|g|2)0 = (gg)0 = 2 re g0 g so integration by parts gives Z β Z β β Z β 2 0 2 2 0 |g(x)| dx = x |g(x)| dx = x|g(x)| − 2 re xg (x)g(x) dx. α α α α Letting α → −∞, β → +∞ and using the fact that g ∈ S1 yields Z ∞ Z ∞ |g(x)|2 dx = −2 re xg0(x)g(x) dx −∞ −∞ Z ∞ (4.4) ≤ 2 |x| |g(x)| |g0(x)| dx −∞ Z ∞ 1/2 Z ∞ 1/2 (4.5) ≤ 2 |x|2|g(x)|2 dx |g0(x)|2 dx −∞ −∞ 32 In the Poisson summation formula we assumed that f ∈ C1(R) and it satisfies (4.1). However, our function f is not C1. The condition (4.1) was required for the function (4.2) to be represented as a Fourier series. Although the condition (4.1) is not satisfied any longer, it is easy to check that the function defined by (4.2) is periodic and Lipschitz continuous so it is still equal to its Fourier series and the proof carries on. HARMONIC ANALYSIS 45 Z ∞ 1/2 Z ∞ 1/2 = 2 |x|2|g(x)|2 dx |gb0(ξ)|2 dξ −∞ −∞ Z ∞ 1/2 Z ∞ 1/2 = 2 |x|2|g(x)|2 dx |2πiξgˆ(ξ)|2 dξ −∞ −∞ from which (4.3) follows. Finally, let us investigate the case of equality. For the equality in the Schwarz inequality (4.5) we need |g0(x)| to be proportional to |xg(x)|, say |g0(x)| = k|xg(x)|, for some k > 0 so g0(x) = keiϕ(x)xg(x) for some real valued function ϕ(x). But then equality in (4.4) implies that −re e−ϕ(x) = 1 so eiϕ(x) = −1 and we have g0(x) = −kxg(x). Solving this differential equation yields g(x) = Ce−kx2 and 2πibx −k(x−a)2 hence f(x) = Ce e . 4.3. Eigenfunctions of the Fourier transform. The Fourier transform applied four times is identity on L2(Rn) so if f is an eigenfunction of the Fourier transform, fˆ = λf, then f = |λ|4f so λ ∈ {−1, 1, −i, i}. We know already one eigenfunction with the eigenvalue 1, namely f(x) = e−πx2 and now we will find now all eigenfunctions in dimension n = 1. It is natural to search among functions of the form p(x)e−πx2 , where p(x) is a polyno- mial.33 Definition 4.8. Hermite functions are defined by the formula n (−1) 2 d n 2 h (x) = eπx e−2πx , for all integers n ≥ 0. n n! dx −πx2 Note that h0(x) = e . −πx2 Lemma 4.9. hn(x) = pn(x)e for a polynomial pn(x) of order n of the form (4π)n p (x) = xn + lower order terms. n n! In particular hn ∈ S1. This lemma easily follows from the definition of hn. 0 Lemma 4.10. hn(x) − 2πxhn(x) = −(n + 1)hn+1(x). Proof. Applying the product rule to n h 2 i h(−1) d n 2 i h (x) = eπx · e−2πx n n! dx 2 2 33Note that evaluation of the Fourier transform of p(x)e−πx is straightforward: since F(e−πx )(ξ) = 2 e−πξ it suffices to apply (3.1). 46 PIOTR HAJLASZ gives n h i h (−1) d n+1 2 i h0 (x) = 2πxh (x) + eπx · e−2πx = 2πxh (x) − (n + 1)h (x). m n n! dx n n+1 | {z } (−1)n+1 −(n+1) (n+1)! Lemma 4.11. If f ∈ C∞(R) and n is a positive integer, then d n d n d n−1 (4πx) f(x) − (4πxf(x)) = −4πn f(x) for all x ∈ . dx dx dx R The result follows from a simple induction. 0 Lemma 4.12. hn(x) + 2πxhn(x) = 4πhn−1(x), where h−1(x) = 0. Proof. For n = 0 we check the equality directly. For n ≥ 1 Lemma 4.10 gives 0 hn(x) + 2πxhn(x) = 4πxhn(x) − (n + 1)hn+1(x) n n+1 (−1) 2 d n 2 (−1) 2 d n+1 2 = 4πx eπx e−2πx −(n + 1) eπx e−2πx n! dx (n + 1)! dx | {z } | {zn } d n −2πx2 +(−1) /n! ( dx ) (−4πxe ) n (−1) 2 h d n 2 d n 2 i = eπx 4πx e−2πx − (4πxe−2πx ) n! dx dx | {z } d n−1 −2πx2 −4πn( dx ) e by Lemma 4.11 n−1 (−1) 2 d n−1 2 = 4π eπx e−2πx = 4πh (x). (n − 1)! dx n−1 ˆ n ˇ n Lemma 4.13. hn = (−i) hn and hn = i hn. Remark 4.14. The lemma yields ˆ ˆ ˆ ˆ h4n = h4n, h4n+2 = −h4n+2, h4n+3 = ih4n+3, h4n+1 = −ih4n+1 Thus the lemma provides examples of eigenfunctions corresponding to all possible eigen- values 1, −1, i, −i. Proof. Since hn ∈ S1 we can compute the Fourier transform directly. Lemma 4.10 gives 0 0 ˆ hn − 2πxhn = −(n + 1)hn+1, hcn − 2\πxhn = −(n + 1)hn+1. Since (−2πixf)∧ = (fˆ)0 and fb0 = 2πiξfˆ(ξ) we have ˆ ˆ 0 ˆ (4.6) 2πiξhn(ξ) − i(hn) (ξ) = −(n + 1)hn+1(ξ) ˆ n We will prove the equality hn = (−i) hn by induction. −πx2 ˆ 0 For n = 0, h0 = e so h0 = h0 = (−i) h0. HARMONIC ANALYSIS 47 Suppose now that the equality holds for n and we will prove it for n + 1. The equality (4.6) yields. n n 0 ˆ 2πiξ(−i) hn(ξ) − i((−i) hn(ξ)) = −(n + 1)hn+1(ξ), n+1 n+1 0 ˆ −(−i) 2πξhn(ξ) + (−i) hn(ξ) = −(n + 1)hn+1(ξ), n+1 0 ˆ (−i) hn(ξ) − 2πξhn(ξ) = −(n + 1)hn+1(ξ). | {z } −(n+1)hn+1(ξ) by Lemma 4.10 ˆ n+1 This proves that hn+1 = (−i) hn+1(ξ) which completes the proof of the claim that ˆ n hn = (−i) hn(ξ) fro all n. Finally ˆ ∨ n ∨ nˇ ˇ n hn = (hn) = ((−i) hn) = (−i) hn, hn = i hn. Let Kf = f 00 − 4π2x2f. 1 Lemma 4.15. Khn = −4π(n + 2 )hn Proof. Lemmas 4.10 and 4.12 yield 0 hn − 2πxhn = −(n + 1)hn+1, 00 0 0 hn − 2πhn − 2πxhn = −(n + 1)hn+1, 00 0 0 hn = 2πhn + 2πxhn − (n + 1)hn+1. Hence 00 2 2 0 0 2 2 Khn = hn − 4π x hn = 2πhn + 2πxhn − (n + 1)hn+1 − 4π x hn 0 0 = 2πhn − (n + 1)hn+1 + 2πx (hn − 2πxhn) | {z } −(n+1)hn+1 0 = 2πhn − (n + 1)(hn+1 + 2πxhn+1) 1 = −4π n + h . 2 n 34 Lemma 4.16. For real valued functions f, g ∈ S1 we have hKf, gi = hf, Kgi Proof. The integration by parts yields Z Z Z hKf, gi = (f 00 − 4π2x2f)g = f 00g − 4π2x2fg R R R Z Z Z = fg00 − 4π2x2fg = f(g00 − 4π2x2g) = hf, Kgi. R R Lemma 4.17. The Hermite functions hn are orthogonal, hhn, hmi = 0, for n 6= m. 34Here h·, ·i denotes the L2 scalar product. 48 PIOTR HAJLASZ Proof. We have 1 1 −4π n + hh , h i = hKh , h i = hh , Kh i = −4π m + hh , h i 2 n m n m n m 2 n m and the result easily follows. Lemma 4.18. n 2 (4π) khnk = √ . 2 2n! 0 Proof. Since by Lemma 4.10 we have (n+1)hn+1 = 2πxhn −hn, integration by parts yields Z Z Z Z 2 0 0 (n + 1)khn+1k2 = 2π xhnhn+1 − hnhn+1 = 2π xhnhn+1 + hnhn+1 R R R R Z 0 2 = hn (hn+1 + 2πxhn+1) = 4πkhnk2. R | {z } 4πhn Replacing n by n − 1 yields 4π kh k2 = kh k2 n 2 n n−1 2 and a simple induction argument yields (4π)n kh k2 = kh k2. n 2 n! 0 2 Now the result follows from a simple observation that Z Z Z 2 −πx2 2 −2πx2 1 −πy2 1 kh0k2 = (e ) dx = e dx = √ e dy = √ . R R 2 R 2 The last two lemmas show that the scaled Hermite functions (4π)n −1/2 en = √ hn 2n! form an orthonormal family in L2(R). ∞ 2 n Theorem 4.19. The fmaily {en}n=0 is an orthonormal basis of L (R ). 2 R Proof. It suffices to prove that if f ∈ L (R) and fhn = 0 for all n ≥ 0, then f = 0 a.e. R R Thus suppose that fhn = 0 for all n ≥ 0. Since R 2 (4π) 2 h (x) = xn + ... e−πx , n n! it easily follows that the functions xne−πx2 , n ≥ 0, are linear combinations of the functions h0, h1, . . . , hn. Hence Z f(x)xne−πx2 dx = 0 for all n ≥ 0. R HARMONIC ANALYSIS 49 Since by the Schwarz inequality f(x)e−πx2 ∈ L1 we can compute the Fourier transform directly Z Z ∞ n 2 ∧ 2 2 X (−2πixξ) f(x)e−πx = f(x)e−πx e−2πixξ dx = f(x)e−πx dx n! R R n=0 ∞ n Z X (−2πiξ) 2 = f(x)xne−πx dx = 0. n! n=0 R Vanishing of the Fourier transform of f(x)e−πx2 implies that f(x)e−πx2 = 0 a.e. and hence f(x) = 0 a.e. ∞ 2 2 Since {en}n=0 is an orthonormal basis of L (R) any f ∈ L can be written as ∞ X f = hf, eni en n=0 n This and the fact thate ˆn = (−i) en yield Theorem 4.20 (Wiener’s definition of the Fourier transform). For f ∈ L2(R) we have ∞ ˆ X n f = hf, eni(−i) en. n=0 2 Let Hk, k = 0, 1, 2, 3 be subspaces of L (R) defined by ∞ Hk = span {e4n+k}n=0. The subspaces Hk are orthogonal and 2 L (R) = H0 ⊕ H1 ⊕ H2 ⊕ H3. Clearly, the subspaces Hk are eigenspaces of the Fourier transform: ˆ k f = (−i) f for f ∈ Hk and k = 0, 1, 2, 3. 4.4. The Bochner-Hecke formula. As was pointed out in Section 4.3 computing the Fourier transform of P (x)e−π|x|2 is straightforward due to formula (3.1) and Theorem 3.7, but the computations are often algebraically complicated and it is not always obvious how to write the answer in a compact and nice form. We have see examples of such computations in Section 4.3 when we found eigenfunctions of the Fourier transform and now we will see another example of this type. The results of this section will be used later in Section 6.3 Definition 4.21. We say that Pk(x) is a homogeneous harmonic polynomial of degree k if X α Pk(x) = aαx and ∆Pk = 0 . |α|=k Theorem 4.22 (Bochner-Hecke). If Pk(x) is a homogeneous harmonic polynomial of de- gree k, then −π|x|2 k −π|ξ|2 F Pk(x)e (ξ) = (−i) Pk(ξ)e . 50 PIOTR HAJLASZ Proof. Since the polynomial Pk is homogeneous of degree k, X α k Pk(−2πix) = aα(2πix) = (−2πi) Pk(x). |α|=k Hence (3.1) applied to f(x) = e−π|x|2 yields35 Z −π|x|2 −2πix·ξ −π|x|2 Pk(x)e e dx = F Pk(x)e (ξ)(4.7) Rn −k −π|ξ|2 −π|ξ|2 = (−2πi) Pk(D)e = Q(ξ)e for some polynomial Q and it remains to prove that Q(ξ) = Pk(−iξ). 36 Lemma 4.23. If P (x) is a polynomial in Rn, then Z Z −π P (x +iξ )2 −π|x|2 (4.8) P (x)e j j j dx = P (x − iξ)e dx . Rn Rn Proof. By the same argument involving the same contour integration as in the proof of Theorem 3.7 for any polynomial P of one variable we have Z ∞ Z ∞ (4.9) P (x)e−π(x+iξ)2 dx = P (x − iξ)e−πx2 dx . −∞ −∞ If P is a polynomial in n variables, then (4.8) follows from (4.9) and the Fubini theorem. Multiplying both sides of (4.7) by eπ|ξ|2 and applying (4.8) we obtain Z P 2 Z 2 −π (xj +iξj ) −π|x| Q(ξ) = Pk(x)e j dx = Pk(x − iξ)e dx . Rn Rn We can write X α Pk(x − η) = η Pα(x) α and then Z Z −π|x|2 X α −π|x|2 Pk(x − η)e dx = η Pα(x)e dx , n n R α R so clearly the integral is a polynomial is η. With this notation we obtain Z X α −π|x|2 Q(ξ) = (iξ) Pα(x)e dx n α R and hence Z Z X α −π|x|2 −π|x|2 Q(ξ/i) = ξ Pα(x)e dx = Pk(x − ξ)e dx . n n α R R 35 Instead of using (3.1) we could apply the differential operator Pk(Dξ) to both sides of the equality Z 2 2 e−π|x| e−2πix·ξ dx = e−π|ξ| n R and differentiate under the sign of the integral. 36Any polynomial, not necessarily harmonic or homogeneous. HARMONIC ANALYSIS 51 37 Since Pk is a harmonic function, it has the mean value property Z n−1 Pk(sθ − ξ) dσ(θ) = |S |Pk(−ξ) . Sn−1 Thus integration in polar coordinates gives Z ∞ Z n−1 −πs2 Q(ξ/i) = s Pk(sθ − ξ) dσ(θ) e ds 0 Sn−1 Z ∞ n−1 n−1 −πs2 = Pk(−ξ) s |S |e ds 0 Z −π|x|2 = Pk(−ξ) e dx = Pk(−ξ) Rn and hence Q(ξ) = Pk(−iξ). The proof is complete. Using homogeneity of Pk and the second part of Theorem 3.2(e) one can easily deduce from the Bochner-Hecke formula the folloiwing result. Corollary 4.24. If Pk is a homogeneous harmonic polynomial of degree k, then for any t > 0 we have −πt|x|2 −k− n k −π|ξ|2/t F Pk(x)e (ξ) = t 2 (−i) Pk(ξ)e . We leave details as an easy exercise. 4.5. Symmetry of the Fourier transform. If f is radially symmetric i.e., f = f ◦ ρ for all ρ ∈ O(n), then fˆ◦ ρ = f[◦ ρ = fˆ for all ρ ∈ O(n) so fˆ is radially symmetric too. Now we will prove a more delicate result about the symmetry of the Fourier transform. Proposition 4.25. Let f ∈ L1(Rn) be a real valued function of the form x f(x) = j g(|x|). |x| Then there is a continuous function h : [0, ∞) → R, h(0) = 0, h(t) → 0 as t → ∞ such that ξ fˆ(ξ) = i j h(|ξ|), ξ ∈ n. |ξ| R While the function f is not radially symmetric a more symmetric object is the map x F (x) = g(|x|) ∈ L1( n, n), x ∈ n |x| R R R so f is the jth component of F . We will prove that ξ Fˆ(ξ) = i h(|ξ|), ξ ∈ n |ξ| R from which Proposition 4.25 will readily follow. In the proof we will use the following result which is interesting on its own. It ill be used later one more time in a proof of Theorem 6.2. 37 We integrate here Pk over the sphere centered at −ξ and of radius s. 52 PIOTR HAJLASZ Lemma 4.26. If m : Rn → Rn is a measurable function that is homogeneous of degree 0, i.e. m(tx) = m(x) for t > 0, and commutes with the orthogonal transformations, i.e. (4.10) m(ρ(x)) = ρ(m(x)) for all x ∈ Rn and ρ ∈ O(n), then there is a constant c such that x (4.11) m(x) = c for all x 6= 0. |x| n Proof. Let e1, e2, . . . , en be the standard orthogonal basis of R . If [ρjk] is the matrix representation of ρ ∈ O(n), then the condition (4.10) reads as n X (4.12) mj(ρ(x)) = ρjkmk(x), j = 1, 2, . . . , n, k=1 where m(x) = (m1(x), . . . , mn(x)). Let m1(e1) = c. Consider all ρ ∈ O(n) such that ρ(e1) = e1. This condition means that the first column of the matrix [ρjk] equals e1, i.e. ρ11 = 1, ρj1 = 0, for j > 1. Since columns are orthogonal, for k > 1 we have n X 0 = ρj1ρjk = ρ1k . j=1 Thus 1 0 ... 0 0 ρ22 . . . ρ2n ρ = . . . . , . . .. . 0 ρn2 . . . ρnn n where [ρjk]j,k=2 is the matrix of an arbitrary orthogonal transformation in the (n − 1)- dimensional subspace orthogonal to e1. For x = e1 = ρ(e1) = ρ(x) and j ≥ 2 identity (4.12) yields n n X X mj(e1) = ρjkmk(e1) = ρjkmk(e1) , k=1 k=2 and hence m2(e1) ρ22 . . . ρ2n m2(e1) . . . . . . = . .. . . . mn(e1) ρn2 . . . ρnn mn(e1) T That means the vector [m2(e1), . . . , mn(e1)] is fixed under an arbitrary orthogonal trans- formation of Rn−1, so it must be a zero vector, i.e. m2(e1) = ... = mn(e1) = 0 . Now formula (4.12) for any ρ ∈ O(n) and x = e1, takes the form mj(ρ(e1)) = ρj1m1(e1) = cρj1 . HARMONIC ANALYSIS 53 By homogeneity it suffices to prove (4.11) for |x| = 1. Let ρ ∈ O(n) be such that ρ(e1) = x. Then ρj1 = xj, j = 1, 2, . . . , n and hence x m (x) = cρ = cx = c j . j j1 j |x| This completes the proof of the lemma. Proof of Proposition 4.25. Since the function F is odd Z Z iFˆ(ξ) = iF (x)(cos(2πx · ξ) − i sin(2πx · ξ)) dx = sin(2πx · ξ)F (x) dx Rn Rn ˆ n n−1 n so the mapping ξ 7→ iF (ξ) takes values in R . Fix k > 0 and define mk : S (0, k) → R ˆ n n by mk(ξ) = iF (ξ) for |ξ| = k. Extend mk to mk : R \{0} → R as a function homogeneous of degree 0, i.e. kξ m (ξ) = i Fˆ for ξ 6= 0. k |ξ| We claim that (4.13) mk(ρ(ξ)) = ρ(mk(ξ)) for ρ ∈ O(n) and ξ 6= 0. 38 Since mk is homogeneous of degree 0, it suffices to check (4.13) for |ξ| = k. We have Z Z x −1 x mk(ρ(ξ)) = sin(2πx · ρ(ξ)) g(|x|) dx = sin(2πρ (x) · ξ) g(|x|) dx Rn |x| Rn |x| Z ρ(x) Z ρ(x) = sin(2πx · ξ) g(|ρ(x)|) dx = sin(2πx · ξ) g(|x|) dx Rn |ρ(x)| Rn |x| Z x = ρ sin(2πx · ξ) g(|x|) dx = ρ(mk(ξ)). Rn |x| According to Lemma 4.26 there is a constant h(k) ∈ R such that mk(ξ) = −h(k)ξ/|ξ|. In particular for |ξ| = k we have ξ iFˆ(ξ) = m (ξ) = − h(|ξ|). k |ξ| ˆ n n Clearly h is continuous, h(0) = 0 and h(t) → 0 as t → ∞, because iF ∈ C0(R , R ). 38First we use equality x·ρ(ξ) = ρ−1(x)·ξ which follows from the fact that the orthogonal transformation ρ−1 preserves the scalar product and then we use the change of variables x = ρ(y) whose absolute value of the Jacobian equals 1. 54 PIOTR HAJLASZ 5. Tempered distributions 5.1. Basic constructions. 0 Definition 5.1. The space Sn of all continuous linear functionals on Sn is called the 0 space of tempered distributions. The evaluation of u ∈ Sn on ϕ ∈ Sn will usually be 0 denoted by u[ϕ]. The space Sn is equipped with the weak-∗ convergence (usually called 0 weak convergence or just convergence). Namely uk → u in Sn if uk[ψ] → u[ψ] for every ψ ∈ Sn. Here are examples: 1. If f ∈ Lp(Rn), 1 ≤ p ≤ ∞, then Z 0 Lf [ϕ] = f(x)ϕ(x) dx defines Lf ∈ Sn. Rn 2. If µ is a measure of finite total variation, then Z 0 Lµ[ϕ] = ϕ dµ defines Lµ ∈ Sn. Rn 3. We say that a function f is a tempered Lp function if f(x)(1+|x|2)−k ∈ Lp(Rn) for some nonnegative integer k. If p = ∞ we call f a slowly increasing function. Then Z 0 Lf [ϕ] = f(x)ϕ(x) dµ defines Lf ∈ Sn for all 1 ≤ p ≤ ∞. Rn Note that slowly increasing functions are exactly measurable functions bounded by poly- nomials. 4. A tempered measure is a Borel measure µ such that Z (1 + |x|2)−k dµ < ∞ Rn 0 for some integer k ≥ 0. As before Lµ ∈ Sn. α 0 5. L[ϕ] = D ϕ(x0) is a tempered distribution L ∈ Sn. The distributions generated by a function or by a measure will often be denoted by Lf [ϕ] = f[ϕ], Lµ[ϕ] = µ[ϕ]. HARMONIC ANALYSIS 55 0 p Suppose that u ∈ Sn. If there is a tempered L function f such that u[ϕ] = f[ϕ] for ϕ ∈ Sn, then we can identify u with the function f and simply write u = f. The identification is possible, because the function f is uniquely defined (up to a.e. equivalence). This follows from a well known result. n 1 R ∞ Lemma 5.2. If Ω ⊂ R is open and f ∈ Lloc(Ω) satisfies Ω fϕ = 0 for all ϕ ∈ C0 (Ω), then f = 0 a.e. Note that not every function f ∈ C∞(Rn) defines a tempered distribution, because it may happen that form some ϕ ∈ Sn the function fϕ is not integrable and hence the integral f[ϕ] = R f(x)ϕ(x) dx does not make sense. Rn Theorem 5.3. A linear functional L on Sn is a tempered distribution if and only if there is a constant C > 0 and a positive integer m such that X |L(ϕ)| ≤ C pα,β(ϕ) for all ϕ ∈ Sn. |α|,|β|≤m Proof. If a linear functional L satisfies the given estimate, then clearly it is continuous 0 on Sn, so it remains to prove the converse implication. Let L ∈ Sn. We claim that there is a positive integer m such that |L(ϕ)| ≤ 1 for all n X 1 o ϕ ∈ ϕ ∈ Sn : pα,β(ϕ) ≤ := Nm . m |α|,|β|≤m To the contrary suppose that there is a sequence ϕk ∈ Sn such that |L(ϕk)| > 1 and X 1 (5.1) p (ϕ ) ≤ , k = 1, 2, 3,... α,β k k |α|,|β|≤k Note that (5.1) implies that ϕk → 0 in Sn, so the inequality |L(ϕk)| > 1 contradicts continuity of L. This proves the claim. Denote X kϕk = pα,β(ϕ) . |α|,|β|≤m Observe that k · k is a norm. For an arbitrary 0 6= ϕ ∈ Sn,ϕ ˜ = ϕ/(mkϕk) satisfies kϕ˜k ≤ 1/m, soϕ ˜ ∈ Nm and hence |L(ϕ)| = mkϕk|L(ϕ ˜)| ≤ mkϕk which proves the theorem. 2 For any function g on Rn we defineg ˜(x) = g(−x). Then it easily follows from the Fubini theorem that for u, ϕ, ψ ∈ Sn Z Z (u ∗ ϕ)(x)ψ(x) dx = u(x)(ϕ ˜ ∗ ψ)(x) dx . Rn Rn Regarding the functions u and u ∗ ϕ as distributions we can rewrite this equality as (u ∗ ϕ)[ψ] = u[ϕ ˜ ∗ ψ] 0 If u ∈ Sn and ϕ ∈ Sn, then ψ 7→ u[ϕ ˜ ∗ ψ] is a tempered distribution so it motivates the following definition. 56 PIOTR HAJLASZ 0 Definition 5.4. If u ∈ Sn and ϕ ∈ Sn, then the convolution of u and ϕ is a tempered distribution defined by the formula (u ∗ ϕ)[ψ] := u[ϕ ˜ ∗ ψ] . The following two results are left as an easy exercise. 0 Proposition 5.5. If u ∈ Sn, ϕ, ψ ∈ Sn, then (u ∗ ϕ) ∗ ψ = u ∗ (ϕ ∗ ψ) . n Proposition 5.6. If µ ∈ B(R ) and ϕ ∈ Sn, then µ ∗ ϕ = ϕ ∗ µ where µ ∗ ϕ is the 0 convolution of a distribution µ ∈ Sn with a function ϕ ∈ Sn and ϕ ∗ µ is the convolution of a function with a measure. 0 Theorem 5.7. If u ∈ Sn and ϕ ∈ Sn, then the tempered distribution u∗ϕ can be identified with a slowly increasing function f defined by n f(x) = u[τ−xϕ˜] = u[ϕ(x − ·)] for all x ∈ R . Moreover f ∈ C∞(Rn) and all its derivatives are slowly increasing. Proof. Note that the equality u[τ−xϕ˜] = u[ϕ(x − ·)] is obvious. First we will prove that ∞ the function f(x) = u[τ−xϕ˜] = u[ϕ(x − ·)] is C and f, as well as all its derivatives, are slowly increasing. It follows from Theorem 3.43(c) that f is continuous. Observe that for a fixed x ∈ Rn f(x + he ) − f(x) ϕ(x + he − ·) − ϕ(x − ·) k = u k → u[(∂ ϕ)(x − ·)] as h → 0 h h k because it is easy to check using Theorem 3.43(e) that for every x ∈ Rn ϕ(x + he − ·) − ϕ(x − ·) k → (∂ ϕ)(x − ·) as h → 0 h k in the topology of Sn. Thus the partial derivatives ∂kf(x) = u[(∂kϕ)(x − ·)] exist and are continuous by Theorem 3.43(c). Since ∂kϕ ∈ Sn, iterating the above process for any multiindex α we obtain (5.2) Dαf(x) = u[(Dαϕ)(x − ·)] . This implies that f ∈ C∞(Rn). Since u is a tempered distribution, Theorem 5.3 gives the estimate X 0 m (5.3) |f(x)| = |u[τ−xϕ˜]| ≤ C pα,β(τ−xϕ˜) ≤ C (1 + |x|) . |α|,|β|≤m Indeed, α β α β pα,β(τ−xϕ˜) = sup |z (D ϕ˜)(z − x)| = sup |(z + x) (D ϕ˜)(z)| z∈Rn z∈Rn ≤ C 1 + |x||α| sup 1 + |z||α| |Dβϕ˜(z)| ≤ C0(1 + |x|)m . z∈Rn HARMONIC ANALYSIS 57 Hence f is slowly increasing, and so it defines a tempered distribution. Since the derivatives of f satisfy (5.2), which is an expression of the same type as the one in the definition of f, we conclude that all derivatives Dαf are slowly increasing. It remains to prove that f = u ∗ ϕ in the sense of tempered distributions i.e., (5.4) (u ∗ ϕ)[ψ] = f[ψ] for all ψ ∈ Sn. Using f(y) = u[ϕ(y − ·)], (5.4) reads as h Z i Z u ϕ(y − ·)ψ(y) dy = u[ϕ(y − ·)]ψ(y) dy. Rn Rn The idea of proving this equality is to approximate the integral on on the left hand side by Riemann sums, use linearity of u to go under the sign of the sum and to pass to the limit. However, the details of this approximation argument are not obvious. They are elemen- tary and tedious, but not obvious and so in most of the textbooks they are summarized as ‘easy to see’. For the sake of completeness we decided to include details. ∞ First of all observe that is suffices to prove (5.4) under the assumption that ϕ, ψ ∈ C0 . Indeed, suppose we know (5.4) for compactly supported functions, but now ϕ, ψ ∈ Sn. 39 ∞ ∞ Let C0 3 ϕk → ϕ in Sn and C0 3 ψ` → ψ in Sn. Observe that fk(x) = u[τ−xϕ˜k] → f(x) pointwise. It follows from the proof of (5.3) that all the functions fk have a common estimate independent of k m |fk(x)| ≤ C(1 + |x|) 0 and hence fk → f in Sn by the Dominated Convergence Theorem. It is also clear from 0 the definition of the convolution that u ∗ ϕk → u ∗ ϕ in Sn, so (u ∗ ϕ)[ψ`] = lim (u ∗ ϕk)[ψ`] = lim fk[ψ`] = f[ψ`] . k→∞ k→∞ ∞ The second equality follows from assuming (5.4) for all ϕk, ψ` ∈ C0 . Now letting ` → ∞ yields (5.4). If the support of ψ is contained in a cube Q with integer edge-length we divide Q into −k cubes {Qki}i of edge-length 2 and centes yki. Then the Riemann sums converge in the topology of Sn (as a function of x) to the integral X Z ϕ(yki − x)ψ(yki)|Qki| → ϕ(y − x)ψ(y) dy n i R and hence h Z i X Z u ϕ(y − ·)ψ(y) dy = lim u[ϕ(yki − ·)]ψ(yki)|Qki| = u[ϕ(y − ·)]ψ(y) dy. n k→∞ n R i R Boring (but elementary) details of the convergence of the Riemann sums in the topology of Sn are moved to Appendix 5.7. Note that formula (5.2) implies 0 Theorem 5.8. If u ∈ Sn and ϕ ∈ Sn, then for any multiindex α we have Dα(u ∗ ϕ)(x) = (u ∗ (Dαϕ))(x) . 39See Theorem 3.43(b). 58 PIOTR HAJLASZ The next result will require the use of the Banach-Steinhaus theorem. 0 Theorem 5.9. If uk → u is Sn and ϕk → ϕ in Sn, then uk[ϕk] → u[ϕ]. Proof. In order to be able to use triangle inequality to prove the convergence we need to 0 know uniform estimates for the sequence uk. Convergence uk → u in Sn means pointwise convergence on a complete metric space Sn so it is natural to use the Banach-Steinhaus Theorem to get uniform estimates for uk. We state it as a lemma. Lemma 5.10 (Banach-Steinhaus). Let X be a complete metric space and let {fi}i∈I be a family of continuous real-valued functions on X. If the functions in the family are pointwise bounded, i.e. sup |fi(x)| < ∞ for x ∈ X, i∈I then there is an open set U ⊂ X and a constant M > 0 such that sup |fi(x)| ≤ M for all x ∈ U. i∈I It easily follows from the linearity that it suffices to consider the case uk → 0 and ϕk → 0. The functions ϕ 7→ |uk[ϕ]| are continuous on the complete metric space (Sn, d). They are pointwise bounded on Sn since |uk[ϕ]| → 0 for every ϕ ∈ Sn. Thus Lemma 5.10 implies that sup |uk[ϕ]| ≤ M for all ϕ ∈ B(ϕ0, r0). k Fix ε > 0. There is k1 such that |uk[ϕ0]| < ε for k ≥ k1. Since ϕk → 0 in Sn, then also −1 ε ϕk → 0 and hence −1 −1 d(ϕ0 + ε ϕk, ϕ0) = d(ε ϕk, 0) → 0 , −1 so there is k2 such that for k ≥ k2, ϕ0 + ε ϕk ∈ B(ϕ0, r0). Hence for k ≥ max{k1, k2} we have −1 −1 |uk[ε ϕk]| ≤ |uk[ϕ0 + ε ϕk]| + |uk[ϕ0]| ≤ M + ε , so |uk[ϕk]| ≤ ε(M + ε). 0 Theorem 5.11. If uk → u in Sn, then there is a constant C > 0 and a positive integer m such that X sup |uk[ϕ]| ≤ C pα,β(ϕ) for all ϕ ∈ Sn. k |α|,|β|≤m 0 Proof. Let vk = uk − u. Clearly vk → 0 in Sn. It suffices to prove that there is C > 0 and a positive integer m such that X sup |vk[ϕ]| ≤ C pα,β(ϕ) for all ϕ ∈ Sn. k |α|,|β|≤m Indeed, this estimate, the inequality sup |uk[ϕ]| ≤ |u[ϕ]| + sup |vk[ϕ]| k k and Theorem 5.3 readily yield the result. HARMONIC ANALYSIS 59 As in the proof of Theorem 5.3 it suffices to prove that there is a positive integer m such that supk |vk[ϕ]| ≤ 1 for n X 1 o ϕ ∈ Nm := ϕ ∈ Sn : pα,β(ϕ) ≤ . m |α|,|β|≤m To the contrary suppose that for any positive integer i we can find ϕi ∈ Ni such that supk |vk[ϕi]| > 1, i.e. (5.5) |vki [ϕi]| > 1 for some ki. Clearly ϕi → 0 in Sn. If the sequence ki is bounded, (5.5) contradicts continuity of functionals vk. If ki is unbounded, then there is a subsequence kij → ∞ and again (5.5) contradicts Theorem 5.9. ∞ n 0 0 Theorem 5.12. C0 (R ) is dense in Sn, i.e. if u ∈ Sn, then there is a sequence wk ∈ ∞ n C0 (R ) such that wk[ψ] → u(ψ) for all ψ ∈ Sn. 0 ∞ n R 40 Proof. Let u ∈ . If ϕ ∈ C ( ), n ϕ = 1, then for every ψ ∈ n,ϕ ˜ε ∗ ψ → ψ in n Sn 0 R R S S so (u ∗ ϕε)[ψ] = u[ϕ ˜ε ∗ ψ] → u[ψ] . ∞ Since uε = u ∗ ϕε ∈ C we obtain a sequence of smooth functions that converge to u in 0 ∞ n Sn. If we take a cut-off function η ∈ C0 (R ), 0 ≤ η ≤ 1, η(x) = 1 for |x| ≤ 1 and η(x) = 0 for |x| ≥ 2, then one can easily prove that for ψ ∈ Sn η(x/k)ψ(x) → ψ(x) in Sn as k → ∞. 41 ∞ Using this fact and Theorem 5.9 we obtain that wk(x) = η(x/k)(u∗ϕk−1 ) ∈ C0 converges 0 to u in Sn. Indeed, wk[ψ] = (u ∗ ϕk−1 )[η(·/k)ψ(·)] → u[ψ]. If ϕ, ψ ∈ Sn, then the integration by parts gives Z Z Dαϕ(x)ψ(x) dx = (−1)|α| ϕ(x)Dαψ(x) dx, Dαϕ[ψ] = ϕ[(−1)αψ]. Rn Rn For h ∈ Rn we have Z Z (τhϕ)(x)ψ(x) dx = ϕ(x)(τ−hψ)(x) dx, τhϕ[ψ] = ϕ[τ−hψ]. Rn Rn Moreover Z Z ϕ˜(x)ψ(x) dx = ϕ(x)ψ˜(x) dx, ϕ˜[ψ] = ϕ[ψ˜], n n ZR ZR ϕˆ(x)ψ(x) dx = ϕ(x)ψˆ(x) dx, ϕˆ[ψ] = ϕ[ψˆ], n n ZR ZR ϕˇ(x)ψ(x) dx = ϕ(x)ψˇ(x) dx, ϕˇ[ψ] = ϕ[ψˇ]. Rn Rn 40Why? 41Instead of using Theorem 5.9 which is somewhat not constructive as it is based on the Banach- Steinhaus theorem, we could prove convergence by a brute force estimates. 60 PIOTR HAJLASZ If η ∈ C∞ is slowly increasing and all derivatives of η are slowly increasing, i.e. every α derivative D η is bounded by a polynomial, then for ψ ∈ Sn, ηψ ∈ Sn and the mapping α ψ 7→ ηψ is continuous in Sn. In particular x ψ(x) ∈ Sn. Moreover Z Z (ηϕ)[ψ] = η(x)ϕ(x)ψ(x) dx = ϕ(x) η(x)ψ(x) dx = ϕ[ηψ]. Rn Rn This motivates the following definition. 0 Definition 5.13. For u ∈ Sn we define • The distributional partial derivative Dαu is a tempered distribution defined by the formula Dαu[ψ] = u[(−1)|α|Dαψ]. 0 • The translation τhu ∈ Sn is defined by (τhu)[ψ] = u[τ−hψ]. 0 • The reflectionu ˜ ∈ Sn is defined by u˜[ψ] = u[ψ˜]. 0 0 • The Fourier transform F(u) =u ˆ ∈ Sn and the inverse Fourier transformu ˇ ∈ Sn are uˆ[ψ] = u[ψˆ] andu ˇ[ψ] = u[ψˇ]. • If η ∈ C∞ is slowly increasing and all derivatives of η are slowly increasing, then we define (ηu)[ψ] = u[ηψ]. In particular (xαu)[ψ] = u[xαψ]. 0 The formulas preceding the definition show that on the subclass Sn ⊂ Sn the partial derivative, the translation, the reflection, the Fourier transform and the multiplication by a function defined in the distributional sense coincide with those defined in the classical way. If f ∈ L1(Rn)+L2(Rn), in particular, if f ∈ Lp(Rn), 1 ≤ p ≤ 2, then the classical Fourier transform coincides with the distributional one. That easily follows from Theorem 3.2(c) and Proposition 3.46. The basic properties of the Fourier transform, distributional derivative and convolution 0 in Sn are collected in the next result whose easy proof is left to the reader. 0 42 Theorem 5.14. The Fourier transform in a homeomorphism of Sn onto itself. Moreover 0 for u ∈ Sn and ϕ ∈ Sn we have (a) (ˆu)∨ = u, (b) (u ∗ ϕ)ˆ =ϕ ˆuˆ, (c) Dα(u ∗ ϕ) = Dαu ∗ ϕ = u ∗ Dαϕ, (d) (Dαu)ˆ = (2πix)αuˆ, (e) ((−2πix)αu)ˆ = Dαuˆ. 42 0 With respect to the weak convergence in Sn. HARMONIC ANALYSIS 61 Just for the illustration we will prove the property (d). For any ψ ∈ Sn we have (Dαu)∧[ψ] = u[(−1)|α|Dαψˆ] = u[(−1)|α|((−2πix)αψ)∧] = u[((2πix)αψ)∧] = (2πix)αuˆ[ψ]. 0 α Note that in the case (c), u ∗ ϕ ∈ Sn and D (u ∗ ϕ) is understood in the distributional sense. On the other hand u ∗ ϕ ∈ C∞ and Theorem 5.8 shows that the distributional derivative of u ∗ ϕ coincides with the classical one. Example 5.15. The Dirac measure δa defines a tempered distribution so Z ˆ ˆ ˆ −2πix·a δa[ψ] = δa[ψ] = ψ(a) = ψ(x)e dx Rn ˆ −2πix·a and hence δa = e . Example 5.16. Let µ be a measure on R defined by µ(E) = #(E ∩ Z), where #A stands for the cardinality of a set A. In other words µ is the counting measure on Z trivially extended to R. Clearly, µ is a tempered measure so it defines a tempered distribution. 0 Moreoverµ ˆ = µ in S1. Indeed, this is a simple consequence of the Poisson summation formula Theorem 4.1 that ∞ ∞ X X µˆ[ϕ] = µ[ϕ ˆ] = ϕˆ(n) = ϕ(n) = µ[ϕ]. n=−∞ n=−∞ 5.2. Tempered distributions as derivatives of functions. Using the notion of distri- 0 butional derivative we can provide a new class of examples of a distributions in Sn. If fα, |α| ≤ m are slowly increasing functions and aα ∈ C for |α| ≤ m, then X α 0 (5.6) u = aαD fα ∈ Sn . |α|≤m α where the derivative D fα is understood in the distributional sense. 0 The aim of this section is to prove a surprising result that every distribution in Sn can be represented in the form (5.6), see Theorem 5.21. Proposition 5.17. Let N be a positive integer, then the operators43 N N 0 0 (I − ∆) : Sn → Sn and (I − ∆) : Sn → Sn are homeomorphisms. The inverse operator in both cases is given by the operator (I − ∆)−N := F −1 (1 + 4π2|ξ|2)−N F. 0 N Remark 5.18. That means for every v ∈ Sn (v ∈ Sn) the equation (I − ∆) u = v has 0 a unique solution u ∈ Sn (u ∈ Sn) given by u = (I − ∆)N v = F −1 (1 + 4π2|ξ|2)−N F(v). 43I − ∆ stands for the identity minus the Laplace operator. 62 PIOTR HAJLASZ N Proof. Suppose that u, v ∈ Sn satisfy (I − ∆) u = v. Taking the Fourier transform yields (1 + 4π2|ξ|2)N F(u) = F(v) and hence F(u) = (1 + 4π2|ξ|2)−N F(v). Since F(u) ∈ Sn, the right hand side is also in Sn so we can take the inverse Fourier transform and we obtain (5.7) u = F −1 (1 + 4π2|ξ|2)−N F(v). Thus if a solution u ∈ Sn exists, it is unique and given by (5.7). To see that the formula 2 2 −N indeed, defines a solution u ∈ Sn observe that the function (1 + 4π |ξ| ) and all its 2 2 −N derivatives are slowly increasing so if f ∈ Sn, then (1 + 4π |ξ| ) f ∈ Sn. Hence for any v ∈ Sn, the right hand side of (5.7) defines a function in Sn and backward computations shows that (I − ∆)N u = v. 0 N Similarly, if u, v ∈ Sn, then (I − ∆) u = v if and only if the Fourier transforms are equal i.e., (1 + 4π2|ξ|2)N F(u) = F(v) which is equivalent to F(u) = (1 + 4π2|ξ|2)−N F(v), u = F −1 (1 + 4π2|ξ|2)−N F(v). Note that we were allowed to multiply both sides by (1 + 4π2|ξ|2)−N because this function and all its derivatives are slowly increasing. The operator (I −∆)N maps functions of class Ck to functions of class Ck−2N so it lowers regularity of functions. That means the inverse operator (I − ∆)−N increases regularity. Hence we should expect that it also increases regularity of distributions. Since every tem- pered distribution has finite order in a sense that it can be estimated by a finite sum as in 0 Theorem 5.3, one could expect that for u ∈ Sn and sufficiently large N, the distribution (I − ∆)−N u is actually a regular function. As we will see this intuition is correct. Remark 5.19. The operator (I −∆)−N is called a Bessel potential and one can prove that it can be represented as an integral operator. In Section ?? we will study Bessel potentials in detail. We will find an explicit integral formula for the Bessel potentials and we will show how use them to characterize Sobolev spaces. k,1 Definition 5.20. By Cloc we will denote the class of k times continuously differentiable 0,1 function whose derivatives of order k are locally Lipschitz continuous. In particular Cloc denotes the class of locally Lipschitz continuous functions. 0 44 Theorem 5.21. Suppose u ∈ Sn satisfies the estimate X |u(ϕ)| ≤ C pα,β(ϕ) . |α|,|β|≤m If k is a nonnegative integer and N > (n + m + k)/2, then the tempered distribution v = (I − ∆)−N u has the following properties: 44See Theorem 5.3. HARMONIC ANALYSIS 63 (a) If k = 0, then v is a slowly increasing function. k−1,1 (b) If k ≥ 1, then v is a slowly increasing function of the class Cloc . Also all deriva- tives of v or order |α| ≤ k are slowly increasing. Proof. We will need Lemma 5.22. If P (x) is a polynomial in Rn of degree p and N > (n + p)/2, then all derivatives of P (x) f(x) = (1 + |x|2)N belong to L1(Rn). Proof. Since 2N − p > n, f ∈ L1(Rn). We have ∂f Q(x) = 2 2N , deg Q = 2N + p − 1. ∂xi (1 + |x| ) The function on the right hand side is of the same form as f. Since n + (2N + p − 1) 2N > 2 1 we conclude that ∂f/∂xi ∈ L . The integrability of higher order derivatives follows by induction. 2 We will prove that the distributional derivatives of v of orders |γ| ≤ k are slowly in- creasing functions. We have (−1)|γ|Dγv[ψ] = uF (1 + 4π2|ξ|2)−N F −1(Dγψ) . Hence X |Dγv[ψ]| ≤ C sup xαDβ F (1 + 4π2|ξ|2)−N F −1(Dγψ) x∈ n |α|,|β|≤m R β+γ X α ξ −1 = C Cα,β,γ sup F D 2 2 N F (ψ) x∈ n (1 + 4π |ξ| ) |α|,|β|≤m R β+γ 0 X α x −1 ≤ C D F (ψ) . (1 + 4π2|x|2)N |α|,|β|≤m 1 Note that xβ+γ Dα F −1(ψ) (1 + 4π2|x|2)N α! xβ+γ X αi βi −1 = D 2 2 N D (F (ψ)) . α1!βi! (1 + 4π |x| ) αi+βi=α Since deg xβ+γ ≤ m + k and N > (n + m + k)/2, Lemma 5.22 gives xβ+γ Dαi ∈ L1( n) , (1 + 4π2|x|2)N R 64 PIOTR HAJLASZ so γ X −1 βi |D v[ψ]| ≤ C kF (x ψ)k∞ i X βi ≤ C kx ψk1 i 0 2 m/2 ≤ C k(1 + |x| ) ψ(x)k1 . This proves that for |γ| ≤ k the functional ψ 7→ Dγv[ψ] 1 2 m/2 ∞ is bounded on L ((1 + |x| ) dx). Thus there are functions gγ ∈ L such that Z γ 2 m/2 D v[ψ] = ψ(x)gγ(x)(1 + |x| ) dx , Rn i.e. γ 2 m/2 D v = gγ(x)(1 + |x| ) 2 m/2 in the distributional sense. In particular, if γ = 0, v(x) = g0(x)(1 + |x| ) is slowly increasing which proves (a). The part (b) follows from the following result whose proof is postponed to Section ??, see Theorem ?? Proposition 5.23. If a function u is slowly increasing and its distributional derivatives k−1,1 of order less than or equal k, k ≥ 1, are slowly increasing functions, then u ∈ Cloc . 5.3. Tempered distributions with compact support. We will investigate now prop- erties of the Fourier transform of tempered distributions with compact support. n 0 ∞ Lemma 5.24. Let Ω ⊂ R be open and u ∈ Sn. If u[ϕ] = 0 for all ϕ ∈ C0 (Ω), then u[ϕ] = 0 for all ϕ ∈ Sn such that supp ϕ ⊂ Ω. ∞ n Proof. Let ϕ ∈ Sn, supp ϕ ⊂ Ω and let η be a cut-off function, i.e. η ∈ C0 (R ), 0 ≤ η ≤ 1, η(x) = 1 for |x| ≤ 1 and η(x) = 0 for |x| ≥ 2. One can easily prove that η(x/R)ϕ(x) → ϕ(x) ∞ in Sn as R → ∞. Since η(x/R)ϕ(x) ∈ C0 (Ω), the lemma follows. 0 Definition 5.25. Let u ∈ Sn. The support of u (supp u) is the intersection of all closed sets E ⊂ Rn such that ∞ n ϕ ∈ C0 (R \ E) ⇒ u[ϕ] = 0. 0 ∞ n Proposition 5.26. If u ∈ Sn and ϕ ∈ C0 (R \ supp u), then u[ϕ] = 0. Proof. Indeed, complement of supp u is the union of all open sets Ui such that if ϕ ∈ ∞ ∞ S C0 (Ui), then u[ϕ] = 0, and we need to prove that if ϕ ∈ C0 ( i Ui), then still u[ϕ] = 0. Since the support of ϕ is compact we can select a finite subfamily of open sets Ui covering the support and the result follows from a simple argument involving partition of unity. Thus the support of u is the smallest closed set such that the distribution vanishes on ∞ C0 functions supported outside that set. ∞ n The lemma shows that we can replace ϕ ∈ C0 (R \E) in the above definition by ϕ ∈ Sn with supp ϕ ⊂ Rn \ E. HARMONIC ANALYSIS 65 Before we state the next result we need some facts about analytic and holomorphic functions in several variables. Definition 5.27. We say that a function f :Ω → C defined in an open set Ω ⊂ Rn is R-analytic, if in a neighborhood on any point x0 ∈ Ω it can be expanded as a convergent power series X α (5.8) f(x) = aα(x − x0) , aα ∈ C, α i.e. if in a neighborhood of any point f equals to its Taylor series. We say that a function f :Ω → C defined in an open set Ω ⊂ Cn is C-analytic if in a neighborhood of any point z0 ∈ Ω it can be expanded as a convergent power series X α f(z) = aα(z − z0) . α Rn has a natural embedding into Cn, just like R into C. n n R 3 x = (x1, . . . , xn) 7→ (x1 + i · 0, . . . , xn + i · 0) = x + i · 0 ∈ C . It is easy to see that an R-analytic function f in Ω ⊂ Rn extends to a C-analytic function f˜ in an open set Ω˜ ⊂ Cn,Ω ⊂ Ω.˜ Namely, if f satisfies (5.8), we set ˜ X α f(z) = aα(z − z0) , z0 = x0 + i · 0. α On the other hand, if f˜ is C-analytic in Ω˜ ⊂ Cn, then the restriction f of f˜ to Ω = Ω˜ ∩ Rn is R-analytic. For example for any ξ ∈ Rn, f(x) = ex·ξ is R-analytic and f˜(z) = ez·ξ is its C-analytic extension. Definition 5.28. We say that a continuous function f :Ω → C defined in an open set Ω ⊂ Cn is holomorphic if ∂f = 0 for i = 1, 2, . . . , n. ∂zi It is easy to see that C-analytic function are holomorphic, but the converse implication is also true. Lemma 5.29 (Cauchy). If f is holomorphic in45 n 1 1 n D (w, r) = D (w1, r1) × ... × D (wn, rn) ⊂ C and continuous in the closure Dn(x, r), then Z Z 1 f(ξ) dξ1 . . . dξn (5.9) f(z) = n ... 1 1 (2πi) ∂D (w1,r1) ∂D (wn,rn) (ξ1 − z1) ... (ξn − zn) for all z ∈ Dn(w, r). 45Product of one dimensional discs. 66 PIOTR HAJLASZ Proof. The function f is holomorphic in each variable separately, so (5.9) follows from one dimensional Cauchy formulas and the Fubini theorem. 2 Just like in the case of holomorphic functions of one variable one can prove that the integral on the right hand side of (5.9) can be expanded as a power series and hence defines a C-analytic function. We proved Theorem 5.30. A function f is holomorphic in Ω ⊂ Cn if and only if it is C-analytic. Now we can state an important result about compactly supported distributions. 0 ∞ Theorem 5.31. If u ∈ Sn has compact support, then uˆ is a slowly increasing C function and all derivatives of uˆ are slowly increasing. Moreover uˆ is R-analytic on Rn and has a holomorphic extension to Cn. ∞ n Proof. Let η ∈ C0 (R ) be such that η(x) = 1 in a neighborhood of supp u. Then u = ηu 0 in Sn and hence for ψ ∈ Sn we have Z uˆ[ψ] = (ηu)ˆ[ψ] = u η(·)ψ(x)e−2πix·(·) dx Rn Z = u η(·)e−2πix·(·) ψ(x) dx . Rn We could pass with u under the sign of the integral, because of an argument with approx- imation of the integral by Riemann sums.46 Note that the function −2πix·(·) F (x1, . . . , xn) = u η(·)e is C∞ smooth and α α −2πix·(·) D F (x1, . . . , xn) = u (−2πi(·)) η(·)e . Indeed, we could differentiate under the sign of u because the corresponding difference quotients converge in the topology of Sn. It also easily follows from Theorem 5.3 that F and all its derivatives are slowly increasing. Thus we may identifyu ˆ with F , sou ˆ ∈ C∞. Moreover F has a holomorphic extension to Cn by the formula −2πiz·(·) F (z1, . . . , zn) = u η(·)e so in particular F (x1, . . . , xn) is R-analytic. 2 0 Remark 5.32. Note that if u ∈ Sn has compact support, then we can reasonably define u[e−2πix·(·)] by the formula ue−2πix·(·) := uη(·)e−2πix·(·) , ∞ n where η ∈ C0 (R ) is such that η = 1 in a neighborhood of supp u. Note that this con- struction does not depend on the choice of η. In the same way we can define u[f] for any f ∈ C∞(Rn) 46Compare with the proof of Theorem 5.7. HARMONIC ANALYSIS 67 0 Theorem 5.33. If u ∈ Sn and supp u = {x0}, then there is an integer m ≥ 0 and complex numbers aα, |α| ≤ m such that X α u = aαD δx0 . |α|≤m Proof. Without loss of generality we may assume that x0 = 0. According to Theorem 5.3 the distribution u satisfies the estimate X |u[ϕ]| ≤ C pα,β(ϕ) . |α|,|β|≤m First we will prove that for ϕ ∈ Sn we have (5.10) Dαϕ[0] = 0 for |α| ≤ m =⇒ u[ϕ] = 0 . Indeed, it follows from Taylor’s formula that ϕ(x) = O(|x|m+1) as x → 0 and hence also (5.11) Dβϕ(x) = O(|x|m+1−|β|) as x → 0 for all |β| ≤ m. ∞ n Let η ∈ C0 (R ) be a cut-off function, i.e. 0 ≤ η ≤ 1, η(x) = 1 for |x| ≤ 1, η(x) = 0 for |x| ≥ 2 and define ηε(x) = η(x/ε). The estimate (5.11) easily implies that pα,β(ηεϕ) → 0 as ε → 0 47 for all |α|, |β| ≤ m. Note that ϕ − ηεϕ = 0 in a neighborhood of 0, so u[ϕ − ηεϕ] = 0 and hence X |u[ϕ]| ≤ |u[ϕ − ηεϕ]| + |u[ηεϕ]| ≤ 0 + pα,β(ηεϕ) → 0 as ε → 0. |α|,|β|≤m This completes the proof of (5.10). Let now ψ ∈ Sn be arbitrary and let X Dαψ(0) h(x) = ψ(x) − xα . α! |α|≤m Clearly (5.12) Dαh(0) = 0 for |α| ≤ m. We have X Dαψ(0) ψ(x) = η(x) xα + η(x)h(x) + (1 − η(x))ψ(x) . α! |α|≤m Since (1−η)ψ vanishes in a neighborhood of 0 we have u[(1−η)ψ] = 0. The equality (5.12) ∞ n implies that ϕ = ηh ∈ C0 (R ) satisfies the assumptions of (5.10), so u[ηh] = 0. Hence h X Dαψ(0) i u[ψ] = u η(x) xα α! |α|≤m 47Check it! 68 PIOTR HAJLASZ X η(x)xα = (−1)|α|u (−1)|α|Dαψ(0) . α! |α|≤m | α{z } | {z } D δ0[ψ] aα The proof is complete. 2 0 2πix·ξ0 Corollary 5.34. Let u ∈ Sn. If suppu ˆ = {ξ0}, then u = P (x)e for some polynomial P (x) in Rn. In particular if suppu ˆ = {0}, then u is a polynomial. We leave the proof as an exercise. 0 Corollary 5.35. If u ∈ Sn satisfies ∆u = 0, then u is a polynomial. Proof. We have 2 2 0 −4π |ξ| uˆ = (∆u)ˆ = 0 in Sn. This implies that suppu ˆ = {0}, so u is a polynomial by Corollary 5.34. 2 5.4. Homogeneous distributions. A very important class of distributions is provided by so called homogeneous distributions that we will study now. Definition 5.36. A function f ∈ C(R \{0}) is homogeneous of degree a ∈ R if for all x 6= 0 and all t > 0 we have f(tx) = taf(x). More generally a measurable function on Rn is homogeneous of degree a ∈ R if a n f(tx) = t f(x) for almost all (x, t) ∈ R × (0, ∞). If f ∈ C(Rn \{0}) is homogeneous of degree a > −n, then it defines a tempered distribution. Indeed, a x a |f(x)| = |x| f ≤ C|x| |x| and |x|a is integrable in a neighborhood of the origin (because a > −n). If f is a measurable homogeneous function of degree a we can assume (by changing f on a set of measure zero) that f is continuous on almost all open rays going out of 0. Then on the remaining rays (which form a set of measure zero) we can make f homogeneous. This is to say that we can find a representative of f (in class of functions that are equal a.e.) such that f(tx) = taf(x) for all x 6= 0 and t > 0. If a measurable function that is not a.e. equal to zero is homogeneous of degree a ≤ −n, then the integration in the spherical coordinate system shows that it is not integrable in any neighborhood of 0 and hence it cannot define a tempered distribution. 1 n Lemma 5.37. Suppose that f ∈ Lloc(R ) defines a tempered distribution. Then the func- 48 tion f is homogeneous of degree a ∈ R if and only if a (5.13) f[ϕt] = t f[ϕ] for all ϕ ∈ Sn and all t > 0. Remark 5.38. If f is not equal to zero a.e., then necessarily a > −n. 48 −n As always ϕt(x) = t ϕ(x/t). HARMONIC ANALYSIS 69 Proof. For ϕ ∈ Sn we have Z Z Z a a f[ϕt] = f(x)ϕt(x) dx = f(tx)ϕ(x) dx and t f[ϕ] = t f(x)ϕ(x) dx. Rn Rn Rn Hence condition (5.13) is equivalent to the equality Z Z a (5.14) f(tx)ϕ(x) dx = t f(x)ϕ(x) dx for all ϕ ∈ Sn and all t > 0 Rn Rn which, in turn, is equivalent to the fact that for all t > 0, equality f(tx) = taf(x) holds a.e. The above result motivates the following definition. 0 49 Definition 5.39. We say that a distribution u ∈ Sn is homogeneous of degree a ∈ R if a u[ϕt] = t u[ϕ] for all ϕ ∈ Sn and all t > 0. 0 0 Proposition 5.40. u ∈ Sn is homogeneous of degree a ∈ R if and only if uˆ ∈ Sn is homogeneous of degree −n − a. Proof. Let ϕ ∈ Sn. Then −n −1 −n −1 −n ϕbt(ξ) =ϕ ˆ(tξ) = t (t ) ϕˆ(ξ/t ) = t (ϕ ˆ)t−1 (ξ) and hence −n −n−a −n−a uˆ[ϕt] = u[ϕbt] = t u[(ϕ ˆ)t−1 ] = t u[ϕ ˆ] = t uˆ[ϕ]. 0 If u ∈ Sn is homogeneous of degree a, the above equality shows thatu ˆ is homogeneous of degree −n − a. Now ifu ˆ is homogeneous of degree −n − a, then its Fourier transform i.e., u˜ is homogeneous of degree −n − (−n − a) = a and hence also u is homogeneous of degree a. Example 5.41. u(x) = |x|a, a ≥ 0, is homogeneous of degree a and it defines a tempered distribution. However,u ˆ is homogeneous of degree −n − a ≤ −n and hence it cannot be represented as a function. The notion of a homogeneous distribution will play a significant role in the proof of the next result. Theorem 5.42. For 0 < a < n, n n−a a− 2 −a π Γ 2 a−n F |x| (ξ) = a |ξ| . Γ 2 Remark 5.43. |x|−a is homogeneous of degree −a > −n and |ξ|a−n is homogeneous of degree a − n > −n so both |x|−a and |ξ|a−n are tempered distributions represented as functions. 49Now a can be any real number and we do not impose restriction a > −n that was necessary in the case of functions. 70 PIOTR HAJLASZ Proof. First we will prove the theorem under the assumption that n/2 < a < n. In that case −a −a −a 1 2 |x| = |x| χ{|x|≤1} + |x| χ{|x|>1} ∈ L + L . −a 2 Thus the Fourier transform of |x| is a function in C0 + L . The Fourier transform commutes with orthogonal transformations f[◦ ρ = fˆ◦ ρ, for ρ ∈ O(n) and f ∈ L1 + L2. If f is radially symmetric i.e., f = f ◦ ρ for all ρ ∈ O(n), then fˆ ◦ ρ = f[◦ ρ = fˆ for all ρ ∈ O(n) so fˆ is radially symmetric too. Thus the Fourier transform of |x|−a ∈ L1 + L2 is 2 a radially symmetric function in C0 + L , homogeneous of degree a − n as a distribution. Hence Lemma 5.37 yields −a a−n F |x| (ξ) = ca,n|ξ| −π|x|2 and it remains to compute the coefficient ca,n. Employing the fact that ϕ(x) = e is a fixed point of the Fourier transform we have a−n −a −a −a ca,n|x| [ϕ] = F(|x| )[ϕ] = |x| [ϕ ˆ] = |x| [ϕ], i.e., Z Z −π|x|2 a−n −π|x|2 −a (5.15) ca,n e |x| dx = e |x| dx. Rn Rn The integrals in this identity are easy to compute. Lemma 5.44. For γ > −n we have Z n+γ −π|x|2 γ Γ 2 e |x| dx = γ n . n 2 R π Γ 2 Proof. For γ > −n we have50 Z Z ∞ −π|x|2 γ −πs2 γ+n−1 e |x| dx = nωn e s ds Rn 0 Z ∞ t=πs2 nωn n+γ −t 2 −1 = n+γ e t dt 2π 2 0 nω n + γ Γ n+γ = n Γ = 2 . n+γ γ n 2 2 2 2π π Γ 2 Applying the lemma to both sides of (5.15) yields a n−a n−a Γ 2 Γ 2 a− n Γ 2 c = so c = π 2 . a,n a−n n − a n a,n a 2 π 2 Γ Γ π Γ 2 2 2 This completes the proof when n/2 < a < n. 50 n Recall that the volume of the unit sphere in R equals nωn so the first equality is just integration in 2πn/2 the spherical coordinates. In the last equality we use ωn = nΓ(n/2) , see (3.9). HARMONIC ANALYSIS 71 Suppose now that 0 < a < n/2. Since n/2 < n − a < n we have n a 2 −a −(n−a) π Γ 2 −a F |x| (ξ) = n−a |ξ| Γ 2 and applying the Fourier transform to this equality yields n a 2 −a a−n π Γ 2 −a | − x| = n−a F |ξ| (x). Γ 2 Replacing x by ξ and ξ by x yields the result when 0 < a < n/2. We are left with the case of a = n/2. Take a sequence n/2 < ak < n, ak → a = n/2. Since the result is true for ak we have n a − n−ak Z π k 2 Γ Z |x|−ak ϕˆ(x) dx = 2 |x|ak−nϕ(x) dx. ak Rn Γ 2 Rn Passing to the limit ak → a = n/2 yields the result for a = n/2. ak−n a−n −ak −a 0 Remark 5.45. Clearly |ξ| → |ξ| and |x| → |x| in Sn. The last convergence −ak −a 0 and the continuity of the Fourier transform yields that F(|x| ) → F(|x| ) in Sn. Hence the result for a = n/2 follows by passing to the limit in the equality n a − n−ak π k 2 Γ F |x|−ak (ξ) = 2 |ξ|ak−n in 0 . ak Sn Γ 2 Remark 5.46. In the case a = n/2 we obtain a particularly simple formula Z Z − n − n − n − n F(|x| 2 ) = |x| 2 , i.e., |x| 2 ϕˆ(x) dx = |x| 2 ϕ(x) dx Rn Rn for all ϕ ∈ Sn. This formula can be regarded as a version of the Poisson summation formula.51 5.5. The principal value. 0 α Proposition 5.47. If u ∈ Sn is homogeneous of degree a, then D u is homogeneous of degree a − |α|. We leave the proof as an (easy) exercise. In particular if f ∈ C∞(Rn \{0}) is homogeneous of degree 1 − n, then it defines a distribution, but its first order distributional partial derivatives are homogeneous of degree −n and hence they cannot be represented as functions. It is however, still possible to represent these distributional derivatives as improper integrals, the so called principal value of the integral. Definition 5.48. If for any ε > 0, a function f is integrable in Rn \B(0, ε), then we define the principal value of the integral as Z Z p.v. f(x) dx = lim f(x) dx ε→0 Rn |x|≥ε provided the limit exists. 51cf. Theorem 4.1. 72 PIOTR HAJLASZ If a function f is slowly increasing, but has singularity at 0, then we can try to define a distribution p.v. f by Z Z (p.v. f)[ϕ] = p.v. f(x)ϕ(x) dx = lim f(x)ϕ(x) dx for ϕ ∈ Sn. ε→0 Rn |x|≥ε In many interesting cases p.v. f defines a tempered distribution, but not always. Theorem 5.49. Let Ω(x/|x|) K (x) = , x 6= 0, Ω |x|n where Ω ∈ L1(Sn−1) satisfies Z Ω(θ) dσ(θ) = 0. Sn−1 Then 0 p.v.KΩ ∈ Sn and |(p.v.KΩ)[φ]| ≤ C(n)kΩkL1(Sn−1) k∇ϕk∞ + k|x|ϕ(x)k∞ for all ϕ ∈ Sn. 0 Remark 5.50. Distributions p.v.KΩ ∈ Sn described in the above result will be called distributions of the p.v. type or simply p.v. distributions. Remark 5.51. Note that the function KΩ is homogeneous of degree −n so the function KΩ does not define a tempered distribution. R Remark 5.52. It is not difficult to verify that if Sn−1 Ω(θ) dσ(θ) 6= 0, then p.v.KΩ does not define a tempered distribution because p.v.KΩ[ϕ] will diverge for some functions ϕ ∈ Sn. Proof. For ϕ ∈ Sn we have Z Z Ω(x/|x|) Ω(x/|x|) (p.v.KΩ)[ϕ] = lim n (ϕ(x) − ϕ(0)) + n ϕ(x) dx ε→0 ε≤|x|≤1 |x| |x|≥1 |x| Z |Ω(x/|x|)| Z |Ω(x/|x|)| ≤ k∇ϕk∞ n−1 dx + kϕ(y)|y|k∞ n+1 dx |x|≤1 |x| |x|≥1 |x| and the result follows upon integration in spherical coordinates. We could subtract ϕ(0) R Ω(x/|x|) in the first integral because ε≤|x|≤1 |x|n dx = 0. In the last step used the following estimates Z 1 Z 1 d |ϕ(x) − ϕ(0)| = ϕ(tx) dt = ∇ϕ(tx) · x dt ≤ |x| k∇ϕk∞ 0 dt 0 and |ϕ(x)| |x| kϕ(y)|y|k |ϕ(x)| = ≤ ∞ . |x| |x| HARMONIC ANALYSIS 73 The next result illustrates a simple situation when the principal value appears in a n+1 natural way. Note that the function xj/|x| satisfies the assumptions of Theorem 5.49. However, the proof of Proposition 5.53 will be straightforward and we will not use Theo- rem 5.49. 1−n 0 Proposition 5.53. For n ≥ 2, the distributional partial derivatives of |x| ∈ Sn satisfy ∂ 1−n xj 0 |x| = (1 − n) p.v. n+1 ∈ Sn, ∂xj |x| i.e., Z ∂ 1−n xj (5.16) |x| [ϕ] = (1 − n) lim n+1 ϕ(x) dx. for ϕ ∈ Sn. ∂xj ε→0 |x|≥ε |x| Before we prove the proposition let us recall the integration by parts formula for functions defined in a domain in Rn. If Ω ⊂ Rn is a bounded domain with piecewise C1 boundary and f, g ∈ C1(Ω), then Z Z (5.17) (∇f(x)g(x) + f(x)∇g(x)) dx = fg~νdσ , Ω ∂Ω where ~ν = (ν1, . . . , νn) is the unit outer normal vector to ∂Ω. Comparing jth components on both sides of (5.17) we have Z ∂f Z ∂g Z (5.18) (x) g(x) dx = − f(x) (x) dx + fg νj dσ . Ω ∂xj Ω ∂xj ∂Ω Proof of Proposition 5.53. Let A(ε, R) = {x : ε ≤ |x| ≤ R}. We have ∂ Z ∂ϕ |x|1−n [ϕ] = − |x|1−n dx ∂xj Rn ∂xj Z ∂ϕ = lim lim − |x|1−n dx ε→0 R→∞ ε≤|x|≤R ∂xj Z Z ∂ 1−n 1−n = lim lim ϕ(x) |x| dx − ϕ(x)|x| νj dσ ε→0 R→∞ ε≤|x|≤R ∂xj ∂A(ε,R) | {z n+1} (1−n)xj /|x| Z Z xj 1−n = lim (1 − n) ϕ(x) n+1 − ϕ(x)|x| νj dσ ε→0 |x|≥ε |x| |x|=ε Indeed, we could pass to the limit as R → ∞, because the part of the integral over ∂A(ε, R) 1−n corresponding to {|x| = R} converges to zero. Since the integral of the function |x| νj over the sphere |x| = ε equals zero we have Z Z 1−n 1−n ϕ(x)|x| νj dσ = (ϕ(x) − ϕ(0))|x| νj dσ ≤ Cε → 0 as ε → 0, |x|=ε |x|=ε because 1−n 1−n 2−n ϕ(x) − ϕ(0) |x| νj ≤ Cε · ε = ε n−1 and the surface area of the sphere {|x| = ε} equals Cε . 74 PIOTR HAJLASZ n+1 Remark 5.54. We proved that p.v. (xj/|x| ) defines a tempered distribution somewhat 1−n 0 indirectly by showing that for ϕ ∈ Sn equality (5.16) is satisfied. Since ∂|x| /∂xj ∈ Sn, n+1 0 equality (5.16) implies that p.v. (xj/|x| ) ∈ Sn. Remark 5.55. When n = 1 the counterpart of Theorem 5.53 is that d x 1 (log |x|) = p.v. = p.v. in S 0 . dx |x|2 x n We leave the proof as an exercise. We plan to generalize Proposition 5.53 to a class of more general functions. Let’s start with the following elementary result. Theorem 5.56. Suppose that K ∈ C1(Rn \{0}) is such that both K and |∇K| have polynomial growth52 for |x| ≥ 1 and there are constants C, α > 0 such that C |K(x)| ≤ for 0 < |x| < 1, |x|n−1−α C |∇K(x)| ≤ for 0 < |x| < 1. |x|n−α 0 Then K ∈ Sn and the distributional partial derivatives ∂K/∂xj, 1 ≤ j ≤ n, coincide with pointwise derivatives, i.e. for ϕ ∈ Sn ∂K Z ∂ϕ Z ∂K [ϕ] := − K(x) (x) dx = (x) ϕ(x) dx . ∂xj Rn ∂xj Rn ∂xj Proof. Let A(ε) = {x : ε ≤ |x| ≤ ε−1}. From (5.18) we have ∂K Z ∂ϕ [ϕ] = lim − K(x) (x) dx ∂xj ε→0 ε≤|x|≤ε−1 ∂xj Z ∂K Z = lim (x) ϕ(x) dx − K(x)ϕ(x) νj dσ(x) . ε→0 ε≤|x|≤ε−1 ∂xj ∂A(ε) Since Z ∂K Z ∂K lim (x) ϕ(x) dx = (x) ϕ(x) dx ε→0 ε≤|x|≤ε−1 ∂xj Rn ∂xj it remains to show that Z lim K(x)ϕ(x) νj dσ = 0 . ε→0 ∂A(ε) The integral on the outer sphere |x| = ε−1 converges to 0 since K has polynomial growth and ϕ rapidly converges to 0 as |x| → ∞ and on the inner sphere |x| = ε we have Z C n−1 α K(x)ϕ(x) νj dσ ≤ n−1−α ε = Cε → 0 as ε → 0. |x|=ε ε The proof is complete. 52i.e., |K(x)| + |∇K(x)| ≤ C(1 + |x|)m for some C > 0, m ≥ 1 and all |x| ≥ 1. HARMONIC ANALYSIS 75 An interesting problem is the case α = 0, i.e. when K and ∇K satisfy the estimates C C (5.19) |K(x)| ≤ , |∇K(x)| ≤ for x 6= 0. |x|n−1 |x|n One such situation was described in Proposition 5.53. Here we make an additional assumption about K. We assume that K ∈ C1(Rn \{0}) is homogenelus of degree 1 − n, i.e. K(x/|x|) K(x) = for x 6= 0. |x|n−1 0 Since K is bounded in {|x| ≥ 1} and integrable in {|x| ≤ 1} we have K ∈ Sn and there is no need to interpret K through the principal value of the integral. The first estimate at (5.19) is satisfied. To see that the second estimate if satisfied too we observe that ∇K is homogeneous of degree −n. Indeed, for 1 ≤ j ≤ n and t > 0 ∂K ∂ ∂ (tx)t = (K(tx)) = t1−n K(x) ∂xj ∂xj ∂xj and hence (∇K)(tx) = t−n∇K(x) . Thus (∇K)(x/|x|) ∇K(x) = , x 6= 0 |x|n from which the second estimate at (5.19) follows. Observe that ∇K(x) is not integrable at any neighborhood of 0, but we may try to con- sider the principal value of ∇K(x), i.e. the principal value of each of the partial derivatives ∂K/∂xj. Theorem 5.57. Suppose that K ∈ C1(Rn \{0}) is homogeneous of degree 1 − n. Then ∇K(x) is homogeneous of degree −n. Moreover Z (5.20) ∇K(θ) dσ(θ) = 0 Sn−1 and 0 p.v. ∇K(x) ∈ Sn is a well defined tempered distribution, i.e. for each 1 ≤ j ≤ n ∂K 0 p.v. (x) ∈ Sn . ∂xj Finally the distributional gradient ∇K satisfies (5.21) ∇K = cδ0 + p.v. ∇K(x) , |{z} | {z } dist. pointwise where Z x c = K(x) dσ(x) . Sn−1 |x| 76 PIOTR HAJLASZ In other words for ϕ ∈ Sn and 1 ≤ j ≤ n we have ∂K Z ∂ϕ Z ∂K [ϕ] := − K(x) (x) dx = cjϕ(0) + lim (x)ϕ(x) dx , ε→0 ∂xj Rn ∂xj |x|≥ε ∂xj where Z xj cj = K(x) dσ(x) . Sn−1 |x| Remark 5.58. Note that Proposition 5.53 follows from this result: since K(x) = |x|1−n, R −n the constant c = Sn−1 |x| x dσ(x) = 0. 0 Remark 5.59. The fact that p.v. ∇K ∈ Sn is a straightforward consequence of (5.20) and Theorem 5.49. However, our argument will be direct without referring to Theorem 5.49. Proof. We already checked that ∇K(x) is homogeneous of degree −n. For r > 1 let A(1, r) = {x : 1 ≤ |x| ≤ r}. From the integration by parts formula (5.17) we have Z Z ∇K(x) dx = K(x) ~ν(x) dσ(x) 1≤|x|≤r ∂A(1,r) Z x Z x = − K(x) dσ(x) + K(x) dσ(x) = 0 . |x|=1 |x| |x|=r |x| Indeed, the last two integrals are equal by a simple change of variables and homogeneity of K. Thus the integral on the left hand side equals 0 independently of r. Hence its derivative with respect to r is also equal zero. d Z Z 0 = ∇K(x) dx = ∇K(θ) dσ(θ) . + dr r=1 1≤|x|≤r |x|=1 0 This proves (5.20). As we will see (5.20) implies that p.v. ∇K(x) ∈ Sn is a well defined tempered distribution as well as that the distributional gradient ∇K satisfies (5.21). Let ϕ ∈ Sn. We have Z ∇K[ϕ] := − K(x)∇ϕ(x) dx Rn Z = lim lim − K(x)∇ϕ(x) dx ε→0 R→∞ ε≤|x|≤R Z Z = lim lim ∇K(x) ϕ(x) dx − K(x)ϕ(x) ~ν(x) dσ(x) ε→0 R→∞ ε≤|x|≤R ∂A(ε,R) Z Z x = lim ∇K(x) ϕ(x) dx + K(x)ϕ(x) dσ(x) . ε→0 |x|≥ε |x|=ε |x| It remains to prove that Z x Z x lim K(x)ϕ(x) dσ(x) = ϕ(0) K(x) dσ(x) . ε→0 |x|=ε |x| |x|=1 |x| Let Z x Z x c = K(x) dσ(x) = K(x) dσ(x) . |x|=1 |x| |x|=ε |x| HARMONIC ANALYSIS 77 The last equality follows from a simple change of variables and homogeneity of K. We have Z x K(x)ϕ(x) dσ(x)(5.22) |x|=ε |x| Z x = cϕ(0) + K(x) ϕ(x) − ϕ(0) dσ(x) |x|=ε |x| → cϕ(0) as ε → 0. Indeed, for |x| = ε