Acta Mathematica Sinica, English Series Oct., 2000, Vol. 16, No.4, pp. 613–626
A Besov Estimate for Multilinear Singular Integrals
Wengu Chen Department of Mathematics, Capital Normal University, Beijing, 100037, P. R. China E-mail:
[email protected] Abstract
The main purpose of this paper is to prove a good λ inequality for a multilinear singular
integral in Rn . Keywords
Multilinear operator, Besov space, Good λ inequality
1991MR Subject Classification 42B20, 42B25
1
Introduction
A well-known result of Coifman, Rochberg and Weiss [1] states that the commutator operator Cf (g) = T (f · g) − f · T (g) (where T is a Calder´ on Zygmund singular integral operator) is bounded on some Lp (Rn ), 1 < p < ∞, if and only if f ∈ BMO. There are other links between the boundedness properties of the operator Cf and the smoothness of f . A particular case of Janson’s [2] result states that Cf : Lp (Rn ) → Lq (Rn ) is bounded, 1 < p < q < ∞, if and only if f ∈ Lipβ, β = n(1/p − 1/q). Here, Lipβ is the homogeneous Lipschitz space determined by the first difference operator. The operators considered here are the maximal and limiting operators of the singular integral C (a, b; f )(x) = where Pm (a; x, y) = a(x) −
|x−y|>
Pm1 (a; x, y)Pm2 (b; x, y)
1 (α) (y)(x |α|<m α! a
Ω(x − y) f (y)dy, |x − y|n+M −2
(1)
− y)α and M = m1 + m2 , Ω satisfies certain
homogeneity, smoothness and symmetry conditions. The BMO estimate for these operators has been obtained by Cohen and Gosselin (see [3]). This paper aims to prove a good λ inequality Received September 30, 1999, Accepted January 14, 2000 The research is supported by Beijing Education Commission Foundation and Beijing Natural Science Foundation (Project No. 1982005)
Wengu Chen
614
for the maximal function C∗ (a, b; f )(x) = sup>0 |C (a, b; f )(x)| if a(α) , |α| = m1 − 1 and b(η) , |η| = m2 − 1 belong to the Besov-Lipschitz class Λ˙ β . As a direct consequence, the limiting operator maps Lp (Rn ) to Lq (Rn ) if 1/q = 1/p − 2β/n.
2
Preliminaries
Throughout this paper we will work in Euclidean space Rn . Let α = (α1 , . . . , αn ) denote a multi-index and let |α| = α1 + · · · + αn denote the order of α. If b is a smooth function |α|
∂ b on Rn , b(α) will denote the partial derivative ∂xα1 ∂x α2 αn . Let Pm (b; x, y) denote the m-th 1 2 ···∂xn order Taylor series remainder of b at x expanded about y. More precisely,
Pm (b; x, y) = b(x) −
b(α) (y) (x − y)α , α!
(2)
|α|<m
where α! = α1 !α2 ! · · · αn ! and (x−y)α = (x1 −y1 )α1 · · · (xn −yn )αn . Let |E| denote the Lebesgue measure of a measurable set E ⊂ Rn . In this paper Q will denote a cube with sides parallel to the axes and if b is an integrable function on Q, mQ (b) will denote the average of b over Q i.e., |Q|−1 Q b(x)dx. Let M f (x) denote the Hardy Littlewood maximal function of f and
Mr f (x) = (M (|f |r )(x))1/r . We also need a variant version of the maximal function Mα, p (f )(x) which was introduced by Muckenhoupt and Wheeden in [4] and defined by 1/p 1 |f (y)|p dy . Mα, p (f )(x) = sup 1−αp/n x∈Q |Q| Q Chanillo [5] proved the following: Lemma 2.1
Let p1 < p < n/α, and 1/q = 1/p − α/n. Then Mα, p1 (f )q ≤ Cf p .
(3)
For β > 0, the Lipschitz space Λ˙ β is the space of functions f such that [β]+1
f Λ˙ β = sup
x, h∈Rn h=0
|∆h
f (x)| < ∞, |h|β
where ∆kh denotes the k-th difference operator. F˙pβ, ∞ is the homogeneous Triebel-Lizorkin space. Finally we will let C be a constant that may vary from line to line.
3
Statements of Results
Let Ω be homogeneous of degree zero, satisfy |Ω(x) − Ω(y)| ≤ C|x − y| for |x| = |y| = 1, and have vanishing moments up to order M − 2 over the unit sphere in Rn . Suppose a and b are functions with derivatives of order m1 − 1 and m2 − 1, respectively, in Λ˙ β , 0 < β < 1, and we
A Besov Estimate for Multilinear Singular Integrals
denote ∇m1 −1 aΛ˙ β =
|α|=m1 −1
615
a(α) Λ˙ β and ∇m2 −1 bΛ˙ β =
|η|=m2 −1
b(η) Λ˙ β . Now we
can state our main results. Theorem 3.1 (Main estimate)
For 1 < r < p < ∞, there exists γ0 > 0 such that for
0 < γ < γ0 and λ > 0 {x ∈ Rn : C∗ (a, b; f )(x) > 3λ, ∇m1 −1 aΛ˙ β ∇m2 −1 bΛ˙ β M2β, p (f )(x) ≤ γλ} ≤ Cγ r {x ∈ Rn : C∗ (a, b; f )(x) > λ}. Theorem 3.2
(4)
If f ∈ Lp (Rn ), 1 < p < ∞, then the limiting operator C(a, b; f )(x) =
lim→0 C (a, b; f )(x) exists almost everywhere and satisfies the estimate C(a, b; f )q ≤ C∇m1 −1 aΛ˙ β ∇m2 −1 bΛ˙ β f p ,
(5)
when 1/q = 1/p − 2β/n and 1/p > 2β/n. Theorem 3.3
If f ∈ Lp (Rn ), 1 < p < ∞, and 0 < β < 1/2, then C(a, b; f )F˙p2β, ∞ ≤ C∇m1 −1 aΛ˙ β ∇m2 −1 bΛ˙ β f p .
4
(6)
Some Lemmas and Proof of the Theorems
Lemma 4.1 (See [3])
Let b(x) be a function on Rn with m-th order derivatives in Lq (Rn )
where q > n. Then |Pm (b; x, y)| ≤ Cm, n |x − y|m
|α|=m
1 |Q(x, y)|
Q(x, y)
|b(α) (t)|q dt
1/q
,
(7)
where Q(x, y) is the cube centered at x with edges parallel to the axes and having diameter √ 5 n|x − y|. (a) For 0 < β < 1, 1 ≤ q < ∞, 1/q 1 1 q f Λ˙ β ≈ sup |f (x) − m (f )| dx . Q β/n |Q| Q |Q| Q
Lemma 4.2 (See [6])
(b) For 0 < β < 1, 1 < p < ∞, hF˙pβ, ∞ ≈ sup ·∈Q
Lemma 4.3 (See [7])
1 |Q|1+β/n
Q
|h(x) − mQ (h)|dx . p
Let Q∗ ⊂ Q. Then |mQ∗ (f ) − mQ (f )| ≤ Cf Λ˙ β |Q|β/n .
We now turn to the proof of the main estimate (4). Using a Whitney argument we write {x ∈ Rn : C∗ (a, b; f )(x) > λ} as a union of cubes {Qj } with mutually disjoint interiors and with distance from each Qj to Rn \ ∪j Qj comparable to the diameter of Qj . It is now sufficient
Wengu Chen
616
to prove the main estimate for each Qj . There exists a constant C = C(n) such that for ˜ j ) = C(n)diam(Qj ) intersects ˜ j with the same center as Qj but with diam(Q each j the cube Q ˜ j such that C∗ (a, b; f )(x0 ) ≤ λ. Rn \ ∪j Qj . Thus for each j there exists a point x0 = x0 (j) ∈ Q We now fix a cube Qj and assume that there exists a point z = z(j) with ∇m1 −1 aΛ˙ β ∇m2 −1 bΛ˙ β M2β, p (f )(z) ≤ γλ. ˜˜ and write f = f + f where ¯j = Q (If no such point exists, the result is trivial for Qj ). Let Q 1 2 j f1 = f χQ¯ j . We now make appropriate estimates on f1 and f2 separately. ¯ j , ϕ(x) ≡ The f1 estimate. We choose a C0∞ function ϕ such that ϕ(x) ≡ 1 for x ∈ Q ¯ , |ϕ(x)| ≤ 1 for all x, and for any multi-index α with |α| ≤ M , |ϕ(α) (x)| ≤ 0 for x ∈ Q j ¯ C(diam(Qj ))−|α| . We note that C is independent of j. We now define 1 mQj (a(α) )(·)α ; y, z ϕ(y), α! |α|=m1 −1 1 mQj (b(α) )(·)α ; y, z ϕ(y). bϕ (y) ≡ Pm2 −1 b(·) − α! aϕ (y) ≡ Pm1 −1 a(·) −
|α|=m2 −1
¯ . We now estimate the derivatives of order m − 1 We note that aϕ and bϕ have support in Q 1 j of aϕ . Let γ be a multi-index of order m1 − 1. Then
1 mQj (a(α) )(·)α ; y, z ϕ(ν) (y) α! γ=µ+ν |α|=m1 −1 µ ∂ 1 (α) α m = Cµ, ν Pm1 −1−|µ| (a )(·) a(·) − ; y, z ϕ(ν) (y). Q j µ ∂y α! γ=µ+ν
a(γ) ϕ (y) =
(µ) Cµ, ν Pm1 −1 a(·) −
|α|=m1 −1
From Lemmas 4.1, 4.2 (a) and 4.3 we have ∂µ a(·) − Pm1 −1−|µ| ∂y µ ≤ C|y − z|m1 −1−|µ|
1 mQj (a(α) )(·)α ; y, z α! |α|=m1 −1 1/q 1 |a(η) (t) − mQj (a(η) )|q dt |Q(y, z)| Q(y, z)
|η|=m1 −1 m1 −1−|µ|
≤ C|y − z|
¯ |β/n . ∇m1 −1 aΛ˙ β |Q j
From the assumptions on ϕ, we have |ϕ(ν) (y)| ≤ C|y − z|−|ν| . So for |γ| = m1 − 1 we have (γ) ¯ |β/n . Since a has support in Q ¯ , we finally obtain, for |γ| = m −1, |aϕ (y)| ≤ C∇m1 −1 a ˙ |Q ϕ 1 j j Λβ
m1 −1 ¯ |β/n+1/q . a(γ) aΛ˙ β |Q j ϕ q ≤ C∇
(8)
Similarly, for |γ| = m2 − 1 we have the estimate m2 −1 ¯ |β/n+1/q . b(γ) bΛ˙ β |Q j ϕ q ≤ C∇
(9)
A Besov Estimate for Multilinear Singular Integrals
617
¯j , We observe that for y ∈ Q Pm1 (a; x, y) = Pm1 Pm1 −1 (a; (·), z)ϕ(·); x, y 1 mQj (a(α) )tα ; (·), z ϕ(·); x, y = Pm1 Pm1 −1 a(t) − α! |α|=m1 −1
= Pm1 (aϕ ; x, y),
(10)
and likewise, Pm2 (b; x, y) = Pm2 (bϕ ; x, y). It now follows that for x ∈ Qj , C∗ (a, b; f1 )(x) = C∗ (aϕ , bϕ ; f1 )(x). By Cohen and Gosselin’s [3, p. 451] result, for 1/r = 1/p + 2/q < 1 we have {x ∈ Qj : C∗ (a, b; f1 )(x) > δλ} = {x ∈ Qj : C∗ (aϕ , bϕ ; f1 )(x) > δλ} C (a , b ; f ) r ∗ ϕ ϕ 1 r ≤C δλ r (η) ≤ C(δλ)−r a(α) b f q q 1 p ϕ ϕ
|α|=m1 −1
|η|=m2 −1
≤ C(δλ)−r ∇m1 −1 aΛ˙ β ∇m2 −1 bΛ˙ β M2β, p (f )(z)
r
¯ |r(2/q+1/p) |Q j
¯ | ≤ C(δλ)−r (γλ)r |Q j γ r |Qj |. ≤C δ
(11)
This completes the estimate on f1 . ˜ j ). Let K = K(n) be a large positive integer depending The f2 estimate for ( ≈ diam(Q ˜ j ) ≤ ( ≤ K(n) diam(Q ˜ j ) and only on n. The estimation for f2 is split into the two cases diam(Q ˜ j ) is ignored since f2 has support outside Q ¯ j . Let ˜ j ). The case of ( < diam(Q ( > K(n) diam(Q aj (x) = a(x) −
|α|=m1 −1
1 mQj (a(α) )xα and bj (x) = b(x) − α!
|η|=m2 −1
1 mQj (b(η) )xη . η!
Ω(x−y) We now set k(x, y) = Pm1 (aj ; x, y)Pm2 (bj ; x, y) |x−y| n+M −2 . n ˜ j with x0 ∈ R \ ∪ Qj , and for x ∈ Qj We now choose x0 in Q
|C (a, b; f2 )(x)| = |C (aj , bj ; f2 )(x)| ≤ [k(x, y) − k(x0 , y)]f2 (y)dy + k(x0 , y)f2 (y)dy |x−y|> |k(x0 , y)f (y)|dy + |k(x0 , y)f (y)|dy + R(x)
R(x0 )
+ |C (a, b; f )(x0 )|,
(13)
˜ j ) ≤ | · −y| ≤ K(n) diam(Q ˜ j ). The last integral where R(·) denotes the integral region diam(Q is bounded by λ since x0 ∈ ∪Qj . The middle integrals are error terms which we will estimate later. We now deal with the first integral. We have k(x, y) − k(x0 , y)
Ω(x − y) Ω(x0 − y) − |x − y|n+M −2 |x0 − y|n+M −2 Ω(x0 − y) + Pm1 (aj ; x, y) − Pm1 (aj ; x0 , y) Pm2 (bj ; x, y) |x0 − y|n+M −2 Ω(x − y) 0 + Pm1 (aj ; x0 , y) Pm2 (bj ; x, y) − Pm2 (bj ; x0 , y) |x0 − y|n+M −2 ≡ k1 (x, x0 , y) + k2 (x, x0 , y) + k3 (x, x0 , y).
= Pm1 (aj ; x, y)Pm2 (bj ; x, y)
For |x − y| > (, standard arguments imply Ω(x − y) |x − x0 | Ω(x0 − y) − . ≤C |x − y|n+M −2 |x0 − y|n+M −2 |x − y|n+M −1 We now write
|x−y|>
4 k1 (x, x0 , y)f2 (y)dy ≤ µ=1
|x−y|>
where
k11 (x, x0 , y) = Pm1 −1 (aj ; x, y)Pm2 −1 (bj ; x, y)
k12 (x, x0 , y) =
|α|=m1 −1
k1µ (x, x0 , y)f2 (y)dy ,
Ω(x − y) Ω(x0 − y) − , |x − y|n+M −2 |x0 − y|n+M −2
(x − y)α (α) aj (y) Pm2 −1 (bj ; x, y) α!
×
Ω(x − y) Ω(x0 − y) − , |x − y|n+M −2 |x0 − y|n+M −2 (x − y)η (η) bj (y) k13 (x, x0 , y) = Pm1 −1 (aj ; x, y) η! |η|=m2 −1
×
Ω(x − y) Ω(x0 − y) − , |x − y|n+M −2 |x0 − y|n+M −2 (x − y)α (x − y)η (α) (η) aj (y) bj (y) k14 (x, x0 , y) = α! η! |α|=m1 −1
×
|η|=m2 −1
Ω(x − y) Ω(x0 − y) − . |x − y|n+M −2 |x0 − y|n+M −2
(14)
(15)
(16)
A Besov Estimate for Multilinear Singular Integrals
619
We estimate these integrals separately. We have k11 (x, x0 , y)f2 (y)dy |x−y|>
P m1 −1 (aj ; x, y) Pm2 −1 (bj ; x, y) |f (y)| dy |x − y|m1 −1 |x − y|m2 −1 |x − y|n+1 |x−y|> 1/q 1 ˜j ) |a(α) (t) − mQj (a(α) )|q dt ≤ C diam(Q |Q(x, y)| |x−y|> Q(x, y) |α|=m1 −1 1/q 1 |f (y)| dy. × |b(η) (t) − mQj (b(η) )|q dt |Q(x, y)| Q(x, y) |x − y|n+1
˜j ) ≤ C diam(Q
|η|=m2 −1
Replacing mQj (a(α) ) by mQ(x, y) (a(α) ) and likewise for b(η) and then using an estimate from [8, Lemma 2.4] we obtain k11 (x, x0 , y)f2 (y)dy ≤ |x−y|>
˜ j )|Q ˜ j |2β/n C∇m1 −1 aΛ˙ β ∇m2 −1 bΛ˙ β diam(Q 2 |x−y| |f (y)| 1 + log diam( ˜ Q )
×
j
|x − y|n+1
|x−y|>
dy.
The last integral can be estimated as ˜ j |2β/n ˜ j )|Q diam(Q ≤
∞
∞ ν=0
2ν
˜j ) diam(Q α!
P (α) (α) m2 −1 (bj ; x, y) |a (y) − mQj (a )| |f (y)| dy |x − y|m2 −1 |x − y|n+1 |x−y|>
˜j ) diam(Q ˜ j |β/n ∇m2 −1 bΛ˙ β |Q α!
(17)
Wengu Chen
620
×
∞ ν=0
1 + log
|x−y| ˜j ) diam(Q
|a(α) (y) − mQj (a(α) )| |f (y)|dy |x − y|n+1
2ν
Ω(x0 − y) f (y)dy . 2 |x0 − y|n+M −2
(20)
The first integral in (20) is majorized by P m1 −1 (aj ; x, x0 ) Pm2 −1 (bj ; x, y) |f (y)|dy C|x − x0 | m1 −2 |x − y|m2 −1 |x − y|n+1 |x−y|> |x − x0 ||x − y| ≤ C∇m1 −1 aΛ˙ β ∇m2 −1 bΛ˙ β M2β, p (f )(z) ≤ Cγλ.
(21)
Each integral in the sum in (20) is majorized by
P (η) m1 −1 (aj ; x, x0 ) |bj (y)||f (y)|dy C|x − x0 | m1 −2 |x − y|n+1 |x−y|> |x − x0 ||x − y| ≤ C∇m1 −1 aΛ˙ β ∇m2 −1 bΛ˙ β M2β, p (f )(z) ≤ Cγλ.
(22)
We note that each of the above estimates follows from the same type of argument used in estimating the integral with k12 above. We now estimate the integral with k22 (x, x0 , y). We have
|x−y|>
k22 (x, x0 , y)f2 (y)dy
≤C
0
+C
1 α!
Pm1 −|α| (aj ; x0 , y)Pm2 −1 (bj ; x, y)
0
(x − x0 )α Ω(x0 − y) f (y)dy 2 |x0 − y|n+M −2
Pm1 −|α| (aj ; x0 , y)
(x − x0 )α (x − y)η Ω(x0 − y) (η) bj (y)f2 (y)dy n+M −2 |x0 − y|
≡ I 22 + I 22 . We now estimate I 22 by breaking up the first remainder. We have 1 |x − x0 | I 22 ≤ C α! 0
(23)
Wengu Chen
622
+C
0
(24)
By familiar arguments, each of these integrals is bounded by ∇m1 −1 aΛ˙ β ∇m2 −1 bΛ˙ β M2β, p (f )(z) ≤ Cγλ. We now estimate I 22 again by breaking up the first remainder. We have
I 22 ≤ C
0 1 1 |x − x0 | +C α!η! γ!
×
0
|γ|=m1 −|α|−1
(α+γ) (η) |aj (y)| |bj (y)| |f (y)| . |x − y|n+1
(25)
Again by familiar arguments each of these integrals is bounded by C∇m1 −1 aΛ˙ β ∇m2 −1 bΛ˙ β M2β, p (f )(z) ≤ Cγλ. This completes the estimate of the integral in (23) involving k22 (x, x0 , y). The same estimate holds for the integral involving k23 (x, x0 , y) by the above argument with the roles of a and b being interchanged. This now completes the estimate of the first term in (13). To summarize, we have shown that
|x−y|>
[k(x, y) − k(x0 , y)]f2 (y)dy ≤ Cγλ.
˜ j ) ≤ ( < K(n) diam(Q ˜ j ), it only remains to estimate To complete the f2 estimate for diam(Q the error terms in (13). We will estimate the first error term while the second one is handled similarly. We have R(x)
|k(x0 , y)f (y)|dy
Ω(x0 − y) |f (y)|dy |x0 − y|n+M −2 R(x) P m1 −1 (aj ; x0 , y) Pm2 −1 (bj ; x0 , y) |f (y)| ≤ dy m −1 m −1 |x0 − y| 1 |x0 − y| 2 |x − y|n+1 R(x) P (α) 1 m2 −1 (bj ; x0 , y) |aj (y)||f (y)| dy + α! R(x) |x0 − y|m2 −1 |x − y|n
=
|Pm1 (aj ; x0 , y)Pm2 (bj ; x0 , y)|
|α|=m1 −1
A Besov Estimate for Multilinear Singular Integrals
+
|η|=m2 −1
+
1 η!
623
P (η) m1 −1 (aj ; x0 , y) |bj (y)||f (y)| dy |x0 − y|m1 −1 |x − y|n R(x)
|α|=m1 −1 |η|=m2 −1
1 α!η!
R(x)
(α)
(η)
|aj (y)||bj (y)||f (y)| dy. |x − y|n
(26)
˜ ) K(n) diam(Q
j We note that for y ∈ R(x), 1 ≤ . Thus each error term is increased by multiplying |x−y| ˜ by K(n) diam(Qj ) on the outside and increasing the exponent of |x−y| under |f (y)| by 1. After
doing so each integral can be estimated by familiar arguments used earlier. Each error term is majorized by CK(n)∇m1 −1 aΛ˙ β ∇m2 −1 bΛ˙ β M2β, p (f )(z). It now follows that the entire error term is majorized by C(n)γλ. In summary we have shown that for any x ∈ Qj , sup≈diam(Q˜ j ) |C (a, b; f2 )(x)| ≤ Cγλ + λ. ˜ j ). Let Q denote the cube with sides parallel to the The f2 estimate for ( > K(n) diam(Q j 1 mQ j (a(α) )xα axes with the same center as Qj and diameter (. Let a (x) = a(x) − |α|=m1 −1 α! Ω(x−y) and let b (x) be defined in a similar manner. Let k (x, y) = Pm1 (a ; x, y)Pm2 (b ; x, y) |x−y| n+M −2 .
Proceeding as before we have |C (a, b; f2 )(x)| ≤ [k (x, y) − k (x0 , y)]f2 (y)dy + |k (x0 , y)f (y)|dy |x−y|> ≤|x−y|≤C |k (x0 , y)f (y)|dy + k (x0 , y)f (y)dy . (27) + ≤|x0 −y|≤C
|x0 −y|>
The last integral in (27) is bounded by λ since x0 ∈ ∪j Qj . The same estimates hold for the error terms since diam(Qj ) = (. For the first term we must be careful since |x − x0 | and diam(Qj ) are no longer comparable. To deal with such terms we select a point x such that |x − x | = 2(. Then ( < |x − x0 | < 3(. We then write Pm1 −1 (a ; x, x0 ) = Pm1 −1 (a ; x, x ) − Pm1 −1 (a ; x0 , x ) (x − x0 )α Pm1 −1−|α| (a(α) − ; x0 , x ). α!
(28)
0
Ω(x0 − y) f (y)dy 2 |x0 − y|n+M −2 P m1 −1 (a ; x, x ) Pm2 −1 (b ; x, y) |f (y)| ≤ C|x − x | dy |x − x |m1 −1 |x − y|m2 −1 |x − y|n+1 |x−y|> P m2 −1 (b ; x, y) |f (y)| ≤ C∇m1 −1 aΛ˙ β |Qj |β/n ( dy m −1 n+1 |x − y| 2 |x − y| |x−y|> ≤ C∇m1 −1 aΛ˙ β ∇m2 −1 bΛ˙ β M2β, p (f )(z) ≤ Cγλ.
(29)
˜ j ), we All the remaining estimates can be handled similarly. Ultimately, for ( > K(n) diam(Q obtain
|x−y|>
[k (x, y) − k (x0 , y)]f2 (y)dy ≤ Cγλ.
(30)
Wengu Chen
624
From (27) it now follows that sup ˜j ) >K(n) diam(Q
|C (a, b; f2 )(x)| ≤ λ + Cγλ.
(31)
The estimates on f2 now yield the pointwise estimate C∗ (a, b; f2 )(x) ≤ λ + Cγλ for all x ∈ Qj . We now choose γ0 such that Cγ0 < 1, where C is the constant in (31). Then from (11) with δ = 1, we have, for γ < γ0 , {x ∈ Qj : C∗ (a, b; f )(x) > 3λ, ∇m1 −1 aΛ˙ β ∇m2 −1 bΛ˙ β M2β, p (f )(x) ≤ γλ} ≤ {x ∈ Qj : C∗ (a, b; f1 )(x) > 2λ − Cγλ} + {x ∈ Qj : C∗ (a, b; f2 )(x) > λ + Cγλ} ≤ {x ∈ Qj : C∗ (a, b; f1 )(x) > λ} ≤ Cγ r |Qj |.
(32)
This establishes the good λ inequality and completes the proof of Theorem 3.1. Theorem 3.2 now follows from the good λ inequality and Lemma 2.1 by a standard good λ argument. We now turn to the proof of Theorem 3.3. For a cube Q and f ∈ Lp (Rn ), we decompose f = f1 + f2 with f1 = f χQ¯ . We fix a cube Q and x ∈ Q. Let x0 be a point on the boundary of ¯ We similarly define Q. 1 mQ (a(α) ); y, x0 ϕ(y), aϕ (y) ≡ Pm1 −1 a(·) − α! |α|=m1 −1 1 mQ (b(η) ); y, x0 ϕ(y), bϕ (y) ≡ Pm2 −1 b(·) − (33) η! |η|=m2 −1
and a ¯(y) ≡ a(y) −
|α|=m1 −1
¯b(y) ≡ b(y) −
|η|=m2 −1
1 mQ (a(α) )y α , α! 1 mQ (b(η) )y η . η!
˜ We have We now choose a point y0 on the boundary of Q. 1 |C(a, b; f )(y) − mQ (C(¯ a, ¯b; f ))|dy 1+2β/n |Q| Q 1 = |C(¯ a, ¯b; f (y) − mQ (C(¯ a, ¯b; f ))|dy |Q|1+2β/n Q 2 ≤ |C(¯ a, ¯b; f )(y) − C(¯ a, ¯b; f2 )(y0 )|dy |Q|1+2β/n Q 2 ≤ |C(¯ a, ¯b; f1 )(y)|dy |Q|1+2β/n Q 2 + |C(¯ a, ¯b; f2 )(y) − C(¯ a, ¯b; f2 )(y0 )|dy |Q|1+2β/n Q ≡ J1 + J2 .
(34)
(35)
A Besov Estimate for Multilinear Singular Integrals
625
To estimate the first term J1 , we use the fact that for y ∈ Q, C(¯ a, ¯b; f1 )(y) = C(aϕ , bϕ ; f1 )(y). If 1/r + 2/q = 1/s < 1, then from the previous results, we have 2 C(aϕ , bϕ ; f1 )s |Q|1−1/s |Q|1+2β/n 2 ≤ aϕ q bϕ q f1 r |Q|1−1/s |Q|1+2β/n C ≤ ∇m1 −1 aΛ˙ β ∇m2 −1 bΛ˙ β |Q|2β/n+2/q f1 r |Q|1−1/s 1+2β/n |Q| 1/r 1 ≤ C∇m1 −1 aΛ˙ β ∇m2 −1 bΛ˙ β |f (y)|r dy |Q| Q¯
J1 ≤
≤ C∇m1 −1 aΛ˙ β ∇m2 −1 bΛ˙ β Mr f (x).
(36)
To get the estimate of J2 , let k(y, z) = Pm1 (¯ a; y, z)Pm2 (¯b; y, z)
Ω(y − z) . |y − z|n+M −2
(37)
Then C(¯ a, ¯b; f2 )(y) − C(¯ a, ¯b; f2 )(y0 ) =
[k(y, z) − k(y0 , z)]f2 (z)dz.
(38)
Ω(y − z) Ω(y0 − z) − |y − z|n+M −2 |y0 − z|n+M −2 Ω(y0 − z) a; y, z) − Pm1 (¯ a; y0 , z) Pm2 (¯b; y, z) + Pm1 (¯ |y0 − z|n+M −2 Ω(y − z) 0 + Pm1 (¯ a; y0 , z) Pm2 (¯b; y, z) − Pm2 (¯b; y0 , z) |y0 − z|n+M −2 ≡ k1 (y, y0 , z) + k2 (y, y0 , z) + k3 (y, y0 , z).
(39)
Proceeding as in the proof of Theorem 3.1, we can obtain [k(y, z) − k(y0 , z)]f2 (z)dz ≤ C|Q|2β/n M f (x) + Mr f (x) .
(40)
Rn
To compute the last integral, we decompose k(y, z) − k(y0 , z)
= Pm1 (¯ a; y, z)Pm2 (¯b; y, z)
Rn
For example, the integrals involving k11 and k12 can be estimated as follows. Ω(y − z) Ω(y0 − z) Pm1 −1 (¯ a; y, z)Pm2 −1 (¯b; y, z) − (z)dz f 2 |y − z|n+M −2 |y0 − z|n+M −2 Rn a; y, z) Pm2 −1 (¯b; y, z) |y − y0 ||f2 (z)| Pm1 −1 (¯ dz ≤ |y − z|m1 −1 |y − z|m2 −1 |y − z|n+1 Rn ≤ C∇m1 −1 aΛ˙ β ∇m2 −1 bΛ˙ β |Q|2β/n diam(Q) 2 |y−z| |f (z)| 1 + log diam(Q) × dz n+1 |y − z| |z−y|>Cdiam(Q) ≤ C∇m1 −1 aΛ˙ β ∇m2 −1 bΛ˙ β |Q|2β/n M f (x)
(41)
Wengu Chen
626
and
Rn
|α|=m1 −1
× ≤
(y − z)α (α) a ¯ (z) Pm2 −1 (¯b; y, z) α!
Ω(y − z) Ω(y0 − z) − (z)dz f 2 |y − z|n+M −2 |y0 − z|n+M −2 P a(α) (z)||f (z)| diam(Q) m2 −1 (¯b; y, z) |¯ dz α! |y − z|m2 −1 |y − z|n+1 |z−y|>Cdiam(Q)
|α|=m1 −1
≤C
|α|=m1 −1
diam(Q) m2 −1 ∇ bΛ˙ β |Q|β/n α!
×
|z−y|>Cdiam(Q)
≤ C∇m2 −1 bΛ˙ β |Q|2β/n × ×
|y − z| |¯ a(α) (z)||f (z)| dz diam(Q) |y − z|n+1 ∞ −β 1 (ν + 2) ν+1 Cdiam(Q) 2 α! ν=0 2ν(1−β)
1 + log
1 (2ν+1 Cdiam(Q))n 1 (2ν+1 Cdiam(Q))n
|α|=m1 −1
|z−y|
