REVISTA MATEMATICA IBEROAMERICANA Vor. 8, N.° 1, 1992 Hardy Space Estimates | for Multilinear | Operators, II be Loukas Grafakos Abstract. We continue the study of multilinear operators given by products of finite vectors of Calderén-Zygmund operators. We determine the set of all <1 for which these operators map products of Lebesgue spaces 1”(R") into the Hardy spaces H’(R"). At the endpoint case r= n/(n + m + 1), where m is the highest vanishing moment of the multilinear operator, we prove a weak type result, 0. Introduction ‘A well-known by now theorem of P. L. Lions says that the derminant of the Jacobian of a function R"—R" maps the product of Sobolev spaces Lx +++ x LTinto the Hardy space H'. Coifman, Lions, Meyer and Semmes, {CLMS], went below H' by showing that for p,q>1, the Jacobian- determinant maps L7(R*) x L4(IR*) into H’(R°), where r~! = p-! + 7, as Jong as r> 2/3, Their result can be generalized to give the n-dimensional version that the determinant of the Jacobian maps L?R") x -+- x L?«(R") into H'(R"), as long as the harmonic mean r of the p/'s is strictly greater than n/(n + 1). In this work we prove a positive result in the endpoint case r=nj(n +1), We treat more general multilinear operators with vanishing o70 Loukas Grasaxos integral since our methods show that this is the only assumption needed. We also study the case of multilinear operators with higher moments vanishing. ‘The number of vanishing moments is related to the lowest r for which these ‘operators map products of Lebesgue spaces into H’. If such an operator has all moments of order n/(n+ m+ 1). Also, a weak type estimate holds in the endpoint case r=n/(n + m+ 1) and no boundedness result holds for renjin+ m+) 1, Statements of Results Throughout this article, N and K will denote fixed integers 32. We are given a matrix of sufficiently smooth convolution Calderén-Zygmund kernels ‘on R". We call 7/ the associated Calderén-Zygmund operator. by LUjy-+++f) the K-linear operator aay Mfr ho = TTI TES) originally defined for smooth compactly supported functions fj, .. fq. For <1, we denote by H” the usual real variable Hardy space as defined in [S] or (FSTI, i.e. the set of all distributions f on R" for which the maximal funetion sup |¢,+/0)| is in L?, where 4.0) = re(*) and @ is smooth, nonzero and compactly supported. We also denote by H”:~ the weak H? as defined in [FRS] (or [FSO] in the case p = 1), i.e. the set of all fin R" for which the maximal function sup,, 9,09] is in weak L? The weak L” (quasi)norm of this maximal function is by definition the 1 |e quasijnorm of f. ‘Our first result concerns the general multilinear operators L of the type above and it presents very clearly the method that will be used in this article, Note however, that there isan unpleasant restriction about the exponents that will be lifted later. ‘Theorem I. Assume that for all (f,,...sf,)€(C3)%, the K-linear operator L satisfies [Line srSgde = 0.Hanpy Space Esrmuarss ror MULTILINEAR OvERATORS, II 71 Suppose that py,..-5P> 1 are given and let r= (po! + +--+ pg)! be their harmonic mean. Assume that the harmonic mean of any proper subset of the p's is greater than 1. Then () If > 1, L maps LP) 0 KIEL, Q) 1dr > n/n My L maps LP x ++ x LPH, G) re njtnt i), L maps LP x «++ x LPR HO, Next, we treat the case of multilinear operators with vanishing higher moments. The significance of the number of vanishing moments is that it gives the lowest r for which such operators map into H". We also get rid of the assumption that the harmonic mean of any subset of the p's is always greater than 1. We are assuming however, that the K-linear operators L have a special form. When K = 2, we consider operators L of the general form (1.1), ie. inner products of two vectors of Calderén-Zygmund operators. For K > 3, we consider operators built inductively as follows: ‘We are assuming that for any j there exist, M= MT SpevSove So (K — 1)}inear operators already defined by the induction hypothesis with the same number of vanishing moments, such that (1.2) Ui Sie efi IDM. Condition (1.2) essetially says that the multilinear operators L look like determinants of matrices. They are built by induction starting from arbitrary bilinear operators as the ones in Theorem I (when K = 2) and at each stage they look like sums of products of multilinear operators of one smaller degree multiplied by a Calderén-Zygmund operator. These sums have a certain degree of symmetry because it follows from a repeated application of (1.2) that for each jy, ... ji, there exist (K ~ /)-linear operators Aje*""~" with the same number of vanishing moments such that Wire Q = BOL) TEM (remaining Fs). In most applications we have in mind, the multilinear operators have this form, for example determinants of matrices. In the case of bilinear operators, K = 2, there are no additional assumptions about the operators L and this is why we state and prove this case separately. ‘Also, this case is going to serve as the first step of an inductive argument that, will be used later.72 Louxas Gaazaxos Theorem HA. Assume that for some m,0 1 are arbitrary and let r = * be their harmonic ‘mean. Then () Ifr>1, B maps L? x L¢> 1, @) If 1d r>n/ln+ m+ 1), B maps L? x Lt H" (3) If r=n/(n + m+ 1), B maps L’ x L?> H"™. Next, we generalize Theorem Ila for K-linear operators of the form (1.2) and for these type of operators we do not have any additional assumption about the p,'s Theorem IB. Asume that for some m, 0 1 are arbitrary and let (pr) be ther harmonic mean. Then () Ifr> 1, L maps LP x +++ x EPL 2) 1 >r>n/(n + m+ 1), L maps LP! x +++ x LPR HT. (8) If r= niin + m1), L maps LP X06 xP HO, Remarks. (@ The assumption m < n(K — 1) ~ 1 is necessary in Theorem II, since other- wise r= n/(n + m +1) <1/K which would contradict that p,> 1. (®) The hypothesis that the harmonic mean of any subset of the ps is greater than 1 seems to be necessary in conclusions (2) and (3) of Theorem I. It is obviously not needed in conclusion 1) of Theorem I and it is always automatically satisfied when r= 1 or when K = 2.Hanby Space Esriwarss ror MULTILINEAR OPERATORS, II 73 ‘an upper bound on the degree K of multilinearity of the K-linear operator L. For, let p, = p > and let r > 1 be the harmonic mean of the p's. Then Kr = p. The assumption on the harmonic mean of any subset of the p/’s ives p/(K ~ 1) > 1. We conclude that K < 1/(1 ~ r) which isa restriction ‘on the size of K. Note, however, that when r= 1 there is no upper bound on K nor any restriction about the exponents and our theorem implies for example, that any K-linear operator as above with mean value zero maps DPX 0+ x DPE HY when Spy t= 1 (©) The vanishing integral hypothesis for Lin Theorem I can be relaxed to the milder condition that for all f, smooth with compact support and Sor some fyy...s,, in the corresponding Lebesgue spaces the integrals [LU fiv sy) vanish, Then conclusion (2) of Theorem I will be that the operator ¢-> L(g, -.-fx) maps L”! to H” with norm not larger than a constant times the product of the L% norms of the Js, , «4K. Conclusion (3) of Theorem I will be similar. 2. Proof of Theorem I We fix py,...,2,> 1 and we let r be their harmonic mean. Clearly only the case r<1 is interesting because the case r>1 is just Holder’s inequality together with the L? boundedness of Calderén-Zygmund operators. Fix a smooth compactly supported function ¢ in R", an x) €IR" and define 7 (54): Without sof enealty we may asm th supported ins <1 We 443409) sup Sx) | Joust is in L/ when r>n/(n + 1) and in L™™ when r= n/(n + 1). We also fix a smooth cut-off 7(x) such that » = 1 on |x| <2 and supported in |x| < 4. We call for simplicity wor=a(25%) and 92)=1—m9 ‘The reader should remember the dependence of np, on f. We now decompose Lys vorSy) = hot hy + Lea74 Loukas Gnaraxos where Lo = Lnofis tofas +++ + 20S) Liv esomdpe sete) tether Mdn bos I Len Simba ++ Si) In each L, above exactly u functions among the f's are multiplie by 7, and the remaining are left intact. To get this decomposition of L we expand LooLiy ++» oSx) = Li = Sir ++ Se— thd and then we solve for L(fis-+ ++.) Note that for any fixed i,k and any x such that |x — x9| < 1 we have sup ITFnAe) - THadea)| < sup fete =~ Ki — mF) dy | is bounded by CA" TT [Selim Term Ly is treated similarly. First write Ly = Ly + Ln + La +L where +1 VO) — LG ++ IVEY ~ (MLMO)s «+ sFids = AVC (MVC ~ (VCO righ -LVO) — MIMO)e OVO Fg + FIs +m16 Loukas Graraxos Same reasoning as before will show that any term Ly,, u = 1,2,3,4 satisfies the following estimate sup vefiten (94 (retan Jeradmar| < > B(. IL ITB yore pete. att where each C,f, is either |f* or (Taf and therefore |Cfylin < Cl Slims Now define a, by oj ' + pj ' + p;' = 1. Hélder’s inequality gives that for each u=1,2,3,4 [oe J easton <3 3 (dhe IG Gln 1GSilin n/(n+ 1), sup Jeratoa| < CII Ul") 2.3) When r=n/(n +1), sup Joractoa n/(n + 1) sup feasted | < ¢| Il cireo”| es CTT MG AI*0)""Lam SII Wilam where we used above that 74/5, > 1. We denote by |A| the measure of the set A. To derive conclusion (3) of Theorem I, let 1 and @,. be arbitrary. It follows from (2.3) that I By the weak type (1,1) result for the Hardy-Littlewood maximal function we get thatthe above is bounded by 1 ) j \glrae. 3 sup fe _tats|>}] <5 {xsagprea> (%+)"} . g ch78 Loukas Graraxos ‘This expression minimizes in ¢,, ... ,¢x > 0 when all the terms that appear in the sum are equal. This happens when ‘With this choice of ¢/s we get the weak type estimate focetat It remains to prove (2.2) and (2.3). We denote by 7* the adjoint operator of T and by Ig,» (T)"T the commutator of ¢,,., and (T')*. Since ¢ is a Lips- chitz function of order 1, it follows that foo:sup >| 1. We get 12.3)] < Pmofilan 2 Mss» TIED vo SOO" filam 3 Milas SCO" nef es 2 ITA) NTE rof dl ve SCO" nofil res dtoSel ++ VtoS el ive K n/(n + 1) ives bent, sh (iy = pet + + ag! = (> Therefore for a suitable selection of s,0 we obtain (2.2). The proof of Theorem I is now ‘complete. 3. Proof of Theorem Ia Clearly, we only need to do the case r < 1. Fix a @ and 7 as in Theorem I and split the bilinear operator B(/, g) as the sum of By + By, + By: + By where80. Loukas GraraKos Bo = Blof, 08), By = BS n8), By = Bh ®, B, = —Bin fg). ‘The arguments presented in Theorem I will give the required estimates for the terms B,,, By, and By. (Note the mean value zero assumption was only used. in the treatment of term Lg.) It remains to get the required estimates for term By which is the main term of the decomposition .We have Jato ae= [frereor0 dy de where 20,2) = mo n0(2) |B KPO — MKT Doug be The following lemma, whose proof we postpone until the end of this section describes the behavior of b,(),2)- Lemma 1, b,(y, 2) is @ smooth function off the diagonal y = z and satisfies the following estimate 10291 cartntiy—apetng(22)9( 222) Sor \y ~ z| small fete dx Assuming the lemma we estimate by ey creme lf Mole =a" ards, eee eee We denote by fy. the potential of order m + 1, ie. convolution with the kernel [x|""""*" in RY. Assume first that m+ 1 n/(n + m + 1) select p,

nfa+m+), 6.2) sup fento| < CUSIP *C)" Cs |*)*C0)" In the case r= n/(n + m + 1) simply repeat the argument above with p = p, and q = q,. We get if r= n/(n + m+ 1) 63) sup CUSIP") g1)*(%0))". fo Bod ‘Conclusions (2) and (3) of Theorem Ila follow as in Theorem I. In fact (3.2) and (3.3) are repetitions of (2.2) and (2.3) in Section 2. When m +1 = 7 only the case r>n/(n +m + 1) = 1/2 can occur. Then (3.2) follows from (3.1) directly from Hélder’s inequality. It remains to prove Lemma 1. We have that 1 (xo-¥, x0- noua fa BEE where 06,2) = nz) J DK iy ~ )K}HE ~ a)9(0) do. The estimate for b, in Lemma | is then equivalent to the estimate \b(y, 2)1 < Cly — 2I"**-*ny)n@2). ‘The vanishing moments assumptions for BCf, 2) are equivalent to the condi- tions JEK10-oK}e-oordo~0 for all. |a| 0. i Then TIM) ~ © Bod *TIO2'). T(@x)(w) is bounded near zero and has rapid decay as |w| ~> 0. Therefore the terms in the sum giving T(W) with 2/ < C/|| contribute C/|w| contribute CyBo, 3 2 Maw) ™ < CB! ‘pztok yn as |w|—>0. This finishes the proof of Lemma 2. ‘We conclude that (T7(a,, .)(2"- »)) is O(|y - 2|"*""") as |y - z| +0. Since het 3D 1TH, NE D) it follows that 1, satisfies the required estimate f, 1, the argument in Theorem I applies. Suppose then that s; < 1. We can assume by induction that VAlliy = LOD% I, < CIT Melr- By Holder's inequality, the L norm in xp of the last term in (4.1) is bounded by cx & VFN LOD Les + TDL IMD Mss kek 1, the argument in Theorem I establishes the result, If sy <1, we can assume by induction that VAM ae = OE La SCOTT Vida wet By Hélder’s inequality, the L’ norm in xp of the right hand side in (4.2) is bounded by Inslfoeae, SOB BIA k inl OAD a HEY" bd < © Pee 2 Vile ilen, ML Melon et

Py, ---.4g > Px such that Let ‘The dy’s are positive numbers and their sum is 1. The following lemma describes the behavior of (v1... +.)- Lemma 3. ,()1,..-43,) is a smooth function off the planes y, = y, and the satisfies the estimate [aor wal er" TL m0 Thain? Assuming the lemma, we prove the theorem as follows. Hélder’s inequality with exponents py! +gz'+-++-+gg'=1 together with (4.3) give the following |Jecntod| SON i, anelon Th Vons ny GXng-yead lanHanoy Seace Estimates rox MULTILINEAR OPERATORS, II 87 where by J, we denote convolution with |x|~*** on R*. By the Hardy- Littlewood-Sobolev fractional integral theorem we get that a wp Th iXig-ajerdn sup Cup" Xa, yl eael um fo. bods x < Couper"! TL sr aeegyeet = CTL USI”, since we are assuming that r = n/(n + m + 1). This estimate is the equivalent of (2.3) in Theorem I and the required weak type result follows as in Theorem I. In the case r > n/(n + m + 1), select s) 2. It follows that In 1 with 1/p + 1/q = 3/2, Jnever maps L? x L* to H®? and therefore our weak type result is sharp. In general, if a bilinear operator has all moments up to and including order m vanishing and one moment of order m + 1 nonzero, it does not map L? x L? to H™*"*” when 1/p + 1/q = (n+ m+ Din. Hence, the number of vanishing moments of the bilinear operator gives the lowest r for which the operator maps products of Lebesgue spaces into H”. The same is true for more general multilinear operators. We now discuss probably the most important example that satisfies the hypotheses of our Theorem Il, a bilinear map that involves sums of products of derivatives of order 2 and is the analogue of the determinant of the Jaco- bian. The Hessian of a map F = (f,g): IR’ > R* is the spacial 2 x 2 x 2 matrix which has on the top the 2 x 2 matrix:90 Loukas Graraxos af ef ox? Gxay a x ayax ay and on the bottom the 2x 2 matrix: Pe eg ax? axay ve ae ‘ayax ay? We denote by H(f,g) the determinant of the 2 x 2 x 2 Hessian matrix above defined as follows ae, Ff Fe of He | oF Hihg) = ff Me OF Oe, oF He (G8) = set ay? ~ Guay ayax ayae dxay * ay? ax? After formally replacing the partial derivatives of F with the corresponding Riesz transforms we get the following bilinear operator FS 8) = (RESMR¥8) — (Ry RoSMR2Rs 8R2Ri8) — (RR SMR RB) {twas shown in the last section of [CG], that AT has integral and first moments zero. By Theorem Ha it follows that Hf maps L? x L* to H’ for all p,q> 1, where ris their harmonic mean. It follows that the determinant of the Hessian ‘H, of a map F: + maps pairs of functions with Laplacean in L? x L# into H for all p,q > 1, where ri their harmonic mean. This result generalizes the correspoding theorem about the Jacobian in the case of second order derivatives and has analogues in higher dimensions. ‘We now discuss generalizations of HT in R*. Let F'= (Fy... F,) be a map from R" to R*. Form the n x --- x n matrix M by stacking the n Xn matrices Ox; 8x, }), on the top of each other. We call M the n-dimensional Hessian of F. The determinant of this matrix is defined by induction on 7 as the sum of its n? jor determinants suitably signed. After formally replacing the 4/4x, derivative of F by the j-th Riesz transform, the n-dimensional determinant of this matrix gives rise to an n-linear operator 7 with vanishing integral and first moments, When m = 2 the resulting bilinear operator His the operator ‘Aldefined above. When n = 3 the resulting trilinear operator A(f,g, h)is the sum of the nine terms (—1)~ */(R)R,A)A,(J, 2) where each 17, correspondsHaapy Space Estmares ror MULTILINEAR OPERATORS, II 91 to the determinant of a Hessian of a 2 x 2 x 2 minor. It is therefore clear that that 7° satisfies the hypotheses of Theorem IIb. Our result then says that that for p,q >1, ? maps L? x L* into H” when the harmonic mean r of p and q> 3/5. In general the operator A" maps L’ x L* into H’ for 1>r> n/(n + 2) since it follows by induction that A” has integral and first moments zero. (k = 1.) It is conceivable that determinants of matrices of higher order derivatives of maps from R" to IR" give rise to multilinear operators with higher moments vanishing but these cases are not investigated in this article. Examples of bilinear operators with moments of all orders vanishing in one dimension: Dy(S, 8) = fe — (HSH) and Dif) (He) + (He, where His the usual Hilbert transform. D, and D, are the real and imaginary parts of holomorphic functions and their mapping properties are well unders- tood. More generally, examples of K-linear operators with all moments vanishing are given by the real and imaginary parts of TI (fe + Hf). References [CG] Coifman, R.R. and Grafakos, L. Hardy Space Estimates for Multilinear Operators, I, Revista Mat. Iberoamericana &(1992), [CLMS} Coifman, R.R., Lions, P.L., Meyer, Y. and Semmes, S. Compacité par ‘compensation et espaces de Hardy, C. R. Acad. Sei. Paris. [CRW] Coifman, R.R., Rochberg, R. and Weiss, G. Factorization Theorems for Hardy Spaces in Several Variables, Annals of Math. 103(1976), 611-635. IFRS] Fefferman, C., Riviere, N.M. and Sagher, Y. Interpolation between H” Spaces: The Real Method, Trans. Amer. Math. Soc. 191(1974), 75-81 [FSO] Fefferman, R. and Soria, F. The Space Weak H', Studia Math. 85(1987), 116. [FST] Fefferman, C. and Stein, E.M. H” Spaces of Several Variables, Acta ‘Math, 229(1972), 137-193. [CMIII] Coifman, R.R. and Meyer, Y. Non Linear Harmonie Analysis, Operator ‘Theory and PDE, Beijing Lectures in Harmonic Analysis, Annals of Math ‘Studies, Princeton Univ. Press, 1986 [CMI] Coifman, R-R. and Meyer, Y. On Commutators of Singular Integrals and Bilinear Singular Integrals, Trans. Amer. Math. Soc. 212(1975), 315-331, [M1] Miyachi Tohoku Math. J. 35(1983), 483-498, [M2] Miyachi Pacific Math. J. 115(1984), 165-175.92, Loukas Graraxos [S] Stein, E.M. Singular Integrals and Differentiablity Properties of Func- tions, Princeton Univ. Press, 1970 [SW] Stein, E.M. and Weiss, G. Introduction to Fourier Analysis on Euclidean Spaces, Princeton Univ. Press, 1971 Recibido: 26 de agosto de 1991. Loukas Grafakos Department of Mathematics ‘Washington University in St Loui 1 Brookings Drive St Louis, MO 63130-4899 USA