Datasets:

kilian-group
/

arxiv-classifier

Dataset card Data Studio Files Files and versions Community

Subset (4)

default · ~99.6k rows (showing the first 120k)

Split (2)

train · ~72.4k rows (showing the first 93.1k)

paper_id stringlengths 9 16	version stringclasses 26 values	yymm stringclasses 311 values	created timestamp[s]	title stringlengths 6 335	secondary_subfield sequencelengths 1 8	abstract stringlengths 25 3.93k	primary_subfield stringclasses 124 values	field stringclasses 20 values	fulltext stringlengths 0 2.84M
1104.3324	1	1104	2011-04-17T16:13:27	S-Bases in S-Linear Algebra	[ "math.FA", "quant-ph" ]	In this paper we define the S-bases for the spaces of tempered distributions. These new bases are the analogous of Hilbert bases of separable Hilbert spaces for the continuous case (they are indexed by m-dimensional Euclidean spaces) and enjoy properties similar to those shown by algebraic bases in the finite dimensional case. The S-bases are one possible rigorous and extremely manageable mathematical model for the "physical" bases used in Quantum Mechanics.	math.FA	math	SBases in SLinear Algebra David Carf`ı Abstract In this paper we define the S bases for the spaces of tempered distri- butions. These new bases are the analogous of Hilbert bases of separable Hilbert spaces for the continuous case (they are indexed by m-dimensional Euclidean spaces) and enjoy properties similar to those shown by algebraic bases in the finite dimensional case. The Sbases are one possible rigorous and extremely manageable mathematical model for the "physical" bases used in Quantum Mechanics. 1 S Linear independence A finite family v of vectors of a vector space is said linearly independent if any zero linear combination of v is given by a zero system of coefficients. This is exactly the definition we generalize in the following. Definition (of Slinear independence). Let v ∈ S(Rm, S′ n) be an Sfamily of tempered distributions in the space S′ n indexed by the Euclidean space Rm. The family v is said to be S linearly independent, if any coefficient distribution a ∈ S′ m such that ZRm av = 0S ′ n m. In other terms the family v is said S linearly must be the zero distribution 0S ′ independent if and only if any zero Slinear combination of the family v has necessarily a zero coefficient system. We shall denote the superposition of a family of distribution v with respect to a distributional system of coefficients a, also by the multiplication symbol a.v. 1 Example (The Dirac family). The Dirac family in S′ n is Slinearly inde- pendent. In fact, we have ZRn uδ = u, for all u ∈ S′ n, and then the relation u.δ = 0S ′ n is equivalent to u = 0S ′ n . Example (the Fourier families). Let a, b be two real non-zero num- bers, recall that the (a, b)-Fourier family in the space of tempered distributions S′(Rn, C) is the family of smooth and bounded regular tempered distributions (recall that [f ] denote the distribution canonically associated to a locally summable function f ) . In the particular case a = 1 and b = −1/ (with the reduced Planck constant) we obtain what we call the De Broglie family, i.e. the family (cid:16)a−nhe−ib(p·)i(cid:17)p∈Rn , uϕ = 0S ′ n(C). . (cid:16)he(i/)(p·)i(cid:17)p∈Rn ZRn 0 = (cid:18)ZRn uϕ(cid:19) (φ) = = u (bϕ (φ)) = = u(cid:0)S(a,b) (φ)(cid:1) = = F(a,b) (u) (φ) , All the Fourier families are Slinearly independent. In fact, let ϕ be the (a, b)- Fourier family, and let For every test function φ in Sn(C), we have where S(a,b) is the Fourier-Schwartz transformation on the space Sn(C), so that the Fourier transform of the distribution u must be zero, i.e. F(a,b) (u) = 0S ′ n(C), and hence u = 0S ′ erator. n(C), being the Fourier transformation F(a,b) an injective op- 2 Linear and Slinear independence A first elementary connection between the classic linear independence and the new Slinear independence is given by the following theorem, which states that 2 the Slinear independence is a stronger requirement for an Sfamily than the requirement of the simple linear independence. Recall that an infinite family v of vectors of a vector space is said to be linearly independent if any finite subfamily of v is linearly independent. Theorem. Let v ∈ S(Rm, S′ n) be an Slinearly independent family. Then, the family v is linearly independent. Consequently, the Slinear hull Sspan (v) is an infinite dimensional subspace of S′ n. Proof. By contradiction, assume the family v linearly dependent. Then there exists a linearly dependent finite subfamily of v. More precisely, there are a positive integer k ∈ N and a k-sequence α ∈ (Rm)k of points belonging to the m-dimensional Euclidean space Rm such that the k-sequence of distributions vα = (vαi )k i=1, extracted from the family v by means of the index selection α is a linearly dependent system of S′ n. Then there exists a non-zero scalar k-tuple a ∈ Kk such that the finite linear combinationP avα is zero, that is such that aivαi = 0S ′ n . kXi=1 = = kXi=1 kXi=1 = 0S ′ n. aivαi = i=1 aiδαi as a coefficient distribution, we have Consider the tempered distribution d =Pk dv = ZRm kXi=1 ai(cid:18)ZRm ZRm aiδαi! v = δαi v(cid:19) = Now, since the distribution d is different from the zero distribution 0S ′ m, the preceding equality contradicts the Slinear independence of the family v, against our assumptions. (cid:4) 3 Topology and Slinear independence The last theorem of the above section shows that, for what concerns the Sfamilies, the Slinear independence implies the usual linear independence. Actually, the 3 Slinear independence is more restrictive than the linear independence, as we shall see later by a simple notable example of family which is linearly indepen- dent but not Slinearly independent. On the contrary the Slinear independence is less restrictive than the β(S′ n)-topological independence, as it is shown below. Topological independence. We recall that a system of vectors v = (vi)i∈I n)-topologically n)-topologically free) if and only if there exists a family n)-continuous) linear forms n, indexed by a non-void index set I, is said β(S′ n)-continuous (respectively, σ(S′ in the space S′ free (respectively, σ(S′ L = (Li)i∈I of β(S′ on S′ n such that Li(vk) = δik, for any pair (i, k) ∈ I 2, where the family δ = (δik)(i,k)∈I 2 is the Kronecker family on the square I 2. Note that the above relation can be written as L ⊗ v = δ, where δ is the Kronecker family. If the family v is not topologically free it is said topologically bound. If the family v is topologically free, any family L of continuous linear forms, satisfying the above relations is said a dual family of the family v. So, to say that a family v is topologically free is equivalent to say that v has a dual family of linear continuous forms. Recalling that (by reflexivity) any continuous linear functional on the space S′ n is canonically and univocally representable by a test function in Sn, to say that the family v is topologically free is equivalent to say that the bi- orthonormality condition hgi, vki = δik, is true, for any pair (i, k) ∈ I 2, where g is a suitable family of test functions. Theorem. Every Sfamily in the space S′ n)-topologically bound and, n)-topologically bound. Consequently, no Sfamily has a dual family of n is β(S′ thus, σ(S′ test functions. Proof. Let v be any Sfamily in the space S′ n indexed by Rm. And let L be an arbitrary family in the dual S′′ n indexed by the same index set. Being the Schwartz space (Sn) reflexive, for every i, there is a test function gi in Sn canonically generating the functional Li, that is such that Li = h., gii . In other terms, the test function gi is such that Li(u) = u(gi), for every tempered distribution u in S′ n. Assume the existence of an index i such that Li(vi) = 1, then we deduce 1 = Li(vi) = vi(gi) = v(gi) (i) , being v an Sfamily, the function v(gi) is continuous, then there is a neighborhood U of the point i in which the function v(gi) is strictly positive. Then, for every point k in the neighborhood U , we have Li(vk) = vk(gi) = v(gi) (k) > 0, 4 and then L cannot verify the condition Li(vk) = δik, for any pair (i, k) ∈ I 2. So we cannot find a dual family of functionals for v and, consequently, v cannot be topologically independent. (cid:4) Note. By the same proof, it is possible to prove that every C0-family of tempered distributions is strongly topologically bound. Consequently every smooth family of tempered distributions is strongly topologically bound. 4 Multiplicity of representations It's simple to prove the following property that characterizes the Slinear depen- dence of a family of distributions by explicit multeplicity of representations of some member of the family itself. Property. An Sfamily v in S′ n indexed by Rm is Slinearly dependent if and only if there is a point index p in Rm and a tempered distribution a in S′ m different from the Dirac delta distribution δp such that the p-th member of the family is representable also by vp =ZRm av. Proof. Necessity. Indeed, if vp fulfills that property we have that the Slinear combination (a − δp).v is zero with a non-zero coefficient distribution, so that the family v is Slinearly dependent. Sufficiency. Vice versa, let, for every index-point p, the term vp of the family be representable in a unique way as the superposition vp = δp.v. Assume v Slinearly dependent, then there is a coefficient system a different from zero such that a.v = 0, hence vp = ZRm = ZRm = ZRm δpv − 0 = δpv −ZRm (δp − a)v, av = since a is a non-zero distribution, the distribution δp − a is different form the Dirac delta δp, and so the member vp is representable in another (different) way, against the assumption. (cid:4) 5 5 Characterizations of Slinear independence By the Dieudonn´e-Schwartz theorem we immediately deduce two characteriza- tions. We say an Sfamily v to be topologically exhaustive (with respect to the weak* n)-closed (or topology or the strong topology) if its Slinear hull Sspan (v) is σ(S′ strongly closed, which is the same). Theorem. Let v ∈ S(Rm, S′ an Sfamily whose Slinear hull Sspan (v) is σ(S′ assertions are equivalent n) be a topologically exhaustive family, that is n)-closed. Then the following • 1) the family v is Slinearly independent; m) and σ(S′ n); morphism for the weak* topologies σ(S′ morphism for the strong topologies β(S′ • 2) the superposition operator RRm (·, v) is an injective topological homo- • 3) the superposition operator RRm (·, v) is an injective topological homo- • 4) the operator bv is a surjective topological homomorphism for the weak • 5) the operator bv is a surjective topological homomorphism of the topolog- ical vector space (Sn) onto the space (Sm). topologies σ(Sn) and σ(Sm); m) and β(S′ n); Theorem. Let v ∈ S(Rm, S′ n) be an Sfamily. Then the following assertions are equivalent • 1) the family v is Slinearly independent and the hull S span (v) is σ(S′ n)- closed; morphism for the weak* topologies σ(S′ • 2) the superposition operator RRm (·, v) is an injective topological homo- • 3) the operator bv is a surjective topological homomorphism for the weak • 4) the operator bv is a surjective topological homomorphism from the space topologies σ(Sn) and σ(Sm); (Sn) onto the space (Sm). m) and σ(S′ n); Remark (on the coordinate operator). In the conditions of the above theorem, if the family v is S linearly independent, we can consider the algebraic isomorphism from the space S′ m onto the Slinear hull Sspan (v) sending every 6 ZRm (·, v) : S′ m → S′ n : (a, v) 7→ a.v tempered distribution a ∈ S′ of the injection m into the superposition a.v, that is the restriction to the pair of sets (S′ phism by the symbol [·v]. theorem that m,S span (v)). We shall denote the inverse of this isomor- It is an important consequence of the preceding • Theorem. The operator [·v] :S span (v) → S′ m is a topological isomorphism, with respect to the topology induced by the n) on the Slinear hull Sspan (v) and to the weak* topol- weak* topology σ(S′ ogy σ(S′ n)-closed, that is if the family v is topologically exhaustive. m), if and only if the Slinear hull Sspan (v) is σ(S′ 6 S Bases Definition (of Sbasis). Let v ∈ S(Rm, S′ a subspace of the space S′ if it is Slinearly independent and it Sgenerates V , that is n and let V be n. The family v is said an S basis of the subspace V n) be an Sfamily in S′ Sspan(v) = V. In other terms, the family v is said an S basis of the subspace V if and only if the superposition operator of the family v is injective and its range coincides with the subspace V . The Dirac family δ of S′ n is an Sbasis of the whole S′ n (indeed its superpo- sition operator is the identity on S′ n). We call the Dirac family δ the canonical Sbasis of S′ n (because of reasons which will appear clear soon, and that can be summarized into the relation u.δ = u, valid for any tempered distribution u) or the Dirac basis of S′ n. Moreover, the following complete version of the Fourier expansion-theorem, allow us to call the Fourier families of S′(Rn, C) by the name of Fourier bases of S′(Rn, C). Theorem (geometric form of the Fourier expansion theorem). In the n(C) the Fourier families are Sbases space of complex tempered distributions S′ (of the entire space S′ n(C)). Proof. Indeed, the superposition operators of the Fourier families in S′ n are the Fourier transforms upon S′ n which are bijective. (cid:4) 7 7 Algebraic characterizations of S bases The following is an elementary but meaningful generalization of the Fourier expansion theorem. Theorem (characterization of an Sbasis). Let v ∈ S(Rm, S′ n) be an Sfamily. Then, • 1) the family v Sgenerates the space S′ n if and only if the superposition • 2) the family v is S linearly independent if and only if the superposition operator t (bv) is surjective; operator t (bv) is injective; operator t (bv) is bijective. • 3) the family v is an Sbasis of the space S′ n if and only if the superposition Proof. The proof follows immediately from the definitions, however we see it. Moreover, it is obvious, by the very definitions, that the family v Sgenerates the space S′ First of all the superposition operator t (bv) is well defined because v is an Sfamily. n if and only if the superposition operator t (bv) is surjective, and that v is Slinearly independent if and only if the superposition operator t (bv) is injective. (cid:4) 7.1 Example The following point gives us an example of linearly independent family which is not Slinearly independent and also an example of a linearly independent system of S generators which is not an Sbasis. Example (a system of linearly independent Sgenerators that is not 1 of the first derivatives of the an Sbasis). Let v = (δ′ Dirac distributions. The family v is of class S, in fact x)x∈R be the family in S′ v(φ)(x) = vx(φ) = δ′ x(φ) = −φ′(x), and the the derivative φ′ is an S function. Consequently, the operator associated with the family v is the derivation in the test function space S1 up to the sign 1. This last superposition operator is a surjective operator (every tempered distribution has a primitive) but it is not injective (every tempered distribution has many primitives), then the family v is a system of Sgenerators for the space S′ 1, but it is not Slinearly independent. Moreover, note that, however, the family v is and, then, the superposition operator t (bv) is the derivation in the space S′ 8 linearly independent. In fact, let P be a finite subset of the real line R, and let, for every point p0 in P , fp0 be a test function in S1 such that f ′ p0 (p) = δp0,p, for every index p in P , here δ(.,.) is the Kronecker delta upon the square P 2. Now, if a = (ap)p∈P is a finite family of scalars such that then apδp0p = ap0, apvp = 0S ′ 1, Xp∈P apvp (fp0 ) =Xp∈P 0 =Xp∈P for every index p0 in P . And so, any finite linear combination of members of v is zero only with respect to a zero-system of coefficients. 8 Totality of Sbases We give another way to characterize an S basis. Theorem (characterization of an Sbasis). Let v ∈ S(Rm, S′ n) be an Sfamily. Then: 1) the family v is a system of S generators of the entire space S′ n if and only if the family v is total in the space Sn (in the sense that if vp(g) = 0, for every index-point p, then g = 0) and the linear hull of the family v is weakly* closed; 2) the family v is Slinearly independent if and only it is total in the distri- bution space S′ m, in the sense that if a.v = 0 then a = 0; 3) a family v is an Sbasis of the space S′ n if and only if the family v is total both in the function space Sn and distribution space S′ m. Proof. 1) Indeed, the condition means that the operator generated by the family v is injective and this (by the Dieudonn`e-Schwartz theorem) implies the surjectivity of the superposition operator of v, since the image of the su- perposition operator of v is closed. 2) This is exactly the definition of linear independence. 3) Follows from the two before. (cid:4) 9 9 Topological characterizations of S bases By the Dieudonn´e-Schwartz theorem we immediately take a characterization. Theorem. Let v ∈ S(Rm, S′ n) be a family of tempered distributions. Then the following assertions are equivalent • 1) the family v is an S basis of the space S′ n; m) and σ(S′ n); m) and β(S′ n); strong topologies β(S′ weak* topologies σ(S′ • 2) the superposition operator RRm(·, v) is a topological isomorphism for the • 3) the superposition operator RRm(·, v) is a topological isomorphism for the • 4) the operator bv is a topological isomorphism for the weak topologies • 5) the operator bv is a topological isomorphism of the topological vector space (Sn) onto (Sm). σ(Sn) and σ(Sm); References [1] J. Barros-Neto, An Introduction to the theory of distributions, Marcel Dekker, Inc. NewYork, 1981 [2] N. Boccara, Functional analysis, an introduction for physicists, Academic press, Inc. 1990 [3] J. Horvath, Topological Vector Spaces and Distributions (Vol.I), Addison- Wesley Publishing Company, 1966 [4] L. Schwartz, Th´eorie des distributions, Hermann, Paris 1966 [5] J. Dieudonn´e, "La dualit´e dans les espaces vectoriels topologiques", Annales scietifiques de l'E.N.S. 3 es´erie, tome 59 p. 107 - 139, 1942. http://www.numdam.org [6] J. Dieudonn´e L. Schwartz, "La dualit´e dans les espaces (F ) and (LF )", An- nales de l'institut Fourier, tome 1 p. 61 - 101, 1949. http://www.numdam.org [7] P. A. M. Dirac, The principles of Quantum Mechanics, Oxford Claredon press, 1930. [8] D. Carf`ı, "S-linear operators in quantum mechanics and in eco- nomics", (APPS), volume 6 pp.7-20, 2004. (no.1 electronic edition, ) http://vectron.mathem.pub.ro/apps/v6/a6.htm 10 [9] D. Carf`ı, "Dirac-orthogonality in the space of tempered distributions", Jour- nal of computational and applied mathematics, vol. 153, pp.99-107, numbers 1-2, 1 april 2003. [10] D. Carf`ı, "S-diagonalizable operators in quantum mechanics", Glasnik Mathematicki, vol. 40 n.2, pp. 267-307, 2005. [11] D. Carf`ı, "On the Schodinger's equation associated with an operator...", Rendiconti del Seminario Matematico di Messina, n. 8 serie II, 2001. [12] D. Carf`ı, "Quantum statistical systems with a continuous range of states", series of Advances in Mathematics for Applied Sciences, vol. 69, pp. 189 - 200, edited by World Scientific. 2005. [13] D. Carf`ı, "Dyson formulas for Financial and Physical evolutions in S′-n", Proceedings of the "VIII Congresso SIMAI", Ragusa, Baia Samuele, 22 - 26 Maj 2006. [14] D. Carf`ı, "Prigogine approach to irreversibility for Financial and Phys- ical applications", Supplemento Atti dell'Accademia Peloritana dei Peri- colanti di Messina, Proceedings Thermocon'05, 2006. ISSN: 0365-0359. http://antonello.unime.it/atti/ [15] D. Carf`ı, "S-convexity in the space of Schwartz distributions and applica- tions", Supplemento ai Rendiconti del Circolo Matematico di Palermo, Serie II - Numero 77 - pp. 107-122, Anno 2006,. ISSN: 0009-725X. [16] D. Carf`ı, "S-Linear Algebra in Economics and Physics", Applied Sciences (APPS), vol. 9, 2007, ISSN 1454-5101, forthcoming volume. David Carf`ı Faculty of Economics University of Messina [email protected] 11
1612.03775	2	1612	2017-07-10T17:02:10	Angular equivalence of normed spaces	[ "math.FA" ]	Angular equivalence is introduced and shown to be an equivalence relation among the norms on a fixed real vector space. It is a finer notion than the usual (topological) notion of norm equivalence. Angularly equivalent norms share certain geometric properties: A norm that is angularly equivalent to a uniformly convex norm is itself uniformly convex. The same is true for strict convexity. Extreme points of the unit balls of angularly equivalent norms occur on the same rays, and if one unit ball is a polyhedron so is the other. Among norms arising from inner products, two norms are angularly equivalent if and only if they are topological equivalent. But, unlike topological equivalence, angular equivalence is able to distinguish between different norms on a finite-dimensional space. In particular, no two $\ell^p$ norms on $\mathbb R^n$ are angularly equivalent.	math.FA	math	ANGULAR EQUIVALENCE OF NORMED SPACES EDER KIKIANTY AND GORD SINNAMON Abstract. Angular equivalence is introduced and shown to be an equivalence relation among the norms on a fixed real vector space. It is a finer notion than the usual (topological) notion of norm equivalence. Angularly equivalent norms share certain geometric properties: A norm that is angularly equivalent to a uniformly convex norm is itself uniformly convex. The same is true for strict convexity. Extreme points of the unit balls of angularly equivalent norms occur on the same rays, and if one unit ball is a polyhedron so is the other. Among norms arising from inner products, two norms are angularly equiv- alent if and only if they are topological equivalent. But, unlike topological equivalence, angular equivalence is able to distinguish between different norms on a finite-dimensional space. In particular, no two ℓp norms on Rn are angu- larly equivalent. 1. Introduction and Definition Two norms on a real vector space are equivalent if both give rise to the same notion of convergence. A wide variety of functional analysis results concern only the topology a norm generates, not the specific values taken by a given norm. In such a setting, choosing the most convenient one among several, or many, topologically equivalent norms can clarify arguments and simplify proofs. In other situations, specific properties of individual norms are central to the theory. For example, uniform convexity is an important property of a norm that may not be shared by a topologically equivalent norm. It is our object here to introduce a finer equivalence of norms, one that preserves certain properties of the norm that simple topological equivalence does not. The idea is straightforward; two norms are angularly equivalent if, over all pairs of non- zero vectors, the angle between the pair, determined by one norm, is comparable to the angle between the same pair, determined by the other norm. This will be made precise shortly. Although the theory of angles in normed spaces cannot match the elegance of its counterpart for inner product spaces, it serves here to give us a means of defining an equivalence of norms that compares vectors two by two rather than one at at time, as topological equivalence does. Thus, angular equivalence emerges as a kind of "second order" equivalence compared to the "first order" topological equivalence. Definition 1. Two norms, k·k1 and k·k2, on a real vector space X are topologically equivalent provided there exist positive constants m, M such that for all x, y ∈ X, (1.1) mkxk1 ≤ kxk2 ≤ Mkxk1. 2010 Mathematics Subject Classification. 46B (primary), 46B20 (secondary). Supported by the Natural Sciences and Engineering Research Council of Canada. 1 2 EDER KIKIANTY AND GORD SINNAMON Very little is needed to give the definition of angular equivalence besides an accessible concept of angle in normed spaces. We define an angle based on the g-functional, introduced in [5] and studied in [2, Chapter 4] and references there. It it closely connected to smoothness and convexity properties of the unit ball and lends itself to straightforward calculations. 1 t→0+ t (kx+tyk−kxk) and g−(x, y) = kxk lim Fix vectors x and y in a real vector space X, with norm k·k. A few applications of the triangle inequality are enough to show that 1 t (kx + tyk − kxk) is a non- decreasing function of t taking (−∞, 0) ∪ (0,∞) into [−kyk,kyk]. It follows that both g+(x, y) = kxk lim exist, and satisfy (1.2) −kxkkyk ≤ kxk(kxk−kx−yk) ≤ g−(x, y) ≤ g+(x, y) ≤ kxk(kx+yk−kxk) ≤ kxkkyk. Definition 2. Suppose k · k is a norm on a real vector space X. The g-functional relative to k · k is the map g : X × X → [0,∞) given by, g = 1 2 (g− + g+). If x and y are non-zero vectors in X, the norm angle1 from x to y is θ(x, y), defined by 0 ≤ θ ≤ π and t (kx+tyk−kxk) t→0− 1 cos θ(x, y) = g(x, y) kxkkyk . We will refrain from referring to the norm angle "between" x and y, since the norm angle from x to y may not coincide with the norm angle from y to x. If the norm in X arises from an inner product, it is easy to see that norm angles agree with angles defined by the inner product. To see that θ(x, y) does not depend on the lengths of x and y, make the substitution s = bt/a in the defining limits to get g(ax, by) = abg(x, y) whenever a, b > 0. A little extra care shows that this equation holds for any a, b ∈ R so, in particular, g(x,−y) = −g(x, y). Definition 3. Two norms, k · k1 and k · k2, on a real vector space X are angularly equivalent provided there exists a constant C such that for all non-zero x, y ∈ X, (1.3) Here θ1(x, y) and θ2(x, y) are the norm angles from x to y relative to k · k1 and k · k2, respectively. Also, tan(π/2) is taken to be +∞. tan(θ2(x, y)/2) ≤ C tan(θ1(x, y)/2). It is clear from the definition that angular equivalence is a reflexive and transitive relation. To see that it is also symmetric, replace y by −y in (1.3). Then, for j = 1, 2, gj(x,−y) = −gj(x, y), cos θj(x,−y) = − cos θj(x, y) and 1 + cos θj(x, y)(cid:19)1/2 tan(θj(x, y)/2) =(cid:18) 1 − cos θj(x, y) 1 − cos θj(x,−y)(cid:19)1/2 =(cid:18) 1 + cos θj(x,−y) So if (1.3) holds for all non-zero x, y then so does (1.4) = 1 tan(θj(x,−y)/2) . tan(θ1(x, y)/2) = 1 tan(θ1(x,−y)/2) ≤ C tan(θ2(x,−y)/2) = C tan(θ2(x, y)/2). 1The term "g-angle" is in use, coined by Pavle Milici´c for an angle based on a symmetrized g-functional. ANGULAR EQUIVALENCE OF NORMED SPACES 3 Thus angular equivalence is an equivalence relation. Since (1.3) implies (1.4) with the same constant C, angular equivalence has a strong form of symmetry. It follows that the constant C is necessarily at least 1. Remark 1. We will see later that for "most" x and y, g− and g+ coincide. When they differ, the choice of g as the average of the two is unimportant for our pur- poses; the notion of angular equivalence does not depend on the value of g at these exceptional pairs. Moreover, at pairs for which g− and g+ coincide, any semi- inner product necessarily agrees with their common value. So angular equivalence is not dependent on the specific functional used to define it. See [2] for a thorough discussion of semi-inner products. This remark is supported by Theorem 3.3 and Corollaries 3.4 and 3.5 below. The use of the trigonometric ratio tan(θ/2) in the definition of angular equiva- lence may seem arbitrary, since there are many possible ways to express the con- dition that two angles be comparable. Our choice is motivated by the situation for inner product spaces, where the ratio tan(θ/2) appears in the following sharp inequality relating the best constants for topological and angular equivalence. Proposition 1.1 ([4]). Let X be a real vector space equipped with inner products h·,·i1 and h·,·i2, giving rise to norms k ·k1 and k ·k2 in the usual way. Let m, M ∈ [0,∞] be the best constants in the inequality mkxk1 ≤ kxk2 ≤ Mkxk1, x ∈ X. Then, for any non-zero x, y ∈ X, tan(θ2(x, y)/2) ≤ (M/m) tan(θ1(x, y)/2). The constant M/m is best possible in this inequality. Here θ1 and θ2 are the angles between x and y relative to k·k1 and k·k2, respectively. (The usual angles coincide with the norm angles in inner product spaces.) This result is a reformulation of the generalized Wielandt inequality. The origi- nal, based on [8] and appearing in [1], involves the action of an invertible matrix on Cn, relating the angle between a pair of vectors and the angle between the trans- formed vectors. Since M/m < ∞ if and only if the two norms are topologically equivalent, the reformulated inequality shows that if two norms arise from inner products on the same vector space, then the norms are angularly equivalent if and only if they are topologically equivalent. This is an indication that angular equivalence is not too much finer than topo- logical equivalence; the two notions coincide on an important class of spaces. As an indication that angular equivalence is nonetheless considerably finer than topo- logical equivalence see Corollaries 2.3 and 2.4 below, where we show that, unlike topological equivalence, norms on a finite-dimensional spaces are not all angularly equivalent. (Of course, we have not yet established that angular equivalence is in fact finer than topological equivalence. But we will; see Theorem 4.2.) 2. A First Look at Angular Equivalence Many properties shared by angularly equivalent norms may be deduced from the definition with minimal calculation, while others emerge only after careful analysis. In this section we present a selection of the former. 4 EDER KIKIANTY AND GORD SINNAMON Recall that if E is a subset of a real vector space X we say x ∈ E is an extreme point of E provided x is not in any open line segment contained in E. When X is a normed space and E is the closed unit ball of X this is of particular interest in functional analysis. Our first result shows that the unit balls of angularly equivalent norms share the same extreme "directions". Theorem 2.1. Let X be a real vector space having two norms, k · k1 and k · k2. Suppose k · k1 and k · k2 are angularly equivalent and x is a non-zero vector in X. Then x/kxk1 is an extreme point of the k · k1-unit ball if and only if x/kxk2 is an extreme point of the k · k2-unit ball. Proof. We argue the contrapositive. Suppose x/kxk2 is not an extreme point of the k · k2-unit ball. Then there are points y and z in X such that (y + z)/2 = x/kxk2 If and the closed line segment from y to z is contained in the k · k2-unit ball. s ∈ [0, 1] then (1 − s)y + sz and sy + (1 − s)z are on the line segment and hence in the k · k2-unit ball. Thus, 2 = ky+zk2 = k(1−s)y+sz+sy+(1−s)zk2 ≤ k(1−s)y+szk2+ksy+(1−s)zk2 ≤ 2. It follows that k(1 − s)y + szk2 = 1. In particular, observe that kyk2 = kzk2 = 1. Now, g± 2 (y, z) = lim t→0± = lim s→0± = lim s→0± 1 1−s t (ky + tzk2 − 1) s (ky + s s (k(1 − s)y + szk2 − 1 + s) = 1. 1−s zk2 − 1) 1 This shows that g2(y, z) = 1, cos(θ2(y, z)) = 1, and tan(θ2(y, z)/2) = 0. By angular equivalence, tan(θ1(y, z)/2) = 0 as well. This implies cos(θ1(y, z)) = 1 and hence g1(y, z) = kyk1kzk1. The last statement, which may be written as g− 1 (y, z) + g+ 1 (y, z) = 2kyk1kzk1, combined with g− 1 (y, z) ≤ g+ 1 (y, z) ≤ kyk1(ky + zk1 − kyk1) ≤ kyk1kzk1, from (1.2), gives Since ky + zk1 = kyk1 + kzk1 and x/kxk2 = (y + z)/2, we have kyk1(ky + zk1 − kyk1) = kyk1kzk1. x kxk1 = y + z ky + zk1 = kyk1 kyk1 + kzk1 y kyk1 + kzk1 kyk1 + kzk1 , z kzk1 which is a convex combination of the points y/kyk1 and z/kzk1. Thus, x/kxk1 is an interior point of the line segment from y/kyk1 to z/kzk1. Since the endpoints of this segment lie in the k · k1-unit ball, convexity shows that the entire line segment does. Thus, x/kxk1 is not an extreme point of the k · k1-unit ball. Reversing the roles of the two norms gives the other implication. (cid:3) A normed space is strictly convex if every boundary point of the unit ball is an extreme point. It is immediate that strict convexity is preserved by angular equiv- alence. The corresponding result for uniform convexity also holds, see Corollary 2.7. ANGULAR EQUIVALENCE OF NORMED SPACES 5 Corollary 2.2. Suppose X be a real vector space having two angularly equivalent norms, k · k1 and k · k2. Then X is strictly convex when equipped with k · k1 if and only if X is strictly convex when equipped with k · k2. The unit ball of finite-dimensional normed space is polygonal if and only if it has only finitely many extreme points, namely, the vertices of the polygon. Corollary 2.3. Let X be a finite-dimensional real vector space having two angularly equivalent norms, k · k1 and k · k2. Then the k · k1-unit ball is polygonal if and only if the k · k2-unit ball is polygonal. In this case, the vertices of the two polygons lie on the same rays. Remark 2. If two norms are angularly equivalent on X then the restrictions of those two norms to any subspace Y of X are also angularly equivalent. This follows easily from the definition. However, the subspace Y may have extreme points that the original space did not have. For every subspace Y , the rays containing extreme points relative to Y are the same for both norms. One of the most telling features of angular equivalence is its ability to distinguish between norms on a finite-dimensional space. Here we see that no two ℓp norms are angularly equivalent. (This only applies in finite dimensions. The angular equivalence of the usual ℓp sequence space norms is a question that does not arise naturally, since they are not norms on the same underlying vector space.) Corollary 2.4. Suppose p, q ∈ [1,∞]. For any integer n ≥ 2, the ℓp and ℓq norms on Rn are angularly equivalent only if p = q. Proof. If the two norms are angularly equivalent then their restrictions to the sub- space R2 of Rn (embedded in the usual way) are also angularly equivalent. So we may assume n = 2. The unit ball in the ℓ1 norm is a square with vertices at (0,±1) and (±1, 0). The unit ball in the ℓ∞ norm is a square with vertices at (±1,±1). When 1 < p < ∞, the unit ball in the ℓp norm is not a polygon. By Corollary 2.3, the ℓ1 norm and the ℓ∞ norm are not angularly equivalent to each other or to any other ℓp norm. It remains to consider p, q ∈ (1,∞). Fix s > 0 and consider the vectors (1, 0) and (1, s). The value of the ℓp-norm g-functional for this pair is, 1 t→0 t (k(1, 0)+t(1, s)kℓp−k(1, 0)kℓp) = lim k(1, 0)kℓp lim Thus θℓp ≡ θℓp ((1, 0), (1, s)) satisfies, t→0 1 t (((1+t)p +tpsp)1/p−1) = 1. cos θℓp = 1 (1 + sp)1/p and tan2 θℓp = (1 + sp)1/p − 1 (1 + sp)1/p + 1 . By angular equivalence of the ℓp and ℓq norms, there exists a C > 0, independent of s such that 1 C ≤(cid:18) (1 + sp)1/p − 1 (1 + sp)1/p + 1(cid:19)(cid:18) (1 + sq)1/q + 1 (1 + sq)1/q − 1(cid:19) ≤ C Taking the limit as s → 0+ it is easy to see that this expression remains bounded above and below only if p = q. (cid:3) The next theorem may be viewed as a stability result for angular equivalence. 6 EDER KIKIANTY AND GORD SINNAMON Theorem 2.5. Suppose k · k1 and k · k2 are angularly equivalent norms on a real vector space X. Then the norm k·k3 = k·k1 +k·k2 is angularly equivalent to k·k1 and k · k2. Proof. Fix non-zero x, y ∈ X. Let gj(x, y) and θj = θj(x, y) be the g-functional and norm angle from x to y with respect to the norm k · kj, for j = 1, 2, 3. The definition of the g-functional shows that g3(x, y) kxk3 = g1(x, y) kxk1 + g2(x, y) kxk2 , kyk3 cos θ3 = kyk1 cos θ1 + kyk2 cos θ2. or equivalently, Therefore, tan2(θ3/2) = 1 − cos θ3 1 + cos θ3 = = kyk1 kyk3 kyk1 kyk3 1 + kyk2 kyk1 1 + kyk2 kyk1 (1 − cos θ1) + kyk2 kyk3 (1 + cos θ1) + kyk2 kyk3 (1 − cos θ2) (1 + cos θ2) 1−cos θ2 1−cos θ1 1+cos θ2 1+cos θ1 tan2(θ1/2). But for any A, B > 0, 1+A 1+B ≤ 1 + A B , so tan2(θ2/2) tan2(θ3/2) ≤(cid:18)1 + tan2(θ1/2)(cid:19) tan2(θ1/2) ≤ (1 + C2) tan2(θ1/2), where C is the constant of angular equivalence of k·k1 and k·k2 from (1.3). Taking square roots shows that k · k3 is angularly equivalent to k · k1, and hence to k · k2 as well. (cid:3) With stability under sums of norms we can easily give an example to show that an angular equivalency class containing inner product norms may contain other norms as well. Example 1. A norm that is angularly equivalent to a norm arising from an inner product need not arise from an inner product: Consider the two norms on R2 defined by k(ξ, η)k1 = (3ξ2 + η2)1/2 and k(ξ, η)k2 = (ξ2 + 3η2)1/2. These both arise from inner products and are topologically equivalent so by Theorem 1.1 they are angularly equivalent. Theorem 2.5 shows that their sum, k(ξ, η)k3 = k(ξ, η)k1 + k(ξ, η)k2 is also angularly equivalent. However, it is easy to check that k·k3 does not arise from an inner product. (For example, the parallelogram law fails for the vectors (1, 0) and (0, 1).) The stability of angular equivalence does not extend to the maximum of two norms. Example 2. The maximum of two angularly equivalent norms is a norm that need not be angularly equivalent to the original two. Let k · k1 and k · k2 be the norms in the previous example, set k(ξ, η)k4 = max(k(ξ, η)k1,k(ξ, η)k2), and let gj(x, y) and θj = θj(x, y) be the g-functional and norm angle from x to y with respect to the norm k · kj, for j = 1, 4. Now let s > 0 and take x = (1, 1) and y = (1 − s, 1 + s). Calculations show that kx + tyk1 = 2p(1 + t)2 − (1 + t)ts + t2s2 and g1(x, y) = 4 − 2s; ANGULAR EQUIVALENCE OF NORMED SPACES 7 and also that kx+tyk4 = 2p(1 + t)2 + (1 + t)ts + t2s2, But kxk1 = kxk4 = 2, kyk1 = 2√1 − s + s2 and kyk4 = 2√1 + s + s2. So, using g± 4 (x, y) = 4±2s, and g4(x, y) = 4. the definition of cos θ1(x, y) and cos θ4(x, y), we get, 3s2 tan2(θ1(x, y)/2) = 4(1 − s/2 + √1 − s + s2)2 and tan2(θ4(x, y)/2) = s + s2 (1 + √1 + s + s2)2 . Clearly, there is no constant C for which tan(θ4(x, y)/2) ≤ C tan(θ1(x, y)/2) as s → 0+. Thus k · k4 is not angularly equivalent to k · k1. We close this section with the promised proof that uniform convexity is preserved by angular equivalence. A real vector space X, with norm k·k, is uniformly convex provided that for all ε > 0 there exists a δ > 0 such that if kxk = kyk = 1 and kx − yk ≥ ε then k(x + y)/2k ≤ 1 − δ. First we show that unform convexity can be characterized in terms of the norm angle. The idea appears in [7], where closely related results are proved. Theorem 2.6. Let X be a real vector space with norm k · k and let θ(x, y) denote the norm angle from x to y. Then X is uniformly convex if and only if for all ε > 0 there exists a δ > 0 such that if kxk = kyk = 1 and kx − yk ≥ ε then tan(θ(x, y)/2) ≥ δ. Proof. Suppose X is uniformly convex. Fix ε > 0 and choose an η > 0 so that k(x + y)/2k ≤ 1 − η whenever kxk = kyk = 1 and kx − yk ≥ ε. Set δ = √η. If kxk = kyk = 1 and kx−yk ≥ ε then (1.2) shows that −1 ≤ g(x, y) ≤ kx+yk−1 ≤ 1 so tan(θ(x, y)/2) =s 1 − g(x, y) 1 + g(x, y) ≥r 1 − g(x, y) 2 ≥r1 −(cid:13)(cid:13)(cid:13) x + y 2 (cid:13)(cid:13)(cid:13) ≥ √η = δ. For the converse, fix ε > 0 and choose an η > 0 such that if kxk = kyk = 1 and kx− yk ≥ ε/4 then tan(θ(x, y)/2) ≥ η. Set δ = min(η2/(1 + η2), ε/4). Now suppose kxk = kyk = 1 and kx − yk ≥ ε, and set z = x + y. If z = 0, then the desired conclusion, kz/2k ≤ 1 − δ, is trivial. Otherwise, k(2 − kzk)x − kzk((z/kzk) − x)k = kx − yk ≥ ε, so either k(2 − kzk)xk ≥ 2δ or kkzk((z/kzk) − x)k ≥ ε − 2δ. The first of the two implies that kz/2k ≤ 1 − δ and completes the proof. The choice of δ ensures that ε − 2δ ≥ ε/2 so the second implies k(z/kzk) − xk ≥ ε/(2kzk) ≥ ε/4, and we have η ≤ tan(θ(z/kzk, x)/2) =s 1 − g(z/kzk, x) 1 + g(z/kzk, x) . But this is equivalent to g(z/kzk, x) ≤ (1 − η2)/(1 + η2). Using (1.2) again, this time with x and y replaced by z and x, respectively, gives kzk − 1 ≤ g(z/kzk, x) so we have, kz/2k ≤ 1 2 (1 + g(z/kzk, x)) ≤ 1 2(cid:16)1 + 1 − η2 1 + η2(cid:17) = 1 − η2 1 + η2 ≤ 1 − δ. 8 EDER KIKIANTY AND GORD SINNAMON The next result uses the previous theorem so it is included here despite its dependence on Theorem 4.2 below. (cid:3) Corollary 2.7. Suppose X be a real vector space having two angularly equivalent norms, k · k1 and k · k2. Then X is uniformly convex when equipped with k · k1 if and only if X is uniformly convex when equipped with k · k2. Proof. Let C be the constant in the definition of angular equivalence so that, using symmetry and (1.3), for all non-zero x and y in X we have tan(θ1(x, y)/2) ≤ C tan(θ2(x, y)/2). Theorem 4.2 shows that k·k1 and k·k2 are also topologically equivalent so there Suppose X is uniformly convex when equipped with k ·k1 and fix ε > 0. Choose η > 0 so that, if kxk1 = kyk1 = 1 and kx− yk1 ≥ mε/(2M ), then tan(θ1(x, y)/2) ≥ η. Set δ = η/C. Let kxk2 = kyk2 = 1 and kx − yk2 ≥ ε, and set x = x/kxk1 and y = y/kyk1. Clearly, kxk1 = kyk1 = 1. Also, (as in the proof of the Dunkl-Williams inequality,) exist constants m and M such that (1.1) holds. x ε ≤ kx − yk2 =(cid:13)(cid:13)(cid:13)(cid:13) kxk2 − ≤ kx − yk2 kxk2 y + y kxk2 − kxk2 + kxk2 − kyk2 kxk2 y kyk2(cid:13)(cid:13)(cid:13)(cid:13)2 ≤ 2kx − yk2 . kxk2 But kxk2 = kxk2/kxk1 ≥ m so kx − yk1 ≥ (1/M )kx − yk2 ≥ mε/(2M ). This ensures that tan(θ1(x, y)/2) ≥ η. But θ1(x, y) = θ1(x, y) so angular equivalence implies tan(θ2(x, y)/2) ≥ η/C = δ. This shows that X is uniformly convex when equipped with k · k2. For the other implication, reverse the roles of k · k1 and k · k2. (cid:3) 3. Norms in the Plane Since angular equivalence is defined in terms of pairs of vectors, it is natural to study it first in two-dimensional spaces. Our analysis in the plane is more than the investigation of a special case, however. It enables us to establish results for real vector spaces of any dimension: It is evident from the definition that, if a pair of norms on a real vector space X are angularly equivalent then the restrictions of those norms to any subspace of X are also angularly equivalent. We will make use of the equally evident converse: If the restrictions of two norms on X to any two-dimensional subspace of X are angularly equivalent with constant C, then the original norms are also angularly equivalent with the same constant C. Uniform control of the constant is essential here. In the plane we can study norms by viewing the boundary of their unit balls as polar functions. The following setup and notation will be used throughout this section and the next. Here and throughout, the symbol "∠" denotes the usual angle in the plane, not the norm angle. Suppose k·k is a norm on R2 and define r > 0 by requiring kr(t)(cos t, sin t)k = 1 for all t ∈ R. Then r is a strictly positive, π-periodic function and the unit ball, {ρ(cos t, sin t) : 0 ≤ ρ ≤ r(t), t ∈ R}, is a closed, convex set with (0, 0) in its interior. The convexity of the unit ball readily implies that r is continuous. Since r is continuous and periodic it attains its minimum and maximum values; both ANGULAR EQUIVALENCE OF NORMED SPACES 9 are strictly positive. Call them rm and rM , respectively. Let O = (0, 0) and parametrize the boundary of the ball by setting Pt = r(t)(cos t, sin t) for all t ∈ R. If 0 < α − β < π, set ψ(α, β) =(∠OPαPβ, α > β, π − ∠OPαPβ, α < β. Also define ϕ−(α) = limβ→α− ψ(α, β) and ϕ+(α) = limβ→α+ ψ(α, β). These exist by property (ii) below. Lemma 3.1. Suppose α, β ∈ R with 0 < α − β < π. Then: (i) ψ(α, β) + α = ψ(β, α) + β; (ii) ψ(α, t) is non-decreasing for t in (α − π, α) ∪ (α, α + π); (iii) 0 < ψ(α, β) < π and, if 0 < α − β ≤ π/2, then 0 < tan−1(rm/rM ) ≤ ψ(α, β) ≤ π − tan−1(rm/rM ) < π; (iv) r(α)/r(β) = cos(α − β) + sin(α − β) cot ψ(α, β) and ψ is continuous; (v) the left derivative of r at α exists and equals r(α) cot ϕ−(α), the right derivative of r at α exists and equals r(α) cot ϕ+(α), and r is differentiable wherever ϕ− = ϕ+; (vi) ϕ−(α) ≤ ϕ+(α) and if α < β then ϕ+(α) + α ≤ ϕ−(β) + β; (vii) ϕ− is left continuous and ϕ+ is right continuous; (viii) ϕ− and ϕ+ are continuous at all but countably many points, and ϕ− is continuous at α if and only if ϕ+ is continuous at α if and only if ϕ−(α) = ϕ+(α). Proof. The angles of triangle PαOPβ add to π. If α > β they are α − β, ψ(α, β), and π − ψ(β, α) and if α < β they are β − α, ψ(β, α) and π − ψ(α, β). This proves (i). To prove (ii), suppose s < t and both lie in (α − π, α) ∪ (α, α + π). If s < t < α then, by convexity, the segment PsPα intersects the segment OPt. It follows that, ψ(α, s) = ∠0PαPs ≤ ∠0PαPt = ψ(α, t). If s < α < t then the segment PsPt intersects the segment OPα. Thus ψ(α, s) + π − ψ(α, t) = ∠OPαPs + ∠OPαPt = ∠PsPαPt ≤ π. If α < s < t then the segment PαPt intersects the segment OPs. It follows that, π − ψ(α, s) = ∠0PαPs ≥ ∠0PαPt = π − ψ(α, t). In each of the three cases we have ψ(α, s) ≤ ψ(α, t), as required. Since r(α) > 0, r(β) > 0, and 0 < α − β < π, △PαOPβ is non-degenerate. In particular, 0 < ψ(α, β) < π, the first statement of (iii). For the other, suppose 0 < α − β < π/2. Then, using (ii), ψ(α, β) ≥ ψ(α, α − π/2) = tan−1(r(α − π/2)/r(α)) ≥ tan−1(rm/rM ) > 0 and ψ(α, β) ≤ ψ(α, α + π/2) = π − tan−1(r(α + π/2)/r(α)) ≤ π − tan−1(rm/rM ) < π. The sine law in △OPαPβ gives sin ψ(α, β) = sin(∠OPαPβ) = r(β) r(β) sin(∠OPβ Pα) r(α) = sin(π − ψ(α, β) − (α − β)) r(α) 10 EDER KIKIANTY AND GORD SINNAMON when α > β and sin(π − ψ(α, β)) r(β) = sin(∠OPαPβ) r(β) = sin(∠OPβ Pα) r(α) = sin(ψ(α, β) − (β − α)) r(α) when α < β. These both reduce to the equation in (iv). Continuity of r now implies continuity of ψ. To prove (v) we use (iv) to get r(α) − r(s) α − s = r(s)(cid:18) cos(α − s) − 1 α − s + sin(α − s) α − s cot ψ(α, s)(cid:19) . The definitions of ϕ+ and ϕ− (and the continuity of r) show that the left and right derivatives of r exist and are equal to, lim s→α− r(α) − r(s) α − s = r(α) cot ϕ−(α) and lim s→α+ r(α) − r(s) α − s = r(α) cot ϕ+(α), respectively. It is immediate that r is differentiable wherever ϕ+ = ϕ−. By (ii), ϕ−(α) = lim s→α− ψ(α, s) ≤ lim t→α+ ψ(α, t) = ϕ+(α). Also, if α < β then, by (i) and (ii), ϕ+(α) + α = lim t→α+ ψ(α, t) + α ≤ ψ(α, β) + α = ψ(β, α) + β ≤ lim s→β− ψ(β, s) + β = ϕ−(β) + β. These prove (vi). For (vii), observe that if γ > α, applying (vi), (ii), and then (iv) gives ϕ+(α) ≤ lim s→α+ = lim s→α+ ϕ−(s) + s − α ≤ lim lim t→s+ ψ(s, t) ≤ lim s→α+ ϕ+(s) s→α+ ψ(s, γ) = ψ(α, γ). Letting γ → α+ shows ϕ+(α) ≤ lims→α+ ϕ+(s) ≤ ϕ+(α). Thus, ϕ+ is right- continuous at α. A similar argument shows that ϕ− is left-continuous. By (vi), the functions s 7→ ϕ−(s) + s and t 7→ ϕ+(t) + t are both non-decreasing and hence continuous except at countably many points. It follows that ϕ− and ϕ+ are both continuous except at countably many points, the first statement of (viii). The second may be deduced from the following consequence of (vi) and (vii): ϕ−(α) = lim s→α− ϕ−(s) ≤ lim s→α− ϕ+(s) ≤ ϕ−(α) ≤ ϕ+(α) ≤ lim t→α+ ϕ−(t) ≤ lim t→α+ ϕ+(t) = ϕ+(α). (cid:3) The g-functional and norm angle for a norm on R2 can be expressed in terms of the functions r, ψ and φ. ANGULAR EQUIVALENCE OF NORMED SPACES 11 Theorem 3.2. Suppose k · k is a norm on R2, define r, ψ and ϕ± as above, and let Qt = (cos t, sin t) for t ∈ R, If a, b > 0 and 0 < α − β < π, then g(aQα, bQβ) = cos θ(aQα, bQβ) = tan2(θ(aQα, bQβ)/2) = ab r(α)2 (cid:0)cos(α − β) + 1 cot(α − β) + 1 2 (cot ϕ−(α) + cot ϕ+(α)) sin(α − β)(cid:1) , 2 (cot ϕ−(α) + cot ϕ+(α)) , and cot(α − β) + cot ψ(α, β) cot ψ(α, β) − 1 2 (cot ϕ−(α) + cot ϕ+(α)) 2 cot(α − β) + cot ψ(α, β) + 1 2 (cot ϕ−(α) + cot ϕ+(α)) . If a, b > 0 and α = β then g(aQα, bQβ) = ab r(α)2 , cos θ(aQα, bQβ) = 1, and tan2(θ(aQα, bQβ)/2) = 0. ab r(α)2 , If a, b > 0 and α − β = π then cos θ(aQα, bQβ) = −1, and tan2(θ(aQα, bQβ)/2) = ∞. g(aQα, bQβ) = − Proof. Since a and b are non-zero, for t sufficiently close to zero we may define c and γ ∈ (α − π/2, α + π/2) as functions of t, by requiring that cQγ = aQα + tbQβ. Evidently, both c and γ are differentiable. Equating components, we have the equations, c cos γ = a cos α + tb cos β and c sin γ = a sin α + tb sin β. Differentiating these with respect to t gives c′ cos γ − cγ′ sin γ = b cos β and c′ sin γ + cγ′ cos γ = b sin β. This system is readily solved to yield c′ = b cos(β − γ) and cγ′ = b sin(β − γ). The definition of r ensures that kr(t)Qtk = 1 for all t, so r(α)(cid:19) = t (kcQγk − kaQαk) = t r(α) − r(γ)−r(α) t (cid:18) c γ−α r(α)r(γ) r(γ) − c−a a γ−α t a 1 1 . As t → 0 we have t → aγ′(0) = b sin(β−α). γ → α, c → a, Notice that γ is a strictly monotone function of t in a neighbourhood of 0. Lemma 3.1(v) shows that t → c′(0) = b cos(β−α) and a γ−α c−a lim t→0± r(γ) − r(α) γ − α = r(α) cot ϕ±(α) or lim t→0± = r(α) cot ϕ∓(α), r(γ) − r(α) γ − α depending on whether γ is increasing or decreasing. In either case, g(aQα, bQβ) = 1 2kaQαk(cid:0) lim t→0+ a 1 t (kcQγk − kaQαk) + lim t→0− 1 t (kcQγk − kaQαk)(cid:1) b cos(β − α)r(α) − r(α) 1 2 (cot ϕ+(α) + cot ϕ−(α))b sin(β − α) r(α)2 = = r(α) ab r(α)2 (cid:0)cos(α − β) + 1 2 (cot ϕ−(α) + cot ϕ+(α)) sin(α − β)(cid:1) . 12 EDER KIKIANTY AND GORD SINNAMON Using this formula and Lemma 3.1(iv), we get cos θ(aQα, bQβ) = = g(aQα, bQβ) kaQαkkbQβk g(aQα, bQβ) ab/(r(α)r(β)) = But tan2(θ/2) = (1 − cos θ)/(1 + cos θ), so, tan2(θ(aQα, bQβ)/2) = cot ψ(α, β) − 1 cot(α − β) + 1 2 (cot ϕ−(α) + cot ϕ+(α)) . cot(α − β) + cot ψ(α, β) 2 cot(α − β) + cot ψ(α, β) + 1 2 (cot ϕ−(α) + cot ϕ+(α)) 2 (cot ϕ−(α) + cot ϕ+(α)) . The formulas in the case α = β and α− β = π are easily established directly from the definitions of the g-functional and norm angle. We omit the details. (cid:3) If we have two norms, k · k1 and k · k2, on R2 we define rj, ψj and ϕ± j as above based on k · kj, for j = 1, 2. The previous theorem shows that the angular equivalence of these two norms can be expressed in the form, (3.1) cot ψ2(α, β) − A2 2 cot(α − β) + cot ψ2(α, β) + A2 ≤ C2 cot ψ1(α, β) − A1 2 cot(α − β) + cot ψ1(α, β) + A1 for all α, β ∈ R with 0 < α − β < π. Here Aj = 1 j (α)) for j = 1, 2. The cases β = α and β = α±π need not be included, as tan(θ2(aQα, bQβ)/2) ≤ C tan(θ1(aQα, bQβ)/2) holds trivially in those cases. 2 . Next we see However, one can avoid discontinuities of the functions ϕ± j (α) + cot ϕ+ 1 and ϕ± 2 (cot ϕ− 2 (β), 2 (β) = ϕ− 2 (α) = ϕ− 1 (β), ϕ+ 2 (α), ϕ+ 1 (α), ϕ+ that it is enough to consider only values of α for which A1 and A2 simplify. Theorem 3.3. The norms k · k1 and k · k2 on R2 are angularly equivalent with constant C if and only if (3.1) holds for all α, β satisfying 1 (β) = ϕ− 1 (α) = ϕ− (3.2) ϕ+ and 0 < α − β < π. Proof. If the two norms are angularly equivalent, then by Theorem 3.2, (3.1) holds for all α, β such that 0 < α − β < π. Conversely, suppose (3.1) holds for all α, β satisfying (3.2) and 0 < α − β < π. We will show that (3.1) holds for all α, β satisfying 0 < α − β < π. 2 (t)} has a countable complement and is therefore dense in R. By hypothesis, for each α ∈ E, inequality (3.1) holds for all β ∈ E∩((α−π, α)∪(α, α+π)). The continuity of ψ1 and ψ2, from Lemma 3.1(iv), shows that it remains valid for all β ∈ (α−π, α)∪(α, α+π). Fix a β ∈ R. Then (3.1) holds for all α ∈ E ∩ ((β − π, β) ∪ (β, β + π)). Taking one-sided limits of (3.1) with respect to α, but only through points of E, gives, for each α ∈ (β − π, β) ∪ (β, β + π), the two inequalities, (one with "+" and one with "−") (3.3) By Lemma 3.1(viii) the set E = {t ∈ R : ϕ+ 2 (t) = ϕ− 1 (t) = ϕ− 1 (t), ϕ+ 2 (α) cot ψ2(α, β) − cot ϕ± 2 cot(α−β) + cot ψ2(α, β) + cot ϕ± 1 (α) Fix an α ∈ (β − π, β) ∪ (β, β + π). A convexity argument will show that for the fixed β and α, (3.3) implies (3.1). C2(cot ψ1(α, β) − cot ϕ± 1 (α)) 2 cot(α−β) + cot ψ1(α, β) + cot ϕ± 2 (α) ≤ . ANGULAR EQUIVALENCE OF NORMED SPACES 13 (3.4) j (α) + z cot ϕ+ Let Aj(z) = (1 − z) cot ϕ− cot ψ2(α, β) − A2(z) 2 cot(α − β) + cot ψ2(α, β) + A2(z) ≤ C2 j (t) for j = 1, 2. The inequality cot ψ1(α, β) − A1(z) 2 cot(α − β) + cot ψ1(α, β) + A1(z) holds for z = 0, 1 because of (3.3). If we show that it holds for z = 1/2 then we have (3.1). It is important to point out that in each of fractions in (3.4) the numerator and denominator cannot both be zero, lest cot(α − β) + cot ψj(α, β) = 0. If this were true, then either ψj(α, β) = β − α or ψj(α, β) = β − α + π. But by Lemma 3.1(i) this is ψj(β, α) = 0 or ψj(β, α) = π, contrary to Lemma 3.1(iii). Define f : [0, 1] → [0,∞) by f (z) = C2(cot ψ1(α, β) − A1(z))(2 cot(α − β) + cot ψ2(α, β) + A2(z)) − (cot ψ2(α, β) − A2(z))(2 cot(α − β) + cot ψ1(α, β) + A1(z)). First consider the case β < α < β + π. Lemma 3.1(ii) shows that for j = 0, 1 and 0 ≤ z ≤ 1, and we get, 0 < ψj(α, β) ≤ ϕ− j (α) ≤ ϕ+ j (α) ≤ ψj(α, β + π) < π cot ψj (α, β) ≥ cot ϕ− j (α) ≥ Aj(z) ≥ cot ϕ+ j (α) ≥ cot ψj(α, β + π). Since r(β) = r(β + π), Lemma 3.1(iv) implies cos(α− β) + sin(α− β) cot ψj(α, β) = cos(α− β− π) + sin(α− β− π) cot ψj(α, β + π) so cot ψj(α, β + π) = −2 cot(α− β)− cot ψj (α, β). We conclude that the numerators and denominators of both sides of inequality (3.4) are all non-negative. It follows that (3.4) holds if and only if f (z) ≥ 0. j = 0, 1 and 0 ≤ z ≤ 1, The second case is similar. If β − π < α < β then Lemma 3.1(ii) shows that for 0 < ψj(α, β − π) ≤ ϕ− j (α) ≤ ϕ+ j (α) ≤ ψj(α, β) < π and we get, cot ψj (α, β − π) ≥ cot ϕ− j (α) ≥ Aj(z) ≥ cot ϕ+ j (α) ≥ cot ψj(α, β). Since r(β) = r(β − π), Lemma 3.1(iv) implies cos(α− β + π) + sin(α− β + π) cot ψj(α, β− π) = cos(α− β) + sin(α− β) cot ψj(α, β) so cot ψj(α, β − π) = −2 cot(α − β) − cot ψj(α, β). This time the numerators and denominators of both sides of inequality (3.4) are all non-positive but again (3.4) holds if and only if f (z) ≥ 0. It remains to show that f (1/2) ≥ 0. The function f (z) is a quadratic polynomial and the coefficient of z2 is −(C2 − 1)(cot ϕ− 1 (t) − cot ϕ+ 1 (t))(cot ϕ− 2 (t) − cot ϕ+ 2 (t)) ≤ 0 so it is a concave function. Since (3.4) holds for z = 0 and z = 1, f (0) ≥ 0 and f (1) ≥ 0. It follows that f (z) ≥ 0 for all z ∈ [0, 1] and in particular f (1/2) ≥ 0. (cid:3) 14 EDER KIKIANTY AND GORD SINNAMON This theorem may be used to simplify verification of angular equivalence of norms on R2 but it also shows that the notion of angular equivalence does not depend on the value of the two g-functionals at the exceptional pairs for which either g-functional fails to satisfy g−(x, y) = g+(x, y). Corollary 3.4. Let k·k1 and k·k2 be norms on the real vectors space X. To show that the norms are angularly equivalent with constant C it is enough to verify (1.3) for pairs x, y satisfying g+ 2 (x, y) = g− Proof. Suppose that (1.3) holds for pairs x, y satisfying g+ j (x, y) for j = 1, 2. A calculation shows that this case includes all linearly dependent pairs. So fix a pair of independent vectors x and y in X. We identify their two-dimensional span with R2 by the map (ξ, η) 7→ ξx + ηy so that k(ξ, η)kj = kξx + ηykj is a norm on R2 for j = 1, 2. This identification is isometric in both norms so it does not affect g-functional calculations or norm angles. 2 (x, y) j (x, y) = g− 1 (x, y) and g+ 1 (x, y) = g− j and ϕ− j j (aQα, bQβ) = g− Using these norms on R2 we define ϕ+ for j = 1, 2. Suppose the condi- tions (3.2) hold for some α, β satisfying 0 < α− β < π. The proof of Theorem 3.2 shows that g+ j (aQα, bQβ) for j = 1, 2 and for any a, b > 0. But the points aQα, bQβ ∈ R2 correspond to vectors in the span of x and y. Our hy- pothesis, shows that tan(θ2(aQα, bQβ)/2) ≤ C tan(θ1(aQα, bQβ)/2). By Theorem 3.2 inequality (3.1) also holds. Now Theorem 3.3 shows that the two norms on R2 are angularly equivalent with constant C. The same is true for the original two norms on the span of x and y. In particular, (1.3) holds for the vectors x, y. This completes the proof. (cid:3) Let X be a real vector space. A semi-inner product in the sense of Lumer-Giles, is a map [·,·] : X × X → [0,∞) that is linear in the first variable, homogeneous in the second, positive definite, and satisfies [x, y]2 ≤ [x, x][y, y]. It follows that x 7→ [x, x]1/2 is a norm on X. See [2, Chapter 2] and the references there. It is natural to define the angle θ[·,·] from x to y, associated with semi-inner product [·,·], by requiring that 0 ≤ θ ≤ π and cos θ[·,·] = [y, x] [y, y]1/2[x, x]1/2 . Corollary 3.6 in [2] shows that any semi-inner product in the sense of Lumer- Giles satisfies g−(x, y) ≤ [y, x] ≤ g+(x, y). In particular, [y, x] = g(x, y) whenever g−(x, y) = g+(x, y) for all x, y ∈ X. This proves the following. Corollary 3.5. Let X be a real vector space and let [·,·]1 and [·,·]2 be semi-inner products in the sense of Lumer-Giles, giving rise to norms k · k1 and k · k2, respec- tively. If (1.3) holds with θ[·,·]1 and θ[·,·]2 replacing the norm angles θ1 and θ2 then k · k1 and k · k2 are angularly equivalent. It is worth pointing out that if X is a normed space, setting [y, x] = g(x, y) defines a semi-inner product in the sense of Lumer-Giles only under additional conditions on the norm. See [2, Definition 4.2.9 and Proposition 4.2.10]. 4. Angular Equivalence Implies Topological Equivalence The techniques developed in the last section are used to prove that angular equivalence is finer than topological equivalence. The definitions of rj , ψj, and ϕ± j , ANGULAR EQUIVALENCE OF NORMED SPACES 15 relative to the norm k·kj, for j = 1, 2, that were introduced at the beginning of Sec- tion 3 will be used throughout. We begin by showing that the angular equivalence of two norms gives a relationship between their ψ and ϕ-functions. Lemma 4.1. Let k · k1 and k · k2 be angularly equivalent norms on R2. If 0 < α − β < π/2, then (α + ϕ− 2 (β)) ≤ M C2((α + ϕ− 1 (α)) − (β + ϕ+ 2 (α)) − (β + ϕ+ 1 (β))), where M = sup{csc2 ψ1(t, s) : s, t ∈ (β, α), s 6= t}. Proof. Let E be the set of points in (β, α) at which both ϕ+ 2 = ϕ− 2 . By Lemma 3.1(viii), E contains all but countably many points of (β, α) and ϕ+ 1 , ϕ− 1 , ϕ+ 2 are continuous there. Fix s, t ∈ E and let ¯t ∈ (β, α). The cosecant function decreases on (0, π/2) and increases on (π/2, π). So if y lies between ψ1(t, s) and ψ1(t, ¯t) we have 2 , and ϕ− 1 and ϕ+ 1 = ϕ− csc2 y ≤ max(csc2 ψ1(t, ¯t), csc2 ψ1(t, s)) ≤ M. The mean value theorem implies that, cot ψ1(t, s) − cot ψ1(t, ¯t) ≤ Mψ1(t, ¯t) − ψ1(t, s). Letting ¯t → t, we have (4.1) cot ψ1(t, s) − cot ϕ+ 1 (t) ≤ M (ϕ+ 1 (t) − ψ1(t, s). On the other hand, the function csc2 is bounded below by 1 so (4.2) cot ψ2(t, s) − cot ϕ+ 2 (t) ≥ ϕ+ 2 (t) − ψ2(t, s). By Lemma 3.1(iii) the functions ψ1 and ψ2 each take values in a compact subset of (0, π). So do their limits, ϕ+ 2 . But the cotangent function is unbounded near 0. Therefore, for each ε > 0 there exists a δ > 0 such that if 0 < t − s < δ then 1 and ϕ+ 2 cot(t − s) + cot ψ2(t, s) + cot ϕ+ 2 cot(t − s) + cot ψ1(t, s) + cot ϕ+ 2 (t) 1 (t) ≤ 1 + ε. Since t ∈ E, the angular equivalence of k · k1 and k · k2, in the form (3.1), becomes cot ψ2(t, s) − cot ϕ+ 2 (t) 2 (t) ≤ C2 2 cot(t − s) + cot ψ2(α, β) + cot ϕ+ Note that both sides of the inequality are non-negative. Combining it with the previous estimate gives 2 cot(t − s) + cot ψ1(t, s) + cot ϕ+ 1 (t) cot ψ1(t, s) − cot ϕ+ 1 (t) . (4.3) cot ψ2(t, s) − cot ϕ+ 2 (t) ≤ (1 + ε)C2 cot ψ1(t, s) − cot ϕ+ 1 (t) whenever 0 < t − s < δ. 16 EDER KIKIANTY AND GORD SINNAMON Now suppose that s, t ∈ E and s < t. By Lemma 3.1(i), t−s = ψj(s, t)−ψj (t, s), for j = 1, 2, so (4.3) combines with (4.2) and (4.1) to give 2 (s)) 2 (t)) − (s + ϕ+ 2 (t) − ψ2(t, s) + ψ2(s, t) − ϕ+ 2 (t) + cot ϕ+ (t + ϕ+ ≤ ϕ+ ≤ cot ψ2(t, s) − cot ϕ+ ≤ (1 + ε)C2( cot ψ1(t, s) − cot ϕ+ ≤ (1 + ε)M C2(ϕ+ = (1 + ε)M C2(ϕ+ = (1 + ε)M C2((t + ϕ+ 1 (t)) − (s + ϕ+ 1 (t) + cot ϕ+ 1 (s) − cot ψ1(s, t)) 1 (t) − ψ1(t, s) + ψ1(s, t) − ϕ+ 1 (s)) 1 (t) − ψ1(t, s) + ψ1(s, t) − ϕ+ 1 (s)) 1 (s))). 2 (s) 2 (s) − cot ψ2(s, t) Removal of the absolute value signs to get the second-last line is justified by Lemma 3.1(ii). Now suppose β < ¯β < ¯α < α with ¯α, ¯β ∈ E. Since E is dense (β, α) we can choose t0 = ¯β < t1 < ··· < tn = ¯α with tk − tk−1 < δ and tk ∈ E for each k. Then (¯α + ϕ+ n 2 (tk−1)) 2 (¯α)) − ( ¯β + ϕ+ 2 ( ¯β)) = (tk + ϕ+ Xk=1 ≤ (1 + ε)M C2 = (1 + ε)M C2((¯α + ϕ+ 2 (tk)) − (tk−1 + ϕ+ Xk=1 (tk + ϕ+ n 1 (tk)) − (tk−1 + ϕ+ 1 ( ¯β))). 1 (¯α)) − ( ¯β + ϕ+ 1 (tk−1)) Letting ¯β decrease to β through E, ¯α increase to α through E, and ε → 0 we have (α + ϕ− 2 (α)) − (β + ϕ+ 2 (β)) ≤ M C2((α + ϕ− 2 (α)) − (β + ϕ+ 2 (β)). (cid:3) Somewhat surprisingly, the next result, initially expected to be easy, turned out to be the main result of the current article because of its involved proof. It is hoped that greater familiarity with angular equivalence will reveal a simpler one. Theorem 4.2. Let X be a real vector space with norms k · k1 and k · k2. If k · k1 and k · k2 are angularly equivalent then they are also topologically equivalent. Proof. Let C ≥ 1 be the constant in the definition of angular equivalence, that is, suppose that for all non-zero x, y ∈ X, tan(θ2(x, y)/2) ≤ C tan(θ1(x, y)/2). We will show that for all non-zero ¯x, ¯y ∈ X, (4.4) k¯xk1k¯yk2 k¯xk2k¯yk1 ≤ 40C2. Topological equivalence follows from this by fixing any non-zero ¯y ∈ X to get mk¯xk1 ≤ k¯xk2 with m = k¯yk2/(40C2kyk1) and, interchanging ¯x and ¯y in (4.4), k¯xk2 ≤ Mk¯xk1 with M = 40C2k¯yk2/k¯yk1. If ¯x and ¯y are multiples of one another then (4.4) holds trivially so we assume henceforth that they are independent. Let Z be the two-dimensional subspace of X spanned by ¯x and ¯y. Choose vectors x, y ∈ Z satisfying three conditions: (4.5) kxk2 = kyk2 = 1; kyk1 ≤ kzk1 kzk2 ≤ kxk1 for 0 6= z ∈ Z; kxk1 ≤ kx+tyk1 for t ≥ 0. ANGULAR EQUIVALENCE OF NORMED SPACES 17 To see that this is possible, take x and y to give the maximum and minimum, respectively, of the continuous function z 7→ kzk1 on the compact set {z ∈ Z : kzk2 = 1}. For any non-zero z ∈ Z, z/kzk2 is in this set so the first two conditions hold. But they also hold with y replaced by −y. The third condition must hold either with y or with −y; otherwise, there would exist s, t > 0 such that kx+ tyk1 < kxk1 and kx − syk1 < kxk1, which leads to the contradiction, s+tkx + tyk1 + t s+tkx − syk1 < kxk1. s+t (x − sy)k1 ≤ s kxk1 = k s s+t (x + ty) + t With x and y now fixed, we identify Z with R2 by the linear map (ξ, η) 7→ ξx+ηy so that k(ξ, η)k1 = kξx + ηyk1 and k(ξ, η)k2 = kξx + ηyk2 give two norms on R2. These make the identification isometric in both norms so g-functional calculations are not affected. In particular, the two norms on R2 are angularly equivalent with constant C. Define rj, ψj and ϕ± j as above for j = 1, 2. Note that the polar graph of rj coincides with the boundary of the k · kj-unit ball-we refer to it simply as the rj-curve. The first condition of (4.5) ensures that r2(0) = 1 and r2(π/2) = 1. For conve- nience we set a = r1(0) and b = r1(π/2). The second condition implies that, k¯xk1k¯yk2 k¯xk2k¯yk1 ≤ kxk1 kyk1 = b a , so we may complete the proof of (4.4) and the theorem by showing b/a ≤ 40C2. Suppose instead that b/a > 40C2. First, we set up to apply Lemma 4.1 with α = tan−1 3 and β = 0. Estimates will be needed for α + ϕ− 1 (0), α + ϕ− 2 (α) − ϕ+ 2 (0), and M = sup{csc2(ψ1(s, t)) : s, t ∈ (0, α), s 6= t}. For the remainder of the proof we will frequently apply the definitions of ψ and ϕ± and the properties in Lemma 3.1(i) and (ii) without explicit reference. Also, as above, "angle" refers to the usual angle in R2, not the norm angle. 1 (α) − ϕ+ The third condition of (4.5) shows that, in the first quadrant, the r1-curve lies to the left of the line x = a. Therefore ψ1(0, t) ≥ π/2 for t > 0 and, in the limit, ϕ+ 1 (0) ≥ π/2. Also, the r1-curve intersects the line y = 3x on the segment from (0, 0) to (a, 3a). Therefore ψ1(π/2, α) is less than or equal to the measure of the angle from (0, 0) to (0, b) to (a, 3a). That is, (4.6) ψ1(π/2, α) ≤ cot−1(b/a − 3). Note that since C ≥ 1 our assumption that b/a > 40C2 implies b/a > 3. Now, (4.7) α + ϕ− 1 (0) ≤ α + ψ1(α, π/2) − π/2 = ψ1(π/2, α) ≤ cot−1(b/a − 3). 1 (α) − ϕ+ To estimate ϕ+ 2 (0) we use the angular equivalence hypothesis applied to the two norm angles from (1, 0) to (cos(−t), sin(−t)), for t > 0. In the form (3.1), it implies cot ψ2(0,−t) − A2 2 cot t + cot ψ2(0,−t) + A2 ≤ C2 cot ψ1(0,−t) − A1 2 cot t + cot ψ1(0,−t) + A1 . 18 EDER KIKIANTY AND GORD SINNAMON where Aj = 1 ψ2 are uniformly bounded away from zero, but cot t → ∞ as t → 0+. Thus, j (0)), j = 1, 2. By Lemma 3.1(iii), A1, A2, ψ1 and j (0) + cot ϕ+ 2 (cot ϕ− cot ϕ− 2 (0) − cot ϕ+ 2 (0) = 2 lim t→0+ t→0+ ≤ 2 lim = C2(cot ϕ− cot ψ2(0,−t) − A2 2 cot t + cot ψ2(0,−t) + A2 cot ψ1(0,−t) − A1 C2 2 cot t + cot ψ1(0,−t) + A1 1 (0) − cot ϕ+ 1 (0)). (2 cot t + cot ψ2(0,−t) + A2) (2 cot t + cot ψ2(0,−t) + A2) Returning to the second condition of (4.5) we see that for each t, r1(t) r2(t) = k(cos t)x + (sin t)yk2 k(cos t)x + (sin t)yk1 ≥ 1 kxk1 = a. Recalling that r2(0) = 1 and applying Lemma 3.1(v), we get t→0− 0 ≤ lim Thus, cot ϕ− 1 t→0− r1(t) r2(t)(cid:19) = lim t (cid:18)a − 2 (0) ≥ cot ϕ− cot ϕ+ 2 (0) ≥ cot ϕ− 1 (0) and we have a r2(t) − 1 tr2(t) − r1(t) − a tr2(t) = a cot ϕ− 2 (0)− a cot ϕ− 1 (0). 2 (0) − C2(cot ϕ− 1 (0) − cot ϕ+ 1 (0) 1 (0)) 1 (0) + C2 cot ϕ+ ≥ (1 − C2) cot ϕ− ≥ (1 − C2) cot ψ1(0,−π/2) + C2 cot ψ1(0, π/2) = (1 − C2)(a/b) + C2(−a/b) ≥ −2C2a/b. We conclude that ϕ+ 2 (0) ≤ cot−1(−2C2a/b) = π − cot−1(2C2a/b). The r2-curve passes through (−1, 0) and (0, 1) so, in the first quadrant, it lies below the line y = x + 1. Thus, the r2-curve intersects the line y = 3x on the segment from (0, 0) to (1/2, 3/2). It follows that ψ2(α, 0) is greater than or equal to the measure of the apex angle of the isosceles triangle with vertices (0, 0), (1/2, 3/2), and (1, 0). The latter is π − 2α so we have (4.8) α + ϕ− 2 (α) − ϕ+ 2 (0) ≥ α + ψ2(α, 0) − ϕ+ 2 (0) ≥ cot−1(2C2a/b) − α. Finally we estimate M . Since b/a > 40C2 > 3 + 1/3, cot−1(b/a − 3) < cot−1(1/3) = α so (4.6) implies ψ1(α, π/2) = ψ1(π/2, α) − α + π/2 ≤ cot−1(b/a − 3) − α + π/2 ≤ π/2. Thus, for s, t ∈ [0, α) with s 6= t, ψ1(t, s) ≤ ψ1(t, α) = ψ1(α, t) + α − t ≤ ψ1(α, t) + α ≤ ψ1(α, π/2) + α ≤ π/2 + α. Combining this with, ψ1(t, s) ≥ ψ1(t, 0) = ψ1(0, t) − t ≥ ϕ+ gives π/2 − α ≤ ψ1(t, s) ≤ π/2 + α and hence 1 (0) − t ≥ π/2 − t ≥ π/2 − α csc2 ψ1(t, s) ≤ csc2(π/2 − α) = 10. ANGULAR EQUIVALENCE OF NORMED SPACES 19 So M ≤ 10. Now Lemma 4.1 gives, Using (4.7) and (4.8), we have α + ϕ− 2 (α) − ϕ+ 2 (0) ≤ 10C2(α + ϕ− 1 (α) − ϕ+ 1 (0)). cot−1(2C2a/b) − α ≤ 10C2 cot−1(b/a − 3). Since cot−1 u ≤ 1/u when u > 2/π, and C ≥ 1, our assumption b/a > 40C2 yields, 10C2 cot−1(b/a − 3) ≤ 10C2 cot−1(40C2 − 3C2) ≤ 10/37 < 0.271 and cot−1(2C2a/b) − α ≥ cot−1(2/40) − tan−1 3 > 0.271. This contradiction completes the proof. (cid:3) 5. Further Work There are many natural, fundamental questions about angular equivalence still to be investigated. We list a few in the hope that interested readers will contribute to the theory. Throughout, k·k1 and k·k2 are angularly equivalent norms on a real vector space X. By Theorem 4.2 the two norms are also topologically equivalent. Since the completions of k · k1 and k · k2 are norms on a common vector space it makes sense to ask, Question 1. Are the completions of angularly equivalent norms again angularly equivalent? If Y is a subspace of X then it is closed with respect to k · k1 if and only if it is closed with respect to k·k2. Each of the two norms give rise to a quotient norm on X/Y defined by, kx + Y kj = inf y∈Y kx + ykj, j = 1, 2. Question 2. Do angularly equivalent norms induce angularly equivalent norms on quotient spaces? The vector space X ∗, of continuous linear functionals on X, is the same for both norms. Thus, their respective dual norms k · k∗ 2 are norms on a common vector space. These dual norms are not, in general, angularly equivalent. See Example 3, below. So it is natural to ask, 1 and k · k∗ Question 3. Under what conditions on angularly equivalent norms are their dual norms also angularly equivalent? Example 3. The dual norms of angularly equivalent norms need not be angu- larly equivalent: Consider the two weighted ℓ1 norms, k(ξ, η)k1 = 2ξ + η and k(ξ, η)k2 = ξ + 2η on R2. A calculation shows that, with θj denoting the k · kj- norm angle from (ξ, η) to (µ, ν), tan2(θ2/2) = µ − µ sgn ξ + 2(ν − ν sgn η) µ + µ sgn ξ + 2(ν + ν sgn η) 2(µ − µ sgn ξ) + ν − ν sgn η 2(µ + µ sgn ξ) + ν + ν sgn η ≤ 4 = 4 tan2(θ1/2). (The operator sgn takes the value 1, −1 or 0 when its argument is positive, negative or zero, respectively.) Thus, the two norms are angularly equivalent. However, their dual norms are given by k(ξ, η)k∗ 2 = max(ξ,η/2), respectively. These have polygonal unit balls in which the vertices are not on the 1 = max(ξ/2,η) and k(ξ, η)k∗ 20 EDER KIKIANTY AND GORD SINNAMON same rays. One unit ball has vertices at (±2,±1) and the other has vertices at (±1,±2). By Corollary 2.3 the dual norms are not angularly equivalent. The Cartesian product and tensor product of normed vector spaces can be given various different, but topologically equivalent, norms. This leads to the question, Question 4. How should the norm of a Cartesian or tensor product be defined to ensure that product norms are angularly equivalent whenever the norms on the factors are angularly equivalent? An Orlicz space can be equipped with either of two topologically equivalent norms, the Luxemburg norm or the Orlicz norm. Both are needed because each arises naturally from the other when considering the norm on the dual space. For the special case of Lp spaces, these two norms coincide (up to a constant multiple) and hence are angularly equivalent. However, the examples following [3, Theorem 10] show that there are Orlicz spaces which are strictly convex when equipped with the Luxemburg norm but are not strictly convex when equipped with the Orlicz norm. By Corollary 2.2 the two norms are not angularly equivalent for such spaces. So we ask, Question 5. For which Orlicz spaces are the Luxemburg norm and the Orlicz norm angularly equivalent? A similar question could be posed for many other families. For example, topo- logically equivalent norms abound on the much-studied families of Banach spaces defined by Hardy, Sobolev, Besov, Triebel, Lizorkin and others. References [1] F. L. Bauer and A. S. Householder 'Some inequalities involving the euclidean condition of a matrix', Numer. Math. 2 (1960), 308–311. [2] S. Dragomir, Semi-Inner Products and Applications (Nova Science Publishers, Inc., Haup- pauge, NY, 2004). [3] A. Kami´nska 'Strict convexity of sequence Orlicz-Musielak spaces with Orlicz norm', J. Funct. Anal. 50 (1983) 285–305. [4] M. Lin and G. Sinnamon 'The generalized Wielandt inequality in inner product spaces', Eurasian Math. J. 3 (2012) 72–85. [5] P. Milici´c 'Sur le semi-produit scalaire dans quelques espaces vectorial norm´es', Mat. Vesnik 8 (1971) 181–185. [6] P. Milici´c 'Sur le g-angle dans un espace norm´e', Mat. Vesnik 45 (1998) 43–48. [7] P. Milici´c 'Characterizations of convexities of normed spaces by means of g-angles', Mat. Vesnik 54 (2002) 37–44. [8] H. Wielandt 'Inclusion theorems for eigenvalues', National Bureau of Standards Appl. Math. Series 29 (1953), 7–78 . E. Kikianty, Department of Mathematics and Applied Mathematics, University of Pretoria, Pretoria, South Africa E-mail address: [email protected] Department of Mathematics, University of Western Ontario, London, Canada E-mail address: [email protected]
1706.10110	2	1706	2017-11-08T08:23:55	On Using Toeplitz and Circulant Matrices for Johnson-Lindenstrauss Transforms	[ "math.FA", "cs.CC", "cs.DS" ]	The Johnson-Lindenstrauss lemma is one of the corner stone results in dimensionality reduction. It says that given $N$, for any set of $N$ vectors $X \subset \mathbb{R}^n$, there exists a mapping $f : X \to \mathbb{R}^m$ such that $f(X)$ preserves all pairwise distances between vectors in $X$ to within $(1 \pm \varepsilon)$ if $m = O(\varepsilon^{-2} \lg N)$. Much effort has gone into developing fast embedding algorithms, with the Fast Johnson-Lindenstrauss transform of Ailon and Chazelle being one of the most well-known techniques. The current fastest algorithm that yields the optimal $m = O(\varepsilon^{-2}\lg N)$ dimensions has an embedding time of $O(n \lg n + \varepsilon^{-2} \lg^3 N)$. An exciting approach towards improving this, due to Hinrichs and Vyb\'iral, is to use a random $m \times n$ Toeplitz matrix for the embedding. Using Fast Fourier Transform, the embedding of a vector can then be computed in $O(n \lg m)$ time. The big question is of course whether $m = O(\varepsilon^{-2} \lg N)$ dimensions suffice for this technique. If so, this would end a decades long quest to obtain faster and faster Johnson-Lindenstrauss transforms. The current best analysis of the embedding of Hinrichs and Vyb\'iral shows that $m = O(\varepsilon^{-2}\lg^2 N)$ dimensions suffices. The main result of this paper, is a proof that this analysis unfortunately cannot be tightened any further, i.e., there exists a set of $N$ vectors requiring $m = \Omega(\varepsilon^{-2} \lg^2 N)$ for the Toeplitz approach to work.	math.FA	math	On Using Toeplitz and Circulant Matrices for Johnson-Lindenstrauss Transforms Casper Benjamin Freksen∗ Kasper Green Larsen∗ Abstract The Johnson-Lindenstrauss lemma is one of the corner stone results in dimensionality reduction. It says that given N , for any set of N vectors X ⊂ Rn, there exists a mapping f : X → Rm such that f (X) preserves all pairwise distances between vectors in X to within (1 ± ε) if m = O(ε−2 lg N ). Much effort has gone into developing fast embedding algorithms, with the Fast Johnson- Lindenstrauss transform of Ailon and Chazelle being one of the most well-known techniques. The current fastest algorithm that yields the optimal m = O(ε−2 lg N ) dimensions has an embedding time of O(n lg n + ε−2 lg3 N ). An exciting approach towards improving this, due to Hinrichs and Vybíral, is to use a random m × n Toeplitz matrix for the embedding. Using Fast Fourier Transform, the embedding of a vector can then be computed in O(n lg m) time. The big question is of course whether m = O(ε−2 lg N ) dimensions suffice for this technique. If so, this would end a decades long quest to obtain faster and faster Johnson-Lindenstrauss transforms. The current best analysis of the embedding of Hinrichs and Vybíral shows that m = O(ε−2 lg2 N ) dimensions suffices. The main result of this paper, is a proof that this analysis unfortunately cannot be tightened any further, i.e., there exists a set of N vectors requiring m = Ω(ε−2 lg2 N ) for the Toeplitz approach to work. 1 Introduction The performance of many geometric algorithms depends heavily on the dimension of the input data. A widely used technique to combat this "curse of dimensionality", is to preprocess the input via dimension- ality reduction while approximately preserving important geometric properties. Running the algorithm on the lower dimensional data then uses less resources (time, space, etc.) and an approximate result for the high dimensional data can be derived from the low dimensional result. Dimensionality reduction approximately preserving pairwise Euclidean distances has found uses in a wide variety of applications, including: Nearest-neighbour search [2, 12], clustering [6, 8], linear program- ming [22], streaming algorithms [19], compressed sensing [7, 11], numerical linear algebra [25], graph sparsification [20], and differential privacy [5]. See more applications in [21, 14]. The most fundamental result in this regime is the Johnson-Lindenstrauss (JL) lemma [15], which says the following: Theorem 1 (Johnson-Lindenstrauss lemma). Let X ⊂ Rn be a set of N vectors, then for any 0 < ε < 1/2, there exists a map f : X → Rm for some m = O(ε−2 lg N ) such that ∀x, y ∈ X, (1 − ε)kx − yk2 2 ≤ kf (x) − f (y)k2 2 ≤ (1 + ε)kx − yk2 2. 7 1 0 2 v o N 8 ] . A F h t a m [ 2 v 0 1 1 0 1 . 6 0 7 1 : v i X r a This result dates back to 1984 and says that to preserve pairwise Euclidean distances amongst a set of N points/vectors in Rn to within a factor (1 ± ε), it suffices to use just m = O(ε−2 lg N ) dimensions. The bound on m was very recently proven optimal [18]. The standard technique for constructing a map with the properties of Theorem 1 is the following: Let A be an m × n matrix with entries independently sampled as either N (0, 1) random variables (as in [10]) or Rademacher (uniform among {−1, +1}) random variables (as in [1]). Once such entries have been drawn, let f : Rn → Rm be defined as: f (x) = Ax. 1 √m ∗This research is supported by a Villum Young Investigator Grant, an AUFF Starting Grant and MADALGO, Center for Massive Data Algorithmics, a Center of the Danish National Research Foundation, grant DNRF84. Department of Computer Science, Aarhus University. {cfreksen,larsen}@cs.au.dk. 1 To prove that the map f satisfies the guarantees in Theorem 1, it is first shown that for any vector x, the probability that kf (x)k2 2 is less than 1/N 2. This probability is called the error probability and denoted δ. Using linearity of f and a union bound over all pairs x, y ∈ X, the probability that all pairwise distances (i.e. the norm of the vector x − y) are preserved can be shown to be at least 1/2. 2 is not within (1 ± ε)kxk2 1.1 Time Complexity Examining the classic Johnson-Lindenstrauss reduction above, we see that to embed a vector, we need to multiply with a dense matrix and the embedding time becomes O(nm) (or equivalently O(nε−2 lg N )). This may be prohibitively large for many applications (recall one prime usage of dimensionality reduction is to speed up algorithms), and much research has been devoted to obtaining faster embedding time. Fast Johnson-Lindenstrauss Transform. Ailon and Chazelle [2] were the first to address the ques- tion of faster Johnson-Lindenstrauss transforms. In their seminal paper, they introduced the so-called Fast Johnson-Lindenstrauss transform for speeding up dimensionality reduction. The basic idea in their paper is to first "precondition" the input data by multiplying with a diagonal matrix with random signs, followed by multiplying with a Hadamard matrix. This has the effect of "spreading" out the mass of the input vectors, allowing for the dense matrix A above to be replaced with a sparse matrix. Since we can multiply with a Hadamard matrix using Fast Fourier Transform, this gives an embed- ding time of O(n lg n + ε−2 lg3 N ) for embedding into the optimal m = O(ε−2 lg N ) dimensions. For m = ε−2 lg N ≤ n1/2−γ for any constant γ > 0, the embedding complexity was improved even further down to O(n lg m) in [3]. Another approach to achieve the O(n lg m) embedding time, but without the restriction on ε−2 lg N ≤ n1/2−γ, is to sacrifice the target dimension. This was done in [4] and later improved in [17], where the embedding complexity was O(n lg m) at the cost of an increased target dimension m = O(ε−2 lg N lg4 n). Sparse Vectors. Another approach to improve the performance of JL transforms, is to assume the input data is sparse, i.e. has few non-zero coordinates. Designing an algorithm based on the work in [24], Dasgupta et al. [9] achieved an embedding complexity of O(kxk0ε−1 lg2(mN ) lg N ), where kxk0 = {i xi 6= 0}. This was later improved to O(kxk0ε−1 lg N ) in [16]. Toeplitz Matrices. Finally, another very exciting approach is to use Toeplitz matrices or partial circulant matrices for the embedding. We first introduce the terminology. An m × n Toeplitz matrix is an m × n matrix, where every entry on a diagonal has the same value: t0 t−1 t−2 ... t1 t0 t−1 ... t2 t1 t0 ... t−(m−1) t−(m−2) t−(m−3)   ··· ··· ··· . . . ··· tn−1 tn−2 tn−3 ... tn−m   A partial circulant matrix is a special kind of Toeplitz matrix, where every row, except the first, is the previous row rotated once: t0 tn−1 tn−2 ... t1 t0 tn−1 ... t2 t1 t0 ... tn−(m−1) tn−(m−2) tn−(m−3) ··· ··· ··· . . . ··· tn−1 tn−2 tn−3 ... tn−m     Hinrichs and Vybíral [13] proposed the following algorithm for generating a JL embedding based on a Toeplitz matrix1: Let t−(m−1), t−(m−2), . . . , tn−1 and d1, . . . , dn be i.i.d. Rademacher random variables, and T be a Toeplitz matrix defined from t−(m−1), t−(m−2), . . . , tn−1 such that entry (i, j) takes values 1[13] uses a partial circulant matrix but notes that a Toeplitz matrix could be used as well. 2 Table 1: Comparison of the performances of various Johnson-Lindenstrauss transform algorithms. N is the number of input vectors, n is the dimension of the input vectors, m is the dimension of the output vectors, ε is the distortion. Embedding time Type Random projection O(nm) Sparse Sparse FFT FFT FFT Toeplitz Toeplitz Target dimension (m) Ref. Notes O(ε−2 lg N ) O(kxk0ε−1 lg2(mN ) lg N ) O(ε−2 lg N ) O(ε−2 lg N ) O(kxk0ε−1 lg N ) O(n lg n + m lg2 N ) O(ε−2 lg N ) O(ε−2 lg N ) O(n lg m) O(ε−2 lg N lg4 n) O(n lg m) O(ε−2 lg3 N ) O(n lg m) O(ε−2 lg2 N ) O(n lg m) [10] [9] [16] [2] [3] [17] [13] [23] m ≤ n1/2−γ tj−i for i = 1, . . . , m and j = 1, . . . , n. Let D be an n × n diagonal matrix with the random variable di giving the i'th diagonal entry. Define the map f as f (x) = 1 √m T Dx. Multiplying with a Toeplitz matrix corresponds to computing a convolution and can be done using Fast Fourier Transform. By appropriately blocking the input coordinates, the complexity of embedding a vector x is just O(n lg m) for any target dimension m. The big question is of course, how low can the target dimension m be, while preserving the distances between vectors up to a factor of 1 ± ε? In the original paper [13], the authors proved that setting the target dimension to m = O(ε−2 lg3(1/δ)), the norm of any vector would be preserved to within (1 ± ε) with probability at least 1 − δ. Setting δ = 1/N 2, a union bound over all pairwise difference vectors (as in the classic construction) shows that dimension m = O(ε−2 lg3 N ) suffices. Later, the analysis was refined in [23], which lowered the target dimension to m = O(ε−2 lg2(1/δ)) for preserving norms to within (1 ± ε) with probability 1 − δ. Again, setting δ = 1/N 2, this gives m = O(ε−2 lg2 N ) target dimension. Now if the analysis could be tightened even further to give the optimal m = O(ε−2 lg N ) dimensions, this would end the decades long quest for faster and faster embedding algorithms! Our Contribution. Our main result unfortunately shows that the analysis of Vybíral [23] cannot be tightened to give an even lower target dimensionality. More specifically, we prove that the upper bound given in [23] is optimal: Theorem 2. Let T and D be the m × n Toeplitz and n × n diagonal matrix in the embedding proposed by [13]. For all 0 < ε < C, where C is a universal constant, and any desired error probability δ > 0, if the following holds for every unit vector x ∈ Rn: then it must be the case that m = Ω(ε−2 lg2(1/δ)). Pr"(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:13)(cid:13)(cid:13)(cid:13) 1 √m 2 2 − 1(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) T Dx(cid:13)(cid:13)(cid:13)(cid:13) < ε# > 1 − δ, While Theorem 2 already shows that one cannot tighten the analysis of Vybíral for preserving the norm of just one vector, Theorem 2 does leave open the possibility that one would not need to union bound over all N 2 pairs of difference vectors when trying to preserve all pairwise distances amongst a set of N vectors. It could still be the case that there somehow was a strong positive correlation between distances being preserved (though this seems extremely unlikely, and would be something not seen in any previous approach to JL). To complete the picture, we indeed show that this is not the case, at least for N somewhat smaller than the dimension n: Theorem 3. Let T and D be the m × n Toeplitz and n × n diagonal matrix in the embedding proposed by [13]. For all 0 < ε < C, where C is a universal constant, if the following holds for every set of N 3 vectors X ⊂ Rn: Pr"∀x, y ∈ X :(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:13)(cid:13)(cid:13)(cid:13) 1 √m T Dx − 1 √m 2 − kx − yk2 ≤ εkx − yk2 2# = Ω(1), 2 T Dy(cid:13)(cid:13)(cid:13)(cid:13) 2(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) then it must be the case that either m = Ω(ε−2 lg2 N ) or m = Ω(n/N ). We remark that our proofs also work if we replace T be a partial circulant matrix (which was also proposed in [13]). Furthermore, we expect that minor technical manipulations to our proof would also show the above theorems when the entries of T and D are N (0, 1) distributed rather than Rademacher (this was also proposed in [13]). 2 Lower Bound for One Vector Let T be m× n Toeplitz matrix defined from random variables t−(m−1), t−(m−2), . . . , tn−1 such that entry (i, j) takes values tj−i for i = 1, . . . , m and j = 1, . . . , n. Let D be an n × n diagonal matrix with the random variable di giving the i'th diagonal entry. This section shows the following: Theorem 4. Let T be m× n Toeplitz and D n× n diagonal. If t−(m−1), t−(m−2), . . . , tn−1 and d1, . . . , dn are independently distributed Rademacher random variables for i = −(m− 1), . . . , n− 1 and j = 1, . . . , n, then for all 0 < ε < C, where C is a universal constant, there exists a unit vector x ∈ Rn such that and furthermore, all but the first O(√m) coordinates of x are 0. Pr"(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:13)(cid:13)(cid:13)(cid:13) 1 √m 2 2 − 1(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) T Dx(cid:13)(cid:13)(cid:13)(cid:13) > ε# ≥ 2−O(ε√m). It follows from Theorem 4 that if we want to have probability at least 1− δ of preserving the norm of any unit vector x to within (1± ε), it must be the case that ε√m = Ω(lg(1/δ)), i.e. m = Ω(ε−2 lg2(1/δ)). This is precisely the statement of Theorem 2. Thus we set out to prove Theorem 4. To prove Theorem 4, we wish to invoke the Paley-Zygmund inequality, which states, that if X is a non-negative random variable with finite variance and 0 ≤ θ ≤ 1, then Pr[X > θE[X]] ≥ (1 − θ)2 E2[X] E[X 2] . We carefully choose a unit vector x, and define the random variable for Paley-Zygmund to be the k'th moment of the difference between the norm of x transformed and 1. Proof. Let k be an even positive integer less than m/4 and define s := 4k. Note that s ≤ m. Let x be an arbitrary n-dimensional unit vector such that the first s coordinates are in {−1/√s, +1/√s}, while the remaining n − s coordinates are 0. Define the random variable parameterized by k Zk := (cid:13)(cid:13)(cid:13)(cid:13) 1 √m T Dx(cid:13)(cid:13)(cid:13)(cid:13) 2 2 − 1!k . Since k is even, the random variable Zk is non-negative. We wish to lower bound E[Zk] and upper bound E[Z 2 bounds we prove are as follows: Lemma 5. If k ≤ √m, then the random variable Zk satisfies: and E[Zk] ≥ m−k/2kk2−O(k) k ] ≤ m−kk2k2O(k). E[Z 2 4 k ] in order to invoke Paley-Zygmund. The Before proving Lemma 5 we show how to use it together with Paley-Zygmund to complete the proof of Theorem 4. We start by invoking Paley-Zygmund and then rewriting the expectations according to Lemma 5, E2[Zk] E[Z 2 k] E2[Zk] E[Z 2 k] =⇒ =⇒ Pr[Zk > E[Zk]/2] ≥ (1/4) Pr[Z 1/k k > (E[Zk]/2)1/k] ≥ (1/4) C0√m# ≥ 2−O(k). > k 1 √m Pr"(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:13)(cid:13)(cid:13)(cid:13) 2 2 − 1(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) T Dx(cid:13)(cid:13)(cid:13)(cid:13) Here C0 is some constant greater than 0. For any 0 < ε < 1/C0, we can now set k such that k/(C0√m) = ε, i.e. we choose k = εC0√m. This choice of k satisfies k ≤ √m as required by Lemma 5. We have thus shown that: Pr"(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:13)(cid:13)(cid:13)(cid:13) 1 √m 2 2 − 1(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) T Dx(cid:13)(cid:13)(cid:13)(cid:13) > ε# ≥ 2−O(ε√m). Remark. Theorem 4 can easily be extended to partial circulant matrices. The difference between partial circulant and Toeplitz matrices is the dependence between the values in the first m and last m columns. However, as only the first s = 4k ≤ 4√m entries in x are nonzero, the last m columns are ignored, and so partial circulant and Toeplitz matrices behave identically in our proof. Proof of Lemma 5. Before we prove the two bounds in Lemma 5 individually, we rewrite E[Zk], as this benefits both proofs. 1 m E[Zk] = E  (cid:13)(cid:13)(cid:13)(cid:13) 1 √m = E      = E      = E    = E    1 m 1 m 1 m m Xi=1 m Xi=1 = 1 mk 2 n n n m m j x2 Xj=1 t2 j−id2 T Dx(cid:13)(cid:13)(cid:13)(cid:13)  Xi=1 Xj=1    Xi=1     Xj=1   Xj=1 Xh∈{1,...,n}\{j} 2 − 1!k  k 2  − 1 tj−idjxj    j  + Xj=1 Xh∈{1,...,n}\{j}  j  + Xj=1 Xh∈{1,...,n}\{j} j − x2  k tj−ith−idj dhxjxh   tj−ith−idjdhxjxh E  Y(i,j,h)∈S  t2 j−id2 X j x2 n n n S∈([m]×[n]×[n])k∀(i,j,h)∈S:h6=j k tj−ith−idj dhxjxh    − 1     k   tj−ith−idjdhxj xh     Observe that for j > s or h > s the product becomes 0, as either xj or xh is 0. By removing all these terms, we simplify the sum to E[Zk] = 1 mk S∈([m]×[s]×[s])k∀(i,j,h)∈S:h6=j X E  Y(i,j,h)∈S tj−ith−idjdhxj xh  5 following two things are true: Observe for an S ∈ ([m] × [s] × [s])k, that the value EhQ(i,j,h)∈S tj−ith−idj dhxjxhi is 0 if one of the • A dj occurs an odd number of times in the product. • A variable ta occurs an odd number of times in the product. To see this, note that by the independence of the random variables, we can write the expectation of the product, as a product of expectations where each term in the product has all the occurrences of the same random variable. Since the dj 's and ta's are Rademachers, the expectation of any odd power of one of these random variables is 0. Thus if just a single random variable amongst the dj 's and ta's occurs an odd number of times, we have EhQ(i,j,h)∈S tj−ith−idjdhxjxhi = 0. Similarly, we observe that if every random variable occurs an even number of times, then the expectation of the product is exactly 1/sk since each xj also occurs an even number of times. If Γk denotes the number of tuples S ∈ ([m]×[s]×[s])k such that ∀(i, j, h) ∈ S we have h 6= j and furthermore: • For all columns a ∈ [s], {(i, j, h) ∈ S j = a ∨ h = a} mod 2 = 0. • For all diagonals a ∈ {−(m − 1), . . . , s − 1}, {(i, j, h) ∈ S j − i = a ∨ h − i = a} mod 2 = 0. Then we conclude Note that Z 2 k = Z2k. Therefore, E[Zk] = Γk skmk . E[Z 2 k ] = E[Z2k] = Γ2k s2km2k . (1) (2) To complete the proof of Lemma 5 we need lower and upper bounds for Γk and Γ2k. The bounds we prove are Lemma 6. If k ≤ √m, then Γk and Γ2k satisfy: and Γk = mk/2skkk2−O(k) Γ2k = mks2kk2k2O(k). The proofs of the two bounds in Lemma 6 are in Sections 2.1 and 2.2. Substituting the bounds from Lemma 6 in (1) and (2) we get which are the bounds we sought for Lemma 5. E[Zk] = m−k/2kk2−O(k) E[Z 2 k ] = m−kk2k2−O(k), 2.1 Lower Bounding Γk We first recall that the definition of Γk is the number of tuples S ∈ ([m] × [s] × [s])k satisfying that ∀(i, j, h) ∈ S we have h 6= j and furthermore: • For all columns a ∈ [s], {(i, j, h) ∈ S j = a ∨ h = a} mod 2 = 0. • For all diagonals a ∈ {−(m − 1), . . . , s − 1}, {(i, j, h) ∈ S j − i = a ∨ h − i = a} mod 2 = 0. We view a triple (i, j, h) ∈ ([m] × [s] × [s]) as two entries (i, j) and (i, h) in an m × s matrix. Furthermore, when we say that a triple touches a column or diagonal, a matrix entry of the triple lie on that column or diagonal, so (i, j, h) touches columns j and h and diagonals j − i and h− i. Similarly, we say that a tuple S ∈ ([m]× [s]× [s])k touches a given column or diagonal l times, if l triples in S touches that column or diagonal. We intent to prove a lower bound for Γk by constructing a big family of tuples F ⊆ ([m] × [s] × [s])k, where each tuple satisfies, that each column and diagonal touched by that tuple is touched exactly twice. As each column and diagonal is touched an even number of times, the number of tuples in the family is a lower bound for Γk. 6 Proof of Γk = mk/2skkk2−O(k). We describe how to construct a family of tuples F ⊆ ([m] × [s] × [s])k satisfying that ∀S ∈ F,∀(i, j, h) ∈ S we have h 6= j and furthermore: • For all columns a ∈ [s], {(i, j, h) ∈ S j = a ∨ h = a} ∈ {0, 2}. • For all diagonals a ∈ {−(m − 1), . . . , s − 1}, {(i, j, h) ∈ S j − i = a ∨ h − i = a} ∈ {0, 2}. From this and the definition of Γk it is clear that F ≤ Γk. When constructing an S ∈ F, we view S as consisting of two halves S1 and S2, such that S1 touches exactly the same columns and diagonals as S2 and both S1 and S2 touches each column and diagonal at most once. To capture this, we give the following definition, where S is meant to be the family of such halves S1 and S2. Definition 7. Let S be the set of all tuples S ∈ ([m] × [s] × [s])k/2 such that • ∀(i, j, h) ∈ S, j 6= h • For all columns a ∈ [s],{(i, j, h) ∈ S j = a ∨ h = a} ≤ 1 • For all diagonals a ∈ {−(m − 1), . . . , s − 1},{(i, j, h) ∈ S j − i = a ∨ h − i = a} ≤ 1 Definition 7 mimics the definition of Γk, and the first item in Definition 7 ensures that the triples in a tuple in S are of the same form as in Γk. The final two items ensure that each column and diagonal, respectively, is touched at most once. This is exactly the properties we wanted of S1 and S2 individually. We can now construct F as all pairs of (half) tuples S1, S2 ∈ S, such that S1 touches exactly the same columns and diagonals as S2. To capture that S1 and S2 touch the same columns and diagonals, we introduce the notion of a signature. A signature of Si is the set of columns and diagonals touched by Si. To have S1 and S2 touch exactly the same columns and diagonals, it is necessary and sufficient that they have the same signature. We introduce the following notation: B denotes the number of signatures with at least one member, and by enumerating the signatures, we let bi denote the number of (half) tuples in S with signature i. We recall that a (half) tuple S1 ∈ S touches each column and diagonal at most once, and if S1 and S2 share the same signature, they touch exactly the same columns and diagonals. Therefore, using ◦ to mean concatenation, S = S1 ◦ S2 ∈ F, as each column and diagonal touched is touched exactly twice. Therefore F is a lower bound for Γk. Note that for a given signature i, the number of choices of S1 and S2 with that signature is b2 i . This gives the following inequality, Γk ≥ F = b2 i . B Xi=1 We now apply the Cauchy-Schwarz inequality: b2 i B Xi=1 B Xi=1 12 ≥(cid:0) B Xi=1 bi(cid:1)2 =⇒ B Xi=1 b2 i=1 bi(cid:1)2 i ≥ (cid:0)PB PB i=1 12 =⇒ Γk ≥ S2 B . (3) To get a lower bound on S2/B (and in turn Γk), we need a lower bound on S and an upper bound on B. These bounds are stated in the following lemmas Lemma 8. S = Ω(mk/2sk2−k). Lemma 9. B = O(cid:16)(cid:0)m+s k(cid:1)(cid:17) k/2(cid:1)sk/2(cid:0)s Before proving any of these lemmas, we show that they together with (3) give the desired lower bound on Γk: Γk = Ω(mk/2sk2−k)2 O(cid:16)(cid:0)m+s k(cid:1)(cid:17) k/2(cid:1)sk/2(cid:0)s = Ω(cid:18) mks2k2−2k(k/2)k/2kk (m + s)k/2sk/2sk (cid:19) . (4) 7 Because s = 4k, we have (k/2)k/2 sk/2 = 2−Θ(k), and because s ≤ m: we can simplify (4), mk (m+s)(k/2) = mk/22−Θ(k). With this Γk = mk/2skkk2−O(k). which is the lower bound we sought. Proof of Lemma 8. Recall that S ⊆ ([m] × [s] × [s])k/2 is the set of (half) tuples that touch each column and diagonal at most once, and, for each triple (i, j, h) in these (half) tuples, we have j 6= h. We prove Lemma 8 by analysing how we can create a large number of distinct S ∈ S by choosing the triples in S iteratively. For each triple, we choose a row and two distinct entries on this row. We choose the row among any of the m rows. However, because S ∈ S, when choosing entries on the row, we cannot choose entries that lie on columns or diagonals touched by previously chosen triples. Instead we choose the two entries among any of the other entries. Therefore, whenever we choose a triple, this triple prevents at most four row entries from being chosen for every subsequent triple, as the two diagonals and two columns touched by the chosen triple intersect with at most four entries on the rows of the subsequent triples. This leads to the following recurrence, describing a lower bound for the number of triples F (r, c, t) =(r · c · (c − 1) · F (r, c − 4, t − 1) 1 if t > 0 otherwise (5) where r is the number of rows to choose from, c is the minimum number of choosable entries in any row, and t is the number of triples left to choose. Inspecting (5), we can see that F can equivalently be defined as F (r, c, t) = rt (c − 4i)(c − 1 − 4i). t−1 Yi=0 (6) If t ≤ c 8 then the terms inside the product in (6) are greater than c 2 , so we can bound F from below: We now insert the values for r, c and t to find a lower bound for S, noting that s = 4k ensures that t ≤ c 8 : = rtc2t 1 4t . c 2(cid:1)2t F (r, c, t) ≥ rt(cid:0) 2(cid:1) ≥ mk/2sk 1 k S ≥ F(cid:0)m, s, 4k/2 =⇒ S = Ω(mk/2sk2−k). Proof of Lemma 9. Recall that for at triple S ∈ S we define the signature as the set of columns and diagonals touched by S. Furthermore, viewing a triple (i, j, h) ∈ ([m] × [s] × [s]) as the two entries (i, j) and (i, h) in an m × s matrix, we define the left endpoint as (i, min{j, h} and the right endpoint as (i, max{j, h}). The claim to prove is This is proven by first showing an upper bound on the number of choices for the diagonals of left endpoints, then diagonals of right endpoints and finally for columns. B = O(cid:18)(cid:18)m + s k/2 (cid:19)sk/2(cid:18)s k(cid:19)(cid:19) . k/2(cid:1) choices for the diagonals corresponding to left endpoints in a triple. In an m× s matrix there are m + s different diagonals and as the chosen diagonals have to be distinct, there are (cid:0)m+s As the right endpoint of a triple has to be in the same row as the left endpoint, there are at most s choices for the diagonal corresponding to the right endpoint when the left endpoint has been chosen (which it has in our case). This gives a total of sk/2 choices for diagonals corresponding to right endpoints. Finally, there are s columns to choose from and the chosen columns have to be distinct, and so the total number of choices of columns is (cid:0)s k(cid:1). The product of these number of choices gives the upper bound sought. 8 2.2 Upper Bounding Γ2k Proof. Recall that Γ2k is defined as the number of tuples S ∈ ([m] × [s] × [s])2k such that ∀(i, j, h) ∈ S we have h 6= j and furthermore: • For all columns a ∈ [s], {(i, j, h) ∈ S j = a ∨ h = a} mod 2 = 0. • For all diagonals a ∈ {−(m − 1), . . . , s − 1}, {(i, j, h) ∈ S j − i = a ∨ h − i = a} mod 2 = 0. Let F ⊆ ([m] × [s] × [s])2k be the family of tuples satisfying these conditions, and so F = Γ2k. To prove an upper bound on Γ2k, we show how to encode a tuple S ∈ F using at most k lg m + 2k lg s + 2k lg k + O(k) bits, such that S can be decoded from this encoding. Since any S ∈ F can be encoded using k lg m + 2k lg s + 2k lg k + O(k) bits and F = Γ2k, we can conclude: Γ2k = 2k lg m+2k lg s+2k lg k+O(k) = mks2kk2k2O(k). Let σ denote the encoding function and σ−1 denote the decoding function. If S ∈ F and t ∈ S, σ(t) denotes the encoding of the triple t, σ(S) denotes the encoding of the entire tuple S, and σ(F ) denotes the image of σ. A tuple S ∈ F consists of triples t1, t2, . . . , t2k such that S = t1 ◦ t2 ◦ ··· ◦ t2k. To encode S ∈ F we We will first describe a graph view of a tuple S which will be useful for encoding and decoding, then encode each of the triples and store them in the same order: σ(S) = σ(t1) ◦ σ(t2) ◦ ··· ◦ σ(t2k). we will show an encoding algorithm and finally a decoding algorithm. Graph A tuple S ∈ F forms a (multi)graph structure, where every triple (i, j, h) ∈ S is a vertex. Since S ∈ F, there lies an even number of triple endpoints on each diagonal. We can thus pair endpoints lying on the same diagonal, such that every endpoint is paired with exactly one other endpoint. When two triples have endpoints that are paired, the triples have an edge between them in the graph. As every triple has two endpoints, every vertex has degree two, and so the graph consists entirely of simple cycles of length at least two. Encoding To encode an S ∈ F, we first encode each cycle by itself by defining the σ(t)'s for the triples t in the cycle. After this, we order the defined σ(t)'s as the t's were ordered in the input. 1. For each cycle we perform the following: (a) We pick any vertex t = (i, j, h) of the cycle and give it the type head. Define σ(t) as the concatenation of its type head, its row i, and two columns j and h. This uses lg m+2 lg s+O(1) bits. (b) We iterate through the cycle starting after head and give vertices, except the last, type mid. The last vertex just before head is given the type last. (c) For each triple t of type mid we store its type and two columns explicitly. However, instead of storing its row we store the index of its predecessor in the cycle order as well as how they are connected: If we typed tr just before ts when iterating through the cycle, when encoding ts we store r as well as whether tr and ts are connected by the left or right endpoint in tr and left or right endpoint in ts. So define σ(t) as the concatenation of mid, the two columns, the predecessor index and how it is connected to the predecessor. All in all we spend lg k + 2 lg s + O(1) bits encoding each mid. (d) Finally, to encode the triple t, which is typed last, we define σ(t) as the concatenation of its type, its predecessors index, how it is connected to its predecessor, and the column of the endpoint on the predecessor's diagonal. We thus spend lg k + lg s + O(1) bits to encode a last. However, since s = 4k the number of bits per encoded last is equivalent to 2 lg k +O(1), which turns out to simplify the analysis later. Note that for each triple, the type is encoded in the first 2 bits2 in the encoding of the triple. This will be important during decoding. 2To encode the type in 2 bits we could use the following scheme: 00 = head, 01 = mid, 10 = last, and 11 is unused. 9 2. After encoding all cycles, we order the encoded triples in the same order as the triples in the input, and output the concatenation of the encoded triples: σ(t1) ◦ σ(t2) ◦ ··· ◦ σ(t2k). To analyse the number of bits needed in total, we look at the average number of bits per triple inside a cycle. Since all cycles have a length of at least two, each cycle has a head and a last triple. These two use an average of lg m 2 + lg s + lg k +O(1) bits. We now claim that the number of bits per mid is bounded by this average. Recall that we assumed √m ≥ k and s = 4k: √m ≥ k =⇒ lg m 2 ≥ lg k =⇒ lg m lg m 2 + lg s + lg k + 2 ≥ lg k + 2 lg s =⇒ 2 + lg s + lg k + O(1) ≥ mid. 2 + lg s + lg k + O(1) bits, the number of bits for all Since the average of all 2k triples is at most lg m triples is at most k lg m + 2k lg s + 2k lg k + O(k). Decoding We now show how to decode any σ(S) ∈ σ(F ) such that σ−1(σ(S)) = S. we restore the order of triples. The following steps describes this algorithm in more detail. First we need to extract the encodings of the individual triples, then we decode each cycle, and finally 1. We first extract the individual σ(ti)'s, by iterating through the bit string σ(S). By looking at the type in the first two bits of an encoded triple we know the length of the encoded triple. From this we can extract σ(ti) as well as know where σ(ti+1) begins. 2. We now wish to decode each cycle to get the row and two columns of every triple, so do the following for each head ti: (a) t ← ti (b) Look at the first two bits in t to determine its type, and while t is not a last, do: i. If t is a head, the row and two columns are stored explicitly. ii. If t is a mid, the two columns are stored explicitly. From the reference to t's predecessor, we can calculate the diagonal shared between t and its predecessor. This is can be done as we have stored which endpoint (left or right) of the predecessor lies on the diagonal, and as we decode in the same order as we encoded, we have already decoded the predecessor. We know which of t's columns the shared diagonal intersects, and so we can calculate the row. If the predecessor is a head, take note of which diagonal is shared with it. iii. t ← the triple that has t as its predecessor. (c) t is now a last, so we know the predecessor's index as well as a column. However, as the graph consists of cycles rather than just linked lists, last shares a diagonal with its predecessor as well as sharing a diagonal with head. We have noted which diagonal has already been used for head, so the other diagonal head touches is shared with last. From these two diagonals and the column, it is possible to calculate the row by the intersection between the predecessor diagonal and the column, and the other column by the intersection of the row and the other diagonal. 3. Finally order the triples as they were, when the encoded versions were extracted in step 1. 3 Lower Bound for N Vectors In this section, we generalize the result of Section 2 to obtain a lower bound for preserving all pairwise distances amongst a set of N vectors. Our proof uses Theorem 4 as a building block. Recall that Theorem 4 guarantees that there is a vector x such that 1 √m Pr"(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:13)(cid:13)(cid:13)(cid:13) 2 2 − 1(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) T Dx(cid:13)(cid:13)(cid:13)(cid:13) 10 > ε# ≥ 2−O(ε√m). Moreover, the vector x has non-zeroes only in the first O(√m) coordinates. From such a vector x, define x→i as the vector having its j'th coordinate equal to the (j − i)'th coordinate of x if j > i and otherwise its j'th coordinate is 0. In words, x→i is just x with all coordinates shifted by i. For i ≤ n − O(√m) (i small enough that all the non-zero coordinates of x stay within the n coordin- ates), we see that T Dx and T Dx→i have the exact same distribution. Furthermore, if i ≥ m + C√m for a big enough constant C, T Dx and T Dx→i are independent. To see this, observe that the only random variables amongst t−(m−1), . . . , tn−1 that are multiplied with a non-zero coordinate of x in T Dx are t−(m−1), . . . , tO(√m). Similarly, only random variables d1, . . . , dO(√m) amongst d1, . . . , dn are multi- plied by a non-zero coordinate of x. For x→i, the same is true for variables t−(m−1)+i, . . . , tO(√m)+i and di+1, . . . , di+O(√m). If i ≥ m + C√m, these sets of variables are disjoint and hence T Dx and T Dx→i are independent. With the observations above in mind, we now define a set of vectors X as follows X := {0, x, x→m+C√m, x→2(m+C√m), . . . , x→⌊(n−C√m)/(m+C√m)⌋(m+C√m)}. 1√m T D as embedding. Furthermore, by the The 0-vector clearly maps to the 0-vector when using arguments above, the embeddings of all the remaining vectors are independent and have the same distribution. It follows that Pr"∀x→i ∈ X :(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:13)(cid:13)(cid:13)(cid:13) 1 √m 2 2 − 1(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) T Dx→i(cid:13)(cid:13)(cid:13)(cid:13) ≤ ε# ≤(cid:16)1 − 2−O(ε√m)(cid:17)X ≤ exp(cid:16)−X2−O(ε√m)(cid:17) . Now since 0 ∈ X, it follows that to preserve all pairwise distances amongst vectors in X to within (1± ε), we also have to preserve all norms to within ±ε. This is true since for all x of unit norm: 1 √m (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:13)(cid:13)(cid:13)(cid:13) This proves the following: 2 2 − 1(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) T Dx(cid:13)(cid:13)(cid:13)(cid:13) =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:13)(cid:13)(cid:13)(cid:13) 1 √m T Dx − 1 √m T D0(cid:13)(cid:13)(cid:13)(cid:13) 2 2 − kx − 0k2 . 2(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Theorem 10. Let T be m×n Toeplitz and D n×n diagonal. If t−(m−1), t−(m−2), . . . , tn−1 and d1, . . . , dn are independently distributed Rademacher random variables for i = −(m− 1), . . . , n− 1 and j = 1, . . . , n, then for all 0 < ε < C, where C is a universal constant, there exists a set of N = Ω(n/m) vectors X ⊂ Rn such that 1 √m T Dx − 1 √m 2 − kx − yk2 ≤ εkx − yk2 2# ≤ exp(cid:16)−N 2−O(ε√m)(cid:17) . It follows from Theorem 10 that if we want to have constant probability of successfully embedding any set of N vectors, then either it must be the case that m = Ω(n/N ), or Pr"∀x, y ∈ X :(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:13)(cid:13)(cid:13)(cid:13) 2 T Dy(cid:13)(cid:13)(cid:13)(cid:13) 2(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) where C0 is a constant. This in turn implies that −N 2−O(ε√m) ≥ −C0, lg N − O(ε√m) ≤ lg C0 =⇒ √m = Ω(ε−1 lg N ) =⇒ m = Ω(ε−2 lg2 N ). This completes the proof of Theorem 3. 11 References [1] Dimitris Achlioptas. Database-friendly random projections: binary coins. doi:10.1016/S0022-0000(03)00025-4. Journal of computer and System Sciences, Johnson–Lindenstrauss with 66(4):671–687, June 2003. [2] Nir Ailon and Bernard Chazelle. The fast Johnson–Lindenstrauss transform and approximate nearest neighbors. SIAM Journal on Computing, 39(1):302–322, May 2009. doi:10.1137/060673096. [3] Nir Ailon and Edo Liberty. Fast dimension reduction using rademacher series on dual BCH codes. Discrete & Computational Geometry, 42(4):615–630, 2009. doi:10.1007/s00454-008-9110-x. [4] Nir Ailon and Edo Liberty. An almost optimal unrestricted fast Johnson–Lindenstrauss transform. ACM Trans. Algorithms, 9(3):21:1–21:12, June 2013. doi:10.1145/2483699.2483701. [5] Jeremiah Blocki, Avrim Blum, Anupam Datta, and Or Sheffet. The Johnson–Lindenstrauss trans- form itself preserves differential privacy. In Proceedings of the 2012 IEEE 53rd Annual Symposium on Foundations of Computer Science, FOCS '12, pages 410–419, Washington, DC, USA, 2012. IEEE Computer Society. doi:10.1109/FOCS.2012.67. [6] C. Boutsidis, A. Zouzias, M. W. Mahoney, and P. Drineas. Randomized dimensionality reduction for k-means clustering. IEEE Transactions on Information Theory, 61(2):1045–1062, February 2015. doi:10.1109/TIT.2014.2375327. [7] E. J. Candes, J. Romberg, and T. Tao. Robust uncertainty principles: exact signal reconstruc- IEEE Transactions on Information Theory, tion from highly incomplete frequency information. 52(2):489–509, February 2006. doi:10.1109/TIT.2005.862083. [8] Michael B. Cohen, Sam Elder, Cameron Musco, Christopher Musco, and Madalina Persu. Di- mensionality reduction for k-means clustering and low rank approximation. In Proceedings of the Forty-seventh Annual ACM Symposium on Theory of Computing, STOC '15, pages 163–172, New York, NY, USA, 2015. ACM. doi:10.1145/2746539.2746569. [9] Anirban Dasgupta, Ravi Kumar, and Tamás Sarlos. A sparse Johnson–Lindenstrauss transform. In Proceedings of the Forty-second ACM Symposium on Theory of Computing, STOC '10, pages 341–350, New York, NY, USA, 2010. ACM. doi:10.1145/1806689.1806737. [10] Sanjoy Dasgupta and Anupam Gupta. An elementary proof of a theorem of Johnson and Linden- strauss. Random Struct. Algorithms, 22(1):60–65, 2003. doi:10.1002/rsa.10073. [11] D. L. Donoho. Compressed sensing. IEEE Transactions on Information Theory, 52(4):1289–1306, April 2006. doi:10.1109/TIT.2006.871582. [12] Sariel Har-Peled, Piotr Indyk, and Rajeev Motwani. Approximate nearest neighbor: To- Theory of Computing, 8(14):321–350, 2012. wards removing the curse of dimensionality. doi:10.4086/toc.2012.v008a014. [13] Aicke Hinrichs and Jan Vybíral. Johnson–Lindenstrauss lemma for circulant matrices. Random Structures & Algorithms, 39(3):391–398, 2011. doi:10.1002/rsa.20360. [14] Piotr Indyk. Algorithmic applications of low-distortion geometric embeddings. In Proceedings of the 42nd IEEE Symposium on Foundations of Computer Science, FOCS '01, pages 10–33, Washington, DC, USA, 2001. IEEE Computer Society. doi:10.1109/SFCS.2001.959878. [15] William B Johnson and Joram Lindenstrauss. Extensions of lipschitz mappings into a hilbert space. Contemporary mathematics, 26:189–206, 1984. doi:10.1090/conm/026/737400. [16] Daniel M. Kane and Jelani Nelson. Sparser Johnson–Lindenstrauss transforms. J. ACM, 61(1):4:1– 4:23, January 2014. doi:10.1145/2559902. [17] Felix Krahmer and Rachel Ward. New and improved Johnson–Lindenstrauss embeddings via the Restricted Isometry Property. SIAM J. Math. Anal., 43(3):1269–1281, 2011. 12 [18] Kasper Green Larsen and Jelani Nelson. Optimality of the Johnson–Lindenstrauss lemma. In Proceedings of the 2017 IEEE 58th Annual Symposium on Foundations of Computer Science, FOCS '17, Washington, DC, USA, October 2017. IEEE Computer Society. [19] S. Muthukrishnan. Data Streams: Algorithms and Applications, volume 1(2) of Foundations and TrendsTM in Theoretical Computer Science. now Publishers Inc., Hanover, MA, USA, January 2005. doi:10.1561/0400000002. [20] Daniel A. Spielman and Nikhil Srivastava. Graph sparsification by effective resistances. SIAM J. Comput., 40(6):1913–1926, December 2011. doi:10.1137/080734029. [21] Santosh S. Vempala. The random projection method, volume 65 of DIMACS - Series in Discrete Mathematics and Theoretical Computer Science. American Mathematical Society, Providence, RI, USA, September 2004. doi:10.1007/978-1-4615-0013-1_16. [22] Ky Vu, Pierre-Louis Poirion, and Leo Liberti. Using the Johnson–Lindenstrauss lemma in linear and integer programming. ArXiv e-prints, July 2015. arXiv:1507.00990. [23] Jan Vybíral. A variant of the Johnson–Lindenstrauss lemma for circulant matrices. Journal of Functional Analysis, 260(4):1096–1105, 2011. doi:10.1016/j.jfa.2010.11.014. [24] Kilian Weinberger, Anirban Dasgupta, John Langford, Alex Smola, and Josh Attenberg. Fea- ture hashing for large scale multitask learning. In Proceedings of the 26th Annual International Conference on Machine Learning, ICML '09, pages 1113–1120, New York, NY, USA, 2009. ACM. doi:10.1145/1553374.1553516. [25] David P. Woodruff. Sketching as a Tool for Numerical Linear Algebra, volume 10(1–2) of Foundations and TrendsTM in Theoretical Computer Science. now Publishers Inc., Hanover, MA, USA, 2014. doi:10.1561/0400000060. 13
1312.7409	3	1312	2015-03-08T05:58:46	Fredholm multiplication conditional type operators on L^p-spaces	[ "math.FA" ]	In this paper, first we investigate closed range multiplication conditional type operators between two Lp-spaces. Then we characterize Fredholm ones when the underlying measure space is non-atomic. Finally we give some examples.	math.FA	math	FREDHOLM MULTIPLICATION CONDITIONAL TYPE OPERATORS ON Lp-SPACE Y. ESTAREMI Abstract. In this paper, first we investigate closed range multiplication con- ditional type operators between two Lp-spaces. Then we characterize Fred- holm ones when the underlying measure space is non-atomic. Finally we give some examples. 1. Introduction and Preliminaries Let (X, Σ, µ) be a complete σ-finite measure space. For any sub-σ-finite algebra A ⊆ Σ with 1 ≤ p ≤ ∞, the Lp-space Lp(X, A, µA) is abbreviated by Lp(A), and its norm is denoted by k.kp. All comparisons between two functions or two sets are to be interpreted as holding up to a µ-null set. The support of a measurable function f is defined as S(f ) = {x ∈ X; f (x) 6= 0}. We denote the vector space of all equivalence classes of almost everywhere finite valued measurable functions on X by L0(Σ). For a sub-σ-finite algebra A ⊆ Σ, the conditional expectation operator associated with A is the mapping f → EAf , defined for all non-negative function f as well as for all f ∈ Lp(Σ), 1 ≤ p ≤ ∞, where EAf , by the Radon-Nikodym theorem, is the unique A-measurable function satisfying ZA f dµ = ZA EAf dµ, ∀A ∈ A. As an operator on Lp(Σ), EA is an idempotent and EA(Lp(Σ)) = Lp(A). If there is no possibility of confusion we write E(f ) in place of EA(f ). This operator will play a major role in our work and we list here some of its useful properties: • If g is A-measurable, then E(f g) = E(f )g. • E(f )p ≤ E(f p). • If f ≥ 0, then E(f ) ≥ 0; if f > 0, then E(f ) > 0. • E(f g) ≤ E(f p) • For each f ≥ 0, σ(f ) ⊆ σ(E(f )). p′ , where 1 p E(gp′ p + 1 p′ = 1 (Holder inequality). 1 1 ) A detailed discussion and verification of most of these properties may be found in [11]. We recall that an A-atom of the measure µ is an element A ∈ A with µ(A) > 0 such that for each F ∈ A, if F ⊆ A then either µ(F ) = 0 or µ(F ) = µ(A). A measure space (X, Σ, µ) with no atoms is called non-atomic measure space. It is well-known fact that every σ-finite measure space (X, A, µA ) can be partitioned 2010 Mathematics Subject Classification. 47B38, 47B37. Key words and phrases. Conditional expectation, multiplication operators, Fredholm, closed range. 1 2 Y. ESTAREMI uniquely as X = (cid:0)Sn∈N An(cid:1)∪B, where {An}n∈N is a countable collection of pairwise disjoint A-atoms and B, being disjoint from each An, is non-atomic (see [14]). Compositions of conditional expectation operators and multiplication operators appear often in the study of other operators such as multiplication operators and weighted composition operators. Specifically, in [10], S.-T. C. Moy characterized all operators on Lp of the form f → E(f g) for g in Lq with E(g) bounded. Eleven years later, R. G. Douglas, [5], analyzed positive projections on L1 and many of his characterizations are in terms of combinations of multiplications and conditional expectations. More recently, P.G. Dodds, C.B. Huijsmans and B. De Pagter, [1], extended these characterizations to the setting of function ideals and vector lat- tices. J. Herron presented some assertions about the operator EMu on Lp spaces in [7]. Also, some results about multiplication conditional type operators can be found in [6, 8]. In [2, 3, 4] we investigated some classic properties of multiplication conditional type operators MwEMu on Lp spaces. Let f ∈ L0(Σ), then f is said to be conditionable with respect to E if f ∈ D(E) := {g ∈ L0(Σ) : E(g) ∈ L0(A)}. Throughout this paper we take u and w in D(E). In this paper, some necessary and sufficient conditions for closeness of range of multiplication conditional type operators between two Lp-spaces are given. Also, Fredholm ones are characterized when the underlying measure space is non-atomic. Some results of this paper are a generalization of some results of [13]. Now we give a definition of multiplication conditional type operators on Lp- spaces. Definition 1.1. Let (X, Σ, µ) be a σ-finite measure space and let A be a σ- subalgebra of Σ such that (X, A, µA) is also σ-finite. Let E be the corresponding conditional expectation operator relative to A. If 1 ≤ p, q < ∞ and u, w ∈ L0(Σ) (the spaces of Σ-measurable functions on X) such that uf is conditionable and wE(uf ) ∈ Lq(Σ) for all f ∈ D ⊆ Lp(Σ), where D is a linear subspace, then the corresponding multiplication conditional type operator is the linear transformation MwEMu : D → Lq(Σ) defined by f → wE(uf ). Here we recall some results of [1] that state our results is valid for a large class of linear operators. Let (X, Σ, µ) be a finite measure space, then L∞(Σ) ⊆ Lp(Σ) ⊆ L1(Σ) and Lp(Σ) is an order ideal of measurable functions on (X, Σ, µ). Thus by propositions (3.1, 3.3, 3.6) of [1] we have theorems A, B, C: Theorem A. If T is a linear operator on Lp(Σ) for which (i) T f ∈ L∞(Σ) whenever f ∈ L∞(Σ). (ii) kT fnk1 → 0 for all sequences {fn}∞ n=1 ⊆ Lp(Σ) such that fn ≤ g (n = 1, 2, 3, ....) for some g ∈ Lp(Σ) and fn → 0 a.e., (iii) T (f.T g) = T f.T g for all f ∈ L∞(Σ) and all g ∈ Lp(Σ), MULTIPLICATION CONDITIONAL TYPE OPERATORS 3 then there exists a σ-subalgebra A of Σ and there exists w ∈ Lp′ (Σ) such that T f = EA(wf ) for all f ∈ Lp(Σ).(p−1 + p′−1 = 1) Theorem B. For a linear operator T : Lp(Σ) → Lp(Σ) the following statement are equivalent. (i) T is positive and order continuous, T 2 = T , T 1 = 1 and the range R(T ) of T is a sublattice. (ii) There exist a σ-subalgebra A of Σ and a function 0 ≤ w ∈ Lp′ (Σ) with EA(w) = 1 such that T f = EA(wf ) for all f ∈ Lp(Σ). Theorem C. For a linear operator T : Lp(Σ) → Lp(Σ) the following statement are equivalent. (i) T is a positive and order continuous projection onto a sublattice such that T 1 is strictly positive. (ii) There exist a σ-subalgebra A of Σ, 0 ≤ w ∈ Lp′ (Σ) and a strictly positive function k ∈ L1(Σ) with EA(wk) = 1 such that T f = kEA(wf ) for all f ∈ Lp(Σ). Moreover, if we choose k such that EA(k) = 1, then both w and k are uniquely determined by T . 2. CLOSED RANGE AND FREDHOLM WEIGHTED CONDITIONAL TYPE OPERATORS In this section first we describe closed range multiplication conditional type op- erators MwEMu between two Lp-spaces. Let 1 ≤ p, q < ∞ and f ∈ Lp, then it is easily seen that kMwEMu(f )kq = kEMv(f )kq, 1 where v = u(E(wq)) q . Thus without loss of generality we can consider the opera- tor EMv instead of MwEMu, in our discussion about closedness of range. Also, we recall that for an operator T on a Banach space X, N (T ) = {x ∈ X : T (x) = 0} and R(T ) = {T (x) : x ∈ X} are called null space and range of T , respectively. Theorem 2.1. Let 1 < p < ∞ and let p′ be conjugate component to p. Then (a) If the operator T = EMu from Lp(Σ) into itself is injective and has closed p′ ≥ δ a.e, On S, where range, then there exists δ > 0 such that v = (E(up′ S = {x ∈ X : v(x) 6= 0}. 1 )) 4 Y. ESTAREMI (b) If σ(E(u)) = σ(E(up′ S = σ(E(up′ closed range on Lp(Σ). )) = {x ∈ X : E(up′ )) and there exists δ > 0 such that E(u) ≥ δ a.e. On )(x) 6= 0}, then the operator T = EMu has Proof. (a) Let f ∈ Lp(Σ). Then kT f kp p = ZX ≤ ZX = ZX E(uf )pdµ (E(up′ )) p p′ E(f p)dµ vpf pdµ = kMvf kp p. Since T is injective and closed range, then there exists δ > 0 such that for f ∈ Lp(Σ), kT f kp ≥ δkf kp. Thus kMvf kLp(S) = kMvf kLp(X) ≥ kT f kp ≥ δkf kLp(X) ≥ δkf kLp(S) and so kMvf kLp(S) ≥ δkf kLp(S), for all f ∈ Lp(Σ). This mean's that Mv has closed range on Lp(X). Thus there exists β > 0 such that v ≥ β a.e. On S. (b) Let fn, g ∈ Lp(Σ) such that kE(ufn) − gkp → 0, when n → ∞. Since E(u) χS ∈ Lp(S) δ a.e. On S. This implies that E(u) ≤ 1 1 g E(u) ≥ δ a.e. On S, then and E(u g g χS)kp kE(ufn) − E(u E(u) χS) = g ∈ Lp(X, A, µ). Hence p = ZX = ZS = ZS E(u) E(ufn) − E(u g E(u) )χSpdµ E(ufn) − E(u g E(u) )χSpdµ E(ufn) − gpdµ ≤ kE(ufn) − gkp p → 0, when n → ∞. So the operator EMu has closed range on Lp(Σ). Theorem 2.2. Let 1 < q < p < ∞ and let p′, q′ be conjugate component to p and q respectively. Then (a) If the operator T = EMu from Lp(Σ) into Lq(Σ) is injective and has closed range, then (1) v = 0 a.e. On B and the set {n ∈ N : v(An) 6= 0} is finite. (2) Mv from Lp(Σ) into Lq(Σ) has finite rank. Where v = (Euq′ ) 1 q′ and S = {x ∈ X : v(x) 6= 0}. (b) Let MULTIPLICATION CONDITIONAL TYPE OPERATORS 5 (1) The operator T = EMu has closed range. (2) The operator T = EMu has finite rank. (3) v = 0 a.e. On B and the set Nv = {n ∈ N : v(An) 6= 0} is finite. (4) E(u) = 0 a.e. On B and the set NN (u) = {n ∈ N : E(u)(An) 6= 0} is finite. Then (3) → (2) → (1) → (4). Proof. (a) Let f ∈ Lp(Σ). Then kT f kq q = ZX ≤ ZX = ZX E(uf )qdµ (E(uq′ )) q q′ E(f q)dµ vqf qdµ = kMvf kq q. Since T is injective and closed range, then there exists δ > 0 such that for f ∈ Lp(Σ), kT f kq ≥ δkf kp. Thus kMvf kq ≥ kT f kp ≥ δkf kp and so kMvf kq ≥ δkf kp, for all f ∈ Lp(Σ). This mean's that Mv from Lp(Σ) into Lq(Σ) has closed range. Thus by [13] we have v = 0 a.e. On B and the set {n ∈ N : v(An) 6= 0} is finite. Also, Mv from Lp(Σ) into Lq(Σ) has finite rank. (b) If µ(S) = 0 then EMu is the zero operator. So we assume that µ(S) > 0. i=1Ani for some integer k > 0. (3) → (2). If (3) holds, then S = ∪n∈Nv An = ∪k Hence EMu(Lp(X, Σ, µ)) ⊆ Lp(S, A, µ), Since for any f ∈ Lp(X, Σ, µ), σ(E(uf )) ⊆ S. This implies that EMu has finite rank. (2) → (1) is trivial. (1) → (4). Suppose that EMu has closed range.First we show that E(u) = 0 a.e. On B. Assume on the contrary that µ({x ∈ B : E(u)(x) 6= 0}) > 0. Then we have µ({x ∈ B : E(u)(x) > δ}) > 0 for some δ > 0. Set G = {x ∈ B : E(u)(x) > δ} and E(u)(x) for x ∈ G. For every f ∈ Lp(G) and define a function v on G by v(x) = g ∈ Lq(G), 1 MvEMu(f ) = v(E(u)G)f = f, ME(u)GMv(g) = g. Thus Mv is the inverse operator of EMuLp(G) = ME(u)G and EMuLp(G) = ME(u)G is a bounded operator from Lp(G) into Lq(G) that has closed rang. For any E ∈ AG = {A ∩ G : A ∈ A} with µ(E) < ∞. put f = E(u)(x) χE(x), then f ∈ Lp(G). Moreover, ME(u)f = χE and so χE ∈ ME(u)(Lp(G)). Hence, the the range ME(u)(Lp(G)) contains all linear combinations of such χE's. Thus 1 6 Y. ESTAREMI ME(u)(Lp(G)) = Lq(G), since ME(u) has closed range and all linear combinations of such χE's are dense in Lq(G). This implies that Mv maps Lq(G) into Lp(G), that is Mv is bounded from Lq(G) into Lp(G). Since G is non-atomic by theorem 1.4 of [13] we have v = 0 a.e on G. But this is a contradiction. 1 E(u)(x) for x ∈ S and consider the operator Mw. put f = Now, we show that NE(u) is finite. Since µ(S) > 0, it follows that NE(u) 6= ∅. Put E(u)(x) χAn(x), w(x) = then f ∈ Lp(G), by the same method of preceding paragraph we see that Mw maps Lq(S) into Lp(S) that is Mw is bounded from Lq(S) into Lp(S). So by theorem 1.4 of [13] we have 1 b = sup n∈NE(u) 1 E(u)(An)rµ(An) = sup n∈NE(u) w(An)r µ(An) < ∞, p + 1 r = 1 where 1 Theorem 1.3 of [13] says E(u) ∈ Lr(X, A, µ). So, q . Since NE(u) 6= ∅, then b > 0 and E(u)(An)rµ(An) ≥ 1 b . While Σn∈NE(u) 1 b ≤ kE(u)kr r < ∞. This implies that NE(u) is finite. Theorem 2.3. Let 1 < p < q < ∞ and let p′, q′ be conjugate component to p and q respectively. (a) If the operator T = EMu from Lp(Σ) into Lq(Σ) is injective and has closed range, then (1) The set {n ∈ N : v(An) 6= 0} is finite. (2) Mv from Lp(Σ) into Lq(Σ) has finite rank. Where v = (Euq′ ) 1 q′ and S = {x ∈ X : v(x) 6= 0}. (b) Let (1) The operator T = EMu has closed range. (2) The operator T = EMu has finite rank. (3) The set Nv = {n ∈ N : v(An) 6= 0} is finite. (4) The set NN (u) = {n ∈ N : E(u)(An) 6= 0} is finite. Then (3) → (2) → (1) → (4). Proof. (a) Let f ∈ Lp(Σ). Then kT f kq q ≤ kMvf kq q. MULTIPLICATION CONDITIONAL TYPE OPERATORS 7 Since T is injective and closed range, then there exists δ > 0 such that for f ∈ Lp(Σ), kT f kq ≥ δkf kp. Thus kMvf kq ≥ kT f kp ≥ δkf kp and so kMvf kq ≥ δkf kp, for all f ∈ Lp(Σ). This mean's that Mv from Lp(Σ) into Lq(Σ) has closed range. Thus by [13] the set {n ∈ N : v(An) 6= 0} is finite. Also, Mv from Lp(Σ) into Lq(Σ) has finite rank. (b) Theorem 2.3 of [4] tells us that v = 0 a.e on B and so E(u) = 0 a.e on B. By the same method that we used in last theorem, it is easy to see that (3) → (2) → (1). Now, we show that (1) → (4). Suppose that NN (u) 6= ∅. If we put S = σ(E(u)), then we can write S = ∪n∈NE(u) An. Define a function w on S by w(x) = E(u)(x) for x ∈ S. By the same method that is used in the proof of last Lp(S). Hence by Theorem 1.3 of [13] we have w ∈ Ls(A), where 1 p . While Theorem 1.4 of µ(An) < ∞. Since NN (u) 6= ∅ implies b > 0 and [13] says that b = supn∈NE(u) since w(An)sµ(An) = µ(An) b for all n ∈ NN (u), it follows that E(u)(An)s ≥ 1 1 E(u)(An)s q + 1 s = 1 Σn∈NE(u) 1 b ≤ kwks s < ∞ this implies that NE(u) is finite. In the sequel we consider the function u is A-measurable. In this case E(u) = u and EMu = Mu Lp(A). Therefore we get the following results. Corollary 2.4. Let 1 < p < ∞ and let p′ be conjugate component to p. If u ∈ L0(A) and T = EMu is injective on Lp(Σ). Then the operator T = EMu has closed range if and only if there exists δ > 0 such that u ≥ δ a.e. On S. Where S = {x ∈ X : u(x) 6= 0}. Corollary 2.5. Let 1 < q < p < ∞ and let p′, q′ be conjugate component to p and q respectively. If u ∈ L0(A) and T = EMu is injective from Lp(Σ) into Lq(Σ). Then the followings are equivalent: (1) The operator Mu has closed range. (2) The operator Mu has finite rank. (3) u = 0 a.e. On B and the set Nu = {n ∈ N : u(An) 6= 0} is finite. Corollary 2.6. Let 1 < p < q < ∞ and let p′, q′ be conjugate component to p and q respectively. If u ∈ L0(A) and T = EMu is injective from Lp(Σ) into Lq(Σ), then the followings are equivalent: (1) The operator EMu has closed range. (2) The operator EMu has finite rank. (3) The set Nu = {n ∈ N : u(An) 6= 0} is finite. 8 Y. ESTAREMI If the multiplication conditional type operator EMu is bounded from Lp(Σ) onto Lp(A), then µ(Z(E(up′ )(x) = 0}) = 0. Suppose that F ⊆ Z(E(up′ )) with F ∈ A and µ(F ) < ∞. Then χF ∈ Lp(A) = R(T ) and there exists f ∈ Lp(Σ) such that T f = χF . So by conditional-type Holder inequality we have )) = {x ∈ X : E(up′ µ(F ) = ZF ≤ ZF E(u.f )pdµ (E(up′ ) p p′ .f pdµ = 0. Hence µ(Z(E(up′ ))) = 0. In the next theorem we characterize Fredholm multiplication conditional type operators on Lp-spaces when the underlying measure space is non-atomic. Theorem 2.7. Let 1 < p < ∞ and (X, Σ, µ) be a non-atomic measure space. If T = EMu is a bounded operator from Lp(Σ) into Lp(A), then T is Fredholm if and only if T is invertible. Proof. Let T be Fredholm, then T has closed range. First we show that T is surjective. Suppose on the contrary. Let f0 ∈ Lp(A) \ R(T ). Then there exists a bounded linear functional Lg0 on Lp(A) for some g0 ∈ Lp′ (A) (p−1 + p′−1 = 1), which is defined as Lg0(f ) = ZX f ¯g0dµ, f ∈ Lp(A), such that Then there exists a positive constant δ such that Lg0 (f0) = 1, Lg0(R(T )) = 0. µ(Eδ = {x ∈ X : f0(x)g0(x) > δ}) > 0. Since the underlying measure space is non-atomic, then we can find a disjoint sequences {En}n such that En ⊆ Eδ and 0 < µ(En) < ∞. Let gn = g0.χEn. Clearly gn ∈ Lp′ (A) and for every f ∈ Lp(A) we have ZX ¯f T ∗(gn)dµ = ZX = ZX ¯f ¯uE(gn)dµ g0E(¯u ¯f .χEn)dµ = Lg0 (T f ) = 0. This implies that gn ∈ N (T ∗) and so N (T ∗) is infinite dimensional. Therefore the codimension of R(T ) is not finite. This is a contradiction, therefore T is surjective. Let 0 6= f ∈ Lp(Σ) such that T (f ) = 0. Since S(f ) ⊆ S(E(f )), µ(S(f )) > 0 and S(E(f )) is A-measurable, then there exists A ∈ A with 0 < µ(A) < ∞ and µ(S(f ) ∩ A) > 0. Also since the underlying measure space is non-atomic, we can find a disjoint sequence of A-measurable subsets {An}n of A with µ(S(f )∩An) > 0. Clearly fn = f.χS(f )∩An ∈ Lp(Σ) and T (fn) = χAn .T (f ) = 0. This means N (T ) is infinite dimensional, that is a contradiction. Thus T should be injective and so is invertible. The converse is obvious. MULTIPLICATION CONDITIONAL TYPE OPERATORS 9 In the sequel we present some examples of conditional expectations and corre- sponding multiplication conditional type operators. Examples. (a) Let (X1, Σ1, µ1) and (X2, Σ2, µ2) be two σ-finite measure spaces and X = X1 × X2, Σ = Σ1 × Σ2 and µ = µ1 × µ2. Put A = {A × X2 : A ∈ Σ1}. Then A is a sub-σ-algebra of Σ and for all f ∈ D(EA) we have EA(f )(x1, x2) = ZX2 f (x1, y)dµ2(y) µ − a.e. on X. Mpreover, if (X, Σ, µ) is a finite measure space and k : X × X → C is a Σ ⊗ Σ- measurable function such that the operator T : Lp(Σ) → Lq(Σ) defined by T f (x) = ZX k(x, y)f (y)dµ, f ∈ Lp(Σ), is bounded. Then T is a weighted conditional type operator. Since kf is a Σ ⊗ Σ- measurable function, when f ∈ Lp(Σ). Then by taking u := k and f ′(x, y) = f (y), we get that EA(uf )(x, y) = EA(uf ′)(x, y) = ZX = ZX u(x, y)f ′(x, y)dµ(y) u(x, y)f (y)dµ(y) = T f (x). Hence T is a weighted conditional type operator. operator In particular the convolution w ∗ u(y) = ZG w(y − x)u(x)dm(x) on an abelian group G with Haar measure m, can be viewed as an integral operator. (b) This example deals with a weighted conditional expectation operator which is in the form of an integral operator. Let X = (0, 1] × (0, 1], dµ = dxdy, Σ the Lebesgue measurable subsets of X and let A be the σ-algebra generated by the family of the sets of the form A × (0, 1] where A is a Lebesgue measurable subset of (0, 1]. For f in L2((0, 1]2), we have (EAf )(x, y) = R 1 0 f (x, t)dt. Let X = (0, ∞) × (0, ∞) and A be the σ-algebra generated by the family of the sets of the form A × (0, ∞) where A is a Lebesgue measurable subset of (0, ∞). Then T f (x, y) = (EAMu(f ))(x, y) = Z ∞ 0 u(x, t)f (x, t)dt. 10 Y. ESTAREMI Hence for every function f : (0, ∞) → R we can define the function f ′ on X as f ′(x, y) = f (y). So, T f (x, y) = (EAMu(f ))(x, y) = (EAMu(f ′))(x, y) = Z ∞ 0 u(x, t)f (t)dt = T ′(f )(x). This implies that T is an integral transform and specially by taking u(x, y) = e−xy we obtain one of the most important classical integral transforms that is widely used in analysis, namely the Laplace integral transform. We refer to [9] for some applications of integral transforms, especially Laplace integral transforms. References [1] P.G. Dodds, C.B. Huijsmans and B. De Pagter, characterizations of conditional expectation- type operators, Pacific J. Math. 141(1) (1990), 55-77. [2] Y. Estaremi, Essential norm of weighted conditional type operators on Lp-spaces, positivity. 18 (2014), 41-52. [3] Y. Estaremi and M.R. Jabbarzadeh, Compact weighted Lambert type operators on Lp-spaces, J. Math. Anal. Appl. 420 (2014)118-123. [4] Y. Estaremi and M.R. Jabbarzadeh, Weighted lambert type operators on Lp-spaces, Oper. Matrices 1 (2013), 101-116. [5] R. G. Douglas, Contractive projections on an L1 space, Pacific J. Math. 15 (1965), 443-462. [6] J. J. Grobler and B. de Pagter, Operators representable as multiplication-conditional expec- tation operators, J. Operator Theory 48 (2002), 15-40. [7] J. Herron, Weighted conditional expectation operators, Oper. Matrices 1 (2011), 107-118. [8] A. Lambert, Lp multipliers and nested sigma-algebras, Oper. Theory Adv. Appl. 104 (1998), 147-153. [9] Y. Luchko, Integral transforms of the Mellin convolution type and their generating operators, Integral Trans. and special Func. 11 (2008), 809-851. [10] Shu-Teh Chen, Moy, Characterizations of conditional expectation as a transformation on function spaces, Pacific J. Math. 4 (1954), 47-63 [11] M. M. Rao, Conditional measure and applications, Marcel Dekker, New York, 1993. [12] J. H. Shapiro, The essential norm of a composition operator, Analls of Math. 125 (1987), 375-404. [13] H. Takagi and K. Yokouchi, Multiplication and composition operators between two Lp-spaces, Contemporary Math. 232(1999), 321-338. [14] A. C. Zaanen, Integration, 2nd ed., North-Holland, Amsterdam, 1967. E-mail address: [email protected] Department of mathematics, Payame Noor university , P. O. Box: 19395-3697, Tehran, Iran
1201.1308	1	1201	2012-01-05T21:27:07	On representing and absolutely representing systems of subspaces in Banach spaces	[ "math.FA" ]	We study properties of representing and absolutely representing systems of subspaces in Banach spaces. We also present sufficient conditions for the system of subspaces to be a representing system of subspaces.	math.FA	math	УДК 517.982.22 О ПРЕДСТАВЛЯЮЩИХ И АБСОЛЮТНО ПРЕДСТАВЛЯЮЩИХ СИСТЕМАХ ПОДПРОСТРАНСТВ В БАНАХОВЫХ ПРОСТРАНСТВАХ И.С.Фещенко Аннотация. В работе изучаются свойства представляющих систем подпространств и абсолютно представляющих систем подпространств в банаховых пространствах. Ключевые слова: банахово пространство, представляющая система подпространств, абсолютно представляющая система подпространств. 1. Введение Пусть X (cid:22) линейное нормированное пространство над полем K действительных или комплексных чисел, Xk, k > 1, (cid:22) система подпространств X (т.е. замкнутых линейных множеств), которую мы будем обозначать S = (X; Xk, k > 1). Определение 1.1. ([4]) Система S называется представляющей системой подпро- странств (ПСП) в X, если произвольный x ∈ X можно представить в виде x = k=1 xk, где xk ∈ Xk, k > 1. Отметим, что в случае банахова пространства X в книге [12] (определение 15.21) P∞ ПСП в X называется псевдоразложением X. Определение 1.2. ([4]) Система S называется абсолютно представляющей систе- мой подпространств в X, если произвольный x ∈ X можно представить в виде x = P∞ k=1 xk, где xk ∈ Xk, k > 1 иP∞ k=1 kxkk < ∞. ПСП и АПСП являются естественным обобщением представляющих и абсолютно представляющих систем (все Xk одномерны) (см., например, [3]). Определения ПСП и АПСП можно давать и в более широких классах пространств, чем линейные нормиро- ванные (см.[4]); ПСП и АПСП в различных классах пространств изучались в работах Ю.Ф. Коробейника, А.В. Абанина, К.А.Михайлова и др. (см., например, [4],[1],[6]). В данной работе мы изучаем ПСП и АПСП в банаховых пространствах. В параграфе 2 изучаются свойства ПСП в X, приведены достаточные условия для того, чтобы счётная система подпространств была ПСП в X, доказана теорема об устойчивости ПСП, а также показана связь между ПСП в гильбертовом пространстве и проблемой Гальперина. В параграфе 3 приводятся различные критерии АПСП, изучаются свойства АПСП в X. 1 2 1 0 2 n a J 5 ] . A F h t a m [ 1 v 8 0 3 1 . 1 0 2 1 : v i X r a 2 И.С.Фещенко 2. Представляющие системы подпространств в банаховых пространствах 2.1. Некоторые свойства ПСП в банаховых пространствах. Пусть X (cid:22) бана- хово пространство, S = (X; Xk, k > 1) (cid:22) система его подпространств. Введём необхо- димые обозначения и определения. Для непустого I ⊂ N определим подпространство S(I) как замыкание линейной оболочки подпространств Xi, i ∈ I. Обозначим X ∗ со- пряженное пространство к X; для ϕ ∈ X ∗ обозначим ϕ(S,I) сужение ϕ на S(I). Ясно, что ϕ(S,I) ∈ (S(I))∗. Для разбиения π = {Ik} (k пробегает конечное или счётное число значений) множества N определим F1(S, π, ϕ) =Xk kϕ(S,Ik)k, ϕ ∈ X ∗. Разбиение π множества N назовём последовательным, если оно имеет один из сле- дующих двух видов: возрастающей последовательности натуральных чисел n(1) < n(2) < . . ., n(0) = 0. (1) Множества Ik = {n(k − 1) + 1, n(k − 1) + 2, . . . , n(k)}, k > 1 для некоторой (2) Множества Ik = {n(k − 1) + 1, . . . , n(k)}, 1 6 k 6 p, Ip+1 = {n(p) + 1, n(p) + 2, . . .} для некоторой возрастающей последовательности натуральных чисел n(1) < . . . < n(p), n(0) = 0. Теорема 2.1. Пусть S (cid:22) ПСП в X. Тогда существует ε > 0 такое, что для произ- вольного последовательного разбиения π множества N выполнено (2.1) F1(S, π, ϕ) > εkϕk, ϕ ∈ X ∗. Доказательство. Определим пространство Dc = {ξ = (x1, x2, . . .) xk ∈ Xk, k > 1, xk сходится} ∞Xk=1 с нормой kξk = supk>1 kx1 + . . . + xkk. Легко проверить, что Dc банахово. Опреде- k=1 xk. Тогда A ограничен (более того, kAk 6 1). Поскольку S является ПСП в X, то Im(A) = X. Из теоремы про открытое отображение следует существование числа M > 0 такого, что для произвольного x ∈ X существует ξ = (x1, x2, . . .) ∈ Dc, для которого x = Aξ k=1 xk и для каждого k > 1 kx1 + . . . + xkk 6 Mkxk. Для натуральных l 6 k определим xl,k = xl + . . . + xk, тогда kxl,kk 6 2Mkxk. Для последовательного разбиения π первого вида и ϕ ∈ X ∗ имеем лим линейный оператор A : Dc → X равенством A(x1, x2, . . .) = P∞ и kξk 6 Mkxk. Тогда x = P∞ ∞Xk=1 ϕ(xn(k−1)+1,n(k)) 6 ϕ(x) = ϕ( xn(k−1)+1,n(k)) 6 ∞Xk=1 6 ∞Xk=1 kϕ(S,Ik)kkxn(k−1)+1,n(k)k 6 2MkxkF1(S, π, ϕ), откуда, в силу произвольности x ∈ X, F1(S, π, ϕ) > (1/(2M))kϕk. Для последователь- ного разбиения π второго вида такая же оценка доказывается аналогично. (cid:3) 3 Для подмножества M ⊂ X обозначим M ⊥ множество всех ϕ ∈ X ∗ таких, что ϕ(x) = 0, x ∈ M. Напомним (см., например, параграф 15 книги [12]), что система под- пространств Gk, k > 1, банахова пространства E называется разложением Шаудера E, если для каждого x ∈ E существуют и единственны xk ∈ Gk, k > 1, такие, что k=1 xk. x =P∞ Теорема 2.2. Пусть S (cid:22) ПСП в X. Тогда система подпространств X ′ 1 является разложением Шаудера в замыкании своей линейной оболочки. k =Tj6=k X ⊥ j , k > kϕ + ψk > εkϕk (см. теорему 15.5 в [12]). Доказательство. Достаточно доказать, что существует ε > 0, такое, что для произ- k выполнено вольных натуральных n, m и произвольных ϕ ∈Pn Из теоремы 2.1 следует, что существует ε > 0 такое, что для произвольного по- следовательного разбиения π выполнено неравенство (2.1). Пусть ϕ ∈Pn k, ψ ∈ Pn+m k. Определим разбиение π так: I1 = {1, 2, . . . , n}, I2 = {n + 1, n + 2, . . .}. Тогда kϕ + ψk > k(ϕ + ψ)(S,I1)k = kϕ(S,I1)k = kϕ(S,I1)k + kϕ(S,I2)k > εkϕk, k=n+1 X ′ k, ψ ∈Pn+m k=n+1 X ′ k=1 X ′ k=1 X ′ откуда следует нужное утверждение. (cid:3) Будем говорить, что система подпространств S является перестановочной ПСП (ППСП) в X, если для произвольной биекции σ : N → N система подпространств Sσ = (X; Xσ(k), k > 1) является ПСП в X. Напомним (см., например, с.534 в [12]), что разложение Шаудера Gk, k > 1 банахова пространства E называется безусловным, если каждый сходящийся ряд видаP∞ Теорема 2.3. Пусть S (cid:22) ППСП в X. Тогда система подпространств X ′ 1 является безусловным разложением Шаудера в замыкании своей линейной оболоч- ки. j , k > k =Tj6=k X ⊥ k=1 xk, xk ∈ Gk, k > 1, сходится безусловно. Доказательство. Воспользуемся следующим утверждением (см. теорему 15.18 в [12]): если Gk, k > 1, (cid:22) система подпространств банахова пространства E, причём замыка- ние линейной оболочки Gk, k > 1, равно E, то Gk, k > 1, является безусловным разло- жением Шаудера E тогда и только тогда, когда для произвольной биекции σ : N → N система подпространств Gσ(k), k > 1, является разложением Шаудера E. Теперь из теоремы 2.2 следует нужное утверждение. (cid:3) 2.2. Достаточное условие для того, чтобы система подпространств была ПСП в X. Пусть X (cid:22) линейное нормированное пространство, S = (X; Xk, k > 1) (cid:22) система его подпространств. Для множества F ⊂ X и элемента x ∈ X обозначим d(x, F ) расстояние от x до F . Определим множество ∆(S) = {x ∈ X lim inf k→∞ d(x, Xk) = 0}. 4 И.С.Фещенко Теорема 2.4. Если замыкание линейной оболочки ∆(S) равно X, то S является ППСП в X. Доказательство. Очевидно, для произвольной биекции σ : N → N ∆(Sσ) = ∆(S). Поэтому достаточно доказать, что произвольная система подпространств S, удовле- творяющая условию теоремы, является ПСП в X. виде x =P∞ Достаточно доказать, что произвольный x ∈ X,kxk < 1 может быть представлен в k=1 xk, где xk ∈ Xk. Итак, пусть x ∈ X,kxk < 1. Для k = 1, 2, . . . проделаем Пусть для некоторого k > 1 у нас уже определены натуральные числа N(l, i, j) для следующие операции. l = 1, . . . , k − 1, i = 1, . . . , r(l), j = 1, . . . , N(l) и элементы yl,i,j ∈ XN (l,i,j), причём kx − k−1Xl=1 r(l)Xi=1 N (l)Xj=1 yl,i,jk < 2−(k−1) (для k = 1 ничего не определено). Обозначим z = x −Pk−1 j=1 yl,i,j, тогда kzk < 2−(k−1) (для k = 1 определяем z = x). Существует N(k) ∈ N и элементы xk,j ∈ ∆(S), j = 1, . . . , N(k), такие, что i=1PN (l) l=1 Pr(l) (2.2) kz − N (k)Xj=1 xk,jk < 2−k. Выберем r(k) ∈ N так, чтобы (2.3) kxk,j/r(k)k < 2−k(N(k))−1, j = 1, . . . , N(k). Неравенство (2.2) перепишем в виде (2.4) kz −(cid:16) xk,1 + . . . + xk,N (k) r(k) {z Из kzk < 2−(k−1) и (2.2) следует, что kPN (k) вольного a = 1, . . . , r(k) − 1 имеем: (2.5) xk,1 + . . . + xk,N (k) + . . . + r(k) r(k) xk,1 + . . . + xk,N (k) + . . . + r(k) (cid:17)k < 2−k. } j=1 xk,jk < 2−(k−1) + 2−k. Поэтому для произ- xk,1 + . . . + xk,N (k) r(k) k < 2−(k−1) + 2−k a {z } k Ясно, что xk,j/r(k) ∈ ∆(S) для j = 1, . . . , N(k). Используя неравенства (2.3),(2.4),(2.5) несложно видеть, что существуют натуральные числа N(k, i, j), i = 1, . . . , r(k); j = 1, . . . , N(k) и элементы yk,i,j ∈ XN (k,i,j), такие, что N(k, i, j) < N(k, i′, j′) если i < i′; N(k, i, j) < N(k, i, j′) если j < j′; (1) N(k − 1, r(k − 1), N(k − 1)) < N(k, 1, 1) (для k = 1 это условие отсутствует); 5 yl,i,jk < 2−k; (2) kz −Pr(k) i=1PN (k) j=1 yk,i,jk < 2−k, т.е. (3) для произвольного a = 1, . . . , r(k) − 1 r(l)Xi=1 N (l)Xj=1 kXl=1 N (k)Xj=1 kx − k aXi=1 yk,i,jk < 2−(k−1) + 2−k; Из неравенств (2.7),(2.8),(2.9) следует δs < 6 · 2−k → 0, s → ∞. Поэтому справедливо равенство (2.6). Дополняя его в нужных местах нулями, получим искомое разложение (cid:3) (4) kyk,i,jk < 2−k(N(k))−1 для всех i = 1, 2, . . . , r(k), j = 1, 2, . . . , N(k). Действительно, сначала выберем N(k, 1, 1) и y(k, 1, 1) (достаточно близко к xk,1/r(k)), затем N(k, 1, 2) и yk,1,2 (достаточно близко к xk,2/r(k)), . . ., затем N(k, 1, N(k)) и yk,1,N (k) (достаточно близко к xk,N (k)/r(k)), затем переходим к выбору ¾второй груп- пы¿: N(k, 2, 1) и y(k, 2, 1) (достаточно близко к xk,1/r(k)) и т.д. Выполнив такие операции, получим набор элементов yl,i,j ∈ XN (l,i,j), где l = 1, 2, . . ., i = 1, . . . , r(l), j = 1, 2 . . . , N(l). По построению N(l, i, j) < N(l′, i′, j′) если l < l′; N(l, i, j) < N(l, i′, j′) если i < i′; N(l, i, j) < N(l, i, j′) если j < j′. Покажем, что (2.6) x = y1,1,1 + y1,1,2 + . . . + y1,1,N (1) + y1,2,1 + . . . + y1,2,N (1) + . . . + + y1,r(1),1 + . . . + y1,r(1),N (1) + y2,1,1 + . . . + y2,1,N (2) + y2,2,1 + . . . Для этого рассмотрим сумму первых s членов ряда в правой части (2.6). Представим s в виде s = r(1)N(1) + . . . + r(k − 1)N(k − 1) + aN(k) + b, где 0 6 a 6 r(k) − 1, 0 6 b 6 N(k) − 1. Оценим δs = kx − k−1Xl=1 yl,i,j − aXi=1 N (k)Xj=1 yk,i,j − bXj=1 yk,a+1,jk. Из построения yl,i,j следуют оценки N (l)Xj=1 r(l)Xi=1 r(l)Xi=1 k−1Xl=1 N (k)Xj=1 aXi=1 kx − N (l)Xj=1 yl,i,jk < 2−(k−1), k yk,i,jk < 2−(k−1) + 2−k, yk,a+1,jk < N(k)2−k(N(k))−1 = 2−k. (2.7) (2.8) (2.9) k bXj=1 x =P∞ k=1 xk, xk ∈ Xk. 6 И.С.Фещенко Приведём пример системы S, для которой выполнено условие теоремы 2.4. Для двух подпространств Y, Z пространства X определим ρ0(Y, Z) = sup{d(y, Z) y ∈ Y,kyk = 1}. (2.10) Пример 2.1. Пусть система подпространств Yj, j ∈ Λ (Λ (cid:22) некоторое множество ин- дексов) такова, что замыкание линейной оболочки Yj, j ∈ Λ равно X. Пусть система S = (X; Xk, k > 1) такова, что для каждого j ∈ Λ существует последовательность натуральных чисел k(1) < k(2) < . . ., для которой limn→∞ ρ0(Yj, Xk(n)) = 0. Тогда для каждого j ∈ Λ Yj ⊂ ∆(S). Поэтому S удовлетворяет условию теоремы 2.4, а значит, является ПСП в X. Замечание 1. Если система S удовлетворяет условию теоремы 2.4, то для каждо- го n ∈ N система S(>n) = (X; Xk+n−1, k > 1) также удовлетворяет условию теоремы 2.4, а поэтому является ПСП в X. Поэтому для каждого n ∈ N замыкание линейной оболочки подпространств Xk, k > n равно X. Последнее условие не является достаточ- ным для того, чтобы S была ПСП в X. Это показывает следующий пример (который относится к математическому фольклору). Действительно, если f (x) = P∞ Пусть X = Lp([0, 1], dx) (p ∈ [1,∞)), подпространство Xk порождено xk, k > 0 (нам удобнее нумеровать подпространства числами 0, 1, 2, . . ., а не 1, 2, . . .). Тогда для всех n > 0 замыкание линейной оболочки Xk, k > n равно X, но S не есть ПСП в X. k=0 akxk (сходимость по норме пространства X), то kakxkk → 0, k → ∞, а поэтому для всех достаточно больших k ak 6 (kp + 1)1/p. Поэтому f (x) ∈ C ∞([0, 1)). 2.3. Об одном достаточном условии для того, чтобы x ∈ X допускал разло- жение по системе подпространств S в случае гильбертова пространства X. Пусть X (cid:22) гильбертово пространство. Для подпространства Y ⊂ X обозначим PY ор- топроектор на Y . Пусть S = (X; Xk, k > 1) (cid:22) система подпространств X. Рассмотрим произвольный x ∈ X и попробуем разложить его по системе подпространств S, т.е. представить в виде x =P∞ k=1 xk, где xk ∈ Xk, k > 1. Естественно попробовать определить искомое разложение следующим образом: x = PX1x + (I − PX1)x = PX1x + PX2(I − PX1)x + (I − PX2)(I − PX1)x = . . . = = PXk(I − PXk−1) . . . (I − PX1)x + (I − PXn) . . . (I − PX1)x = . . . nXk=1 Определим операторы E0 = I, En = (I − PXn) . . . (I − PX1), n > 1. Обозначим xn = k=1 xk +Enx, n > 1. Таким образом, получаем следующее утверждение. PXnEn−1x, n > 1, тогда x =Pn Утверждение 2.1. Если Enx → 0, n → ∞, то x =P∞ Таким образом, если последовательность операторов En сходится к 0 сильно при n → ∞, то система подпространств S является ПСП в X. Однако вопрос о сильной сходимости En к 0 может оказаться очень сложным. Рассмотрим следующий пример. k=1 xk. 7 Пусть N ∈ N, Hk, 1 6 k 6 N (cid:22) подпространства X, причём TN k=1 Hk = 0. Пусть отображение i(·) : N → {1, 2, . . . , N} таково, что i(k + 1) 6= i(k), k > 1 и для каждого m ∈ {1, 2, . . . , N} существует бесконечно много k, для которых i(k) = m. Определим Xk = H ⊥ i(k), k > 1. Тогда En = PHi(n) . . . PHi(1), n > 1. Известно, что En сходится к 0 слабо при n → ∞ (см.[9]). Вопрос о сильной сходимости En к 0 при n → ∞ называется проблемой Гальперина и является чрезвычайно сложным (см., например, [10]). В то же время из теоремы 2.4 следует, что S является ПСП в X. 2.4. Устойчивость ПСП в банаховых пространствах. Пусть X (cid:22) банахово про- странство, S = (X; Xk, k > 1) (cid:22) система подпространств X. Мы покажем, что ес- соответствующих результатов мы обобщим результаты параграфов 2,3 работы [7], в которой рассматриваются системы одномерных подпространств, на произвольные си- стемы подпространств. ли подпространства Xk, eXk достаточно ¾близки¿, k > 1, то система подпространств eS = (X; eXk, k > 1) также является ПСП в X. За меру ¾близости¿ подпространств выберем величину ρ0(Xk, eXk), определённую формулой (2.10). Для доказательства [7]). Для набора P = (x1, . . . , xn), где xk ∈ Xk, 1 6 k 6 n, определим Σ(P ) =Pn Введём необходимые определения (обобщающие определения параграфа 2 работы k=1 xk, а также ΘS(P ) = max 16k6nk xjk. kXj=1 Для x ∈ X, ε > 0 определим ΘS(x, ε) = inf{ΘS(P ) kΣ(P ) − xk 6 ε}. (Мы считаем, что inf(∅) = ∞.) Для x ∈ X определим Θ∗ S(x) = sup{ΘS(x, ε) ε > 0} = lim ε→0+ ΘS(x, ε). Наконец, определим Следующие две леммы и теорема доказываются точно так же, как леммы 1,2 и ΘS = sup{Θ∗ S(x) kxk 6 1}. теорема 1 в [7]. S(x) < ∞ для произвольного x ∈ X, то Θ∗ Лемма 2.1. Если Θ∗ эквивалентная k · k. Лемма 2.2. Если для некоторых α ∈ (0, 1), B > 0 выполнено ΘS(x, αkxk) 6 Bkxk, x ∈ X, то Θ∗ S(x) (cid:22) норма на X, S(x) 6 B 1−αkxk, x ∈ X. Теорема 2.5. Следующие утверждения эквивалентны: (1) S является ПСП в X, (2) существуют α ∈ (0, 1), B > 0, такие, что для произвольного x ∈ X (3) Θ∗ ΘS(x, αkxk) 6 Bkxk, S(x) < ∞ для произвольного x ∈ X, 8 И.С.Фещенко (4) ΘS < ∞. Теперь установим теорему об устойчивости ПСП в X. Теорема 2.6. Пусть S (cid:22) ПСП в X. Если система подпространств eS = (X; eXk, k > 1) такова, чтоP∞ k=1 ρ0(Xk, eXk) < (2ΘS)−1, то eS является ПСП в X. Доказательство. Доказательство аналогично доказательству теоремы 2 в [7]. Пусть x ∈ X, x 6= 0. Пусть ε > 0. Существует P = (x1, . . . , xn), xk ∈ Xk, такой, что kx − Σ(P )k 6 ε, ΘS(P ) 6 ΘS(1 + ε)kxk. Тогда kxkk 6 2ΘS(1 + ε)kxk, 1 6 k 6 n. Обозначим dk = ρ0(Xk, eXk), k > 1. Для произвольного k, 1 6 k 6 n, существует exk ∈ eXk, для которого kxk −exkk 6 dk(1 + ε)kxkk. Определим eP = (ex1, . . . ,exn). Тогда nXk=1 kx−Σ(eP )k 6 kx−Σ(P )k+kΣ(P )−Σ(eP )k 6 ε+ dk(1+ε)kxkk 6 ε+2ΘS(1+ε)2kxk Для произвольного m, 1 6 m 6 n, имеем nXk=1 dk. k dk. (cid:3) mXk=1 xkk + k mXk=1 mXk=1exkk 6 k (exk − xk)k 6 ΘS(1 + ε)kxk + 2ΘS(1 + ε)2kxk Зафиксируем α ∈ (2ΘSP∞ k=1 dk, 1), B > ΘS + 2ΘSP∞ ε > 0 из доказаных неравенств имеем Θ eS(x, αkxk) 6 Bkxk. Поэтому eS является ПСП k=1 dk. При достаточно малом mXk=1 в X. 3. Абсолютно представляющие системы подпространств в банаховых пространствах ностей ξ = (x1, x2, . . .), xk ∈ Xk, для которых kξk = P∞ равенством A(x1, x2, . . .) = P∞ 3.1. Критерии АПСП. Пусть X (cid:22) банахово пространство, S = (X; Xk, k > 1) (cid:22) система подпространств X. Определим l1(X1, X2, . . .) как множество последователь- k=1 kxkk < ∞. Ясно, что l1(X1, X2, . . .) (cid:22) банахово пространство. Определим оператор A : l1(X1, X2, . . .) → X k=1 xk. Система S является АПСП в X тогда и только тогда, когда Im(A) = X. Хорошо известно, что это равносильно тому, что A∗ : X ∗ → (l1(X1, X2, . . .))∗ является изоморфным вложением, т.е. для некоторого ε > 0 kA∗ϕk > εkϕk, ϕ ∈ X ∗. Легко видеть, что (l1(X1, X2, . . .))∗ = l∞(X ∗ 2 , . . .) (cid:22) множество всех последовательностей η = (ϕ1, ϕ2, . . .), ϕk ∈ X ∗ k , для которых kηk = supk kϕkk < ∞. k=1 ϕk(xk). Легко видеть, что A∗ϕ = (ϕ(S,1), ϕ(S,2), . . .), где ϕ(S,k) обозначено сужение ϕ на Xk. Таким образом, получаем следующую теорему. При этом действие η(ξ) =P∞ 1 , X ∗ Теорема 3.1. S является АПСП в X тогда и только тогда, когда существует ε > 0 такое, что (3.1) k kϕ(S,k)k > εkϕk, ϕ ∈ X ∗. sup 9 Отметим, что теорему 3.1 можно сформулировать следующим образом (уменьшив ε): S является АПСП в X тогда и только тогда, когда существует ε > 0 такое, что для произвольного ϕ ∈ X ∗,kϕk = 1 существуют k > 1, x ∈ Xk,kxk = 1 такие, что ϕ(x) > ε. Понятие АПСП тесно связано с понятием абсолютно представляющего семейства (АПСм). Напомним (см., например, [5]), что множество D ⊂ X называется АПСм в X j=1 ajxj j=1 kajxjk < ∞. АПСм в банаховых и гильбертовых пространствах изучались в k=1{x ∈ Xk,kxk = 1}. Ясно, что S является АПСП в X тогда и только тогда, когда D(S) является АПСм в X. если для произвольного x ∈ X существуют aj ∈ K, xj ∈ D, такие, что x =P∞ и P∞ [13],[2],[8]. Для системы подпространств S определим D(S) =Sn Приведенный критерий для АПСП (см. абзац после теоремы 3.1) можно получить из следующего хорошо известного критерия для АПСм (см., например, теорему 1 в [5], теорему 2.1 в [13]), который доказывается аналогично теореме 3.1. (Множество D называется нормированным, если kxk = 1, x ∈ D.) Теорема 3.2. Пусть D (cid:22) нормированное множество в X. D является АПСм в X тогда и только тогда, когда существует ε > 0 такое, что для произвольного ϕ ∈ X ∗,kϕk = 1, существует x ∈ D такой, что ϕ(x) > ε. Далее мы докажем критерий для АПСм в X, из которого непосредственно следует критерий для АПСП в X (вместо D надо взять D(S)). Следующая теорема обобщает теорему 3 в [8] и показывает, что в теореме 3.2 условие произвольности ϕ ∈ X ∗ можно ослабить. (Будем говорить, что D тотально в X, если замыкание линейной оболочки D равно X). Теорема 3.3. Пусть X (cid:22) банахово пространство, Y (cid:22) конечномерное подпростран- ство X, D (cid:22) тотальное в X нормированное множество. D является АПСм в X тогда и только тогда, когда существует ε > 0 такое, что для произвольного ϕ ∈ Y ⊥,kϕk = 1 существует x ∈ D такой, что ϕ(x) > ε. Для доказательства нам нужна следующая лемма (которая относится к математи- ческому фольклору). Лемма 3.1. Пусть Y, Z (cid:22) подпространства банахова пространства X. Если Y ∩Z = 0 и Y + Z = X, то существует c > 0 такое, что для произвольных y ∈ Y, z ∈ Z ky + zk > c(kyk + kzk). Доказательство. Определим пространство Y ⊕ Z как множество пар ξ = (y, z), y ∈ Y, z ∈ Z, с нормой kξk = kyk + kzk. Ясно, что Y ⊕ Z банахово. Определим оператор A : Y ⊕ Z → X равенством A(y, z) = y + z. Тогда A ограничен, ker A = 0, ImA = X. Поэтому A обратим, откуда непосредственно следует нужное утверждение. (cid:3) Доказательство теоремы 3.3. Необходимость очевидна. Докажем достаточность. Пред- положим, что D (cid:22) не АПСм в X. Поскольку Y конечномерно, то оно дополняемо в X, т.е. существует подпространство Z такое, что Y ∩ Z = 0 и Y + Z = X. Существует Оценим kηk. Пусть y ∈ Y , y =Pm ϕ(y) 6 tkϕ(ek) 6 mXk=1 mXk=1 Поэтому для произвольных y ∈ Y, z ∈ Z tk(ϕ(ek − fk) + ϕ(fk)) 6 2δ mXk=1 tk 6 2δc−1 2 kyk. 10 И.С.Фещенко c1 > 0 такое, что (3.2) ky + zk > c1(kyk + kzk), y ∈ Y, z ∈ Z. Выберем в Y нормированный базис e1, . . . , em. Существует c2 > 0 такое, что для про- извольных t1, . . . , tm ∈ K kPm k=1 tkekk > c2Pm Рассмотрим произвольное δ > 0. Поскольку D тотально в X, существуют элементы f1, . . . , fm из линейной оболочки D, такие, что kek − fkk < δ,kfkk = 1 для k = 1, . . . , m. Ясно, что {f1, . . . , fm}∪D не является АПСм в X. Поэтому существует ϕ ∈ X ∗, kϕk = 1 такой, что ϕ(fk) 6 δ для 1 6 k 6 m, ϕ(x) 6 δ для x ∈ D. Определим линейные функционалы ψ, η равенствами ψ(y + z) = ϕ(z), η(y + z) = ϕ(y), y ∈ Y, z ∈ Z. Из неравенства (3.2) следует, что ψ, η ∈ X ∗. Более того, ψ ∈ Y ⊥, η ∈ Z ⊥. k=1 tk. k=1 tkek. Тогда Положим c3 = 2c−1 η(y + z) = ϕ(y) 6 2δc−1 1 c−1 2 , тогда kηk 6 c3δ. 2 kyk 6 2c−1 1 c−1 2 δky + zk. Поскольку kϕk = 1, то kψk > (1 − c3δ). Для произвольного x ∈ D ψ(x) 6 δ(1 + c3). Положим eψ = ψ/kψk. Тогда eψ ∈ Y ⊥, keψk = 1. Для произвольного x ∈ D eψ(x) 6 δ(1 + c3)/(1 − c3δ). При достаточно малых δ получим противоречие. Рассмотрим АПСП в равномерно гладких пространствах (АПСм в равномерно глад- ких пространствах изучались в [13],[2]). Напомним определение равномерно гладкого пространства. Определим модуль гладкости пространства X равенством ρ(τ ) = sup{(kx + yk + kx − yk)/2 − 1 kxk = 1,kyk = τ}, τ > 0. (cid:3) X называется равномерно гладким если ρ(τ )/τ → 0 при τ → 0. Для нас равномерно гладкие пространства важны по следующей причине: как мы увидим при доказатель- стве следующей теоремы, если S (cid:22) АПСП в равномерно гладком X, то для каждого x ∈ X разложение x в абсолютно сходящийся ряд по системе подпространств S может быть получено простым ¾естественным¿ образом. Будем говорить, что множество A является λ-сетью для множества B, если для произвольного x ∈ B существует y ∈ A такой, что kx − yk 6 λ. Обозначим VX = {x ∈ X,kxk = 1}. Напомним, что для системы подпространств S = (X; Xk, k > 1) Теорема 3.4. Пусть X (cid:22) равномерно гладкое банахово пространство. Тогда утвер- ждения равносильны: D(S) =S∞ k=1 VXk. (1) S = (X; Xk, k > 1) является АПСП в X, (2) существуют τ, λ ∈ (0, 1) такие, что τ D(S) (cid:22) λ-сеть для VX, (3) λS = supkxk=1 inf k>1 d(x, Xk) < 1. Доказательство. (1) ⇒ (2). Для действительного пространства X нужное утвержде- ние следует из теоремы 3.1 в [13]. Для комплексного X рассмотрим X как простран- ство над R и из упомянутой теоремы получим нужное. 11 (2) ⇒ (3). Очевидно. (3) ⇒ (1). Доказательство аналогично доказательству теоремы 3 в [2]. Фиксируем λ ∈ (λS, 1). Рассмотрим произвольный x ∈ X. Существуют i(1) ∈ N, x1 ∈ Xi(1) такие, что kx − x1k 6 λkxk. Определим y1 = x − x1, тогда ky1k 6 λkxk и x = x1 + y1. Далее проделаем аналогичную процедуру. Пусть мы имеем разложение x = x1 + . . . + xk + yk. Существуют i(k + 1) ∈ N, xk+1 ∈ Xi(k+1) такие, что kyk − xk+1k 6 λkykk. Определим yk+1 = yk − xk+1, тогда kyk+1k 6 λkykk и x = x1 + . . . + xk+1 + yk+1. Индукцией по k легко установить, что kykk 6 λkkxk, k > 1. Поэтому (3.3) x = xk. k=1 kxkk 6 ((1 + λ)/(1− λ))kxk. Для Ясно, что kxkk 6 λk−1(1 + λ)kxk, k > 1, поэтомуP∞ k > 1 определим zk =Pj:i(j)=k xj, тогда zk ∈ Xk. Из (3.3) имеем x =P∞ Замечание 2. В работе [2] доказано, что каждая АПС (одномерных подпространств) в равномерно гладком X является ¾быстрой¿ ПС (см. определение 3 и теорему 3 в [2]). Аналогичное утверждение верно для АПСП. Как следует из доказательства теоремы 3.4, (3) ⇒ (1), каждая АПСП S в равномерно гладком X является ¾быстрой¿ ПСП (наше определение согласовано с определением 3 в [2]): существуют C > 0 и λ ∈ (0, 1) такие, что для произвольного x ∈ X существует инъективное отображение k 7→ n(k) k=1 zk. (cid:3) k=1 yk и kykk 6 Cλkkxk, k > 1. и элементы yk ∈ Xn(k) такие, что x =P∞ ∞Xk=1 Рассмотрим теперь АПСП в гильбертовых пространствах; как мы увидим далее, критерий для АПСП в гильбертовых пространствах приобретает геометрическую на- глядность (см. также теоремы 1,2 в [8]). Итак, пусть X гильбертово. Тогда X ∗ можно отождествить с X: ϕ(·) = (·, ϕ). S является АПСП в X тогда и только тогда, когда существует ε > 0 такое, что для произвольного ϕ ∈ VX существует x ∈ D(S) такой, что (x, ϕ) > ε. Теорема 3.5. Пусть τ > 0, ε ∈ (0, 1]. Следующие условия равносильны: (1) для произвольного ϕ ∈ VX существует x ∈ D(S) такой, что (x, ϕ) > ε, (2) τ D(S) является √1 + τ 2 − 2τ ε-сетью для VX . Доказательство. Для произвольных ϕ ∈ VX, x ∈ D(S) имеем kϕ − τ xk2 = 1 + τ 2 − 2τ Re(x, ϕ). Из этого равенства очевидным образом следует нужное утверждение. (cid:3) Следствие 3.1. Пусть τ > 0. Система подпространств S является АПСП в X то- гда и только тогда, когда существует λ ∈ (0,√1 + τ 2) такое, что τ D(S) является λ-сетью для VX . 3.2. Об одном свойстве АПСП. Перед тем, как сформулировать и доказать следу- ющую теорему, напомним определение C-выпуклого пространства и некоторые свой- ства C-выпуклых пространств. 12 И.С.Фещенко Пусть Y (cid:22) банахово пространство над R. Обозначим c0(R) множество последова- тельностей ξ = (z1, z2, . . .), zk ∈ R, для которых zk → 0 при k → ∞, с нормой kξk = supk zk. Y называется C-выпуклым, если c0(R) не является финитно пред- ставимым в Y (см., например, параграфы 5.1, 5.2 книги [11] и библиографию в конце параграфа 5.2). Для натурального n определим (3.4) C(n, Y ) = inf(max(k αkykk αk = ±1, 1 6 k 6 n) yk ∈ Y,kykk > 1, 1 6 k 6 n) . nXk=1 В из лемм 5.2.1, 5.2.2 и теоремы 5.2.2 книги [11] следует, что Y C-выпукло тогда и только тогда, когда C(n, Y ) → ∞ при n → ∞. Пусть Y (cid:22) банахово пространство над C. Обозначим c0(C) множество последова- тельностей ξ = (z1, z2, . . .), zk ∈ C, для которых zk → 0 при k → ∞, с нормой kξk = supk zk. Y называется C-выпуклым, если c0(C) не является финитно предста- вимым в Y . Для натурального n определим C(n, Y ) формулой (3.4); величину CC(n, Y ) формулой (3.4), только максимум берётся по αk = 1, αk ∈ C. Перенося леммы 5.2.1, 5.2.2 и теорему 5.2.2 книги [11] на случай комплексных пространств, получим, что Y C-выпукло тогда и только тогда, когда CC(n, Y ) → ∞ при n → ∞. Поскольку для произвольных α1, . . . , αn ∈ C,αk 6 1 и произвольных y1, . . . , yn ∈ Y то CC(n, Y ) 6 2C(n, Y ), а поэтому Y C-выпукло тогда и только тогда, когда C(n, Y ) → ∞ при n → ∞. Для I ⊂ N обозначим min(I) наименьший элемент множества I. Теорема 3.6. Предположим, что X ∗ C-выпукло. Если S (cid:22) АПСП в X, то существу- ет N0, такое, что для произвольного конечного I ⊂ N, удовлетворяющего min(I) > N0, система подпространств Xk, k /∈ I является АПСП в X. Теорема 3.6 очевидным образом следует из следующей леммы. Лемма 3.2. Пусть I1, I2, . . . (cid:22) подмножества N, m ∈ N. Предположим, каждое на- туральное n принадлежит не более чем m множествам Ij. Тогда для некоторого j система подпространств Xk, k /∈ Ij является АПСП в X. Доказательство. Существует ε > 0, для которого выполнено (3.1). Предположим, утверждение леммы неверно. Тогда для произвольного j существует ϕj ∈ X ∗,kϕjk = 1, такой, что kϕ(S,k) k 6 2−j для всех k /∈ Ij. Рассмотрим произвольное n ∈ N. Существуют α1, . . . , αn ∈ {±1}, для которых kPn j=1 αjϕjk > C(n, X ∗). Определим ϕ =Pn k 6 m + 1. Из неравенства (3.1) следует m + 1 > εC(n, X ∗), что, в силу произвольности n, противоречит C-выпуклости X ∗. (cid:3) j=1 αjϕj. Тогда kϕk > C(n, X ∗) и для произвольного k kϕ(S,k)k 6Pn j=1 kϕ(S,k) j j k nXk=1 nXk=1 nXk=1 αkykk 6 k Re(αk)ykk + k Im(αk)ykk 6 2 max βk=±1k βkykk, nXk=1 3.3. Устойчивость АПСП. Теорема 3.7. Если S является АПСП в X, то существует δ > 0 такое, что произ- 13 АПСП в X. Доказательство. Пусть ε > 0 такое, что выполнено неравенство (3.1). Покажем, что Выберем ε1 < ε, d1 > d, причём d1 < ε1. Рассмотрим произвольный ϕ ∈ X ∗,kϕk = 1. вольная система eS = (X; eXk, k > 1), удовлетворяющая supk ρ0(Xk, eXk) < δ, является если для системы подпространств eS d = supk ρ0(Xk, eXk) < ε, то eS является АПСП в X. Существует k ∈ N и x ∈ Xk,kxk = 1, такие, что ϕ(x) > ε1. Существует ex ∈ eXk, для которого kx −exk 6 d1. Имеем откуда kϕ( eS,k)k > (ε1 − d1)/(1 + d1). Поэтому eS (cid:22) АПСП в X. ϕ(ex) > ϕ(x) − ϕ(x −ex) > (ε1 − d1), kexk 6 (1 + d1), Список литературы (cid:3) [1] Абанин А.В. Индуктивные абсолютно представляющие системы подпространств // Комплекс- ный анализ. Теория операторов. Математическое моделирование.(cid:22) Владикавказ: Изд-во ВНЦ РАН, 2006, С. 27-34. [2] Вершинин Р.В. О представляющих и абсолютно представляющих системах в банаховых про- странствах // Матем. физ., анал., геом.(cid:22) 1998.(cid:22) Т.5, №1/2.(cid:22) С.3-14. [3] Коробейник Ю.Ф. Представляющие системы // УМН.(cid:22) 1981.(cid:22) Т.36,Вып. 1 (217).(cid:22) С.73-126. [4] Коробейник Ю.Ф. О представляющих системах подпространств // Мат. заметки.(cid:22) 1985.(cid:22) Т.38,№5.(cid:22) С.741-755. [5] Коробейник Ю.Ф. Об абсолютно представляющих семействах в некоторых классах локально выпуклых пространств // Изв. вузов. Матем.(cid:22) 2009.(cid:22) № 9.(cid:22) С.25-35. [6] Михайлов К.А. Абсолютно представляющие системы подпространств в пространствах пробных ультрадифференцируемых функций // Изв. Вузов. Сев. Кав. регион. Естественные науки.(cid:22) 2009.(cid:22) № 6.(cid:22) С. 8-11. [7] Слепченко А.Н. О некоторых обобщениях базисов банаховых пространств // Матем. сб.(cid:22) 1983.(cid:22) Т. 121 (163),№ 2 (6).(cid:22) С.272-285. [8] Шрайфель И.С. Об абсолютно представляющих системах в гильбертовых пространствах // Изв. вузов. Матем.(cid:22) 1995.(cid:22) №9.(cid:22) С.78-82. [9] Amemiya I., Ando T. Convergence of random products of contractions in Hilbert space // Acta Sci. Math. Szeged.(cid:22) 1965.(cid:22) V.26.(cid:22) P.239-244. [10] Bauschke H.H. Projection algorithms: results and open problems // Inherently Parallel Algorithms in Feasibility and Optimization and their Applications (Haifa 2000).(cid:22) D. Butnariu, Y. Censor, S. Reich (editors).(cid:22) Elsevier, 2001.(cid:22) P.11-22. [11] Kadets M.I., Kadets V.M. Series in Banach spaces. Conditional and Unconditional convergence.(cid:22) Birkhauser Verlag, Basel, Boston,Berlin, 1997. [12] Singer I. Bases in Banach spaces II.(cid:22) Springer Verlag, Berlin, Heidelberg, New York, 1981.(cid:22) 880 p. [13] Vershynin R. Absolutely representing systems, uniform smoothness, and type // arXiv: math/ 9804044v1 [math.FA] 8 Apr 1998.
1102.3097	1	1102	2011-02-15T15:21:15	Phase Space Localization of Riesz bases for L^2(R^d)	[ "math.FA" ]	We prove a strong uncertainty principle for Riesz bases in L^2(R^d) and show that the orthonormal basis constructed by Bourgain possesses the optimal phase-space localization.	math.FA	math	(1) (2) PHASE SPACE LOCALIZATION OF RIESZ BASES FOR L2(Rd) KARLHEINZ GR OCHENIG AND EUGENIA MALINNIKOVA Abstract. We prove a strong uncertainty principle for Riesz bases in L2(Rd) and show that the orthonormal basis constructed by Bourgain possesses the optimal phase-space localization. In [B] J. Bourgain constructed an orthonormal basis for L2(R) consisting of functions 1. Introduction fn ∈ L2(R), such that a∈RZR n∈N(cid:16) inf sup x − a2fn(x)2dx + inf b∈RZR ξ − b2bfn(ξ)2dξ(cid:17) < ∞ . Bourgain remarked that the exponent 2 of x − a and ξ − b is optimal and that there are no orthonormal bases with a better phase-space localization. In this paper we prove the following strong uncertainty principle for Riesz bases for L2(Rd). Theorem 1. If {fn}∞ n=1 is a Riesz basis for L2(Rd) and s > d, then n∈N(cid:16) inf a∈RdZRd sup x − a2sfn(x)2dx + inf b∈RdZRd ξ − b2sbfn(ξ)2dξ(cid:17) = ∞ . This theorem therefore asserts that the Bourgain basis possesses the best possible phase-space localization. For the case of an orthonormal basis for L2(R) in dimension d = 1, Bourgain outlines a proof strategy for Theorem 1. Precisely, he writes that "it has been shown by T. Steger that L2(R) does not admit a basis of the form fj = eibj xgj(x−aj), where gj satisfies supj kgjkAǫ < ∞, defining kgk2 Aǫ =Z (1 + x2)1+ǫgj(x)2dx +Z (1 + ξ2)1+ǫbgj(ξ)2dξ . Here ǫ > 0 is any strictly positive number. His argument is based on the fact that the operations x (x-multiplication) and d/dx in the latter basis would become "almost" diagonal operators, violating the non-commutation property [d/dx, x] = I. He also makes use of a density computation due to Y. Meyer of the set Λ of pairs (aj, bj) in phase space. The condition ǫ > 0 is important in Steger's argument as well as for Meyer's distribution result to be valid", see [B]. 2000 Mathematics Subject Classification. 81B99, 81S05, 42C15. Key words and phrases. Uncertainty principle, Balian-Low theorem, Riesz basis, localized frame. K. G. was supported in part by the project P22746-N13 of the Austrian Science Foundation (FWF). E. M. was supported by the Research Council of Norway grant 185359/V30. 1 2 KARLHEINZ GR OCHENIG AND EUGENIA MALINNIKOVA Some of these arguments have made their way into the literature. A density argument related to Meyer's argument has appeared in the fundamental paper of Ramanathan and Steger [RS] on the density of Gabor frames and has become the main technique to inves- tigate the density of frames. See [BCHL, GR, H] for some variations of the Ramanathan- Steger technique. The canonical commutation relations were used in Battle's elegant proof of the Balian-Low theorem [Bat]. However, a full proof of the uncertainty principle of Theorem 1 has not yet been given. Research has focused mainly on bases consisting of phase-space shifts fn(x) = e2πibnxg(x−an) of a single generating function g, so-called Gabor systems. The theorem of Balian-Low asserts that a basis satisfying (1) cannot consist of a (regular) Gabor system. We refer the reader to proofs of the theorem in [D] and [Bat], to the survey articles on the Balian-Low theorem and its generalizations [BHW, CP] and to the monograph [G1] for detailed discussions of the subject. Gabor systems are somewhat easier to handle, because one needs to control the localization of only one function in contrast to Bourgain's case. In this paper we offer a complete proof of Theorem 1 which extends the result men- tioned in [B] to higher dimensions and to Riesz bases instead of orthonormal bases. For orthogonal bases our proof follows the outline of Bourgain. The case of Riesz bases requires additional ideas. We will apply the theory of localized frames [FG, G2] to verify that the biorthogonal basis possesses the same localization properties as the original basis. In a second step we use a bootstrap argument. We will show that if a Riesz basis violates condition (2) for s > d, then we can construct a new Riesz basis with optimal phase-space localization, for instance, with all functions in a Gelfand-Shilov space of test functions. It may seem a lot of effort to prove the non-existence of well-localized phase-space bases, but several arguments are of interest in themselves. The proof combines tools from the density theory of frames, the canonical commutation relations, the theory of localized frames, recent phase-space methods, and a new argument of how to improve the quality of a given basis. One of the corollaries of Theorem 1 is that there is no Riesz basis of phase-space shifts of the Gaussian function in L2(Rd). This fact implies that there is no subset Λ ⊂ Cd that is both sampling and interpolating for the Bargmann-Fock space F 2(Cd). This statement is well-known in dimension d = 1, but seems to have been open in higher dimensions. The paper is organized as follows: In Section 2 we show that a well-localized phase- space basis must be indexed by a set of density one. In Section 3 we prove Theorem 1 for the special case when its biorthogonal basis is also well-localized, it includes the case of an orthonormal basis. Section 4 contains some preparations from time-frequency analysis. In Section 5 we develop the necessary arguments to prove the uncertainty principle of Theorem 1 for Riesz bases. Section 6 elaborates the non-existence of sets of simultaneous sampling and interpolation and concludes with further remarks. LOCALIZATION PROPERTIES OF BASES FOR L2(Rd) 3 2. Density Conditions We say that a sequence of functions {fn}∞ n=1 ⊂ L2(Rd) has phase-space localization of magnitude s, if In this case there exist points (an, bn) ∈ R2d, such that sup a∈RdZRd n∈N(cid:16) inf n∈N(cid:16)ZRd sup x − a2sfn(x)2dx + inf x − an2sfn(x)2dx +ZRd b∈RdZRd ξ − b2sbfn(ξ)2dξ(cid:17) < ∞ . ξ − bn2sbfn(ξ)2dξ(cid:17) < ∞ . Then the set Λ = {(an, bn)}∞ are localized. Note that there is some freedom in the choice of points (an, bn) ∈ R2d. n=1 is the set in the phase-space where the functions {fn}n We will first estimate the density of the set Λ = {(an, bn)}∞ n=1 ⊂ R2d both for Riesz bases and frames for L2(Rd) which have phase-space localization. The ideas we follow are well known, see [RS, SP, LP, S]. Let Λ be a subset of R2d, we denote by D+(Λ) and D−(Λ) its upper and lower Beurling densities, D+(Λ) = lim sup r→∞ sup x∈R2d card(λ ∩ Q(x, r)) Q(x, r) , D−(Λ) = lim inf r→∞ inf x∈R2d card(λ ∩ Q(x, r)) Q(x, r) , where x = (x1, x2) ∈ Rd × Rd and Q(x, r) = {(y1, y2) ∈ Rd × Rd : x1 − y1 < r, x2 − y2 < r} . These densities can be also defined by using dilations of cubes or balls in R2d instead of Q(x, r), as was proved by Landau [L]. A set Λ ⊂ R2d is relatively separated, if supx∈R2d card(cid:0)Λ ∩ (x + [0, 1]2d)(cid:1) < ∞. Clearly, if D+(Λ) < ∞, then Λ is relatively separated. Lemma 1. Suppose that {fn}∞ ization of magnitude s, s > 0, at points {(an, bn)}∞ n=1, i.e., n=1 is a Riesz basis for L2(Rd) that has phase-space local- n∈N(cid:16)ZRd sup x − an2sfn(x)2dx +ZRd ξ − bn2sbfn(ξ)2dξ(cid:17) = S < ∞ . Then Λ = {(an, bn)}∞ Proof. Fix ǫ > 0. We say that a function g ∈ L2(Rd) is ǫ-concentrated on some set E ⊂ Rd if n=1 ⊂ R2d is relatively separated and D+(Λ) ≤ 1. ZE g(x)2dx ≥ (1 − ǫ2)kgk2. Since Zx−an≥r fn(x)2 dx ≤ r−2sZx−an≥r x − an2sfn(x)2 dx ≤ r−2sS , there exists r = r(ǫ) such that fn is ǫ-concentrated on B(an, r) uniformly in n. Likewise and denote QR = Q((x0, ξ0), R). Remark that if (an, bn) ∈ QR, then fn is ǫ-concentrated bfn is ǫ-concentrated on B(bn, r) for every n. We fix (x0, ξ0) ∈ Rd×Rd, consider any R > 0 on B(x0, R + r), and bfn is ǫ-concentrated on B(ξ0, R + r), where r = r(ǫ) as above. 4 KARLHEINZ GR OCHENIG AND EUGENIA MALINNIKOVA We now apply a standard estimate of the trace of a time-frequency restriction operator to conclude that D+(Λ) ≤ 1, see [RS]. Let F be the Fourier transform and PE be the projection operator PEf = χE f (multiplication of f by the characteristic function of E). The phase-space restriction operator is defined by L = PB(x0,R+r)(F −1PB(ξ0,R+r)F )PB(x0,R+r) = P1P2P1 . It is well-known, see for example [FS], that tr(L) = B(x0, R + r)B(ξ0, R + r) = QR+r. For each fn such that (an, bn) ∈ QR, we have kfn − Lfnk ≤ kfn − P1fnk + kP1kkfn − P2fnk + kP1P2kkfn − P1fnk ≤ 3ǫkfnk. Now let {gn} be the biorthogonal basis for {fn}, i.e., (fn, gm) = δnm. Then tr(L) ≥ X(an,bn)∈QR (Lfn, gn) ≥ X(an,bn)∈Qr ((fn, gn) − (fn − Lfn, gn)) ≥ (1−3Cǫ)card(Λ∩QR), where C = supn kfnkkgnk < ∞ (since {fn} is a Riesz basis). Thus card(cid:16)Λ ∩ Q((x0, ξ0), R)(cid:17) ≤ (1 − 3Cǫ)−1Q((x0, ξ0), R + r). Taking the limit R → ∞, we obtain D+(Λ) ≤ (1 − 3Cǫ)−1 for every ǫ > 0, and thus D+(Λ) ≤ 1, and Λ is relatively separated. (cid:3) Remark. If {fn}∞ n=1 is a frame that has phase-space localization of magnitude s > 0 at points {(an, bn)}∞ n=1 and satisfies kfnk2 ≤ C, then it is still true that Λ = {(an, bn)}∞ n=1 is a relatively separated set and that D+(Λ) < ∞. This follows by compactness arguments, see Theorem 3.5 in [JP] for a similar result in dimension d = 1. Lemma 2. Suppose that {fn}∞ n=1 is a frame for L2(Rd) with supn∈N kfnk2 = C < ∞. If s > d and {fn}n has phase-space localization of magnitude s at points {(an, bn)}n, then Λ = {(an, bn)}∞ n=1 is relatively separated and D−(Λ) ≥ 1. We remark that the Lemma does not hold for s = d. This can be seen from the construction of an orthonormal basis in [B]. Proof. Let K(y, l) denote the cube with center y ∈ Rq and the side length 2l, K(y, l) = {z ∈ Rq : ky − zk∞ < l}, where kzk∞ = max1≤s≤q zs, z = (z1, . . . , zq) ∈ Rq. Fix ǫ > 0 and choose δ in the open interval (d/s, 1). This is possible by the hypothesis s > d. Step 1. An estimate for the coefficients (ψ, fm) of a localized function. Assume that ψ ∈ L2(Rd), kψk2 = 1, ψ is ǫ-concentrated on K(a, R − Rδ) and its Fourier transform is supported on K(b, R − Rδ). Set η = ψ(1 − χK(a,R−Rδ)), so that kηk2 ≤ ǫ. LOCALIZATION PROPERTIES OF BASES FOR L2(Rd) 5 If kam − ak∞ > 2kR, then the following estimate holds: (ψ, fm)2 ≤ 2(ψ(1 − χK(a,R−Rδ )), fm)2 + 2(cid:18)ZK(a,R−Rδ) ≤ 2(η, fm)2 + 2(cid:18)((2k − 1)R + Rδ)−sZK(a,R−Rδ) ψ(x)fm(x)dx(cid:19)2 x − amsψ(x)fm(x)dx(cid:19)2 ≤ 2(η, fm)2 + 2((2k − 1)R + Rδ)−2sS . If kbm − bk∞ > 2kR, then (ψ, fm)2 = (bψ,cfm)2 =(cid:18)ZK(b,R−Rδ ) bψ(ξ)bfn(ξ)dξ(cid:19)2 ≤ ((2k − 1)R + Rδ)−2sS . Let M0 = {n : (an, bn) ∈ K(a, R) × K(b, R)} and M be the complement of M0, M = {n : (an, bn) 6∈ K(a, R) × K(b, R)}. We further partition M into the sets Mk as follows: Mk = {n : max(kan − ak∞, kbn − bk∞) ∈ [2kR, 2k+1R)}, k ≥ 0. Since D+(Λ) < ∞ by Lemma 1 (see also the remark after the lemma), we find that card(Mk) ≤ C1(2kR)2d for some constant C1 large enough.Thus Xm∈M (ψ, fm)2 = (ψ, fm)2 ∞Xk=0 Xm∈Mk ≤ 2 Xm∈M (η, fm)2 + 2S ∞Xk=0 C122kdR2d(cid:0)(cid:0)2k − 1(cid:1) R + Rδ(cid:1)−2s ≤ 2Bkηk2 2 + 2SC1R2d−2sδ + 2SC1R2d−2s ∞Xk=1 22kd(2k − 1)−2s , where B is the upper frame bound of {fn}n. Since s > d by assumption, the last sum converges. Further, sδ > d and, by choosing R large enough, the second and third terms can be made arbitrarily small. Given ǫ > 0 and δ ∈ (d/s, 1), we find that (3) Xm∈M (ψ, fm)2 ≤ C 2 0 ǫ2 for R ≥ R0(ǫ, δ, Λ, S) , with the constant C0 depending only on the frame bound B of {fn}n. Step 2. Comparison with a basis of prolate spheroidal functions. For given ǫ > 0, δ ∈ (d/s, 1), and R ≥ R0(ǫ, δ, Λ, S), we now consider those prolate spheroidal functions φ1, ..., φN with N = N(R) that are ǫd−1/2 concentrated on (−R + Rδ, R − Rδ) and whose Fourier transforms are supported on (−R + Rδ, R − Rδ). We refer the reader to [SP] and [LP] for definitions and properties of these functions. According to [LP] the number of φj with these concentration properties satisfies limR→∞ N(R)R−2 = 1. 6 KARLHEINZ GR OCHENIG AND EUGENIA MALINNIKOVA In higher dimensions we take tensor products of phase-space shifts of these prolate spheroidal functions. Let σ = (n1, . . . , nd) ∈ {1, 2, . . . , N(R)}d and define ψσ(x) = dYj=1 e−2πibj xj φnj (xj − aj), then we obtain an orthonormal set of N d functions {ψσ}σ that are ǫ-concentrated on K(a, R − Rδ) and whose Fourier transforms are supported on K(b, R − Rδ). Now let {gn}n be the dual frame of {fn}n. If A > 0 is the lower frame bound of {fn}n, then we have kXn cngnk2 2 ≤ A−1kck2 2 for every c ∈ ℓ2 . Step 3. Density estimate. We now follow the argument of Ramanathan and Steger in [RS]. Let S be the orthogonal projection of L2(Rd) onto Ψ = span{ψσ : σ ∈ {1, . . . , N(R)}d} and T be the orthogonal projection onto G = span{gn : n ∈ M0}. We consider U : Ψ → Ψ, U = S ◦ T . For each ψ ∈ Ψ we obtain kψ − Uψk2 = kS(ψ − T ψ)k2 ≤ kψ − T ψk = inf g∈G (ψ, fn)gnk = k Xm∈M (ψ, fm)gmk ≤ A−1/2(cid:16) Xm∈M kψ − gk ≤ kψ − Xn∈M0 (ψ, fm)2(cid:17)1/2 . Since each basis function ψσ is in Ψ and satisfies the concentration assumptions from Step 1, the estimate (3), implies that kψσ − Uψσk2 ≤ A−1/2C0ǫ . Consequently, tr(U) ≥Xσ (Uψσ, ψσ) =Xσ (cid:16)kψσk2 2 − (ψσ − Uψσ, ψσ)(cid:17) ≥ (1 − A−1/2C0ǫ)N(R)d . On the other hand, since U is the composition of two projections, all eigenvalues of U belong to (0, 1), and therefore tr(U) ≤ rank(U) ≤ dim(G). Thus (1 − A−1/2C0ǫ)N(R)d ≤ tr(U) ≤ card(Λ ∩ K(a, R) × K(b, R)) . We now use the definition of the Beurling density with cubes in R2d instead of balls, and obtain D−(Λ) = lim R→∞ inf (a,b)∈R2d card(cid:0)Λ ∩ K(a, R) × K(b, R)(cid:1) N(R)d R2d = 1 − A−1/2C0ǫ . R2d ≥ (1 − A−1/2C0ǫ) lim R→∞ As ǫ > 0 was arbitrary, we conclude that D−(Λ) ≥ 1. (cid:3) Combining Lemmas 1 and 2, we obtain the density result for localized Riesz bases (recall, however, that our aim is to prove that there are no such bases). LOCALIZATION PROPERTIES OF BASES FOR L2(Rd) 7 Corollary. If s > d and {fn}∞ calization of magnitude s at points {(an, bn)}∞ D(Λ) = D+(Λ) = D−(Λ) = 1. n=1 is a Riesz basis for L2(Rd) that has phase-space lo- n=1, then the density of Λ = {(an, bn)}n is 3. Uncertainty identity We first prove Theorem 1 under the additional condition that the dual basis is also well-localized. The proof extends Battle's elegant proof of the Balian-Low theorem [Bat] and rediscovers Steger's argument mentioned by Bourgain in [B] (see the quote above). The core of the argument is the following uncertainty identity (the canonical commu- tation relations) (xf, ∇g) + (∇f, xg) = dXj=1 (xjf, ∂g ∂xj ) + ( ∂f ∂xj , xjg) = −d(f, g), which holds provided that f, g, ∂f ∂xj , ∂g ∂xj , xjg, xjf ∈ L2(Rd) for j = 1, . . . , d. Lemma 3. Assume that {fn}∞ {gn}∞ n=1. If the bases satisfy the localization estimates n=1 is a Riesz basis for L2(Rd) with the biorthogonal basis (a) RRd x − an2sfn(x)2dx +RRd ξ − bn2sbfn(ξ)2dξ ≤ S2 < ∞ for every n; (b) RRd x − an2sgn(x)2dx +RRd ξ − bn2sbgn(ξ)2dξ ≤ T 2 < ∞ for every n; and n=1 ⊂ R2d is relatively separated and 0 < D−(Λ) ≤ D+(Λ) < ∞, (c) Λ = {(an, bn)}∞ then s ≤ d. Proof. We assume that s > d and use the uncertainty identity to derive a contradiction −−−−→ (xf, g) ∈ Cd for the vector with components from (a)−(c). In the following we will write −−−−→ (∇f, g) = ( ∂f , g)d (xjf, g), j = 1, . . . , d. Likewise ∂xj j=1. Step 1. An estimate for non-diagonal coefficients. Condition (a) implies xfn ∈ L2(Rd)d, and then the sequence of vectors cn m = −−−−−−→ (xfn, gm) = (xjfn, gm)d j=1 ∈ Cd is well defined. By the biorthogonality condition, for m 6= n −−−−−−−−−−−→ ((x − an)fn, gm). Since {gm} is a Riesz basis, assumption (b) implies that cn m = cn m2 ≤ Bkx − an fnk2 2 ≤ BS2 . Xm:m6=n Xn:n6=m Next, since cn m = −−−−−−→ (xgm, fn), we also have cn m2 ≤ Bkx − am gmk2 2 ≤ BT 2. Here B is the upper basis constant for both Riesz bases {gm}m and {fn}n. The coefficients dn m = −−−−−−→ (ζbfn,cgm) = (2πi)−1−−−−−−→ (∇fn, gm) 8 KARLHEINZ GR OCHENIG AND EUGENIA MALINNIKOVA enjoy similar properties. Step 2. Commutation relations. We now apply the uncertainty identity to each pair {fn, gn} and obtain (4) (xfn, gm) · d = −Xm (cid:16)−−−−−−→ = −2πiXm m · dm (cn n − dn m · cm n ), −−−−−−→ (∇gn, fm) + −−−−−−→ (∇fn, gm) · −−−−−−→ (xgn, fm)(cid:17) where λ · µ =Pd j=1 λjµj is the standard scalar product in Cd. For each R > 0 define N (R) = {n : an ≤ R, bn ≤ R} and N(R) = card N (R). Now we sum up the identities (4) for all n ∈ N (R), (−cn m · dm n + dn m · cm n ) d 2πi (5) N(R) = Xn∈N (R)Xm dXj=1 = Xn,m∈N (R) (−(cn m)j(dm n )j + (dn m)j(cm n )j) + Xn∈N (R) Xm6∈N (R) (−cn m · dm n + dn m · cm n ). Clearly, the first sum equals zero. We will derive a contradiction by showing that the second sum possesses a slower growth than N(R). We divide the necessary estimates into several steps. Step 3. Points (an, bn) near the boundary. To estimate the second sum, we partition N (R) into two sets, N (R) = N (R − r) ∪(cid:0)N (R) \ N (R − r)(cid:1) , where r = Rδ for some δ ∈ (d/s, 1). First, for n ∈ R(R − r, R) = N (R) \ N (R − r) we get Xn∈R(R−r,R) Xm6∈N (R) cn mdm n ≤ (6) ≤ Xn∈R(R−r,R)(cid:16) Xm:m6=n cn m2(cid:17)1/2(cid:16) Xm:m6=n dm n 2(cid:17)1/2 ≤ (N(R) − N(R − r))BS1T1 . The sum of dn mcm n admits the same estimate. Step 4. Further partition of N (R) for interior (an, bn). Next we partition the complement of N (R) into the rings Nk = N (Rk+1) \ N (Rk) where Rk = 2kR, k ≥ 0. LOCALIZATION PROPERTIES OF BASES FOR L2(Rd) 9 Then for each n ∈ N (R − r) we get Xm6∈N (R) cn mdm n = ≤ cn mdm n ∞Xk=0 Xm∈Nk ∞Xk=0 Xm∈Nk:am>Rk ≤(cid:16) ∞Xk=0 Xm∈Nk:am>Rk +(cid:16) Xm:m6=n cn cn mdm n + ∞Xk=1 Xm∈Nk:bm>Rk n 2(cid:17)1/2 n 2(cid:17)1/2 dm dm . cn m2(cid:17)1/2(cid:16) Xm:m6=n m2(cid:17)1/2(cid:16) ∞Xk=0 Xm∈Nk:bm>Rk cn mdm n + Step 5. Main estimate. Now we write down an estimate for cn m when an < R − r and am > Rj. Set h(j) n (x) = (x − an)jfn(x)(1 − χB(R−r/2)(x)), j = 1, 2, ..., d. Then kh(j) n k2 2 ≤Zx>R−r/2 (x − an)j2fn(x)2dx. Further, for x ≥ r/2 and an ≤ R − r, we have x − an ≥ r/2, and therefore kh(j) n k2 2 ≤ (r/2)2−2sZRd (7) dXj=1 Then we have x − an2sfn(x)2dx ≤ (r/2)2−2sS2. (8) cn m2 = ≤ dXj=1 (cid:0)(x − an)jfn, gm(cid:1)2 dXj=1(cid:0)2((x − an)jfnχB(R−r/2), gm)2 + 2(h(j) n , gm)2(cid:1) ≤ 2S2kgmχB(R−r/2)k2 2 + 2 dXj=1 (h(j) n , gm)2 (9) ≤ 2S2T 2(Rk − R + r/2)−2s + 2 dXj=1 (h(j) n , gm)2. And since {gm}m is a Riesz basis, (10) Xm (h(j) n , gm)2 ≤ Bkh(j) n k2 2 . 10 KARLHEINZ GR OCHENIG AND EUGENIA MALINNIKOVA Summing up the estimates (8) over all k and all m ∈ Nk such that am > Rk and taking into account (7) and (10), we obtain (11) ∞Xk=0 Xm∈Nk:am>Rk cn m2 ≤ 2S2T 2 ∞Xk=0 N(Rk+1)(Rk − R + r/2)−2s + B(r/2)2−2sS2. n where the summation is over all k and m, where the sums are over all k and m ∈ Nk such that bm > Rk and over all k and m ∈ Nk such that am > Rk, by We can derive similar estimates ofPkPm dm m ∈ Nk such that bm > Rk by using the localization inequality for bgn. Likewise, we obtain the estimates for PkPm cm using the localization conditions on gn and bfn. Step 6. Comparison of the densities. Finally we combine the inequality obtained in Step 4 with (11) and similar inequalities with other combinations of indices. Then we obtain for every n ∈ N (R − r) n and PkPm dn Xm6∈N (R) (cn mdn m + cm n dm n ) ≤(cid:16)C1r2−2s + C2 ∞Xk=0 N(Rk+1)(Rk − R + r/2)−2s(cid:17)1/2 , where C1 and C2 depend on S, T, B, s, and supn kfnk2 and supn kgnk2. Assumption (c) (the estimate of the upper density D+(Λ) < ∞) implies that N(Rk+1) ≤ D122d(k+1)R2d and N(R − r) ≤ D1(R − r)2d for some D1 > 0 and all R large enough. Then for R large enough we obtain (cn mdm n + cm Xn∈N (R−r) Xm6∈N (R) N(R − r)(cid:16)C1r2−2s + C2D122dR2d(r/2)−2s + C2D1R2d n dn m) ≤ ∞Xk=1 22d(k+1)(2kR − R + r/2)−2s(cid:17)1/2 ≤ C(R − r)2dRdr−s, where C depends on S, T, B, s, D1, d as well as supn kfnk2 and supn kgnk2. To finish the proof we recall that r = Rδ and δ ∈ (d/s, 1). Observe that for r large enough the estimate of the upper density implies N(R) − N(R − r) ≤ D1R2d−1r (we just cover the set Q(0, R) \ Q(0, R − r) by cubes with side length r). Now, combining (5), (6), and the last inequality, we obtain N(R) ≤ C3(cid:16)N(R) − N(R − r) + R3d−δs(cid:17) ≤ C4(R2d−1+δ + R3d−s). If we now let R go to infinity, we see that D−(Λ) ≤ lim R→∞ N(R) R2d ≤ C4 lim R→∞(cid:16)Rδ−1 + Rd−s(cid:17) = 0 . This conclusion contradicts the assumption (c) that the lower density estimate D−(Λ) is strictly positive. (cid:3) LOCALIZATION PROPERTIES OF BASES FOR L2(Rd) 11 Lemma 3 concludes the proof of Theorem 1 for the case of an orthonormal basis. For a Riesz basis we are not able to prove that the phase-space localization of magnitude s (condition (a) of the lemma) with s > d implies the required localization for its biorthogonal basis (stated in (b)). For the general case a more complicated argument is presented in the next sections. 4. Some preliminaries on modulation spaces The proof of Theorem 1 for Riesz bases requires some tools from time-frequency anal- ysis. We give a minimalistic account of the required facts on the short-time Fourier transform and modulation spaces. The reader can find the details and a much more general theory of modulation spaces in [G1]. Short-time Fourier Transform. For (a, b) ∈ R2d we write π(a, b)f (t) = e2πib·tf (t − a) for the phase-space shift of a function f on Rd. Let g(x) = 2−d/4e−πx2 be the normalized Gaussian function on Rd. We consider the short-time Fourier transform of a function φ ∈ L2(Rd) with respect to g, [G1, Chapter 3] Vgφ(x, ξ) = (φ, π(x, ξ)g) =ZRd φ(t)g(t − x)e−2πit·ξdt, x, ξ ∈ Rd. The inversion formula for the short-time Fourier transform yields φ(t) =ZZR2d Vgφ(x, ξ)e2πiξ·tg(t − x)dx, for every φ ∈ L2(Rd), with a weak interpretation of the vector-valued integral. Modulation Spaces. For each s ≥ 0 let L2 s(Rm) = {f ∈ L2(Rm) : kf k2 L2 s f (x)2(1 + x)2sdx < ∞} . =ZRm The modulation space M 2 s (Rd) is defined by M 2 s (Rd) = {φ ∈ L2(Rd) : kφkM 2 s = kVgφkL2 s(R2d) < ∞}. The following norm equivalence identifies the modulation space M 2 Lebesgue space L2 s, see [G1, Prop. 11.3.1, 12.1.6]: s ∩ F L2 s with the Fourier- The adjoint operator V ∗ s ≤ kφkL2 c1kφkM 2 g of the short-time Fourier transform is defined on L2(R2d) by s ≤ c2kφkM 2 s . s + kbφkL2 g F (t) =ZZR2d V ∗ F (x, ξ)e2πiξ·tg(t − x)dxdξ =ZZR2d F (x, ξ)π(x, ξ)g(t)dxdξ . If F ∈ L2 s(R2d), then by [G1, Prop. 11.3.2] V ∗ s (Rd) and g F ∈ M 2 s ≤ CkF kL2 s. kV ∗ g F kM 2 (12) 12 KARLHEINZ GR OCHENIG AND EUGENIA MALINNIKOVA We note that the phase-space localization of magnitude s can be rephrased as sup n∈N inf (a,b)∈R2d kπ(a, b)fnkM 2 s < ∞ . Amalgam spaces. We define the amalgam space W (L2 the space of all continuous function on R2d for which the norm s) ⊂ L2 s(R2d) ∩ L∞(R2d) as kF k2 W (L2 s) = Xk,n∈Zd sup F (x + k, ξ + n)2 (1 + k + n)2s x,ξ∈[0,1]d is finite. For s = 0, kF kW (L2) ≥ kF k2 obviously. The continuity of F implies the existence of points xkn, ξkn ∈ [0, 1]d, such that kF k2 W (L2 s) = Xk,n∈Zd F (k + xkn, n + ξkn)2(1 + k + n)2s . The definition of W (L2 relatively separated, z ∈ R2d, and F ∈ W (L2 s) implies the following sampling inequality: If Λ = {λn} ⊆ R2d is s), then ≤ sup k∈Z2d card(cid:0)Λ ∩ (k + [0, 1]2d)(cid:1) kF kWL2 s . (13) (14) (15) (16) (17) (cid:16)Xn F (z + λn)2 (1 + z + λn)2s(cid:17)1/2 s(Rd) with bφ ∈ L2 kVgφkW (L2 s) ≤ CkVgφkL2 s = CkφkM 2 s . The following important inequality links modulation spaces with amalgam spaces: For every φ ∈ L2 s(Rd) we have, e.g., by [G1, Theorem 12.2.1], 5. Basis modification To finish the proof of Theorem 1, we will modify a given Riesz basis {fn}n for L2(Rd) that has phase-space localization of magnitude s > d into a Riesz basis with much better localization properties. The argument in this section may be of independent interest and can also be used to prove positive results about frames and bases. Proposition 1. Assume that {fn}∞ n=1 is a Riesz basis for L2(Rd) that satisfies for some s > d. Then there exists a Riesz basis {hn}∞ sup n∈N(cid:16)ZRd n∈N(cid:16)ZRd sup x − an2sfn(x)2 +ZRd x − an2thn(x)2 +ZRd ξ − bn2sbfn(ξ)2(cid:17) < ∞ ξ − bn2tchn(ξ)2(cid:17) < ∞ n=1 that satisfies for every t > 0. Proof. The new basis is obtained by a modification of {fn}n. We use the inversion formula for the short-time Fourier transform and truncate it. In the language of time- frequency analysis we apply a localization operator to fn. Precisely, let R > 0 and Q(R) = Q(0, R) = B(0, R) × B(0, R) ⊂ R2d. Then the localization operator AR is defined by ARf (t) =ZZQ(R) Vgf (x, ξ)e2πiξ·tg(t − x)dxdξ, f ∈ L2(Rd) . LOCALIZATION PROPERTIES OF BASES FOR L2(Rd) 13 Intuitively, ARf is the part of f that is concentrated on the set Q(R) in the phase space. For more on localization operators see for instance [W, CG]. We recast the assumption as follows: {fn}n is a Riesz basis for L2(Rd), fn(x) = e2πibnxφn(x − an), s > d and kφnk2 M 2 s sup n∈N ≤ S < ∞. We now define (18) ψn = ARφn =ZZQ(R) Vgφn(x, ξ)π(x, ξ)g dxdξ and the modified basis hn(x) = hR n (x) = e2πibnxψn(x − an). Claim. For R large enough {hn}∞ To prove the claim, it suffices to show that for every ǫ > 0 there exists R such that n=1 is a Riesz basis for L2(Rd). for hn = hR n and every sequence {cn}n ∈ ℓ2 the inequality ≤ ǫkcnk2 cn(fn − hn)(cid:13)(cid:13)(cid:13)2 (cid:13)(cid:13)(cid:13)Xn cnfnk2 − kXn cnhnk2 ≥ kXn kXn holds. If {fn}n is a Riesz basis with the lower basis constant A > 0, then cn(fn − hn)k2 ≥ (A − ǫ)kck2 , and so {hn}n is a Riesz basis. Using once again the crucial assumption s > d, we now choose a number σ such that d < σ < s. Using the inversion formula for the short-time Fourier transform and (18), we write φn − ψn =ZR2d(cid:0)1 − χQ(R)(x, ξ)(cid:1)Vgφn(x, ξ)π(x, ξ)g dxdξ , and estimate the M 2 σ -norm of φn − ψn with (12) as kφn − ψnk2 M 2 σ ≤ZR2d(cid:0)1 − χQ(R)(x, ξ)(cid:1)Vgφn(x, ξ)2(1 + x + ξ)2σ dxdξ ≤ (1 + R)2(σ−s)ZR2d(cid:0)1 − χQ(R)(x, ξ)(cid:1)Vgφn(x, ξ)2(1 + x + ξ)2s dxdξ ≤ C(1 + R)2(σ−s)kφnk2 ≤ C(1 + R)2(σ−s)S . M 2 s Choosing now R large enough, we have kφn − ψnkM 2 σ < ǫ for all n . 14 KARLHEINZ GR OCHENIG AND EUGENIA MALINNIKOVA Then we have, with (13) and a suitable choice of points (xk,m, ξk,m) ∈ [0, 1]2d, that (cid:13)(cid:13)(cid:13)Xn 2 2 2 2 2 ≤ C(cid:13)(cid:13)(cid:13)Xn cn(fn − hn)(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)Xn = X(k,m)∈Z2d(cid:12)(cid:12)(cid:12)Xn ≤ X(k,m)∈Z2d(cid:16)Xn ≤ X(k,m)∈Z2d(cid:16)Xn cnVg(fn − hn)(cid:13)(cid:13)(cid:13) cnVg(fn − hn)(cid:13)(cid:13)(cid:13) cnVg(fn − hn)(xk,m + k, ξk,m + m)(cid:12)(cid:12)(cid:12) cn Vg(φn − ψn)(xk,m + k − an, ξk,m + m − bn)(cid:17)2 (1 + k − an + m − bn)−2σ(cid:17) × W (L2) 2 uniformly bounded independent of k and m. Thus we obtain ×(cid:16)Xn cn2Vg(φn − ψn)(xk,m + k − an, ξk,m + m − bn)2(1 + k − an + m − bn)2σ(cid:17) . Since σ > d and Λ is relatively separated, the sum Pn(1 + k − an + m − bn)−2σ is cn(fn − hn)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)Xn cn2Xk,m ≤ CXn Vg(φn − ψn)(xk,m + k − an, ξk,m + m − bn)2(1 + k − an + m − bn)2σ . ≤ 2 2 By (14) and (15) we estimate further that Xk,m Vg(φn − ψn)(xk,m + k − an, ξk,m + m − bn)2(1 + k − an + m − bn)2σ ≤ C2kVg(φn − ψn)k2 σ) ≤ C3kφn − ψnk2 M 2 σ W (L2 < C3ǫ2 . Collecting all estimates, we arrive at (cid:13)(cid:13)(cid:13)Xn cn(fn − hn)(cid:13)(cid:13)(cid:13) 2 2 ≤ C3ǫ2Xn cn2 = C3kck2 2 ǫ2 . Consequently, {hn}n is a Riesz basis for L2(Rd). Finally, applying (12) once again, we obtain, for arbitrary t > 0, kψnkL2 t + k ψnkL2 t ≤ CkψnkM 2 t ≤ C ′kVgφnχQ(R)kL2 t ≤ CtRt , which is (17). (cid:3) REMARK: The construction of ψn implies that ψn(t) ≤ Ce−αt2 for some α, β, C > 0. Thus the perturbed basis belongs to the Gelfand-Shilov space S1/2,1/2, the smallest space of test functions that is invariant under the Fourier transform. and cψn(ξ) ≤ Ce−βξ2 To complete the proof of Theorem 1 we will show that the biorthogonal basis {fhn}n satisfies (16) for some s large enough and then apply Lemma 3. LOCALIZATION PROPERTIES OF BASES FOR L2(Rd) 15 The modified basis {hn}n possesses enough phase-space localization so that the theory of localized frames [FG,G2] is applicable. We say that a frame {hλ : λ ∈ Λ} is s-localized over the index set Λ ⊆ Rd, if its Gramian satisfies (19) (hµ, hλ) ≤ C(1 + λ − µ)−s for all λ, µ ∈ Λ . The main result about localized frames asserts that the dual frame possesses the same type of localization. Specifically we need the following result taken from [FG, Thm. 1.1, Cor. 3.7] Proposition 2. Let Λ ⊆ Rd be a relatively separated set and {hλ : λ ∈ Λ} be a frame for L2(Rd). Assume that {hλ} is s-localized for s > d. Then the (canonical) dual frame is also s-localized, i.e., (20) for all λ, µ ∈ Λ . (fhµ,fhλ) ≤ C ′(1 + λ − µ)−s For a Riesz basis this result can be proved directly. The Gramian matrix G of the basis with entries (hµ, hλ) is invertible with inverse (G−1)λµ = (fhµ,fhλ). By a theorem of Jaffard [Jaf] the polynomial off-diagonal decay is preserved under inversion, whence follows the statement of the proposition (for a Riesz basis). To apply Proposition 2, we need to compare the phase-space localization of magnitude s in Bourgain's uncertainty principle with the localization defined by the off-diagonal decay of the Gramian. Lemma 4. (i) If {hn}n is a Riesz basis that has phase-space localization of magnitude s > 0, then {hn}n is s-localized over the index set Λ in the sense of (19). (ii) If {hn}n is a Riesz basis that has phase-space localization of magnitude s > 3d, then the biorthogonal basis {fhn}n has phase-space localization of magnitude t for any t ∈ (d, s). Proof. (i) We choose the set Λ = {(an, bn)}n ⊂ R2d as the appropriate index set. By Lemma 1, Λ is relatively separated. Then (hn, hm) ≤ ≤ sup x∈Rd(cid:16)(1 + x − am)(1 + x − an)(cid:17)−s ZRd n∈NZRd hn(x)(1 + x − an)shm(x)(1 + x − am)s dx ≤ C(1 + am − an)−s sup hn(x)2(1 + x − an)2s dx = (1 + am − an)−sS2 . The argument for the Fourier transforms yields that By combining both estimates we arrive at (hn, hm) = (chn,chm) ≤ C(1 + bm − bn)−sS2 . (hn, hm) ≤ C(1 + λ − µ)−s for all λ, µ ∈ Λ , in other words, {hn} is s-localized over Λ. 16 KARLHEINZ GR OCHENIG AND EUGENIA MALINNIKOVA (ii) By (i) {hn} is s-localized on Λ ⊆ R2d and s > 3d. Therefore Proposition 2 applies and the dual frame is also s-localized as in (20). Now we have Next let d < t < s − 2d (which is possible by s > 3d). A straightforward calculation and (17) give fhn =Xm x − an2thm(x)2dx ≤ZRd ZRd (fhn,fhm)hm. Ct(cid:0)x − am2t + an − am2t(cid:1)hm(x)2dx ≤ C(S + an − am2t) ≤ C ′(1 + λn − λm)2t. With the triangle inequality for the L2 t -norm we obtain (cid:16)ZRd x − an2thn(x)2dx(cid:17)1/2 Xm (hn, hm)hm(x)2x − an2t dx(cid:17)1/2 =(cid:16)ZRd ≤ CXm ≤ CXm (hn, hm)(1 + λn − λm)t (1 + λn − λm)t−s . Since Λ is relatively separated and t − s < −2d, the last sum is uniformly bounded. Similar estimates hold for the Fourier transforms. Consequently, the dual basis is (cid:3) localized in the sense of (16) for t ∈ (d, s). We now can finish the proof of Theorem 1 for Riesz bases. We start with a Riesz basis {fn}n that satisfies (16) for some s > d. Then we modify this basis by means of Proposition 1 to a Riesz basis {hn}n that satisfies the localization estimates (16) for all t > 0. Finally, Lemma 4 guarantees that the dual basis {fhn}n also satisfies the localization estimates (16) for all t > 0. Thus all assumptions of Lemma 3 are satisfied whence we conclude that the localization of {hn}n is t ≤ d. This is a contradiction to the construction of {hn}n. This means that the original basis {fn}n cannot satisfy the strong localization estimate s > d, and the proof of Theorem 1 is complete. 6. Sampling in Bargmann-Fock spaces and Concluding Remarks F =RCd F (z)2 e−πz2 We finally give an application of Theorem 1 to several complex variables. Recall that the Bargmann-Fock space F consists of all entire functions on Cd with norm kF k2 dz. The Bargmann-Fock space possesses the reproducing kernel Kw(z) = eπ ¯w·z for w, z ∈ Cd, so that F (w) = (F, Kw). Consequently, a set {Kλ : λ ∈ Λ} is a Riesz basis for F , if and only if Λ ⊆ Cd is simultaneously sampling F, and for every c ∈ ℓ2(Λ) there and interpolating for F , i.e., Pλ F (λ)2e−πλ2 ≍ kF k2 exists a (unique) F ∈ F , such that F (λ)e−πλ2/2 = aλ. The following result is an immediate consequence of Theorem 1. Theorem 2. The Bargmann-Fock space does not admit a set Λ ⊆ Cd that is simultane- ously sampling and interpolating. LOCALIZATION PROPERTIES OF BASES FOR L2(Rd) 17 Proof. We use the Bargmann transform defined as Bf (z) = F (z) = 2d/4e−πz2/2ZRd f (t)e−πt·te2πt·zdt z ∈ Cd , and translate Theorem 1 into a statement of complex analysis. The Bargmann trans- form is unitary from L2(Rd) onto F and maps the phase-space shifts of the Gaussian e−2πiw2·xe−π(x−w1)2 = π(w1, −w2)g(x) to the reproducing kernel eiαKw e−w2/2 for some phase factor c = 1. Thus Λ ⊆ Cd is a set of sampling and interpolation, if and only if {Kλ : λ ∈ Λ} is a Riesz basis for F . Clearly, the Gaussian satisfies the localization condition (16) for all s > 0. By Theo- rem 1 a set of phase-space shifts of the Gaussian cannot from a Riesz basis for L2(Rd). Consequently, no set Λ ⊆ Cd can be simultaneously sampling and interpolating. (cid:3) REMARK: Theorem 2 is a statement of complex analysis. Indeed, in dimension d = 1 is it well-known and can be proved with classical methods from complex variables. In higher dimensions the result was expected, but seems to have been open so far. A proof with different methods has been announced in [AFK]. Final remarks. It is interesting to compare the critical value of the localization parameter s in higher dimensions with the higher dimensional versions of the Balian- Low theorem. It is known that for d = 1 and every s < 2 there exists a function f with ZR ZR xsf 2 < ∞, ξs f2 < ∞, such that {e2πintf (x − m)}n,m∈Z is an orthogonal basis for L2(R). (A more precise result is obtained in [BCGP], we refer the reader to [Jan] also.) Thus in one-dimensional setting the restrictions on the localization properties of an arbitrary orthogonal basis and of a Gabor system can be observed only at one point of our scale, p = 2. The situation changes drastically when we consider higher dimensional spaces. There exists an orthogonal basis {fn} for L2(Rd) that satisfies but for every orthogonal basis of the form {e2πintf (x − m)}n,m∈Zd a multidimensional version of Balian-Low theorem (see [CP] and references therein) implies n∈N(cid:16) inf a∈RdZRd sup x − a2dfn(x)2 + inf b∈RdZRd a∈RdZRd inf sup n∈N x − a2fn(x)2 = ∞ or sup n∈N ξ − b2dbfn(ξ)2(cid:17) < ∞, b∈RdZRd inf ξ − b2bfn(ξ)2 = ∞. The reason for it could lie in the choice of the radial weight vp = xp. It seems that might be a more natural weight in the product weight ws(x) =(cid:16)(1 + x1)...(1 + xd)(cid:17)s higher dimensions. There are a (p, q)-version of Bourgain's theorem [BP] and a (p, q) version of the Balian- Low theorem [Ga] for 1/p + 1/q = 1. It would be interesting to obtain a (p, q)-version of Theorem 1. 18 KARLHEINZ GR OCHENIG AND EUGENIA MALINNIKOVA Acknowledgment. K. G. would like to thank the Department of Mathematics of NTNU Trondheim for its hospitality and great research environment during the research on this pa- per. Both authors would like to thank the Centre de Recerca Matem`atica Barcelona for ideal conditions to complete this work. References [AFK] G. Ascensi, H. G. Feichtinger, and N. Kaiblinger, Dilation of the Weyl symbol and Balian-Low Theorem, in preparation. [BCHL] R. Balan, P. G. Casazza, C. Heil, and Z. Landau, Density, overcompleteness, and localization of frames. I. Theory, J. Fourier Anal. Appl., 12(2) (2006), 105 -- 143. [Bat] G. Battle, Heisenberg proof of the Balian-Low theorem, Lett. Math. Phys., 15 (1988), no. 2, 175 -- 177. [BCGP] J. Benedetto, W. Czaja, P. Gadzin'ski, and A. Powell, The Balian-Low theorem and regularity of Gabor systems, J. Geom. Anal., 13 (2003), no. 2, 239 -- 254. [BHW] J. Benedetto, Ch. Heil, and D. Walnut, Differentiation and the Balian-Low theorem, J. Fourier Anal. Appl., 1 (1995), no. 4, 355 -- 402 [BP] J. Benedetto and A. Powell, A (p, q) version of Bourgain's theorem, Trans. Amer. Math. Soc., 358 (2006), no. 6, 2489 -- 2505. [B] J. Bourgain, A Remark on the uncertainty principle for Hilbertian basis, J. Funct. Anal., 79 (1988), 136 -- 143. [CG] E. Cordero and K. Grochenig, Time-frequency analysis of localization operators, J. Funct. Anal., 205(1) (2003), 107 -- 131. [CP] W. Czaja and A. Powell, Recent developments in the Balian-Low theorem. Harmonic analysis and [D] applications, 79 -- 100, Appl. Numer. Harmon. Anal., Birkhauser, Boston, MA, 2006. I. Daubechies, The wavelet transform, time-frequency localization and signal analysis, IEEE Trans. Inform. Theory, 36 (1990), no. 5, 961 -- 1005. [FS] G. B. Folland and A. Sittaram, The Uncertainty Principle: A Mathematical survey, J. Fourier Anal. and Appl., 3 (1997), no. 3, 208 -- 238. [FG] M. Fornasier and K. Grochenig, Intrinsic localization of frames, Constr. Approx., 22(3) (2005), 395 -- 415. [Ga] S. Z. Gautam, A critical-exponent Balian-Low theorem, Math. Res. Letters, 15 (2008), no. 3, 471-483. [G1] K. Grochenig, Foundations of Time-Frequency Analysis, Birkhauser, Boston, 2001. [G2] K. Grochenig, Localization of Frames, Banach Frames and the Invertibility of the Frame Operator, J. Fourier Anal. and Appl., 10 (2004), no. 2, 105 -- 132. [GR] K. Grochenig and H. Razafinjatovo, On Landau's necessary density conditions for sampling and interpolation of band-limited functions, J. London Math. Soc. (2), 54(3) (1996), 557 -- 565. [H] C. Heil. History and evolution of the density theorem for Gabor frames, J. Fourier Anal. Appl. , 13(2) (2007), 113 -- 166. [Jaf] S. Jaffard, Propri´et´es des matrices "bien localis´ees" pr`es de leur diagonale et quelques applications, Ann. Inst. H. Poincar´e Anal. Non Lin´eaire, 7 (1990), no. 5, 461 -- 476. [JP] P. Jaming and A. M. Powell, Time-frequency concentration of generating systems, Proc. Amer. Math. Soc., to appear. [Jan] A. J. E. M Janssen, A decay result for certain windows generating orthogonal Gabor bases, J. Fourier Anal. Appl. 14 (2008), no. 1, 1 -- 15. [LP] H. J. Landau and H. O. Pollak, Prolate spheroidal wave functions, Fourier analysis and uncertainty. III. The dimension of the space of essentially time- and band-limited signals, Bell System Tech. J., 41 (1962) 1295 -- 1336. [L] H. J. Landau, Necessary density conditions for sampling and interpolation of certain entire func- tions, Acta Math., 117 (1967), 37 -- 52. LOCALIZATION PROPERTIES OF BASES FOR L2(Rd) 19 [RS] J. Ramanathan and T. Steger, Incompleteness of sparse coherent states, Appl. Comput. Har- mon. Anal., 2 (1995), no. 2, 148 -- 153. [SP] D. Slepian and H. O. Pollak, Prolate spheroidal wave functions, Fourier analysis and uncertainty, I, Bell System Tech. J., 40 (1961), 43 -- 64. [S] D. Slepian, Prolate spheroidal wave functions, Fourier analysis and uncertainity. IV. Extensions to many dimensions; generalized prolate spheroidal functions. Bell System Tech. J., 43 (1964) 3009 -- 3057. [W] M. W. Wong, Localization operators, Seoul National University Research Institute of Mathematics Global Analysis Research Center, Seoul, 1999. Faculty of Mathematics, University of Vienna, Nordbergstrasse 15, A-1090 Vienna, Austria E-mail address: [email protected] Department of Mathematics, Norwegian University of Science and Technology, 7491, Trondheim, Norway E-mail address: [email protected]
1003.1329	1	1003	2010-03-05T19:40:04	Quantization for an elliptic equation of order 2m with critical exponential non-linearity	[ "math.FA", "math.AP" ]	On a smoothly bounded domain $\Omega\subset\R{2m}$ we consider a sequence of positive solutions $u_k\stackrel{w}{\rightharpoondown} 0$ in $H^m(\Omega)$ to the equation $(-\Delta)^m u_k=\lambda_k u_k e^{mu_k^2}$ subject to Dirichlet boundary conditions, where $0<\lambda_k\to 0$. Assuming that $$\Lambda:=\lim_{k\to\infty}\int_\Omega u_k(-\Delta)^m u_k dx<\infty,$$ we prove that $\Lambda$ is an integer multiple of $\Lambda_1:=(2m-1)!\vol(S^{2m})$, the total $Q$-curvature of the standard $2m$-dimensional sphere.	math.FA	math	Quantization for an elliptic equation of order 2m with critical exponential non-linearity Luca Martinazzi∗ Centro De Giorgi, Pisa Michael Struwe ETH Zurich [email protected] [email protected] November 18, 2018 Abstract On a smoothly bounded domain Ω ⊂ R2m we consider a sequence ⇁ 0 in H m(Ω) to the equation (−∆)muk = k subject to Dirichlet boundary conditions, where 0 < λk → 0. w of positive solutions uk λkukemu2 Assuming that Λ := lim k→∞ZΩ uk(−∆)m ukdx < ∞, we prove that Λ is an integer multiple of Λ1 := (2m − 1)! vol(S 2m), the total Q-curvature of the standard 2m-dimensional sphere. 1 Introduction Given a smoothly bounded domain Ω ⊂ R2m, suppose that for each k ∈ N we have a smooth function uk > 0 satisfying the equation (−∆)muk = λkukemu2 k in Ω (1) with uk = ∂ν uk = . . . = ∂m−1 (2) where 0 < λk → 0 as k → ∞. We assume that (uk) is bounded in H m(Ω). Hence, after passing to a subsequence and integrating by parts we may assume that as k → ∞ we have uk = 0 on ∂Ω, ν ZΩ ∇muk2dx = ZΩ uk(−∆)mukdx = λkZΩ kemu2 u2 k dx → Λ > 0. (3) Note that by elliptic estimates the quantity kuk := (cid:0)ZΩ ∇muk2dx(cid:1)1/2 = (cid:0)ZΩ Xα=m ∂αuk2dx(cid:1)1/2 defines a norm on the Beppo-Levi space H m standard Sobolev norm. 0 (Ω) which is equivalent to the ∗The first author was supported by the ETH Research Grant no. ETH-02 08-2 and by the Italian FIRB Ideas "Analysis and beyond". 1 Generalising previous results by Adimurthi and Struwe [3], Adimurthi and Druet [1] and Robert and Struwe [11], the first author proved in [8] the following theorem. Theorem 1 Let (uk) be a sequence of positive solutions to (1), (2) with 0 < λk → 0 as k → ∞ and satisfying (3) for some Λ > 0. Then supΩ uk → ∞ as k → ∞ and there exist a subsequence (uk) and sequences of points x(i) k → x(i) ∈ Ω, 1 ≤ i ≤ I, for some integer I ≤ CΛ, such that the following is true. For every 1 ≤ i ≤ I, letting r(i) k(x(i) k )2mu2 λk(r(i) k > 0 be given by k )emu2 k(x(i) k ) = 22m(2m − 1)! (4) (5) (6) and setting η(i) k (x) := uk(x(i) k )(uk(x(i) k )) + log 2, k + r(i) k x) − uk(x(i) k → ∞ as k → ∞, and 1 + x2 in C2m−1 loc 2 (R2m) as k → ∞. we have r(i) k , ∂Ω)/r(i) k → 0, dist(x(i) η(i) k (x) → η0(x) = log Moreover, for i 6= j there holds k − x(j) x(i) k r(i) k → ∞ as k → ∞. such that there holds In addition, with Rk(x) := inf 1≤i≤I x − x(i) k(x)emu2 k (x)u2 λkR2m k there exists a constant C > 0 k(x) ≤ C (7) uniformly for all x ∈ Ω, k ∈ N. Finally uk → 0 in C2m−1 We remark that the function η0 given by (5) satisfies the Q-curvature equa- (Ω\S), where S = {x(1), . . . , x(I)}. loc tion and (−∆)mη0 = (2m − 1)!e2mη0 (8) (2m − 1)!ZR2m e2mη0dx = ZS2m QS2mdvolgS2m = (2m − 1)!S2m =: Λ1. (9) For a discussion of the geometric meaning of (8) we refer to [4] or to the intro- duction of [7]. The purpose of this paper is to prove the following quantization result. Theorem 2 Under the hypothesis of Theorem 1 we have Λ = I ∗Λ1 for some I ∗ ∈ N\{0}. 2 The analogue of Theorem 2 was proven by O. Druet [5] in dimension 2 (m = 1) and by the second author [13] in dimension 4 (m = 2) in the case of the Navier boundary condition uk = ∆uk = 0 on ∂Ω. Note that in the latter case the maximum principle implies that ∆uk ≤ 0 in Ω whereas such an estimate is not available in the case of the Dirichlet boundary condition. Quantization results similar to Theorem 2 previously have also been obtained for concentrating sequences of solutions uk to the Q-curvature equation (−∆)muk = λke2muk in Ω ⊂ R2m. (10) In the case of the Navier boundary condition, assuming that λk → 0 and Λ := lim k→∞ZΩ λke2muk dx < ∞, J. Wei [14] proved that when m = 2 and when Ω is convex the quantity Λ is an integer multiple of Λ1. Moreover concentration points are simple and isolated, in the sense that x(i) 6= x(j) for i 6= j, and I ∗ = I in the notation of Theorems 1 and 2 above. Robert and Wei [12] proved the analogous result for a general domain Ω and in the case of Dirichlet boundary conditions. In [9], the first author and Petrache generalized the result of Robert and Wei to arbitrary dimensions. Equation (1) is more difficult to deal with analytically than equation (10); the analogous questions whether for a blowing up sequence of solutions to (1) the concentration points are isolated, simple and stay away from the boundary are still open, even in dimension 2. Our paper is organized as follows. In the next section we present the proof of Theorem 2 in the case when Ω = BR is a ball and each function uk is radially symmetric. In Section 3 we prove the theorem in the general case. Some useful technical results are collected in the Appendix. The overall strategy of the proof is very similar to the approach followed in [13], and some of the results in [13] can be carried over almost literally to the present setting. Several key steps in the proof, however, require conceptually new ideas in the case when m ≥ 3. These ideas also shed new light on the previous approaches in low dimensions and have a unifying feature. Throughout the paper the letter C denotes a generic constant independent of k which can change from line to line, or even within the same line. 2 Proof of Theorem 2 in the radial case Let Ω = BR = BR(0) and assume that each uk is radially symmetric. By slight abuse of notation we write uk(x) = uk(r) if x = r. In the notation of Theorem 1 we then have I = 1 and we can choose x(1) k = 0 for every k > 0. In fact, as shown in assertion (17) of Lemma 4 below, we have uk(0) = maxΩ uk. 2.1 Strategy of the proof Set ek := λku2 k and let kemu2 Λk(r) := ZBr ekdx, Nk(s, t) := Λk(t) − Λk(s) = ZBt\Bs ekdx 3 as in [13]. We shall say that the property (Hℓ) is satisfied if there exist sequences s(0) k := 0 < r(1) k < s(1) k < . . . < r(ℓ) k < s(ℓ) k ≤ R, k ∈ N, such that the following holds: (Hℓ,1) limk→∞ r(j) k s(j) k = limk→∞ s(j−1) k r(j) k = 0 for 1 ≤ j ≤ ℓ, (Hℓ,2) limk→∞ uk(s(j) k ) uk(Lr(j) k ) = 0 for 1 ≤ j ≤ ℓ, L > 0, , r(j) k /L) + Nk(Lr(j) k , s(j) k )(cid:1) = 0 for 1 ≤ j ≤ ℓ. (Hℓ,3) limk→∞ Λk(s(j) (Hℓ,4) limL→∞ limk→∞(cid:0)Nk(s(j−1) k k ) = jΛ1 for 1 ≤ j ≤ ℓ, For the proof of Theorem 2 we proceed via induction from the following two claims: (H1) holds, and if (Hℓ) holds then either (Hℓ+1) holds as well, or lim k→∞ Nk(s(ℓ) k , R) = 0. (11) By (3) and (Hℓ,3) the induction terminates when ℓ > Λ Λ1 largest integer such that (Hℓ0 ) holds, (Hℓ0,3) and (11) imply . Letting ℓ0 be the Λ = lim k→∞ Λk(s(ℓ0) k ) + lim k→∞ Nk(s(ℓ0) k , R) = ℓ0Λ1, and Theorem 2 in the radial case follows. 2.2 Proof of (H1) Let rk > 0 be defined as in Theorem 1 such that λkr2m k u2 k(0)emu2 k(0) = 22m(2m − 1)!, and set We have wk(x) := uk(0)(uk(x) − uk(0)) in BR. (−∆)mwk = λkuk(0)ukemu2 = λkuk(0)ukemu2 k k(0)e 2m(cid:0)1+ wk 2u2 k (0)(cid:1)wk =: fk in BR. Letting also σk(r) := ZBr fkdx ≥ Λk(r), then by (5) of Theorem 1 and (9) clearly we have lim L→∞ lim k→∞ Λk(Lrk) = lim L→∞ lim k→∞ = lim L→∞ lim k→∞ 4 σk(Lrk) (2m − 1)!ZBL e2mηk dx = Λ1. (12) For 0 < t ≤ R let gk solve the equation ∆mgk = ∆mwk in Bt with homogeneous Dirichlet boundary data gk = ∂ν gk = . . . = ∂m−1 ν gk = 0 on ∂Bt. Then Lemma 22 in the Appendix gives the identity (−1)m∂m ν gk(t) = Ak(t) ω2m−1t3m−2 (13) similar to (20) in [13], where Ak(t) := Z t 0 t2 ···Z tm−1 0 tmσk(tm)dtm . . . dt2, (14) and where ω2m−1 is the (2m − 1)-dimensional volume of S2m−1. Lemma 3 For every b < 2 we can find L = L(b) and k0 = k0(b) such that for k ≥ k0 we have (−∂ν)mgk(t) ≥ Proof. Noting that 2m−1(m − 1)!b tm for Lrk ≤ t ≤ R. (15) Λ1 ω2m−12m−1(m − 1)! = 2m(m − 1)!, from (12) and (13) together with the identity Z t 0 t2 ···Z tm−1 0 tmdtm ··· dt2 = t2m−2 2m−1(m − 1)! we obtain the claim. (cid:3) These estimates now yield the following result analogous to Lemma 2.1 in [13]. Note, however, that the statement (17) below in the present case no longer can simply be deduced from the maximum principle, as was the case in [13]. In addition, the higher order nature of equation (1) requires substantial technical modifications of the approach used in [13]. Lemma 4 For any b < 2 there is L = L(b) and k0 = k0(b) such that for k ≥ k0 there holds b t + tP (t) w′ k(t) ≤ − w′ k(t) ≤ 0 wk(t) ≤ b log(cid:16) rk in BR\BLrk, t (cid:17) + C in BR, in BR, (16) (17) (18) where P is a polynomial independent of k. In particular uk is monotone de- creasing. For any ε ∈]0, 1[ let Tk > 0 be such that uk(Tk) = εuk(0). Then we have (19) = 0 lim k→∞ rk Tk 5 and lim k→∞ Λk(Tk) = lim k→∞ σk(Tk) = Λ1. (20) Proof. Fix t > 0 and write wk = gk + hk, where ∆mhk = 0 in Bt, and gk = ∂ν gk = . . . = ∂m−1 ν gk = 0 on ∂Bt. Step 1. We claim that ∂m ν gk(t) = tm−1 (t−1(t−1 ··· (t−1 } m−1 times {z w′ k(t) )′ ··· )′)′ {z } m−1 times . (21) Indeed, subtracting ∂m ν wk(t) from both sides of (21) we need to show −∂m ν hk(t) = tm−1 (t−1(t−1 ··· (t−1 } m−1 times {z w′ k(t) )′ ··· )′)′ {z } m−1 times −∂m ν wk(t). Using the boundary condition ∂j observing that on the right-hand side the terms involving ∂m can replace wk by hk and it suffices to prove ν hk(t) for 0 ≤ j ≤ m − 1, and ν wk(t) cancel, we νwk(t) = ∂j −∂m ν hk(t) = tm−1 (t−1(t−1 ··· (t−1 } m−1 times {z k(t) )′ ··· )′)′ h′ {z } m−1 times −∂m ν hk(t). But ∆mhk = 0 and radial symmetry imply that h(r) = Pm−1 i=0 αir2i; so (t−1(t−1 ··· (t−1 } m−1 times {z h′ k(t) )′ ··· )′)′ {z } m−1 times = 0, and (21) follows. Step 2. Inserting now (15) into (21), for any given b < 2 we infer (−1)m−1tm−1(cid:16)t−1(cid:16)t−1 ···(cid:16)t−1 } (cid:16)w′ {z 2m−1(m − 1)!b = (−1)m−1∂m ν gk(t) + m−1 times tm k(t) + b t(cid:17)(cid:17)′ ···(cid:17)′(cid:17)′ {z } m−1 times ≤ 0 for Lrk ≤ t ≤ R, (22) provided that we fix L = L(b) sufficiently large and then also choose k large enough. We now prove by induction over 1 ≤ j ≤ m that t(cid:17)(cid:17)′ ϕj,k(t) := (−1)m−jt−1(cid:16)t−1(cid:16)t−1 ···(cid:16)t−1 } ···(cid:17)′(cid:17)′ {z } ≤ Pj(t), (cid:16)w′ k(t) + m−j times m−j times {z (23) b for Lrk ≤ t ≤ R, where Pj(t) ≥ 0 is a polynomial in t independent of k. The case j = 1 follows at once from (22) with P1 ≡ 0. Using the Dirichlet boundary condition (which implies ∂j νwk(R) = 0 for 1 ≤ j ≤ m − 1) we get ϕj,k(R) ≤ Cj 6 for some constant Cj ≥ 0, 2 ≤ j ≤ m. Observing that ϕ′ 2 ≤ j ≤ m, we then obtain j,k(t) = −tϕj−1,k(t) for ϕj,k(t) = ϕj,k(R) +Z R t rϕj−1,k(r)dr ≤ Cj +Z R t rPj−1(r)dr =: Pj(t), that is, (23). For j = m we get w′ k(t) ≤ − b t + tPm(t), Lrk ≤ t ≤ R. Integrating once more, recalling that L depends on b, and that wk(Lrk) → η0(L) as k → ∞, for sufficiently large k we find wk(t) ≤ wk(Lrk) − b log(cid:16) t Lrk(cid:17) + C ≤ b log(cid:16) rk t (cid:17) + C for Lrk ≤ t ≤ R. For 0 < t < Lrk (18) already follows from Theorem 1. In order to prove (17), observe that (13) implies (−∂ν)mgk(t) ≥ 0 for 0 < t ≤ R, k ∈ N, and (21) yields (−1)m−1tm−1 (t−1(t−1 ··· (t−1 } m−1 times {z w′ k(t) )′ ··· )′)′ {z } m−1 times ≤ 0. (24) In analogy with (23), for 1 ≤ j ≤ m we can show by induction that ψj,k(t) := (−1)m−jt−1 (t−1(t−1 ··· (t−1 } w′ k(t) )′ ··· )′)′ {z } m−j times m−j times {z ≤ 0 for all 0 < t ≤ R. Indeed ψ1,k(t) ≤ 0 by (24), while for 2 ≤ j ≤ m, we have ψj,k(R) = 0 thanks to the boundary condition. Hence ψj,k(t) = Z R t rψj−1(r)dr ≤ 0 for all 0 < t ≤ R, and the case j = m implies (17). Step 3. In order to prove (19), assume by contradiction that lim inf k→∞ Tk rk = L ∈ [0,∞[. Then from Theorem 1 for a suitable subsequence on the one hand we have uk(0)(uk(Tk) − uk(0)) + log 2 = ηk(cid:18) Tk rk(cid:19) → log(cid:18) 2 But on the other hand, since uk(0) → ∞ we also have that 1 + L2(cid:19) as k → ∞. uk(0)(uk(Tk) − uk(0)) = u2 k(0)(ε − 1) → −∞ 7 as k → ∞, a contradiction. It thus remains to prove (20). Using (18) and observing that k(0) ≤ wk(r) ≤ 0 for 0 ≤ r ≤ Tk, (ε − 1)u2 from (4) for k ≥ k0 we get k(0)emu2 k u2 fk(r) ≤ λku2 ≤ λkr2m k(0)e k(0)emu2 2m(cid:0)1+ wk (r) 2u2 k (0)(cid:1)wk(r) em(ε+1)wk(r) ≤ Cr−2m k k(0)r−2m k (cid:16) rk r (cid:17)m(ε+1)b . Choosing now b < 2 such that m(ε + 1)b = 2m + ε, and integrating over BTk , we find Λ1 ≤ lim k→∞ Λk(Tk) ≤ lim k→∞ σk(Tk) = Λ1 + lim L→∞ lim k→∞ZBTk \BLrk fkdx ≤ Λ1 + C lim L→∞ lim k→∞ ≤ Λ1 + C ε lim L→∞ 1 r2m k ZBTk \BLrk (cid:16) rk Lrk(cid:17)ε = Λ1, lim k→∞(cid:16) rk r (cid:17)2m+ε dx hence (20). (cid:3) According to Lemma 4 we can now choose a sequence εk → 0 as k → ∞ and corresponding numbers sk = Tk(εk) such that uk(sk) → ∞ as k → ∞ and lim k→∞ rk sk = 0, lim k→∞ Λk(sk) = Λ1, lim L→∞ lim k→∞ Nk(Lrk, sk) = 0. (25) Observing that Theorem 1 implies limk→∞ uk(Lrk) uk(0) = 1 for every L ≥ 0, we get lim k→∞ uk(sk) uk(Lrk) = lim k→∞ uk(sk) uk(0) = 0, for all L > 0. (26) We also claim that lim L→∞ lim k→∞ Nk(sk, Lsk) = 0. (27) To see this, remember that for 0 < s < t < R Nk(s, t) = ZBt\Bs ekdx = ω2m−1Z t s λkr2m−1u2 kemu2 k dr. Now set Pk(t) := t ∂ ∂t Nk(s, t) = tZ∂Bt ekdσ = ω2m−1λkt2mu2 k(t)emu2 k(t). Using the monotonicity of uk that we proved in Lemma 4 we immediately obtain the estimate Pk(t) = Cω2m−1λku2 k(t)emu2 k(t)Z t t/2 r2m−1dr ≤ CNk(t/2, t) ≤ CPk(t/2) (28) 8 analogous to (26) in [13]; hence we also conclude that Nk(t, 2t) ≤ CNk(t/2, t) for t ∈ [0, R/2]. (29) Now (25) and (29) imply that for any M ∈ N Nk(2M−1sk, 2M sk) ≤ C lim lim k→∞ k→∞ Nk(2M−2sk, sM−1sk) ≤ ··· ≤ CM lim k→∞ Nk(sk/2, sk) = 0. Therefore if 2M ≥ L we have lim k→∞ Nk(sk, Lsk) ≤ M Xj=1 Nk(2j−1sk, 2jsk) = 0, as claimed. Setting r(1) k := rk, s(1) k 1 we see that the property (H1) is satisfied. := sk and taking into account (25) - (27) and Theorem 2.3 The inductive step We now assume that (Hℓ) holds for some integer ℓ ≥ 1 and fix numbers s(0) k = 0 < r(1) k < s(1) k < . . . < r(ℓ) k < s(ℓ) k , k ∈ N such that (Hℓ,1), (Hℓ,2), (Hℓ,3) and (Hℓ,4) hold true. To complete the proof of Theorem 2 it suffices to show that either (Hℓ+1) or (11) holds. The proof requires the following analogue of (29) in [13]. Lemma 5 There is a constant C0 = C0(Λ) such that for tk > s(ℓ) k there holds Pk(tk) m + C0N 2 k (s(ℓ) k , tk) + o(1), (30) Nk(s(ℓ) k , tk) ≤ with error o(1) → 0 as k → ∞ Proof. For s = s(ℓ) k < t we integrate by parts to obtain r2m−1λku2 kemu2 k dr s Nk(s, t) = ω2m−1Z t 2m (cid:0)r2mu2 kemu2 2m − ω2m−1Z t ω2m−1 Pk(t) = λk ≤ s t k(cid:1)(cid:12)(cid:12) s − λkr2m(cid:16) 1 m ω2m−1 2m Z t k(cid:17) uk + u2 s λkr2m(2uk + 2mu3 k)u′ kemu2 k dr uk(0) kemu2 w′ k dr. (31) Define gk(t) as in the beginning of the proof of Lemma 4. Then (13) and (21) imply (t−1(t−1 ··· (t−1 } m−1 times {z = (−1)m Ak(t) ω2m−1t4m−3 , k(t) )′ ··· )′)′ w′ {z } m−1 times 9 where Ak is as in (14). Integrating this relation m − 1 times from t to R, and using the Dirichlet boundary condition ∂j νwk(R) = 0 for 1 ≤ j ≤ m − 1 we get w′ k(t) t = −Z R t t1Z R t1 t2 ···Z R tm−2 Ak(tm−1) ω2m−1t4m−3 m−1 dtm−1 ··· dt1; hence −tw′ k(t) uk(t) uk(0) = t2 uk(t) uk(0) Z R t t1Z R t1 t2 ···Z R tm−2 Ak(tm−1) ω2m−1t4m−3 m−1 dtm−1 ··· dt1 =: I. More explicitly, I = t2Z R ×Z tm−1 t 0 t1 t2 ···Z R t1Z R ρ1 ···Z ρm−2 0 tm−2 1 ω2m−1t4m−3 m−1 ρm−1τk(ρm−1, t)dρm−1 ··· dρ1 dtm−1 ··· dt1, where τk(ρ, t) = uk(t)σk(ρ) uk(0) = ZBρ λkuk(t)ukemu2 k dx. We now show that I can be bounded in terms of Nk(s, t) up to a small error. From this the desired inequality (30) will be immediate. Split I =: II + III, where II corresponds to ρm−1 ≤ t. Since u′ k ≤ 0, for ρ ≤ t we have τk(ρ, t) = ZBρ ≤ ZBs λkuk(t)ukemu2 λkuk(s)ukemu2 λkuk(ρ)ukemu2 k dx ≤ ZBρ k dx + Nk(s, ρ) ≤ Nk(s, t) + o(1) k dx (32) with error o(1) → 0 as k → ∞. Here we used that for arbitrary L > 1 we can bound ZBs λkuk(s)ukemu2 k dx ≤ Nk(Lr(ℓ) k , s) + uk(s) uk(Lr(ℓ) k ) Λk(Lr(ℓ) k ), and by (Hℓ,2), (Hℓ,4) the latter tends to 0, if first k → ∞ and then L → ∞. Since t2Z ∞ ρm−1dρm−1 ··· dt1 ≤ C ρ1 ···Z ρm−2 t1 ···Z ∞ m−1 Z tm−1 ω2m−1t4m−3 tm−2 1 0 0 t uniformly in t, we conclude that II ≤ CNk(s, t) + o(1). In order to obtain a similar bound for III, for t ≤ ρ we estimate τk(ρ, t) = uk(t)σk(ρ) uk(0) = uk(t) uk(ρ) + 1 ZBρ λk(uk(ρ) + 1)ukemu2 k dx . 10 Recalling (32), we have ZBρ λk(uk(ρ) + 1)ukemu2 k dx ≤ τk(ρ, ρ) + o(1) ≤ Nk(s, ρ) + o(1). Also note that by Holder's inequality we can estimate uk(t) − uk(ρ) ≤ Z ρ t k(r)dr ≤ k∇ukkL2m(cid:0) log u′ Thus, with a constant C = C(Λ) for all t ≤ ρ we obtain ρ 2m . t(cid:1) 2m−1 tuk(t) ρ(uk(ρ) + 1) = ≤ t uk(ρ) + 1 ρ(cid:18) uk(t) − uk(ρ) ρ(cid:18)C(cid:0) log t(cid:1) 2m−1 ρ t + uk(ρ) uk(ρ) + 1(cid:19) 2m + 1(cid:19) ≤ C (33) and with C1 = C1(Λ) we can bound t ρ τk(ρ, t) ≤ C1Nk(s, ρ) + o(1). It follows that 1 ω2m−1t4m−3 m−1 1 ω2m−1t4m−3 m−1 t t t t1 tm−2 III = t2Z R ×Z tm−1 = tZ R t1Z R ×Z tm−1 ≤ C1tZ R ×Z tm−1 t2 ···Z R t1Z R ρ1 ···Z ρm−2 t2 ···Z R ρ1 ···Z ρm−2 t2 ···Z R t1Z R ρ1 ···Z ρm−2 tm−2 t1 t1 t t t t t t ρm−1τk(ρm−1, t)dρm−1 ··· dρ1 dtm−1 ··· dt1 ρ2 m−1 τk(ρm−1, t)dρm−1 ··· dρ1 dtm−1 ··· dt1 t ρm−1 1 tm−2 ω2m−1t4m−3 m−1 ρ2 m−1(Nk(s, ρm−1) + o(1))dρm−1 ··· dt1. For any L ≥ 1 we split the integral with respect to t1 and use the obvious inequality Nk(s, ρm−1) ≤ 2Λ for large k to estimate 1 t3 ···Z R tm−2 ω2m−1t4m−3 m−1 t t1 t2 III ≤ C1tZ Lt ×Z tm−1 + 2C1ΛtZ R ×Z tm−1 t1Z R t2Z R ρ1 ···Z ρm−2 t1Z R t2Z R ρ1 ···Z ρm−2 Lt t2 t1 t t t t ρ2 m−1(Nk(s, ρm−1) + o(1))dρm−1 ··· dt1 t3 ···Z R ρ2 m−1dρm−1 ··· dt1 + o(1) ω2m−1t4m−3 m−1 tm−2 1 11 Observing the uniform bound LtZ ∞ Lt t1 ···Z ∞ tm−2 1 m−1 Z tm−1 0 ω2m−1t4m−3 ρ1 ···Z ρm−2 0 ρ2 m−1dρm−1 ··· dt1 ≤ C, with a constant C2 = C2(Λ) we obtain III ≤ C1tZ Lt ×Z tm−1 t1Z R t2Z R ρ1 ···Z ρm−2 t1 t2 t t t C2Λ L + + o(1) t3 ···Z R tm−2 1 ω2m−1t4m−3 m−1 ρ2 m−1(Nk(s, ρm−1) + o(1))dρm−1 ··· dt1 To proceed we successively split the integral also with respect to t2, . . . , tm−1 and use the uniform bounds LtZ Lt 0 t1 ···Z Lt 0 tj−1Z ∞ Lt tj ···Z ∞ ρ1 ···Z ρm−2 0 tm−2 ×Z tm−1 0 1 ω2m−1t4m−3 m−1 ρ2 m−1dρm−1 ··· dtj ≤ C t t t t t1 t2 Lt III ≤ C1tZ Lt ×Z tm−1 + 2C1ΛtZ Lt ×Z tm−1 for 2 ≤ j < m to estimate t1Z Lt t2Z R ρ1 ···Z ρm−2 t2Z R t1Z R ρ1 ···Z ρm−2 t1Z Lt ρ1 ···Z ρm−2 t1Z Lt t2Z Lt ρ1 ···Z ρm−2 ≤ ··· ≤ C1tZ Lt ×Z tm−1 + 2C1ΛtZ Lt ×Z tm−1 t2 t2 t1 t1 t2 t t t t t t t t2Z Lt ≤ CmNk(s, Lt) + + o(1), t CmΛ L t3 ···Z R tm−2 1 ω2m−1t4m−3 m−1 1 L tm−2 + o(1) ω2m−1t4m−3 m−1 C2Λ ρ2 m−1(Nk(s, ρm−1) + o(1))dρm−1 ··· dt1 t3 ···Z R ρ2 m−1dρm−1 ··· dt1 + t3 ···Z Lt m−1(cid:0)Nk(s, Lt) + o(1)(cid:1)dρm−1 ··· dt1 t3 ···Z R ρ2 m−1dρm−1 ··· dt1 + ω2m−1t4m−3 m−1 ω2m−1t4m−3 m−1 Cm−1Λ + o(1) tm−2 ρ2 L Lt 1 1 with constants Cj = Cj (Λ), 2 ≤ j ≤ m. Using (27) in case t ≤ 2s and (28) in case t > 2s we get Nk(s, Lt) ≤ C(L)Nk(s, t) + o(1), 12 and with the constant Cm+1 = CmΛ = Cm+1(Λ) there results −tw′ k(t) uk(t) uk(0) ≤ C(L, Λ)Nk(s, t) + Cm+1 L + o(1). Inserting this into (31) we infer Nk(s, t) ≤ Pk(t) 2m − ω2m−1Z t λkr2m(cid:16) 1 +(cid:16)C(L, Λ)Nk(s, t) + s m Pk(t) 2m ≤ + u2 k dr uk(0) k(cid:17) uk kemu2 w′ L (cid:17)Nk(s, t) + o(1). Cm+1 (34) Choosing L = 2Cm+1 we finally get (30) for an appropriate C0 = C0(Λ). (cid:3) Lemma 6 Let C0 = C0(Λ) be the constant appearing in (30). tk ∈]s(ℓ) k , R] there holds If for some 0 < lim k→∞ Nk(s(ℓ) k , tk) =: α < 1 2C0 , (35) then s(ℓ) k tk lim k→∞ = 0, lim inf k→∞ Pk(tk) ≥ Proof. Assume that for some tk ∈]s(ℓ) as in the proof of (27) also yields that 2 mα , and lim L→∞ lim k→∞ Nk(s(ℓ) k , tk/L) = 0. k , R] we have (35). Since the same reasoning lim L→∞ lim k→∞ Nk(s(ℓ) k , Ls(ℓ) k ) = 0, necessarily s(ℓ) Pk(tk) lim inf k→∞ k /tk → 0 as k → ∞. Moreover, (30) yields k→∞(cid:16)Nk(s(ℓ) k , tk) − C0N 2 Nk(s(ℓ) k , tk) ≥ lim 2 m ≥ lim α 2 k→∞ = , k (s(ℓ) k , tk)(cid:17) as claimed. Now we show that Indeed, if we assume lim L→∞ lim k→∞ Nk(s(ℓ) k , tk/L) = 0. lim L→∞ lim sup k→∞ Nk(s(ℓ) k , tk/L) = β > 0, we have β 2 ≤ Nk(s(ℓ) k , tk/L) ≤ Nk(s(ℓ) k , tk) < 1 2C0 (36) (37) for any L ≥ 1 and sufficiently large k. Therefore we can apply (36) with tk/L instead of tk for any L ≥ 1 to get lim k→∞ Pk(tk/L) ≥ mβ 2 . 13 Then (28) yields C lim k→∞ Nk(tk/(2L), tk/L) ≥ lim k→∞ Pk(tk/L) ≥ mβ 2 . Choosing L = 2j and summing over j for 0 ≤ j ≤ M − 1, we get C lim k→∞ Nk(2−M tk, tk) ≥ which contradicts (3). Therefore (37) is proven. Λk(tk) ≥ C lim k→∞ mM β 2 → ∞ as M → ∞, (cid:3) Suppose now that for some tk ≥ s(ℓ) k there holds lim sup k→∞ Nk(s(ℓ) k , tk) > 0. We then want to show that (Hℓ+1) holds. We can choose numbers rℓ+1 such that for a subsequence there holds k ∈]s(ℓ) k , tk[ 0 < lim k→∞ Nk(s(ℓ) k , r(ℓ+1) k ) < 1 2C0 , (38) where C0 is as in Lemma 6. Observe that Lemma 6 then implies lim k→∞ s(ℓ) k /r(ℓ+1) k = lim L→∞ lim k→∞ Nk(s(ℓ) k , r(ℓ+1) k /L) = 0, (39) and lim k→∞ Pk(r(ℓ+1) k ) > 0. (40) Proposition 7 We have η(ℓ+1) k (x) := uk(r(ℓ+1) k )(cid:0)uk(r(ℓ+1) k x) − uk(r(ℓ+1) k )(cid:1) → η(ℓ+1) in C2m−1 loc η(ℓ+1) 0 (R2m\{0}). Moreover, for a suitable constant c(ℓ+1) the function := η(ℓ+1) + c(ℓ+1) satisfies (−∆)mη(ℓ+1) 0 = (2m − 1)!e2mη(ℓ+1) 0 , ZR2m (2m − 1)!e2mη(ℓ+1) 0 dx = Λ1. The above proposition, which will be proven in the following section, implies that hence (39) yields lim L→∞ lim k→∞ Nk(r(ℓ+1) k /L, Lr(ℓ+1) k ) = Λ1; lim L→∞ lim k→∞ Nk(s(ℓ) = lim L→∞ lim k→∞ ) k , Lr(ℓ+1) k Nk(s(ℓ) k , r(ℓ+1) k /L) + lim L→∞ lim k→∞ Nk(r(ℓ+1) k /L, Lr(ℓ+1) k ) (41) = 0 + Λ1 = Λ1. Then the inductive hypothesis (Hℓ,3) gives lim L→∞ lim k→∞ Λk(Lr(ℓ+1) k ) = lim L→∞ lim k→∞(cid:16)Λk(s(ℓ) k ) + Nk(s(ℓ) k , Lr(ℓ+1) k )(cid:17) = (ℓ + 1)Λ1. 14 Now set w(ℓ+1) k (x) = uk(r(ℓ+1) k (−∆)mw(ℓ+1) k )(uk(x) − uk(r(ℓ+1) = λkuk(r(ℓ+1) k )ukemu2 k )) so that k =: f (ℓ+1) . k Similar to Lemma 4 and with the same proof (except that instead of Theorem 1 one needs to use Proposition 7) we have Lemma 8 For any 0 < ε < 1, letting T (ℓ+1) εuk(r(ℓ+1) ), we have k k (ε) > 0 be such that uk(T (ℓ+1) k ) = lim k→∞ Nk(s(ℓ) k , T (ℓ+1) k ) = Λ1. (42) Moreover r(ℓ+1) k /T (ℓ+1) k → 0 as k → ∞. According to Lemma 8 and (41) we can choose numbers εk → 0 and a ) → ∞ as k → ∞ (εk) we have uk(s(ℓ+1) subsequence so that for s(ℓ+1) and := T (ℓ+1) k k k lim k→∞ r(ℓ+1) k s(ℓ+1) k = 0, while lim k→∞ Λk(s(ℓ+1) k ) = (ℓ + 1)Λ1, lim L→∞ lim k→∞ Nk(Lr(ℓ+1) k , s(ℓ+1) k ) = 0. Again reasoning as in the proof of (27) we also infer lim k→∞ Nk(s(ℓ+1) k , Ls(ℓ+1) k ) = 0 for every L ≥ 1. Finally, observe that the definition of s(ℓ+1) k implies that lim k→∞ uk(s(ℓ+1) ) uk(Lr(ℓ+1) k k ) = 0 for every L ≥ 0. Together with (39) this completes the proof of (Hℓ+1), and hence of Theorem 2 in the radially symmetric case. 2.4 Proof of Proposition 7 As preparation for the proof of Proposition 7 we need the following two lemmas. Lemma 9 For r(ℓ+1) k as above, we have k vk(x) := uk(cid:0)r(ℓ+1) Proof. We write rk = r(ℓ+1) . Moreover, we consider only the case m > 1, the case m = 1 being considerably easier. As in the proof of Lemma 3.2 in [13] we have x(cid:1) − uk(cid:0)r(ℓ+1) (R2m\{0}). (cid:1) → 0 in C2m−1 loc k k (−∆)mvk(x) = λkr2m k uk(rkx)emu2 k(rkx) =: gk(x) ≥ 0, 15 with gk → 0 in L∞ have loc(R2m\{0}). By scaling and Sobolev's embedding we also k∇2vkkLm(BR/rk k∇mvkkL2(BR/rk ) = k∇2ukkLm(BR) ≤ C, ) = k∇mukkL2(BR) ≤ C. (43) loc loc Set wk := ∆vk. Then a subsequence wk → w weakly in H m−2 (R2m) and in C2m−3,α (R2m \ {0}) for some function w ∈ Lm(R2m) with ∇m−2w ∈ L2(R2m). Clearly ∆m−1w = 0 in R2m\{0}. In fact, since the point x = 0 has vanishing H m-capacity, as in [13] we have ∆m−1w = 0 in R2m. Recalling that w ∈ Lm(R2m) we conclude that w ≡ 0; see Lemmas 23 and 24 in the appendix. Recalling that (∆vk) is bounded in Lm(R2m) and noting the condition vk(1) = 0, from standard elliptic estimates we infer that (vk) is bounded in W 2,m(B1). Hence a subsequence vk → v weakly in W 2,m(B1) and in C2m−1,α away from x = 0. We then have ∆v = 0 and v(1) = 0, therefore v ≡ 0 on B1. By elliptic estimates, from (43) and the condition vk(1) = 0 we also infer that (vk) is bounded in W 2,m loc (R2m). Therefore, we also have that vk → v weakly in W 2,m (R2m \ {0}), with ∆v = 0. By unique continuation it follows that v ≡ 0. This completes the proof. loc (R2m) and in C2m−1,α loc (cid:3) Lemma 10 For any L > 0 there exists k0 = k0(L) such that for all k ≥ k0 and any 1 ≤ j ≤ 2m − 1 there holds uk(r(ℓ+1) k )ZB ∇jukdx ≤ C(Lr(ℓ+1) k )2m−j . (ℓ+1) k \B r (ℓ+1) k /L Lr Proof. The proof is identical to the proof of Lemma 6 in [8], using Lemma 9 above instead of Lemma 3 in [8]. (cid:3) Proof of Proposition 7. For simplicity of notation, we now drop the index ℓ + 1. Step 1. We claim that ηk → η in C2m−1,α (R2m\{0}) for some smooth function η. For any L > 1 let ΩL := BL(0)\B1/L(0). Recall that by Lemma 9 we have uk(x) := uk(rkx) uk(rk) → 1 uniformly on ΩL as k → ∞. Thus by (7) with error o(1) → 0 as k → ∞ we have loc 0 ≤ (−∆)mηk(x) = λkr2m ≤ (L2m + o(1))λk(rkx)2mu2 k u2 k(rk)uk(x)emu2 k(rkx)emu2 k(rkx) k(rkx) ≤ CL2m + o(1). (44) Split ηk = hk + lk on Ω2L, where ∆mhk = 0 on Ω2L, and lk = ∆lk = . . . = ∆m−1lk = 0 on ∂Ω2L. Since k∆mηkkL∞(Ω2L) ≤ C = C(L), by elliptic estimates we get that lk → l in C2m−1,α(Ω2L). Together with Lemma 10 this implies k∇hkkL1(Ω2L) ≤ k∇lkkL1(Ω2L) + k∇ηkkL1(Ω2L) ≤ C. Moreover, since ηk = 0 on ∂B1(0), we have hk = lk ≤ C on ∂B1(0). (45) 16 Then, from a Poincar´e-type inequality, we easily get khkkL1(Ω2L) ≤ C. By virtue of Proposition 21, we infer that khkkC j(ΩL) ≤ Cj for every j ∈ N. Hence a subsequence hk → h smoothly on ΩL, and ηk → η := h + l in C2m−1,α(ΩL), proving our claim. Step 2. With uk(x) := uk(rkx) uk(rk) as above, from (44) we get (−∆)mηk = λkr2m k u2 k(rk)emu2 = µkukem(uk+1)ηk , k(rk)uk(x)em(u2 k(rk · )−u2 k(rk)) (46) where by (40) we may assume µk := λkr2m k u2 k(rk)emu2 k(rk) = ω−1 2m−1Pk(rk) → µ0 > 0. Since uk → 1 locally uniformly on R2m\{0} we may pass to the limit k → ∞ in (46) to see that η solves the equation (−∆)mη = µ0e2mη on R2m\{0} (47) in the distribution sense. In fact, we now show that (47) holds on all of R2m. Note that by Step 1 for any L > 1 we have ZΩL e2mηdx = lim k→∞ZΩL k→∞ZΩL = lim u2 kem(uk+1)ηk dx µ−1 k uk(−∆)mηkdx ≤ µ−1 0 lim inf k→∞ ZBLrk uk(−∆)mukdx ≤ µ−1 0 Λ. As L → ∞, by Fatou's lemma, we get e2mη ∈ L1(R2m). Moreover η ≥ 0 on B1, hence η ∈ Lp(B1) for every p ∈ [1,∞[. Also note that (−∆)mηk ≥ 0 and that from (32) we can bound lim sup k→∞ ZB1/L(0) (−∆)mηk dx = lim sup k→∞ ZBrk /L(0) λkuk(rk)ukemu2 k dx ≤ lim sup k→∞ Nk(s(ℓ) k , rk/L) → 0 (48) as L → ∞. Since by Lemma 4 we have uk ≥ 1, ηk ≥ 0 on B1, from (46) and (48) we also find that lim sup k→∞ ZB1/L(0) ηk dx → 0 as L → ∞ . (49) 17 By (47) and (48) for any test function ϕ ∈ C∞ ZR2m(cid:0)(−∆)mη − µ0e2mη(cid:1)ϕ dx = lim L→∞ZR2m (−∆)mηϕτL dx k→∞ ZR2m(cid:0)(−∆)mη − (−∆)mηk(cid:1)ϕτL dx, lim inf 0 (R2m) we now obtain = lim L→∞ (50) where for L ∈ N we let τL(x) = τ (Lx) with a fixed cut-off function τ ∈ C∞ such that 0 ≤ τ ≤ 1 and τ ≡ 1 in B1. But by Step 1 for any L ≥ 1 we have 0 (B2) lim inf k→∞ ZR2m (cid:0)(−∆)mη − (−∆)mηk(cid:1)ϕτL dx = lim inf k→∞ ZR2m (η − ηk)(cid:0)(−∆)mϕ(cid:1)τL dx, and since η ∈ L1(B1) and on account of (49) the latter converges to 0 as L → ∞ for any fixed ϕ ∈ C∞ 0 (R2m). From (50) we thus see that η solves (47) in the distribution sense on R2m. By elliptic estimates, η is smooth on all of R2m; see for instance [7], Corollary 8. The function η0 := η + 1 (2m−1)! then satisfies 2m log µ0 (−∆)mη0 = (2m − 1)!e2mη0 in R2m, ZR2m e2mη0dx < ∞. (51) Solutions to (51) have been classified in [7], where it was shown that either 2σ (i) η0(x) = log (ii) m > 1 and there exist 1 ≤ j ≤ m − 1 and a 6= 0 such that 1+σx−x02 for some σ > 0, x0 ∈ R2m, or lim x→∞ ∆jη0(x) = a, and hence for sufficiently large L, with error o(1) → 0 as k → ∞, (Lrk)2j−2muk(rk)ZBLrk \Brk /L ∇2jukdx = L2j−2mZBL\B1/L ∇2jηkdx = L2j−2mZBL\B1/L ∇2jη0dx + o(1) ≥ CL2j + o(1) (52) for some constant C > 0 independent of L. But (52) is incompatible with the estimate of Lemma 10 when L and k are large. Hence case (i) occurs (with x0 = 0, by radial symmetry). In particular, (cid:3) we have RR2m(2m − 1)!e2mη0 dx = Λ1. 3 The general case The following gradient bound analogous to [5], Proposition 2, and generalizing [13], Proposition 4.1, will be crucial in the sequel. The proof will be given in the next section. Proposition 11 There exists a uniform constant C such that sup x∈Ω inf 1≤j≤I x − x(j) k ℓuk(x)∇ℓuk(x) ≤ C for all 1 ≤ ℓ ≤ 2m − 1, k ∈ N. 18 Fix an index i ∈ {1, . . . , I} and let xk = x(i) k → 0 as given by Theorem 1. After a translation we may assume that x(i) = 0. Set as before k → x(i), rk = r(i) ek := λku2 kemu2 k , fk := λkuk(0)ukemu2 k , and Λk(r) := ZBr ekdx. In the following we will use the notation ¯f (r) := Z ∂Br f dσ, for any function f . Set also ek := λk ¯u2 kem¯u2 k ≤ ¯ek. (Here we used Jensen's inequality.) Again we let wk(x) := uk(0)(uk(x)− uk(0)), satisfying (−∆)m ¯wk = λkuk(0)ukemu2 k = ¯fk. Finally set Λk(r) := ZBr ekdx ≤ Λk(r), σk(r) := ZBr ¯fkdx. (53) Again Theorem 1 implies lim L→∞ lim k→∞ Λk(Lrk) = lim L→∞ lim k→∞ Λk(Lrk) = lim L→∞ lim k→∞ σk(Lrk) = Λ1. (54) Recalling that x(i) k = 0 we let ρk = ρ(i) k := min(cid:8) inf j6=i x(j) k 2 , dist(0, ∂Ωk)(cid:9); that is, we set ρk = dist(0, ∂Ωk) if the (x(i) Observe that by Theorem 1 we have rk = o(ρk) as k → ∞. Note that Proposition 11 implies the uniform bound k ) are the only concentration points. 0 ≤ sup r/2≤x≤r u2 k(x) − inf r/2≤x≤r u2 k(x) ≤ Cr sup x=r∇u2 k(x) ≤ C (55) for 0 ≤ r ≤ ρk. Lemma 12 Let 0 < ε < 1 and assume that for k ≥ k0 = k0(ε) there holds inf 0≤r≤ρk ¯uk(r) ≤ ε¯uk(0) 2 . Let Tk = Tk(ε) ≤ Sk = Sk(ε) ∈]0, ρk] be the smallest numbers such that ¯uk(Tk) = εuk(0), ¯uk(Sk) = εuk(0)/2, respectively. Then lim k→∞ rk Tk = lim k→∞ Tk Sk = 0. 19 (56) Moreover for any b < 2 and k ≥ k0 = k0(b) there holds ¯wk(r) ≤ b log(cid:18) rk r (cid:19) + C for 0 ≤ r ≤ Tk, and we have Λk(Tk) = Λ1. lim k→∞ (57) (58) Proof. Property (56) follows from (55) and our choice of Tk and Sk. As in the proof of Lemma 4 for a given t ≤ Tk we decompose ¯wk = gk + hk on Bt, with ∆mhk = 0 in Bt, and gk = ∂ν gk = . . . = ∂m−1 ν gk = 0 on ∂Bt. By (54), we get the analogues of Lemma 3 and of (22); that is, for L ≥ L0 = L0(b), k ≥ k0 = k0(L) there holds (−1)m−1tm−1(cid:16)t−1(cid:16)t−1 ···(cid:16)t−1 } m−1 times {z (cid:16) ¯w′ k(t) + b t(cid:17)(cid:17)′ ···(cid:17)′(cid:17)′ } {z m−1 times ≤ 0 for all t ∈ [Lrk, Sk]. We now inductively integrate from t to Sk as in Lemma 4. Using Proposition 11 to bound ∂j r ¯wk(Sk) = uk(0) ¯uk(Sk) Sj k ¯uk(Sk)∂j Sj k r ¯uk(Sk) ≤ C εSj k , and recalling (56), for L ≥ L0 and k ≥ k0 we get t ¯w′ k(t) ≤ −b + C ε t2 S2 k = −b + o(1) for all Lrk ≤ t ≤ Tk, with error o(1) → 0 as k → ∞. Since b < 2 is arbitrary, (57) follows as before. In order to prove (58) observe that the definition of rk gives ek(r) ≤ Cλku2 ≤ Cλkr2m k(0)emu2 k(0)e 2m(cid:0)1+ k u2 k(0)emu2 k(0)r−2m k ¯wk (r) 2u2 k (0)(cid:1) ¯wk(r) em(ε+1) ¯wk(r) ≤ Cr−2m k (cid:18) rk r (cid:19)m(ε+1)b for Lrk ≤ r ≤ Tk. We then complete the proof as in the radial case. (cid:3) For 0 ≤ s < t ≤ ρk set Nk(s, t) := Λk(t) − Λk(s) = ZBt\Bs λku2 kemu2 k dx, and let Nk(s, t) := Λ(t) − Λ(s) = Z t s From (55) we infer ω2m−1λkr2m−1 ¯u2 kem¯u2 k dr ≤ Nk(s, t). sup x=r emu2 k(x) ≤ Cem¯u2 k(r) for 0 ≤ r ≤ ρk; 20 (59) (60) hence we obtain sup x=r k(x)emu2 u2 k(x) ≤ C(1 + ¯u2 k(r))em¯u2 k (r) for 0 ≤ r ≤ ρk. (61) Then (61) implies Nk(s, t) ≤ C Nk(s, t) + o(1) for 0 ≤ s ≤ t ≤ ρk, (62) with o(1) → 0 as k → ∞. Similarly, setting ekdσ = ω2m−1λkt2m ¯u2 Pk(t) = tZ∂Bt k(t)em¯u2 k(t) ≤ Pk(t) := tZ∂Bt ekdσ, we can estimate Pk(t) ≤ C Pk(t) + o(1) for 0 ≤ t ≤ ρk, (63) with o(1) → 0 as k → ∞. Finally, from (61) we also obtain the analogue of (28); that is, we have Pk(t) ≤ CNk(t/2, t) + o(1) ≤ CPk(t/2) + o(1), (64) with error o(1) → 0 as k → ∞. In particular, we obtain the following improvement of Lemma 12. Lemma 13 For any 0 < ε < 1, if Tk = Tk(ε) ≤ ρk is as in Lemma 12, then we have Proof. Indeed (58) and (62) imply lim k→∞ Λk(Tk) = Λ1. Nk(Lrk, Tk) ≤ C lim which together with (54) implies the lemma. lim L→∞ lim k→∞ L→∞ Nk(Lrk, Tk) = 0, lim k→∞ (cid:3) If the assumptions of Lemma 12 hold for any 0 < ε < 1 we may proceed to resolve secondary concentrations at scales o(ρk) as in the radially symmetric case. Indeed, by Lemmas 12 and 13 we may then choose a subsequence (uk), numbers εk → 0 as k → ∞ and corresponding numbers sk = Tk(εk) ≤ ρk with rk/sk → 0 as k → ∞ and such that Λk(sk) = Λ1, Nk(Lrk, sk) = 0, lim L→∞ lim k→∞ lim k→∞ while in addition ¯uk(sk) → ∞ and ¯uk(sk) ¯uk(Lrk) lim k→∞ = 0 for every L > 0. As before, by slight abuse of notation, we set rk = r(1) k , so that the analogue of (H1) holds, and iterate. Suppose that for some integer ℓ ≥ 1 we already have determined numbers k , sk = s(1) s(0) k := 0 < r(1) k < s(1) k < ··· < r(ℓ) k < s(ℓ) k = o(ρk) satisfying the analogues of (Hℓ,1) up to (Hℓ,4). Similar to Lemma 5 we then have the following result. 21 Lemma 14 There is a constant C0 = C0(Λ) such that for s(ℓ) there holds k ≤ tk = o(ρk) + C0 N 2 k (s(ℓ) k , tk) + o(1), (65) Pk(tk) Nk(s(ℓ) k , tk) ≤ with error o(1) → 0 as k → ∞. Proof. For ease of notation we write s = s(ℓ) proof of Lemma 5, similar to (31) we find m k . Replacing wk with ¯wk in the Nk(s, t) ≤ Pk(t) 2m −Z t s ω2m−1r2m ¯uk(r) uk(0) ¯w′ k(r)ekdr + o(1), with error o(1) → 0 as k → ∞, uniformly in s ≤ t. Proceeding as in Lemma 5, from the equation (t−1(t−1 ··· (t−1 } m−1 times {z ¯w′ k(t) )′ ··· )′)′ {z } m−1 times = (−1)m Ak(t) ω2m−1t4m−3 , where Ak is defined by (14), with σk now given by (53), we get t ¯w′ k(t) = −t2Z ρk t t1Z ρk t1 t2 ···Z ρk tm−2 Ak(tm−1) ω2m−1t4m−3 m−1 dtm−1 ··· dt1 + Bk(t, ρk), where Bk(t, ρk) corresponds to the boundary terms. By arguing as in the proof of Lemma 4 we see that Bk is a linear combination of terms of the form t2l+2 ρ2l+2 k ρj k∂j r ¯wk(ρk), 0 ≤ l ≤ m − 2, 1 ≤ j ≤ m − 1. After multiplication with ¯uk(t) uk(0) , the resulting terms can be written as t2l+2 ρ2l+2 k ¯uk(t)ρj k∂j r ¯uk(ρk) = t2l+1 ρ2l+1 k t¯uk(t) ρk(¯uk(ρk) + 1) ρj k(¯uk(ρk) + 1)∂j r ¯uk(ρk). But by Proposition 11 and the analogue of (33) we have ρj k(¯uk(ρk) + 1)∂j r ¯uk(ρk) ≤ C, t¯uk(t) ρk(¯uk(ρk) + 1) ≤ C. Hence for t = tk = o(ρk) we have ¯uk(t) error o(1) → 0 as k → ∞ we obtain the identity t2 ···Z ρk uk(0) Z ρk t1Z ρk = t2 ¯uk(t) ¯uk(t) uk(0) −t ¯w′ k(t) t t1 tm−2 Ak(tm−1) ω2m−1t4m−3 m−1 dtm−1 ··· dt1. uk(0) Bk(t, ρk) → 0 as k → ∞, and up to an The rest of the proof is similar to the proof of Lemma 5. (cid:3) On account of (62) and (63) we now obtain the analogue of Lemma 6. The proof is the same as in the radially symmetric case. 22 Lemma 15 Let C0 = C0(Λ) be the constant appearing in (65), and let tk > s(ℓ) k be such that for a subsequence lim k→∞ tk ρk = 0, 0 < lim k→∞ Nk(s(ℓ) k , tk) =: α < 1 2C0 . (66) Then s(ℓ) k tk lim k→∞ = 0, lim inf k→∞ Pk(tk) ≥ mα 2 , and lim L→∞ lim k→∞ Nk(s(ℓ) k , tk/L) = 0. We now closely follow [6]. By the preceding result it suffices to consider the following two cases. In Case A for any sequence tk = o(ρk) we have sup s(ℓ) k <t<tk Pk(t) → 0 as k → ∞, and then in view of Lemma 15 also lim L→∞ lim k→∞ Nk(s(ℓ) k , ρk/L) = 0, (67) thus completing the concentration analysis at scales up to o(ρk). In Case B for some s(ℓ) k < tk ≤ ρk there holds lim k→∞ k , tk) > 0, Nk(s(ℓ) lim sup k→∞ tk ρk = 0. Then, as in the radial case, from Lemma 15 we infer that for a subsequence (uk) and suitable numbers r(ℓ+1) k , tk[ we have k ∈]s(ℓ) Nk(s(ℓ) lim k→∞ s(ℓ) k r(ℓ+1) k = 0, lim k→∞ k , r(ℓ+1) k ) > 0, lim inf k→∞ Pk(r(ℓ+1) k ) > 0; (68) in particular, ¯uk(r(ℓ+1) k ) → ∞ as k → ∞. Also note that lim L→∞ lim sup k→∞ Nk(s(ℓ) k , r(ℓ+1) k /L) = lim k→∞ r(ℓ+1) k ρk = lim k→∞ tk ρk = 0. (69) Moreover, analoguous to Proposition 7 we have the following result, which is a special case of Proposition 17 below. Proposition 16 There exists a subsequence (uk) such that η(ℓ+1) k (x) := ¯uk(r(ℓ+1) k )(uk(r(ℓ+1) k x) − ¯uk(r(ℓ+1) k )) → η(ℓ+1)(x) in C2m−1 for a suitable constant c(ℓ+1). (R2m \ {0}) as k → ∞, where η(ℓ+1) loc 0 := η(ℓ+1) + c(ℓ+1) solves (8), (9) From Proposition 16 the desired energy quantization result at the scale r(ℓ+1) k follows as in the radial case. If ρk ≥ ρ0 > 0 we can argue as in [13], p. 416, to obtain numbers s(ℓ+1) k satisfying lim L→∞ lim k→∞ Λk(s(ℓ+1) k ) = (ℓ + 1)Λ1, (70) 23 and such that lim L→∞ lim k→∞ (Λk(s(ℓ+1) k ) − Λk(Lr(ℓ+1) k )) = lim k→∞ r(ℓ+1) k s(ℓ+1) k = lim k→∞ s(ℓ+1) k = 0, while ¯uk(s(ℓ+1) k ) → ∞ as k → ∞. Moreover, for any L ≥ 1 we have lim k→∞ ¯uk(s(ℓ+1) ) ¯uk(Lr(ℓ+1) k k ) = 0. (71) By iteration we then establish (70), (71) up to ℓ + 1 = ℓ0 for some maximal index ℓ0 ≥ 1 where Case A occurs and thus complete the concentration analysis near the point x(i), getting lim k→∞ Λk(ρ0) = ℓ0Λ1. k k k k k k k ). Then as before we can define numbers s(ℓ+1) , ρk] there holds ¯uk(t) ≥ ε0 ¯uk(r(ℓ+1) If ρk → 0 as k → ∞, we distinguish the following two cases. In Case 1 for some ε0 ∈]0, 1[ and all t ∈ [r(ℓ+1) ). The decay estimate that we established in Lemma 12 then remains valid through- out this range and (70) holds true for any choice s(ℓ+1) = o(ρk). Again the concentration analysis at scales up to o(ρk) is complete. In Case 2, for any ε ∈]0, 1[ there is a minimal Tk = Tk(ε) ∈ [r(ℓ+1) , ρk] as in Lemma 12 such that ¯uk(Tk) = ε¯uk(r(ℓ+1) < ρk with ¯uk(s(ℓ+1) ) → ∞ as k → ∞ so that (70), (71) also hold true, and we proceed by iteration up to some maximal index ℓ0 ≥ 1 where either Case 1 or Case A holds with final radii r(ℓ0) For the concentration analysis at the scale ρk first assume that for some number L ≥ 1 there is a sequence (xk) such that ρk/L ≤ Rk(xk) ≤ xk ≤ Lρk and (72) By Proposition 11 we may assume that xk = ρk. Moreover, (55) implies that dist(0, ∂Ωk)/ρk → ∞ as k → ∞. As in [13], Lemma 4.6, we then have ¯uk(ρk)/¯uk(r(ℓ0) ) → 0 as k → ∞, ruling out Case 1; that is, at scales up to o(ρk) we end with Case A. The desired quantization result at the scale ρk then is a consequence of the following result similar to [13], Proposition 4.7, whose proof may be easily carried over to the present situation. k(xk) ≥ ν0 > 0. λkxk2mu2 k(xk)emu2 , respectively. , s(ℓ0) k k k Proposition 17 Assuming (72), there exist a finite set S0 ⊂ R2m and a sub- sequence (uk) such that ηk(x) := uk(xk)(uk(ρkx) − uk(xk)) → η(x) in C2m−1 η0 = η + c0 solves (8), (9). loc (R2m \ S0) as k → ∞, where for a suitable constant c0 the function By Proposition 17 in case of (72) there holds lim L→∞ k→∞Z{x∈Ω; lim ρk L ≤Rk(x)≤x≤Lρk} ekdx = Λ1. (73) 24 Letting Xk,1 = X (i) k,1 = {x(j) k ;∃C > 0 : x(j) k ≤ Cρk for all k} and carrying out the above blow-up analysis up to scales of order o(ρk) also on all balls of center x(j) k ∈ Xk,1, then from (71) and (73) we have lim L→∞ lim k→∞ Λk(Lρk) = Λ1(1 + I1), where I1 is the total number of bubbles concentrating at the points x(j) at scales o(ρk). k ∈ X (i) k,1 On the other hand, if (72) fails to hold clearly we have lim L→∞ lim sup k→∞ Z{x∈Ω; ρk L ≤Rk(x)≤x≤Lρk} ekdx = 0, (74) and the energy estimate at the scale ρk again is complete. In order to deal with secondary concentrations around x(i) k = 0 at scales exceeding ρk, with Xk,1 defined as above we let x(j) k 2 ρk,1 = ρ(i) k,1 = min(cid:8) inf k /∈Xk,1} {j;x(j) , dist(0, ∂Ωk)(cid:9); that is, we again set ρk,1 = dist(0, ∂Ωk), if {j; x(j) /∈ Xk,1} = ∅. From this definition it follows that ρk,1/ρk → ∞ as k → ∞. Then, using the obvious analogue of Lemma 15, either we have k lim L→∞ lim sup k→∞ Nk(cid:16)Lρk, ρk,1 L (cid:17) = 0, and we iterate to the next scale; or there exist radii tk ≤ ρk,1 such that tk/ρk → ∞, tk/ρk,1 → 0 as k → ∞ and a subsequence (uk) such that Pk(tk) ≥ ν0 > 0 for all k. (75) The argument then depends on whether (72) or (74) holds. In case of (72), as in [13], Lemma 4.6, the bound (75) and Proposition 7 imply that ¯uk(tk)/¯uk(ρk) → 0 as k → 0. Then we can argue as in Case A for r ∈ [Lρk, ρk,1] for sufficiently large L, and we can continue as before to resolve concentrations in this range of scales. In case of (74) we further need to distinguish whether Case A or Case 1 holds at the final stage of our analysis at scales o(ρk). In fact, for the following estimates we also consider all points x(j) k . Recalling that in Case A we have (71) (with index ℓ0 instead of ℓ + 1) and (67), on account of (74) for a suitable sequence of numbers s(0) k,1 → ∞ as k → ∞ we find k ∈ X (i) k,1 such that s(0) k,1/ρk → ∞, tk/s(0) k,1 in place of x(i) lim L→∞ lim k→∞(cid:16)Λ(s(0) k,1) − Xx(j) k ∈X (i) k,1 k (Lr(ℓ(j) Λ(j) 0 ) k )(cid:17) = 0, 0 ) k (r) and r(ℓ(j) k . In particular, with such a choice of s(0) are computed as above with respect to the concentration k,1 we find the intermediate k where Λ(j) point x(j) quantization result lim k→∞ Λk(s(0) k,1) = Λ1I1 25 analogous to (70), where I1 is defined as above. In Case 1 we can obtain the same conclusion by our earlier reasoning. Moreover, in Case 1 we can argue as in [13], Lemma 4.8, to conclude that ¯uk(tk)/¯uk(Lr(ℓ(j) 0 ) ) → 0 for any L ≥ 1 as k → 0; therefore, similar to (71) in Case A, we can achieve that for any L ≥ 1 we have k lim k→∞ ¯uk(s(0) k,1) ¯uk(Lr(ℓ(j) 0 ) k ) = lim k→∞ r(ℓ(j) 0 ) k s(0) k,1 = lim k→∞ ρk s(0) k,1 = lim k→∞ s(0) k,1 tk = 0 for all x(j) k ∈ X (i) k,1 where Case 1 holds, similar to (Hℓ). We then finish the argument by iteration. For ℓ ≥ 2 we inductively define the sets Xk,ℓ = X (i) k,ℓ = {x(j) k ;∃C > 0 : x(j) k ≤ Cρk,ℓ−1 for all k} and we let ρk,ℓ = ρ(i) k,ℓ = min(cid:8) inf k /∈Xk,1} {j;x(j) x(j) k 2 , dist(0, ∂Ωk)(cid:9); that is, as before, we set ρk,ℓ = dist(0, ∂Ωk), if {j; x(j) k,ℓ} = ∅. Iteratively performing the above analysis at all scales ρk,ℓ, thereby exhausting all concen- tration points x(j) k , upon passing to further subsequences, we finish the proof of Theorem 2. /∈ X (i) k 3.1 Proof of Proposition 11 Our proof of Proposition 11 is modelled on the proof of [5], Proposition 2. In fact, the first steps of the proof seem almost identical to the corresponding arguments in [5]. The special character of the present problem only enters at the last stage, where we also need to distinguish the cases ℓ = 1 and 2 ≤ ℓ ≤ 2m− 1. Fix any index 1 ≤ ℓ ≤ 2m − 1. The following constructions will depend on this choice; however, for ease of notation we suppress the index ℓ in the sequel. Set Rk(x) := inf 1≤j≤I x − x(j) k and choose points yk such that Rℓ k(yk)uk(yk)∇ℓuk(yk) = sup Ω Rℓ kuk∇ℓuk =: Lk. Suppose by contradiction that Lk → ∞ as k → ∞. From Theorem 1 then it follows that sk := Rk(yk) → 0 as k → ∞. Set and let Ωk := {y; yk + sky ∈ Ω} vk(y) := uk(yk + sky), y ∈ Ωk. Observe that for 1 ≤ j ≤ m via Sobolev's embedding from (3) we obtain k∇jvkk2 L 2m j (Ωk) ≤ Ck∇mvkk2 L2(Ωk) = CZΩk vk(−∆)mvkdx ≤ C. (76) 26 Also let and set Clearly then we have y(i) k := x(i) k − yk sk , 1 ≤ i ≤ I Sk := {y(i) k ; 1 ≤ i ≤ I}. dist(0, Sk) = inf 1≤i≤I y(i) k = 1 and sup y∈Ωk (dist(y, Sk)ℓvk(y)∇ℓvk(y)) = vk(0)∇ℓvk(0) = Lk → ∞ (77) as k → ∞. Moreover (7) implies 0 ≤ vk(−∆)mvk = λks2m k v2 kemv2 k ≤ C dist(y, Sk)2m . (78) Since limk→∞ sk = 0, we may assume that as k → ∞ the domains Ωk exhaust a half-space Ω0 = R2m−1×] − ∞, R0[, where 0 < R0 ≤ ∞. We may also assume that either limk→∞ y(i) limk→∞ y(i) points of Sk, satisfying dist(0, S0) = 1. For R > 0 denote k = ∞ or k = y(i), 1 ≤ i ≤ I, and we let S0 be the set of these accumulation Observing that λks2m B1/R(y). Kk,R := Ωk ∩ BR(0)\ [y∈S0 k → 0, from (78) we obtain that k→∞k∆mvkkL∞(Kk,R) = 0 for every R > 0. lim (79) Lemma 18 We have R0 = ∞, hence Ω0 = R2m. Proof. Suppose by contradiction that R0 < ∞. Choosing R = 2R0 and observ- ing that by (2) for 0 ≤ j < ℓ ≤ 2m − 1 we have ∂j k = 0 on ∂Ωk, from Taylor's formula and (77) we conclude ν v2 v2 k sup Kk,R vk(0)∇ℓvk(0) ≤ C = C(R). , we then have 0 ≤ wk ≤ C on Kk,R. Using (76), Letting wk := vk √vk(0)∇vk(0) Sobolev's embedding, (77) and (79) we infer k∇wkkL2m(Ωk) + k∇2wkkLm(Ωk) + k∆mwkkL∞(Kk,R) → 0 as k → ∞. Since ∂j that wk → 0 in C2m−1,α wk(0)∇ℓwk(0) = 1. νwk = 0 on ∂Ωk for 0 ≤ j ≤ m − 1, it follows from elliptic regularity (Kk,R) for 0 < α < 1, contradicting the fact that (cid:3) loc 27 Lemma 19 As k → ∞ we have vk(0) → ∞ and vk vk(0) → 1 in C2m−1,α loc (R2m\S0). Proof. First observe that ck := sup B1/2 vk → ∞ as k → ∞. Indeed, otherwise (76), (79) and elliptic regularity would contradict (77). Let- ting wk := vk ck , from (76) and (79) for any R > 0 we have k∇wkkL2m(Ωk) + k∇2wkkLm(Ωk) + k∆mwkkL∞(Kk,R) → 0 as k → ∞, whence wk → w ≡ const in C2m−1,α we obtain loc (R2m\S0). Recalling that dist(0, S0) = 1, w ≡ sup B1/2 w = lim k→∞ sup B1/2 wk = 1. In particular we conclude that vk(0) ck vk(0) = ckwk(0) → ∞, = wk(0) → 1 as k → ∞ and therefore (R2m\S0), as claimed. (cid:3) For the final argument now we need to distinguish the cases ℓ = 1 and wk(0) → 1 in C2m−1,α vk(0) = wk loc vk 2 ≤ ℓ ≤ 2m − 1. Consider first the case ℓ = 1. Set vk(y) − vk(0) vk(y) := ∇vk(0) . From (77) and Lemma 19 we infer ∇vk(y) = vk(0) vk(y) vk(y)∇vk(y) vk(0)∇vk(0) ≤ 1 + o(1) dist(y, S0) , (80) with error o(1) → 0 in C2m−1,α (R2m\S0) as k → ∞. Since vk(0) = 0, from (80) we conclude that vk is bounded in C1(Kk,R) for every R > 0, uniformly in k. Moreover, (78) and Lemma 19 give loc ∆mvk = vk(0) vk vk∆mvk vk(0)∇vk(0) ≤ C(R) vk(0) Lkvk → 0 (81) loc loc ∆mv = 0, uniformly on Kk,R as k → ∞, for any R > 0. The sequence vk then is bounded in C2m−1,α (R2m\S0) for any α < 1, and by Arzel`a-Ascoli's theorem we can assume that vk → v in C2m−1,α v(0) = 0, (R2m\S0), where v satisfies ∇v(0) = 1, ∇v(y) ≤ dist(y, S0) Fix a point x0 ∈ S0. For any r ∈]0, dist(x0, S0\{x0})/2[ let ϕ ∈ C∞ 0 (Br(x0)) be a function 0 ≤ ϕ ≤ 1 such that ϕ ≡ 1 in Br/2(x0), and satisfying ∇jϕ ≤ Cr−j for 0 ≤ j ≤ m. Integration by parts yields (∇ϕvk · ∇∆m−1vk + ϕvk∆mvk)dx ZBr (x0) (82) 1 . = −ZBr (x0) ϕ∇vk · ∇∆m−1vk dx =: I. 28 (83) Again integrating by parts m − 1 times, we obtain I = (−1)mZBr(x0) Xα=m−1 ∂α(ϕ∇vk) · ∇∂αvk dx, so that by Holder's inequality and (76) this term may be bounded I ≤ C X1≤j≤m ≤ C X1≤j≤m rj−mZBr(x0) ∇jvk∇mvkdx j k∇mvkkL2 ≤ C. k∇jvkkL 2m Similarly, we have 0 ≤ ZBr (x0) ϕvk(−∆)mvk dx ≤ C, and from (83) we conclude the bound ZBr (x0)∇ϕvk · ∇∆m−1vkdx(cid:12)(cid:12)(cid:12) ≤ C. (cid:12)(cid:12)(cid:12) (84) Observe that ∇ϕ = 0 in Br/2(x0). By Lemma 19 therefore the integral on the left-hand side equals ZBr (x0)∇ϕvk · ∇∆m−1vkdx = (1 + o(1))vk(0)∇vk(0)ZBr (x0) ∇ϕ · ∇∆m−1vkdx = −(1 + o(1))vk(0)∇vk(0)ZBr(x0) ϕ∆mvkdx. Since (−∆)mvk ≥ 0, it follows that ZBr/2(x0) (−∆)mvkdx ≤ C vk(0)∇vk(0) = CL−1 k → 0 as k → ∞. loc(R2m). Therefore ∆mv ≡ 0 Recalling (81), we infer that ∆mvk → 0 in L1 in R2m. Since from (82) we have v(y) ≤ C(1 + y) for y ∈ R2m, we may now invoke a Liouville-type theorem as in [7], Theorem 5, to see that v is a polynomial of degree at most 2m− 2 if m > 1 and of degree at most 1 if m = 1. But then (82) implies that v ≡ 0, contradicting the fact that ∇vk(0) = 1. This completes the proof in the case ℓ = 1. In the case when 2 ≤ ℓ ≤ m − 1 we set vk(y) := vk(y) − vk(0) ∇ℓvk(0) . As shown above we have dist(y, Sk)vk(y)∇vk(y) ≤ C sup x∈Ω Rk(x)uk(x)∇uk(x) ≤ C; 29 ∇vk ≤ C(1 + o(1)) vk(0)∇ℓvk(0) dist(y, S0) = C(1 + o(1)) Lk dist(y, S0) → 0. (85) hence Lemma 19 implies with error o(1) → 0 in C2m−1,α that loc (R2m\S0) as k → ∞ Notice that this is stronger than its analogue (80). As in the case ℓ = 1 we have ∆mvk = vk(0) vk vk∆mvk vk(0)∇ℓvk(0) ≤ C(R) Lk → 0 (86) uniformly on Kk,R as k → ∞, for any R > 0, hence vk → v in C2m−1,α where v satisfies loc (R2m\S0), ∆mv = 0, v(0) = 0, ∇ℓv(0) = 1. On the other hand (85) implies ∇v ≡ 0, contradiction. This completes the proof. (cid:3) Appendix We collect here some technical results used in the above sections. The proof of the following proposition can be found in [7], Prop. 4. Proposition 20 Let ∆mh = 0 in B2 ⊂ Rn. For every 0 ≤ α < 1 and ℓ ≥ 0 there is a constant C(ℓ, α) independent of h such that khkC ℓ,α(B1) ≤ C(ℓ, α)khkL1(B2). By a simple covering argument Proposition 20 can be extended to the case of annuli. Proposition 21 Let ∆mh = 0 in B2L(0)\B1/2L(0) ⊂ Rn for some L ≥ 1. For every 0 ≤ α < 1 and ℓ ≥ 0 there is a constant C = C(ℓ, α, L) such that khkC ℓ,α(BL(0)\B1/L(0)) ≤ CkhkL1(B2L(0)\B1/2L(0)). Lemma 22 Let g ∈ C∞(Bt), where Bt = Bt(0) ⊂ Rn for some n ∈ N, t > 0. Assume that g is radially symmetric and satisfies g = ∂νg = . . . = ∂m−1 ν g = 0 on ∂Bt. Then Z∂Bt tm−1∂m ν g dσ = Z t 0 t2 ···Z tm−1 0 tm(cid:18)ZBtm ∆mgdx(cid:19)dtm . . . dt2. Proof. For m = 1 equation (88) simply reduces to Z∂Bt ∂νg dσ = ZBt ∆g dx. For m = 2 consider the function ϕ(x) = x · ∇g(x) with ∆(x · ∇g)dx = ZBt ∂νϕ dσ = ZBt ∆ϕ dx = ZBt Z∂Bt (x · ∇∆g + 2∆g)dx 30 (87) (88) (89) and note that the condition ∂νg = 0 on ∂Bt and (89) imply Z∂Bt ∂ν ϕ dσ = Z∂Bt ∂ν(x · ∇g)dσ = Z∂Bt t∂2 ν g dσ, and ZBt ∆g dx = 0. Thus from Fubini's theorem we obtain the desired identity Z∂Bt t∂2 ν g dσ = ZBt = Z t 0 x · ∇∆g dx t2(cid:18)Z∂Bt2 ∂ν ∆g dσ(cid:19)dt2 = Z t 0 t2(cid:18)ZBt2 ∆2g dx(cid:19)dt2. We now proceed by induction. Assume that the lemma is true for m − 1. Choosing ϕ(x) = x · ∇g(x) with ϕ = ∂νϕ = . . . = ∂m−2 ν ϕ = 0 on ∂Bt we get tm−1∂m Z∂Bt = Z t 0 ν gdσ = Z∂Bt t2 ···Z tm−2 0 tm−1ZBtm−1 tm−2∂m−1 ν tm−2∂m−1 (t∂ν g)dσ = Z∂Bt ∆m−1(x · ∇g)dxdtm−1 . . . dt2 =: I. ν (x · ∇g)dσ Observe that ∆m−1(x · ∇g) = x · ∇∆m−1g + 2(m − 1)∆m−1g, hence I = Z t 0 t2 ···Z tm−2 + 2(m − 1)Z t tm−1ZBtm−1 t2 ···Z tm−2 0 0 0 (x · ∇∆m−1g)dxdtm−1 . . . dt2 tm−1ZBtm−1 ∆m−1gdxdtm−1 . . . dt2 = II + III. By inductive hypothesis and (87) the contribution from the second term is III = 2(m − 1)Z∂Bt tm−2∂m−1 ν gdσ = 0, and our claim follows from writing ZBtm−1 x · ∇∆m−1gdx = Z tm−1 = Z tm−1 0 0 tmZ∂Btm tmZBtm ∂ν∆m−1gdσdtm ∆mgdxdtm. (90) (cid:3) Lemma 23 Let u ∈ C∞(Rn) ∩ Lp(Rn), for some p ≥ 1, satisfy ∆ju = 0 for some integer j > 0. Then u ≡ 0. 31 Proof. We first claim that R→∞ Z lim BR(ξ) udx = 0 for every ξ ∈ Rn. Indeed by Jensen's inequality updx(cid:19) 1 udx ≤ (cid:18) Z Z (cid:12)(cid:12)(cid:12)(cid:12) udx(cid:12)(cid:12)(cid:12)(cid:12) ≤ Z p 1 Rn/pkukLp(Rn) → 0, ≤ BR(ξ) BR(ξ) BR(ξ) as R → ∞. By Pizzetti's formula (see [10]) we have constants c1, . . . , cj−1 such that Z BR(ξ) udx = u(ξ) + c1R2∆u(ξ) + ··· + cj−1R2j−2∆j−1u(ξ) =: P (R). Taking the limit as R → ∞ we see at once that the polynomial P (R) is iden- tically 0, and in particular u(ξ) = P (0) = 0. Since ξ was arbitrary the proof is complete. (cid:3) Lemma 24 There holds capHm ({0}) = inf{k∇mϕkL2; ϕ ∈ X} = 0, where X = {ϕ ∈ C∞ 0 (B1(0)); 0 ≤ ϕ ≤ 1,∃r > 0 : ϕ(x) = 1 for x ≤ r}. Proof. Let f (x) = log log log(1/x) with ∇mf ∈ L2(Be−e (0)) and fix g ∈ C∞(R) with 0 ≤ g ≤ 1 satisfying g(s) = 0 for s ≤ 0, g(s) = 1 for s ≥ 1. Letting ϕk(x) = g(f (x) − k), k ∈ N, we find ϕk ∈ X for all k and k∇mϕkkL2 → 0 as k → ∞. (cid:3) References [1] Adimurthi, O. Druet, Blow-up analysis in dimension 2 and a sharp form of Trudinger-Moser inequality, Comm. Partial Differential Equations 29 (2004), 295-322. [2] Adimurthi, F. Robert, M. Struwe, Concentration phenomena for Li- ouville's equation in dimension 4, J. Eur. Math. Soc. 8 (2006), 171-180. [3] Adimurthi, M. Struwe, Global compactness properties of semilinear el- liptic equations with critical exponential growth, J. Functional Analysis 175 (2000), 125-167. [4] S-Y. A. Chang Non-linear Elliptic Equations in Conformal Geometry, Zurich lecture notes in advanced mathematics, EMS (2004). 32 [5] O. Druet, Multibumps analysis in dimension 2: quantification of blow-up levels, Duke Math. J. 132 (2006), 217-269. [6] T. Lamm, F. Robert, M. Struwe The heat flow with a critical exponential nonlinearity, J. Functional Anal. 257 (2009) 2951-2998. [7] L. Martinazzi Classification of solutions to the higher order Liouville's equation in R2m, Math. Z. 263 (2009), 307-329. [8] L. Martinazzi A threshold phenomenon for embeddings of H m 0 into Orlicz spaces, Calc. Var. Partial Differential Equations 36 (2009), 493-506. [9] L. Martinazzi, M. Petrache Asymptotics and quantization for a mean- field equation of higher order, Comm. Partial Differential Equations 35 (2010), 1-22. [10] P. Pizzetti Sulla media dei valori che una funzione dei punti dello spazio assume alla superficie di una sfera, Rend. Lincei 18 (1909), 182-185. [11] F. Robert, M. Struwe Asymptotic profile for a fourth order PDE with critical exponential growth in dimension four, Adv. Nonlin. Stud. 4 (2004), 397-415. [12] F. Robert, J.-C. Wei Asymptotic behavior of a fourth order mean field equation with Dirichlet boundary condition, Indiana Univ. Math. J. 57 (2008), 2039-2060. [13] M. Struwe Quantization for a fourth order equation with critical expo- nential growth, Math. Z. 256 (2007), 397-424. [14] J.-C. Wei Asymptotic behavior of a nonlinear fourth order eigenvalue prob- lem, Comm. Partial Differential Equations 21 (1996), 1451-1467. 33
1809.08465	2	1809	2019-12-04T07:32:19	The Segal-Bargmann Transform on Classical Matrix Lie Groups	[ "math.FA" ]	We study the complex-time Segal-Bargmann transform $\mathbf{B}_{s,\tau}^{K_N}$ on a compact type Lie group $K_N$, where $K_N$ is one of the following classical matrix Lie groups: the special orthogonal group $\mathrm{SO}(N,\mathbb{R})$, the special unitary group $\mathrm{SU}(N)$, or the compact symplectic group $\mathrm{Sp}(N)$. Our work complements and extends the results of Driver, Hall, and Kemp on the Segal-Bargman transform for the unitary group $\mathrm{U}(N)$. We provide an effective method of computing the action of the Segal-Bargmann transform on \emph{trace polynomials}, which comprise a subspace of smooth functions on $K_N$ extending the polynomial functional calculus. Using these results, we show that as $N\to\infty$, the finite-dimensional transform $\mathbf{B}_{s,\tau}^{K_N}$ has a meaningful limit $\mathscr{G}_{s,\tau}^{(\beta)}$ (where $\beta$ is a parameter associated with $\mathrm{SO}(N,\mathbb{R})$, $\mathrm{SU}(N)$, or $\mathrm{Sp}(N)$), which can be identified as an operator on the space of complex Laurent polynomials.	math.FA	math	The Segal-Bargmann Transform on Classical Matrix Lie Groups Alice Z. Chan∗a aDepartment of Mathematics, University of California, San Diego, La Jolla, CA 92093 Abstract We study the complex-time Segal-Bargmann transform BKN s,τ on a compact type Lie group KN , where KN is one of the following classical matrix Lie groups: the special orthogonal group SO(N, R), the special unitary group SU(N ), or the compact symplectic group Sp(N ). Our work complements and extends the results of Driver, Hall, and Kemp on the Segal-Bargman transform for the unitary group U(N ). We provide an effective method of computing the action of the Segal-Bargmann transform on trace polynomials, which comprise a subspace of smooth functions on KN extending the polynomial functional calculus. Using these results, we show that as N → ∞, the finite-dimensional transform BKN s,τ has a meaningful limit Gs,τ which can be identified as an operator on the space of complex Laurent polynomials. Keywords: Segal-Bargmann transform; Heat kernel analysis on Lie groups Contents 1 Introduction 1.1 The classical Segal-Bargmann transform . . . . . . . . . . . . . . . . . . . 1.2 Main definitions and theorems . . . . . . . . . . . . . . . . . . . . . . . . 2 Background 2.1 Heat kernels on Lie groups . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Notation and definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 The Segal-Bargmann transform on SO(N, R) 3.1 Magic formulas and derivative formulas 3.2 3.3 Intertwining formulas for ∆SO(N,R) Intertwining formulas for ASO(N,C) s,τ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 2 3 7 7 8 10 10 16 23 ∗Supported by NSF GRFP under Grant No. DGE-1650112 Email address: [email protected] (Alice Z. Chan) Preprint submitted to Journal of Functional Analysis December 5, 2019 4 Limit theorems for the Segal-Bargmann transform on SO(N, R) 4.1 Concentration of measures . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 The free Segal-Bargmann transform for SO(N, R) . . . . . . . . . . . . . . 5 The Segal-Bargmann transform on Sp(N ) 5.1 Two realizations of the compact symplectic group . . . . . . . . . . . . . . 5.2 Magic formulas and derivative formulas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.3 . . . . . . . . . . . . . . . . . . . . . . 5.4 5.5 Limit theorems for the Segal-Bargmann transform on Sp(N ) . . . . . . . Intertwining formulas for ∆Sp(N ) Intertwining formulas for ASp(N,C) s,τ 5.6 Extending the magic formulas and intertwining formulas to fSp(N ) . . . . 6 The Segal-Bargmann transform on SU(N ) 29 29 33 37 38 40 45 47 50 51 53 1. Introduction 1.1. The classical Segal-Bargmann transform In this paper, we consider a complex-time generalization of the classical Segal-Bargmann transform on the matrix Lie groups SO(N, R), SU(N ), and Sp(N ), and analyze its limit as N → ∞. We begin with a brief discussion of the classical Segal-Bargmann transform and its generalizations to Lie groups of compact type. For t > 0, let ρt denote the heat kernel with variance t on Rd: This has an entire analytic continuation to Cd, given by ρt(x) = (2πt)−d/2 exp(−x2/2t). 2t (cid:17) . (ρt)C(z) = (2πt)−d/2 exp(cid:16)− (Btf )(z) :=ZRd (ρt)C(z − y)f (y) dy. z · z If f ∈ L1 loc(Rd) is of sufficiently slow growth, we can define (1.1) (1.2) The map f 7→ Btf is called the classical Segal-Bargmann transform, due to the work of the eponymous authors in [2, 1, 17, 15, 16]. If we let µt denote the heat kernel on Cd, now with variance t 2 : µt(z) = (πt)−d exp(−z2/t), then the main result is that the map Bt : L2(Rd, ρt) → HL2(Cd, µt) is a unitary isomorphism, where HL2(Cd, µt) is the space of square-integrable holomor- phic functions on Cd. 2 1.2. Main definitions and theorems In [10], Hall showed that the Segal-Bargmann transform can be extended to compact Lie groups, and in [6], the transform was further extended by Driver to Lie groups of compact type. Later, the behavior of the transform on the unitary group U(N ) as N → ∞ was analyzed independently by C´ebron in [5] and Driver, Hall, and Kemp in [8]. The authors of the latter paper subsequently introduced a complex-time generalization of the Segal-Bargmann transform for all compact type Lie groups in [9]. We use this version of the transform (see Definition 1.1) in our analysis of the special orthogonal group SO(N, R), the special unitary group SU(N ), and the compact symplectic group Sp(N ). We now outline the main results of this paper, deferring a fuller discussion of the requisite background and definitions to Section 2. Let K denote a Lie group of compact type with Lie algebra k. Then K possesses a left-invariant metric whose associated Laplacian ∆K is bi-invariant (cf. (2.1)). For each s > 0, there is an associated heat kernel ρK s ∈ C∞(K, (0,∞)) satisfying f (k)ρK s (xk−1) dk for all f ∈ L2(K). (1.3) s 2 ∆K f(cid:1) (x) =ZK (cid:0)e f (xk)ρK s (k) dk =ZK Let KC denote the complexification of K and C+ = {τ = t + iu : t > 0, u ∈ R} denote the right half-plane. As shown in [9, Theorem 1.2], the heat kernel (ρK s )s>0 possesses a unique entire analytic continuation ρK C : C+ × KC → C such that ρK C (s, x) = ρs(x) for all s > 0 and x ∈ K ⊆ KC. Thus we can proceed as in the classical case and define the Segal-Bargmann transform for compact type Lie groups using a group convolution formula similar to (1.1) (see [6] for details). Definition 1.1 (Complex-time Segal-Bargmann transform). Let s > 0 and τ = t + iθ ∈ D(s, s), and let ρK C be as above. For z ∈ KC and f ∈ L2(K, ρK s and ρK s ), define (BK s,τ f )(z) :=ZK ρK C (τ, zk−1)f (k) dk for z ∈ KC. (1.4) The complex-time Segal-Bargmann transform BK certain L2 and holomorphic L2 spaces on K and KC: s,τ is an isometric isomorphism between Theorem 1.2 (Driver, Hall, and Kemp, compact-type Lie group, and set s > 0 and τ ∈ D(s, s). Let µKC rescaled) heat kernel on KC defined by (2.4). Then BK f ∈ L2(K, ρK [9, Theorem 1.6]). Let K be a connected, s,τ denote the (time- s,τ f is holomorphic on KC for each s ). Moreover, BK s,τ : L2(K, ρK s ) → HL2(KC, µKC s,τ ) (1.5) is a unitary isomorphism. 3 If we set K = Rd and s = τ > 0, we recover the main result for the classical Segal- Bargmann transform from Section 1.1. Every compact Lie group K has a faithful representation. Hence, viewing K as a matrix group, we can extend BK s,τ to matrix-valued functions. Let MN (R) denote the algebra of N × N real matrices with unit IN , and let MN (C) denote the algebra of N × N complex matrices. Let GLN (C) denote the group of all invertible matrices in MN (C). Define a scaled Hilbert-Schmidt norm on MN (C) by kAk2 MN (C) := 1 N Tr(AA∗), A ∈ MN (C), (1.6) where Tr denotes the usual trace. Definition 1.3 (Boosted Segal-Bargmann transform). Let K ⊆ GLN (C) be a compact matrix Lie group with complexification KC. Let F be an MN (C) valued function on K or KC, and let kFkMN (C) denote the scalar-valued function A 7→ kF (A)kMN (C). Fix s > 0 and τ ∈ D(s, s), and let L2(K, ρK L2(KC, µKC s ; MN (C)) = {F : K → MN (C); kFkMN (C) ∈ L2(K, ρK s )}, s,τ ; MN (C)) = {F : KC → MN (C); kFkMN (C) ∈ L2(KC, µKC s,τ )}. The norms defined by L2(K,ρK kFk2 kFk2 L2(KC,µ s ;MN (C)) :=ZK kF (A)k2 :=ZKC kF (A)k2 KC s,τ ;MN (C)) MN (C) ρK s (A) dA MN (C) µKC s,τ (A) dA and (1.7) (1.8) make these spaces into Hilbert spaces. Let HL2(KC, µKC denote the subspace of matrix-valued holomorphic functions. s,τ ; MN (C)) ⊆ L2(KC, µKC s,τ ; MN (C)) The boosted Segal-Bargmann transform BK s,τ : L2(K, ρK s ; MN (C)) → HL2(KC, µKC s,τ ; MN (C)) is the unitary isomorphism determined by applying BN s,τ componentwise; i.e., BK s,τ (f · A) = BK s,τ f · A for f ∈ L2(K, ρK s ) and A ∈ MN (C). In this paper, we consider the case when K = KN is one of SO(N, R), SU(N ), or Sp(N ). We provide an effective method of computing BKN s,τ on the space of polynomials, and show that in this setting, the transform has a meaningful limit as N → ∞. However, the boosted Segal-Bargmann transform typically does not preserve the space of polynomials on K. For example, consider the polynomial PN (A) = A2 on SO(N, R). Let Tr denote the usual trace and set tr = 1 N Tr. From Example 4.11, we see that (BSO(N,R) s,t PN )(A) = 1 N (1 − e−t)IN + 1 2 e−t(1 + e 4 2t N )A2 − N 2 e−t(−1 + e 2t N )A tr A, (1.9) This example highlights that in general, BK s,τ maps the space of polynomial functions on K into the larger space of trace polynomials, first introduced in [8, Definition 1.7], which we now briefly describe (see Section 3.2 for the full details). Let C[u, u−1; v] denote the algebra of polynomials in the commuting variables u, u−1, and v = {v±1, v±2, . . .}. For each P ∈ C[u, u−1; v], there is an associated trace polynomial N on SO(N, R). For now, it is enough to think of P (1) P (1) N (A) as a certain "polynomial" function of A and tr(A). If P ∈ C[u, u−1; v] is a function of the variable u alone, then P (1) is simply the corresponding polynomial on SO(N, R). For example, the trace polynomial on SO(N, R) associated with P (u; v) = u5 is P (1) N (A) = A5. On the other hand, (1.9) is the trace polynomial on SO(N, R) associated with N f (u; v) = 1 N (1 − e−t) + 1 2 e−t(1 + e 2t N )u2 − N 2 e−t(−1 + e 2t N )uv1 ∈ C[u, u−1; v]. In addition, for P ∈ C[u, u−1; v], there are analogous trace polynomials P (2) on SU(N ), and Sp(N ), resp. (see Definition 3.5 for precise definitions). To simplify the statement of our results, we introduce the following notation: for N ∈ N, let N and P (4) N K (1) N = SO(N, R), K (2) N = SU(N ), K (2′) N = U(N ), K (4) N = Sp(N ), (1.10) N,C = (K (β) and set K (β) K (β) s ρN,(β) s and µN,(β) = ρ N s,τ = µ N )C and ∆(β) N = ∆K (β) N K (β) s,τ N,C for the heat kernels on K (β) for β = 1, 2, 2′, 4. In addition, we write N,C, respectively. N and K (β) N P (β) Our first theorem is an intertwining result showing that ∆(β) N can be computed by applying a certain pseudodifferential operator on C[u, u−1; v] to P and then considering the corresponding trace polynomial. Theorem 1.4 (Intertwining formulas for ∆SO(N,R), ∆SU(N ), and ∆Sp(N )). For β ∈ {1, 2, 4} and N ∈ N, there are pseudodifferential operators D(β) N acting on C[u, u−1; v] (see Definition 3.7) such that, for each P ∈ C[u, u−1; v], N P ]N . N P (β) (1.11) ∆(β) N = [D(β) Moreover, for all τ ∈ C and P ∈ C[u, u−1; v], N = [e N P (β) 2 ∆(β) e τ τ 2 D(β) N P ]N For each such β, D(β) N takes the form D(β) N = L0 + 1 N L(β) 1 + 1 N 2L(β) 2 , where L0,L(β) 1 are first order pseudodifferential operators and L(β) 2 differential operator, all independent of N . 5 (1.12) (1.13) is a second order τ 2 D(β) N is well defined on C[u, u−1; v] (see Corollary 3.11). The proof of The exponential e Theorem 1.4 for SO(N, R) and Sp(N ) can be found on p. 20 and p. 45, resp. We do not provide a full proof of the SU(N ) version of the theorem, as it is very similar to the SO(N, R) and Sp(N ) cases; the proof is discussed on p. 55. Our next result shows that we can use e transform on trace polynomials. Theorem 1.5. Let P ∈ C[u, u−1; v], N ∈ N, s > 0 and τ ∈ D(s, s). For β ∈ {1, 2, 4}, the Segal-Bargmann transform on the trace polynomial P (β) N to explicitly compute the Segal-Bargmann N can be computed as τ 2 D(β) K (β) s,τ P (β) N N = [e B τ 2 D(β) N P ]N . (1.14) The proof of Theorem 1.5 for SO(N, R) and Sp(N ) can be found on p. 23 and p. 47, resp. The proof of this theorem for the SU(N ) case is similar, and is discussed on p. 55. K (β) N s,τ on polynomials as N → ∞. While the range Our main result concerns the limit of B of the boosted Segal-Bargmann transform is the space of trace polynomials, as N → ∞ the image concentrates back onto the smaller subspace of Laurent polynomials. Theorem 1.6. Let s > 0, τ ∈ D(s, s), and f ∈ C[u, u−1]. Then there exist unique gs,τ , hs,τ ∈ C[u, u−1] such that K (β) kB K (β) N s,τ N s,τ fN − [gs,τ ]Nk2 )−1fN − [hs,τ ]Nk2 L2(K (β) N,C,µN,(β) s,τ L2(K (β) N ,ρN,(β) s ;MN (C)) ;MN (C)) k(B N 2(cid:19) , = O(cid:18) 1 = O(cid:18) 1 N 2(cid:19) . (1.15) (1.16) for β ∈ {1, 2, 4}. Definition 1.7. The map Gs,τ : C[u, u−1] → C[u, u−1] given by f 7→ gs,τ is called the free Segal-Bargmann transform, and the map Hs,τ : C[u, u−1] → C[u, u−1] given by f 7→ hs,τ is called the free inverse Segal-Bargmann transform. Explicit formulas for Gs,τ and Hs,τ are given in (4.22) and (4.23), respectively. The proof of Theorem 1.6 for SO(N, R) can be found on p. 35. The proof of the theorem for the SU(N ) and Sp(N ) cases is entirely analogous to the SO(N, R) case, so we do not provide the full details; the proofs for SU(N ) and Sp(N ) are discussed on pp. 55 and 50. Remark 1.8. The fact that the Segal-Bargmann transform on matrices in MN (C) has a meaningful limit G t as N → ∞ is due to Biane [4, Proposition 13]. For the special case when s = t > 0, the free Segal-Bargmann transform Gt,t of this paper is equivalent to G t, which Biane called the free Hall transform. A two-parameter generalization of this transform, called the free unitary Segal-Bargmann transform, Gs,t, where s > t/2 > 0, was introduced in [8]. This version of the free Segal-Bargmann transform was also investigated by Ho in [11]; in particular, he gave an integral kernel for the large-N limit of the Segal- Bargmann transform over U(N ). In this paper, we study a slight generalization of Gs,t by allowing the time parameter τ to be complex, incorporating results from [9] on the complex-time Segal-Bargmann transform. When τ = t > 0, the free Segal-Bargmann transform Gs,τ of Definition 1.7 is precisely the same as the operator Gs,t of [8]. 6 Remark 1.9. Many of the results in this paper on the Segal-Bargmann transform for SO(N, R), SU(N ), and Sp(N ) correspond to work done by Driver, Hall, and Kemp on the Segal-Bargmann transform on U(N ). The U(N ) analogue of our intertwining formulas, Theorem 1.4, is contained in [8, Theorem 1.18], where it is shown that ∆U(N ) has the intertwining formula 1 1 1 2 in the SU(N ) case). 1 2 1 D(2′) N = L0 + on the trace polynomial space is identically zero (as it is for N 2L(2′) N L(2′) 1 + , and L(2′) for certain pseudodifferential operators L0,L(2′) C[u, u−1; v] (cf. Remark 3.13), where the operator L(2′) L(2) The same operator L0 appears in the intertwining formulas for SO(N, R), U(N ), SU(N ), and Sp(N ). As N → ∞, it is the operator L0 that drives the large-N behavior of the transform; this is the key to our main result on the free Segal-Bargmann transform, Theorem 1.6, which states that the large-N limit of the Segal-Bargmann transform on SO(N, R), SU(N ), and Sp(N ) is the same operator Gs,τ . Moreover, we emphasize that the rate of convergence in all cases is O(1/N 2), as it is in the analogous result for the U(N ) case, [8, Theorem 1.11]. This is somewhat surprising, since there is a nontrivial term of order 1/N in the intertwining formulas for SO(N, R) and Sp(N ) that does not appear in the U(N ) and SU(N ) cases. That we still obtain O(1/N 2) convergence for SO(N, R) and Sp(N ) is due to the fact that in the scalar version of their intertwining formulas (cf. Theorems 3.20 and 5.5), the term of order 1/N is a first order differential operator. As a result, the 1/N terms are "washed away" in the covariance estimates (cf. the proof of Theorem 4.10 in the SO(N, R) case for a more precise explanation; the Sp(N ) case is similar). 2. Background In this section, we expand on the background required to prove our main results. In particular, we provide a brief overview of the heat kernel results used in constructing the Segal-Bargmann transform, and conclude by setting some notation that will be used for the remainder of the paper. 2.1. Heat kernels on Lie groups Definition 2.1. Let K be a Lie group with Lie algebra k. An inner product h·,·ik on k is Ad-invariant if, for all X1, X2 ∈ k and all k ∈ K, hAdk X1, Adk X2ik = hX1, X2ik . A group whose Lie algebra possesses an Ad-invariant inner product is called compact type. The Lie groups SO(N, R), SU(N ), and Sp(N ) studied in this paper are compact type (and are, in fact, compact). Compact type Lie groups have the following property. 7 Proposition 2.2 ([[14], Lemma 7.5]). If K is a compact type Lie group with a fixed Ad-invariant inner product, the K is isometrically isomorphic to a direct product group, i.e. K ∼= K0 × Rd for some compact Lie group K0 and some nonnegative integer d. Let K be a compact type Lie group with Lie algebra k, and fix an Ad-invariant inner product on k. Choose a right Haar measure λ on K; we will write dx for λ(dx) and L2(K) for L2(K, λ). If βk is an orthonormal basis for k, we define the Laplace operator ∆K on K by where for any X ∈ k, eX is the left-invariant vector field given by ∆K = XX∈βk eX 2, f (ketX )(cid:12)(cid:12)(cid:12)(cid:12)t=0 (eXf )(k) = d dt for any smooth real or complex-valued function f on K. The operator is independent of orthonormal basis chosen. For a detailed overview of left-invariant Laplacian operators, see [14]; see also [6, 7, 9] for further background on heat kernels on compact type Lie groups. The operator ∆K is left-invariant for any Lie group K. When K is compact type, ∆K is also bi-invariant. In addition, it is well known that ∆K on K is elliptic and essentially self-adjoint on L2(K) with respect to any right invariant Haar measure (see [9, Section 3.2]). As a result, there exists an associated heat kernel ρK t ∈ C∞(K, (0,∞)) satisfying (1.3). Definition 2.3. Let s > 0 and τ = t + iθ ∈ D(s, s). We define a second order left- invariant differential operator AKC s,τ on KC by (2.1) (2.2) (2.3) AKC s,τ = dXj=1(cid:20)(cid:18)s − j + t 2(cid:19) eX 2 t j − θeXjeYj(cid:21) , 2eY 2 j=1 is any orthonormal basis of k, and Yj = JXj where J is the operation of where {Xj}d multiplication of i on kC = Lie(KC). The operator AKC c (KC) is essentially self-adjoint on L2(KC) with respect to any right Haar measure (see [9, Section 3.2]). There exists a corresponding heat kernel density µKC s,τ ∈ C∞(KC, (0,∞)) such that s,τC∞ (cid:16)e 1 2 A KC s,τ f(cid:17) (z) =ZKC µKC s,τ (w)f (zw) dz for all f ∈ L2(KC). (2.4) 2.2. Notation and definitions The classical Segal-Bargmann transform is an isometric isomorphism from L2(Rd) onto HL2(Cd). In order to extend the Segal-Bargmann transform to Lie groups, we note that every connected Lie group has a complexification defined by a certain representation- theoretic universal property, mimicking the relationship between Rd and Cd (see [9, 8 Section 2]). For the present purposes, it is enough to know the concrete complexifications of SO(N, R), SU(N ), and Sp(N ), which we describe below. Let SO(N, R) denote the special orthogonal group, defined by SO(N, R) = {A ∈ MN (R) : AA⊺ = IN , det(A) = 1}. Its complexification, SO(N, R)C, is given by SO(N, C) = {A ∈ MN (C) : AA⊺ = IN , det(A) = 1}. The Lie group SO(N, R) has Lie algebra so(N, R) = {X ∈ MN (R) : X ⊺ = −X}, while SO(N, C) has Lie algebra so(N, C) = {X ∈ MN (C) : X ⊺ = −X}. The unitary group, U(N ), is defined by U(N ) = {A ∈ MN (C) : AA∗ = IN}. The Lie algebra of U(N ) is u(N ) = {X ∈ MN (C) : X ∗ = −X}. The special unitary group SU(N ) is the subgroup of U(N ) consisting of matrices with determinant 1. The complex- ification of SU(N ) is the special linear group SL(N, C) = {A ∈ GLN (C) : det(A) = 1}. The Lie algebra of SU(N ) is su(N ) = {X ∈ MN (C) : X ∗ = −X, Tr(X) = 0}, and the Lie algebra of SL(N, C) is sl(N, C) = {X ∈ MN (C) : Tr(X) = 0}. There are two standard realizations of the compact symplectic group Sp(N ), first as a group of N × N matrices over the quaternions, and second as a group of 2N × 2N matrices over C. It is more convenient for us to work with the latter realization, which we introduce here; we address the former realization in Section 5.6. Our choice of realization means that we will often have to add a normalization factor of 1 2 when working with Sp(N ) ⊆ M2N (C). Set Ω0 =(cid:20) 1 0 −1 0 (cid:21) ∈ M2(C). } {z N times ) ∈ M2N (C). We define the com- For N ∈ N, we define Ω = ΩN := diag(Ω0, Ω0, . . . , Ω0 plex symplectic group by Sp(N, C) = {A ∈ M2N (C) : A⊺ΩA = Ω}. The Lie algebra of Sp(N, C) is sp(N, C) = {X ∈ M2N (C) : ΩX ⊺Ω = X}. The compact symplectic group Sp(N ) consists of the elements of Sp(N, C) which are also unitary, i.e. Sp(N ) = Sp(N, C) ∩ U(2N ) = {A ∈ M2N (C) : A⊺ΩA = Ω, A∗A = I2N}. (2.5) The Lie algebra of Sp(N ) is sp(N ) = {X ∈ M2N (C) : ΩX ⊺Ω = X, X ∗ = −X}. The complexification of Sp(N ) is Sp(N, C). The groups SO(N, R), SU(N ), and Sp(N ) can be thought of as (special) unitary groups over R, C, and H, resp. For notational convenience, throughout the paper we associate the parameters β = 1, 2, 2′, 4 with SO(N, R), SU(N ), U(N ), and Sp(N ), resp., where β refers to the dimension of R, C, and H as associative algebras over R. In order for the large-N limit of the Segal-Bargmann transform to converge in a meaning- ful way, we require the following normalizations of the trace and Hilbert-Schmidt inner products. Notation 2.4. For A ∈ MN (C), define Tr(A), Tr(A), (2.6) (2.7) 1 N 1 2N 9 tr(A) = etr(A) = where Tr denotes the usual trace. We use tr when applying the Segal-Bargmann trans- Sp(N ). form on SO(N, R) and SU(N ), and etr when applying the Segal-Bargmann transform on We also define inner products on so(N, R), su(N ), and sp(N ) by N 2 N 2 2 Tr(XY ∗) = tr(XY ∗), X, Y ∈ so(N, R), hX, Y iso(N,R) = hX, Y isu(N ) = N Tr(XY ∗) = N 2 tr(XY ∗), X, Y ∈ su(N ), hX, Y isp(N ) = N Tr(XY ∗) = 2N 2etr(XY ∗), X, Y ∈ sp(N ). (2.8) (2.9) (2.10) 3. The Segal-Bargmann transform on SO(N, R) In this section and the next, we prove Theorem 1.4, Theorem 1.5, and Theorem 1.6 for the Segal-Bargmann transform on SO(N, R). We begin with a set of "magic formulas" which are the key ingredient to proving the SO(N, R) version of the intertwining formula for ∆SO(N,R) of Theorem 1.4. These formulas, which give simple expressions for certain quadratic matrix sums, are the analogues of the "magic formulas" of [8, Proposition 3.1] for U(N ). They allow us to compute the Segal-Bargmann transform for trace polynomials on SO(N, R) (1.5 for SO(N, R)). We then prove a second intertwining formula for the operator ASO(N,C) s,τ which is used in proving the limit theorems of Section 4. 3.1. Magic formulas and derivative formulas Lemma 3.1 (Elementary matrix identities). Let Ei,j ∈ MN (C) denote the N ×N matrix with a 1 in the (i, j)th entry and zeros elsewhere. For 1 ≤ i, j, k, ℓ ≤ N and A = (Ai,j ) ∈ MN (C), we have Ei,jEk,ℓ = δj,kEi,ℓ Ei,jAEi,j = Aj,iEi,j Ei,j AEj,i = Aj,j Ei,i Tr(AEi,j ) = Aj,i (3.1) (3.2) (3.3) (3.4) Proof. For 1 ≤ a, b,≤ N , [Ei,jEk,ℓ]a,b = [Ei,j]a,h[Ek,ℓ]h,b. NXh=1 Note that [Ei,j]a,h[Ek,ℓ]h,b equals 1 precisely when a = i, b = ℓ, and j = k, and h = j = k, and equals 0 otherwise, so [Ei,j Ek,ℓ]a,b = δj,kδa,iδb,ℓ, which proves (3.1). We now make the following observations: for any N × N matrix A, AEi,j is the matrix which is all zeros except for the jth column, which is equal to the ith column of A. Hence the only (possibly) nonzero diagonal entry is the (j, j)th entry, which is Aj,i. This proves (3.4). Similarly, Ei,jA is the matrix which is all zeros except for the i row, which is the jth row of A. Putting these observations together yields (3.2) and (3.3). 10 Proposition 3.2 (Magic formulas for SO(N, R)). For any A, B ∈ MN (C), X 2 = − 1 N XAX = XX∈βso(N,R) XX∈βso(N,R) XX∈βso(N,R) tr(XA)X = tr(XA) tr(XB) = N − 1 N IN = −IN + 1 N IN A⊺ − tr(A)IN 1 N 2 (A⊺ − A) 1 N 2 (tr(A⊺B) − tr(AB)) XX∈βso(N,R) (3.5) (3.6) (3.7) (3.8) XAX is independent of the choice of orthonormal basis. Hence to see (3.6), we may set βso(N,R) to be the orthonormal basis Proof. As shown in Theorem 3.3 of [8], the quantity PX∈βso(N,R) (Ei,j − Ej,i)(cid:27)1≤i<j≤N . By Lemma 3.1, Xi<j βso(N,R) =(cid:26) 1 √N (Ei,j − Ej,i)A(Ei,j − Ej,i) =Xi<j =Xi6=j =Xi6=j =Xi6=j [Ei,jAEi,j − Ei,jAEj,i − Ej,iAEi,j + Ej,iAEj,i] (Ei,j AEi,j − Ei,jAEj,i) Aj,iEi,j −Xi6=j NXi=1 Aj,iEi,j + Aj,jEi,i Ai,iEi,i − NXi=1 Ai,iEi,i −Xi6=j Aj,jEi,i = A⊺ − Tr(A)IN , and dividing by N proves (3.6). Equation (3.5) then follows from (3.6) by setting A = IN . 11 Similarly, we compute Xi<j Tr(A(Ei,j − Ej,i))(Ei,j − Ej,i) =Xi<j =Xi6=j =Xi6=j =Xi6=j [Tr(AEi,j )Ei,j − Tr(AEi,j )Ej,i − Tr(AEj,i)Ei,j + Tr(AEj,i)Ej,i] [Tr(AEi,j )Ei,j − Tr(AEj,i)Ei,j ] Ai,jEi,j Aj,iEi,j + Aj,iEi,j −Xi6=j NXi=1 Ai,iEi,i −Xi6=j NXi=1 Ai,iEi,i − Ai,jEi,j = A⊺ − A, which is equivalent to (3.7). Equation (3.8) now follows from (3.7) by multiplying by B and taking tr. Remark 3.3. Our proof of the magic formulas for SO(N, R) is primarily computational, relying on the elementary matrix identities of Lemma 3.1. These results can also be (β) obtained more abstractly. For β ∈ {1, 2, 2′, 4}, let k N denote so(N, R), su(N ), u(N ), (β) N . In [13, Lemma 1.2.1], and sp(N ), resp., and let βk , defined as the L´evy provides an explicit decomposition of the Casimir element Ck tensor be an orthonormal basis for k (β) N (β) N Ck (β) N = XX∈β k (β) N X ⊗ X. Let Ea,b ⊗ Eb,a and P = Ea,b ⊗ Ea,b. For K ∈ {R, C, H}, set I(K) = {1, i, j, k} ∩ K and define two elements ReK and CoK of K ⊗R K by γ ⊗ γ−1 and γ ⊗ γ. Then for β ∈ {1, 2′4} (where 2′ corresponds to the value 2), the Casimir element Ck given by is (β) N Ck (β) N = 1 βN (cid:16)−T ⊗ ReK +P ⊗ CoK(cid:17) , and for β = 2, Csu(N ) = Cu(N ) − 1 compute any expression of the formPX∈β k (β) N N 2 iIN ⊗ iIN . This decomposition can be used to B(X, X), where B is an R-bilinear map. In particular, the magic formulas for SO(N, R), SU(N ), and Sp(N ) (cf. Propositions 3.2, 6.2, and 5.10) can be obtained in this way. 12 T = NXa,b=1 ReK = Xγ∈I(K) NXa,b=1 CoK = Xγ∈I(K) Proposition 3.4 (Derivative formulas for SO(N, R)). Let X ∈ βso(N,R). The following identities hold on SO(N, R) and SO(N, C): AjXAm−j, m < 0 mXj=1 eXAm = AjXAm−j, m ≥ 0 0Xj=m+1 eXAm = − eX tr(Am) = m · tr(XAm), m ∈ Z (m − j) tr(Aj )Am−j ∆SO(N,R)Am = 21m≥2 1 m−1Xj=1 m−1Xj=1 (m − j)Am−2j − Am, m ≥ 0 (−m + j) tr(Aj)Am−j ∆SO(N,R)Am = 21m≤−2 1 −1Xj=m+1 −1Xj=m+1 XX∈βso(N,R) eX tr(Am) · eXAp = N mp N 2 (Ap−m − Ap+m), m, p ∈ Z. (−m + j)Am−2j − m(N − 1) + m(N − 1) N − Am, m < 0 N N (3.9) (3.10) (3.11) (3.12) (3.13) (3.14) We emphasize that these derivative formulas hold true only in the case on SO(N, R) and SO(N, C), in which case A−1 = A⊺. Proof. The proof of (3.9), (3.10), and (3.11) for the U(N ) case are contained in [8]; the proof of these equations for the SO(N, R) case is identical. Next, by (3.9), we have, for m ≥ 0, eX 2Am = 21m≥2 Xj,k6=0 j+k+ℓ=m AjXAkXAℓ + mXj=1 AjX 2Am−j. (3.15) 13 Summing over all X ∈ βSO(N,R) and using the magic formulas (3.5) and (3.6), we have ∆SO(N,R)Am = 21m≥2 XX∈βso(N,R) Xk,j6=0 k+j+ℓ=m AkXAjXAℓ + XX∈βso(N,R) Ak(Aj)⊺Aℓ − tr(Aj)Ak+ℓ(cid:21) − AjX 2Am−j mXj=1 m(N − 1) Am N (cid:20) 1 N k+j+ℓ=m = 21m≥2 Xk,j6=0 = 21m≥2 1 m−1Xj=1 (m − j)Am−2j − N m(N − 1) Am, N (m − j) tr(Aj )Am−j m−1Xj=1 − which proves (3.12). Similarly, if m < 0, we use (3.10) to get eX 2Am = 21m≤−2 Xk,ℓ<0,j≤0 j+k+ℓ=m AjXAkXAℓ + 0Xj=m+1 AjX 2Am−j. Summing over all X ∈ βSO(N,R) and using the magic formulas (3.5) and (3.6), we have ∆SO(N,R)Am = 21m≥2 XX∈βso(N,R) Xk,ℓ<0,j≤0 j+k+ℓ=m AkXAjXAℓ + XX∈βso(N,R) Ak(Aj )⊺Aℓ − tr(Aj)Ak+ℓ(cid:21) + 0Xj=m+1 AjX 2Am−j m(N − 1) Am N (−m − j)Am−2j − (−m − j) tr(Aj )Am−j −1Xj=m+1 N j+k+ℓ=m (cid:20) 1 = 21m≥2 Xk,ℓ<0,j≤0 = 21m≥2 1 −1Xj=m+1 N m(N − 1) + Am, N which proves (3.13). For (3.14), we first assume m, p ≥ 0. Using equation (3.7) and the fact that A⊺ = A−1 14 for A ∈ SO(N, R), XX∈βso(N,R) eX tr(Am) · eXAp = XX∈βso(N,R) AkXAp−k! AjXAm−j pXk=1 tr mXj=1 tr(XAm) XAp−k  mXj=1 Ak XX∈βso(N,R) mAk XX∈βso(N,R) tr(XAm)X Ap−k mp N 2 (Ap−m − Ap+m). pXk=1 pXk=1 = = = For m < 0 and p ≥ 0, we have tr− Aj XAm−j pXk=1 0Xj=m+1 XX∈βso(N,R) eX tr(Am) · eXAp = XX∈βso(N,R)  tr(XAm) XAp−k pXk=1 0Xj=m+1 Ak XX∈βso(N,R) (−m)Ak XX∈βso(N,R) tr(XAm)X Ap−k pXk=1 = − mp N 2 (Ap−m − Ap+m). = − = AkXAp−k! For m ≥ 0 and p < 0, we have XX∈βso(N,R) eX tr(Am) · eXAp = XX∈βso(N,R) 0Xk=p+1 0Xk=p+1 = − mp N 2 (Ap−m − Ap+m). AjXAm−j− tr 0Xk=p+1 mXj=1 tr(XAm) XAp−k  mXj=1 Ak XX∈βso(N,R) mAk XX∈βso(N,R) tr(XAm)X Ap−k = − = AkXAp−k 15 Finally, for m, p < 0, we have XX∈βso(N,R) eX tr(Am) · eXAp = XX∈βso(N,R) 0Xk=p+1 0Xk=p+1 = = Aj XAm−j− − tr 0Xk=p+1 0Xj=m+1 tr(XAm) XAp−k  0Xj=m+1 Ak XX∈βso(N,R) (−m)Ak XX∈βso(N,R) tr(XAm)X Ap−k (−m)(−p) N 2 (Ap−m − Ap+m). = This proves (3.14). AkXAp−k 3.2. Intertwining formulas for ∆SO(N,R) In this section, we describe the trace polynomial functional calculus that will be necessary for the proof of Theorem 1.4, which contains the intertwining formulas for ∆SO(N,R), ∆SU(N ), and ∆Sp(N ). The space of trace polynomials was first introduced in [8, Definition 1.7], where it was used to prove analogous intertwining results for ∆U(N ). Using this framework, we then prove Theorem 1.4 for the SO(N, R) case. The computations in this section are related to, but not quite the same as, those used to establish the intertwining formulas for ∆U(N ) in [8, Sections 3-4]. Definition 3.5. Let C[u, u−1] denote the algebra of Laurent polynomials in a single variable u: C[u, u−1] =(Xk∈Z akuk : ak ∈ C, ak = 0 for all but finitely many k) , (3.16) with the usual polynomial multiplication. Let C[u] and C[u−1] denote the subalgebras consisting of polynomials in u and u−1, respectively. Let C[v] denote the algebra of polynomials in countably many commuting variables v = {v±1, v±2, . . .}, and let C[u, u−1; v] denote the algebra of polynomials in the commuting variables {u, u−1, v±1, v±2, . . .}. We now define the trace polynomial functional calculus: for P ∈ C[u, u−1; v], we have P (1) N (A) = [P (1)]N (A) := P (u; v)u=A,vk=tr(Ak),k6=0, P (2) N (A) = [P (2)]N (A) := P (u; v)u=A,vk=tr(Ak),k6=0, P (4) N (A) = [P (4)]N (A) := P (u; v)u=A,vk= etr(Ak),k6=0, 16 A ∈ SO(N, R) or SO(N, C), A ∈ SU(N ) or SL(N, C), A ∈ Sp(N ) or Sp(N, C). N , where P ∈ C[u, u−1; v], are called trace polynomials. We N and write PN when it is clear from context Functions of the form P (β) will often suppress the superscripts for P (β) which version of the trace polynomial functional calculus is being used. As an illustrative example, the trace polynomial P (1) P (u; v) = v1v6 N = PN on SO(N, R) associated with −3u2 − 3v2 5 is PN (A) = tr(A) tr(A−3)6A2 − 3 tr(A5)2IN . We now introduce the pseudodifferential operators on C[u, u−1; v] that appear in our intertwining formulas. Definition 3.6. Let R± denote the positive and negative projection operators and R− : C[u, u−1; v] → C[u−1; v] R+ : C[u, u−1; v] → C[u; v] defined by R+ ∞Xk=−∞ ukqk(v)! = ukqk(v), R− ∞Xk=−∞ ∞Xk=0 ukqk(v)! = −1Xk=0−∞ ukqk(v). In addition, for any k ∈ Z, let Muk denote the multiplication operator defined by Muk P (u; v) = ukP (u; v). 17 Definition 3.7. Define the following operators on C[u, u−1; v]: 1 = 1 = 2 = Z1 = Z + 2 − Y − 1 − Y − Y2 = Y + N = N0 + N1 = Xk≥1 ∞Xk=1 Y1 = Y + ∞Xk=2 ∞Xk=2 ∞Xk=1 1 = Xk≥1 2 = Xj,k≥1 2 − Z − 1 − Z − 1 − K− Z2 = Z + K1 = K+ K2 = K+ 2 − K− J = 2 Xk≥1 ∂u∂vk 2 = ku ∂ kvk ∂ ∂vk ∂ ∂u + u ∂uMu−kR+ − (R+ − R−), −1Xk=−∞ jvjvk−j ∂ −2Xk=−∞ (k − j)vk−2j ∂ ∂vk − ∂vk − vkuR+ ∂  k−1Xj=1  k−1Xj=1 u−k+1R+ ∂ ku−k+1 ∂2 ∂uMu−kR+ − ∂vk∂u − Xk≥1 ∂vj∂vk − Xj,k≥1 + Xk≥1 ∂vj∂vk jkvjvk ∂2 ∂2 jkvk−j + Xj,k≥1 , ∂vk ∂uMu−kR−, vkuR− ∂  jvjvk−j ∂ −1Xj=k+1  (k − j)vk−2j ∂ −2Xk=−∞ −1Xj=k+1 −1Xk=−∞ u−k+1R− ∂ ∂uMu−kR−, , ∂vk kuk+1 ∂2 ∂vk∂u , jkvk+j ∂2 ∂vj∂vk , (3.17) (3.18) (3.19) (3.20) (3.21) (3.22) (3.23) k2vk ∂ ∂vk + u ∂2 ∂u2 + u ∂ ∂u . (3.24) We set L0 = −N − 2Y1 − 2Y2, L(1) 1 = N + 2Z1 + 2Z2, L(2) 1 = 0, L(4) 1 = − L(2) 2 = −2K− 2L(1) L(4) 1 , 2 = 1 L(1) 2 = 2K1 + K2, 1 − K− 2 + J , 1 4L(1) 2 . For β ∈ {1, 2, 4}, we define D(β) N = L0 + 1 N L(β) 1 + 1 N 2L(β) 2 . (3.25) (3.26) (3.27) (3.28) (3.29) Notation 3.8. For m ∈ Z and A ∈ MN (C), let Wm(A) = Am, Vm(A) = tr(Am), and V(A) = {Vm(A)}m≥1. The functions Wm, Vm, and V implicitly depend on N , but we suppress the index, which should not cause confusion. Theorem 3.9 (Partial product rule). Let α, β ∈ C. For P ∈ C[u, u−1; v] and Q ∈ C[v], the operator αL0 + βL(1) satisfies 1 (cid:16)αL0 + βL(1) 1 (cid:17) (P Q) =(cid:16)(cid:16)αL0 + βL(1) 1 (cid:17) P(cid:17) Q + P(cid:16)(cid:16)αL0 + βL(1) 1 (cid:17) Q(cid:17) . 18 (3.30) Proof. From Definition 3.7, recall that L0 = −(N0 + N1) − 2Y1 − 2Y2, L(1) 1 = N0 + N1 + 2Z1 + 2Z2. Observe that N0, Y2, Z1 are first order differential operators on C[v], and so satisfy the product rule on C[u, u−1; v]. On the other hand, N1, Y1, and Z2 annihilate C[v] and satisfy, for P ∈ C[u, u−1; v] and Q ∈ C[v], N1(P Q) = (N1P )Q, Y1(P Q) = (Y1P )Q, Z2(P Q) = (Z2P )Q. Putting this together shows (3.30). Definition 3.10. The trace degree of a monomial in C[u, u−1; v] is defined to be deg(cid:16)uk0vk1 1 vk−1 −1 ··· vkm m vk−m −m (cid:17) = k0 + X1≤j≤m jkj. (3.31) We define the trace degree of any element in C[u, u−1; v] to be the highest trace degree of any of its monomial terms. For m ≥ 0, we define the subspace Cm[u, u−1; v] ⊆ C[u, u−1; v] to be Cm[u, u−1; v] = {P ∈ C[u, u−1; v] : deg P ≤ m}, (3.32) Cm[u, u−1; v]. consisting of polynomials of trace degree ≤ m. Note that Cm[u, u−1; v] is finite-dimensional and C[u, u−1; v] =Sm≥0 Corollary 3.11. Let m, N ∈ N and α, β ∈ C. The operators αL0+βL(1) N , and D(4) N preserve the finite dimensional subspace Cm[u, u−1; v] ⊆ C[u, u−1; v]. It follows that eαL0+βL(1) N are well-defined operators on each Cm[u, u−1; v] (via power series, cf. Remark 3.14) and hence are well-defined on all of C[u, u−1; v]. N , and eαD(4) N , D(2) 1 , D(1) N , eαD(2) 1 , eαD(1) 1 , D(1) N , D(2) N , and D(4) Proof. Recall that αL0 + βL(1) N are linear combinations of the operators J , N , Y1, Y2, Z1, Z2, K1, and K2 described in Definition 3.7. Each of these operators, in turn, are linear combinations of compositions of multiplication, dif- ferentiation, and projection operators. The multiplication and differentiation operators raise and lower trace degree, respectively, while projection operators R+ and R− leaves Cm[u, u−1; v] invariant. Keeping track of multiplication and differentiation operators in J , N , Y1, Y2, Z1, Z2, K1, and K2 shows that these operators preserve trace de- gree. Hence αL0 + βL(1) N preserve the finite dimensional subspace Cm[u, u−1; v] ⊆ C[u, u−1; v], and the result for eαL0+βL(1) N , and eαD(4) N , and D(4) N , D(2) 1 , D(1) N , eαD(2) 1 , eαD(1) N follows. Corollary 3.12. For any P ∈ C[u, u−1; v], Q ∈ C[v], and α, β, τ ∈ C, 1 )Q. 1 )(P Q) = e 2 (αL0+βL(1) 2 (αL0+βL(1) 2 (αL0+βL(1) e τ τ τ 1 )P · e (3.33) 19 Proof. Suppose L is a linear operator on C[u, u−1; v] which leaves C[v] and Cm[u, u−1; v] invariant and satisfies the partial product rule (3.30). Applying (3.30) repeatedly, we get 1 k!Lk(P Q) = 1 k! ∞Xk=0 1 = 1 1 1 j!k! ℓ!(k − ℓ)! (LℓP )(Lk−ℓQ) ∞Xk=0 ∞Xj,k=0 = eLP · eLQ. 2(cid:16)αL0 + βL(1) kXℓ=0(cid:18)k ℓ(cid:19)(LℓP )(Lk−ℓQ) = ∞Xk=0 kXℓ=0 j!LjP ∞Xk=0 (Lj P )(LkQ) = ∞Xj=0 1 (cid:17) leaves C[v] and Cm[u, u−1; v] invari- 2(cid:16)αL0 + βL(1) 1 (cid:17) concludes the proof. k!LkQ! By Definition 3.7, it is clear that τ ant, so setting L = τ We can now prove our main intertwining result for ∆SO(N,R). Proof of Theorem 1.4 for SO(N, R). It suffices to show that ∆SO(N,R)PN = [DN P ]N holds for P (u; v) = umq(v), where m ∈ Z \ {0} and q ∈ C[v]. Then PN = Wm · q(V), and by the product rule, ∆SO(N,R)PN = XX∈βso(N,R) eX 2(Wm · q(V)) = XX∈βso(N,R) eX(cid:16)(eXWm) · q(V) + Wm · eXq(V)(cid:17) = XX∈βso(N,R) (eX 2Wm) · q(V) + 2eXWm · eXq(V) + Wm · eX 2q(V) + 2 XX∈βso(N,R) eXWm · eXq(V) + Wm · (∆SO(N,R)q(V)) = (∆SO(N,R)Wm) · q(V) (3.34) For the first term in (3.34), we consider the cases m ≥ 0 and m < 0 separately. For 20 m ≥ 0, we apply (3.12) to get (∆SO(N,R)Wm) · q(V) (m − k)Wm−2k − 2 (m − k)VkWm−k# q(V) m−1Xk=1 N " ∞Xk=1 ∂uMu−kR+P!#N u−k+1R+ ∂ ∂uMu−kR+! P#N + vkuR+ ∂ m(N − 1) N = − Wm · q(V) ∂ N N − 1 + 21m≥2" 1 m−1Xk=1 N (cid:20)u ∂uR+P(cid:21)N − 2" ∞Xk=1 N (cid:20)u ∂uR+P(cid:21)N N (cid:20)u N − 1 N − 1 ∂ = − = − 2 N + 2 P ]N − 2[Y + [Z + ∂uR−P(cid:21)N − 2 N 2 annihilates C[u−1] while Y − 1 ,Z − ∂ 1 P ]N . (3.35) 1 P ]N . 2 P ]N + 2[Y − [Z − 2 annihilates C[u], so for all A similar computation, using (3.13), shows that for m < 0, (∆SO(N,R)Wm) · q(V) = We observe that Y + m ∈ Z, 1 ,Z + XX∈βso(N,R) eXWm · eXq(V) (∆SO(N,R)Wm) · q(V) = − N − 1 N [N1P ]N + 2 N [Z2P ]N − 2[Y1P ]N . (3.36) Next, by the chain rule and (3.14), the middle term in (3.34) is mk ∂vk ∂vk = XX∈βso(N,R) eXWm · Xk≥1(cid:18) ∂ q(cid:19) (V) · eXVk  XX∈βso(N,R) eXWm · eXVk(cid:18) ∂ q(cid:19) (V) = Xk≥1 N 2 (Wm−k − Wm+k)(cid:18) ∂ q(cid:19) (V) = Xk≥1 k(cid:18)W−k+1(cid:20) ∂ um(cid:21)N − Wk+1(cid:20) ∂ N 2 Xk≥1 N 2Xk≥1 ∂u∂vk PN k(u−k+1 − uk+1) ∂vk ∂u ∂u ∂2 = = 1 1 = 1 N 2 [K1P ]N . 21 um(cid:21)N(cid:19)(cid:18) ∂ ∂vk q(cid:19) (V) (3.37) (3.38) ∂u um(cid:3)N in line (3.37). Finally, for the last term in (3.34), we have, for each X ∈ βso(N,R), where we have used the fact that mWm+j = Wj+1(cid:2) ∂ eX 2q(V) = eXXk≥1(cid:18) ∂ = Xk≥1(cid:18) ∂ q(cid:19) (V) · eXVk q(cid:19) (V) · eX 2Vk + Xj,k≥1(cid:18) ∂2 ∂vk ∂vk ∂vj∂vk q(cid:19) (V) · eXVj · eXVk. (3.39) In summing the first term in (3.39) over X ∈ βso(N,R), we break up the sum for positive and negative k. Using (3.12) and (3.13), we have (k − ℓ)Vk−2ℓ(cid:18) ∂ ∂vk q(cid:19) (V) 2 N k−1Xℓ=1 ∞Xk=2 q(cid:19) (V) ∂vk XX∈βso(N,R) ∞Xk=1(cid:18) ∂ ∂vk N − 1 N = − and XX∈βso(N,R) −1Xk=−∞(cid:18) ∂ ∂vk = N − 1 N − 2 − 2 ∂vk q(cid:19) (V) · eX 2Vk kVk(cid:18) ∂ q(cid:19) (V) + ∞Xk=1 (k − ℓ)VℓVk−ℓ(cid:18) ∂ k−1Xℓ=1 ∞Xk=2 q(cid:19) (V) · eX 2Vk kVk(cid:18) ∂ −1Xk=−∞ −1Xℓ=k+1 −2Xk=−∞ 2 N ∂vk (3.40) q(cid:19) (V) (3.41) (3.42) (3.43) Summing the second term in (3.39) over X ∈ βso(N,R) and using (3.14) yields (k − ℓ)Vk−2ℓ(cid:18) ∂ ∂vk ∂vk q(cid:19) (V) − −2Xk=−∞ (−k + ℓ)VℓV−k−ℓ(cid:18) ∂ −1Xℓ=k+1 q(cid:19) (V). q(cid:19) (V) · eXVj · eXVk jk(Vk−j − Vk+j )(cid:18) ∂2 ∂vj∂vk ∂vj∂vk N − 1 N N0 + 2 N Z1 − 2Y2 + q(cid:19) (V) N 2K2(cid:19) P(cid:21)N 1 XX∈βso(N,R) Xj,k≥1(cid:18) ∂2 N 2 Xj,k≥1 Wm · (∆SO(N,R)q(V)) =(cid:20)(cid:18)− = 1 Adding (3.40), (3.41), and (3.42) together and multiplying by Wm, we get Combining (3.36), (3.38), and (3.43) proves (1.11) for SO(N, R) (β = 1). Equation (1.12) now follows from Corollary (3.11). 22 Remark 3.13. A similar intertwining formula is proven in [8] for the U(N ) case, where for any P ∈ C[u, u−1; v], , (3.44) (3.45) ∆U(N )PN =(cid:20)(cid:18)L0 − 1 N 2 (2K− 1 + K− 2 )(cid:19) P(cid:21)N and, subsequently, for τ ∈ C, τ 2 ∆U(N )PN = [e e τ 2 (L0− 1 N 2 (2K− 1 +K− 2 ))P ]N . Remark 3.14. In this paper, we are primarily concerned with computing the Segal- Bargmann transform on trace polynomials. In this setting, if D is a left-invariant differ- ential operator on K, where K is a compact type matrix Lie group, then we can compute the action of eD on a trace polynomial f via the power series (eDf )(x) = ∞Xn=0 Dnf! (x). 1 n! (3.46) In particular, if D = τ series definition. In [9], it is shown that e KC. Moreover, this analytic continuation is equal to BK 2 ∆K, we can define e τ τ 2 ∆K f as a function on K using the power 2 ∆K f has a unique analytic continuation to s,τ f . Proof of Theorem 1.5 for SO(N, R). By the intertwining formula (1.12), we have τ 2 ∆SO(N,R)PN = [e e τ 2 D(1) N P ]N . τ 2 D(1) Corollary 3.11 shows that [e N P ]N is a trace polynomial on SO(N, R), so its ana- lytic continuation to SO(N, C) is given by the same trace polynomial function. Hence N P ]N , viewed as a holomorphic function on SO(N, C), is the analytic continuation [e 2 ∆SO(N,R)PN , and so is equal to BSO(N,R) of e 2 D(1) PN . s,τ τ τ 3.3. Intertwining formulas for ASO(N,C) s,τ We now prove an intertwining formula for SO(N, C) that is analogous to the intertwining formula for SO(N, R) in Theorem 1.4. To do so, we use the word polynomial space introduced in [8, Notation 3.21], which was used to prove a similar intertwining formula for GLN (C) = U(N )C. Notation 3.15. For m ∈ N, define Em to be the set of words ε : {1, . . . , m} → {±1,±∗}. We denote the length of a word ε ∈ Em by ε = m and set E = ∪m≥0Em. We define the word polynomial space W to be the space of polynomials in the commuting variables {vε}ε∈Em. For j, k ∈ Z not both zero, we define the words W = C [{vε}ε∈E ] , ε(j, k) = z } j times k times ±∗, . . . ,±∗) ∈ Ej+k, (±1, . . . ,±1, 23 { z } { (3.47) where the first j slots contain +1 if j > 0 and −1 if j < 0, and the last k slots contain +∗ if k > 0 and −∗ if k < 0. We set vε(0,0) = 1. Notation 3.16. For ε = (ε1, . . . , εm) ∈ Em and A ∈ SO(N, C), we define Aε = Aε1 ··· Aεm , where A+∗ := A∗ and A−∗ := (A∗)−1. If P ∈ W , we define a function PN : SO(N, C) → C by PN (A) = P (V(A)), where and V(A) = {Vε(A) : ε ∈ E } Vε(A) = tr(Aε) = tr(Aε1 ··· Aεm ). Notation 3.17. We define the inclusion maps ι, ι∗ : C[v] ֒→ W , with ι linear and ι∗ conjugate linear, by (3.48) The maps ι and ι∗ intertwine with the evaluation map in the following way: for Q ∈ C[v], (3.49) [ι∗(Q)]N (A) = QN (A)∗. [ι(Q)]N (A) = QN (A), ι∗(vk) = vε(0,k). ι(vk) = vε(k,0), Definition 3.18. The trace degree of a monomialQm j=1 vkj εj ∈ W is defined to be deg mYj=1 vkj εj = mXj=1 kjεj, and the trace degree of an arbitrary element of W is the highest trace degree of any of its monomial terms. For each m ∈ N, we let Wm denote the finite dimensional subspace of W Wm = {P ∈ W : deg(P ) ≤ m} ⊆ C[{vε}ε≤m], Wm. so that W =S∞ m=1 ε such that ε Theorem 3.19. Fix s ∈ R and τ = t + iθ ∈ C. There are collections {Qs,τ {Rs,τ ε,δ : ε, δ ∈ Eo in W such that: : ε ∈ E }, andnSs,τ 1. for each ε ∈ E , Qs,τ and Rs,τ ε ε ε are certain finite sums of monomials of trace degree : ε ∈ E }, ASO(N,C) s,τ Vε = [Qs,τ ε ]N + 1 N [Rs,τ ε ]N , (3.50) 2. for ε, δ ∈ E , Ss,τ that ε,δ is a certain finite sum of monomials of trace degree ε + δ such dSO(N,R)Xℓ=1 (cid:20)(cid:18)s − t 2(cid:19) (eXℓVε)(eXℓVδ) + t 2 (eYℓVε)(eYℓVδ) − θ(eXℓVε)(eYℓVδ)(cid:21) = 1 N 2 [Ss,τ ε,δ ]N . (3.51) 24 For the proof below, we use the following conventions: let βso(N,R) = {Xℓ} denote an orthonormal basis for so(N, R), with β+ = βso(N,R) and β− = iβso(N,R). If X ∈ MN (C) and A ∈ SO(N, C), (AX)1 := AX, (AX)∗ := X ∗A∗, (3.52) In addition, we will be liberal in our use of ± to denote a sign that may vary for different terms and on different sides of an equation, since we will not require a precise formula for our word polynomials beyond what is described in the theorem statement. dSO(N,R) ℓ=1 (AX)−1 := −XA−1, (AX)−∗ := −A∗X ∗. Proof. Fix a word ε = (ε1, . . . , εm) ∈ E . Then for each X ∈ β± and A ∈ SO(N, C), so tr(Aε1 ··· (AX)εj ··· Aεm ), mXj=1 (eXVε)(A) = mXj=1 tr(Aε1 ··· (AX 2)εj ··· Aεm ) + 2 X1≤j<k≤m tr(Aε1 ··· (AX)εj ··· (AX)εk ··· Aεm ) (eX 2Vε)(A) = (3.53) (3.54) (3.55) Applying magic formula (3.5) to sum each term in (3.54) over β±, we have XX∈β± tr(Aε1 ··· (AX 2)εj ··· Aεm ) = ± N − 1 N tr(Aε1 ··· Aεj ··· Aεm ) = ± N − 1 N Vε(A). (3.56) Summing over 1 ≤ j ≤ m now gives XX∈β± mXj=1 tr(Aε1 ··· (AX 2)εj ··· Aεm ) = N − 1 N n±(ε)Vε(A), (3.57) where n±(ε) ∈ Z and n±(ε) ≤ ε. For (3.55), we can express each term in the sum as (3.58) j,k XAε1 j,k XAε2 j,k ), tr(Aε1 ··· (AX)εj ··· (AX)εk ··· Aεm ) = ± tr(Aε0 j,kε1 j,k, and ε2 j,k are substrings of ε such that ε0 j,k, ε1 so ε0 X ∈ β± by magic formula (3.6), we have j,kε2 j,k = ε. Summing (3.58) over XX∈β± tr(Aε1 ···(AX)εj ··· (AX)εk ··· Aεm ) = ±(cid:20) 1 N 1 N = ± tr(Aε0 j,k (Aε1 j,k )−1Aε2 j,k ) − tr(Aε0 j,k Aε2 j,k ) tr(Aε1 j,k )(cid:21) Vε3 j,k (A) ± Vε0 j,kε2 j,k (A)Vε1 j,k (A), 25 (3.59) (3.60) j,k is the word of length ε such that Vε3 where ε3 summing (3.55) over X ∈ β± gives j,k (A) = tr(Aε0 j,k (Aε1 j,k )−1Aε2 j,k ). Thus XX∈β± X1≤j<k≤m tr(Aε1 ··· (AX)εj ··· (AX)εk ··· Aεm ) N X1≤j<k≤m (A) + X1≤j<k≤m ±Vε3 = 1 j,k ±Vε0 j,kε2 j,k (A)Vε1 j,k (A). We define word polynomials corresponding to (3.61) by Q± Q± ε,0 = X1≤j<k≤m ε,1 = X1≤j<k≤m ±vε3 j,k ±vε0 j,kε2 j,k vε1 j,k . (3.61) (3.62) Now if X ∈ β+ and Y = iX ∈ β−, then (eY Vε)(A) = and mXj=1 tr(Aε1 ··· (AY )εj ··· Aεm ) = i mXj=1 ± tr(Aε1 ··· (AX)εj ··· Aεm ), (3.63) ± tr(Aε1 ··· (AX 2)εj ··· Aεm ) (eXeY Vε)(A) =i mXj=1 + 2i X1≤j<k≤m ± tr(Aε1 ··· (AX)εj ··· (AX)εk ··· Aεm ). (3.64) (3.65) A similar argument to the above shows that summing (3.64) over X ∈ β+ gives ± tr(Aε1 ··· (AX 2)εj ··· Aεm ) = N − 1 N η(ε)Vε(A) (3.66) mXj=1 where η(ε) ∈ Z and η(ε) ≤ ε. In addition, summing (3.65) over X ∈ β+, we have XX∈β+ XX∈β+ X1≤j<k≤m We define corresponding word polynomials ± tr(Aε1 ··· (AX)εj ··· (AX)εk ··· Aεm ) (A) + X1≤j<k≤m = 1 j,k ±Vε0 j,kε2 j,k Vε3 N X1≤j<k≤m Qε,2 = i X1≤j<k≤m Qε,3 = i X1≤j<k≤m 26 ±vε3 j,k ±vε0 j,kε2 j,k vε1 j,k . (A)Vε1 j,k (A). (3.67) (3.68) (3.69) Putting this together, we define Qs,τ ε =(cid:18)s − ε =(cid:18)s − Rs,τ t t 2(cid:19)(cid:0)n+(ε)vε + 2Q+ ε,1(cid:1) + 2(cid:19)(cid:0)−n+(ε)vε + 2Q+ ε,0(cid:1) + t 2(cid:0)n−(ε)vε + 2Q− 2(cid:0)−n−(ε)vε + 2Q− ε,1(cid:1) − θ (iη(ε) + 2Qε,3) , ε,0(cid:1) − θ (−iη(ε) + 2Qε,2) . t (3.71) (3.70) By construction, Qs,τ For part (2), fix δ ∈ E with n = δ. For each X ∈ β±, and Rs,τ ε ε are of the required form and satisfy (3.50). (eXVδ)(A)(eXVε)(A) = mXj=1 nXk=1 tr(Aε1 ··· (AX)εj ··· Aεm ) tr(Aδ1 ··· (AX)δk ··· Aδn ). (3.72) Using the invariance of the trace under cyclic permutations, we can write the terms in the above sum as ± tr(XAε(j)) tr(XAδ(k)), where ε(j) and δ(k) are particular cyclic permutations of ε and δ. Summing over X ∈ β± and applying magic formula (3.8), along with the fact that X ∈ so(N, C) is skew- symmetric, XX∈β± (eXVδ)(A)(eXVε)(A) = = 1 N 2 1 N 2 mXj=1 mXj=1 nXk=1 nXk=1 ±(cid:16)tr((Aε(j))−1Aδ(k)) − tr(Aε(j)Aδ(k)(cid:17) ±(Vε(j)⊺δ(k)(A) − Vε(j)δ(k)(A)), (3.73) where ε(j)⊺ is the word satisfying Vε(j)⊺ (A) = Vε(j)(A)−1. (For example, if ε(j) = (−∗,−1,∗, 1, 1), then ε(j)⊺ = (−1,−1,−∗, 1,∗).) We define the associated word polyno- mials S± ε,δ = mXj=1 nXk=1 ±(vε(j)⊺δ(k) − vε(j)δ(k)). (3.74) Using magic formula (3.8) to sum over X ∈ β+ gives Next, if X ∈ β+ and Y = iX ∈ β−, then dSO(N,R)Xℓ=1 (eXVε)(A)(eY Vδ)(A) = i mXj=1 nXk=1 (eXℓVε)(A)(eYℓVδ)(A) = mXj=1 ε,δ = i i N 2 S′ Hence we define the associated word polynomial mXj=1 nXk=1 ± tr(XAε(j)) tr(XAδ(k)). (3.75) nXk=1 ±(Vε(j)⊺δ(k)(A) − Vε(j)δ(k)(A)). (3.76) ±(vε(j)⊺δ(k) − vε(j)δ(k)) 27 (3.77) (where the signs in (3.77) do not necessarily correspond to those in (3.74)). Finally, we define Ss,τ ε,δ =(cid:18)s − t 2(cid:19) S+ ε,δ + t 2 S− ε,δ − θS′ ε,δ, (3.78) where we have constructed Ss,τ and satisfies (3.51). ε,δ so that it is a sum of monomials of trace degree ε + δ s,τ ). Fix s ∈ R and τ = t + iθ ∈ C, ε,δ : ε, δ ∈ Eo be as in Theorem 3.19. t 2(cid:19) (eXℓVε)(eXℓVδ) (eYℓVε)(eYℓVδ) − θ(eXℓVε)(eYℓVδ)(cid:21). t 2 The following result will be useful for the proof of Theorem 1.6. It appears in [8] for the GL(N, C) case, with the proof virtually unchanged for SO(N, C). 28 Rs,τ ε ∂2 ε ε Qs,τ ε Ss,τ ε,δ 1 1 1 2 = 1 = 0 = ∂vε∂vδ ∂ ∂vε ∂ ∂vε : ε ∈ E }, {Rs,τ Then for all N ∈ N and P ∈ W , Theorem 3.20 (Intertwining formula for ASO(N,C) : ε ∈ E }, and nSs,τ and let {Qs,τ Define first and second order differential operators 2Xε∈E eLs,τ 2Xε∈E eLs,τ 2 Xε,δ∈E eLs,τ PN =(cid:20)(cid:18)eLs,τ N eLs,τ ∂vε(cid:19) (V) · (eXVε)(cid:21) ∂vε(cid:19) (V) · (eX 2Vε) + Xε,δ∈E(cid:18) ∂2P ∂vε(cid:19) (V) · ASO(N,C) ∂vε∂vδ(cid:19) (V) + Xε,δ∈E(cid:18) ∂2P eX 2PN =Xε∈E eX(cid:20)(cid:18) ∂P =Xε∈E(cid:18) ∂P PN =Xε∈E(cid:18) ∂P dSO(N,R)Xℓ=1 (cid:20)(cid:18)s − ASO(N,C) s,τ 1 2ASO(N,C) s,τ Hence 1 + 0 + s,τ Vε 1 Proof. For each X ∈ MN (C), we apply the chain rule to get The result now follows from Theorem 3.19. + (3.79) (3.80) (3.81) . (3.82) 2 (cid:19) P(cid:21)N 1 N 2 eLs,τ ∂vε∂vδ(cid:19) (V) · (eXVε)(eXVδ). Lemma 3.21 ([8, Lemma 3.28]). There exists a sesquilinear form (with the second argument conjugate linear) B : C[u, u−1; v] × C[u, u−1; v] → W such that, for all P, Q ∈ C[u, u−1; v], we have deg(B(P, Q)) = deg(P ) + deg(Q) and [B(P, Q)]N (A) = tr[PN (A)QN (A)∗] for all A ∈ SO(N, C). 4. Limit theorems for the Segal-Bargmann transform on SO(N, R) We now use the results from the previous section to identify the limit of the Segal- Bargmann transform on SO(N, R) as N → ∞. In Section 4.1, we prove a few prelimi- and µSO(N,C) nary results regarding the way in which the heat kernel measures ρSO(N,R) concentrate their mass as N → ∞. Then, in Section 4.2, we prove the SO(N, R) version of Theorem 1.6, the main limit theorem for BSO(N,R) s,τ . s s,τ 4.1. Concentration of measures The following lemma is an essential component of our main limit theorems. It is a known result (see [12, Corollary 6.2.32]); we include a direct proof here for convenience. Lemma 4.1. Let X, Y ∈ M (N, C) and suppose k · k is a submultiplicative matrix norm. Then keX+Y − eXk ≤ kY kekXkekY k. Proof. For n ≥ 0, note that k(X + Y )n − X nk ≤ (kXk + kY k)n − kXkn, where the inequality follows by expanding (X +Y )n, which includes an X n, then applying the submultiplicativity of the matrix norm, and then recombining terms. Hence X n n! (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) n! − (X + Y )n keX+Y − eXk =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ∞Xn=0 ∞Xn=0 ∞Xn=0 1 n!k(X + Y )n − X nk ∞Xn=0 1 ≤ n! = ekXk+kY k − ekXk. ≤ [(kXk + kY k)n − kXkn] It remains to show that ekXk+kY k− ekXk ≤ kY kekXkekY k, which is equivalent to showing that ey − 1 ≤ yey for y ≥ 0. Rearranging, this inequality is equivalent to 1 − y ≤ e−y, which holds, in fact, for all y ∈ R: by Bernoulli's inequality, ≥ 1 − y. e−y = lim y n→∞(cid:16)1 − 29 n(cid:17)n Corollary 4.2. Let V be a finite-dimensional normed vector space over C and suppose that that L0, L1, and L2 are operators on V . Then there exist constants C1, C2, C3 < ∞ depending on L0, L1, L2,k · kV such that for N sufficiently large, N L1+ 1 keL0+ 1 N 2 L2 − eL0kEnd(V ) ≤ N L1+ 1 keL0+ 1 N 2 L2 − eL0+ 1 N L1 − eL0kEnd(V ) ≤ C3 N keL0+ 1 C1 N N L1kEnd(V ) ≤ , C2 N 2 (4.1) (4.2) (4.3) where k · kEnd(V ) is the operator norm on End(V ) induced by k · kV . Hence if ϕ ∈ V ∗ is a linear functional, then for N sufficiently large, ϕ(eL0+ 1 N L0+ 1 N 2 L2v) − ϕ(eL0v) ≤ C1 N kϕkV ∗kvV , ϕ(eL0+ 1 N L1+ 1 N 2 L2v) − ϕ(eL0+ 1 N L1v) − ϕ(eL0 v) ≤ ϕ(eL0+ 1 N L1v) ≤ C2 N 2kϕkV ∗kvV C3 N kϕkV ∗kvV (4.4) (4.5) (4.6) where k · kV ∗ is the dual norm on V ∗. Proof. We set X = L0 and Y = 1 N L1+ 1 keL0+ 1 1 N 2 L2 as in Lemma 4.1, so N L1 + 1 N kL1kEnd(V )ekL0kEnd(V ) ek 1 1 N kL1kEnd(V )ekL0kEnd(V ) e N 2 L2 − eL0kEnd(V ) ≤ ≤ kL1kEnd(V )+ 1 N 2 0 1 N0 N L1+ 1 N 2 L2kEnd(V ) 1 N kL1kEnd(V )+ 1 N 2 kL2kEnd(V ) . kL2kEnd(V ) ≤ 2. Then (4.1) holds for all Choose N0 ∈ N such that e N ≥ N0 by setting C1 = 2kL1kEnd(V )ekL0kEnd(V ) . Equation (4.4) then follows. The proofs of (4.2), (4.3), and subsequently (4.5) and (4.6), are similar. We now introduce notation and establish a few preliminary results which we will use for proving Theorem 1.6 for SO(N, R) in the next section. Definition 4.3. Define a family of holomorphic functions {νk}k∈Z on C by setting ν0(τ ) ≡ 1 and for k 6= 0, νk(τ ) = e− k 2 τ (−τ )j j! kj−1(cid:18) k j + 1(cid:19). k−1Xj=0 For τ ∈ C, define the trace evaluation map πτ : C[u, u−1; v] → C[u, u−1] by (πτ P )(u) = P (u; v)vk=νk(τ ),k6=0. 30 (4.7) (4.8) For a concrete example, if P (u; v) = v3v2 −7u6 + 9v1v5 −4, then (πτ P )(u) = ν3(τ )ν−7(τ )2u6 + 9ν1(τ )ν−4(τ )5. The following result is a key tool in proving our main limit theorems. Theorem 4.4 (Biane, [3]). For each s > 0 and k ∈ Z, N→∞ZU(N ) lim tr(U k) ρU(N ) s (U ) dU = νk(s). For P ∈ C[u, u−1; v], let P (u; 1) = P (u; v)vk=1,k6=0. Using the intertwining formula (3.45) for the U(N ) case and Theorem 4.4, we have νk(s) = lim N→∞ZU(N ) = lim N→∞ (e s 2 (L0− 1 N 2 (2K− 1 +K− 2 ))vk)(1). tr(U k) ρU(N ) s (U ) dU = lim N→∞ (e s 2 ∆U(N ) tr[(·)k])(IN ) (4.9) (4.10) 1 , and K− Since L0, K− 2 are linear operators which preserve the finite-dimensional vector 2 ))vk)(1) to be the evaluation of the space Ck[u, u−1; v], we consider (e linear functional ϕ(Q) := Q(1) on the finite-dimensional space Cm[v]. Hence Corollary 4.2 applies, so taking the limit as N → ∞, we have N 2 (2K− 2 (L0− 1 1 +K− s νk(s) = (e s 2 L0 vk)(1). (4.11) This yields the following version of Theorem 4.4 for the SO(N, R) case, which is due to L´evy [13]. Lemma 4.5 (L´evy, [13]). For s > 0 and k ∈ Z, tr(Ak) ρSO(N,R) s (A) dA = νk (s) . (4.12) N→∞ZSO(N,R) lim Proof. We have N→∞ZSO(N,R) lim tr(Ak) ρSO(N,R) s (A) dA = lim N→∞ (e = lim N→∞ (e = νk (s) , s 2 ∆SO(N,R) tr[(·)k])(IN ) 2 (L0+ 1 N L(1) N 2 L(1) 1 + 1 s 2 )vk)(1) = (e where we have again applied Corollary 4.2 as in (4.11). Lemma 4.6. For Q ∈ C[v] and τ ∈ C, we have (cid:0)e τ 2 L0Q(cid:1) (1) = πτ Q. 31 s 2 L0 vk)(1) (4.13) Proof. By Theorem 3.9, e τ 2 L0 is a homomorphism of C[v]. Hence it suffices to show that (cid:0)eτ L0vk(cid:1) (1) = πτ (vk) = νk (τ ) (4.14) τ k (cid:12)(cid:12)(cid:0)e for each k ∈ Z. To see that it holds for all τ ∈ C, we expand the left-hand side as a power series in τ . Fix a norm k·kWk on the finite-dimensional vector space Wk := Ck[u, u−1; v]. Let k·kEnd(Wk) denote the operator norm on End(Wk) induced by k·kWk, and let k·kW ∗ denote the corresponding dual norm on Wk. Let ϕ be the linear functional acting on Wk defined by ϕ(P ) = P (1). Then for any τ ∈ C, (L0)nvk(cid:19) (1)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 2 L0vk(cid:1) (1)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ∞Xn=0(cid:18) 1 n!(cid:16) τ 2(cid:17)n n!(cid:18)τ 2 (cid:19)n ∞Xn=0 ϕ((L0)nvk)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) and the right-hand side converges by the ratio test. Thus the function τ 7→(cid:0)e 2 L0 vk(cid:1) (1) is analytic on the complex plane. By (4.11), this function agrees with the entire function τ 7→ νk (τ ) for τ ∈ R, and so it also agrees for all τ ∈ C. Lemma 4.7. Let s > 0 and τ = t + iθ ∈ D(s, s). For any Q ∈ C[v], n!(cid:16) τ 2(cid:17)n End(Wk)kvkkWk , =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) k kL0kn ∞Xn=0 kϕkW ∗ ≤ 1 1 τ eLs,τ 0 eLs,τ 0 (cid:16)e (cid:16)e ι(Q)(cid:17) (1) = πs−τ Q, ι∗(Q)(cid:17) (1) = πs−τ Q. (4.15) (4.16) (4.18) (4.19) (4.20) Proof. First, observe that if f : SO(N, C) → MN (C) is holomorphic, then gJXf = ieXf for all X ∈ so(N, R). Thus 2 eX 2 − iθeX 2(cid:21) f = (s − τ )∆SO(N,R)f. ASO(N,C) fSO(N,R) = XX∈βso(N,R)(cid:20)(cid:18)s − 2(cid:19) eX 2 − s,τ t t In particular, since QN is a trace polynomial, 1 2 ASO(N,C) s,τ e QN = e 1 2 (s−τ )∆SO(N,R)QN . (4.17) Applying intertwining formulas (3.82) and (1.12) to the left side and using (3.49) gives eLs,τ 0 + 1 N he eLs,τ 1 + 1 N 2 eLs,τ 2 ι(Q)iN = [e 1 2 (s−τ )D(1) N Q]N . Hence for Q ∈ C[v], eLs,τ eLs,τ 0 + 1 N 1 + 1 N 2 eLs,τ 2 (cid:16)e ι(Q)(cid:17) (1) =(cid:16)e 1 2 (s−τ )(L0+ 1 N L(1) 1 + 1 N 2 L(1) 2 )Q(cid:17) (1). If deg Q = m, then we can view the left and right sides as the evaluation of a linear functional ϕ(P ) = P (1) on the finite-dimensional spaces Wm and Cm[v], respectively. Thus we may apply Corollary 4.2 to take the limit as N → ∞, which yields eLs,τ 0 (cid:16)e ι(Q)(cid:17) (1) =(cid:16)e 32 1 2 (s−τ )L0Q(cid:17) (1). By Lemma 4.6, the expression on the right is πs−τ Q. Equation (4.16) is proved similarly, using the fact that if f : SO(N, C) → MN (C) is antiholomorphic, thengJXf = −ieXf for all X ∈ so(N, R). Theorem 4.8. Let s > 0, τ ∈ D(s, s), and k ∈ Z. Then N→∞ZSO(N,C) lim tr(Ak) µSO(N,C) s,τ (A) dA = νk (s − τ ) . (4.21) Proof. We use (2.4) and intertwining formula (3.82) to get ZSO(N,C) tr(Ak) µSO(N,C) s,τ 1 2 ASO(N,C) s,τ tr[(·)k](cid:17) (IN ) eLs,τ 0 + 1 N eLs,τ 1 + 1 N 2 eLs,τ 2 (A) dA =(cid:16)e =(cid:16)e ι(vk)(cid:17) (1). Viewing the last expression as a linear functional acting on ι(vk) ∈ Wk, we apply Corollary 4.2 to take the limit as N → ∞. Combining this with Lemma 4.7, we have N→∞ZSO(N,C) lim tr(Ak) µSO(N,C) s,τ (A) dA =(cid:16)e eLs,τ 0 ι(vk)(cid:17) (1) = πs−τ vk = νk (s − τ ) . 4.2. The free Segal-Bargmann transform for SO(N, R) In this section, we prove our main result on the large-N limit of the Segal-Bargmann transform for SO(N, R). The free Segal-Bargmann transform Gs,τ and the free inverse Segal-Bargmann transform Hs,τ appearing in Theorem 1.6 have explicit formulas, which we first describe. Definition 4.9. Let L0 be the pseudodifferential operator on C[u, u−1; v] from Definition 3.7. We define the free Segal-Bargmann transform to be the map Gs,τ : C[u, u−1] → C[u, u−1] given by (4.22) 2 L0 , τ Gs,τ = πs−τ ◦ e and we define the free inverse Segal-Bargmann transform to be the map Hs,τ : C[u, u−1] → C[u, u−1] given by Hs,τ = πs ◦ e− τ 2 L0. Before proving Theorem 1.6, we require the following result. Theorem 4.10. Let s > 0 and τ = t + iθ ∈ D(s, s). For any P ∈ C[u, u−1; v], L2(ρSO(N,R) s ) kPN − [πsP ]Nk2 kPN − [πs−τ P ]Nk2 L2(µSO(N,C) s,τ and = O(cid:18) 1 N 2(cid:19) , = O(cid:18) 1 N 2(cid:19) . ) (4.23) (4.24) (4.25) This result shows that with respect to the heat kernel measure, the space of trace poly- nomials concentrates onto the space of polynomials as N → ∞. 33 Proof. We first show (4.25). It suffices to prove the result for polynomials of the form P (u; v) = ukQ(v) for k ∈ Z and Q ∈ C[v]; (4.25) then holds on all of C[u, u−1; v] by the triangle inequality. For such a polynomial, P (u; v) − πs−τ P (u; v) = uk[Q(v) − πs−τ Q] = ukRs−τ , where we let Rs−τ = Q − πs−τ Q. Hence for A ∈ SO(N, C), kPN (A) − [πs−τ P ]N (A)k2 MN (C) = tr(cid:0)Ak[Rs−τ ]N (A)[Rs−τ ]N (A)∗A∗k(cid:1) = tr(AkA∗k)[Rs−τ ]N (A)[Rs−τ ]N (A)∗ = [vε(k,k)ι(Rs−τ )ι∗(Rs−τ )]N (A). Along with intertwining formula (3.82), this allows us to compute (4.26) (4.27) (cid:16)kPN − [πs−τ P ]Nk2 MN (C)(cid:17) (IN ) eLs,τ eLs,τ 1 + 1 N 2 (4.28) 2 (vε(k,k)ι(Rs−τ )ι∗(Rs−τ ))(cid:17) (1) 1 (vε(k,k)ι(Rs−τ )ι∗(Rs−τ ))(cid:17) (1)(cid:12)(cid:12)(cid:12)(4.29) (4.30) kPN − [πs−τ P ]Nk2 L2(µSO(N,C) s,τ ) 1 2 ASO(N,C) s,τ = e eLs,τ 0 + 1 N =(cid:16)e By the triangle inequality, the last line is bounded by N N eLs,τ eLs,τ eLs,τ eLs,τ eDs,τ 0 + 1 0 + 1 (cid:12)(cid:12)(cid:12)(cid:16)e 1 (vε(k,k)ι(Rs−τ )ι∗(Rs−τ ))(cid:17) (1)(cid:12)(cid:12)(cid:12) . N (vε(k,k)ι(Rs−τ )ι∗(Rs−τ ))(cid:17) (1) −(cid:16)e +(cid:12)(cid:12)(cid:12)(cid:16)e N (vε(k,k)ι(Rs−τ )ι∗(Rs−τ ))(cid:17) (1) −(cid:16)e N 2(cid:19) . = O(cid:18) 1 (cid:12)(cid:12)(cid:12)(cid:16)e To bound (4.30), recall from Theorem 3.20 that eLs,τ operator on Wm, so e 0 + 1 0 + 1 0 + 1 eDs,τ eLs,τ eLs,τ eLs,τ eLs,τ N N 1 eLs,τ 0 + 1 N e eLs,τ 1 (vε(k,k)ι(Rs−τ )ι∗(Rs−τ )) eLs,τ 1 vε(k,k) · e 0 + 1 = e eLs,τ N If m = deg vε(k,k)ι(Rs−τ )ι∗(Rs−τ ), we can interpret (4.29) as the evaluation of the linear functional ϕ(R) = R(1) on Wm, so by Corollary 4.2, 1 (vε(k,k)ι(Rs−τ )ι∗(Rs−τ ))(cid:17) (1)(cid:12)(cid:12)(cid:12) N eLs,τ 1 is a first-order differential (4.31) is an algebra homomorphism, i.e. eLs,τ 0 + 1 N eLs,τ 1 ι(Rs−τ ) · e eLs,τ 0 + 1 N eLs,τ 1 ι∗(Rs−τ ). (4.32) For any T ∈ W , 0 + 1 eLs,τ N (cid:12)(cid:12)(cid:12)(e eLs,τ 1 T )(1)(cid:12)(cid:12)(cid:12) ≤(cid:12)(cid:12)(cid:12)(e By Corollary 4.2, the first term on the right side of (4.33) is O(1/N ), while the second term is a constant not depending on N . Moreover, by Lemma 4.7, eLs,τ 0 + 1 N eLs,τ 1 T )(1) − (e eLs,τ 0 T )(1)(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)(cid:12)(e eLs,τ 0 T )(1)(cid:12)(cid:12)(cid:12) (4.33) eLs,τ 0 (e ι(Rs−τ ))(1) = (e 34 eLs,τ 0 ι∗(Rs−τ ))(1) = 0. (4.34) Applying (4.33) to each term in the product in (4.32) and using (4.34), we have that (4.30) is O(1/N 2). Combining this with (4.31) shows (4.25). Finally, observe that s SO(N, R) shows (4.24). 2 ∆SO(N,R) = 1 2ASO(N,C) s,0 . Setting τ = 0 in (4.27) and restricting to We can now prove Theorem 1.6 for SO(N, R). Proof of Theorem 1.6 for SO(N, R). Let f ∈ C[u, u−1]. Using Theorem 1.5 to rewrite BSO(N,R) s,τ fN and applying the triangle inequality, we have kBSO(N,R) s,τ ) τ s,τ 2 D(1) fN − [Gs,τ f ]NkL2(µSO(N,C) N f ]N − [πs−τ ◦ e = k[e N f ]N − [e ≤ k[e + k[e 2 D(1) τ τ τ τ 2 L0 f ]NkL2(µSO(N,C) s,τ ) 2 L0f ]NkL2(µSO(N,C) s,τ ) 2 L0 f ]N − [πs−τ ◦ e 2 L0 f ]NkL2(µSO(N,C) s,τ τ By Theorem 4.10, (4.36) is O(1/N ), so it remains to show that k[e τ 2 D(1) N f ]N − [e τ 2 L0f ]Nk2 L2(µSO(N,C) s,τ N 2(cid:19) . = O(cid:18) 1 ) To this end, let m = deg f and R(N ) = e and the sesquilinear form B of Lemma 3.21, we have N f − e τ 2 D(1) τ 2 L0f ∈ Cm[u, u−1; v]. Recalling (1.8) k[e τ 2 D(1) N f ]N − [e τ 2 L0f ]Nk2 L2(µSO(N,C) s,τ ) ). (4.35) (4.36) (4.37) 1 ) s,τ 2 AN L2(µSO(N,C) = k[R(N )]Nk2 = e =(cid:16)e s,τ(cid:16)k[R(N )]Nk2 N B(R(N ), R(N ))(cid:17) (1). MN (C)(cid:17) (IN ) eDs,τ Fix norms k · kW2m and k · kCm[u,u−1;v] on W2m and Cm[u, u−1; v], respectively. By Corollary 4.2 applied to the linear functional ϕ(P ) = P (1) on W2m, there exists a constant C = C(m, s, τ ) (not dependent on N ) such that eDs,τ N B(R(N ), R(N ))(cid:17) (1) −(cid:16)e (cid:12)(cid:12)(cid:12)(cid:16)e Next, using the linear functional ψ(P ) =(cid:0)e 0 B(R(N ), R(N ))(cid:17) (1)(cid:12)(cid:12)(cid:12) ≤ eLs,τ τ 2 L0 P )(cid:1) (1) on W2m, we have 2mkB(R(N ), R(N ))kW2m . C N B(R(N ), R(N ))kW2m. Putting (4.37), (4.39), and (4.40) together yields (cid:12)(cid:12)(cid:12)(cid:16)e eLs,τ 0 B(R(N ), R(N ))(cid:17) (1)(cid:12)(cid:12)(cid:12) ≤ kψkW ∗ ) ≤(cid:18)kψkW ∗ 2 L0f ]Nk2 L2(µSO(N,C) s,τ τ 35 k[e τ 2 D(1) N f ]N − [e 2m + C N(cid:19)kB(R(N ), R(N ))kW2m . (4.38) (4.39) (4.40) (4.41) Since B : Cm[u, u−1; v] × Cm[u, u−1; v] → W2m is a sesquilinear form with finite domain and range, there exists a constant C′ = C′(m) (not dependent on N ) such that for all P, Q ∈ Cm[u, u−1; v]. (4.42) kB(P, Q)kW2m ≤ C′kPkCm[u,u−1;v]kQkCm[u,u−1;v] A final application of Corollary 4.2 shows that there exists a constant C′′ depending on L0, L(1) 1 , L(1) 2 , τ , k · kCm[u,u−1;v], and f , but again, not on N , such that C′′ N kR(N )kCm[u,u−1;v] = ke 2 L0 fkCm[u,u−1;v] ≤ . (4.43) N f − e 2 D(1) τ τ Hence kB(R(N ), R(N ))kW2m ≤ kR(N )k2 Cm[u,u−1;v] ≤ (C′′)2 N 2 . (4.44) Combining (4.41) and (4.44) proves (4.37). To prove (1.16), we observe that (BSO(N,R) triangle inequality argument using (4.24), it suffices to show s,τ 2 ∆SO(N,R)fN . By a similar k[e− τ 2 D(1) N f ]N − [e− τ 2 L0 f ]Nk2 (4.45) )−1fNSO(N,R) = e− τ = O(cid:18) 1 N 2(cid:19) . L2(ρSO(N,R) ) s Replacing τ with −τ in the definition of R(N ) and (s, τ ) with (s, 0) from (4.38) onwards, the same argument as above proves (4.45). For uniqueness, we first define seminorms on C[u, u−1; v] by kPk(1) kPk(1) s,N = kPNkL2(ρSO(N,R) ), s,τ,N = kPNkL2(µSO(N,C) s,τ s ), In addition, for β = 1, 2, define the seminorms kPk(2) kPk(2) s,N = kPNkL2(ρU(N ) s,τ,N = kPNkL2(µ s ), GLN (C) s,τ kPk(β) kPk(β) s,N N→∞kPk(β) N→∞kPk(β) s,τ,N s = lim s,τ = lim . ) (4.46) (4.47) (4.48) (4.49) In [8, Lemma 4.7], it was shown that for β = 2, the seminorms (4.48) and (4.49) are norms when restricted to C[u, u−1]. Since kPk(1) s = kPk(2) s (4.50) it follows that the seminorm (4.48) is also a norm on C[u, u−1] for β = 1. An argument completely analogous to the one used in [8, Lemma 4.7] to prove that kPk(2) s,τ is a norm on C[u, u−1] shows that kPk(1) s,τ ∈ C[u, u−1] s,τ is a norm on C[u, u−1]. Hence, if gs,τ , g′ both satisfy kBSO(N,R) s,τ fN − [gs,τ ]Nk2 L2(µSO(N,C) s,τ = O(cid:18) 1 N 2(cid:19) = kBSO(N,R) s,τ 36 ) fN − [g′ s,τ ]Nk2 L2(µSO(N,C) s,τ , ) the triangle inequality implies that kgs,τ − g′ s,τkL2(µSO(N,C) s,τ ) = O(cid:18) 1 N 2(cid:19) . s,τ = 0. Since k · k(1) (4.51) s,τ is a norm on C[u, u−1], it s,τ . A similar argument shows that Hs,τ f is the unique polynomial Taking N → ∞ shows that kgs,τ − g′ follows that gs,τ = g′ satisfying (1.16). Example 4.11. Set P (u; v) = u2 ∈ C[u, u−1]. Then we can show that s,τk(1) BSO(N,R) s,τ τ 2 ∆SO(N,R) PN (A) PN (A) = e 1 (1 − e−τ )IN + N = 1 2 e−τ (−1 + e via the same method used in [8, Example 3.5]. By the intertwining formula e [e N P ]N , this corresponds to the trace polynomial N )A2 − e−τ (1 + e 2 D(1) τ 2τ 2τ N )A tr A, τ 2 ∆SO(N,R) PN = N 2 τ 2 D(1) N P )(u; v) = (e e−τ (1 + e 2τ N )u2 − N 2 e−τ (−1 + e 2τ N )uv1 1 N (1 − e−τ ) + 1 2 = e−τ (u2 − τ uv1) + O(cid:18) 1 N 2(cid:19) . N P at the moments By Theorem 4.10, we compute Gs,τ P by evaluating the traces in e νs−τ and taking the large-N limit of the resulting polynomial. Since ν1(s−τ ) = e−(s−τ )/2, this yields τ 2 D(1) In [8, Example 3.5], it was shown that for the unitary case (with τ = t > 0), Gs,τ P (u) = e−τ (u2 − τ e−(s−τ )/2u). t 2 D(2′ ) (e N P )(u; v) = e−t cosh(t/N )u2 − N e−t sinh(t/N )uv1 = e−τ (u2 − τ uv1) + O(cid:18) 1 N 2(cid:19) , which also results in the same expression for Gs,tP . 5. The Segal-Bargmann transform on Sp(N ) In this section, we prove the Sp(N ) version of Theorems 1.4, 1.5, and 1.6. which comprise our main results. We begin with a slight digression and show that Sp(N ) can also be realized as a subset of N × N quaternion matrices. This allows us to easily identify a par- ticular orthonormal basis of sp(N ) ⊆ M2N (C) which we use to compute a set of magic 2 ASp(N,C) . formulas used in the proofs of our intertwining formulas for e These magic formulas share key similarities with the magic formulas for SO(N, R). Con- sequently, many of the results for the Sp(N ) case can proven using techniques similar to those used in the SO(N, R) case, and so we do not always provide the full details. We conclude by showing that when Sp(N ) is realized as a subset of MN (H), a version of the magic formulas and intertwining formulas for ∆Sp(N ) also hold. 2 ∆Sp(N ) and e s,τ 1 τ 37 5.1. Two realizations of the compact symplectic group The quaternion algebra H is the four-dimensional associative algebra over R spanned by 1, i, j, and k satisfying the relations and i2 = j2 = k2 = −1 ij = k, jk = i, ki = j, ji = −k, kj = −i, ik = −j. We denote the quaternion conjugate of q ∈ H by q∗. That is, if q = a1 + bi + cj + dk, then q∗ = a1 − bi − cj − dk. If A ∈ MN (H), we define the adjoint of A to be the matrix A∗ defined by (A∗)i,j = (Aj,i)∗. For each N ≥ 1, we define fSp(N ) as (5.1) Its Lie algebra is which we endow with the real inner product fSp(N ) = {A ∈ MN (H) : A∗A = IN}. esp(N ) = {X ∈ MN (H) : X ∗ = −X}, (5.2) (5.3) , Let hX, Y i esp(N ) = 2N Re Tr(X ∗Y ) CN =(cid:26) 1 √4N for all X, Y ∈ esp(N ). Ea,a(cid:27)1≤a≤N An orthonormal basis for esp(N ) with respect to this inner product is (Ea,b + Eb,a)(cid:27)1≤a<b≤N ∪(cid:26) 1 √2N (Ea,b − Eb,a)(cid:27)1≤a<b≤N ∪ iCN ∪ jCN ∪ kCN . β esp(N ) =(cid:26) 1 √4N where Ea,b ⊆ MN (H) is the N ×N matrix with a 1 in the (a, b) entry and zeros elsewhere. (5.4) From (5.1), we see that fSp(N ) is the unitary group over the quaternions. We now explicitly construct a (real) Lie group isomorphism between fSp(N ) and Sp(N ). While fSp(N ) is, in some sense, the more natural realization of the compact symplectic group, for our purposes it is easier to work with the definition given in (2.5). In particular, the complexification of the compact symplectic group is more readily understood when Sp(N ) is realized as a subset of M2N (C) rather than of MN (H). The material presented below is standard (cf. [18]); we include it here for convenience. First, observe that we can realize H as a subalgebra of M2(C). We do so by letting ψ : H → M2(C) denote the injective homomorphism H ∋ q = a1 + bi + cj + dk 7→ (cid:20) a + di −b − ci a − di (cid:21) ∈ M2(C). b − ci 38 (5.5) The map ψ preserves conjugation: ψ(q∗) = ψ(q)∗ for all q ∈ H, where we take the quaternion conjugate on the left and the complex conjugate transpose on the right. Moreover, note that we can write a matrix of the form (5.5) as (cid:20) α −β α (cid:21) β (5.6) for some α, β ∈ C, and any matrix of this form corresponds to a q ∈ H. Now define a map Φ : MN (H) → M2N (C) by Φ([qi,j ]1≤i,j≤N ) = [ψ(qi,j)]1≤i,j≤N . By the properties of block multiplication, Φ is an algebra homomorphism. Since ψ is one-to-one, it follows that Φ is also one-to-one. Moreover, since ψ preserves conjugation, if A ∈ MN (H), then Φ(A∗) = Φ(A)∗, where again we take the quaternion so does Φ: conjugate transpose on the left and the complex conjugate transpose on the right. Consequently, if A ∈fSp(N ) ⊆ MN (H), so that A∗A = IN , then Φ(A)∗Φ(A) = Φ(A∗)Φ(A) = Φ(A∗A) = I2N . Hence Φ(fSp(N )) = {Z ∈ M2N (C) : Z ∗Z = I2N and ∃ A ∈ MN (H) such that Φ(A) = Z}. We would like to show that Φ(fSp(N )) = Sp(N ), which would imply that Sp(N ) ∼=fSp(N ) as groups. Recall that since A ∈ Sp(N ) is unitary and Ω−1 = −Ω, we can rewrite the definition of Sp(N ) as follows: (5.7) Sp(N ) = {Z ∈ M2N (C) : Z ∗Z = I2N ,−ΩZΩ = Z}. Comparing (5.7) and 5.8, we see that it suffices to show that Z = Φ(A) for some A ∈ MN (H) ⇔ −ΩZΩ = Z. (5.8) (5.9) We need the following lemma. Lemma 5.1. Let B ∈ M2(C). Then −Ω0BΩ0 = B if and only if B =(cid:20) α −β α (cid:21) for β some α, β ∈ C. Proof. Direct computation. Consequently, for Z ∈ M2N (C), [−ΩZΩ]i,j = X1≤k,ℓ≤N [Ω]i,k[Z]k,ℓ[Ω]ℓ,j = −Ω0[Z]i,jΩ0, (5.10) where [Z]i,j denotes the (i, j)th 2 × 2 block of Z. By Lemma 5.1, we see that for each 1 ≤ i, j ≤ N , there exists qi,j ∈ H such that ψ(qi,j) = [Z]i,j precisely when −ΩZΩ = Z. 39 Theorem 5.2. The map ΦfSp(N ) :fSp(N ) → Sp(N ) is a Lie group isomorphism. Proof. The preceding discussion shows that ΦfSp(N ) is a group isomorphism. It remains to be shown that ΦfSp(N ) and Φ−1 are smooth, but this is clear from the definition of Φ. fSp(N ) The inner products (2.10) and (5.3) for sp(N ) and esp(N ), resp., are related by Φ. For X ∈ MN (H), note that Re Tr X = Re Tr Φ(X) = Tr Φ(X). (5.11) 1 2 1 2 Hence for X, Y ∈ esp(N ), hX, Y i esp(N ) = 2N Re Tr(XY ∗) = N Tr(Φ(XY ∗)) = N Tr(Φ(X)Φ(Y )∗) = hΦ(X), Φ(Y )isp(N ) . (5.12) In particular, if β esp(N ) is an orthonormal basis for esp(N ) ⊆ MN (H) with respect to (5.2), then Φ(β esp(N )) is an orthonormal basis for sp(N ) ⊆ M2N (C) with respect to (2.10). 5.2. Magic formulas and derivative formulas Proposition 5.3 (Magic formulas for Sp(N )). Let βsp(N ) be any orthonormal basis for sp(N ) with respect to the inner product (2.10). For any A, B ∈ M2N (C), X 2 = − 2N + 1 2N I2N = −I2N − 1 2N I2N XAX = − XX∈βsp(N ) XX∈βsp(N ) XX∈βsp(N )etr(XA)X = XX∈βsp(N )etr(XA)etr(XB) = 1 2N ΩA⊺Ω−1 −etr(A)I2N 1 4N 2 (ΩA⊺Ω−1 − A) 1 4N 2 (etr(ΩA⊺ΩB) −etr(AB)) (5.13) (5.14) (5.15) (5.16) Proof. As with the magic formulas for SO(N, R), the quantities on the left hand side of (5.13), (5.14), (5.15), and (5.16) are independent of choice of orthonormal basis. Thus we may compute with respect to the orthonormal basis βsp(N ) := Φ(β esp(N )), where β esp(N ) is the orthonormal basis described in (5.4). We will use the following notation: for A ∈ M2N (C), let [A]i,j denote the (i, j)th 2 × 2 submatrix, and let Ai,j If Ea,b ⊆ MN (H) is an elementary matrix and γ ∈ {1, i, j, k} ⊆ H, k,ℓ denote the (k, ℓ)th entry of [A]i,j. F γ a,b := Φ(γEa,b) ∈ M2N (C). 40 (5.17) Observe that [F γ one of a,b]i,j contains all zeros unless (i, j) = (a, b), in which case it is equal to 0 ψ(1) =(cid:20) 1 0 −i −i depending on whether γ equals 1, i, j, or k, respectively. 1 (cid:21) , ψ(i) =(cid:20) 0 −1 0 (cid:21) , ψ(j) =(cid:20) 0 1 0 (cid:21) , ψ(k) =(cid:20) i 0 −i (cid:21) , 0 Equation (5.13) follows from (5.14). For (5.14), we have XX∈βsp(N ) XAX = 1 4N X1≤a,b≤N Xγ∈{1,i,j,k} F 1 4N X1≤a,b≤N − 1 F γ a,bAF γ a,b a,bAF 1 b,a − Xγ∈{i,j,k} F γ a,bAF γ b,a For 1 ≤ a, b, c, d, i, j ≤ N , we have [F γ [F γ a,bAF γ c,d]i,j = X1≤k,ℓ≤N =(02 ψ(γ)[A]b,cψ(γ) i 6= a or j 6= d i = a and j = d a,b]i,k[A]k,ℓ[F γ c,d]ℓ,j = [F γ a,b]i,b[A]b,c[F γ c,d]c,j A straightforward computation shows that Hence for the first term on the right hand side of (5.18), 1,2 Ab,c 1,1 Ab,c 2,1 Ab,c 2,2 −Ab,c 1,2 −Ab,c ψ(1)[A]b,cψ(1) =" Ab,c ψ(j)[A]b,cψ(j) =" −Ab,c −Ab,c  1 4N X1≤a,b≤N Xγ∈{1,i,j,k} 2,1 2,2 # , ψ(i)[A]b,cψ(i) =" −Ab,c 1,1 # , ψ(k)[A]b,cψ(k)" −Ab,c 4N  Xγ∈{1,i,j,k} a,b 2N " −Aj,i 1,1 # Ab,c 2,2 2,1 1,2 −Ab,c Ab,c 2,2 # Ab,c 1,1 1,2 2,1 −Ab,c Ab,c i,j 1,1 # . Aj,i 1,2 2,2 2,1 −Aj,i Aj,i F γ a,bAF γ F γ i,jAF γ = = i,j i,j 1 1 (5.18) (5.19) (5.20) (5.21) (5.22) (5.23) For the second term on the right hand side of (5.18), the (i, j)th block is 02 for i 6= j. 41 For i = j, i,j i,j F γ i,bAF γ F γ a,bAF γ b,a b,i ψ(γ)[A]b,bψ(γ) 2,2 # 0 1,1 + Ab,b 1 = = 1 4N i,bAF 1 a,bAF 1  1 F 1 4N X1≤a,b≤N b,a − Xγ∈{i,j,k} 4N  F 1 NXb=1 b,i − Xγ∈{i,j,k} [A]b,b − Xγ∈{i,j,k} NXb=1 NXb=1" Ab,b  " −Aj,i " −Aj,i XAX 2,2 0 Ab,b 1,1 + Ab,b Tr(A)I2. 1 2N 1 2N 1 2N = = = i,j 1,1 # Aj,i 1,2 2,2 2,1 −Aj,i Aj,i 1,1 # − Tr(A)I2 Aj,i 2,2 1,2 2,1 −Aj,i Aj,i i 6= j i = j. Putting this together, we have  XX∈βsp(N ) On the other hand, we can compute (cid:20)− 1 2N ΩA⊺Ω−1 −etr(A)I2N(cid:21)i,j 1 2N (Ω0([A]j,i)⊺Ω−1 0 Ω0([A]j,i)⊺Ω−1 0 − Tr(A)I2 = − Multiplying out (5.29) and comparing with (5.28) proves (5.14). Next, for (5.15), we again use the basis βsp(N ) to compute (5.24) (5.25) (5.26) (5.27) (5.28) i 6= j, i = j. (5.29) (5.30) b,a . (5.31) XX∈βsp(N )etr(XA)X = 1 1 − 4N X1≤a,b≤N Xγ∈{1,i,j,k}etr(F γ etr(F 1 4N X1≤a,b≤N etr(F i etr(F k 1,2 + Ab,a 2,1), 1,1 + Ab,a (Ab,a (Ab,a 2,2), i 2N 42 a,bA) = 1 2N a,bA) = − etr(F 1 etr(F j For 1 ≤ a, b ≤ N , a simple calculation shows that a,bA)F γ a,b a,bA)F 1 b,a − Xγ∈{i,j,k}etr(F γ a,bA)F γ a,bA) = a,bA) = 1 2N (Ab,a 1,2 − Ab,a 2,1) 1,1 − Ab,a (Ab,a 2,2), i 2N (5.32) (5.33) The (i, j)th block of the right hand side of (5.30) is thus  1 4N X1≤a,b≤N Xγ∈{1,i,j,k}etr(F γ a,bA)F γ a,b i,j = = 1 4N etr(F 1 4N 2" Aj,i −Aj,i 1 i,j A)F 1 i,j + Xγ∈{i,j,k}etr(F γ 1,1 # . 2,2 −Aj,i 1,2 Aj,i 2,1 i,jA)F γ Similarly, the (i, j)th block of (5.31) is 1 =  1 4N X1≤a,b≤N etr(F 1 4N etr(F 1 4N 2" Ai,j  XX∈βsp(N )etr(XA)X Ai,j = = i,j 1 a,bA)F 1 b,a − Xγ∈{i,j,k}etr(F γ i,j − Xγ∈{i,j,k}etr(F γ 2,2 # . 1,2 j,iA)F 1 1,1 Ai,j 2,1 Ai,j a,bA)F γ j,iA)F γ i,j b,a i,j i,j 1 4N 2" Aj,i −Aj,i 2,2 − Ai,j 2,1 − Ai,j 2,2 # 1,2 − Ai,j 1,1 −Aj,i 1,1 − Ai,j Aj,i 1,2 i,j Combining our computations, i,j i,j (5.34) (5.35) (5.36) (5.37) Again, we can expand (cid:20) 1 4N 2 (ΩA⊺Ω−1 − A)(cid:21)i,j B ∈ M2N (C) and taking etr proves (5.16). = 1 4N 2 [−Ω0([A]j,i)⊺Ω−1 0 − A]i,j to see that this is equivalent to (5.36). This proves (5.15); multiplying on the right by While Proposition 5.3 holds for all A, B ∈ M2N (C), we now restrict to the case in which A and B are in Sp(N ) or Sp(N, C). Note that for A ∈ Sp(N, C), ΩA⊺Ω−1 = A−1. The magic formulas for Sp(N ) allow us to compute the following derivative formulas, as in the SO(N, R) case. Proposition 5.4 (Derivative formulas for Sp(N )). The following identities hold on 43 Sp(N ) and Sp(N, C): 1 1 N (m − j)etr(Aj )Am−j ∆Sp(N )Am = 1m≥2− m−1Xj=1 m−1Xj=1 (m − j)Am−2j − 2 +(cid:18)−1 − 2N(cid:19) mAm, m ≥ 0 (−m + j)etr(Aj)Am−j ∆Sp(N )Am = 1m≤−2− −1Xj=m+1 −1Xj=m+1 (−m + j)Am−2j − 2 −(cid:18)−1 − 2N(cid:19) mAm, m < 0 XX∈βsp(N ) eXetr(Am) · eXAp = mp 4N 2 (Ap−m − Ap+m), m, p ∈ Z. 1 N 1 (5.39) (5.40) (5.38) Proof. For m ≥ 0, we use (3.15) and magic formulas (5.83) and (5.84) to sum over all X ∈ βSp(N ): ∆Sp(N )Am = 21m≥2 XX∈βso(N,R) Xk,j6=0 k+j+ℓ=m Aj X 2Am−j (5.41) mXj=1 AkXAjXAℓ + XX∈βso(N,R) Ak(Aj)∗Aℓ −etr(Aj )Ak+ℓ(cid:21) − (cid:20)− 1 2N m(2N + 1) 2N k+j+ℓ=m = 21m≥2 Xk,j6=0 = 1m≥2− +(cid:18)−1 − 1 N m−1Xj=1 (m − j)Am−2j − 2 2N(cid:19) mAm. 1 (m − j)etr(Aj )Am−j m−1Xj=1 Am (5.42) (5.43) For m < 0, we use (3.10) and magic formulas (5.83) and (5.84) to sum over all X ∈ βSp(N ): ∆Sp(N )Am = 21m≥2 XX∈βso(N,R) Xk,ℓ<0,j≤0 j+k+ℓ=m AjX 2Am−j 0Xj=m+1 AkXAjXAℓ + XX∈βso(N,R) Ak(Aj )∗Aℓ −etr(Aj)Ak+ℓ(cid:21) + (−m + j)etr(Aj)Am−j −1Xj=m+1 m(2N + 1) Am 2N (−m + j)Am−2j − 2 j+k+ℓ=m 1 2N = 21m≥2 Xk,ℓ<0,j≤0 = 1m≤−2− −(cid:18)−1 − (cid:20)− −1Xj=m+1 2N(cid:19) mAm. 1 N 1 44 Finally, comparing the magic formulas (3.7) and (5.15), we see that the proof of (5.40) is identical to that of (3.14). 5.3. Intertwining formulas for ∆Sp(N ) The derivative formulas allow us to prove the intertwining formula for ∆Sp(N ). Proof of Theorem 1.4 for Sp(N ). We proceed as in the proof of Theorem 1.4 for the SO(N, R) case. Again, it suffices to show that ∆Sp(N )PN = [D(4) N P ]N holds for P (u; v) = umq(v), where m ∈ Z \ {0} and q ∈ C[v]. Then PN = Wm · q(V), and by the product rule, ∆Sp(N )PN = (∆Sp(N )Wm) · q(V) + 2 XX∈βsp(N ) eXWm · eXq(V) + Wm · (∆Sp(N )q(V)) (5.44) For the first term in (5.44), we consider the cases m ≥ 0 and m < 0 separately. For m ≥ 0, we apply (5.38) to get (∆Sp(N )Wm) · q(V) = − m(2N + 1) 2N 2N + 1 Wm · q(V) 1 N + 1m≥2(cid:20) − m−1Xk=1 2N (cid:20)u ∂uR+P(cid:21)N 2N "−2 ∞Xk=1 − 2" ∞Xk=1 vkuR+ ∂ 2N (cid:20)u ∂uR+P(cid:21)N − 2N + 1 + 1 ∂ ∂ = − = − (m − k)Wm−2k − 2 (m − k)VkWm−k(cid:21)q(V) m−1Xk=1 ∂uMu−kR+! P#N u−k+1R+ ∂ ∂uMu−kR+! P#N 1 N [Z + 2 P ]N − 2[Y + 1 P ]N . (5.45) Similarly, for m < 0, we apply (5.39) to get 2N + 1 (∆Sp(N )Wm) · q(V) = 2N (cid:20)u 2 annihilates C[u−1] while Y − ∂uR−P(cid:21)N 1 ,Z − 1 ,Z + ∂ Since Y + m ∈ Z, + 1 N [Z − 2 P ]N + 2[Y − 1 P ]N . 2 annihilates C[u], we have that for all (∆Sp(N )Wm) · q(V) = − 2N + 1 2N [N1P ]N − 1 N [Z2P ]N − 2[Y1P ]N . (5.46) 45 For the middle term in (5.44), we note the similarity between the derivative formulas (3.14) and (5.40) for the SO(N, R) and Sp(N ) cases, resp. Hence, as in (3.38), 1 4N 2 [K1P ]N . XX∈βsp(N ) eXWm · eXq(V) = q(cid:19) (V) · eX 2Vk + Xj,k≥1(cid:18) ∂2 ∂vj∂vk q(cid:19) (V) · eXVj · eXVk., (5.47) (5.48) For the last term in (5.44), we have, for each X ∈ βsp(N ), eX 2q(V) = Xk≥1(cid:18) ∂ ∂vk as with (3.39). In summing the first term in (5.48) over X ∈ βsp(N ), we break up the sum for positive and negative k. Using (5.38) and (5.39), we have XX∈βsp(N ) ∞Xk=1(cid:18) ∂ ∂vk q(cid:19) (V) · eX 2Vk = − − 1 N 2N + 1 2N ∂vk kVk(cid:18) ∂ q(cid:19) (V) ∞Xk=1 (k − ℓ)Vk−2ℓ(cid:18) ∂ k−1Xℓ=1 ∞Xk=2 (k − ℓ)VℓVk−ℓ(cid:18) ∂ k−1Xℓ=1 ∞Xk=2 ∂vk ∂vk − 2 q(cid:19) (V) q(cid:19) (V) (5.49) and XX∈βsp(N ) −1Xk=−∞(cid:18) ∂ ∂vk q(cid:19) (V) · eX 2Vk = 2N + 1 2N ∂vk kVk(cid:18) ∂ q(cid:19) (V) −1Xk=−∞ (k − ℓ)Vk−2ℓ(cid:18) ∂ −1Xℓ=k+1 −2Xk=−∞ (−k + ℓ)VℓV−k−ℓ(cid:18) ∂ −2Xk=−∞ −1Xℓ=k+1 ∂vk 1 N − − 2 q(cid:19) (V) q(cid:19) (V) ∂vk (5.50) Summing the second term in (5.48) over X ∈ βsp(N ) and using magic formula (5.40) yields XX∈βsp(N ) Xj,k≥1(cid:18) ∂2 ∂vj∂vk q(cid:19) (V) · eXVj · eXVk = 1 4N 2 Xj,k≥1 jk(Vk−j − Vk+j )(cid:18) ∂2 ∂vj∂vk q(cid:19) (V). (5.51) 46 Adding (5.49), (5.50), and (5.51) together and multiplying by Wm, we get Wm · (∆Sp(N )q(V)) =(cid:20)(cid:18)− 2N + 1 2N N0 − 1 N Z1 − 2Y2 + 1 4N 2K2(cid:19) P(cid:21)N . (5.52) Combining (5.46), (5.47), and (5.52) proves (1.11) for Sp(N ). Equation (1.12) now follows. τ 2 D(4) N P ]N . Corollary 3.11 shows that [e Proof of Theorem 1.5 for Sp(N ). The proof for Sp(N ) is entirely analogous to the proof 2 ∆Sp(N )PN = for SO(N, R). Using the intertwining formula for e [e N P ]N is a well-defined trace polynomial on Sp(N ), so its analytic continuation to Sp(N, C) is given by the same trace polyno- mial function. Hence [e N P ]N , viewed as a holomorphic function on Sp(N, C), is the 2 ∆Sp(N )PN , and so is equal to BSp(N ) analytic continuation of e 2 ∆Sp(N ) (1.12), we have e s,τ PN . 2 D(4) 2 D(4) τ τ τ τ t 5.4. Intertwining formulas for ASp(N,C) Theorem 5.5. Fix s ∈ R and τ = t + iθ ∈ C. There are collections {T s,τ {U s,τ monomials of trace degree ε such that : ε ∈ E } such that for each ε ∈ E , T s,τ and U s,τ s,τ ε ε ε ε : ε ∈ E }, are certain finite sums of ASp(N,C) s,τ Vε = [T s,τ ε ]N + 1 N [U s,τ ε ]N , (5.53) Moreover, let {Ss,τ 3.19. Then for ε, δ ∈ E , ε,δ : ε, δ ∈ E } ⊆ W be the collection of word polynomials from Theorem dsp(N )Xℓ=1 (cid:20)(cid:18)s − t 2(cid:19) (eXℓVε)(eXℓVδ) + t 2 (eYℓVε)(eYℓVδ) − θ(eXℓVε)(eYℓVδ)(cid:21) = 1 N 2 [Ss,τ ε,δ ]N , (5.54) dsp(N ) ℓ=1 where βsp(N ) = {Xℓ} For the proof, we will employ the conventions, mutandis mutatis, as those used in the proof of Theorem 3.19 (see the remarks following the theorem statement). and Yℓ = iXℓ. Proof. Fix a word ε = (ε1, . . . , εm) ∈ E . Then for each X ∈ β± and A ∈ Sp(N ), we apply the product rule twice to get (eX 2Vε)(A) = mXj=1 tr(Aε1 ··· (AX 2)εj ··· Aεm ) + 2 X1≤j<k≤m tr(Aε1 ··· (AX)εj ··· (AX)εk ··· Aεm ). (5.55) (5.56) 47 Applying magic formula (5.13) to each term in (5.56), we have XX∈β± tr(Aε1 ··· (AX 2)εj ··· Aεm ) = ± 2N + 1 2N tr(Aε1 ··· Aεj ··· Aεm ) = ± 2N + 1 2N Vε(A). (5.57) Summing over 1 ≤ j ≤ m now gives XX∈β± mXj=1 tr(Aε1 ··· (AX 2)εj ··· Aεm ) = 2N + 1 2N n±(ε)Vε(A), (5.58) where n±(ε) ∈ Z and n±(ε) ≤ ε (where the integers n±(ε) are not necessarily the same as in the SO(N, R) case). For (5.56), we can express each term in the sum as tr(Aε1 ··· (AX)εj ··· (AX)εk ··· Aεm ) = ± tr(Aε0 j,kε1 j,k, and ε2 j,k are substrings of ε such that ε0 j,k, ε1 so ε0 X ∈ β± by magic formula (5.14), we have tr(Aε1 ··· (AX)εj ··· (AX)εk ··· Aεm ) = ±(cid:20) − XX∈β± j,k XAε1 j,k XAε2 j,k ), (5.59) j,kε2 j,k = ε. Summing (5.59) over 1 2N tr(Aε0 j,k (Aε1 j,k )−1Aε2 j,k ) j,k Aε2 j,k ) tr(Aε1 j,k )(cid:21) − tr(Aε0 Vε3 j,k 1 2N = ± (A) ± Vε0 j,kε2 j,k (A)Vε1 j,k (A), (5.60) (A) = tr(Aε0 j,k (Aε1 j,k )−1Aε2 j,k ). Thus j,k is the word of length ε such that Vε3 where ε3 summing (5.56) over X ∈ β± gives j,k XX∈β± X1≤j<k≤m tr(Aε1 ··· (AX)εj ··· (AX)εk ··· Aεm ) 2N X1≤j<k≤m (A) + X1≤j<k≤m ±Vε3 = 1 j,k ±Vε1 j,k (A)Vε0 j,kε2 j,k (A). (5.61) We define word polynomials corresponding to (5.61) by T ± ε,0 = X1≤j<k≤m ε,1 = X1≤j<k≤m T ± ±vε3 j,k ±vε0 j,kε2 j,k vε1 j,k . Now if X ∈ β+ and Y = iX ∈ β−, then ± tr(Aε1 ··· (AX 2)εj ··· Aεm ) (eXeY Vε)(A) = i 2 mXj=1 + i X1≤j<k≤m ± tr(Aε1 ··· (AX)εj ··· (AX)εk ··· Aεm ). 48 (5.62) (5.63) (5.64) (5.65) Summing (5.64) over X ∈ β+ and applying (5.83) gives XX∈β+ mXj=1 ± tr(Aε1 ··· (AX 2)εj ··· Aεm ) = 2N + 1 2N η(ε)Vε(A) (5.66) where η(ε) ∈ Z and η(ε) ≤ ε. In addition, summing (5.65) over X ∈ β+, we have XX∈β+ X1≤j<k≤m ± tr(Aε1 ··· (AX)εj ··· (AX)εk ··· Aεm) (A) − X1≤j<k≤m 2N X1≤j<k≤m = − Vε3 1 j,k ±Vε0 j,kε2 j,k (A)Vε1 j,k (A). (5.67) We define corresponding word polynomials Tε,2 = i X1≤j<k≤m Tε,3 = i X1≤j<k≤m Putting this together, we define ±vε3 j,k ±vε0 j,kε2 j,k vε1 j,k . T s,τ ε =(cid:18)s − ε =(cid:18)s − U s,τ t 2(cid:19)(cid:0)n+(ε)vε + 2T + ε,1(cid:1) + 2(cid:19)(cid:18)− ε,0(cid:19) + vε + T + n+(ε) 2 t t 2(cid:0)n−(ε)vε + 2T − 2(cid:18)− n−(ε) 2 t ε,1(cid:1) − θ (iη(ε) + 2Tε,3) , η(ε) ε,0(cid:19) − θ(cid:18)−i 2 vε + T − (5.68) (5.69) (5.70) + Tε,2(cid:19) . (5.71) By construction, T s,τ ε and U s,τ ε are of the required form and satisfy (5.53). The proof of (5.54) is essentially identical to that of Theorem 3.19(2); this is due to the similarity of magic formulas (3.7) and (5.85) for SO(N, R) and Sp(N ). ε ε s,τ Theorem 5.6 (Intertwining formula for ASp(N,C) {T s,τ be as in Theorem 3.20. Define first and second order differential operators : ε ∈ E }, {U s,τ ). Fix s ∈ R and τ = t + iθ ∈ C. Let : ε ∈ E }, and nSs,τ ε,δ : ε, δ ∈ Eo be as in Theorem 5.5, and eLs,τ 2Xε∈E 2Xε∈E Cs,τ 1 = Cs,τ 0 = ∂ ∂vε ∂ ∂vε (5.72) (5.73) T s,τ ε U s,τ 1 1 2 ε Then for all N ∈ N and P ∈ W , 1 2ASp(N,C) s,τ PN =(cid:20)(cid:18)Cs,τ 0 + 1 N Cs,τ 1 + 49 2 (cid:19) P(cid:21)N 1 N 2 eLs,τ . (5.74) Proof. For each X ∈ sp(N ), we apply the chain rule to get eX 2PN =Xε∈E eX(cid:20)(cid:18) ∂P =Xε∈E(cid:18) ∂P PN =Xε∈E(cid:18) ∂P ∂vε(cid:19) (V) · (eXVε)(cid:21) ∂vε(cid:19) (V) · (eX 2Vε) + Xε,δ∈E(cid:18) ∂2P ∂vε(cid:19) (V) · ASp(N,C) ∂vε∂vδ(cid:19) (V) + Xε,δ∈E(cid:18) ∂2P dSO(N,R)Xℓ=1 (cid:20)(cid:18)s − Vε s,τ Hence ASp(N,C) s,τ The result now follows from Theorem 5.5. + ∂vε∂vδ(cid:19) (V) · (eXVε)(eXVδ). t 2(cid:19) (eXℓVε)(eXℓVδ) (eYℓVε)(eYℓVδ) − θ(eXℓVε)(eYℓVδ)(cid:21). t 2 5.5. Limit theorems for the Segal-Bargmann transform on Sp(N ) In this section, we outline the proof of Theorem 1.6 for Sp(N ). Since the techniques used are similar to those used in Section 4 to prove the SO(N, R) case, we do not provide the full details. The following two lemmas are the analogues of Lemmas 4.5, 4.6, and 4.7 for Sp(N ). The proofs are very similar, and so are not included: the key concentration result is again Lemma 4.2, and the only change required is to replace intertwining formulas (1.11) and (3.82) for SO(N, R) with intertwining formulas (1.11) and (5.74) for Sp(N ) wherever applicable. Lemma 5.7 (L´evy, [13]). For s > 0 and k ∈ Z, tr(Ak) ρSp(N ) (A) dA = νk (s) . (5.75) lim s Lemma 5.8. Let s > 0 and τ = t + iθ ∈ D(s, s). For any Q ∈ C[v], N→∞ZSp(N ) (cid:16)eCs,τ 0 ι(Q)(cid:17) (1) = πs−τ Q. As in the proof of Theorem 1.6 for SO(N, R), we first require the following analogue of Theorem 4.10. Theorem 5.9. Let s > 0 and τ = t + iθ ∈ D(s, s). For any P ∈ C[u, u−1; v], L2(ρSp(N ) s ) kPN − [πsP ]Nk2 kPN − [πs−τ P ]Nk2 L2(µSp(N,C) s,τ 50 and N 2(cid:19) , = O(cid:18) 1 N 2(cid:19) . = O(cid:18) 1 ) (5.76) (5.77) (5.78) Proof. The proof is entirely analogous to the proof of Theorem 4.10. Again, we replace with intertwining formula (5.74) for ASp(N,C) intertwining formula (3.82) for ASO(N,C) wherever necessary. To show the Sp(N ) version of (4.34) (where eLs,τ is replaced with Cs,τ 0 ), we use Lemma 5.8 in place of Lemma 4.7. The remainder of the proof is the same. s,τ s,τ 0 Proof of Theorem 1.5 for Sp(N ). The proof of the limit results (1.15) and (1.16) is very similar to the SO(N, R) case (see p. 35), where we now use the polynomial S(N ) = 2 L0 f ∈ Cm[u, u−1; v] in place of R(N ) and intertwining formula (5.74) in e place of (3.82) in (4.38). N f − e 2 D(4) τ τ To show uniqueness, we define seminorms on C[u, u−1; v] by kPk(4) kPk(4) s,N = kPNkL2(ρSp(N ) N→∞kPk(4) s = lim s,N , s ), kPk(4) kPk(4) s,τ,N = kPNkL2(µSp(N,C) s,τ = lim N→∞kPk(4) s,τ,N . s,τ Noting that kPk(4) s = kPk(2) s , ), (5.79) (5.80) (5.81) the remainder of the proof proceeds as in the proof of uniqueness in Theorem 1.5 for SO(N, R). We conclude this section by showing a version of the magic formulas and intertwining 5.6. Extending the magic formulas and intertwining formulas to fSp(N ) results for Sp(N ) hold for fSp(N ) ⊆ MN (H). We have the following key relation: (5.82) which can be seen by the definition of ψ (see (5.5)). etr(Φ(A)) = Re tr(A), A ∈ MN (H), Proposition 5.10 (Magic formulas forfSp(N )). Let β esp(N ) be any orthonormal basis for esp(N ) with respect to the inner product (5.3). For any A, B ∈ MN (H), 2N + 1 (5.83) X 2 = − 2N IN = −IN − 1 2N IN XX∈β fsp(N ) XX∈β fsp(N ) XX∈β fsp(N ) XX∈β fsp(N ) Re tr(XA)X = Re tr(XA) Re tr(XB) = 51 XAX = − 1 2N A∗ − Re tr(A)IN 1 4N 2 (A∗ − A) 1 4N 2 (Re tr(A∗B) − Re tr(AB)) (5.84) (5.85) (5.86) Proof. We begin with the observation that for A ∈ MN (H), ΩΦ(A)⊺Ω−1 = Φ(A∗). (5.87) To see why this is the case, note that Lemma 5.1 directly implies that for q ∈ H, Ω0ψ(q)⊺Ω−1 0 = ψ(q∗). Hence if A = [qi,j], [ΩΦ(A)⊺Ω−1]i,j = X1≤k,ℓ≤N [Ω]i,k[Φ(A)⊺]k,ℓ[Ω]ℓ,j = [Ω]i,i[Φ(A)⊺]i,j [Ω]j,j = Ω0ψ(qj,i)⊺Ω−1 0 = ψ(q∗ j,i) = [Φ(A∗)]i,j , which shows (5.87). Recall that if β esp(N ) is an orthonormal basis for esp(N ), Φ(β esp(N )) is an orthonormal basis for sp(N ). Let A ∈ MN (H). Using (5.14) and (5.82), we have Φ(X)Φ(A)Φ(X) XX∈β fsp(N ) Φ(XAX) = XX∈β fsp(N ) 1 2N 1 2N = − = − = Φ(cid:18)− ΩΦ(A)⊺Ω−1 −etr(Φ(A))I2N Φ(A∗) − Re tr(A)Φ(IN ) A∗ − Re tr(A)IN(cid:19) . 1 2N Since Φ is injective, (5.84) follows. The proof of (5.85) is similar, and (5.83) and (5.86) follow from (5.84) and (5.85), resp. defined by AnfSp(N ) version of Theorem 1.4 also holds for an appropriately defined trace polynomial functional calculus: for P ∈ C[u, u−1; v], we let ePN : fSp(N ) → MN (H) be the function Proposition 5.11. Let P ∈ C[u, u−1; v] and A ∈fSp(N ) ⊆ MN (H). Then ePN (A) := P (u; v)u=A,vk=Re tr(Ak),k6=0. (5.89) (5.88) Proof. It suffices to prove the result for P (u; v) = umvn Using (5.82), we have k , with m, n, k ∈ Z and k 6= 0. PN (Φ(A)) = Φ(ePN (A)). PN (Φ(A)) = Φ(A)metr(Φ(A)k)n = Φ(Am)(Re tr(Ak))n = Φ(Am(Re tr(Ak))n) = Φ(ePN (A)). Using this proposition, we see that the intertwining formula for Sp(N ) ⊆ MN (H) is a direct consequence of the intertwining formula (1.11) for Sp(N ) ⊆ M2N (C). 52 Theorem 5.12 (Intertwining formulas). For any P ∈ C[u, u−1; v], Moreover, for all τ ∈ C, e . N PiN ∆fSp(N )ePN =hD(4) 2 ∆ fSp(N )ePN = [e 2 D(4) τ τ N P ]N . (5.90) (5.91) (5.92) Remark 5.13. We have seen that the magic formulas (5.10) for fSp(N ) can be derived from the magic formulas (5.3) for Sp(N ). However, the converse is not true. Suppose we only know that the magic formula (5.84) holds. Applying Φ to both sides yields XX∈β fsp(N ) Φ(X)Φ(A)Φ(X) = − 1 2N Φ(A)∗ −etr(Φ(A))I2N . We cannot replace Φ(A) with an arbitrary B ∈ M2N (C) in the formula above. For example, consider Then we can compute that XY ∈βsp(1) Y BY = 1 2(cid:20) −3 1 0 B =(cid:20) 1 1 −3 (cid:21) 6= 1 (cid:21) ∈ Sp(1, C). 2(cid:20) −3 −1 1 0 0 −3 (cid:21) = − 1 2 B∗ −etr(B)I2. This is another reason why it is more convenient to work with Sp(N ) rather thanfSp(N ). 6. The Segal-Bargmann transform on SU(N ) In this final section, we analyze the Segal-Bargmann transform on the special unitary group SU(N ). This case uses many of the techniques from the U(N ) case, studied in [8]. Consequently, we outline the main ideas required for the proofs of these results and do not provide the full details. We first prove the intertwining formulas for SU(N ). As in the SO(N, R) and Sp(N ) cases, the basis of these results is a set of magic formulas. We recall the following set of magic formulas for U(N ), proven in [8]. Proposition 6.1 ([8, Proposition 3.1]). Let βu(N ) be any orthonormal basis for u(N ) with respect to the inner product defined by hX, Y iu(N ) = N 2 tr(XY ∗). For any A, B ∈ 53 MN (C), X 2 = −IN XAX = − tr(A)IN XX∈βu(N ) XX∈βu(N ) XX∈βu(N ) tr(XA)X = − tr(XA) tr(XB) = − 1 N 2 A 1 N 2 tr(AB) XX∈βu(N ) The magic formulas for SU(N ) follow easily from the magic formulas for U(N ). Proposition 6.2 (Magic formulas for SU(N )). Let βsu(N ) be any orthonormal basis for su(N ) with respect to the inner product (2.9). For any A, B ∈ MN (C), (6.1) (6.2) (6.3) (6.4) (6.5) (6.6) (6.7) (6.8) X 2 =(cid:18)−1 + 1 N 2(cid:19) IN XAX = − tr(A)IN + 1 N 2 A XX∈βsu(N ) XX∈βsu(N ) XX∈βsu(N ) tr(XA)X = − tr(XA) tr(XB) = − 1 N 2 A + 1 N 2 tr(A)IN 1 N 2 tr(AB) + 1 N 2 tr(A) tr(B) XX∈βsu(N ) Proof. The expressions on the left side of (6.5), (6.6), (6.7), and (6.8) are independent of orthonormal basis chosen. Fix an orthonormal basis βsu(N ) for su(N ). Observe that the matrix i N IN is an element of u(N ) of unit norm such that (cid:28) i N IN , X(cid:29)u(N ) = 0 for all X ∈ su(N ) ⊆ u(N ). Hence βu(N ) = βsu(N ) ∪ { i XX∈βsu(N ) X 2 = XX∈βu(N ) N IN} is an orthonormal basis for u(N ). Thus X 2 −(cid:18) i N IN(cid:19)2 =(cid:18)−1 + 1 N 2(cid:19) IN , which proves (6.5). The remaining magic formulas are proven similarly. Proposition 6.3 (Derivative formulas for SU(N )). The following identities hold on 54 SU(N ) and SL(N, C): ∆SU(N )Am = 21m≤−2 ∆SU(N )Am = −21m≥2 m−1Xj=1 −1Xj=m+1 XX∈βsu(N ) eX tr(Am) · eXAp = jAj tr(Am−j) − mAm + jAj tr(Am−j ) + mAm + m2 N 2 Am, m ≥ 0 m2 N 2 Am, m < 0 (6.9) (6.10) (6.11) mp N 2 (−Am+p + Ap tr(Am)), m, p ∈ Z. Proof. The proof of these results is essentially identical to the proofs of the corresponding derivative formulas for U(N ) in [8, Theorem 3.3]; one need only to replace the magic formulas for U(N ) by the magic formulas for SU(N ) in the proof, wherever applicable. The derivative formulas of Proposition 6.3 comprise the key ingredient for the intertwin- ing formulas for ∆SU(N ) and BSU(N ) (Theorems 1.4 and 1.5). The proofs of these results are entirely analogous to the corresponding results for U(N ) (cf. [8, Theorems 1.8, 1.9]). The only change required in the proof is to replace the magic formulas and derivative formulas for U(N ) by the corresponding formulas for SU(N ) from Propositions 6.2 and 6.3. By keeping track of these changes, we see that for P ∈ C[u, u−1; v], the intertwining formula for ∆SU(N ) is s,τ ∆SU(N )PN = [D(2) N P ]N =(cid:20)(cid:18)L0 − 1 N 2 (2K− 1 + K− 2 − J )(cid:19) P(cid:21)N . The only difference between this and (3.44), the intertwining formula for ∆U(N ), is that for SU(N ), the 1/N 2 term on the right hand side contains the additional operator J . This is a consequence of the close relationship between the magic formulas for U(N ) and SU(N ). Finally, our main result regarding the free Segal-Bargmann transform for SU(N ), Theo- rem 1.6, is proven in the same way as the corresponding result for U(N ) (see [8, Theorem 1.11]); again, the only change necessary is the substitution of the magic and derivative formulas for SU(N ) in place of those for U(N ) wherever required. Acknowledgments The author thanks their PhD advisor Todd Kemp for suggesting the topic of the present paper, as well as his invaluable advice and extensive comments in completing this work. In addition, the author thanks the referee, whose suggestions greatly improved the pre- sentation of the results. References [1] Bargmann, V. On a Hilbert space of analytic functions and an associated integral transform. Comm. Pure Appl. Math. 14 (1961), 187 -- 214. 55 [2] Bargmann, V. Remarks on a Hilbert space of analytic functions. Proc. Nat. Acad. Sci. U.S.A. 48 (1962), 199 -- 204. [3] Biane, P. Free Brownian motion, free stochastic calculus and random matrices. In Free probability theory (Waterloo, ON, 1995), vol. 12 of Fields Inst. Commun. Amer. Math. Soc., Providence, RI, 1997, pp. 1 -- 19. [4] Biane, P. Segal-Bargmann transform, functional calculus on matrix spaces and the theory of semi-circular and circular systems. J. Funct. Anal. 144, 1 (1997), 232 -- 286. [5] C´ebron, G. Free convolution operators and free Hall transform. J. Funct. Anal. 265, 11 (2013), 2645 -- 2708. [6] Driver, B. K. On the Kakutani-Ito-Segal-Gross and Segal-Bargmann-Hall isomorphisms. J. Funct. Anal. 133, 1 (1995), 69 -- 128. [7] Driver, B. K., Gross, L., and Saloff-Coste, L. Holomorphic functions and subelliptic heat kernels over Lie groups. J. Eur. Math. Soc. (JEMS) 11, 5 (2009), 941 -- 978. [8] Driver, B. K., Hall, B. C., and Kemp, T. The large-N limit of the Segal-Bargmann transform on UN . J. Funct. Anal. 265, 11 (2013), 2585 -- 2644. [9] Driver, B. K., Hall, B. C., and Kemp, T. The complex-time Segal -- Bargmann transform. J. Funct. Anal. 278, 1 (2020). [10] Hall, B. C. The Segal-Bargmann "coherent state" transform for compact Lie groups. Journal of Functional Analysis 122, 1 (1994), 103 -- 151. [11] Ho, C.-W. The two-parameter free unitary Segal-Bargmann transform and its Biane-Gross- Malliavin identification. J. Funct. Anal. 271, 12 (2016), 3765 -- 3817. [12] Horn, R. A., and Johnson, C. R. Topics in matrix analysis. Cambridge University Press, Cambridge, 1994. Corrected reprint of the 1991 original. [13] L´evy, T. The master field on the plane. Ast´erisque, 388 (2017). [14] Milnor, J. Curvatures of left invariant metrics on Lie groups. Advances in Math. 21, 3 (1976), 293 -- 329. [15] Segal, I. E. Mathematical characterization of the physical vacuum for a linear Bose-Einstein field. (Foundations of the dynamics of infinite systems. III). Illinois J. Math. 6 (1962), 500 -- 523. [16] Segal, I. E. Mathematical problems of relativistic physics, vol. 1960 of With an appendix by George W. Mackey. Lectures in Applied Mathematics (proceedings of the Summer Seminar, Boulder, Col- orado. American Mathematical Society, Providence, R.I., 1963. [17] Segal, I. E. The complex-wave representation of the free boson field. In Topics in functional analysis (essays dedicated to M. G. Kreın on the occasion of his 70th birthday), vol. 3 of Adv. in Math. Suppl. Stud. Academic Press, New York-London, 1978, pp. 321 -- 343. [18] Stillwell, J. Naive Lie Theory. Undergraduate Texts in Mathematics. Springer New York, 2008. 56
1707.04663	1	1707	2017-07-14T23:30:13	Mixing inequalities in Riesz spaces	[ "math.FA", "math.CA", "math.PR" ]	Various topics in stochastic processes have been considered in the abstract setting of Riesz spaces, for example martingales, martingale convergence, ergodic theory, AMARTS, Markov processes and mixingales. Here we continue the relaxation of conditional independence begun in the study of mixingales and study mixing processes. The two mixing coefficients which will be considered are the $\alpha$ (strong) and $\varphi$ (uniform) mixing coefficients. We conclude with mixing inequalities for these types of processes. In order to facilitate this development, the study of generalized $L^1$ and $L^\infty$ spaces begun by Kuo, Labuschagne and Watson will be extended.	math.FA	math	Mixing inequalities in Riesz spaces ∗ Wen-Chi Kuo† School of Computational and Applied Mathematics University of the Witwatersrand Private Bag 3, P O WITS 2050, South Africa Michael J. Rogans School of Statistics and Actuarial Science University of the Witwatersrand Private Bag 3, P O WITS 2050, South Africa Bruce A. Watson ‡ School of Mathematics University of the Witwatersrand Private Bag 3, P O WITS 2050, South Africa June 18, 2021 Abstract Various topics in stochastic processes have been considered in the abstract setting of Riesz spaces, for example martingales, martingale convergence, ergodic theory, AMARTS, Markov processes and mixingales. Here we continue the relaxation of conditional inde- pendence begun in the study of mixingales and study mixing processes. The two mixing coefficients which will be considered are the α (strong) and ϕ (uniform) mixing coeffi- cients. We conclude with mixing inequalities for these types of processes. In order to facilitate this development, the study of generalized L1 and L∞ spaces begun by Kuo, Labuschagne and Watson will be extended. ∗Keywords: Riesz spaces, vector lattices; mixing processes, dependent processes, conditional expectation operators. Mathematics subject classification (2000): 47B60, 60G20. †Supported in part by NRF grant number CSUR160503163733. ‡Supported in part by the Centre for Applicable Analysis and Number Theory and by NRF grant number IFR2011032400120. 1 1 Introduction Mixing processes are stochastic processes in which independence assumptions are re- placed by a measure of independence called the mixing coefficient, see [1, 3, 6, 16] for measure theoretic essentials of mixing processes. In the Riesz space (measure free) set- ting, processes which require independence, such as Markov processes, were considered with independence replaced by conditional independence, see [21]. In line with the above approach, mixingales (processes with independence/conditional independence in the limit) were considered in the Riesz space setting in [14]. In this work we will pose α (strong) and ϕ (uniform) mixing processes in the Riesz space setting. Core to the theory of mixing is the family of inequalities generally referred to as the mixing inequal- ity, the conditional Riesz space analogues of which form the focus of this paper, see Section 4. We will give mixing inequalities for both α and ϕ mixing (the one being an easy consequence of the other). We refer the reader to [1, 6, 10, 16, 18] for the measure, non-conditional analogues. To facilitate this study the Riesz space analogues of the Lp spaces introduced in [15] will be revisited. In [12], it was shown that a conditional expectation operator, T , on a Riesz space, E, admits a unique maximal extension to a conditional expectation operator, also denoted T , in the universal completion, Eu, of E, with domain a Dedekind complete Riesz space, which will be denoted L1(T ). The procedure used there was based on that of de Pagter and Grobler, [8], for the measure theoretic setting. We observe here that the range of the maximal extension of the conditional expectation operator, i.e. R(T ) := {T f f ∈ L1(T )}, is a universally complete f -algebra and that L1(T ) is an R(T )-module. This prompts the definition of an R(T ) (vector valued) norm k · kT,1 := T · on L1(T ). Here the homogeneity is with respect to multiplication by elements of R(T )+. Following in a similar manner L∞(T ) is taken to be the subspace of L1(T ) composed of R(T ) bounded elements. An R(T ) valued norm k · kT,∞ := inf{g ∈ R(T )+ · ≤ g}, is defined on L∞(T ). This extends on the concepts of L∞(T ) defined in [15]. In [4, Sections 39, 42 and 43] Dellacherie and Meyer gave an extension of martingale theory to σ-finite processes. As a direct application of the material presented in Sections 2 to 4, we give, in Section 5, an extension of the theory of mixing theory to σ-finite processes. The extension of mixing theory even to this special case, to the knowledge of the authors, has not been considered in the literature. Natural connections with the theory presented here are to laws of large numbers and other convergence theorems. For the conditionally independent case in Riesz spaces we refer the reader to Stoica [19]. 2 2 Preliminaries For general background on Riesz spaces we refer the reader to [23] and for the foundations of stochastic processes in Riesz spaces to [11, 12, 13, 14, 15]. We will present only the essential results from the theory of Riesz space stochastic processes required for our consideration of mixing processes. Definition 2.1 Let E be a Dedekind complete Riesz space with weak order unit. A positive order continuous linear projection T on E with range R(T ), a Dedekind complete Riesz subspace of E, is said to be a conditional expectation operator if T e is a weak order unit of E for each weak order unit e of E. A Riesz space E is said to be universally complete if E is Dedekind complete and every subset of E which consists of mutually disjoint elements has a supremum in E. A Riesz space Eu is said to be a universal completion of the Riesz space E if Eu is universally complete and contains E as an order dense subspace. If e is a weak order unit of E then e is also a weak order unit of Eu, see [23]. We say that a conditional expectation operator, T , on a Riesz space is strictly positive if T f = 0 implies that f = 0. As shown in [12], a strictly positive conditional expectation operator, T , on a Riesz space can be extended to its so called natural domain (maximal domain to which it can be extended as a conditional expectation operator) denoted dom(T ). We set L1(T ) := dom(T ) and we denote the extension of T to L1(T ) again by T . This is consistent with the special case of T an expectation operator in the measure theoretic setting, see [8]. In particular, if E is a Dedekind complete Riesz space with weak order unit and conditional expectation operator, T , we say that E is T -universally complete if E = L1(T ). From the definition of dom(T ), E is T -universally complete if and only if for each upwards directed net (fα)α∈Λ in E+ such that (T fα)α∈Λ is order bounded in Eu, we have that (fα)α∈Λ is order convergent in E. If e is a weak order unit of E then we denote the f -algebra of e bounded elements by Ee := {f ∈ E f ≤ ke, for some k ∈ R+} and set L∞(T ) := {f ∈ L1(T ) f ≤ g, for some g ∈ R(T )+}. We recall from [12, Theorem 5.3] that each conditional expectation operator T is an averaging operator in the sense that if f ∈ R(T ) and g ∈ E with f g ∈ E then T (f g) = f T (g). Theorem 2.2 Let E be a T -universally complete Riesz space, where T is a conditional expectation operator on E, and let e be a weak order unit for E with T e = e. Then R(T ) 3 is universally complete and hence an f -algebra. In addition E = L1(T ) and L∞(T ) are R(T )-modules. In order to show that R(T ) is universally complete, we need to show that for Proof: each W ⊂ R(T )+ consisting of mutually disjoint elements, i.e. if u, v ∈ W with u 6= v then u ∧ v = 0, we have that w := ∨v∈W v ∈ R(T ). To this end, let W be as above and set w := ∨v∈W v, which exists in Eu. Now u ∧ ne ∈ R(T )+ and u ∧ ne ≤ ne, for all n ∈ N, u ∈ W . Here {u ∧ ne u ∈ W } ⊂ R(T ) is bounded above by ne ∈ R(T ) and so the Dedekind completeness of R(T ) allows us to conclude that Now w ∧ ne ↑ w and w ∧ ne = ∨u∈W (u ∧ ne) ∈ R(T ). T (w ∧ ne) = w ∧ ne ≤ w ∈ Eu, n ∈ N, so the T -universal completeness of E gives that w ∈ E. The order continuity of T gives T w = w and w ∈ R(T ). Thus R(T )u = R(T ) from which it follows that R(T ) is an f -algebra. Since f g = f +g+ + f −g− − f −g+ − f +g−, to show that L1(T ) is an R(T )-module, it suffices to show that f g ∈ L1(T ) for each f ∈ L1(T )+ and g ∈ R(T )+. Now from the averaging properties of T and as (f ∧ ne)g ∈ L1(T ) we have that T ((f ∧ ne)g) = T (f ∧ ne)g ≤ T (f )g ∈ Eu + for all n ∈ N. However L1(T ) is T-universally complete and (f ∧ne)g ↑ f g, so f g ∈ L1(T ). Finally, to show that L∞(T ) is an R(T )-module we, take f ∈ L∞(T )+ and g ∈ R(T )+. Here there is F ∈ R(T ) with 0 ≤ f ≤ F . Now as R(T ) is an algebra. Hence f g ∈ L∞(T ). f g ≤ F g ∈ R(T )+ We note the connection with the work of Grobler and de Pagter in [8], where in the measure theoretic case, it is shown that the range space of the maximal extension of a classical conditional expectation operator is an algebra. 3 The Lp(T ), p = 1, ∞, spaces with T -norms In the previous section we define the Lp(T ), p = 1, ∞, spaces, see also [12, 15]. Here we present the corresponding vector valued generalizations of the Lp-norms. In particular, 4 these norms take their values in the positive cone of the universally complete algebra R(T ). We also refer the reader to the Riesz semi-norm approach used by Grobler and Labuschagne, [9], to study the space L2(T ). For some recent progress in this area we refer the reader to [2]. Definition 3.1 Let E be a Dedekind complete Riesz space with weak order unit and T be a strictly positive conditional expectation operator on E. If E is an R(T )-module and φ : E → R(T )+ with (a) φ(f ) = 0 if and only if f = 0, (b) φ(gf ) = gφ(f ) for all f ∈ E and g ∈ R(T ), (c) φ(f + h) ≤ φ(f ) + φ(h) for all f, h ∈ E, then φ will be called an R(T )-valued norm on E. Theorem 3.2 Let E be a T -universally complete Riesz space with weak order unit, where T is a strictly positive conditional expectation operator on E. The map defines an R(T )-valued norm on L1(T ), and the map f 7→ T f =: kf kT,1 f 7→ kf kT,∞ = inf {g ∈ R(T )+ f ≤ g} defines an R(T )-valued norm on L∞(T ). Proof: For L1(T ), condition (a) of Definition 3.1 follows from the strict positivity of T and while (c) of Definition 3.1 follows from the linearity of T and · obeying the triangle inequality. For (b) of Definition 3.1, let f ∈ E and g ∈ R(T ), then we observe that the terms f +g+, f −g+, f +g−, f −g− are disjoint and positive. Thus gf = f +g+ − f −g+ − f +g− + f −g− = f +g+ + f −g+ + f +g− + f −g− = gf . Here g ∈ R(T ), so T gf = T (gf ) = gT f , from which Definition 3.1 part (b) follows. We now consider the case of L∞(T ). Here (a) and (b) follow directly from the definition of k · kT,∞. For (c), consider f, g ∈ L∞(T ). As f + g ≤ f + g, it follows that {h ∈ R(T )+ f + g ≤ h} ⊂ {h ∈ R(T )+ f + g ≤ h}. Therefore kf + gkT,∞ = inf{h ∈ R(T )+ f + g ≤ h} ≤ inf{h ∈ R(T )+ f + g ≤ h}. 5 Writing h = hf + hg, for hf , hg ∈ R(T )+, it follows that {hf + hg f ≤ hf , g ≤ hg, hf , hg ∈ R(T )+} ⊂ {h ∈ R(T )+ f + g ≤ h}, giving inf{hf + hg f ≤ hf , g ≤ hg, hf , hg ∈ R(T )+} ≥ inf{h ∈ R(T )+ f + g ≤ h}. Combining the above and noting that the conditions f ≤ hf and g ≤ hg are indepen- dent, we have kf + gkT,∞ ≤ inf{hf + hg f ≤ hf , g ≤ hg, hf , hg ∈ R(T )+} = inf{hf f ≤ hf , hf ∈ R(T )+} + inf{hg g ≤ hg, hg ∈ R(T )+} = kf kT,∞ + kgkT,∞. Theorem 3.3 If f ∈ L1(T ) and g ∈ L∞(T ), then gf ∈ L1(T ) and kgf kT,1 ≤ kgkT,∞kf kT,1. Proof: From linearity, it suffices to show that gf ∈ L1(T ) for f ∈ L1(T )+ and g ∈ L∞(T )+. Here there exists h ∈ R(T )+ such that g ≤ h, and we note from earlier that hf ∈ L1(T )+, but 0 ≤ gf ≤ hf , giving gf ∈ L1(T )+. Now for each h ∈ R(T )+ with g ≤ h we have from T being an averaging operator that kgf kT,1 = T gf = T (gf ) ≤ T (hf ) = hT f = hkf kT,1. However, from the compatability of the multiplicative structure with the order structure, inf{hkf kT,1 g ≤ h, h ∈ R(T )+} = inf{h ∈ R(T )+ g ≤ h}kf kT,1 = kf kT,1kgkT,∞, from which the result follows. Setting f = e, where e is a weak order unit with T e = e and which has been chosen as the algebraic unit for the f -algebra structure, we have the following corollary to Theorem 3.3. Corollary 3.4 If g ∈ L∞(T ) then kgkT,1 ≤ kgkT,∞. To conclude this section, we give a variant of the conditional Jensen's inequality. For additional details on conditional Jensen's inequalities in Riesz spaces, see [7]. 6 Theorem 3.5 If S is a conditional expectation operator on L1(T ) compatible T (in the sense that T S = T = ST ), then for all f ∈ Lp(T ), p = 1, ∞. kSf kT, p ≤ kf kT,p, Proof: For p = 1, as S is a positive operator, kSf kT,1 = T Sf ≤ T Sf = T f = kf kT,1. For p = ∞, if f ≤ g ∈ R(T )+, then g = T g and from the positivity of S we have Sf ≤ Sf so Sf ≤ Sf ≤ Sg = ST g = T g = g. Thus {g ∈ R(T )+ f ≤ g} ⊂ {g ∈ R(T )+ Sf ≤ g}, from which it follows that kSf kT,∞ ≤ kf kT,∞. 4 Mixing inequalities In this section we consider conditional versions of strong mixing (α-mixing) and uniform mixing (ϕ-mixing) in the Riesz space setting. At the core of mixing processes is the family of inequalities generally termed the mixing inequalities, see [1, 6, 10, 16, 18] for the classical mixing inequalities. We begin by giving conditional definitions of strong and uniform mixing in the measure theoretic setting. These conditional definitions of mixing admit direct generalizations to Riesz spaces with conditional expectation operators. We conclude with a conditional mixing inequality for conditionally strong mixing processes in Riesz spaces. This yields, directly, conditional mixing inequalities for Riesz space con- ditionally uniform mixing processes and conditional mixing inequalities for conditionally strong and conditionally uniform mixing processes in the measure space setting. In the classical measure theoretic setting, the strong mixing coefficient is defined as follows. Let (Ω, F , µ) be a probability space and A and B be sub-σ-algebras of F. The strong mixing coefficient between A and B is α(A, B) = sup{µ(A ∩ B) − µ(A)µ(B) A ∈ A, B ∈ B}. (4.1) In place of the expectation, we could condition on a sub-σ-algebra, say C, of A ∩ B which would result in the C-conditioned strong mixing coefficient αC(A, B) = sup{E[IAIBC] − E[IAC]E[IBC] A ∈ A, B ∈ B]. (4.2) 7 Definition 4.1 Let E be a Dedekind complete Riesz space with weak order unit, say e, and conditional expectation operator, T , with T e = e. If U is a conditional expectation operators on E, with T U = T = U T , then we say that U is compatible with T . If U is a conditional expectation on E compatible with T then we denote by B(U ) the set of band projections P on E with P e ∈ R(U ). In light of (4.2), we define the strong mixing coefficient in a Riesz space with conditional expectation operator as follows. Definition 4.2 Let E be a Dedekind complete Riesz space with weak order unit, say e, and conditional expectation operator, T , with T e = e. We define the T -conditional strong mixing coefficient with respect to the conditional expectation operators U and V on E compatible with T , by αT (U, V ) := sup{T P Qe − T P e · T Qe P ∈ B(U ), Q ∈ B(V )}. We can now give bounds for the T -conditional strong mixing coefficient in terms of the T -conditional norm. Theorem 4.3 Let E be a T -universally complete Riesz space, E = L1(T ), where T is a conditional expectation operator on E where E has a weak order unit, say e, with T e = e. Let U and V be conditional expectation operators on E compatible with T , then αT (U, V ) ≤ sup kU Qe − T QekT,1 ≤ 2αT (U, V ). Q∈B(V ) Proof: Let P ∈ B(U ) and Q ∈ B(V ) then as T is an averaging operator T P e · T Qe = T (P e · T Qe). The f -algebra structure gives that P e · Qe = P Qe and hence T (P e · Qe) − T P e · T Qe = T [P e · (Qe − T Qe)] = T [P (Qe − T Qe)]. From the Ando-Douglas-Radon-Nikod`ym Theorem, see [22], it follows that T [P (Qe − T Qe)] = T [P U (Qe − T Qe)] (4.3) which is maximized, over P ∈ B(U ), when P = P+ is the band projection onto the band generated by [U (Qe − T Qe)]+, in which case T P+U (Qe − T Qe) = T [U (Qe − T Qe)]+. (4.4) Hence sup{T P Qe − T P e · T Qe P ∈ B(U )} = T [U (Qe − T Qe)]+. 8 We note that (4.3) is minimized, over P ∈ B(U ), when P = P− is the band projection onto the band generated by [U (Qe − T Qe)]−, in which case T P−U (Qe − T Qe) = −T [U (Qe − T Qe)]−. (4.5) Hence Now sup{−T P Qe + T P e · T Qe P ∈ B(U )} = T [U (Qe − T Qe)]−. sup{T P Qe − T P e · T Qe P ∈ B(U )} = sup{(T P Qe − T P e · T Qe) ∨ (−T P Qe + T P e · T Qe) P ∈ B(U )} ≤ sup{(T P1Qe − T P1e · T Qe) ∨ (−T P2Qe + T P2e · T Qe) P1, P2 ∈ B(U )} = sup{(T P Qe − T P e · T Qe) P ∈ B(U )} ∨ sup{(−T P Qe + T P e · T Qe) P ∈ B(U )} and applying (4.4) to each of the expressions in the last line of the above gives sup{T P Qe − T P e · T Qe P ∈ B(U )} = T ([U (Qe − T Qe)]+) ∨ T ([U (Qe − T Qe)]−). From the linearity of the expectation and conditional expectation operators, U [(I − Q)e − T (I − Q)e] = −U [Qe − T Qe]. (U [(I − Q)e − T (I − Q)e])+ = [U (Qe − T Qe)]− Hence and thus sup{T P Qe − T P e · T Qe P ∈ B(U )} ≤ T [U (Qe − T Qe)]+ ∨ T (U [(I − Q)e − T (I − Q)e])+. (4.6) Since Q ∈ B(V ) implies I − Q ∈ B(V ), the first inequality of the Theorem follows from taking the supremum of (4.6) over V ∈ B(V ). Combining (4.3), (4.4) and (4.5), we have kU Qe − T QekT,1 = T U (Qe − T Qe) = T (U (Qe − T Qe))+ + T (U (Qe − T Qe))− = T P+U (Qe − T Qe) − T P−U (Qe − T Qe). As U is an averaging operator and P−, P+ ∈ B(U ) we have T P±U (Qe − T Qe) = T U P±(Qe − T Qe) = T P±(Qe − T Qe). The positivity of T and the definition of αT (U, V ) give T P±(Qe − T Qe) ≤ T P±(Qe − T Qe) ≤ αT (U, V ). 9 Combining the above gives kU Qe − T QekT,1 ≤ 2αT (U, V ) which proves the second inequality of the Theorem. It should be noted here that the product T P e · T Qe exists in E, as shown in the previous section. We now consider the uniform mixing coefficient. In the measure theoretic setting the uniform mixing coefficient is defined as follows. Let (Ω, F , µ) be a probability space and A and B be sub-σ-algebras of F. The uniform mixing coefficient between A and B is ϕ(A, B) = sup{µ(BA) − µ(B) A ∈ A, B ∈ B, µ(A) > 0}. (4.7) As with the strong mixing coefficient, the uniform mixing coefficient has an interesting formulation in terms of Lp norms. Lemma 4.4 Let (Ω, F , µ) be a probability space and A and B be sub-σ-algebras of F, then ϕ(A, B) = sup B∈B kE[IB − E[IB]A]k∞. Proof: We begin by observing that, since E[IBA] − E[IB] is A-measurable, kE[IBA] − E[IB]k∞ = sup(cid:26) E[IA(E[IBA] − E[IB])] µ(A) Now for A ∈ A with µ(A) > 0 we have A ∈ A, µ(A) > 0(cid:27) . (cid:12)(cid:12)(cid:12)(cid:12) E[IA(E[IBA] − E[IB])] µ(A) = E[IAIB] − E[IA]E[IB] µ(A) = µ(BA) − µ(B), from which the Lemma follows. The above Lemma leads naturally to conditional and Riesz space variants of the uniform mixing coefficient. Definition 4.5 Let E be a Dedekind complete Riesz space with weak order unit, say e, T a conditional expectation operator on E and E having weak order unit say e with T e = e. Let U and V be conditional expectation operators on E compatible with T , then ϕT (U, V ) = sup kU Qe − T QekT,∞. Q∈B(V ) 10 Combining Corollary 3.4 with Theorem 4.3 and Definition 4.5 we have the following theorem. Theorem 4.6 Let E be a T -universally complete Riesz space, E = L1(T ), where T is a conditional expectation operator on E where E has a weak order unit, say e, with T e = e. Let U and V be conditional expectation operators on E compatible with T , then αT (U, V ) ≤ ϕT (U, V ). The mixing inequalities now give bounds on the norm of the differences between the composition of the conditional expectation operators, say U and V , compatible with the conditional expectation operator T . It should be noted that if U and V are conditionally independent with respect to T then U V = T = V U . The measure theoretic version was proved in [17], wherein results from [5] were used. Theorem 4.7 Let E be a T -universally complete Riesz space, T a conditional expec- tation operator on E and e a weak order unit for E with e = T e. Let U and V be conditional expectation operators on E compatible with T , then, for f ∈ R(V ) ∩ L∞(T ), we have kU f − T f kT,1 ≤ 4αT (U, V )kf kT,∞. Proof: Let g := kf kT,∞ ∈ R(T )+, then f is in the order interval [−g, g]. Hence, from [12, Theorem 4.2], there are sequences (f ± n ) ⊂ R(V ) with 0 ≤ f ± n ↑ f ± of the form N ± n i,ng, f ± n = θ± i,nP ± Xi=1 i,n have P ± i,ne ∈ R(V ) and P ± 1,n < θ± 0,n < θ± 2,n < · · · < θ± where the band projections P ± real numbers θ± i,n have 0 =: θ± n ≤ g and P ± f ± i,ng ∈ R(V ). Set i,nP ± Nn,n. Also θ± j,n = 0 for i 6= j, and the Nn,n ≤ 1 since then Q± i,n = N ± n Xj=i P ± i,n, f ± n = N ± n Xi=1 i,nQ± β± i,ng, where 0 < β± i,n := θ± i,n − θ± i−1,n. Here N ± n Xi=1 β± i,n ≤ 1 and Q± i,ne ∈ R(V ). 11 Now U Q± i,ng − T Q± i,ng = U (g · Q± i,ne) − T (g · Q± i,ne). Since U and T are averaging operators with g ∈ R(T ) ⊂ R(U ) it follows that U (g · Q± i,ne) − T (g · Q± i,ne) = g · (U Q± i,ne − T Q± i,ne) = g · U Q± i,ne − T Q± i,ne. (4.8) Hence T U Q± i,ng − T Q± i,ng = T (g · U Q± i,ne − T Q± i,ne) = g · T U Q± i,ne − T Q± i,ne. Theorem 4.3 gives and hence T U Q± i,ne − T Q± i,ne ≤ 2αT (U, V ) T U Q± i,ng − T Q± i,ng ≤ 2αT (U, V ) · g. Applying the above to f ± n gives n − T f ± n T U f ± N ± n β± i,nT U Q± i,ng − T Q± i,ng ≤ Xi=1 ≤ 2 N ± n Xi=1 β± i,nαT (U, V ) · g ≤ 2αT (U, V ) · g. Taking the order limit as n → ∞ and using the order continuity of conditional expecta- tion operators gives T U f ± − T f ± ≤ 2αT (U, V ) · g. Finally T U f − T f ≤ T U f + − T f + + T U f − − T f − ≤ 4αT (U, V ) · g which can be rewritten as in the statement of the theorem. Applying Theorem 4.7 to probability spaces we have the following corollary. Corollary 4.8 Let (Ω, F , µ) be a probability space, C, G, H be sub-σ-algebras of F with G and H containing C. For f ∈ L0(Ω, H, µ) with f bounded by g ∈ L0(Ω, C, µ) we have E[E[f G] − E[f C] C] ≤ 4αC(G, H)g. Setting f = V g in Theorem 4.7 and using Theorem 3.5, we obtain the following corollary. 12 Corollary 4.9 Let E be a T -universally complete Riesz space with weak order unit e = T e, where T is a conditional expectation operator on E. Let U and V be conditional expectation operators on E compatible with T . Then for g ∈ L∞(T ) kU V g − T gkT,1 ≤ 4αT (U, V )kgkT,∞. The next theorem, see [18] for the measure theoretic case, arises using a similar procedure as the above for theorem but we now proceed from (4.8) and the definition of ϕT (U, V ) as follows: U Q± i,ne) ≤ g · ϕT (U, V ). i,ng − T Q± i,ng = g · (U Q± i,ne − T Q± Here U f ± n − T f ± n ≤ N ± n Xi=1 β± i,nU Q± i,ng − T Q± i,ng ≤ N ± n Xi=1 β± i,nϕT (U, V ) · g ≤ ϕT (U, V ) · g. Letting n → ∞ gives and thus U f ± − T f ± ≤ ϕT (U, V ) · g kU f − T f kT,∞ ≤ 2ϕT (U, V ) · kf kT,∞ from which the following theorem follows. Theorem 4.10 Let E be a T -universally complete Riesz space with weak order unit e = T e, where T is a conditional expectation operator on E. Let U and V be conditional expectation operators on E compatible with T . Then for f ∈ L∞(T ) ∩ R(V ) kU f − T f kT,1 ≤ kU f − T f kT,∞ ≤ 2ϕT (U, V )kf kT,∞. Applying Theorem 4.10 to probability spaces we have the following corollary. Corollary 4.11 Let (Ω, F , µ) be a probability space, C, G, H be sub-σ-algebras of F with G and H containing C. For f ∈ L0(Ω, H, µ) with f bounded by g ∈ L0(Ω, C, µ) we have E[f G] − E[f C] ≤ 2ϕC(G, H)g. 5 Mixing for σ-finite processes In this section we consider the simplest non-trivial application, that is, to σ-finite pro- cesses, hence giving a theory of conditional mixing for such processes. In this concrete 13 example the spaces and operators can be clearly identified. A consideration of σ-finite processes in the context of martingale theory can be found in the work of Dellacherie and Meyer, [4, Sections 39, 42 and 43]. Let (Ω, A, µ) be a σ-finite measure space, which to be interesting should have µ(Ω) = ∞, and let (Ωi)i∈N be a µ-measurable partition of Ω into sets of finite positive measure. Let A0 be the sub-σ-algebra of A generated by (Ωi)i∈N. We take the Riesz space E = L∞(Ω, A, µ) and the conditional expectation operator T = E[ · A0]. For f ∈ E we have T f (ω) = RΩi f dµ µ(Ωi) , for ω ∈ Ωi. (5.9) Here we have that the universal completion, Eu, of E is the space of all A-measurable functions. The T -universal completion of E is the space L1(T ) = (cid:26)f ∈ Eu (cid:12)(cid:12)(cid:12)(cid:12) ZΩi f dµ < ∞ for all i ∈ N(cid:27) , which is characterized by f Ωi ∈ L1(Ω, A, µ), for each i ∈ N. Here T can be extended to an L1(T ) conditional expectation operator as per (5.9). We note that E has weak order unit e = 1, the function identically 1 on Ω, which again is a weak order unit for L1(T ), but is not in L1(Ω, A, µ). The range of the generalized conditional expectation operator T is R(T ) = {f ∈ Eu f a.e. constant on Ωi, i ∈ N} , which is an f -algebra. The last of the spaces to be considered is L∞(T ) = {f ∈ Eu f essentially bounded on Ωi for each i ∈ N} . The vector norms on L1(T ) and L∞(T ) are kf k1(ω) = T f (ω) = RΩi f dµ µ(Ωi) , for ω ∈ Ωi, f ∈ L1(T ), kf k∞(ω) = ess supΩif , for ω ∈ Ωi, f ∈ L∞(T ). (5.10) (5.11) We note that L1(Ω, A, µ) ( L1(T ), L∞(Ω, A, µ) ( L∞(T ), L∞(T ) ⊂ L1(T ) while L∞(Ω, A, µ) 6⊂ L1(Ω, A, µ). Let C and D be sub-σ-algebras of A which contain A0. The α-mixing coefficient of C and D conditioned on A0 (which in measure theoretic terms could be denote αA0(C, D) is αT (U, V ). Here U and V are the restrictions to L1(T ) of the extensions to L1(U ) and L1(V ) respectively of the conditional expectation operators U and V on E conditioning with respect to the σ-algebras C and D. In this example case these operators can be given explicitly by U (f ) = V (f ) = Ei[f IΩiC], Ei[f IΩiD], ∞ Xi=1 Xi=1 ∞ 14 (5.12) (5.13) for f ∈ L1(T ). Here the conditional expectation Ei[f IΩiC] = Ei[f C] is the conditional expectation on Ωi of f Ωi with respect to the probability measure µi(A) := µ(A∩Ωi) µ(Ωi) and the σ-algebra {C ∩ ΩiC ∈ C}, and similarly for C replaced by D. Since the structure of the example is extremely simple, an explicit computation can be carried out to give αT (U, V ) = αA0(C, D) = = ∞ Xi=1 Xi=1 ∞ IΩi sup {µi(C ∩ D) − µi(C)µi(D) C ∈ C, D ∈ D} αi(C, D)IΩi , where αi(C, D) is the α-mixing coefficient of σ-algebras C and D with respect to the prob- ability measure µi. Corollary 4.9 gives that if g is µ-measurable and essential bounded on each Ωi, i ∈ N, then kU V g − T gkT,1 ≤ 4αT (U, V )kgkT,∞, which in this example case can be written as 1 µ(Ωi) ZΩi (cid:12)(cid:12)(cid:12)(cid:12) 1 µ(Ωi) ZΩi g dµ(cid:12)(cid:12)(cid:12)(cid:12) Ei[Ei[gD]C] − dµ ≤ 4αi(C, D)ess supΩig for each i ∈ N. The conditional uniform mixing coefficient is given by ϕT (U, V ) = ϕA0(C, D) = = ∞ Xi=1 Xi=1 ∞ IΩi sup D∈D ess supΩi Ei[IDC] − µi(D) ϕi(C, D)IΩi , where ϕi(C, D) is the ϕ-mixing coefficient of C and D relative to the probability measure µi. For g as above, Theorem 4.10 gives kU V g − T gkT,∞ ≤ 2ϕT (U, V )kgkT,∞, which in the special case under consideration yields Ei[Ei[gD]C] − (cid:12)(cid:12)(cid:12)(cid:12) a.e. on Ωi, for each i ∈ N. 1 µ(Ωi) ZΩi g dµ(cid:12)(cid:12)(cid:12)(cid:12) ≤ 2ϕi(C, D) ess sup Ωig It should be noted that the work presented here also applies to processes where the random variables are Riesz space valued, say Lp, and the conditional expectation, T , is generated by an arbitrary sub-σ-algebra of A. In this case we obtain a generalization of mixing to the context of vector measure. 15 References [1] P.P. Billingsley, Probability and Measure, John Wiley and Sons, 3rd edition, 1995. [2] S. Cerreia-Vioglio, M. Kupper, F. Maccheroni, M. Marinacci, N. Vo- gelpoth, Conditional Lp-spaces and the duality of modules over f-algebras, J. Math. Anal. Appl., (2016), http://dx.doi.org/10.1016/j.jmaa.2016.06.018. [3] P. Doukhan, Mixing: properties and examples, Lecture Notes in Statistics, 85 (1994), 15-23. [4] C. Dellacherie, P.-A. Meyer, Probabilities and Potentials: B, Theory of Mar- tingales, North Holland Publishing Company, 1982. [5] A. Dvoretzky, Asymptotic normality for sums of dependent random variables, Proc. Sixth Berkeley Symp. on Math. Statist. and Prob., 2 (1972), 513-535. [6] G.A. Edgar, L. Sucheston, Stopping times and directed processes, Cambridge University Press, 1992. [7] J.J. Grobler, Jensen's and martingale inequalities in Riesz spaces, Indag. Math., 25 (2014), 275-295. [8] J.J. Grobler, B. de Pagter, Operators representable as multiplication- conditional expectation operators, J. Operator Theory, 48 (2002), 15-40. [9] J.J. Grobler, C.C.A. Labuschagne, The Ito integral for Brownian motion in vector lattices: Part 2, J. Math. Anal. Appl., 423 (2015), 820-833. [10] I.A. Ibragimov, Some limit theorems for stationary processes, Theory Probab. Appl., 7 (1962), 349-382. [11] W.-C. Kuo, C.C.A. Labuschagne, B.A. Watson, Discrete-time stochastic pro- cesses on Riesz spaces, Indag. Math., 15 (2004), 435-451. [12] W.-C. Kuo, C.C.A. Labuschagne, B.A. Watson, Conditional expectations on Riesz spaces, J. Math. Anal. Appl., 303 (2005), 509-521. [13] W.-C. Kuo, C.C.A. Labuschagne, B.A. Watson, Convergence of Riesz space martingales, Indag. Math., 17 (2006), 271-283. [14] W.-C. Kuo, J.J. Vardy, B.A. Watson, Mixingales on Riesz spaces, J. Math. Anal. Appl., 402 (2013), 731-738. [15] C.C.A. Labuschagne, B.A. Watson, Discrete stochastic integration in Riesz spaces, Positivity, 14 (2010), 859-875. [16] Z. Lin, Z. Bai, Probability Inequalities, Science Press Beijing and Springer-Verlag, Berlin, Heidelberg, 2010. [17] D.L. McLeish, A maximal inequality and dependent strong laws, Ann. Probab., 3 (1975), 829-839. [18] R.J. Serfling, Contributions to central limit theory for dependent variables, Ann. Math. Stat., 39 (1968), 1158-1175. 16 [19] G. Stoica, Limit laws in vector lattices, Quaetiones Math., 38 (2015), 829-833. [20] V.G. Troitsky, F. Xanthos Spaces of regular abstract martingales, J. Math. Anal. Appl., (2016), http://dx.doi.org/10.1016/j.jmaa.2015.09.082. [21] J.J. Vardy, B.A. Watson, Markov processes on Riesz spaces, Positivity, 16 (2012), 373-391. [22] B.A. Watson, An Ando-Douglas type theorem in Riesz spaces with a conditional expectation, Positivity, 13 (2009), 543-558. [23] A.C. Zaanen, Introduction to Operator Theory in Riesz Spaces, Springer Verlag, 1997. 17
1206.1613	1	1206	2012-06-07T20:56:27	Average Number of Lattice Points in a Disk	[ "math.FA" ]	The difference between the number of lattice points in a disk of radius $\sqrt{t}/2\pi$ and the area of the disk $t/4\pi$ is equal to the error in the Weyl asymptotic estimate for the eigenvalue counting function of the Laplacian on the standard flat torus. We give a sharp asymptotic expression for the average value of the difference over the interval $0 \leq t \leq R$. We obtain similar results for families of ellipses. We also obtain relations to the eigenvalue counting function for the Klein bottle and projective plane.	math.FA	math	Average Number of Lattice Points in a Disk Sujay Jayakar Robert S. Strichartz ∗ Abstract The difference between the number of lattice points in a disk of radius √t/2π and the area of the disk t/4π is equal to the error in the Weyl asymptotic estimate for the eigenvalue counting function of the Laplacian on the standard flat torus. We give a sharp asymptotic expression for the average value of the difference over the interval 0 ≤ t ≤ R. We obtain similar results for families of ellipses. We also obtain relations to the eigenvalue counting function for the Klein bottle and projective plane. 1 The simplest case Consider the standard flat torus [0, 1] × [0, 1] with boundaries identified. The eigenfunctions of the Laplacian are e2πin·x for n ∈ Z2 with eigenvalues (2π)2n2, so the eigenvalue counting function is N (t) = # (1.1) t/2π of radius √t/2π about the the number of lattice points inside the disk B√ origin. To first approximation N (t) is the area of the disk t/4π, and this is exactly the Weyl asymptotic law. The problem of estimating the difference , (cid:110) n ∈ Zd : n ≤ √t/2π (cid:111) D(t) = N (t) − t 4π (1.2) is notoriously difficult (conjectured to be O(t1/4+) for every > 0). Here ∗Research supported in part by NSF grant #DMS-0652440 2010 Mathematics Subject Classification. Primary 35J05; Primary 42B99. Key words and phrases: lattice points, Weyl asymptotics, Bessel function 1 Figure 1: D(t) Figure 2: t1/4D(t) Figure 3: A(t) Figure 4: t1/4A(t) we study the simpler problem of approximating the average value (cid:90) R A(R) = 1 R D(t) dt. 0 (1.3) Note that we are not taking the absolute value of D(t) in the average, so we may exploit the cancellation from regions where N (t) is greater than and less than t/4π. We will show that A(R) = O(R−1/4) as R → ∞, and more precisely A(R) = g(R −1/2)R −1/4 + O(R −3/4) as R → ∞ (1.4) where g(R) is an explicit uniformly almost periodic function of mean value zero. Somewhat different but related ideas are given in Bleher [2, 3]. The 2 0200040006000800010000t−15−10−5051015D(t)0200040006000800010000t−2.0−1.5−1.0−0.50.00.51.01.52.0t−1/4D(t)0200040006000800010000t−0.4−0.20.00.20.4A(t)0200040006000800010000t−2.0−1.5−1.0−0.50.00.51.01.52.0t1/4A(t)following Lemma is well-known (see [4], p. 74), but we include the proof for the convenience of the reader. Lemma 1. We have A(R) = (cid:88) n(cid:54)=0 √R) 1 πn2 J2(n (1.5) where n = (n1, n2) is a variable in Z2, and J2 denotes the Bessel function. The series in (1.5) converges uniformly and absolutely. Proof. Let χt denote the characteristic function of the ball B√ well-known that t/2π. It is χt(z) = z (cid:54)= 0 z = 0 (1.6) Following standard methods (see [7] or [4]) we apply the Poisson summation formula to FR,δ = 1 R χt ∗ ψδ dt, 0 (1.7) where ψδ is a smooth approximate identity. The ψδ convolution makes FR,δ smooth, but eventually we will let δ → 0. Note that t (cid:40) √ 2πz J1(z√t) (cid:90) R t 4π (cid:90) R 1 R n∈Z2 (cid:88) (cid:88) (cid:90) R 1 R n∈Z2 1 R 0 (cid:88) n∈Z2 N (t) dt = lim δ→0 0 χt(n) ψ(δn) dt (cid:90) R 0 1 R FR,δ(n) (cid:90) R dt + (cid:88) (cid:90) R n(cid:54)=0 0 t 4π (cid:88) FR,δ(n). (1.8) (1.9) √t 2πn J1(n√t) dt ψ(δn) J1(n√t) dt ψ(δn). (1.10) The Poisson summation formula gives (cid:88) n∈Z2 FR,δ(n) = = = by (1.6). Combining (1.8) and (1.9) yields √t 2πn A(R) = lim δ→0 1 R n(cid:54)=0 0 3 (cid:90) R 0 (cid:90) R 0 1 R sα+1Jα(s) ds = Rα+1Jα+1(R) (1.11) for α = 1, together with the change of variables s = n√t, to evaluate the integral in (1.10) √t 2πn J1(n√t) dt = 1 R R s2 πn4 J1(s) ds (1.12) (cid:90) n√ 0 1 = πn2 J2(nR), Now we use the property of Bessel functions ([6]) (cid:88) n(cid:54)=0 and substitute this into (1.10) to obtain A(R) = lim δ→0 √R) ψ(δn). 1 πn2 J2(n (1.13) The estimate J2(n√R) = O( n1/2R1/4 ) shows the convergence of the sum in (1.13) without the term ψ(δn), so we can take the limit in (1.13) and obtain (1.5). 1 Theorem 2. Consider the uniformly almost periodic function with mean value zero g(x) = − −5/2 cos n nx − π 4 . (1.14) (cid:16) (cid:17) (cid:88) n(cid:54)=0 √2 π3/2 We have A(R) = g(R1/2)R −1/4 + O(R −3/4) as R → ∞. (1.15) More generally, there exists a sequence of uniformly almost periodic functions g1, g2, . . . with g1 = g such that for any n, A(R) = gj(R1/2)R 1 4− j 2 + O(R − 1 4− n 2 ). (1.16) (cid:114) Proof. We use the well-known asymptotic expression for Bessel functions Jα(x) = −1/2 cos x 2 π + O(x −3/2) as x → ∞. (1.17) n(cid:88) (cid:18) j=1 x − (cid:19) 1 2 απ − π 4 4 Figure 5: g(√t) (cid:114) When α = 2 this is Figure 6: t1/4A(t) − g(√t) (cid:17) x 2 π −1/2 cos J2(x) = − −3/2), and we substitute this into (1.5) with x = n√R to obtain n2 O((n√R) −3/2). A(R) = g(R1/2)R −1/4 + x − + O(x n(cid:54)=0 π 4 1 (1.18) (1.19) It is easy to see that the remainder term in (1.19) is O(R−3/4), so (1.19) yields (1.15). To obtain the more refined asymptotic expression (1.16) we use the known more refined asymptotic expansion for Bessel functions (see [6]). In particular we note that it is possible to obtain explicit series expansions of the functions gj; for example, 15√2 8π3/2 (cid:88) −7/2 sin nx − g2(x) = (1.20) (cid:16) (cid:17) n π 4 . n(cid:54)=0 It is also reasonable to consider the function N ((2πr)2) that counts the number of lattice points inside the ball Br of radius r, the difference D((2πr)2) = N ((2πr)2) − πr2, and the average with respect to the radius variable A(R) = 1 R D((2πr)2) dr. (1.21) (cid:16) (cid:88) (cid:90) R 0 5 0200040006000800010000t−2.0−1.5−1.0−0.50.00.51.01.52.0g(√t)0200040006000800010000t−0.4−0.20.00.20.4t1/4A(t)−g(√t)Figure 7: √t(t1/4A(t) − g(√t)) Figure 8: A(t) Figure 9: √ 1 2 2π A(2πt2) Figure 10: √ 1 2 This is a different average, but a change of variable shows that (cid:32) (cid:90) 2πR2 2π A(2πt2) − A(t) (cid:33) . (1.22) A(R) = 1 2 R 1 2πR2 0 D(t) dt t1/2 Since most of the contribution to the integral occurs for values of t near 2πR2, we see that A(R) has the same asymptotics as √ 1 2 2π A(2πR2). In Figure 1 we show the graph of D(t) and in Figure 2 the graph of t−1/4D(t). This illustrates the rough t1/4 growth rate of D(t). In Figure 3 we show the graph of A(t), and Figure 4 the graph of t1/4A(t). Figure 5 shows the graph of g(√t), which is almost identical to Figure 4 for large t. Figure 6 shows the difference of t1/4A(t) and g(√t), and Figure 7 shows this 6 0200040006000800010000t−3−2−10123√t(t1/4A(t)−g(√t))020406080100t−0.4−0.20.00.20.4A(t)020406080100t−0.4−0.20.00.20.4A(2πt2)/2√2π020406080100t−2.0−1.5−1.0−0.50.00.51.01.52.0A(2πt2)/2√2π−A(t)difference multiplied by t1/2. Figure 8 shows the graph of A(t). Figure 9 2π A(2πt2), which agrees with Figure 8 for large t, and shows the graph of Figure 10 shows the difference. For more data see the website [5]. √ 1 2 (2π)2 2 The general case Consider the general flat 2-dimensional torus, R2/L for some lattice L. The eigenfunctions of the Laplacian (restriction of the standard R2 Laplacian) have the form e2πix·ξ for ξ in the dual lattice L (cid:48), with eigenvalues (2π)2ξ · ξ. (cid:48) we can find an orthonormal (cid:19)2(cid:33) By diagonalizing the quadratic form ξ · ξ on L (cid:18)n · v2 basis v1, v2 in R2 and positive constants a1, a2, such that the eigenvalues are for n ∈ Z2.  . (cid:19)2(cid:33)1/2 (cid:18) n · v2 (cid:19)2 ≤ √t/2π (cid:41) (cid:19)2(cid:33) (cid:32)(cid:18) x · v1 (cid:19)2 (cid:18) x · v2 (cid:32)(cid:18) n · v1 (cid:19)2 n ∈ Z2 : (cid:32)(cid:18) n · v1 (cid:40) x ∈ R2 : (2π)2 In place of disks we consider the family of ellipses Thus the eigenvalue counting function is N (t) = # ≤ t . + a2 + a2 Et = a1 + a2 (2.2) (2.1) a1 a1 Of course N (t) is just the number of lattice points in Et, and the volume of Et is a1a2t for the difference and define the average A(R) by (1.3). The analog of (1.6) is 4π . Again we write D(t) = N (t) − a1a2t 4π a1a2 2π√(a1z·v1)2+(a2z·v2)2 J1( χEt(z) = a1a2t 4π Lemma 3. We have A(R) = n(cid:54)=0 π[(a1n · v1)2 + (a2n · v2)2] a1a2 the series converging uniformly and absolutely. 7 (cid:40) (cid:88) (cid:112) (a1z · v1)2 + (a2z · v2)2√t) z (cid:54)= 0 z = 0 (2.3) (cid:112) (a1n · v1)2 + (a2n · v2)2√R), J2( (2.4) Figure 11: g(√t), a1 = 2, a2 = 1/2 Figure 12: t1/4A(t) − g(√t) Proof. The same proof as for Lemma 1, with (2.3) used in place of (1.6). Theorem 4. The asymptotic expansions (1.15) and (1.16) hold, where now g(x) = − √2 π3/2 a1a2 (cid:0)(a1n · v1)2 + (a2n · v2)2(cid:1)−5/4 (cid:17) (cid:88) (cid:16)(cid:0)(a1n · v1)2 + (a2n · v2)2(cid:1)1/2 n(cid:54)=0 · cos π 4 x − . (2.5) Proof. Same as for Theorem 2, using Lemma 3 in place of Lemma 1. See Figure 11 for g(√t) with a1 = 2 and a2 = 1/2 and Figure 12 for the difference t1/4A(t) − g(√t) for the same family of ellipses. Plots for different families of ellipses are available on the website [5]. 3 The Klein bottle and projective plane If we identify the vertical boundaries of the square directly, and the horizontal boundaries with reflection, we obtain the standard flat Klein bottle KB. In terms of functions defined on the square, we are imposing the boundary conditions u(0, y) = u(1, y) and u(x, 0) = u(1 − x, 1) in order to have a function on KB. We may cover KB by the rectangular torus [0, 1] × [0, 2] with the identities (cid:40) (3.1) u(x + 1, y) = u(x, y) u(1 − x, y + 1) = u(x, y) 8 0200040006000800010000t−4−3−2−101234g(√t)0200040006000800010000t−2.0−1.5−1.0−0.50.00.51.01.52.0t1/4A(t)−g(√t)describing the lifts of functions on KB to R2. The eigenfunctions of the Laplacian on KB lift to eigenfunctions on the rectangular torus, and so are linear combinations of functions of the form e2πi(jx+ k 2 y) with eigenvalue 2 (y+1)) = (−1)ke2πi(−jk+ k (2π)2(j2 + ( k 2 y). Thus there are two families of eigenfunctions 2 )2). Now we observe that e2πi(j(1−x)+ k e2πi k 2 y e2πi(jx+ k for k even (corresponding to j = 0), and 2 y) + (−1)ke2πi(−jx+ k 2 y) for j > 0. (3.2) (3.3) We can therefore see that the eigenvalue function NKB is close to one half the counting function NT1,2 for the [0, 1] × [0, 2] torus. Theorem 5. NKB(t) = 1 2NT1,2(t) ± 1 2. Proof. NT1,2(t) counts all integers j, k such that j2 + ( k (2π)2 . When j (cid:54)= 0 the pair ±j contributes just a single eigenvalue to NKB(t). When √ t j = 0 we count all k such that k ≤ π in NT1,2(t), but just the even values √ t π } = 1 of k in NKB(t), and #{k even : k ≤ √ π } ± 1 2. 2{k : k ≤ 2 )2 ≤ t t It is interesting to compare the Klein bottle with the projective plane (PP) obtained from [0, 1] × [0, 1] by identifying both sets of boundary edges with reflections. Functions on PP lift to R2 with the identities (cid:40) u(1 − x, y + 1) = u(x, y) u(x + 1, 1 − y) = u(x, y) (3.4) and the torus [0, 2] × [0, 2] is a four-fold covering of PP. However, while it is possible to pull back the standard Laplacian to PP, the pairs {(0, 0), (1, 1)} and {(0, 1), (1, 0)} of identified points on PP are singularities (cone points with total angle π) with respect to the otherwise flat metric. Reasoning as in the KB example, we know that eigenfunctions of the 2 y) 2 )2). Imposing the conditions (3.4) leads to Laplacian on PP must be linear combinations of the functions e2πi( j with eigenvalue (2π)2(( j 2)2 + ( k 2 x+ k 9 four families of eigenfunctions: −2πi k constants (corresponding to j = 0 and k = 0) 2 y for k > 0 even (corresponding to j = 0 but k (cid:54)= 0) 2 x for j > 0 even (corresponding to k = 0 but j (cid:54)= 0) 2 y)(cid:17) 2 x− k e2πi(− j 2 x+ k 2 y) + e2πi( j e2ıi k e2πi j e2πi( j 2 y + e 2 x + e 2 x+ k −2πi j 2 y) + e2πi(− j 2 x− k 2 y) + (−1)j+k(cid:16) (3.5) (3.6) (3.7) for j > 0 and k > 0. (3.8) This leads to the identity NPP(t) = 1 4 NT2,2(t) + 1 4 ± 1 2 = 1 4 NT1,1(4t) + 1 4 ± 1 2 . (3.9) Similar results hold for KB and PP constructed from the tori considered in section 2. Related questions in the context of fractal Laplacians with the Sierpinski carpet replacing the square are discussed in [1]. References [1] M. Begu´e, T. Kalloniatis, and R. Strichartz. Harmonic functions and the spectrum of the Laplacian on the Sierpinski carpet. Preprint, 2012. [2] P.M. Bleher. On the distribution of the number of lattice points inside a family of convex ovals. Duke Math J. 67, pages 461 -- 481, 1991. [3] P.M. Bleher. Distribution of the error term in the Weyl asymptotics for the Laplace operator on a two-dimensional torus and related lattice problems. Duke Math J. 70, pages 655 -- 682, 1993. [4] H. Iwaniec and E. Kowalski. Analytic Number Theory. AMS Colloq. Publ. vol 53, 2004. [5] S. Jayakar and R. Strichartz. Average number of lattice points in a disk. http://www.math.cornell.edu/~sujay/lattice, June 2012. [6] N.N Lebedev. Special functions and their applications. Dover Publica- tions, New York, 1965. 10 [7] E.M. Stein and R. Shakarchi. Functional Analysis. Princeton Univ. Press, 2011. 563 Malott Hall, Cornell University, Ithaca, NY 14853, USA Email addresses: [email protected], [email protected] 11
1807.11865	1	1807	2018-07-31T15:28:08	On extensions of symmetric operators	[ "math.FA", "math.SP" ]	We give an explicit description of all minimal self-adjoint extensions of a densely defined, closed symmetric operator in a Hilbert space with deficiency indices $(1, 1)$.	math.FA	math	ON EXTENSIONS OF SYMMETRIC OPERATORS NAMIG J. GULIYEV Abstract. We give an explicit description of all minimal self-adjoint exten- sions of a densely defined, closed symmetric operator in a Hilbert space with deficiency indices (1, 1). There is a widely used linearization technique in the theory of Sturm -- Liouville problems with boundary and/or discontinuity conditions polynomially dependent on the eigenvalue parameter. One considers a Hilbert (or Pontryagin) space of the form L2 ⊕ Ck and constructs a self-adjoint operator in this space such that the eigenvalue problem for this operator and the original boundary value problem become equivalent, in the sense that their eigenvalues coincide, the eigenfunctions of the latter problem are in one-to-one correspondence with the first components of the eigenvectors of the former problem, and so on (see, e.g., [1], [2], [8], and the references therein). Fulton [6, Remark 2.1] attributes this technique to Friedman [5, pp. 205 -- 207]. The purpose of this short paper is to show that a straightforward generalization of this technique gives an explicit description of all minimal self- adjoint extensions of a densely defined, closed symmetric operator with deficiency indices (1, 1) (see below for definitions). Let A be a densely defined, closed symmetric operator in a separable Hilbert space H with deficiency indices (1, 1). Let {C, Γ0, Γ1} be a boundary triplet for A∗. This means that Γ0, Γ1 : D(A∗) → C are two linear mappings such that abstract Green's identity hA∗x, yiH − hx, A∗yiH = Γ1x · Γ0y − Γ0x · Γ1y, x, y ∈ D(A∗) holds and the mapping Γ : D(A∗) → C2, x 7→ (Γ0x, Γ1x) is surjective [4], [12, Chapter 14]. The domain of A then coincides with the kernel of Γ: D(A) = {x ∈ D(A∗) Γ0x = Γ1x = 0}. (1) 8 1 0 2 l u J 1 3 ] . A F h t a m [ 1 v 5 6 8 1 1 . 7 0 8 1 : v i X r a If eA is a self-adjoint operator in a Hilbert space eH ⊃ H such that A ⊂ eA, then eA is called a self-adjoint extension of A. A self-adjoint extension eA is called minimal if no nontrivial subspace of eH ⊖ H is reducing for eA [10, Section 4], or equivalently [12, Lemma 5.17] eH = span(cid:26)(cid:16)eA − λI(cid:17)−1 x(cid:12)(cid:12)(cid:12)(cid:12) x ∈ H, λ ∈ C \ R(cid:27), where span denotes the linear span. 2010 Mathematics Subject Classification. 47A20, 47B25. Key words and phrases. symmetric operator, self-adjoint extension, generalized resolvent. 1 R(λ) = PH(cid:16)eA − λI(cid:17)−1(cid:12)(cid:12)(cid:12)(cid:12)H for some self-adjoint extension eA, where PH is the orthogonal projection onto H. For every generalized resolvent there is a unique (up to unitary equivalence) min- imal self-adjoint extension with this property (see [10, Theorem 8]). On the other hand, there is a one-to-one correspondence between generalized resolvents R(λ) and functions ω(λ) holomorphic on the open upper half-plane C+ with ω(λ) ≤ 1: for every x ∈ H the value y := R(λ)x satisfies the equation and the condition (see [3]). Denoting A∗y − λy = x (ω(λ) − 1) Γ1y − i (ω(λ) + 1) Γ0y = 0 f (λ) := iω(λ) + i 1 − ω(λ) we obtain a one-to-one correspondence between generalized resolvents R(λ) and Herglotz -- Nevanlinna functions f , i.e. functions holomorphic on C+ with Im f (λ) ≥ 0 (cf. [13, Subsection 1.2]). The above condition then becomes Γ1y + f (λ)Γ0y = 0. (2) The "Dirichlet" condition Γ0y = 0 corresponds to f = ∞. We will denote the gener- alized resolvent corresponding to f by Rf (λ). It should be noted that complete pa- rameterizations of all generalized resolvents in terms of Herglotz -- Nevanlinna func- tions were first obtained independently by Naimark [11] and Krein [9]. We are now ready to construct our minimal self-adjoint extension corresponding to a Herglotz -- Nevanlinna function f . This function has a unique representation of the form [12, Appendix F] f (λ) = h0λ + h +Z +∞ −∞ (cid:18) 1 t − λ t 1 + t2(cid:19) dσ(t), − where h0 ≥ 0, h ∈ R, and Z +∞ −∞ dσ(t) 1 + t2 < ∞ (the reader may refer to [7, Appendix A] for some examples of such representations). If h0 > 0 then we consider the Hilbert space eH := H ⊕ L2(R; dσ) ⊕ C with inner product given by 2 NAMIG J. GULIYEV A holomorphic operator-valued function R(λ) on C \ R is called a generalized resolvent of A if for x1(t)y1(t) dσ(t) + x2y2 h0 −∞ hex,eyi eH := hx0, y0iH +Z +∞ ey =  ex =    , x0 x1 x2 y0 y1 y2   ∈ eH, ON EXTENSIONS OF SYMMETRIC OPERATORS 3 and define the operator eAex :=  Γ1x0 + hΓ0x0 +R +∞ A∗x0 tx1(t) − Γ0x0 −∞ (cid:16)x1(t) − t 1+t2 Γ0x0(cid:17) dσ(t)   with and D(eA) := {ex ∈ eH x0 ∈ D(A∗), tx1(t) − Γ0x0 ∈ L2(R; dσ), x2 = −h0Γ0x0}. If h0 = 0 then we set eH := H ⊕ L2(R; dσ), and define eA by tx1(t) − Γ0x0(cid:19) A∗x0 eAex :=(cid:18) D(eA) :=(cid:26)ex ∈ eH(cid:12)(cid:12)(cid:12)(cid:12) x0 ∈ D(A∗), tx1(t) − Γ0x0 ∈ L2(R; dσ), 1 + t2 Γ0x0(cid:19) dσ(t) = 0(cid:27) . The extension eA is canonical (i.e. eH = H) if and only if f is a real constant (or ∞). Theorem. For each Herglotz -- Nevanlinna function f the operator eA defined above is a minimal self-adjoint extension of the operator A with the corresponding gener- alized resolvent Rf (λ). Γ1x0 + hΓ0x0 +Z +∞ −∞ (cid:18)x1(t) − t Proof. We will consider the case h0 > 0; the other case can be proved similarly. To prove the self-adjointness, let ey, ez ∈ eH be such that heAex,eyi eH = hex,ezi eH for all ex ∈ D(eA). Taking into account (1) and setting x1(t) ≡ 0 = x2 we obtain hAx0, y0iH = hx0, z0iH, x0 ∈ D(A). (3) Hence y0 ∈ D(A∗) and z0 = A∗y0. By surjectivity of Γ, there exists x0 ∈ D(A∗) with Γ0x0 = 0 and Γ1x0 = 1. Then for this x0 and x1(t) ≡ 0 the equality (3) gives y2 = −h0Γ0y0. Setting x0 = 0 in (3) we get Z +∞ −∞ x1(t)(ty1(t) − Γ0y0 − z1(t)) dσ(t) = 0 for all x1(t) ∈ L2(R; dσ) with tx1(t) ∈ L2(R; dσ). Since such functions are dense in L2(R; dσ), we obtain ty1(t) − Γ0y0 = z1(t) ∈ L2(R; dσ). Finally choosing x0 ∈ D(A∗) with Γ0x0 = 1, the equality (3) yields z2 = Γ1y0 + hΓ0y0 +Z +∞ −∞ (cid:18)y1(t) − t 1 + t2 Γ0y0(cid:19) dσ(t). check that if Thus ey ∈ D(eA) and eAey =ez. To see that the generalized resolvent corresponding to eA is Rf (λ), it suffices to ey =    =(cid:16)eA − λI(cid:17)−1    y0 y1 y2 x 0 0 4 NAMIG J. GULIYEV then y0 satisfies (2), and this is straightforward. Finally, to check the minimality, we need to verify that the linear span of all ey of this form with all possible values of x ∈ H and λ ∈ C \ R is dense in eH. To this end, let ez ∈ eH be orthogonal to all such ey, i.e. dσ(t) − z2Γ0y0 = 0 z1(t) hez,eyi eH = hz0, y0iH +Z +∞ −∞ Γ0y0 t − λ for all y0 ∈ D(A∗). Surjectivity of Γ implies z0 = 0 and Z +∞ −∞ z1(t) t − λ dσ(t) = z2, λ ∈ C \ R. Now the Stieltjes -- Perron inversion formula [12, Theorem F.2] applied to the regular complex Borel measure z1 dσ yields z1(t) = 0 for dσ-a.e. t and consequently z2 = 0. (cid:3) References [1] C. Bartels, S. Currie, M. Nowaczyk, and B. A. Watson, Sturm -- Liouville problems with trans- fer condition Herglotz dependent on the eigenparameter: Hilbert space formulation, Integral Equations Operator Theory 90 (2018), no. 3, Art. 34, 20 pp. arXiv:1804.07149 [2] J. Behrndt and F. Philipp, Finite rank perturbations in Pontryagin spaces and a Sturm -- Liouville problem with λ-rational boundary conditions, Indefinite inner product spaces, Schur analysis, and differential equations, Birkhauser/Springer, Cham, 2018, pp. 163 -- 189. [3] V. M. Bruk, A certain class of boundary value problems with a spectral parameter in the boundary condition (Russian), Mat. Sb. (N.S.) 100(142) (1976), no. 2, 210 -- 216. [4] V. Derkach, Boundary triplets, Weyl functions, and the Krein formula, Operator theory, Springer, Basel, 2015, pp. 183 -- 218. [5] B. Friedman, Principles and techniques of applied mathematics, John Wiley & Sons, Inc., New York, 1956. [6] C. T. Fulton, Two-point boundary value problems with eigenvalue parameter contained in the boundary conditions, Proc. Roy. Soc. Edinburgh Sect. A 77 (1977), no. 3-4, 293 -- 308. [7] F. Gesztesy and E. Tsekanovskii, On matrix-valued Herglotz functions, Math. Nachr. 218 (2000), 61 -- 138. arXiv:funct-an/9712004 [8] N. J. Guliyev, Essentially isospectral transformations and their applications, preprint. arXiv:1708.07497 [9] M. G. Krein, On Hermitian operators with deficiency indices one (Russian), Dokl. Akad. Nauk SSSR 43 (1944), 339 -- 342. [10] M. A. Naimark, Spectral functions of a symmetric operator (Russian), Izv. Akad. Nauk SSSR. Ser. Mat. 4 (1940), no. 3, 277 -- 318. [11] M. A. Naimark, On spectral functions of a symmetric operator (Russian), Izv. Akad. Nauk SSSR. Ser. Mat. 7 (1943), no. 6, 285 -- 296. [12] K. Schmudgen, Unbounded self-adjoint operators on Hilbert space, Springer, Dordrecht, 2012. [13] A. V. Shtraus, On spectral functions of differential operators (Russian), Izv. Akad. Nauk SSSR. Ser. Mat. 19 (1955), no. 4, 201 -- 220. Institute of Mathematics and Mechanics, Azerbaijan National Academy of Sciences, 9 B. Vahabzadeh str., AZ1141, Baku, Azerbaijan. E-mail address: [email protected]
1308.6227	1	1308	2013-08-28T17:45:08	On the Convergence of Regular Families of Cardinal Interpolators	[ "math.FA" ]	In this note a general way to develop a cardinal interpolant for $l^2$-data on the integer lattice $Z^n$ is shown. Further, a parameter is introduced which allows one to recover the original Paley-Weiner function from which the data came.	math.FA	math	ON THE CONVERGENCE OF REGULAR FAMILIES OF CARDINAL INTERPOLATORS JEFF LEDFORD Abstract. In this note a general way to develop a cardinal in- terpolant for l2-data on the integer lattice Zn is shown. Further, a parameter is introduced which allows one to recover the original Paley-Weiner function from which the data came. Keywords. cardinal interpolation, multiquadric, multivariate in- terpolation, Paley-Wiener functions . A F h t a m [ 1 v 7 2 2 6 . 8 0 3 1 : v i X r a 1. Introduction In this article we prove a result similar to the ones found in [2], [3], [4], [5], and [6]. These results may be thought of as outgrowths of Schoenberg's work on splines (see for instance [7]). The general set up is as follows. Suppose that we have data {f (j)} from some class, we wish to interpolate this data with an interpolant that depends on some parameter. In each case what is investigated is what happens as the parameter approaches a limiting value. The papers mentioned above all fix a particular type of what it called in this article a cardinal interpolator. The goal of this paper is to unify the procedures in those and similar papers in order to provide conditions on an interpolator which allows one to recover Paley-Wiener functions from their samples on the integer lattice by allowing the parameter to approach a limiting value. In doing so, more examples are uncovered. This article is organized as follows. In the next, section several pre- liminary definitions are laid out. We define what is meant by cardinal interpolator in Section 3 and introduce the corresponding fundamental function. Section 4 deals with properties of the fundamental function, while in Section 5 regular families are introduced and contains the proof of the main theorem. The final section consists of three examples of regular families of cardinal interpolators. Finally, our calculations will often involve a positive constant C, whose value depends on its occurrence but not on various parameters unless otherwise stated. 1 2 JEFF LEDFORD 2. Definitions and Basic Facts In this section, we collect several definitions and general facts that will be used in what follows. Our methods require the use of the Fourier transform, for which we adopt the following convention. Definition 1. If f ∈ L1(Rn) we define its Fourier transform, denoted f (ξ) or F (f )(ξ), to be: (1) f (x)e−ix·ξdx, F (f )(ξ) = f (ξ) = (2π)−n/2ZRn where x · ξ =Pn j=1 xjξj. We will encounter examples which are not integrable in this case we adopt the convention of [8], particularly Definition 8.9 on page 103, and include it here for convenience changing the name a little to avoid confusion. Definition 2. For m ∈ Z+, the set of all functions γ ∈ S that satisfy γ(ξ) = O(kξkm) for kξk → 0 will be denoted by Sm. Here S is the Schwartz space, and k · k is the Euclidean norm on Rn. Definition 3. Suppose that φ : Rn → C is continuous and slowly increasing. A measurable function φ ∈ L2 loc(Rn \ {0}) is called the specialized Fourier transform of φ if there exists an integer m ∈ Z+ such that ZRn φ(x)γ(x)dx =ZRn φ(ξ)γ(ξ)dξ is satisfied for all γ ∈ S2m. The least such m is called the order of φ. Before we continue, a few comments are needed. In [8] this is called the generalized Fourier transform. However, there is already a cus- tomary definition of the generalized Fourier transform, which uses the entire Schwartz class as test functions. To avoid this confusion, I have If a function φ(x) ∈ L1(Rn) then its specialized changed the name. and classical Fourier transform coincide, and the order is 0. The same is true for φ(x) ∈ L2(Rn). Finally, the specialized and the generalized (defined in the usual way) Fourier transform coincide on the set S2m. We turn to constraints on the data that we will use. It is not nec- essary to start out assuming that the data comes from any particular function, however it serves our purposes later. Thus throughout the paper we will sample functions exclusively from the Paley-Wiener space P Wπ. REGULAR CARDINAL INTERPOLATORS 3 Definition 4. We define the Paley Wiener space, denoted P Wπ, as (2) P Wπ = {f ∈ L2(Rn) : supp( f ) ⊆ [−π, π]n}. We note that the norm on this space is given by kfkP Wπ = kfkL2(Rn). Lemma 1. Suppose that f ∈ P Wπ, then {f (j)} ∈ l2(Zn). Proof. This is a direct consequence of Plancherel's theorem. (cid:3) 3. Cardinal Interpolators Here we set out hypotheses on a function φ(x) which will be used throughout the rest of the paper. Definition 5. We say that a function φ(x) is a cardinal interpolator if it satisfies the following conditions: (H1) φ(x) is a real valued slowly increasing function on Rn, (H2) φ(ξ) ≥ 0 and φ(ξ) ≥ δ > 0 in [−π, π]n, (H3) φ(ξ) ∈ C n+1(Rn \ {0}), (H4) There exists ǫ > 0 such that if α ≤ n + 1, (H5) For any multi-index α, with α ≤ n + 1, Dα φ(ξ) = O(kξk−(n+ǫ)) as kξk → ∞, α Dαj φ(ξ) Yj=1 [ φ(ξ)]α+1 ∈ L∞([−π, π]n) where αj = α. α Xj=1 For a cardinal interpolant φ(x), we form the corresponding so-called fundamental function of interpolation, denoted Lφ(x), by its Fourier transform. (3) Lφ(ξ) = (2π)−n/2 φ(ξ)"Xj∈Zn φ(ξ − 2πj)#−1 4. Properties of the Fundamental Function The main goal of this section is to establish the integrability Lφ(x). This is done through several lemmas. First we rewrite the transform of the fundamental function as (2π)n/2 Lφ(ξ) = φ(ξ) φ(x) + u(ξ) , 4 JEFF LEDFORD φ(ξ − 2πj). Condition (H4) implies that for all where u(ξ) = Pj6=0 α ≤ n + 1, Dαu is well defined on [−π, π]n. Notice also that (H2) provides the transparent bound Lφ(ξ) ≤ 1. Hypotheses (H4) and (H5) will allow us to control the growth of the derivative of the above expression. Before stating our first lemma we adopt the notation N0 = N ∪ {0}, and follow the convention D0f = f . Lemma 2. If α ∈ N0 is a multi-index, with α ≥ 1, we have (4) Dα Lφ(ξ) =h φ(ξ) + uξi−α−1 Daγ φ(ξ)Yδ /∈B Ca,B Yγ∈B Xa∈Aα,B⊆Sα where Aα =(cid:8)(a1, a2, . . . , aα+1) : aj ∈ Nn 0 , a1 + · · · + aα+1 = α(cid:9), Sα = {1, 2, . . . ,α+1}, and Ca,B are constants. Furthermore, the coefficients corresponding to B = ∅ and B = Sα are 0. Proof. We will induct on α. If α = 1, we use the quotient rule to see that Daδ u(ξ), Dα Lφ(ξ) = (2π)−n/2 u(ξ)Dα φ(ξ) − φ(ξ)Dαu(ξ) ( φ(ξ) + u(ξ))2 . This is the desired form, and since all of the terms on top are mixed products of φ(ξ) and u(ξ) the coefficient condition holds as well. Now if we assume that (4) holds for all α ≤ n, then applying the Leibniz rule and collecting terms yields the result. (cid:3) We have the following result. Corollary 1. If α ≤ n + 1, then Dα Lφ ∈ L∞([−π, π]n). Proof. Combine (H5) with (4). (cid:3) In order to determine the growth of Lφ(ξ) away from the origin, we examine the function P (ξ) = 1 φ(ξ) + u(ξ) . Lemma 3. For ξ ∈ [−π, π]n and α a multi-index, we have (5) DαP (ξ) =h φ(ξ) + u(ξ)i−α−1 Xa∈Aα where Aα =(cid:8)(a1, a2, . . . , aα) : aj ∈ Nn 0 , a1 + · · · + aα = α(cid:9). Yj=1 α Ca Daj ( φ(ξ) + u(ξ)), REGULAR CARDINAL INTERPOLATORS 5 Proof. We again induct on α. If α = 1, we use the quotient rule to obtain DαP (ξ) = −Dα( φ(ξ) + u(ξ)) ( φ(ξ) + u(ξ))2 . This is the desired result. Suppose now that (5) holds for all α ≤ n, then applying the quotient rule and collecting terms yields the result. (cid:3) Corollary 2. For α ≤ n + 1, DαP (ξ) ∈ L∞(Rn). Proof. Since P (ξ) is periodic, we need only combine (H5) and (5). (cid:3) Proposition 1. If φ is a cardinal interpolator and α ≤ n + 1, then Dα Lφ ∈ L1(Rn). Proof. Since Lφ(ξ) = φ(ξ)P (ξ), we see that near the origin Corollary 1 controls the growth, while away from the origin Corollary 2 and (H4) imply the result. (cid:3) Corollary 3. If φ is a cardinal interpolator, then Lφ(x) is continuous and there exists a constant C > 0 such that (6) Proof. Continuity follows from Proposition 1 by letting α = 0. The bound (6), follows as well since (cid:12)(cid:12)(1 + kxkn+1)Lφ(x)(cid:12)(cid:12) ≤ C. (cid:12)(cid:12)(1 + kxkn+1)Lφ(x)(cid:12)(cid:12) ≤ Xα≤n+1(cid:13)(cid:13)(cid:13) Dα Lφ(ξ)(cid:13)(cid:13)(cid:13)L1(Rn) . (cid:3) We now show why Lφ(x) is called a fundamental function of inter- polation. Lemma 4. Lφ(x) satisfies Lφ(k) = δ0,k for k ∈ Zn. Proof. The proof is an easy calculation invoking the inversion theorem: Lφ(ξ)eik·ξdξ Lφ(k) = = = = 1 1 1 (2π)n/2 ZRn (2π)n/2 Xm∈ZnZ[−π,π]n (2π)n/2 Z[−π,π]n (2π)n Z[−π,π]n 1 Lφ(ξ − 2πm)eik·ξdξ Lφ(ξ − 2πm)dξ eik·ξ Xm∈Zn eik·ξdξ = δ0,k. 6 JEFF LEDFORD We turn now to constructing an interpolant consisting of translates of Lφ(x). Claim 1. If {f (j)}j∈Zn ∈ l∞(Zn) and we define Iφf (x) pointwise as follows: (cid:3) Iφf (x) = Xj∈Zn f (j)Lφ(x − j). where φ(ξ) is a cardinal interpolator, then Iφf (x) is well defined and continuous. Proof. As a result of (6) we have the following: Iφf (x) ≤ Ck{f (j)}kl∞(Zn) Xj∈Zn 1 1 + kx − jkn+1 < ∞. Continuity follows from the Weierstrass M test and the continuity of Lφ(x). Notice this bound holds for all x ∈ Rn because the last sum is periodic. (cid:3) We are now in position to prove our first theorem. Theorem 1. Let 1 ≤ p ≤ ∞. If {f (j)} ∈ lp(Zn) and Iφf (x) is defined by (7), then Iφf (x) ∈ Lp(Rn). Proof. We will use the Riesz-Thorin interpolation theorem. Consider the linear map given by {f (j)} 7→ Iφf (x). From the above claim, we can see that Iφ maps l∞(Zn) to C(Rn). In order to apply the theorem, we must check the p = 1 case and the p = ∞ case. For p = 1 we have: (7) (8) f (j)Lφ(x − j)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)L1(Rn) kIφf (x)kL1(Rn) =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xj∈Zn ≤k{f (j)}kl1(Zn)kLφ(x − j)kL1(Rn) 1 + kxkn+1(cid:13)(cid:13)(cid:13)(cid:13)L1(Rn)k{f (j)}kl1(Z) ≤C(cid:13)(cid:13)(cid:13)(cid:13) ≤C1k{f (j)}kl1(Zn) 1 Here we've used Minkowski's inequality and Corollary 3. The p = ∞ case is the content of (8). Thus the Riesz-Thorin interpolation theorem guarantees a constant Cp independent of {f (j)} such that kIφf (x)kLp(Rn) ≤ Cpk{f (j)}klp(Zn). (cid:3) REGULAR CARDINAL INTERPOLATORS 7 5. Regular Families of Cardinal Interpolators The goal of this section is to introduce regularity conditions on a family of cardinal interpolators {φα} so that we may get recovery re- sults similar to those found in [2], [3], [4], [5], and [6]. Definition 6. We call a family of functions {φα(ξ)}α∈A a regular fam- ily of cardinal interpolators if for each α ∈ A ⊂ (0,∞), φα(x) is a cardinal interpolator and in addition to this, we have: φα(ξ + 2πj) (R1) For j ∈ Zn \ {0}, ξ ∈ [−π, π]n, define Mj,α(ξ) = , φα(ξ) Mj,α(ξ) = 0 for almost every ξ ∈ (R2) There exists Mj ∈ l1(Zn \ {0}), independent of α, such that for then for j ∈ Zn \ {0}, [−π, π]n. all j ∈ Zn \ {0}, Mj,α(ξ) ≤ Mj. lim α→∞ Depending on the interpolator φ(x), the parameter α may be contin- uous or discrete. We will encounter examples of both in the following section. Proposition 2. If {φα} is a regular regular family of cardinal inter- polators, then Lφα(ξ) → (2π)−n/2χ[−π,π]n(ξ) almost everywhere. Proof. For ξ ∈ [−π, π]n we have: φα(ξ)(cid:27)−1 If we can show Lφα(ξ) = (2π)−n/2(cid:26)1 + uα(ξ) φα(ξ) → 0 we will be done with this part. Conditions (R1) and (R2) allow us to apply the dominated convergence theorem for series, which gives us the result. Now we have only to check ξ /∈ [−π, π]n. We may write ξ = 2πj + r, where j ∈ Zn \ {0} and r ∈ [−π, π]n. Then we have: uα(ξ) . Lφα(ξ) = (2π)−n/2 φα(2πj + r) φα(r) (cid:26)1 + uα(r) φα(r)(cid:27)−1 ≤ (2π)−n/2Mj,α(ξ). The last inequality comes from (H2), thus (R1) provides the desired result. (cid:3) Proposition 3. If {φα} is regular a regular family of cardinal inter- Lφα(ξ + 2πj) = 0 almost everywhere and in polators, then lim α→∞Xj6=0 L1([−π, π]n). 8 JEFF LEDFORD Proof. We note that if j ∈ Zn \ {0}, then Lφα(ξ + 2πj) ≤ Mj,α(ξ) by (R1), thus (R2) allows us to use the dominated convergence theorem for series which gives the result. (cid:3) These propositions allow us to prove the main theorem. Theorem 2. Suppose that { φα(ξ)}α∈A is a regular family of cardinal interpolators and let f (x) ∈ P Wπ. Then we have the following limits: (a) lim (b) lim α→∞(cid:13)(cid:13)(cid:13)(cid:13)Xj∈Zn α→∞(cid:12)(cid:12)(cid:12)(cid:12)Xj∈Zn f (j)Lφα(x − j) − f (x)(cid:13)(cid:13)(cid:13)(cid:13)L2(Rn) f (j)Lφα(x − j) − f (x)(cid:12)(cid:12)(cid:12)(cid:12) = 0, = 0 uniformly in Rn. Proof. (a) Since f ∈ P Wπ, we have that {f (j)}j∈Zn ∈ l2(Zn). This, in turn, implies that Xj∈Zn f (j)Lφα(x − j) ∈ L2(Rn). So we may use Plancherel's theorem to estimate the norm. 2 L2(Rn) 2 f (j)Lφα(x − j) − f (x)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xj∈Zn f (j)e−ij·ξ − f (ξ)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Lφα(ξ)Xj∈Zn =Z[−π,π]n(cid:12)(cid:12)(cid:12) f (ξ)(cid:16)(2π)n/2 Lφα(ξ) − 1(cid:17)(cid:12)(cid:12)(cid:12) f (j)e−ij·ξ(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) +ZRn\[−π,π]n(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Lφα(ξ)Xj∈Zn 2 dξ L2(Rn) 2 dξ We have used that f ∈ P Wπ and the well known L2(Rn) equality (2π)n/2 f (ξ) = Xj∈Zn f (j)e−ij·ξ for ξ ∈ [−π, π]n. Proposition 2 shows that (2π)n/2 Lφα(ξ) → χ[−π,π]n(ξ) almost everywhere, thus the first integral tends to 0 by the dominated convergence theorem, since we have the elementary estimate (2π)n/2 Lφα(ξ) − 1 ≤ (2π)n/2 + 1. We estimate the second integral. REGULAR CARDINAL INTERPOLATORS 9 ZRn\[−π,π]n Lφα(ξ)Xj∈Zn f (j)e−ij·ξ2dξ = Xm∈Zn\{0}Z[−π,π]n Lφα(ξ + 2πm)Xj∈Zn =(2π)n Xj∈Zn\{0}Z[−π,π]n Lφα(ξ + 2πj)2 f (ξ)2dξ ≤(2π)nZ[−π,π]n Xj∈Zn\{0} Mj,α(ξ)2 f (ξ)2dξ f (j)e−ij·(ξ+2πm)2dξ This last term tends to 0 by the dominated convergence theorem as in Proposition 3, noting that l1(Zn \ {0}) ⊂ l2(Zn \ {0}). Schwarz inequality. Part (b) is proved in a similar manner, with help from the Cauchy- f (j)Lφα(x − j) − f (x)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Xj∈Zn f (j)e−ijξ − f (ξ)! eix·ξdξ(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) =(2π)−(n/2)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ZRn Lφα(ξ)Xj∈Zn ≤Z[−π,π]n(cid:12)(cid:12)(cid:12) f (ξ)(cid:12)(cid:12)(cid:12) Lφα(ξ) − (2π)−n/2(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) +Xj6=0Z[−π,π]n(cid:12)(cid:12)(cid:12) f (ξ)(cid:12)(cid:12)(cid:12) Lφα(ξ + 2πj)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Lφα(ξ) − (2π)−n/2(cid:13)(cid:13)(cid:13)L2([−π,π]n) ≤kfkP Wπ(cid:13)(cid:13)(cid:13) Mj,α(ξ)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)L2([−π,π]n) +kfkP Wπ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xj6=0 dξ dξ From Proposition 2 above we can see that the first term tends to 0. By (R2) we have that the sum in the second term is bounded above by kMjk2 l1(Zn\{0}), so Proposition 3 and the dominated convergence theo- rem finish the proof. (cid:3) 6. Examples of Regular Families of Interpolators In this section, we exhibit three families of regular cardinal interpo- lators. Our first example will be that of polyharmonic cardinal splines. 10 JEFF LEDFORD The result that we attain is most similar to the one found in [3]. Our second example will be a family of Gaussians, which has also been studied previously, see for instance [4], [5]. Finally, we will consider a family of multiquadrics and extend a result which is found in [6], as well a provide what appears to be a novel example. For each example we work out the one dimensional case, followed by the multivariate case often suppressing the straightforward yet tedious details involved in checking (H5). 6.1. Polyharmonic Cardinal Splines. Our first example was ex- plored in [3], and motivated this treatment. We begin by considering the one dimensional version of the example found there. A function or distribution f is called k-harmonic, where k ∈ {1, 2, 3, . . .}, if it satisfies (9) ∆kf = 0 on Rn, where ∆ is the Laplacian operator and ∆k(f ) = ∆(∆k−1f ). A function which satisfies (9) with k ≥ 1 is called polyharmonic. The fundamental solution of (9) is given by a constant multiple of φk(x) = x2k−1, whose Fourier transform is given by φk(ξ) = (2π)−1/2ξ−2k. which satisfies A polyharmonic cardinal spline is a function or distribution on R (i) f ∈ C 2k−2(Rn), (ii) ∆kf = 0 on Rn \ Zn. The fundamental functions Lk(x) use φk(ξ) = (2π)−1/2ξ−2k. Several properties of the fundamental function may be found in [3], including the fact that it is a k-harmonic spline. We show that {ξ−2k}∞ k=1 is a regular family of cardinal interpolators. Checking (H1)-(H5) is an elementary exercise in calculus and will be omitted. We check (R1) and (R2). For j 6= 0 and ξ < π we have Mj,k(ξ) = ξ2k ξ + 2πj2k ≤ (2j − 1)−2k. These terms tend to 0, hence (R1) is satisfied. For (R2) we note that (2j − 1)−2k ≤ (2j − 1)−2. Thus, if n = 1, polyharminic cardinal splines form a regular family of cardinal interpolators. For n > 1, we have φk(ξ) = kξk−2k, where ξ ∈ Rn, in this case we must take k ≥ n+1. (H1)-(H4) are straightforward, so we will check (H5), (R1), and (R2). We need to find the derivatives REGULAR CARDINAL INTERPOLATORS 11 in question. To this end, we note that for j = 1, . . . , n we have Repeating this argument for higher derivatives we arrive at . d dξj (cid:0)kξk−2k(cid:1) = −2kkξk−2k−1 ξj kξk Dβ(cid:0)kξk−2k(cid:1) = kξk−2k−βΩ(ξ) where β is a multi-index such that β ≤ n + 1, and Ω(λξ) = Ω(ξ). This information allows us to check (H5) easily. We move on to (R1) and (R2). Mimicking the one dimension case we see that Mj,k(ξ) =(cid:18) kξ + 2πjk(cid:19)2k kξk . If ξ ∈ (−π, π)n and j 6= 0, then we have kξk < kξ + 2πjk hence these terms go to 0. Thus we have shown that (R1) holds. For (R2) we must be a little more careful. In this case, we see that the terms grow like kjk−2k. Thus we may find a constant and C such that Mj,k(ξ) ≤ Ckjk−2k as long as j 6= 0. Recalling that 2k ≥ n + 1 we have that Mj,k(ξ) ≤ Ckjk−n−1. 6.2. Gaussians. Our second example is the Gaussian. This example was studied in [4, 5], although we change the notation somewhat. We begin with φα(x) = e−x2/(4α), where x ∈ R, thus omitting constants we have φα(ξ) = e−αξ2 . We take α ≥ 1 as a convenience, the arguments go through with α ≥ α0 > 0. Again, checking conditions (H1)-(H5), is a straightforward calculus exercise, whose details are omitted. We check (R1) and (R2). For ξ < π and j 6= 0, we have Mj,α(ξ) = eαξ2−α(ξ+2πj)2 . We can see that these terms go to 0 since the exponent is negative. Since this is true for all j 6= 0, (R1) holds. As for (R2) we note that if α ≥ 1 we have eαπ2−α(ξ+2πj)2 = e−απ2(4j)(j−1) ≤ e−4π2j(j−1). is a regular family of cardinal interpolators in the Hencene−x2/(4α)oα≥1 univariate case. In the multivariate case, we have φα(ξ) = e−αkξk2 , and again consider α ∈ [1,∞) This function is smooth so all of the relevant properties are easily checked. (D1) and (D2) are checked as above by replacing the absolute values with norms. Hence, ne−x2/(4α)oα≥1 is a regular family of cardinal interpolators in the multivariate case as well. 12 JEFF LEDFORD 6.3. Multiquadrics. Our next example is a family of 'multiquadrics', meaning φ(x) = (c2 +kxk2)α. In this example we have a choice between which parameter to limit over and which to keep fixed. We will work both cases. In the univariate case, allowing c to vary is the subject of [6]. We begin by finding the specialized Fourier transform of φ(x) = (c2 + kxk2)α. We have 21+α 2 K n 2 +α(ckξk), ξ 6= 0, φ(ξ) = Γ(−α)(cid:18)kξk where Kβ(r) = e−r cosh teβtdt, c (cid:19)−α− n 2ZR 1 for r > 0 and β ∈ R. The details of this may be found in Theorem 8.15 of [8]. We must omit the case when α ∈ {0, 1, 2, . . .} because in this case we would get a measure in the transform space and we would like a function. This is why we use the specialized Fourier transform and not the more standard generalization. We first consider the univariate case and list some properties of Kβ(r) which will be useful is our analysis. These properties may be found on page 52 of [8]. The first is that Kβ(r) is decreasing on (0,∞). We find the following asymptotic expansions in [1]: (10) e−reβ2/(2r) 1 √r Kβ(r) ∼ Kβ(r) ∼ r−β (11) Here f ∼ g means that f = (C + o(1))g, where C is a constant in- dependent of the limiting parameter. We also have the differentiation formula (r → ∞), (r → 0). (12) d dz(cid:2)zβKβ(z)(cid:3) = −zβKβ−1(z). 2 and r > 0 And because we will need it later we have the following bounds for β ≥ 1 (13) √2πr−1/2e−re−β2/(2r) r−1/2e−r ≤ Kβ(r) ≤ r π 2 All of this information may be found in [1] and Chapter 5 of [8], and has been restated here for convenience and to aide in the calculations that follow. We will consider two different families of multiquadrics. In the first example we fix a positive number c and take {αj}∞ 2,∞), αj = ∞. We show such that dist({αj},{1, 2, 3, . . .}) > 0 and lim {φαj (x)}∞ j=1 is a regular family of cardinal interpolators. We begin j=1 ⊂ [ 1 j→∞ REGULAR CARDINAL INTERPOLATORS 13 in the univariate case checking (H1)-(H5). (H1)-(H4) follow from the definition, the differentiation formula, and the asymptotic estimates. (H5) also follows from the asymptotic estimates, but we will show these calculations. First, we find the derivatives in question omitting the associated constants since they have no bearing on the overall result. αj (ξ) = − ξ−αj− 1 φ(1) αj (ξ) =ξ−αj− 1 φ(2) 2 Kαj + 3 2 (cξ) (cξ) Using the asymptotic expansion, we find the following limits. (cξ) + ξ−αj− 3 2 Kαj + 3 2 Kαj− 5 2 2 1 φαj (ξ) φ(1) αj (ξ) φ2 αj (ξ) φ(2) αj (ξ) φ2 αj (ξ) = 0 2 2 = ξαj+ 1 Kαj + 1 = ξ2αj+1 C + o(1) (cξ) = ξαj+ 1 (cξ) = ξ2αj−1(cid:26) C3 + o(1) Kαj + 3 K 2 αj + 1 2 2 2 [C1 + o(1)]2 − = ξ2αj C2 + o(1) [C1 + o(1)]2 = 0 [C1 + o(1)]2(cid:27) C2 + o(1) (ξ → 0) (ξ → 0) (ξ → 0) 1 2 We can see that in the last expression we must take αj ≥ in order for the limit to be bounded. Now we turn to (R1). Let ξ < π and i 6= 0, then Mi,αj (ξ) ≤ παj+ 1 2 Kαj + 1 2 (cπ) Kαj + 1 2 (cξ + 2πi) ξ + 2πiαj+ 1 2 ≤ 1 . (2i − 1)αj+ 1 2 j=1 ⊂ [ 1 We have used the fact that Kβ(r) is decreasing. This estimate allows us to prove (D2) as well, since the condition for every α may be replaced with for every α large enough. In this case, large enough means αj > 1 2. Thus we have that {(c2 + x2)αj} is a well behaved family for fixed c > 0 and{αj}∞ 2,∞), such that dist({αj},{1, 2, 3, . . .}) > 0 and αj = ∞. Before moving on, we make note of the specific example lim j→∞ αk = 2k−1 for k ∈ {1, 2, 3, . . .}. This is the family of odd powers }. This family corresponds to 'smoothed out' polyharmonic splines of order 2k. In light of the results in [3], one may well expect that our family of multiquadrics shares a similar convergence property. This is indeed the case as our work above shows. of the multiquadric {(cid:16)pc2 + x2(cid:17)2k−1 2 We now explore what happens if we fix α and allow c to vary. Again we will need to bound the parameter away from the origin so we take c ≥ 1. Let φc(x) = (c2 + x2)α for some fixed α ∈ (−∞,− 3 2,∞) \ 2] ∪ [ 1 14 JEFF LEDFORD {1, 2, 3, . . .}, then, omitting constants because they will divide out in the fundamental function, we have φc(ξ) = ξ−α− 1 2 +α(cξ). The bounds on α are a result of the restrictions imposed by the derivatives. Since φc(0) is a real positive number if α < − 1 2, (H5) places restrictions on the growth of φc(ξ) near the origin. To satisfy this condition, we need to increase the rate of decay of φc(x). Most of the work to check that this is a regular family of cardinal interpolators has already been done. The same arguments used above can be recycled, however we must check (R1) and (R2) separately. We check (R1) presently. 2 K 1 Mj,c(ξ) ≤ 2 2 πα+ 1 2 Kα+ 1 (cξ + 2πj) Kα+ 1 Kα+ 1 (cπ) (cξ + 2πj) Kα+ 1 ξ + 2πjα+ 1 ≤ (2j − 1)−(α+ 1 ≤ (2j − 1)−(α+1) ecπe−cξ−2πje−(α+ 1 ≤ (2j − 1)−(α+1) ecπe−cξ−2πj (cπ) 2 ) 2 2 2 2 )2/(2cξ+2πj) We can see, for j 6= 0 and ξ < π, that lim Mj,c(ξ) = 0. This calculation also shows that (R2) is satisfied since each term there is bounded by (2j − 1)−(α+1)e−2π(j−1). Thus we have shown that {(c2 + x2)α}c≥1 is a well behaved family of interpolators for a fixed α ∈ (−∞,− 3 2 is examined in [6]. 2 ,∞)\{1, 2, 3, . . .}. The case where α = 1 2]∪[ 1 c→∞ We now visit the n-dimensional multiquadric φ(x) = (c2 + kxk2)α. We will again consider both the case when α is the parameter and c is fixed and the case when c is the parameter and α is fixed. We begin by recalling the specialized n-dimensional Fourier transform of the multiquadric, φ(ξ) = 21+α Γ(−α)(cid:18)kξk c (cid:19)−α− n 2 K n 2 +α(ckξk), ξ 6= 0. Conditions (H1)-(H4) follow as before. We need to calculate the deriva- tives in order to check (H5). To this end, we may use the differentiation formula (12) together with repeated use of the product rule and the as- ymptotic estimate (11) to see that near the origin we have the following asymptotic estimate (14) Dβ φ(ξ) ∼ kξk−α− n 2 −α+ n 2 +β, kξk → 0. REGULAR CARDINAL INTERPOLATORS 15 Recall that f ∼ g means that f = g(C + o(1)) for an appropriate constant C. This calculation holds no matter which parameter we are considering fixed. Thus we can begin by fixing c > 0 and checking (H5), (R1) and (R2) for φα(x). For (H5) we note that α > 0, hence the asymptotic estimate (14) becomes Dβ φ(ξ) ∼ kξk−2α−n−β, kξk → 0. Using this estimate, checking (H5) is now a routine matter. We note that we must have α ≥ 1 Mj,α(ξ) =(cid:18) 2. To see (R1), note that kξk kξ + 2πjk(cid:19)−α− n 2 (ckξ + 2πjk) Kα+ n 2 (ckξk) 2 Kα+ n , and if ξ ∈ (−π, π)n we have lim α→∞ α→∞(cid:18) Mj,α(ξ) ≤ lim kξk kξ + 2πjk(cid:19)−α− n 2 . These terms tend to 0 as α increases. As in the one variable case, we must be careful about how we choose the parameter α, in particular we take α ∈ {αj} ⊂ [ 1 2,∞) such that dist({αj},{0, 1, 2, 3, . . .}) > 0 αj = ∞. For the bounding sequence in (R2) we take α = n+1 2 . and lim j→∞ These terms grow like kjk−n−1, hence we may find a constant C such that Mj,α(ξ) ≤ Ckjk−n−1. This bound holds for all αj ≥ n+1 We now will fix α and let c → ∞. That is, we consider φc(x) = (c2 + kxk2)α. (H1)-(H4) hold as before. We take advantage of (14) to find that if α ∈ (−∞,− 2n+1 2 ,∞) \ {1, 2, 3, . . .} then (H5) holds. As for (R1) and (R2), we must be a bit more careful. We have 2 , hence (R2) holds. ] ∪ [ 1 2 and we can see from (13) that Mj,c(ξ) =(cid:18) Mj,c(ξ) ≤ C(cid:18) 2 Kα+ n kξk kξ + 2πjk(cid:19)−α− n kξ + 2πjk(cid:19)−α− n−1 kξk 2 2 (ckξ + 2πjk) Kα+ n 2 (ckξk) , ec(kξk−kξ+2πjk). Here C is a positive constant independent of c whose particular value is unimportant. If ξ ∈ (−π, π)n, then Mj,c(ξ) → 0 because the exponent in the exponential is negative. To check (R2), note that these terms decay exponentially hence we can find a bounding sequence which is in l1(Zn \ {0}). 16 JEFF LEDFORD References [1] M. Abramowitz and I. Stegun, eds., Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables, Dover, New York, New York, 1965 pp. 375-378. [2] B.J.C. Baxter, The asymptotic cardinal function of the multiquadratic ϕ(r) = [3] W. Madych, S. Nelson, Polyharmonic Cardinal Splines, J. Approximation The- (r2 + c2)1/2 as c → ∞, Comput. Math. Appl. 24 (1992), no. 12, 1-6. ory 60 (1990), 141-156. [4] S. Riemenschneider and N. Sivakumar, Gaussian radial-basis functions: cardi- nal interpolation of lp and power-growth data. Radial basis functions and their applications Adv. Comput. Math. 11, (1999), 229-251. [5] S. Riemenschneider and N. Sivakumar, On cardinal interpolation by Gaussian radial-basis functions: properties of fundamental functions and estimates for Lebesgue constants J. Anal. Math. 79, (1999), 33-61. [6] S. Riemenschneider and N. Sivakumar, On the Cardinal-Interpolation Operator Associated with the One-Dimensional Multiquadric East J. Approx. 7, (2001), no. 4, 485-514. [7] I.J. Schoenberg, Cardinal Spline Interpolation SIAM, Philadelphia, PA, 1973. [8] H. Wendland, Scattered Data Approximation, Cambridge University Press, Cambridge, UK, 2005.
1312.0986	2	1312	2016-12-12T11:09:33	The short-time Fourier transform of distributions of exponential type and Tauberian theorems for shift-asymptotics	[ "math.FA" ]	We study the short-time Fourier transform on the space $\mathcal{K}_{1}'(\mathbb{R}^n)$ of distributions of exponential type. We give characterizations of $\mathcal{K}_{1}'(\mathbb{R}^n)$ and some of its subspaces in terms of modulation spaces. We also obtain various Tauberian theorems for the short-time Fourier transform.	math.FA	math	THE SHORT-TIME FOURIER TRANSFORM OF DISTRIBUTIONS OF EXPONENTIAL TYPE AND TAUBERIAN THEOREMS FOR SHIFT-ASYMPTOTICS SANJA KOSTADINOVA, STEVAN PILIPOVI ´C, KATERINA SANEVA, AND JASSON VINDAS Abstract. We study the short-time Fourier transform on the space K′ 1(Rn) of dis- tributions of exponential type. We give characterizations of K′ 1(Rn) and some of its subspaces in terms of modulation spaces. We also obtain various Tauberian theorems for the short-time Fourier transform. 1. Introduction The short-time Fourier transform (STFT) is a very effective device in the study of function spaces. The investigation of major test function spaces and their duals through time-frequency representations has attracted much attention. For example, the Schwartz class S(Rn) and the space of tempered distributions S ′(Rn) were studied in [10] (cf. [9]). Characterizations of Gelfand-Shilov spaces and ultradistribution spaces by means of the short-time Fourier transform and modulation spaces are also known [11, 17, 26] (cf. [4, 5]). The purpose of this paper is two folded. On the one hand we study the short-time Fourier transform in the context of the space K′ 1(Rn) of distributions of exponential type, the dual of the space of exponentially rapidly decreasing smooth functions K1(Rn) (see Section 2 for the definition of all spaces employed in this article). We will obtain various characterizations of K′ 1(Rn) and related spaces via the short-time Fourier trans- form. The space K′ 1(Rn) was introduced by Silva [24] and Hasumi [12] in connection with the so-called space of Silva tempered ultradistributions U ′(Cn). Let us mention 1(Rn) and U ′(Rn) were also studied by Morimoto through the theory of ultra- that K′ hyperfunctions [16] (cf. [18]). We refer to [7, 14, 25, 30] for some applications of the Silva spaces. Our second goal is to present a new kind of Tauberian theorems. In such theorems the exponential asymptotics of functions and distributions can be obtained from those of the short-time Fourier transform. Let us state a sample of our results. In the next statement L stands for a locally bounded Karamata slowly varying function 2010 Mathematics Subject Classification. Primary 81S30, 40E05; Secondary 26A12, 41A60, 46F05, 46F12. Key words and phrases. short-time Fourier transform; Tauberian theorems; modulation spaces; 1 and K1; distributions of exponential type; Silva tempered ultradistributions (tempered ultra- K′ hyperfunctions); S-asymptotics; regularly varying functions. Research supported by the project 174024 of the Ministry of Education and Sciences of Serbia. S. Pilipovi´c and J. Vindas also acknowledge support from Ghent University, through the BOF-grant number 01T00513. 1 2 S. KOSTADINOVA, S. PILIPOVI ´C, K. SANEVA, AND J. VINDAS [2, 15], namely, a positive function that is asymptotically self-similar in the sense: lim x→∞ L(ax) L(x) = 1, ∀a > 0. Theorem 1.1. Let f be a positive non-decreasing function on [0, ∞) and let ψ be a positive function such that ψ′′ ∈ L1 −∞(ψ(t) + ψ′(t) + ψ′′(t))eβt+εtdt < ∞, where β ≥ 0 and ε > 0. Suppose that the limits loc(R) andR ∞ (1.1) lim x→∞ f (t)ψ(t − x)e−2πiξt dt = J(ξ) e2πiξx eβxL(ex)Z ∞ 0 exist for every ξ ∈ R, then f (x) = J(0) . (1.2) lim x→∞ eβxL(ex) −∞ ψ(t)eβtdt R ∞ the requirements over ψ can be relaxed toR ∞ Furthermore, if L satisfies L(xy) ≤ AL(x)L(y) for all x, y > 0 and some constant A, −∞(ψ(t) + ψ′(t) + ψ′′(t))L(et)eβtdt < ∞. It turns out that Theorem 1.1 can be deduced from a more general type of Tauberian theorems. In Section 6 we shall give precise descriptions of the S-asymptotic properties [21] of a distribution in terms of the asymptotic behavior of its short-time Fourier transform (S-asymptotics stands for shift-asymptotics). The notion of S-asymptotics measures the asymptotic behavior of the translates of a distribution T−hf with respect the parameter h and it is closely related to the extension of Wiener's Tauberian ideas [15, 29] to the context of Schwartz distributions [20]. We would like to point out that there is an extensive literature about Tauberian theorems for Schwartz distributions, see, e.g., the monographs [21, 28] and references therein. We also mention the article [22], where Tauberian theorems for the short-time Fourier transform were also studied, though with a different approach. The plan of this article is as follows. In Section 3 we shall present continuity theorems for the STFT and its adjoint on the test function space K1(Rn) and the topological 1(Rn). We also introduce in this paper the space B′ tensor product K1(Rn)b⊗U(Cn), where U(Cn) is the space of entire rapidly decreasing functions in any horizontal band of Cn. We then use such continuity results to develop a framework for the STFT on K′ ω(Rn) of ω-bounded distributions and its subspace B′ ω(Rn) with respect to an exponentially moderate weight ω; when ω = 1, these spaces coincide with the well-known Schwartz spaces [23, p. 200] of bounded distributions B′(Rn) and B′(Rn), which are of great importance in the study of convolution and growth properties of distributions. Notice that the distribution space B′(Rn) also plays an important role in Tauberian theory; see, for instance, Beurling's theorem [6, p. 230] and the distributional Wiener Tauberian theorem from [20]. The spaces B′ ω(Rn) will be characterized in Section 4 in terms of the short-time Fourier transform and also in terms properties of the set of translates of their elements. Section 5 is devoted to the characterization of K′ 1(Rn) and related spaces via modulation spaces. The conclusive Section 6 deals with Tauberian theorems, where in particular we give a proof of Theorem 1.1. Our Tauberian ω(Rn) and B′ THE STFT OF DISTRIBUTIONS OF EXPONENTIAL TYPE 3 hypotheses are actually in terms of membership to suitable modulation spaces, this allows us to reinterpret the S-asymptotics in the weak∗ topology of modulation spaces. 2. Preliminaries 2.1. Notation. We use the constants in the Fourier transform as (2.1) e−2πix·ξϕ(x)dx. F (ϕ)(ξ) = bϕ(ξ) =ZRn The translation and modulation operators are defined by Txf ( · ) = f ( · − x) and Mξf ( · ) = e2πiξ · f ( · ), x, ξ ∈ Rn. The operators MξTx and TxMξ are called time- frequency shifts and we have MξTx = e2πix·ξTxMξ. The notation hf, ϕi means dual pairing whereas (f, ϕ)L2 stands for the L2 inner product. All dual spaces in this article are equipped with the strong dual topology. We denote by f the function (or distribution) f (t) = f (−t). 2.2. The STFT. The short-time Fourier transform (STFT) of a function f ∈ L2(Rn) with respect to a window function ψ ∈ L2(Rn) is defined as (2.2) f (t)ψ(t − x)e−2πiξ·t dt, x, ξ ∈ Rn. There holds kVψf k2 = kf k2 kψk2. The adjoint of Vψ is given by the mapping Vψf (x, ξ) = hf, MξTxψi =ZRn ψ F (t) =ZZR2n V ∗ F (x, ξ)ψ(t − x)e2πiξ·tdxdξ, interpreted as an L2(Rn)-valued weak integral. If ψ 6= 0 and γ ∈ L2(Rn) is a synthesis window for ψ, namely, (γ, ψ)L2 6= 0, then for any f ∈ L2(Rn), (2.3) f = Vψf (x, ξ)MξTxγdξdx. 1 (γ, ψ)L2ZZR2n Whenever the dual pairing in (2.2) is well-defined, the definition of Vψf can be generalized for f in larger classes than L2(Rn), for instance: f ∈ D′(Rn) and ψ ∈ D(Rn). In fact, it is enough to have ψ ∈ A(Rn) and f ∈ A′(Rn), where A(Rn) is a time-frequency shift invariant topological vector space. Note also that the inversion formula (2.3) holds pointwise when f is sufficiently regular, for instance, for function in the Schwartz class S(Rn). For a complete account on the STFT, we refer to [9]. 2.3. Spaces. The Hasumi-Silva [24, 12] test function space K1(Rn) consists of those ϕ ∈ C ∞(Rn) for which all norms νk(ϕ) := sup t∈Rn, α≤k ektϕ(α)(t), k ∈ N0, are finite. The elements of K1(Rn) are called exponentially rapidly decreasing smooth It is easy to see that K1(Rn) is an FS-space and therefore Montel and functions. reflexive. The space K1(Rn) is also nuclear [12]. Note that if ϕ ∈ K1(Rn), then the Fourier transform (2.1) extends to an entire In fact, the Fourier transform is a topological isomorphism from K1(Rn) function. 4 S. KOSTADINOVA, S. PILIPOVI ´C, K. SANEVA, AND J. VINDAS onto U(Cn), the space of entire functions which decrease faster than any polynomial in bands. More precisely, a entire function φ ∈ U(Cn) if and only if (1 + z2)k/2φ(z) < ∞, ∀k ∈ N0, νk(φ) := sup z∈Πk where Πk is the tube Πk = Rn + i[−k, k]n. The dual space K′ 1(Rn) consists of all distributions f of exponential type, i.e., those of the form f =Pα≤l(es · fα)(α), where fα ∈ L∞(Rn) [12]. The Fourier transform extends to a topological isomorphism F : K′ 1(Rn) → U ′(Cn), the latter space is known as the space of Silva tempered ultradistributions [12] (also called the space of tempered ultra-hyperfunctions [16]). The space U ′(Cn) contains the space of analytic functionals. See also the textbook [14] for more information about these spaces. We introduce a generalization of the Schwartz space of bounded distributions B′(Rn) [23, p. 200]. Let ω : Rn → (0, ∞) be an exponentially moderate weight, namely, ω is measurable and satisfies the estimate (2.4) ω(x + y) ≤ Aω(y)eax, x, y ∈ Rn, for some constants A > 0 and a ≥ 0. For instance, any positive measurable function ω which is submultiplicative, i.e., ω(x + y) ≤ ω(x)ω(y), and integrable near the origin must necessarily satisfy (2.4), as follows from the standard results about subadditive functions [1, 13]. Extending the Schwartz space DL1(Rn), we define the Fr´echet space 0 }, provided with the family of DL1 norms ω (Rn) = {ϕ ∈ C ∞(Rn) : ϕ(α) ∈ L1 ω(Rn), ∀α ∈ Nn kϕk1,ω,k := sup ϕ(α)(t)ω(t)dt, k ∈ N0. α≤kZRn ω(Rn) stands for the strong dual of DL1 Then, B′ we have the dense embedding K1(Rn) ֒→ DL1 ω(Rn) the space of ω-bounded distributions. We also define B′ B′ D(Rn) in B′ ω (Rn), we have B′ ω (Rn), i.e., B′ ω(Rn). ω(Rn) = (DL1 ω(Rn) ⊂ K′ ω(Rn))′. Since 1(Rn). We call ω(Rn) as the closure of norms tained as the completion of K1(Rn) ⊗ U(Cn) in, say, the π- or the ε- topology [27]. Next, we shall consider K1(Rn)b⊗U(Cn), the topological tensor product space ob- Explicitly, the nuclearity of K1(Rn) implies that K1(Rn)b⊗U(Cn) = K1(Rn)b⊗πU(Cn) = K1(Rn)b⊗εU(Cn). Thus, the topology of K1(Rn)b⊗U(Cn) is given by the family of the and we also obtain (K1(Rn)b⊗U(Cn))′ = K′ Finally, let m be a weight on R2n, that is, m : R2n → (0, ∞) is measurable and m (R2n) consists ∂xα Φ(x, z)(cid:12)(cid:12)(cid:12)(cid:12) , k ∈ N0, locally bounded. Then, if p, q ∈ [1, ∞], the weighted Banach space Lp,q of all measurable functions F such that (x,z)∈Rn×Πk, α≤k ρk(Φ) := sup ∂α ekx(1 + z2)k/2(cid:12)(cid:12)(cid:12)(cid:12) 1(Rn)b⊗U ′(Cn). F (x, ξ)pm(x, ξ)pdx(cid:19)q/p kF kLp,q m := ZRn(cid:18)ZRn dξ!1/q < ∞. (With the obvious modification when p = ∞ or q = ∞.) THE STFT OF DISTRIBUTIONS OF EXPONENTIAL TYPE 5 3. Short-time Fourier transform of distributions of exponential type In this section we study the mapping properties of the STFT on the space of distri- butions of exponential type. Note that the STFT extends to the sesquilinear mapping (f, ψ) 7→ Vψf and its adjoint induces the bilinear mapping (F, ψ) 7→ V ∗ We start with the test function space K1(Rn). If f, ψ ∈ K1(Rn), then we immedi- ately get that (2.2) extends to a holomorphic function in the second variable, namely, Vψf (x, z) is entire in z ∈ Cn. We write in the sequel z = ξ + iη with ξ, η ∈ Rn. Observe ψ ∈ K1(Rn), then for arbitrary η ∈ Rn we may write V ∗ also that an application of the Cauchy theorem shows that if Φ ∈ K1(Rn)b⊗U(Cn) and ψ Φ as ψ F . V ∗ ψ Φ(t) =ZZR2n (3.1) Φ(x, ξ + iη)ψ(t − x)e2πi(ξ+iη)·tdxdξ. Our first proposition deals with the range and continuity properties of V and V ∗ on test function spaces. Proposition 3.1. The following mappings are continuous: ∂α Proof. For part (i), let ϕ, ψ ∈ K1(Rn). Let k be an even integer. If (x, z) ∈ Rn × Πk and α ≤ k, then (i) V : K1(Rn) × K1(Rn) → K1(Rn)b⊗U(Cn). (ii) V ∗ : (K1(Rn)b⊗U(Cn)) × K1(Rn) → K1(Rn). ∂xα Vψϕ(x, z)(cid:12)(cid:12)(cid:12)(cid:12) (1 − ∆t)k/2(ϕ(t)ψ(α)(t − x)e2πη·t)dt(cid:12)(cid:12)(cid:12)(cid:12) ekxZRn(cid:12)(cid:12)(cid:12)ϕ(β1)(t)ψ(α+β2)(t − x)(cid:12)(cid:12)(cid:12) e2πktdt, ekx(1 + z2)k/2(cid:12)(cid:12)(cid:12)(cid:12) ≤ (1 + nk2)k/2ekx(cid:12)(cid:12)(cid:12)(cid:12)ZRn ≤ Ck Xβ1+β2≤k ψ Φ(t)(cid:12)(cid:12)(cid:12)(cid:12) ≤ (2π)αXβ≤α(cid:18)α ≤ (4π)ανk(ψ)ZZR2n ξkekxΦ(x, ξ)dxdξ ∂α ∂tα V ∗ K1(Rn), and α ≤ k, we obtain which shows that ρk(Vψϕ) ≤ Ckν8k(ϕ)νk(ψ). For (ii), if Φ ∈ K1(Rn)b⊗U(Cn), ψ ∈ ekt(cid:12)(cid:12)(cid:12)(cid:12) β(cid:19)ektZZR2n ξkΦ(x, ξ)ψ(β)(t − x)dxdξ ≤ Ak,nνk(ψ)ρk+n+1(Φ); hence ρk(V ∗ ψ Φ) ≤ Ak,nνk(ψ)ρk+n+1(Φ). (cid:3) Observe that if the window ψ ∈ K1(Rn) \ {0} and γ ∈ K1(Rn) is a synthesis window, the reconstruction formula (2.3) reads as: (3.2) 1 (γ, ψ)L2 V ∗ γ Vψ = idK1(Rn). 6 S. KOSTADINOVA, S. PILIPOVI ´C, K. SANEVA, AND J. VINDAS We now study the STFT on K′ 1(Rn). Notice that the modulation operators Mz operate continuously on K1(Rn) even when z ∈ Cn. Thus, if f ∈ K′ 1(Rn) and ψ ∈ K1(Rn) then Vψf , defined by the dual pairing in (2.2), also extends in the second variable as an entire function Vψf (x, z) in z ∈ Cn. Furthermore, it is clear that Vψf (x, z) is C ∞ in x ∈ Rn. We begin with a lemma. Lemma 3.2. Let ψ ∈ K1(Rn). (a) Let B′ ⊂ K′ 1(Rn) be a bounded set. There is k = kB′ ∈ N0 such that (3.3) sup f ∈B′, (x,z)∈Rn×Πλ e−kx−2πx·ℑm z(1 + z)−kVψf (x, z) < ∞, ∀λ ≥ 0. Proof. Part (a). By the Banach-Steinhaus theorem, B′ is equicontinuous, so that there are C > 0 and k ∈ N0 such that hf, ϕi ≤ Cνk(ϕ), ∀f ∈ B′, ∀ϕ ∈ K1(Rn). Hence, for all f ∈ B′ (z = ξ + iη), Vψf (x, z) ≤ C sup (b) For every f ∈ K′ (3.4) 1(Rn) and Φ ∈ K1(Rn)b⊗U(Cn), hVψf, Φi =Df, V ∗ ψ ΦE . ∂tα(cid:16)e−2πiz·tψ(t − x)(cid:17)(cid:12)(cid:12)(cid:12)(cid:12) ekt(cid:12)(cid:12)(cid:12)(cid:12) ekt+2πη·tXβ≤α(cid:18)α t∈Rn,α≤k sup ∂α t∈Rn,α≤k ≤ (2π)kC(1 + z2)k/2 ≤ (4π)kC(1 + z2)k/2ekx+2πη·xνk+1+⌊2πη⌋(ψ), β(cid:19)(cid:12)(cid:12)ψ(β)(t − x)(cid:12)(cid:12) where ⌊2πη⌋ stands for the integral part of 2πη. Part (b). We first remark that the left hand side of (3.4) is well defined because of part (a). To show (3.4), notice that the integral in (3.1), with η = 0, can be approximated by a sequence of convergent Riemann sums in the topology of K1(Rn); this justifies the exchange of integral and dual pairing in (cid:28)f (t),ZZR2n Φ(x, ξ)e−2πiξ·tψ(t − x)dxdξ(cid:29)t =ZZR2n which is the same as (3.4). Φ(x, ξ)hf, MξTxψidxdξ, (cid:3) In particular, if B′ is a singleton, part (a) of Lemma 3.2 gives the growth order of the function Vψf on every set Rn × Πλ. Let us define the adjoint STFT on ∈ K′ 1(Rn)b⊗U ′(Cn). Definition 3.3. Let ψ ∈ K1(Rn). The adjoint STFT V ∗ the distribution V ∗ ψ F ∈ K′ ψ of F ∈ K′ 1(Rn) whose action on test functions is given by hV ∗ ψ F, ϕi :=(cid:10)F, Vψϕ(cid:11) , ϕ ∈ K1(Rn). (3.5) The next theorem summarizes our results. Theorem 3.4. The two STFT mappings 1(Rn)b⊗U ′(Cn) is THE STFT OF DISTRIBUTIONS OF EXPONENTIAL TYPE 7 (i) V : K′ (ii) V ∗ : (K′ 1(Rn) × K1(Rn) → K′ 1(Rn)b⊗U ′(Cn) 1(Rn)b⊗U ′(Cn)) × K1(Rn) → K′ 1(Rn) are hypocontinuous. Let ψ ∈ K1(Rn) \ {0} and let γ ∈ K1(Rn) be a synthesis window for it. The following inversion and desingularization formulas hold: (3.6) 1 (γ, ψ)L2 V ∗ γ Vψ = idK′ 1(Rn), and, for all f ∈ K′ 1(Rn), ϕ ∈ K1(Rn), and η ∈ Rn, (3.7) hf, ϕi = Vψf (x, ξ + iη)Vγϕ(x, −ξ − iη)dxdξ. 1 (γ, ψ)L2ZZR2n Proof. That V and V ∗ are hypocontinuous on these spaces follows from Proposition 3.1 and the formula (3.4) from Lemma 3.2; we leave the details to the reader. By the Cauchy theorem, it is enough to show (3.7) for η = 0. Using (3.5), (3.4), and (3.2), we have hV ∗ ψ Vγϕi = (γ, ψ)L2hf, ϕi, namely, (3.6) and (3.7). (cid:3) γ Vψf, ϕi = hVψf, Vγϕi = hf, V ∗ The next corollary gives the converse to part (a) of Lemma 3.2 under a weaker 1(Rn) in terms inequality than (3.3), namely, a characterization of bounded sets in K′ of the STFT. Corollary 3.5. Let B′ ⊂ K′ k ∈ N0 such that 1(Rn) and ψ ∈ K1(Rn) \ {0}. If there are η ∈ Rn and (3.8) sup f ∈B′,(x,ξ)∈R2n e−kx(1 + ξ)−kVψf (x, ξ + iη) < ∞, then the set B′ is bounded in K′ k ∈ N0 such that (3.3) holds. 1(Rn). Conversely, if B′ is bounded in K′ 1(Rn) there is Proof. In view of the Banach-Steinhaus theorem, we only need to show that B′ is weakly bounded. Let γ be a synthesis window for ψ and let ϕ ∈ K1(Rn). Then, by the desingularization formula (3.7), we have hf, ϕi ≤ sup f ∈B′ Cη (γ, ψ)L2ZZR2n ekx(1 + ξ)k Vγϕ(x, −ξ − iη) dxdξ < ∞, because Vγϕ ∈ K1(Rn)b⊗U(Cn). The converse was already shown in Lemma 3.2. 4. Characterizations of B′ ω(Rn) and B′ ω(Rn) (cid:3) We now turn our attention to the characterization of the space of ω-bounded dis- ω(Rn). Recall that ω stands for an exponentially tributions B′ moderate weight, i.e., a positive and measurable function satisfying (2.4). ω(Rn) and its subspace B′ Theorem 4.1. Let f ∈ K′ 1(Rn) and ψ ∈ K1(Rn) \ {0}. (i) The following statements are equivalent: (a) f ∈ B′ (b) The set {T−hf /ω(h) : h ∈ Rn} is bounded in K′ ω(Rn). 1(Rn). 8 S. KOSTADINOVA, S. PILIPOVI ´C, K. SANEVA, AND J. VINDAS (c) There is s ∈ R such that (4.1) sup (x,ξ)∈R2n (1 + ξ)−s Vψf (x, ξ) ω(x) < ∞. (ii) The next three conditions are equivalent: ω(Rn). (a)′ f ∈ B′ (b)′ limh→∞ T−hf /ω(h) = 0 in K′ (c)′ There is s′ ∈ R such that 1(Rn). (4.2) lim (x,ξ)→∞ (1 + ξ)−s′ Vψf (x, ξ) = 0. ω(x) Remark 4.2. Theorem 4.1 remains valid if we replace K′ 1(Rn) and K1(Rn) by D′(Rn) and D(Rn) everywhere in the statement. Schwartz has shown in [23, p. 204] the equivalence between (a) and (b) for ω = 1 by using a much more complicated method involving a parametrix technique. Proof. Part (i). (a) ⇒ (b). Let f ∈ B′ ω(Rn), since K1(Rn) is barreled, we only need to show that the f ∗ ϕ is bounded by ω for fixed ϕ ∈ K1(Rn). Let B := {φ ∈ D(Rn) : RRn φ(x)ω(x)dx ≤ 1}. By the assumption (2.4), α≤kZRn k ϕ ∗ φk1,ω,k ≤ A max ϕ(α)(x)eaxdx, ∀k ∈ N0, ∀φ ∈ B, namely, the set ϕ ∗ B is bounded in DL1 supφ∈B hf, ϕ ∗ φi < ∞. Since D(Rn) is dense in L1 (L1 ω(Rn). Consequently, supφ∈B hf ∗ ϕ, φi = ω(Rn), this implies that f ∗ ϕ ∈ ω(Rn))′, i.e., suph∈Rn (f ∗ ϕ)(h)/ω(h) < ∞, as claimed. (b) ⇒ (c). Notice that (VψT−hf )(x, z) = e2πiz·hVψf (x + h, z). Fix λ ≥ 0. By Corollary 3.5 (cf. (3.3)), there are k ∈ N0 and Cλ > 0 such that, for all x, h ∈ Rn and z ∈ Πλ, e2πiz·hVψf (x + h, z) ≤ Cλω(h)(1 + z)ke(k+2πλ)x. Taking x = 0 and ℑm z = 0, one gets (4.1). (c) ⇒ (a). Fix a synthesis window γ ∈ K1(Rn). In view of (2.4), one has that if j is any non-negative even integer and λ ≥ 0, then, for all ϕ ∈ DL1 ω(Rn), We may assume that s is an even integer. By (4.1) and the previous estimate, we obtain, for every ϕ ∈ K1(Rn), hf, ϕi ≤ (1 + ξ)sω(x) Vγϕ(x, −ξ) dxdξ ≤ Cskϕk1,ω,s+n+1, e−2πx·ℑm zω(x)Vγϕ(x, z)dx ω(x)ϕ(β1)(t)γ(β2)(t − x)e2πλt−xdtdx γ(β)(x)e(2πλ+a)xdx ≤ Cj,λkϕk1,ω,j. sup z∈Πλ (1 + z2)j/2ZRn ≤ Cj Xβ1+β2≤jZZR2n β≤jZRn ≤ A Cjkϕk1,ω,j max C (ψ, γ)L2ZZR2n THE STFT OF DISTRIBUTIONS OF EXPONENTIAL TYPE 9 which yields f ∈ B′ ω(Rn). Part (ii). Any of the conditions implies that f ∈ B′ ω(Rn). (a)′ ⇒ (b)′. Fix ϕ ∈ K1(Rn). Given fixed ε > 0, we must show that lim suph→∞ hT−hf, ϕi/ω(h) ≤ ε. ω(Rn). Since f is in the Notice that {Thϕ/ω(h) : h ∈ Rn} is a bounded set in DL1 closure of D(Rn) in B′ ω(Rn), there is φ ∈ D(Rn) such that hT−h(f − φ), ϕ ≤ εω(h) for every h ∈ Rn. Consequently, lim sup h→∞ hT−hf, ϕi ω(h) ≤ ε + lim h→∞ 1 ω(h)(cid:12)(cid:12)(cid:12)(cid:12)ZRn ϕ(t − h)φ(t)dt(cid:12)(cid:12)(cid:12)(cid:12) ≤ ε. (b)′ ⇒ (c)′. If ξ remains on a compact of K ⊂ Rn, then {Mξψ : ξ ∈ K} is compact in K1(Rn), thus, by the Banach-Steinhaus theorem, 0 = lim x→∞ hT−xf, Mξψi ω(x) = lim x→∞ Vψf (x, ξ) ω(x) , uniformly in ξ ∈ K. There is s such that (4.1) holds. Taking into account that the above limit holds for arbitrary K, we obtain that (4.2) is satisfied for any s′ > s. (c)′ ⇒ (a)′. We may assume that s′ is a non-negative even integer. Consider the weight ωs′(x, ξ) = ω(x)(1 + ξ)s′. The limit relation (4.2) implies that Vψf is in the closure of K1(Rn) ⊗ S(Rn) with respect to the norm k kL∞,∞ . Since we have the dense embedding U(Cn) ֒→ S(Rn), there is a sequence {Φj}∞ that limj→∞ Φj = Vψf in L∞,∞ φj = V ∗ for any ϕ ∈ K1(Rn), 1/ωs′ (R2n). Let γ ∈ K1(Rn) be a synthesis window and set γ Φj ∈ K1(Rn) (cf. Proposition 3.1). By the relations (3.7) and (3.5), we have j=1 ⊂ K1(Rn)b⊗U(Cn) such 1/ω s′ hf − φj, ϕi ≤ Ckϕk1,ω,s+n+1 (γ, ψ)L2 kVψf − ΦjkL∞,∞ 1/ωs′ , where C does not depend on j. Thus, φj → f in B′ f ∈ B′ ω(Rn) because D(Rn) ֒→ K1(Rn). ω(Rn), which in turn implies that We immediately get the ensuing result, a corollary of Theorem 4.1. (cid:3) ω(Rn). In particular, f ∈ D′(Rn) belongs 1(Rn) if and only if there is s ∈ R such that {e−shT−hf : h ∈ Rn} is bounded in 1(Rn) =Sω B′ ω(Rn) =Sω Corollary 4.3. K′ to K′ D′(Rn). B′ We present here the characterization of the spaces K1(Rn), K′ 5. Characterizations through modulation spaces 1(Rn), B′ ω(Rn), B′ ω(Rn), U(Cn), and U ′(Cn) in terms of modulation spaces. Let us recall the definition of the modulation spaces. There are several equivalent ways to introduce them [9]. Here we follow the approach from [4, 5] based on Gelfand- Shilov spaces. We are interested in modulation spaces with respect to weights that are 10 S. KOSTADINOVA, S. PILIPOVI ´C, K. SANEVA, AND J. VINDAS exponentially moderate. We denote by M the class of all weight functions m on R2n that satisfy inequalities (for some constants A > 0 and a ≥ 0): m(x1 + x2, ξ1 + ξ2) m(x1, ξ1) ≤ Aea(x2+ξ2), (x1, ξ1), (x2, ξ2) ∈ R2n. Observe that any so-called v-moderate weight [9] belongs to M. We also consider the Gelfand-Shilov space Σ1 1(Rn) of Beurling type (sometimes also denoted as S (1)(Rn) or G(Rn)) and its dual (Σ1 1(Rn) consists [3] of all entire functions ϕ 1)′(Rn). The space Σ1 such that ϕ(x)eλx < ∞ and sup x∈Rn sup ξ∈Rn bϕ(ξ)eλξdξ < ∞, ∀λ > 0. We refer to [19] for topological properties of Σ1 1)′(Rn) is also known as the space of Silva ultradistributions of exponential type [14, 25] or the space of Fourier ultra-hyperfunctions [18]. If m ∈ M, ψ ∈ Σ1 1(Rn) \ {0}, and p, q ∈ [1, ∞], the modulation space M p,q m (Rn) is defined as the Banach space 1(Rn). The dual space (Σ1 (5.1) M p,q m (Rn) = {f ∈ (Σ1 1)′(Rn) : kf kM p,q m := kVψf kLp,q m < ∞}. This definition does not depend on the choice of the window ψ, as different windows lead to equivalent norms. If p = q, then we write M p m (Rn). The m(Rn) (for m = 1) was original introduced by Feichtinger in [8]. We shall space M 1 m (Rn) = {f ∈ also define (Σ1 1)′(Rn) : lim(x,ξ)→∞ m(x, ξ)Vψf (x, ξ) = 0}. We now connect the space of exponential distributions with the modulation spaces. For it, we consider the weight subclass M1 ⊂ M consisting of all weights m such that (for some s, a ≥ 0 and A > 0) m (Rn) as the closed subspace of M ∞ m (Rn) given by M ∞ m(Rn) instead of M p,q M ∞ (5.2) m(x1 + x2, ξ1 + ξ2) m(x1, ξ1) ≤ Aeax2(1 + ξ2)s, (x1, ξ1), (x2, ξ2) ∈ R2n. 1(Rn) ֒→ K1(Rn), Let m ∈ M1. By Proposition 3.1, K1(Rn) ⊂ M p,q m (Rn) we obtain that K1(Rn) is dense (weakly∗ dense if p = ∞ or q = ∞) in M p,q 1(Rn). It follows from the results of [9] that we may use and therefore M p,q ψ ∈ K1(Rn) \ {0} in (5.1). Also, if f ∈ M p,q m (Rn) and ψ ∈ K1(Rn) then Vψf is an entire function in the second variable (cf. Section 3); the next proposition describes the norm behavior of Vψf (x, z) in the complex variable z ∈ Cn. m (Rn). Since Σ1 m (Rn) ⊂ K′ Proposition 5.1. Let m ∈ M1, p, q ∈ [1, ∞], and ψ ∈ K1(Rn) \ {0}. If f ∈ M p,q then (∀λ ≥ 0) m (Rn), (5.3) η≤λ ZRn(cid:18)ZRn sup e−2πx·ηVψf (x, ξ + iη)m(x, ξ)pdx(cid:19)q/p dξ!1/q < Cλkf kM p,q m . (With obvious changes if p = ∞ or q = ∞.) Proof. Assume that m satisfies (5.2) and set v(x, ξ) = (1 + ξ)seax. Notice first that e−2πx·ηVψf (x, ξ + iη) = Vψη f (x, ξ), where ψη(t) = e2πη·tψ(t). As in the proof of [9, THE STFT OF DISTRIBUTIONS OF EXPONENTIAL TYPE 11 Prop. 11.3.2, p. 234], Vψη f Lp,q m = 1 kψk2 L2 (VψηV ∗ ψ )Vψf Lp,q m ≤ CVψη ψL1 vVψf Lp,q m . Since {ψη : η ≤ λ} is bounded in K1(Rn), we obtain that {Vψη ψ : η ≤ λ} is bounded v < ∞. in K1(Rn)b⊗U(Cn); hence supη≤λ VψηψL1 e−2πix·ξV bψbf (ξ, −x), we can transfer results from K′ Using the fundamental identity of time-frequency analysis, i.e. [9, p. 40] Vψf (x, ξ) = 1(Rn) into U ′(Rn) by employing the weight class M2 = {m ∈ M : m(x, ξ) = m(ξ, x) ∈ M1}. For s, a ≥ 0, we employ the following special classes of weights (ω satisfies the conditions imposed in Subsection 2.3): (cid:3) vs,a(x, ξ) := eax(1 + ξ)s and ωs(x, ξ) := ω(x)(1 + ξ)s. Clearly vs,a, ωs ∈ M1. Obviously, for every m ∈ M1 there are s, a ≥ 0 such that M p,q m (Rn) ⊆ M p,q vs,a(Rn) ⊆ M p,q (Rn). 1/vs,a Proposition 5.2. Let p, q ∈ [1, ∞]. Then, (5.4) (5.5) (5.6) M p,q m (Rn), M p,q m (Rn), K′ 1(Rn) = [m∈M1 K1(Rn) = \m∈M1 ω(Rn) =[s>0 B′ M ∞ 1/ωs(Rn), and U ′(Cn) = [m∈M2 U(Cn) = \m∈M2 ω(Rn) =[s>0 B′ M p,q m (Rn), M p,q m (Rn), M ∞ 1/ωs(Rn). Proof. The results for U(Cn) and U ′(Cn) follow from those for K1(Rn) and K′ 1(Rn). The equalities in (5.6) are a reformulation of the equivalences (a) ⇔ (c) and (a)′ ⇔ (c)′ from Theorem 4.1. By (5.2) and [9, Cor. 12.1.10, p. 254], given m ∈ M1, there are (Rn) hold. s, a > 0 such that the embeddings M ∞ Thus, part (a) from Lemma 3.2 gives the equality K′ 1/vs,a(Rn) = vs+n+1,a+ε(Rn) ⊆ M p,q m (Rn) ⊆ M ∞ 1/vs,a m (Rn). In view of Proposition 3.1, it only remains to show that 1(Rn) = Ss,a>0 M ∞ Sm∈M1 M p,q \m∈M1 M p,q m (Rn) = \s,a>0 M ∞ vs,a(Rn) ⊆ K1(Rn). (5.7) We show the latter inclusion by proving that if f ∈ M ∞ is holomorphic in the tube Rn + i{η ∈ Rn : η < a/(2π)} and satisfies vs,a(Rn) (with s, a > 0), then bf In fact, choose a positive window ψ ∈ D(Rn) such that Pj∈Zn ψ(t − j) = 1 for all t ∈ Rn. Since f =Pj∈Zn f Tjψ, we obtain bf =Pj∈Zn Vψf (j, · ), with convergence in (1 + z2)s/2bf (z) < ∞, a 2π ∀λ < ℑm z≤λ sup . 12 S. KOSTADINOVA, S. PILIPOVI ´C, K. SANEVA, AND J. VINDAS U ′(Cn). In view of Proposition 5.1, each Vψf (j, z) is entire in z and satisfies the bounds sup (1 + z2)s/2Vψ(j, z) < Cλe−(a−2πλ)j. ℑm z≤λ The Weierstrass theorem implies that bf (z) = Pj∈Zn Vψf (j, z) is holomorphic in the stated tube domain and we also obtain (5.7). Summing up, if f ∈ Ts,a>0 M ∞ then bf ∈ U(Cn), i.e., f ∈ K1(Rn). The following corollary collects what was shown in the proof of Proposition 5.2. vs,a(Rn), (cid:3) Corollary 5.3. Let s, a > 0. If f ∈ M ∞ Rn + i{η ∈ Rn : η < a/(2π)} and satisfies the bounds (5.7). vs,a(Rn), then bf is holomorphic in the tube We make a remark concerning Proposition 5.2. Remark 5.4. Employing [26, Thrms. 3.2 and 3.4], Proposition 5.2 can be extended for p, q ∈ (0, ∞]. 6. Tauberian theorems for S-asymptotics of distributions In this section we characterize the so-called S-asymptotic behavior of distributions in terms of the STFT. We briefly explain this notion; we refer to [21] for a complete treatment of the subject. Let f ∈ K′ 1(Rn). The idea of the S-asymptotics is to study the asymptotic properties of the translates T−hf with respect to a locally bounded and measurable comparison function c : Rn → (0, ∞). It is said that f has S-asymptotic behavior with respect to c if there is g ∈ D′(Rn) such that (6.1) lim h→∞ 1 c(h) T−hf = g in D′(Rn). The distribution g is not arbitrary; in fact, one can show [21] that the relation (6.1) forces it to have the form g(t) = Ceβ·t, for some C ∈ R and β ∈ Rn. If C 6= 0, one can also prove [21] that c must satisfy the asymptotic relation (6.2) lim h→∞ c(t + h) c(h) = eβ·t, uniformly for t in compact subsets of Rn. ¿From now on, we shall always assume that c satisfies (6.2). A typical example of such a c is any function of the form c(t) = eβ·tL(et), where L is a Karamata slowly varying function [2]. The assumption (6.2) implies [21] that (6.1) actually holds in the space K′ 1(Rn). We will use the more suggestive notation (6.3) f (t + h) ∼ c(h)g(t) in K′ 1(Rn) as h → ∞ for denoting (6.1), which of course means that (f ∗ ϕ)(h) ∼ c(h)RRn ϕ(t)g(t)dt as h → ∞, for each ϕ ∈ D(Rn) (or, equivalently, ϕ ∈ K1(Rn)). In order to move further, we give an asymptotic representation formula and Potter type estimates [2] for c: THE STFT OF DISTRIBUTIONS OF EXPONENTIAL TYPE 13 Lemma 6.1. The locally bounded measurable function c satisfies (6.2) if and only if there is b ∈ C ∞(Rn) such that limx→∞ b(α)(x) = 0 for every multi-index α > 0 and (6.4) c(x) ∼ exp (β · x + b(x)) as x → ∞. In particular, for each ε > 0 there are constants aε, Aε > 0 such that (6.5) aε exp(β · t − εt) ≤ c(t + h) c(h) ≤ Aε exp(β · t + εt), t, h ∈ Rn. Proof. By considering e−β·tc(t), one may assume that β = 0. Let ϕ ∈ D(Rn) be such that RRn ϕ(t)dt = 1. Set b(x) = RRn log c(t + x)ϕ(t)dt. Clearly, b ∈ C ∞(Rn) and the relation (6.2) implies that b(x) = log c(x)+o(1) and b(α)(x) = o(1) as x → ∞, for each multi-index α > 0. This gives (6.4). Conversely, since c is locally bounded, we may assume that actually c(x) = eβ·x+b(x), but b(t + h) − b(h) ≤ t maxξ∈[h,t+h] ∇b(ξ), which gives (6.2). Using the fact that ∇b is bounded, the same argument yields (6.5). (cid:3) Observe that Lemma 6.1 also tells us that the space B′ c(Rn) is well-defined for c. We can now characterize (6.3) in terms of the STFT. The direct part of the following theorem is an Abelian result, while the converse may be regarded as a Tauberian theorem. If f ∈ K′ 1(Rn) has the S- Theorem 6.2. Let f ∈ K′ asymptotic behavior (6.3) then, for every λ ≥ 0, e2πiz·h Vψf (x + h, z) (6.6) 1(Rn) and ψ ∈ K1(Rn)\{0}. lim h→∞ c(h) = Vψg(x, z), uniformly for z ∈ Πλ and x in compact subsets of Rn. Conversely, suppose that the limits (6.7) lim x→∞ e2πiξ·x Vψf (x, ξ) c(x) = J(ξ) ∈ C exist for almost every ξ ∈ Rn. If there is s ∈ R such that (6.8) sup (x,ξ)∈R2n (1 + ξ)−s Vψf (x, ξ) c(x) < ∞, then f has the S-asymptotic behavior (6.3) with g(t) = Ceβ·t, where the constant is completely determined by the equation J(ξ) = Cbψ(−ξ + iβ/(2π)). Remark 6.3. Assume (6.8). Consider a weight of the form mε(x, ξ) = eβ·x+εx(1 + ξ)s It will be shown below that the asymptotics (6.3) holds in the weak∗ with ε > 0. (Rn), i.e., (f ∗ ϕ)(h) ∼ c(h)hg, ϕi as h → ∞ for every ϕ in the topology of M ∞ mε(Rn). Furthermore, one may use in (6.7) and (6.8) a window modulation space M 1 ψ ∈ M 1 1/mε mε(Rn)\{0}. 14 S. KOSTADINOVA, S. PILIPOVI ´C, K. SANEVA, AND J. VINDAS Proof. Fix λ ≥ 0 and a compact K ⊂ Rn. Note that the set {MzTxψ : (x, z) ∈ K × Πλ} is compact in K1(Rn). By the Banach-Steinhaus theorem, lim h→∞ e2πizh Vψf (x + h, z) c(h) = lim h→∞D T−hf c(h) , MzTxψE =Dg, MzTxψE, uniformly with respect to (x, z) ∈ K × Πλ, as asserted in (6.6). Conversely, assume (6.7) and (6.8). Let H = {ξ ∈ Rn : 1(Rn), T−hf /c(h) converges strongly to a distribution g in K′ In view c(Rn) or, equivalently, {T−hf /c(h) : h ∈ Rn} of Theorem 4.1, we have that f ∈ B′ 1(Rn). By the Banach-Steinhaus theorem and the Montel property is bounded in K′ of K′ 1(Rn) if and only if limh→∞hT−hf, ϕi/c(h) exists for ϕ in a dense subspace of K1(Rn). Let D be the linear span of {MξTxψ : (x, ξ) ∈ Rn × H}. By the desingularization formula (3.7) and the Hahn-Banach theorem, we have that D is dense in K1(Rn). Thus, it suffices to verify that limh→∞hT−hf, MξTxψi/c(h) exists for each (x, ξ) ∈ Rn × H. But in this case (6.2) and (6.7) yield (6.7) holds}. lim h→∞ hT−hf, MξTxψi c(h) = lim h→∞ e2πiξ·h Vψf (x + h, ξ) c(h) e2πiξ·(x+h) Vψf (x + h, ξ) c(h + x) = e(β−2πiξ)·x lim h→∞ = e(β−2πiξ)·xJ(ξ), as required. We already know that g(t) = Ceβ·t. Comparison between (6.6) and (6.7) leads to J(ξ) = Vψg(0, ξ) = CRRn ψ(t)eβ·t−2πiξ·tdt. To show the assertion from Remark 6.3, note first that, by using (6.5), one readily verifies that T−hf M ∞ 1/mε c(h) sup h∈Rn < ∞. mε(Rn), we also have that D is dense mε(Rn) and the assertion follows at once. The fact that one may use a window mε(Rn)\{0} in (6.7) and (6.8) follows in a similar fashion because in this case Since we have the dense embedding K1(Rn) ֒→ M 1 in M 1 ψ ∈ M 1 the desingularization formula (3.7) still holds. (cid:3) Let us make two addenda to Theorem 6.2. The ensuing corollary improves Remark 6.3, provided that c satisfies the extended submultiplicative condition (for some A > 0): (6.9) c(t + h) ≤ Ac(t)c(h). Corollary 6.4. Assume that c satisfies (6.9) and set cs(x, ξ) = c(x)(1 + ξ)s, s ∈ R. cs(Rn) \ {0} such that the limits (6.7) exist for If f ∈ M ∞ 1/cs almost every ξ ∈ Rn, then, for some g, the S-asymptotic behavior (6.3) holds weakly∗ in M ∞ 1/cs(Rn), that is, (f ∗ ϕ)(h) ∼ c(h)hg, ϕi as h → ∞ for every ϕ ∈ M 1 (Rn) and there is ψ ∈ M 1 cs(Rn). THE STFT OF DISTRIBUTIONS OF EXPONENTIAL TYPE 15 Proof. We retain the notation from the proof of Theorem 6.2. The assumption f ∈ (Rn) of course tells us that (6.8) holds. Employing the hypothesis (6.9), one M ∞ 1/cs readily sees that suph∈Rn kT−hf kM ∞ /c(h) < ∞. A similar argument to the one used 1/cs in the proof of Theorem 6.2 yields that the set D associated to ψ is dense in M 1 cs(Rn), which as above yields the result. (cid:3) In dimension n = 1, the next theorem actually obtains the ordinary asymptotic behavior of f in case it is a regular distribution on (0, ∞) satisfying an additional Tauberian condition. We fix mε as in Remark 6.3 and cs as in Corollary 6.4. Theorem 6.5. Let f ∈ M ∞ (6.10) 1/cs(R). Suppose that e2πiξ·x Vψf (x, ξ) lim x→∞ c(x) = J(ξ) ∈ C, for almost every ξ ∈ R, where ψ ∈ M 1 cs(R) \ {0} if c satisfies (6.9)). If there is α ≥ 0 such that eαtf (t) is a positive non-decreasing function on the interval (0, ∞), then mε(R) \ {0} (resp. ψ ∈ M 1 (6.11) lim t→∞ f (t) c(t) = C, where C is the constant from Theorem 6.2. Proof. Using (6.10), the same method from Theorem 6.2 applies to show that f (t+h) ∼ 1(R) as h → ∞, where g(t) = Ceβt. We may assume that α ≥ −β. Set Cg(t) in K′ f (t) = eαtf (t), b(t) = eαtc(t), and r = α+β ≥ 0. It is enough to show that f (t) ∼ Cb(t) as t → ∞, whence (6.11) would follow. By (6.3), we have that f (t + h) ∼ b(h)Cert as h → ∞ in K′ 1(R), h f (t + h), ϕ(t)i ∼ Cb(h)Z ∞ −∞ ertϕ(t)dt, ∀ϕ ∈ K1(Rn). Let ε > 0 be arbitrary. Choose a non-negative test function ϕ ∈ D(R) such that 0 ϕ(t)dt = 1. Using the fact that f is non-decreasing on (0, ∞) i.e., (6.12) supp ϕ ⊆ (0, ε) and R ε and (6.12), we obtain f (h) b(h) lim sup h→∞ = lim sup h→∞ ϕ(t)dt ≤ lim h→∞ f (h) b(h)Z ε 0 h f (t + h), ϕ(t)i = lim h→∞ b(h) = CZ ε 1 b(h)Z ε 0 f (t + h)ϕ(t)dt ertϕ(t)dt ≤ Cerε, taking ε → 0+, we have shown that lim suph→∞ in (6.12) a non-negative ϕ such that supp ϕ ⊆ (−ε, 0) andR 0 lim inf h→∞ f (h)/b(h) ≥ C. This shows that f (t) ∼ Cb(t) as t → ∞, as claimed. 0 f (h)/b(h) ≤ C. Similarly, choosing −ε ϕ(t)dt = 1, one obtains (cid:3) We conclude this article with a proof of Theorem 1.1. 16 S. KOSTADINOVA, S. PILIPOVI ´C, K. SANEVA, AND J. VINDAS Proof of Theorem 1.1. Set c(t) = eβtL(et) and, as before (with s = 0), c0(x, ξ) = c(x) and mε(x, ξ) = eβx+εx. Note that (1.2) is the same as (6.11). Let us first verify that ψ ∈ M 1 mε(R). In fact, if we take another window γ ∈ K1(R), we have Vγψ(x, ξ)eβx+εxdxdξ =ZZR2 ψ(t − x)γ(t)eβx+εxdtdx + ZZR2 ≤ C ZZR2 (1 + ξ3)Vγψ(x, ξ)eβx+εxdx dξ 1 + ξ3 3Xj=0ZZR2 ψ(j)(t − x)γ(3−j)(t)eβx+εxdtdx! , which is finite (a similar argument shows that ψ ∈ M 1 ψ′′(t))L(et)eβtdt < ∞). M ∞ 1/c0 Since f is non-decreasing, we have (R). Let us first show the crude bound f (t) = O(c(t)). Set A1 =R ∞ −∞(ψ(t) + ψ′(t) + In view of Theorem 6.5, it is enough to establish f ∈ 0 ψ(t)dt < ∞. c0(R) if R ∞ f (x) ≤ f (t)ψ(t − x)dt ≤ A2c(x), 1 0 0 1 f (t + x)ψ(t)dt ≤ A1Z ∞ A1Z ∞ c(t)ψ(t−x)dt ≤ c(x) AεZ ∞ −∞ because of (1.1) with ξ = 0. Thus Vψf (x, ξ) ≤ A2Z ∞ 0 eβt+εtψ(t)dt < A3c(x), ∀(x, ξ) ∈ R2 (likewise in the other case using L(xy) ≤ AL(x)L(y)), which completes the proof. (cid:3) References [1] A. Beurling, Sur les int´egrales de Fourier absolument convergentes et leur application `a une transformation fonctionelle, in: IX Congr. Math. Scand., Helsingfors, (1938) 345 -- 366. [2] N. H. Bingham, C. M. Goldie, J. L. Teugels, Regular Variation, Cambridge University Press, Cambridge, 1987. [3] S.-Y. Chung, D. Kim, S. Lee, Characterizations for Beurling-Bjorck space and Schwartz space, Proc. Amer. Math. Soc. 125 (1997), 3229-3234. [4] E. Cordero, Gelfand-Shilov window classes for weighted modulation spaces, Integral Transforms Spec. Funct. 18 (2007), 829 -- 837. [5] E. Cordero, S. Pilipovi´c, L. Rodino, N. Teofanov, Localization operators and exponential weights for modulation spaces, Mediterr. J. Math. 2 (2005), 381 -- 394. [6] W. F. Donoghue, Distributions and Fourier Transforms, Academic Press, New York-London, 1969. [7] R. Estrada, J. Vindas, On Borel summability and analytic functionals, Rocky Mountain J. Math. 43 (2013), 895 -- 903. [8] H. G. Feichtinger, On a new Segal algebra, Monatsh. Math. 92 (1981), 269 -- 289. [9] K. Grochenig, Foundations of time-frequency analysis, Birkhauser Boston, Inc., Boston, MA, 2001. [10] K. Grochenig, G. Zimmermann, Hardy's theorem and the short-time Fourier transform of Schwartz functions, J. London Math. Soc. 63 (2001), 205 -- 214. [11] K. Grochenig, G. Zimmermann, Spaces of test functions via the STFT, J. Funct. Spaces Appl. 2 (2004), 25 -- 53. [12] M. Hasumi, Note on the n-dimensional tempered ultra-distributions, Tohoku Math. J. 13 (1961), 94 -- 104. [13] E. Hille, R. S. Phillips, Functional analysis and semi-groups, American Mathematical Society, Providence, R. I., 1957. THE STFT OF DISTRIBUTIONS OF EXPONENTIAL TYPE 17 [14] R. F. Hoskins, J. Sousa Pinto, Theories of generalised functions. Distributions, ultradistributions and other generalised functions, Horwood Publishing Limited, Chichester, 2005. [15] J. Korevaar, Tauberian theory. A century of developments, Springer-Verlag, Berlin, 2004. [16] M. Morimoto, Theory of tempered ultrahyperfunctions. I, II, Proc. Japan Acad. 51 (1975), 87 -- 91; 51 (1975), 213 -- 218. [17] H. M. Obiedat, Z. Mustafa, F. Awawdeh, Short-time Fourier transform over the Silva space, Inter. J. Pure Appl. Math. 44 (2008), 755 -- 764. [18] Y. S. Park, M. Morimoto, Fourier ultra-hyperfunctions in the Euclidean n−space, J. Fac. Sci. Univ. Tokyo Sect. IA Math. 20 (1973), 121 -- 127. [19] S. Pilipovi´c, Tempered ultradistributions, Boll. Un. Mat. Ital. B (7) 2 (1988), 235 -- 251. [20] S. Pilipovi´c, B. Stankovi´c, Wiener Tauberian theorems for distributions, J. London Math. Soc. (2) 47 (1993), 507 -- 515. [21] S. Pilipovi´c, B. Stankovi´c, J. Vindas, Asymptotic behavior of generalized functions, Series on Analysis, Applications and Computation, 5, World Scientific Publishing Co. Pte. Ltd., Hacken- sack, NJ, 2012. [22] K. Saneva, R. Aceska, S. Kostadinova, Some Abelian and Tauberian results for the short-time Fourier transform, Novi Sad J. Math. 43 (2013), 81 -- 89. [23] L. Schwartz, Th´eorie des distributions, Hermann, Paris, 1966. [24] J. Sebastiao e Silva, Les fonctions analytiques comme ultra-distributions dans le calcul op´erationnel, Math. Ann. 136 (1958), 58 -- 96. [25] J. Sebastiao e Silva, Les s´eries de multipoles des physiciens et la th´eorie des ultradistributions, Math. Ann. 174 (1967), 109 -- 142. [26] J. Toft, The Bargmann transform on modulation and Gelfand-Shilov spaces, with applications to Toeplitz and pseudo-differential operators, J. Pseudo-Differ. Oper. Appl. 3 (2012), 145 -- 227. [27] F. Tr`eves, Topological vector spaces, distributions and kernel, Academic Press, New York, 1967. [28] V. S. Vladimirov, Yu. N. Drozzinov, B. I. Zavialov, Tauberian theorems for generalized functions, Kluwer Academic Publishers Group, Dordrecht, 1988. [29] N. Wiener, Tauberian theorems, Ann. of Math. (2) 33 (1932), 1 -- 100. [30] Z. Ziele´zny, Hypoelliptic and entire elliptic convolution equations in subspaces of the space of distributions. II, Studia Math. 32 (1969), 47 -- 59. Faculty of Electrical Engineering and Information Technologies, Ss. Cyril and Methodius University, Rugjer Boshkovik bb, 1000 Skopje, Macedonia E-mail address: [email protected] Department of Mathematics and Informatics, University of Novi Sad, Trg Dositeja Obradovi´ca 4, 21000 Novi Sad, Serbia E-mail address: [email protected] Faculty of Electrical Engineering and Information Technologies, Ss. Cyril and Methodius University, Rugjer Boshkovik bb, 1000 Skopje, Macedonia E-mail address: [email protected] Department of Mathematics, Ghent University, Krijgslaan 281 Gebouw S22, 9000 Gent, Belgium E-mail address: [email protected]
1606.08606	1	1606	2016-06-28T08:32:20	Mityagin's Extension Problem. Progress Report	[ "math.FA", "math.CA" ]	Given a compact set $K\subset {\Bbb R}^d,$ let ${\mathcal E}(K)$ denote the space of Whitney jets on $K$. The compact set $K$ is said to have the extension property if there exists a continuous linear extension operator $W:{\mathcal E}(K) \longrightarrow C^{\infty}({\Bbb R}^d)$. In 1961 B. S. Mityagin posed a problem to give a characterization of the extension property in geometric terms. We show that there is no such complete description in terms of densities of Hausdorff contents or related characteristics. Also the extension property cannot be characterized in terms of growth of Markov's factors for the set.	math.FA	math	1 MITYAGIN'S EXTENSION PROBLEM. PROGRESS REPORT ALEXANDER GONCHAROV AND ZELIHA URAL Abstract. Given a compact set K ⊂ Rd, let E(K) denote the space of Whitney jets on K. The compact set K is said to have the extension property if there exists a continuous linear extension operator W : E(K) −→ C∞(Rd). In 1961 B. S. Mityagin posed a problem to give a characterization of the extension property in geometric terms. We show that there is no such complete description in terms of densities of Hausdorff contents or related characteristics. Also the extension property cannot be characterized in terms of growth of Markov's factors for the set. 1. introduction By the celebrated Whitney theorem [23], for each compact set K ⊂ Rd, by means of a continuous linear operator one can extend jets of finite order from E p(K) to functions defined on the whole space, preserving the order of differentiability. In the case p = ∞, the possibility of such extension crucially depends on geometry of the set. Following [20], let us say that K has the extension property (EP) if there exists a linear continuous extension operator W : E(K) −→ C ∞(Rd). For example, any set K with an isolated point does not have EP , since here each neighborhood of the space E(K) contains a linear subspace, but this is not the case for C ∞(Rd). B. S. Mityagin posed in 1961 ([13], p.124) the following problem (in our terms): What is a geometric characterization of the extension property? We show that there is no complete characterization of that kind in terms of densities of Hausdorff contents of sets or analogous functions related to Hausdorff measures. This is similar to the state in Potential Theory where R. Nevanlinna [14] and H. Ursell [21] proved that there is no complete characterization of polarity of compact sets in terms of Hausdorff measures. The scale of growth rate of functions h, which define the Hausdorff measure Λh, can be decomposed into three zones. For h from the first zone of small growth, if 0 < Λh(K) then the set K is not polar. For h from the zone of fast growth, if Λh(K) < ∞ then the set K is polar. But between them there is a zone of uncertainty. It is possible to take two functions with h2 ≺ h1 from this zone and the corresponding Cantor-type sets Kj with 0 < Λhj (Kj) < ∞ for j ∈ {1, 2}, such that the large (with respect to the Hausdorff measure) set K2 is polar, whereas the smaller K1 is not polar. Here we present a similar example of two Cantor-type sets: the smaller set has EP whereas the larger set does not have it. 1 Alexander Goncharov, Zeliha Ural (Bilkent University, 06800, Ankara, Turkey) e-mail: [email protected], [email protected] 2010 Mathematics Subject Classification. 46E10, 31A15, 41A10. Key words and phrases. Whitney functions, extension problem, Hausdorff measures, Markov's factors. 1 2 ALEXANDER GONCHAROV AND ZELIHA URAL Of course, such global characteristics as Hausdorff measures or Hausdorff contents cannot be used, in general, to distinguish EP, which we observe if a compact set is not "very small" near each its point. One can suggest for this reason to characterize EP in terms of lower densities of Hausdorff contents of sets, because, clearly, densities of Hausdorff measures cannot be used for this aim. We analyze a wide class of dimension functions and show that lower densities of Hausdorff contents do not distinguish EP . Neither EP can be characterized in terms of growth rate of Markov's factors (Mn(·))∞ for sets. Two sets are presented, K1 with EP and K2 without it, such that Mn(K1) grows essentially faster than Mn(K2) as n → ∞. It should be noted that, by W. Ple´sniak [16], any Markov compact set (with a polynomial growth rate of Mn(·)) has EP . All examples are given in terms of the sets K(γ) introduced in [9]. n=1 Our paper is organized as follows. Section 2 is a short review of main methods of extension. Also we consider there the Tidten-Vogt linear topological characteriza- tion of EP . In Section 3 we give some auxiliary results about the weakly equilibrium Cantor-type set K(γ). In Section 4 we use local Newton interpolations to construct an extension operator W . Sections 5 contains the main result, namely a characterization of EP for E(K(γ)) in terms of a sequence related to γ. In sections 6 we compare W with the extension operator from [11], which is given by individual extensions of ele- ments of Schauder basis for the space E(K(γ)). In Section 7 we consider two examples that correspond to regular and irregular behaviour of the sequence γ. In Section 8 we calculate the Hausdorff h−measure of K(γ) for a siutable dimension function h and present Ursell's type example for EP. In Section 9 we consider Hausdorff contents and related characteristics. In Section 10 we compare the growth of Markov's factors and EP for K(γ). For the basic facts about the spaces of Whitney functions defined on closed subsets of Rd see e.g. [3], the concepts of the theory of logarithmic potential can be found in [17]. Throughout the paper, log denotes the natural logarithm. Given compact set K, Cap(K) stands for the logarithmic capacity of K, Rob(K) = log(1/Cap(K)) ≤ ∞ is the Robin constant for K. If K is not polar then µK is its equilibrium measure. For each set A, let #(A) be the cardinality of A, A be the diameter of A. Also, [a] is k=m(· · · ) = 1 if m > n. The symbol ∼ denotes the strong equivalence: an ∼ bn means that an = bn(1 + o(1)) for n → ∞. the greatest integer in a, Pn k=m(· · · ) = 0 and Qn 2. Three methods of extension j=1 ∈ Zd Let K ⊂ Rd be a compact set, α = (αj)d + be a multi-index. Let I be a closed cube containing K and F (K, I) = {F ∈ C ∞(I) : F (α)K = 0, ∀α} be the ideal of flat on K functions. The Whitney space E(K) of extendable jets consists of traces on K of C ∞-functions defined on I, so it is a factor space of C ∞(I) and the restriction operator R : C ∞(I) −→ E(K) is surjective. This means that the sequence J−→ C ∞(I) R−→ E(K) −→ 0 is exact. If it splits then the right inverse 0 −→ F (K, I) to R is the desired linear continuous extension operator W and K has EP . We see that there always exists a linear extension operator (for example one can individually extend the elements of a vector basis in E(K)) and a continuous extension operator, MITYAGIN'S EXTENSION PROBLEM. PROGRESS REPORT 3 by Whitney's construction. Numerous examples show that a set K has EP if K is not "very small" near each its point, but the exact geometric meaning of "smallness" has not been comprehended yet. In [20] M. Tidten applied D. Vogt's theory of splitting of short exact sequences of Fr´echet spaces (see e.g [12], Chapter 30) and presented the following important linear topological characterization of EP : a compact set K has the extension property if and only if the space E(K) has a dominating norm (satisfies the condition (DN)). Recall that a Fr´echet space X with an increasing system of seminorms ( · k)∞ k=0 has a dominating norm · p if for each q ∈ N there exist r ∈ N and C ≥ 1 such that · 2 q ≤ C · p · r. Concerning the question "How to construct an operator W if it exists?", we can select three main methods that can be applied for wide families of compact sets. ements (en)∞ W (f ) = P∞ n=1 of a topological basis of E(K). Then for f = P∞ The first method goes back to B. S. Mityagin [13]: to extend individually the el- n=1 ξn · en take n=1 ξn · W (en). See Theorem 2.4 in [22] about possibility of suitable si- multaneous extensions of en in the case when K has nonempty interior. The main problem with this method is that we do not know whether each space E(K) has a topological basis, even though E(K) is complemented in C ∞(I). This is a particular case of the significant Mityagin-Pe lczy´nski problem: Suppose X is a nuclear Fr´echet space with basis and E is a complemented subspace of X. Does E possess a basis? The space X = s of rapidly decreasing sequences, which is isomorphic to C ∞(I), presents the most important unsolved case. The second method was suggested in [15], where W. Paw lucki and W. Ple´sniak constructed an extension operator W in the form of a telescoping series containing La- grange interpolation polynomials with Fekete nodes. The authors considered the family of compact sets with polynomial cusps, but later, in [16], the result was generalized to any Markov sets. In fact (see T.3.3 in [16]), for each C ∞ determining compact set K, the operator W is continuous in the so-called Jackson topology τJ if and only if τJ coincides with the natural topology τ of the space E(K) and this happens if and only if the set K is Markov. We remark that τJ is not stronger then τ and that τJ always has the dominating norm property, see e.g. [2]. Thus, in the case of non-Markov compact set with EP ([5], [2]), the Paw lucki-Ple´sniak extension operator is not continuous in τJ , but this does not exclude the possibility for it to be bounded in τ . At least for some non-Markov compact sets, the local version of this operator is bounded in τ ([2]). In [4] L. Frerick, E. Jord´a, and J. Wengenroth showed that, provided some condi- tions, the classical Whitney extension operator for the space of jets of finite order can be generalized to the case E(K). Instead of Taylors polynomials in the Whitney con- struction, the authors used a kind of interpolation by means of certain local measures. A linear tame extension operator was presented for E(K), provided K satisfies a local form of Markov's inequality. 4 ALEXANDER GONCHAROV AND ZELIHA URAL There are some other methods to construct W for closed sets, for example Seeley's extension [18] from a half space or Stein's extension ([19], Ch 6) from sets with the Lipschitz boundary. However these methods, in order to define W (f, x) at some point x, essentially require existence of a line through x with a ray where f is defined, so these methods cannot be applied for compact sets. Here we consider rather small Cantor-type sets that are neither Markov no local Markov. We follow [2] in our construction, so W is a local version of the Paw lucki- Ple´sniak operator. It is interesting that, at least for small sets, W can be considered as an operator extending basis elements of the space. Thus, for such sets, the first method and a local version of the second method coincide. 3. Notations and auxiliary results In what follows we will consider only perfect compact sets K ⊂ I = [0, 1], so the Fr´echet topology τ in the space E(K) can be given by the norms k f kq = f q,K + sup((Rq yf )(k)(x) x − yq−k : x, y ∈ K, x 6= y, k = 0, 1, ...q) for q ∈ Z+, where f q,K = sup{f (k)(x) : x ∈ K, k ≤ q} and Rq is the Taylor remainder. yf (x) = f (x) − T q y f (x) Given f ∈ E(K), let f q = inf F q,I, where the infimum is taken over all possible extensions of f to F ∈ C ∞(I). By the Lagrange form of the Taylor remainder, we have f q ≤ 3 F q,I for any extension F . The quotient topology τQ, given by the norms ( · ∞ q=0), is complete and, by the open mapping theorem, is equivalent to τ. Hence for any q there exist r ∈ N, C > 0 such that (1) f q ≤ C f r for any f ∈ E(K). In general, extensions F that realize f q for a given function f , essentially depend on q. Of course, the extension property of K means the existence of a simultaneous extension which is suitable for all norms. Our main subject is the set K(γ) introduced in [9]. For the convenience of the reader we repeat the relevant material. Given sequence γ = (γs)∞ s=1 with 0 < γs < 1/4, let r0 = 1 and rs = γsr2 s−1 for s ∈ N. Define P2(x) = x(x − 1), P2s+1 = P2s(P2s + rs) and Es = {x ∈ R : P2s+1(x) ≤ 0} for s ∈ N. Then Es = ∪2s j=1Ij,s, where the s-th level basic intervals Ij,s are disjoint and max1≤j≤2s Ij,s → 0 as s → ∞. Here, (P2s + rs/2)(Es) = [−rs/2, rs/2], so the sets Es are polynomial inverse images of intervals. Since Es+1 ⊂ Es, we have a Cantor-type set K(γ) := ∩∞ s=0Es. In what follows we will consider only γ satisfying the assumptions ∞ (2) γk ≤ 1/32 for k ∈ N and γk < ∞. The lengths lj,s of the intervals Ij,s of the s−th level are not the same, but, provided (2), we can estimate them in terms of the parameter δs = γ1γ2 · · · γs ([9], L.6): (3) δs < lj,s < C0 δs for 1 ≤ j ≤ 2s, Xk=1 MITYAGIN'S EXTENSION PROBLEM. PROGRESS REPORT 5 where C0 = exp(16 P∞ k=1 γk). Each Ij,s contains two adjacent basic subintervals I2j−1,s+1 and I2j,s+1. Let hj,s = lj,s − l2j−1,s+1 − l2j,s+1 be the distance between them. By Lemma 4 in [9], hj,s > (1 − 4γs+1)lj,s. Therefore, (4) hj,s ≥ 7/8 · lj,s > 7/8 · δs for all j. In addition, by T.1 in [9], the level domains Ds = {z ∈ C : P2s(z) + rs/2 < rs/2} k=1 2−k log 1 form a nested family and K(γ) = ∩∞ γk represents the Robin constant of Ds. Therefore, the set K(γ) is non-polar if and only s=0Ds. The value Rs = 2−s log 2 +Ps =P∞ n=1 2−n−1 log 1 δn < ∞. if Rob(K(γ)) =P∞ n=1 2−n log 1 γn We decompose all zeros of P2s into s groups. Let X0 = {x1, x2} = {0, 1}, X1 = {x3, x4} = {l1,1, 1 − l2,1}, · · · , Xk = {l1,k, l1,k−1 − l2,k, · · · , 1 − l2k,k} for k ≤ s − 1. Thus, Xk = {x : P2k (x) + rk = 0} contains all zeros of P2k+1 that are not zeros of P2k. Set Ys = ∪s #(Ys) = 2s+1 for s ∈ Z+. We refer s−th type points to the elements of Xs. k=0Xk. Then P2s(x) = Qxk∈Ys−1(x − xk). Clearly, #(Xs) = 2s for s ∈ N and The points from Ys can be ordered using, as in [7], the rule of increase of the type. First we take points from X0 and X1 in the ordering given above. The set X2 = {x5, · · · , x8} consists of the points of the second type. We take xj+4 as the point which is the closest to xj for 1 ≤ j ≤ 4. Here, x5 = x1 + l1,2, x6 = x2 − l4,2, etc. Simi- larly, Xk = {x2k+1, · · · , x2k+1} can be defined by the previous points. For 1 ≤ j ≤ 2k, the point xj is an endpoint of a certain basic interval of k-th level. Let us take xj+2k as its another endpoint. Thus, xj+2k = xj ± li,k, where the sign and i are uniquely defined by j. In the same way, any N points can be chosen on each basic interval. For example, suppose 2n ≤ N < 2n+1 and the points (zk)N k=1 are chosen on Ij,s by this rule. Then the set includes all 2n zeros of P2s+n on Ij,s (points of the type ≤ s + n − 1) and some N − 2n points of the type s + n. We use two technical lemmas from [11]. We suppose that γ satisfies (2). Let 2n ≤ N < 2n+1 and Z = (zk)N +1 k=1 be chosen on a given I = Ij,s by the rule of k=1 and C1 = 8/7 · (C0 + 1). For fixed x ∈ R and increase of the type. Write ZN = (zk)N finite A = (am), let dk(x, A) := x − amk ր . Lemma A. For each x ∈ R with δ = dist(x, ZN ) ≤ δs+n and z ∈ Z we have δs+n QN k=2 dk(x, ZN ) ≤ C N k=2 dk(z, Z). 1 QN +1 In the next lemma we consider the same N and Z, as above, but now we arrange zk in increasing order. For q = 2m − 1 with m < n and 1 ≤ j ≤ N + 1 − q, let J = {zj, · · · , zj+q} be 2m consecutive points from Z. Given j, we consider all possible chains of strict embeddings of segments of natural numbers: [j, j + q] = [a0, b0] ⊂ [a1, b1] ⊂ · · · ⊂ [aN −q, bN −q] = [1, N + 1], where ak = ak−1, bk = bk−1 + 1 or ak = ak−1 − k=1 (zbk − zak ). 1, bk = bk−1 for 1 ≤ k ≤ N − q. Every chain generates the product QN −q Lemma B. For each J ⊂ Z there exists z ∈ J such that QN +1 Potential Theory meaning: Rob(K(γ)) = P∞ We will characterize EP of K(γ) in terms of the values Bk = 2−k−1 · log 1 δk that have k=1 Bk. The main condition is (compare For fixed J, let Π(J) denote the minimum of these products for all possible chains. k=q+2 dk(z, Z) ≤ Π(J). 6 ALEXANDER GONCHAROV AND ZELIHA URAL with (3) in [8]): (5) ⇒ 0 as n → ∞ uniformly with respect to s. Bn+s k=s Bk Pn+s We see that this condition allows polar sets. Example 1. Let γ1 = exp(−4B) and γk = exp(−2kB) for k ≥ 2, where B ≥ 1 4 log 32, so (2) is valid. Here, Bk = B for all k. Hence (5) is satisfied and the set K(γ) is polar. The condition (5) means that (6) ∀ε ∃s0, ∃n0 : Bs+n < ε(Bs + · · · + Bs+n) for n ≥ n0, s ≥ s0. Clearly, instead of ∃s0 one can take above ∀s0. Let us show that (6) is equivalent to (7) ∀ε1 ∀m ∈ Z+∃N : Bs+n−m + · · · + Bs+n < ε1(Bs + · · · + Bs+n−m−1), n ≥ N, s ≥ 1. Indeed, the value m = 0 in (7) gives (6) at once. For the converse, remark that in (7) we can take on the right side ε1(Bs + · · · + Bs+n), so here we consider (7) in this new form. Suppose (6) is valid. Given ε1 and m, take ε = ε1/(m + 1) and the corresponding value n0 from (6). Take N = n0 + m. Then for n ≥ N and 0 ≤ k ≤ m we have n − k ≥ n0, so Bs+n−k < ε(Bs + · · · + Bs+n−k) < ε(Bs + · · · + Bs+n). Summing these inequalities, we obtain a new form of (7). It follows that the negation of the main condition can be written as (8) ∃ε ∃m : ∀N ∃n > N : Also, (6) is equivalent to s+n Xs+n−m Bk > ε s+n−m−1 Xs Bk for s = sj ↑ ∞. (9) ∀ε ∃m, n0, s0 : Bs+n < ε(Bs+n−m + · · · + Bs+n−1) for n ≥ n0, s ≥ s0. Indeed, comparison of right sides of inequalities shows that (9) implies (6). Conversely, given ε, take n0 such that (6) is valid with ε/(1 + ε) instead of ε. Take m = n0. Then for n ≥ n0, s ≥ s0 we have s = s + n − m ≥ s0 and, by (6), Bs+n = Bs+m < ε 1+ε (Bs + · · · Bs+m), which is (9). We will use a "geometric" version of (9) in terms of (δk) (10) ∀M ∃m, n0, s0 : δs+n−1 δ2 s+n−2 · · · δ2m−1 s+n−m < δM s+n for n ≥ n0, s ≥ s0. 4. Extension operator for E(K(γ)) Here, as in [2], we use the method of local Newton interpolations. Let K be shorthand for K(γ). We fix a nondecreasing sequence of natural numbers (ns)∞ s=0 with ns ≥ 2 and ns → ∞. Given function f on K, we interpolate f at 2n0 points that are chosen by the rule of increase of the type on the whole set. A half of points are located on K ∩ I1,1. We continue interpolation on this set up to the degree 2n1. Separately we do the same on K ∩ I2,1. Continuing in this fashion, we interpolate f with higher and higher degrees on smaller and smaller basic intervals. At each step the additional points are chosen by the rule of increase of the type. Interpolation on Ij,s does not affect other intervals of the same level due to the following function. MITYAGIN'S EXTENSION PROBLEM. PROGRESS REPORT 7 Let t > 0 and a compact set E on the line be given. Then u(·, t, E) is a C ∞− function with the properties: u(·, t, E) ≡ 1 on E, u(x, t, E) = 0 for dist(x, E) > t and supx∈K u(p) xp (x, t, K) ≤ cp t−p, where the constant cp depends only on p. Let cp ր . Given N + 1 points (zk)N +1 ωk(x) = ΩN +1(x) (x−zk)Ω′ k=1 on K ∩ Ij,s let LN (f, x, Ij,s) =PN +1 k=1 (x − zk). N +1(zk) with ΩN +1(x) =QN +1 Let Ns = 2ns − 1 and Ms = 2ns−1−1 − 1 for s ≥ 1, M0 = 1. Then, for fixed s, we take Ms + 1 ≤ N ≤ Ns, so 2n ≤ N < 2n+1 with n ∈ {ns−1 − 1, · · · , ns − 1}. For such N and s we take tN := δs+n. Let, in addition, 1 ≤ j ≤ 2s be fixed. Then we choose N + 1 points on the interval Ij,s by the rule of increase of the type and consider for given f k=1 f (zk) ωk(x), where AN,j,s := [LN (f, x, Ij,s) − LN −1(f, x, Ij,s)] u(x, tN , Ij,s ∩ K). We call Aj,s(f, x) := PNs N =Ms+1 AN,j,s the accumulation sum. The last term here corresponds to the interpolation on Ij,s at 2ns points. In order to continue interpolation on subintervals of Ij,s, let us consider the transition sum Tk,s(f, x) := [LMs+1(f, x, Ik,s+1) − LNs(f, x, Ij, s)] u(x, δs+ns−1, Ik, s+1 ∩ K), where we suppose 1 ≤ k ≤ 2s+1, j = [ k+1 2 ] and Ij, s ⊃ Ik,s+1 ∪ Ii,s+1. As above, we represent the difference in brackets in the telescoping form: [LMs+1 − LNs] = − 2ns −1 XN =2ns −1 [LN (f, x, Ij, s) − LN −1(f, x, Ij, s)]. Here, the interpolating set Z for LN consists of Ms+1 + 1 points of Ys+ns−1 ∩ Ik,s+1 and N − Ms+1 points, chosen by the rule of increase of the type on Ii,s+1. The second parameter of u is smaller than the mesh size of Z, so Tk,s(f, x) 6= 0 only near Ik,s+1. Consider a linear operator W (f, ·) = LM0(f, ·, I1, 0) u(·, 1, K) + ∞ 2s Xs=0(cid:2) Xj=1 Aj,s(f, ·) + 2s+1 Xk=1 Tk,s(f, ·)(cid:3). We remark at the outset that, for fixed x ∈ R and s, because of the choice of parameters for the function u, at most one value Aj,s does not vanish. The same is valid for Tk,s. Let us show that W extends functions from E(K), provided a suitable choice of (ns)∞ s=0. Define n0 = n1 = 2 and ns = [log2 log 1 δs ] for s ≥ 2. Then ns ≤ ns+1 and (11) 1 2 log 1 δs < 2ns ≤ log 1 δs for s ≥ 2. Lemma 4.1. Let (ns)∞ we have W (f, x) = f (x). s=0 be given as above. Then for any f ∈ E(K(γ)) and x ∈ K(γ) Proof. Let us fix a natural number q with q > 2 + log(8C0/7), where C0 is defined in (3). By the telescoping effect, (12) W (f, x) = lim s→∞ LMs(f, x, Ij,s), 8 ALEXANDER GONCHAROV AND ZELIHA URAL where j = j(s, x) is chosen in a such way that x ∈ Ij,s. As in [EvI], 2n (13) LMs(f, x, Ij,s) − f (x) ≤ f q x − zk q ωk(x) . Xk=1 Here n is shorthand for ns−1 − 1 and s is such that Ms = 2n − 1 > q. The interpolating set (zk)2n k=1 for LMs consists of all points of the type ≤ s + n − 1 on Ij,s. Given point x, we consider the chain of basic intervals containing it: x ∈ Ijn,s+n ⊂ · · · ⊂ Ij1,s+1 ⊂ Ij,s. We see that Ijn,s+n contains one interpolating point, Ijn−1,s+n−1 \ Ijn,s+n does one more zi, Ijn−2,s+n−2 \ Ijn−1,s+n−1 contains two such points, etc. Thus, for fixed k, we get 2n Yi=1,i6=k x − zk q x − zi ≤ l q−1 j,s · ljn,s+n · ljn−1,s+n−1 · l2 jn−2,s+n−2 · · · l 2 n−1 j,s . By (3), this does not exceed C 2n+q−1 On the other hand, by a similar argument, for the denominator of ωk(x) we have s+n−2 · · · δ2n−2 s+1 δ2n−1+q−1 δs+n δs+n−1 δ2 . 0 s zk − z1 · · · zk − zk−1 · zk − zk+1 · · · zk − z2n ≥ lqn−1,s+n−1 · h2 qn−2,s+n−2 · · · h 2 n−1 j,s for some indices qn−1, qn−2, · · · . The last product exceeds (7/8)2n−2δs+n−1 δ2 by (4). It follows that s+n−2 · · · δ2n−1 s , LHS of (13) ≤ f q 2n C q−1 0 (8C0/7) 2 n δs+n δq−1 s . The expression on the right side approaches zero as s → ∞. Indeed, 2n < log(1/δs−1), by (11), and 2n(8C0/7) 2 n δq−1 s < 1 due to the choice of q. Thus the limit in (12) exists and equals f (x). (cid:3) 5. Extension property of weakly equilibrium Cantor-type sets We need two more lemmas. Lemma 5.1. Let γ satisfy (2), q = 2m, r = 2n with m < n and Z = (zk)r points of the type ≤ s + n − 1 on I1,s for some s ∈ Z+. Let f (x) = Qr for x ∈ K(γ) ∩ I1,s and f = 0 on K(γ) \ I1,s. Then f 0,K(γ) ≤ C r n+s−2 · · · δ2n−1 δ2 , f (q)(0) ≥ q! · (7/8)r−q · δ2m n+s−m−1 · · · δ2n−1 and f r ≤ 2 · r!. k=1 be all k=1(x − zk) 0 · δn+s · δn+s−1 · s s Proof. Fix x that realizes f 0,K(γ) and a chain of basic intervals containing this point: x ∈ Ij0,n+s ⊂ Ij1,n+s−1 ⊂ · · · ⊂ Ijn,s = I1,s. Arguing as in Lemma 4.1, we see that f 0,K(γ) ≤ lj0,n+s · lj1,n+s−1 · l2 , which, by (3), gives the desired bound. j2,n+s−2 · · · l2n−1 In order to estimate f (q)(0), let us remark that f (q)(x) is a sum of(cid:0)r is g(x) := Qr the location of points from Z, we get g(0) = Qr product has a coefficient q! and consists of r − q terms (x − zk). One of these products k=q+1(x − zik ), where zi1 < zi2 < · · · < zir . All products are nonnegative at x = 0, since r − q is even. From here, f (q)(0) ≥ q! · g(0). Taking into account 1,s > (cid:3) k=q+1 zik > h2m , by (4). The bound of kf kr is evident. q(cid:1) products, each 1,n+s−m−1 · · · h2n−1 n+s−m−1 · · · δ2n−1 (7/8)r−q · δ2m 1,s s MITYAGIN'S EXTENSION PROBLEM. PROGRESS REPORT 9 In the next Lemma, for given 2n ≤ N < 2n+1, we consider ΩN (x) = QN k=1(x − zk) k=1, where the points are chosen on Ij,s by the rule of increase of the with ZN = (zk)N type. Let u(x) = u(x, δs+n, Ij,s ∩ K(γ)) and, as above, di(x, ZN ) := x − zki ր . Lemma 5.2. The bound (ΩN · u)(p)(x) ≤ 2p (C0 + 1) cp δ−p+1 k=2 dk(x, ZN ) is valid for each p < N and x ∈ R. Proof. By Leibnitz's formula, (ΩN · u)(p)(x) ≤ Pp increases, we have Ω(i) N (x) ≤ N ! i=0(cid:0)p k=i+1 dk(x, ZN ). This gives N (x) cp−iδ−p+i s+n . Since dk s+n N p QN i(cid:1) Ω(i) Yk=i+1 N dk(x, ZN ). (14) (ΩN · u)(p)(x) ≤ 2p cp δ−p s+n · max 0≤i≤p (N δs+n)i (N −i)!QN The set ZN consists of 2n endpoints of subintervals of the level s+n−1 covered by Ij,s and N −2n points of the type s+n. Here, dist(x, Ij,s ∩K) = x−x0 ≤ δs+n for some x0. Let x0 ∈ Ii,s+n ⊂ Im,s+n−1. Then Im,s+n−1 contains from 2 to 4 points of ZN . In all cases, d1(x, ZN ) ≤ li,s+n + δs+n ≤ (C0 + 1)δs+n, by (3). Also, δs+n/2 ≤ d2 ≤ (C0 + 1)δs+n−1. Here the lower bound corresponds to the case #(Ii,s+n ∩ ZN ) = 2, whereas the upper bound deals with #(Im,s+n−1 ∩ ZN ) = 2. Similarly, d3 ≥ hm,s+n−1 − δs+n. From (4) and (2) it follows that d3 ≥ 7/8 δs+n−1 − δs+n ≥ 27 δs+n. This gives δi−1 s+ndi+1 · · · dN ≤ (C0 + 1)d2 · · · dN for 0 ≤ i ≤ p and, by (14), the lemma follows. (cid:3) We can now formulate our main result. Theorem 5.3. Suppose γ satisfies (2). Then K(γ) has the extension property if and only if (5) is valid. Proof. Recall that the extension property of a set is equivalent to the condition (DN) of the corresponding Whitney space. Due to L. Frerick [Fr, Prop. 3.8], E(K) satisfies (DN) if and only if for any ε > 0 and for any q ∈ N there exist r ∈ N and C > 0 such that · 1+ε r. Hence, in order to prove that (5) is necessary for EP of K(γ), we can show that (8) implies the lack of (DN) for E(K(γ)), that is there exist ε > 0 and q such that for any r ∈ N one can find a sequence (fj) ⊂ E(K(γ)) with q ≤ C · 0,K · ε fj1+ε q fj−1 0,K(γ) fj− ε r → ∞ as j → ∞. Let us fix ε and m from the condition (8) and take q = 2m. For each fixed large r (clearly, we can take it in the form r = 2n) and sj defined by (8), we consider the function fj given in Lemma 5.1 for s = sj. Then C fj1+ε r ≥ (δn+s · δn+s−1 · δ2 n+s−m−1 · · · δ2n−1 n+s−2 · · · δ2m−1 n+s−m)−1(δ2m 0,K(γ) fj− ε fj−1 )ε, s q where C does not depend on j. The right side here goes to infinity. Indeed, its logarithm is 2n+s {2Bn+s + Bn+s−1 + · · · + Bn+s−m − ε[Bn+s−m−1 + · · · + Bs]} and the expression in braces exceeds Bn+s by (8). Therefore the whole value exceeds 2n+sBn+s = 1 , which goes to infinity when s = sj increases. Thus, EP of K(γ) implies (5). 2 log 1 δs+n For the converse, we consider the extension operator W from Section 5, where (ns) are chosen as in (11). We proceed to show that W is bounded provided (10). Let us fix any natural number p. This p and C1 from Lemma A define M = 2p + 2 + log(2C1). 10 ALEXANDER GONCHAROV AND ZELIHA URAL We fix m ∈ N that corresponds to M it the sense of (10). Let q = 2m − 1 and r = r(q) be defined by (1). We will show that the bound (W (f, x))(p) ≤ C f r is valid for some constant C = C(p) and all f ∈ E(K), x ∈ R. Given f and x, let us consider terms of accumulation sums. For fixed s ∈ N we choose j ≤ 2s such that x ∈ Ij,s. Fix N with 2n ≤ N < 2n+1 for ns−1 − 1 ≤ n ≤ ns − 1, so Ms + 1 ≤ N ≤ Ns. For large enough s the value N exceeds p and q. As in Lemma A, let Z = (zk)N +1 k=1 = ZN ∪ {zN +1}. By Newton's representation of interpolating operator in terms of divided differences, we have AN,j,s(f, x) = [z1, · · · , zN +1]f · ΩN (x) u(x), where ΩN and u are taken as in Lemma 5.2. We aim to show that (15) Ns A(p) N,j,s(f, x) ≤ s−2 f r for large s. This gives convergence of the accumulation sums. Since divided differences are symmetric in their arguments, we can use (4) from [2] : (16) [z1, · · · , zN +1]f ≤ 2N − q f q (Π(J0))−1, where Π(J0) = min1≤j≤N +1−qΠ(J) for Π(J) defined in Lemma B. Fix z ∈ J0 that corresponds to this set in the sense of Lemma B. Applying Lemma 5.2 and Lemma A for z = z yields N +1 (ΩN · u)(p)(x) ≤ C δ−p s+n N p C N 1 dk(z, Z) with C = 2p(C0 + 1) cp. Yk=2 Recall that the set Z includes all points of the type ≤ s + n − 1 on Ij,s and N − 2n k=2 dk(z, Z) if we will consider only distances from z to points from Ys+n−1 ∩ Ij,s. Arguing as in Lemma 4.1, s+n−m. We observe that d1(z, Z) = 0 is not points of the type s + n. We can only enlarge the product Qq+1 we getQq+1 included into the product on the left side. By (10), Qq+1 In order to get (15), it is enough to show that k=2 dk(z, Z) ≤ C q k=2 dk(z, Z) ≤ C q 0 δM s+n. 0δs+n−1 δ2 s+n−2 · · · δ2m−1 (17) s2 Ns N p (2C1)N δM −p s+n → 0 as s → ∞. . Also, (2C1)N < δ− log(2C1) Here, by (11), Ns N p < 2ns(p+1) ≤ log(1/δs)p+1 < δ−p−1 . Clearly, we can replace δs+n in (17) by δs. Then, because of the choice of M, the product in (17) does not exceed s2 δs, which approaches 0 as s → ∞, since, by (2), δs ≤ 32−s. s s Similar arguments are used for terms of the transition sums. (cid:3) On the other hand, (16) and Lemma B for J0 give [z1, · · · , zN +1]f ≤ 2N − q f q Combining these we see that N +1 Yk=q+2 d−1 k (z, Z). q+1 A(p) N,j,s(f, x) ≤ C f q δ−p s+n N p (2C1)N dk(z, Z). Yk=2 MITYAGIN'S EXTENSION PROBLEM. PROGRESS REPORT 11 6. Extension of basis elements Bases in the spaces E(K(γ)) 7. Two examples First we consider regular sequences (Bk)∞ (Bk)∞ Recall that (Bk)∞ k=1 is regular if, for some k0, both sequences (Bk)∞ k=k0 and (βk)∞ k=1 has subexponential growth if βk → 0 as k → ∞. k=1. Let βk = (log Bk)/k. We say that k=k0 are monotone. k = k−a, γ(2) k = a−k, γ(3) For example, given a > 1, let γ(1) k = exp(−ak) for large enough k. Then γ(j) for 1 ≤ j ≤ 3 generate regular B(j) with B(1) k ∼ 2−k−2 k2 log a, B(3) k → − log(a/2), so B(j) are not of subexponential growth, except B(3) for a = 2. We see that (5) is valid in the first two cases and in the third case with a ≤ 2. k ∼ (a/2)k+1/(a−1). Here, β(1) k ∼ 2−k−1 a k log k, B(2) k ր − log 2 and β(3) k , β(2) More generally, (5) is valid for each monotone convergent (Bk)∞ k=1. Indeed, if Bk ց B ≥ 0, then LHS of (5) does not exceed (n + 1)−1. If Bk ր B, then we take s0 with Bs > B/2 for s ≥ s0. Then Bs+· · · Bs+n ≥ (n+1)B/2 and LHS of (5) < 2(n+1)−1. This covers the case of regular sequences (Bk)∞ k=1 when βk are negative. Let us show that (5) is valid as well for divergent regular sequences (Bk)∞ k=1 of subexponential growth. Theorem 7.1. Let (Bk)∞ and only if (Bk)∞ k=1 be regular with positive values of βk. Then (5) is valid if k=1 has subexponential growth. Proof. A regular sequence (Bk)∞ k=1 is not of subexponential growth, provided βk > 0, in the following three cases: βk ր β < ∞, βk ր ∞ and βk ց ε0 > 0. We aim to show that (5) is not valid under the circumstances. In the first case, given s and n, let b = exp βs+n. Then b − 1 ≥ exp β1 − 1 > β1 > 0 k=s Bk < bs+n+1/(b − 1) as Bk = exp(kβk) ≤ bk for such k. k=s Bk > (b − 1)/b > β1/ exp β, which contradicts (5). and b ≤ exp β. Here, Ps+n Therefore, Bs+n/Ps+n If βk ր ∞ then, by the same argument, Bs+n/Ps+n where s0 is fixed with exp βs0 > 2. k=s Bk > (b−1)/b > 1/2 for s ≥ s0, Suppose βk ց ε0. We fix indices s1 < s2 < · · · such that the intervals Ij connecting points (sj, βsj ) and (sj+1, βsj+1) form a convex envelope of the set (k, βk) on the plane. We start from s1 = max{s : βs = β1}. If sj is chosen, then we take sj+1 with he property: for each k with sj ≤ k ≤ sj+1 the point (k, βk) is not over Ij. At any step we can take the next value so large that the slopes of Ij increases to zero. In addition, given sj, we take sj+1 such that (18) (4 − 2 sj/sj+1)βsj+1 ≥ (3 − sj/sj+1)βsj , which is possible as βk decreases to a positive limit. For fixed j, we take s = sj and s+n = sj+1. Let βk = ak +b with a = −(βs −βs+n)/n and b = βs + (βs − βs+n) s/n for s ≤ k ≤ s + n, so the points (k, βk) are located just on the interval Ij. Also, let g(x) = ax2 + bx and Bk = exp g(k) = exp(k βk) on [s, s + n]. Of course, Bs = Bs and Bs+n = Bs+n. It is easy to check that the function g increases on this interval. Hence, Ps+n Bk < R s+n Ps+n [2a(n + s) + b]−1 − g(s) · [2as + b]−1 + 2aR s+n g(x) dx + Bs+n. By integration by parts, R s+n k=s Bk ≤ g(x) dx = g(n + s) · g(x)(2ax + b)−2dx. We neglect the last k=s s s s 12 ALEXANDER GONCHAROV AND ZELIHA URAL s term, as a < 0, and the second term, as 2as + b = g′(s) > 0. Also, 2a(n + s) + b = g(x) dx < 2 Bs+n/ε0 (2 + s/n)βs+n − (1 + s/n)βs ≥ βs/2 ≥ ε0/2, by (18). Hence R s+n and Bs+n/Ps+n k=s Bk > ε0/(2 + ε0), so (5) is not valid. growth of (Bk)∞ We proceed to show that (5) is valid for βk ց 0, that is in the case of subexponential k=s Bk from below. Let k=1. Here, for fixed large s and n, we estimatePs+n b = exp βs+n. Then Bk ≥ bk for s ≤ k ≤ s + n. Therefore, Bs+n/Ps+n If bn < 2 for the given s and n then bn+1 − 1 > (n + 1)βs+n. On the other hand, exp βs+n − 1 < 2 βs+n for βs+n < 1. Thus the fraction above does not exceed 4/(n + 1). Otherwise, bn ≥ 2 and bn < 2(bn+1 − 1). Here the fraction does not exceed 4 βs+n. It k=s Bk ≤ max{4/(n + 1), 4 βn}, which is the desired conclusion. (cid:3) k=s Bk ≤ bn(b−1) bn+1−1 . follows that Bs+n/Ps+n Our next objective is to consider irregular sequences (Bk)∞ k=1 (compare with Ex.6 in [10]). Given two sequences, (kj)∞ j=1 with εj ց 0, let γk = (k + 5)−2 for k 6= kj and γkj = (kj + 5)−2εj. Then γ satisfies (2) with δk = (5!/(k + 5)!)2 ε1ε2 · · · εj for kj ≤ k < kj+1. Let Aj := log . We will consider only sequences with the property j=1 ⊂ N with kj+1 − kj ր ∞ and (εj)∞ ε1ε2···εj 1 (19) j+1 · A−1 k2 j → 0 as j → ∞. Provided this condition, Bk = 2−k log (k+5)! In addition, an easy computation shows that for large j, 5! + 2−k−1 Aj ∼ 2−k−1 Aj for kj ≤ k < kj+1. (20) Bkj + Bkj+1 + · · · + Bkj+1−1 < 3 Bkj . Now we can construct different examples of compact sets K(γ) without extension property. Example 2. Let Aj = 2kj , so εj = exp(−2kj + 2kj−1) for j ≥ 2 and ε1 = exp(−2k1). In this case, (19) is valid under mild restriction 2−kj k2 j+1 → 0 as j → ∞. Let us take s = kj, n = kj+1 − kj. Then Bs+n > 2−kj+1−1 Aj+1 = 1/2 and, by (20), Bs + · · · + Bs+n−1 < 3 Bs < 4 · 2−kj−1 Aj = 2. This gives (8) with ε = 1/4 and m = 0. 8. Extension Property of K(γ) and Hausdorff measures From now on, h is a dimension function, which means that h : (0, T ) → (0, ∞) is continuous, nondecreasing and h(t) → 0 as t → 0. The h−Hausdorff content of E ⊂ R is defined as and the h−Hausdorff measure of E is Mh(E) = inf{X h(Gi) : E ⊂ ∪Gi} Λh(E) = lim inf δ→0 {X h(Gi) : E ⊂ ∪Gi, Gi ≤ δ}. Here we consider at most countable coverings of E by intervals (open or closed). It is easily seen that Mh(E) = 0 if and only if Λh(E) = 0. We write h1 ≺ h2 if h1(t) = o(h2(t)) as t → 0. Let h1 ≈ h2 if C −1h1(t) ≤ h2 ≤ Ch1(t) for some constant C ≥ 1 and 0 < t ≤ t0 < T. We will denote by h0 the function h0(t) = (log 1 t )−1 with 0 < t < 1, which defines the logarithmic measure of sets. MITYAGIN'S EXTENSION PROBLEM. PROGRESS REPORT 13 A set E is called dimensional if there is at least one dimension function h that makes E an h−set, that is 0 < Λh(E) < ∞. In our case, the set K(γ) is dimensional. In [1], following Nevanlinna [Nev], the corresponding dimension function was presented. Let η(δk) = k for k ∈ Z+ with δ0 := 1 and η(t) = k + log δk for δk+1 < t < δk. Then h(t) := 2−η(t) for 0 < t ≤ 1. Clearly, h(δk) = 2−k. t / log δk δk+1 Proposition 8.1. Let γ satisfy (2) and h be defined as above. Then Λh(K(γ)) = 1. Proof. Take t = C0 δk, where C0 is given in (3). Then δk < t = C0 γk δk−1 < δk−1 for large enough k. Here, η(t) = k − log C0/ log(1/γk) and h(t) = 2−k ak with ak := exp log C0·log 2 log(1/γk ) . Since γk → 0, given ε > 0, there is k0 such that ak < 1 + ε for k ≥ k0. j=1 h(lj,k) < 2k h(t) < 1 + ε provided that j=1 h(lj,k) for each k. Since ε is arbitrary, we get From (3) it follows that 1 = 2k h(δk) < P2k k ≥ k0. Of course, Λh(K(γ)) ≤ P2k Let us show that Λh(K(γ)) ≥ 1. Fix ε > 0 and choose k0 such that Λh(K(γ)) ≤ 1. (21) ε log 1/γk > − log 2 · log(1 − 4γk) for k ≥ k0. This can be done as γk → 0. Take any open covering ∪Gi of K(γ). Given ε, we can consider only coverings with Gi < δk0 for each i. We choose a finite subcover ∪N i=1Gi of K(γ). Fix i ≤ N and k with δk+1 < Gi ≤ δk. Recall that, for each j ≤ 2k, we have hj,k > (1 − 4γk+1)lj,k. Therefore the distance between any two basic intervals from Ek+1 exceeds (1 − 4γk+1)δk. If Gi < (1 − 4γk+1)δk then Gi can intersect at most one interval from Ek+1. In this case we can consider only Gi ≤ max1≤j≤2j+1 lj,k+1 ≤ C0δk+1, by (3). Thus there are two possibilities: δk+1 < Gi ≤ C0δk+1 or (1 − 4γk+1)δk < Gi ≤ δk. In the first case we have h(Gi) > 2−k−1. Here, Gi intersects at most one interval from Ek+1 and, by construction, at most 2m−k−1 interval from Em for m > k. In turn, in the latter case, h(Gi) > 2−k(1 − ε). Indeed, here, η(Gi) < k − log(1 − 4γk+1)/ log(1/γk+1) and h(Gi) > 2−k a, where a = exp log(1−4γk+1)·log 2 log(1/γk+1) > (1−ε), by (21). Now Gi intersects at most two interval from Ek+1 and at most 2m−k interval from Em. Let us choose m so large that each basic interval from Em belongs to some Gi, perhaps not to unique. We decompose all intervals from Em into two groups corresponding to i 2m−k < the cases considered above. Counting intervals gives 2m ≤ P′ 2m[P′ i h(Gi)(1 − ε)−1]. From this we see that Pi h(Gi) > 1 − ε, which is the desired conclusion, as ε and (Gi) here are arbitrary. i h(Gi) +P′′ i 2m−k−1 +P′′ (cid:3) The same reasoning applies to a part of K(γ) on each basic interval. Corollary 8.2. Let γ and h be as in Proposition above, k ∈ N, 1 ≤ j ≤ 2k. Then Λh(K(γ) ∩ Ij,k) = 2−k. Suppose, in addition, that K(γ) is not polar. Then, by Corollary 3.2 in [1], µK(γ)(Ij,k) = 2−k, so the values of µK(γ) and the restriction of Λh on K(γ) coincide on each basic interval. From here, by Lemma 3.3 in [1], these measures are equal on K(γ). Thus, a non-polar set K(γ) satisfying (2) is indeed equilibrium Cantor-type set if we accept for definition of this concept the condition µK = ΛhK(γ), which is more natural than the 14 ALEXANDER GONCHAROV AND ZELIHA URAL definition suggested in in [WE], Section 6. We recall that there is no complete characterization of polarity of compact sets in terms of Hausdorff measures, see e.g Chapter V in [14]. On the one hand, a set is polar if its logarithmic measure is finite. This defines a zone Zpol in the scale of growth rate of dimension functions consisting of h with lim inf t→0 h(t)/h0(t) > 0. If h ∈ Zpol and Λh(K) < ∞ then Cap(K) = 0. On the other hand, functions with R0 h(t)/t dt < ∞ form a nonpolar zone Znp : if h ∈ Znp and Λh(K) > 0 then Cap(K) > 0. But, by Ursell [21], the remainder makes up a zone Zu of uncertainty. One can take two functions in this zone with h2 ≺ h1 and sets K1, K2, where Kj is a hj−set, such that K2 is polar, K1 is not, though in the sense of Hausdorff measure the set K2 is larger than K1. Indeed, Λh2(K2) > 0, but Λh2(K1) = 0 or Λh1(K2) = ∞, but Λh1(K1) < ∞. Let us show that a similar circumstance is valid with the extension property. Proposition 8.3. There are two dimension functions h2 ≺ h1 and two sets K1, K2, where Kj is an hj−set for j ∈ {1, 2}, such that the smaller set K1 has the extension property, whereas the larger set K2 does not have. Proof. Take K1 from Example 1. Let us show that the corresponding function h1 = 2−η1 is equivalent to h0. It is enough to find C > 0 such that η0(t) − C ≤ η1(t) ≤ η0(t) + C for small t. Here, η0(t) = (log log 1/t)/ log 2, so h0(t) = 2−η0(t). For the set K1 we have δk = exp(−2k+1B) and η0(δk) = k + log 2B/ log 2. If δk+1 < t ≤ δk for some k, then k ≤ η1(t) < k + 1 and k + log 2B/ log 2 ≤ η0(t) < k + 1 + log 2B/ log 2, which gives h1 ≈ h0. j=1 satisfies 2−kj 2j k2 In turn, let K2 be as in Example 2 with Aj = 2kj 2−j and εj = exp(−Aj + Aj+1) for j ≥ 2. Here we suppose that (kj)∞ j+1 → 0 as j → ∞. Then (19) and (20) are valid, which, as in Example 2, gives the lack of the extension property for K2. Let us show that h2 ≺ h0. It is enough to check that η2(t) − η0(t) → ∞ as t → 0. Let δk < t ≤ δk−1 with kj ≤ k < kj+1 for large enough j. Then log 1/δk = 2 log((k + 5)!/5!) + Aj < 2 Aj and η0(t) < η0(δk) < kj + 1 − j. On the other hand, η2(t) ≥ η2(δk−1) = k − 1 ≥ kj − 1. Therefore, η2(t) − η0(t) > j − 2, which completes the proof. (cid:3) One can suppose that, for the considered family of sets, the scale of growth rate of dimension functions can be decomposed as above into three zones. If K(γ) is an h−set for a function h with moderate growth then the set has EP . If the corresponding function h is large enough, then EP fails. Proposition above shows that the zone of uncertainty here is not empty. We see that h = h0 is not the largest function which allows EP for h−sets K(γ). If, as in the regular case, we take Bk ր ∞ of subexponential growth, then δk = exp(−2k+1Bk) and h0(δk) = 2−k−1B−1 k , which is essentially smaller than h(δk) = 2−k for the corresponding function h. MITYAGIN'S EXTENSION PROBLEM. PROGRESS REPORT 15 Example 3. Let log(m) t denote the m−th iteration log · · · log t for large enough t. The sequence Bk = exp(k/ log(m) k) has subexponential growth. Then the correspond- ing sequence (γk)∞ k=1 satisfies (2), as for large k we have γk = δk/δk−1 < exp(−2kBk) < exp(−2k) and for the previous k we can take γk = 1/32. By Theorem 7.1, the set K(γ) has EP . Let us find a dimension function h that corresponds to this set. We will search it in the form h(t) = hα(t) (t). Let t = δk. Then log 1/t = 2k+1Bk, so k ∼ (log log 1/t)/ log 2. On the other hand, h(t) = 2−k = (2k+1Bk)−α(t), which gives α(t) ∼ 1 − (log 2 · log(m) k)−1 ∼ 1 − (log 2 · log(m+2) 1/t)−1. Clearly, h ≻ h0. 0 The next Proposition generalizes Example 3. We restrict our attention to strictly increasing functions h of the form h = hα 0 , where α is a monotone function on [0, t0]. Since in the next section we shall be interested in considering of dimension functions exceeding h0, let us suppose that α(t) ≤ 1. Then h ≻ tσ for each fixed σ > 0. In addition we assume that asymptotically (22) h(t) ≤ 2 h(t2), which is valid for typical dimension functions corresponding to the cases a) α(t) = α0 ∈ (0, 1], b) α(t) = α0 + ε(t) with α0 ∈ [0, 1), c) α(t) = 1 − ε(t). Here, (23) ε(t) ց 0 with ε(t) log log 1/t ր ∞ as t ց 0, since for slowly increasing ε we get hα0±ε ≈ hα0 0 . By (22), for the inverse function h−1, we have h−1(τ ) ≤ (h−1(2τ ))2 and h−1 ≺ τ M for M given beforehand. From this, γk = h−1(2−k)/h−1(2−k+1) defines a sequence satisfying (2). We denote the corresponding set by K α(γ). Our aim is to check EP for this set provided regularity of the sequence Bk = 2−k−1 log(1/h−1(2−k)). We see at once that Bk increases. In its turn, βk ց 0 if α0 = 1 in the case (a), βk ր 1/α0 − 1 in (b) and in (a) with α0 < 1. Concerning (c), the monotonicity of βk requires additional rather technical restrictions on ε. At least for ε(t) = εm(t) := (log(m) 1/t)−1 we have βk ց 0. Here, m ≥ 3, as h ≈ h0 for m ∈ {1, 2}. Proposition 8.4. Let K α(γ) be defined by a function h, as above, with a regular sequence (Bk)∞ k=1. Then K α(γ) has the extension property if and only if (cid:18)log 1 h−1(2−k)(cid:19)1/k → 2 as k → ∞. Proof. Let us find h−1 for the case α(t) = 1−ε(t). If h(t) = τ then [1−ε(t)] log log 1/t = log 1/τ. Let us define a function δ by the condition log log 1/t = [1 + δ(τ )] log 1/τ. Then [1 − ε(t)][1 + δ(τ )] = 1, so δ(τ ) ց 0 as τ ց 0. Then t = h−1(τ ) = exp[−(1/τ )1+δ(τ )] and log(1/h−1(2−k)) = 2k(1+δ(2−k )). The k−th root of this expression tends to 2. On the other hand, (Bk)∞ k=1 here has subexponential growth as βk = (δ(2−k) − 1/k) log 2 → 0. By Theorem 7.1, K α(γ) has EP. 16 ALEXANDER GONCHAROV AND ZELIHA URAL Similarly, if α(t) = α0 + ε(t) with 0 < α0 < 1 then h−1(τ ) = exp[−(1/τ )1/α0−δ(τ )]. Here, (log(1/h−1(2−k)))1/k = 2(1/α0−δ(2−k)) 9 2 and βk 9 0, there is no EP. In the case (a), the function δ vanishes. Lastly, α0 = 0 in (b) gives h−1(τ ) = exp[−(1/τ )∆(τ )] with ∆(τ ) ր ∞ as τ ց 0. (cid:3) Here, (log(1/h−1(2−k)))1/k → ∞ and βk → ∞. 9. Extension Property and densities of Hausdorff contents To decide whether a set K has EP, we have to consider a local structure of the most rarefied parts of K. Obviously, such global characteristics as Hausdorff measures or Hausdorff contents cannot be applied in general for this aim. Instead, one can suggest to describe EP in terms of lower densities of Mh or related functions. Given a dimension function h, a compact set K, x ∈ K and r > 0, let ϕh,K(x, r) := Mh(K ∩ B(x, r)) and ϕh,K(r) := inf x∈K ϕh,K(x, r), where B(x, r) = [x−r, x+r]. One can suppose that K has EP if and only if the corresponding function ϕh,K is not very small, in a sense, as r → 0. Essentially, this is similar to analysis of the lower density of the Hausdorff content, which can be defined as φh(K) := lim inf r→0 inf x∈K . Indeed, Mh(B(x, r)) = Mh(K∩B(x,r)) Mh(B(x,r)) h(2r) for h with h(t) ≻ t and the expression above is lim inf r→0 ϕh,K (r) h(2r) . In order to distinguish EP by means of φh, we have to consider large enough di- mension functions h. Indeed, if for some h1 with h1 ≻ h there exists h1−set K1 with EP , then h cannot be used for this aim, because Λh(K1) = 0 implies Mh(K1) = 0 and the corresponding density vanishes contrary to our expectations. Therefore, we can consider only functions exceeding h0. We remark that Λh−analogs of ϕh,K or φh cannot be applied in general for distin- guishing EP, since for fat sets (K = Int(K)) we have Λh(K ∩ B(x, r)) = ∞ provided h(t) ≻ t. Interestingly, it turns out that the lower density φh can be used to characterize EP for the family of compact sets considered in [6]. with Qk ≥ 2, let K = {0} ∪S∞ Example 4. Given two sequences bk ց 0 (for brevity, we take bk = e−k) and Qk ր k . In what follows we will consider two cases: Qk ≤ Q with some Q and Qk ր ∞ with Qk < log k for large k. By Theorem 4 in [6], K has the extension property in the first case and does not have it for unbounded (Qk). k=1 Ik, where Ik = [ak, bk], Ik = bQk In the next lemma we consider concave dimension functions h = hα 0 for the cases (a), (b), as above, and for more general c ′) α(t) = α0 − ε(t) with α0 ∈ (0, 1]. We suppose now that ε is a monotone differentiable function on [0, t0] with 0 < ε(t) < 1 − α0 in (b) and 0 < ε(t) < α0/2 in (c ′). As before, we assume (23). A direct computation shows that (24) h′(t) < h(t) h0(t) α(t)/t for the cases (a), (b) and h′(t) < h(t) h0(t)/t for (c ′). MITYAGIN'S EXTENSION PROBLEM. PROGRESS REPORT 17 k=nIk) = h(bn). This means that the covering of the set ∪∞ Lemma 9.1. Suppose intervals Ik are given as in Example 4 and n is large enough. Then Mh(∪∞ k=nIk by the interval [0, bn] is optimal in the sense of definition of Mh. Proof. Let us fix any open covering of K, choose its finite subcovering ∪M i=1Gi and enumerate sets Gi from left to right. We can suppose that G1 covers ∪∞ k=N Ik for some N ≥ n. Indeed, if G1 covers as well some part of IN −1, then other part of IN −1 is covered by G2. In this case, association of G1 and G1 into one interval will give better covering, since h(b) ≤ h(x) + h(b − x) for 0 ≤ x ≤ b, by concavity of h. For the same reason, we suppose that each Gi covers entire number of Ik. After this we reduce each Gi to the minimal closed interval Fi containing the same intervals Ik. Thus, F1 = [0, bN ] and F2 = [aN −1, bq] with some N − 1 ≤ q ≤ n. Our aim is to show that (25) h(bq) < h(bN ) + h(bq − aN −1), so replacing F1 ∪ F2 with [0, bq] is preferable. We use the mean value theorem and the decrease of h′. Note that h(bk) = k−α(bk). Consider first the value q = N − 1. We will show h(bN −1) − h(bN ) < h(IN −1). In the cases (a), (b), by (24), LHS < h′(bN )e−N (e − 1) < N −1−α(bN )α(bN )(e − 1). On the other hand, h(IN −1) = [QN −1(N − 1)]−α(IN −1). Here, α(IN −1) < α(bN ), so we reduce the desired inequality to (QN −1/N)α(bN )α(bN )(e − 1) < 1. It is valid, since for α0 > 0 the first term on the left goes to zero, whereas for α0 = 0 in (b) we have α(bN ) = ε(bN ) → 0 as N → ∞. Similarly, in the case (c ′) the inequality [QN −1(N −1)]α0−ε(IN −1)(e−1) < N 1+α0−ε(bN ) is valid, as is easy to check. Suppose now that q ≤ N − 2. We write (25) as h(bq) − h(bq − aN −1) < h(bN ). Here, in all cases, by (24), LHS < h′(bq−aN −1) aN −1 < h(bq)h0(bq) aN −1 bq−aN −1 , where the last fraction does not exceed bN −1 . On the other hand, h(bN ) ≥ N −1 as α(bN ) ≤ 1. bq−bN −1 Hence it is enough to show that N < (eN −q−1 − 1) q1+α(bq). We neglect α(bq) and notice that (eN −q−1 − 1) q ≥ (e − 1)(N − 2), which completes the proof of (25). Continuing in this manner, we see that h(bn) ≤PM Corollary 9.2. Suppose bn+1 ≤ r ≤ bn − bn+1. Then ϕh,K(r) = h(In). i=1 h(Fi). (cid:3) Proof. Clearly, ϕh,K(x, r) = h(In) for each x ∈ In. If x ∈ K ∩ [0, bn+1] then B(x, r) covers all intervals Ik with k ≥ n + 1. By Lemma, ϕh,K(x, r) = h(bn+1) > h(In). Of course, for x ∈ Ik with k < n the value ϕh,K(x, r) also exceed h(In). (cid:3) Remark. The covering of two (or small number of) intervals Ik by one interval is not optimal, since Mh(Ik ∪ Ik+1) = h(Ik) + h(Ik+1) < h(bk − ak+1). We proceed to characterize EP for given compact sets in terms of lower densities φh for h = hα 0 , where α(t) = α0 ∈ (0, 1] or α(t) = α0 ± εm(t) (26) with 0 < α0 < 1 and εm(t) = (log(m) 1/t)−1 for m > 2, so (23) is valid. Proposition 9.3. Let K be from the family of compact sets given in Example 4 and h be as above. Then K has the extension property if and only if φh(K) > 0. 18 ALEXANDER GONCHAROV AND ZELIHA URAL Proof. Suppose first that Qk ≤ Q with some Q, so K has EP . We aim to show ϕh,K (r) h(2r) > 0. Let e−k−1 ≤ r < e−k for some k. Then, as ϕh,K increases, ϕh,K(r) ≥ limr→0 ϕh,K(e−k−1), which is h(Ik) = (k · Qk)−α(Ik), by Corollary 9.2. On the other hand, h(2r) < h(2e−k) = (k − log 2)−α(2e−k). Therefore, ϕh,K(r)/h(2r) > Q−α(Ik) k kα(2e−k)−α(Ik) (1 − log 2/k)−α(2e−k). The first term on the right converges to Q−α0 as k → ∞. The second and the third terms converge to 1. Hence, φh(K) ≥ Q−α0. Besides, this value is achieved in the case Qk = Q by the sequence rk = bk − bk+1. Thus, φh(K) > 0. In addition, for given σ > 0 we have a compact sets K with EP such that 0 < φh(K) < σ. Similar arguments apply to the case Qk ր ∞, when K does not have EP. Here, φh(K) ≤ limk ϕh,K(rk)/h(2rk) for rk as above. By Corollary 9.2, ϕh,K(rk) = h(Ik). Also, h(2rk) > h(e−k). Hence, ϕh,K(rk)/h(2rk) < Q−α0/2 kα(e−k)−α(Ik), which converges to 0 as k increases. (cid:3) k Remark. For this family of sets, the extension property can be characterized as well in terms of the Lebesgue linear measure λ. Let λ(r) := inf x∈K λ(K ∩ [x − r, x + r]). Then K has the extension property if and only if lim inf r→0 λ(r) · r−Q > 0 for some Q. Nevertheless, at least for dimension functions h = hα 0 with α as in (26), there is no general characterization of EP in terms of lower densities φh. In view of Example 3 and the discussion in the beginning of the section, the value α0 = 1 can be omitted from consideration We treat now regular sets K(γ) with δk = exp(−bk). Here, Bk = 2−1(b/2)k. By Theorem 7.1, K(γ) has EP if b = 2 and does not have it for b > 2. Lemma 9.4. For each constants C ≥ 1 and h, as above, there is b > 2 such that h(Cδk) < 2 h(δk+1) for large enough k. This inequality is valid also for b = 2. Proof. In all cases we have h(δk) = b−k·α(δk) and the desired inequality has the form (27) b(k+1)α(δk+1) < 2 (bk − log C)α(Cδk). Suppose α ≡ α0. Then (27) is valid as bα0 < 2 (1−b−k log C) for large k and b = 2+σ with small enough σ. All the more it is valid for b = 2. The same reasoning applies to the case α = α0 + ε(t) with ε ր as ε(δk+1) < ε(Cδk). In the last case α(t) = α0 − εm(t) we use the following simple inequality log(m)(Cx) − log(m)(x) < log C · [log x log(2)(x) · · · log(m−1)(x)]−1, which is valid for all x from the domain of definition of log(m). From this we have k · [ε(Cδk) − ε(δk+1)] → 0 as k → ∞ and (27) can be treated as in the first case. (cid:3) Corollary 9.5. Let k be large enough.Then the covering of each basic interval Ij,k of K(γ) by one interval is better (in the sense of definition of Mh) than covering by two adjacent subintervals. Indeed, by (3), h(lj,k) < h(C0δk) < 2 h(δk+1) < h(l2j−1,k+1) + h(l2j,k+1). MITYAGIN'S EXTENSION PROBLEM. PROGRESS REPORT 19 Remark. It is essential that coverings of a whole basic interval are considered. For example, for the set I1,k∪I3,k+1 we have h(l1,k)+h(l3,k+1) < h(b3,k+1), which corresponds to the covering of the set by one interval. Proposition 9.6. Let h = hα Then φh(K(γ)) = b−α0. 0 with α as in (26) and K(γ) be defined by δk = exp(−bk). Proof. For brevity, we denote here K(γ) by K. Fix x ∈ K. Let x ∈ Ij,k ⊂ Ii,k−1 and C0 δk ≤ r ≤ 7/8 · δk−1. Then, by (3), lj,k ≤ r < hi,k−1 and K ∩ [x − r, x + r] = K ∩ Ij,k. Arguing as in Lemma 9.1, by Lemma 9.4, we get ϕh,K(x, r) = h(lj,k). Therefore, by monotonicity, h(δk) < ϕh,K(x, r) < h(C0δk) for each x ∈ K. We proceed to estimate φh(K) from both sides. Suppose that C0 δk ≤ r ≤ C0 · δk−1 for some k. Then h(δk) < ϕh,K(r) < h(C0δk−1) and h(δk) h(2C0δk−1) < ϕh,K(r) h(2r) < h(C0δk−1) h(2C0δk) . Here, δk = δb k−1. Analysis similar to that in the proof of Lemma 9.4 shows that the first fraction above has the limit b−α0, whereas the last fraction tends to bα0 as k → ∞. Moreover, the value b−α0 can be achieved as limk ϕh,K(rk)/h(2rk) for rk = 7/8·δk−1. (cid:3) Comparison of Propositions 9.3 and 9.6 shows that, for given dimension functions, lower densities of Hausdorff contents cannot be used in general to characterize the extension property. 10. Extension Property and growth of Markov's factors Let Pn denote the set of all holomorphic polynomials of degree at most n. For any infinite compact set K ⊂ C we consider the sequence of Markov's factors Mn(K) = inf{M : P ′0,K ≤ M P 0,K, P ∈ Pn} for n ∈ N. We see that Mn(K) is the norm of the operator of differentiation in the space (Pn, · 0,K). We say that a set K is Markov if the sequence (Mn(K)) is of polynomial growth. This class of sets is of interest to us, since, by W.Ple´sniak [16], any Markov set has EP. On the other hand, there exist non-Markov compact sets with EP ([5], [2]). We guess that there is some extremal growth rate (mn)∞ n=1 with the property: if, for some compact set K, Mn(K)/mn → ∞ as n → ∞ then K does not have EP. The next proposition asserts that here, as above, there is a zone of uncertainty, in which growth rate of Markov's factors is not related with EP. In this sense, it is an analog of Proposition 8.3. Proposition 10.1. There are two sets K1 with EP and K2 without it, such that Mn(K1) grows essentially faster than Mn(K2) as n → ∞. Proof. By Theorem 6 in [we], M2k (K(γ)) ∼ 2/δk. By monotonicity, δ−1 4 δ−1 Example 1, so δ(1) Aj = 2kj . For simplicity, we fix kj = j2 that satisfies (19). Here, δ(2) k < Mn(K(γ)) < k+1 for 2k ≤ n < 2k+1 with large enough k. As in Proposition 8.3, we take K1 from k = exp(−2k+1B) with B > 1. Also, we use K2 from Example 2 with k > k−2k ε1ε2 · · · εj 20 ALEXANDER GONCHAROV AND ZELIHA URAL for kj ≤ k < kj+1. We aim to show that Mn(K2)/Mn(K1) → 0 as n → ∞. Let us fix large n with 2k ≤ n < 2k+1. For this k we fix j with kj ≤ k < kj+1. Then (28) Mn(K2)/Mn(K1) < 4 δ(1) k /δ(2) k+1. Suppose first that k ≤ kj+1 − 2. Then RHS of (28) does not exceed 4 exp[−2k+1B + j+1, 2(k + 1) log(k + 1) + Aj]. The expression in brackets is smaller than 2kj (1 − 2B) + k2 which is (j + 1)4 − (2B − 1) 2j 2, so it tends to −∞ as j → ∞. If k = kj+1−1 then RHS of (28) is smaller than 4 exp[−2kj+1B+2kj+1 log kj+1+Aj+1], (cid:3) which goes to 0, since B > 1. This completes the proof. Existence of a zone of uncertainty (for the extension property) in the scale of growth rate of Markov's factors implicates the problem to find boundaries of this zone. References [1] G. Alpan- A. Goncharov, Two measures on Cantor sets, J. Approx. Theory 186(2014), 28-32. [2] M. Altun-A. Goncharov, A local version of the Pawlucki-Plesniak extension operator, Journal of Approx. Theory, 132(2004), 34-41. [3] L. Frerick, Extension operators for spaces of infinite differentiable Whitney jets, J. Reine Angew. Math. 602 (2007), 123-154. [4] L. Frerick, E. Jorda and J. Wengenroth, Tame linear extension operators for smooth Whitney functions,J. Funct. Anal., 261 (2011), no. 3, 591-603. [5] A. Goncharov, A compact set without Markov's property but with an extension operator for C∞-functions, Studia math., 119(1996), 27-35. [6] A. Goncharov-M. Kocatepe, Isomorphic Classification of the Spaces of Whitney Functions, Michigan Math. J., 44(1997), 555-577. [7] A. Goncharov, Bases in the spaces of C∞-functions on Cantor-type sets, Constr. Approx. 23 (2006), no. 3, 351-360. [8] A. Goncharov, On the geometric characterization of the extension property, Bull. Belg. Math. Soc. Simon Stevin, 14(2007), 513-520. [9] A. Goncharov,Weakly Equilibrium Cantor-Type Sets, Potential Anal., 40(2014),143-161. [10] A. Goncharov, Best exponents in Markov's inequalities, Math. Inequal. Appl. 17(2014), no. 4, 1515-1527. [11] A. Goncharov- Z. Ural, Bases in some spaces of whitney functions, in process... [12] R. Meise- D. Vogt, Introduction to functional analysis, Oxford Graduate Texts in Mathematics, 2. The Clarendon Press, Oxford University Press, New York (1997), x+437 pp. [13] B.S. Mityagin, Approximate dimension and bases in nuclear spaces, Russian Math. Surveys, 16:4(1961), 59-127. [14] R. Nevanlinna, Analytic Functions, Die Grundlehren der mathematischen Wissenschaften, vol. 162, Springer-Verlag, New York, Berlin (1970), viii+373 pp. [15] W. Pawlucki and W. Plesniak,Extension of C∞ functions from sets with polynomial cusps, Studia Math., 88(1988), 279-287. [16] W. Plesniak,Markov's Inequality and the Existence of an Extension Operator for C∞ functions, Jour. of Approx. Theory, 61(1990), 106-117. [17] T. Ransford, Potential theory in the complex plane, London Mathematical Society Student Texts, 28. Cambridge University Press, Cambridge, (1995), x+232 pp. [18] R.T. Seeley, Extension of C∞ functions defined in a half space, Proc. Amer. Math. Soc. 15(1964), 625-626. [19] E.M.Stein, Singular integrals and differentiability properties of functions,Princeton Univ.Press (1970). [20] M. Tidten, Fortsetzungen von C∞- Funktionen, welche auf einer abgeschlossenen Menge in Rn definiert sind, Manuscripta Math., 27(1979), 291-312. MITYAGIN'S EXTENSION PROBLEM. PROGRESS REPORT 21 [21] H. D. Ursell, Note on the Transfinite Diameter, J. London Math. Soc.,s1-13 (1938),34-37. [22] D. Vogt,Sequence space representations of spaces of test functions and distributions in Func- tional Analysis, Holomorphy and Approximation Theory, G.I. Zapata (ed.), Lecture Notes in Pure and Appl. Math., 83(1983), 405-443. [23] H. Whitney,Analytic extensions of differentiable functions defined in closed sets, Trans. Amer. Math. Soc., 36(1934),63-89.
1301.2606	1	1301	2013-01-11T21:10:17	Affine invariant points	[ "math.FA" ]	We answer in the negative a question by Gruenbaum who asked if there exists a finite basis of affine invariant points. We give a positive answer to another question by Gruenbaum about the "size" of the set of all affine invariant points. Related, we show that the set of all convex bodies K, for which the set of affine invariant points is all of n-dimensional Euclidean space, is dense in the set of convex bodies. Crucial to establish these results, are new affine invariant points, not previously considered in the literature.	math.FA	math	Affine invariant points. ∗ Mathieu Meyer, Carsten Schutt and Elisabeth M. Werner † Abstract We answer in the negative a question by Grunbaum who asked if there exists a finite basis of affine invariant points. We give a positive answer to another question by Grunbaum about the "size" of the set of all affine invariant points. Related, we show that the set of all convex bodies K, for which the set of affine invariant points is all of Rn, is dense in the set of convex bodies. Crucial to establish these results, are new affine invariant points, not previously considered in the literature. 1 Introduction. A number of highly influential works (see, e.g., [8, 10, 12], [14]-[18], [20]-[31], [37, 44, 51, 52, 58]) has directed much of the research in the theory of convex bodies to the study of the affine geometry of these bodies. Even questions that had been considered Euclidean in nature, turned out to be affine problems - among them the famous Busemann-Petty Problem (finally laid to rest in [6, 9, 56, 57]). The affine structure of convex bodies is closely related to the symmetry structure of the bodies. From an affine point of view, ellipsoids are the most symmetric convex bodies, and simplices are considered to be among the least symmetric ones. This is reflected in many affine invariant inequalities (we give examples below) where ellipsoids and simplices are the extremal cases. However, simplices have many affine symmetries. Therefore, a more systematic study for symmetry of convex bodies is needed. Grunbaum, in his seminal paper [13], initiated such a study. A crucial notion in his work, the affine invariant point, allows to analyze the symmetry situation. In a nutshell: the more affine invariant points, the fewer symmetries. In this paper, we address several issues that were left open in Grunbaum's paper. For instance, it was not even known whether there are "enough" affine invariant points. We settle this in Theorem 3 below. Let Kn be the set of all convex bodies in Rn (i.e., compact convex subsets of Rn with nonempty interior). Then (see Section 2 for the precise definition) a map p : Kn → Rn is called an affine invariant point, if p is continuous and if for every nonsingular affine map T : Rn → Rn one has, p(T (K)) = T (p(K)). ∗Keywords: affine invariant points, symmetry. 2010 Mathematics Subject Classification: 52A20, 53A15 †Partially supported by an NSF grant, a FRG-NSF grant and a BSF grant 1 An important example of an affine invariant point is the centroid g. More examples will be given throughout the paper. Let Pn be the set of affine invariant points on Kn, Pn = {p : Kn → Rn(cid:12)(cid:12) p is affine invariant}. Observe that Pn is an affine subspace of C(Kn, Rn), the continuous functions on Kn with values in Rn. We denote by V Pn the subspace parallel to Pn. Thus, with the centroid g, Grunbaum [13] posed the problem if there is a finite basis of affine invariant points, i.e. affine invariant points pi ∈ Pn, 1 ≤ i ≤ l, such that every p ∈ Pn can be written as V Pn = Pn − g. p = l Xi=1 αipi, with αi ∈ R and αi = 1. l Xi=1 We answer this question in the negative and prove: Theorem 1. V Pn is infinite dimensional for all n ≥ 2. In fact, we will see that, with a suitable norm, V Pn is a Banach space. Hence, by Baire's theorem, a basis of Pn is not even countable. For a fixed body K ∈ Kn, we let Pn(K) = {p(K) : p ∈ Pn}. Then Grunbaum conjectured [13] that for every K ∈ Kn, Pn(K) = Fn(K), (1) where Fn(K) = {x ∈ Rn : T x = x, for all affine T with T K = K}. We give a positive answer to this conjecture, when Pn(K) is (n − 1)-dimensional. Note also that if K has enough symmetries, in the sense that Fn(K) is reduced to one point xK , then Pn(K) = {xK}. Theorem 2. Let K ∈ Kn be such that Pn(K) is (n-1)-dimensional. Then Pn(K) = Fn(K). Symmetry or enough symmetries, are key in many problems. The affine invariant in- equalities connected with the affine geometry often have ellipsoids, respectively simplices as extremal cases. Examples are the Lp affine isoperimetric inequalities of the Lp Brunn Minkowski theory, a theory initiated by Lutwak in the groundbreaking paper [25]. For related results we refer to e.g. [3, 32, 33], [45]-[49], [53], [54]. The corresponding Lp affine isoperimetric inequalities, established by Lutwak [25] for p > 1 and in [53] for all other p - the case p = 1 being the classical affine isoperimetric inequality [1] - are stronger than the celebrated Blaschke Santal´o inequality (see e.g., [7, 43]; and e.g., [2, 36] for recent results): the volume product of polar reciprocal convex bodies is maximized precisely by ellipsoids. It is an open problem which convex bodies are minimizers for the Blaschke Santal´o inequality. Mahler conjectured that the minimum is attained for the simplex. A ma- jor breakthrough towards Mahler's conjecture is the inequality of Bourgain-Milman [4], 2 which has been reproved with completely different methods by Kuperberg [19] and by Nazarov [35]. See also [11, 39, 40, 42] for related results. Even more surprising is that it is not known whether the minimizer is a polytope. The strongest indication to date that it is indeed the case is given in [41]. Another example is the Petty projection inequality [38], a far stronger inequality than the classical isoperimetric inequality, and its Lp analogue, the Lp Petty projection inequality, established by Lutwak, Yang, and Zhang [27] (see also Campi and Gronchi [5]). These inequalities were recently strengthened and extended by Haberl and Schuster [15]. It is precisely the ellipsoids that are maximizers in all these inequalities. On the other hand, the reverse of the Petty projection inequality, the Zhang projection inequality [55], has the simplices as maximizers. Grunbaum [13] also asked , whether Pn(K) = Rn, if Fn(K) = Rn. A first step toward solving this problem, is to clarify if there is a convex body K such that Pn(K) = Rn. Here, we answer this question in the affirmative and prove that the set of all K such that Pn(K) = Rn, is dense in Kn and consequently the set of all K such that Pn(K) = Fn(K), is dense in Kn. Theorem 3. The set of all K ∈ Kn such that Pn(K) = Rn is open and dense in (Kn, dH ). Here, dH is the Hausdorff metric on Kn, defined as dH (K1, K2) = min{λ ≥ 0 : K1 ⊆ K2 + λBn 2 ; K2 ⊆ K1 + λBn 2 }, 2 is the Euclidean unit ball centered at 0. More generally, Bn where Bn 2 (a, r), is the Euclidean ball centered at a with radius r. We shall use the following well known fact. Let Km, K ∈ Kn. Then dH (Km, K) → 0 if and only if for some εm → 0 one has (1 − εm)(cid:0)K − g(K)(cid:1) ⊂ Km − g(Km) ⊂ (1 + εm)(cid:0)K − g(K)(cid:1) for every m. To establish Theorems 1 - 3, we need to introduce new examples of affine invariant (2) (3) points, that have not previously been considered in the literature. 2 Affine invariant points and sets: definition and properties. Let K ∈ Kn. Throughout the paper, int(K) will denote the interior, and ∂K the boundary of K. The n-dimensional volume of K is voln(K), or simply K. K◦ = {y ∈ Rn : hx, yi ≤ 1 ∀x ∈ K} is the polar body of K. More generally, for x in Rn, the polar of K with respect to x is K x = (K − x)◦ + x. A map p : Kn → Rn is said to be continuous if it is continuous when Kn is equipped with the Hausdorff metric and Rn with the Euclidean norm k · k. Grunbaum [13] gives the following definition of affine invariant points. Please note that formally we are considering maps, not points. Definition 1. A map p : Kn → Rn is called an affine invariant point, if p is continuous and if for every nonsingular affine map T : Rn → Rn one has p(T (K)) = T (p(K)). (4) 3 Let Pn the set of affine invariant points in Rn, and for a fixed body K ∈ Kn, Pn(K) = {p(K) : p ∈ Pn}. Pn = {p : Kn → Rn(cid:12)(cid:12) p is affine invariant}, (5) We say an affine invariant point p ∈ Pn proper, if for all K ∈ Kn, one has p(K) ∈ int(K). Examples. Well known examples (see e.g. convex body K in Rn are (i) the centroid [13]) of proper affine invariant points of a g(K) = RK xdx K ; (6) (ii) the Santal´o point, the unique point s(K) for which the volume product KK x attains its minimum; (iii) the center j(K) of the ellipsoid of maximal volume J (K) contained in K, or John ellipsoid of K; (iv) the center l(K) of the ellipsoid of minimal volume L(K) containing K, or Lowner ellipsoid of K. Note that if T (K) = K for some affine map T : Rn → Rn and some K ∈ Kn, then for every p ∈ Pn, one has p(K) = p(T (K)) = T (p(K)). It follows that if K is centrally symmetric or is a simplex, then p(K) = g(K) for every p ∈ Pn, hence Pn(K) = {g(K)}. The continuity property is an essential part of Definition 1, as, without it, pathological affine invariant points can be constructed. The next example illustrates this. Example 1. Let Pn be the set of all convex polytopes in Kn and define for P ∈ Pn, p(P ) = 1 m m Xi=1 vi(P ), where v1(P ), . . . vm(P ) are the vertices of P . For K ∈ Kn \ Pn, let p(K) = g(K), the centroid of K. Then p : Kn → Rn is affine invariant, but it is not continuous at any point. Indeed, let K ∈ Kn. We approximate K by a polytope P , and, in turn, approximate P by a polytope Pl by replacing one vertex v of P by sufficiently many vertices v1, . . . , vl near v. When l → ∞, p(Pl) → v ∈ ∂P . Pl is near K, but p(K) = g(K). Next, we introduce the notion of affine invariant set mappings, or, in short, affine invariant sets. There, continuity of a map A : Kn → Kn is meant when Kn is equipped on both sides with the Hausdorff metric. Our Definition 2 of affine invariant sets differs from the one given by Grunbaum [13]. Definition 2. A map A : Kn → Kn is called an affine invariant set mapping, if A is continuous and if for every nonsingular affine map T of Rn, one has A(T K) = T (A(K)). 4 We then call A(K), or simply the map A, an affine invariant set mappings. We denote by Sn the set of affine invariant set mappings, Sn = {A : Kn → Kn(cid:12)(cid:12)A is affine invariant and continuous}. We say that A ∈ Sn is proper, if A(K) ⊂ int(K) for every K ∈ Kn. Known examples (see e.g., [13]) of affine invariant sets are the John ellipsoid and the (7) Lowner ellipsoid. Further examples will be given all along this paper. Remarks. (i) It is easy to see that if λ ∈ R, p, q ∈ Pn and A ∈ Sn, then p ◦ A ∈ Pn and (1 − λ)p + λq ∈ Pn. Thus, Pn is an affine space and for every K ∈ Kn, Pn(K) is an affine subspace of Rn. Moreover, for A, B ∈ Sn, the maps K → (A ◦ B)(K), are affine invariant set mappings. (1 − λ)A(K) + λB(K) and conv[A, B](K) = conv[A(K), B(K)] (ii) Properties (4) and (2) imply in particular that for every translation by a fixed vector x0 and for every convex body K ∈ Kn, and p(K + x0) = p(K) + x0, for every p ∈ Pn A(K + x0) = A(K) + x0, for every A ∈ Sn (8) (9) (iiii) Unless p = q, it is not possible to compare two different affine invariant points p and q via an inequality of the following type (10) where k · k is a norm on Rn and c > 0 a constant. Indeed, by (ii), p(K − p(K)) = 0 and q(L − q(L)) = 0. Therefore, if (10) would hold, then kp(K) − p(L)k ≥ c kq(K) − q(L)k, 0 = kp(K − p(K)) − p(L − p(L))k ≥ c kq(K − p(K)) − q(L − p(L))k = c kq(K) − p(K) − q(L) + p(L)k. Choose now for L a symmetric convex body. Then kq(K)− p(K)k = 0, or p(K) = q(K). Remark (i) provides examples of non-proper affine invariant points: once there are two different affine invariant points, there are affine invariant points p(K) /∈ K, i.e. non- proper affine invariant points. An explicit example is the convex body Cn constructed in [34], for which the centroid and the Santal´o point differ. The next results describe some properties of affine invariant points and sets. Proposition 1. Let p, q ∈ Pn and suppose that p is proper. For K ∈ Kn, define φq(K) = inf {t ≥ 0 : q(K) − p(K) ∈ t (K − p(K))} . Then φq : Kn → R+ is continuous and (i) there exist c = c(q) > 0 such that q(K) − p(K) ∈ c (K − p(K)) for every K ∈ Kn. (ii) If moreover q is proper, then one can chose c ∈ (0, 1) in (i). 5 Proof. Since p(K) ∈ int(K), Rn = ∪t≥0t(K − p(K)). Therefore φq is well defined. Now we show that φq is continuous. Suppose that Km → K in (Kn, dH ). By definition, we have By continuity of p and q it follows that q(Km) − p(Km) ∈ φq(Km) (Km − p(Km)) for all m. and thus q(K) − p(K) ∈ lim inf m φq(Km) (K − p(K)) , φq(K) ≤ lim inf m φq(Km). Since p is proper, there exists d > 0, such that Bn 2 ⊆ d(K − p(K)). Since K is bounded, there exists D > 0 such that d(K − p(K)) ⊆ DBn 2 . Let η > 0 and fix ε = η/d > 0. Since Km → K, and by continuity of p and q, there exists m0 > 0 such that for every m ≥ m0, K − p(K) ⊆ Km − p(Km) + εBn 2 ⊆ Km − p(Km) + η(K − p(K)) and q(Km) − p(Km) ∈ q(K) − p(K) + εBn 2 ⊆ (φq(K) + η)(K − p(K)). Now we observe that, if two convex bodies A and B in Rn satisfy A ⊆ B + tA for some 0 < t < 1, then A ⊆ B/(1 − t). It then follows that for every m ≥ m0, K − p(K) ⊆ q(Km) − p(Km) ∈ Hence and thus 1 1 − η φq(K) + η 1 − η (Km − p(Km)). (Km − p(Km)) , φq(K) ≥ lim sup m φq(Km), 2 }, which is compact in Kn. ✷ and the continuity of φq is proved. Assertions (i) and (ii) follow from the continuity of φq. Indeed, by affine invariance, we may reduce the problem to the set {K ∈ Kn : Bn 2 ⊆ K ⊆ nBn Lemma 1. Let p, q ∈ Pn and suppose that p is proper. Then there exists a proper r ∈ Pn such that q is an affine combination of p and r. Proof. By the preceding proposition, there is c > 0 such that q(K)−p(K) ∈ c(K−p(K) for all K ∈ Kn. Put r = q−p 2c + p. Then r ∈ Pn and q = 2cr + (1 − 2c)p is an affine combination of p and r. Since p(K) ∈ int(K), (K + p(K)) ⊆ int(K), for all K ∈ Kn. (cid:3) r(K) ∈ 1 2 Analogous results to Proposition 1 for affine invariant sets are also valid. We omit their proofs. Proposition 2. Let A ∈ Sn, p, q ∈ Pn and suppose that p is proper. Then there exists a constant c1 > 0 such that for every K ∈ Kn, If moreover A is proper and p = q, one can choose c1 < 1. A(K) − q(cid:0)K(cid:1) ⊆ c1 (K − p(K)). 6 Lemma 2. Let A ∈ Sn and p ∈ Pn be proper. Then there exists t > 0 such that K → t(A(K) − p(K)) + p(K) = tA(K) + (1 − t)p(K) is a proper affine invariant set mapping. The next proposition gives a reverse inclusion for affine invariants sets. We need first another lemma, where, as in the proposition, g denotes the center of gravity. Lemma 3. For every D, d > 0 and n ≥ 1, there exists c > 0 such that, whenever K ∈ Kn satisfies K ⊆ DBn Proof. Suppose that K ∈ Kn satisfies the two assumptions. Define 2 and K ≥ d, then cBn 2 ⊆ K − g(K). cK = sup{c ≥ 0 : cBn 2 ⊆ K − g(K)}. Then cK > 0, and there exists x ∈ ∂K such that kx − g(K)k = cK. Since K − g(K) ⊆ n(cid:0)g(K) − K(cid:1), the length of the chord of K passing through g(K) and x is not bigger than (n + 1)cK. Let u ∈ Sn−1 be the direction of the segment [g(K), x] and let PuK be the orthogonal projection of K onto u⊥, the subspace orthogonal to u. Then, d ≤ K ≤ k(g + Ru) ∩ Kk PuK ≤ (n + 1)cK Dn−1Bn−1 2 . The second inequality follows from a result by Spingarn [50]. Thus we get a strictly positive lower bound c for cK which depends only on n, d and D. (cid:3) Proposition 3. Let A be an affine invariant set mapping. Then there exist c > 0 such that c(cid:0)K − g(K)(cid:1) ⊆ A(K) − g(cid:0)A(K)(cid:1), for every K ∈ Kn. Proof. We first prove that there exists d > 0 such that A(K) ≥ dK for every K ∈ Kn. By affine invariance, it is enough to prove that inf {K∈Kn:Bn 2 ⊂K⊆nBn 2 } A(K) K > 0. Since K → A(K) 2 } is compact in Kn, K this infimum is a minimum and it is strictly positive. By Proposition 2, applied with q = g ◦ A, there exists c > 0 such that is continuous and since {K ∈ Kn : Bn 2 ⊆ K ⊂ nBn Therefore, By Lemma 3, there is c0 > 0 such that A(K) − g(cid:0)A(K)(cid:1) ⊆ c(cid:0)K − g(K)(cid:1), A(K) − g(cid:0)A(K)(cid:1) ⊆ 2ncBn 2 ⊆ A(K) − g(cid:0)A(K)(cid:1), c0Bn 2 , for every K ∈ Kn. for every K ∈ K. for every K ∈ K. Now, since K − g(K) ⊆ 2nBn for K ∈ Kn with Bn 2 ⊆ K ⊂ nBn 2 for every K ∈ Kn with Bn 2 ⊆ K ⊂ nBn 2 , and thus for all K ∈ Kn by affine invariance. (cid:3) 2 , we get the result 7 3 Several questions by Grunbaum. We now give the proof of Theorems 1 - 3. To do so, we first need to introduce new affine invariant points. 3.1 The convex floating body as an affine invariant set mapping. n+1(cid:17)n Let K ∈ Kn and 0 ≤ δ <(cid:16) n . For u ∈ Rn and a ∈ R, H = {x ∈ Rn : hx, ui = a} is the hyperplane orthogonal to u and H + = {x ∈ Rn : hx, ui ≥ a} and H− = {x ∈ Rn : hx, ui ≤ a} are the two half spaces determined by H. Then the (convex) floating body Kδ [46] of K is the intersection of all halfspaces H + whose defining hyperplanes H cut off a set of volume at most δK from K, Kδ = \ {H:H−∩K≤δK} H +. (11) Clearly, K0 = K and Kδ ⊆ K for all δ ≥ 0. The condition δ < (cid:16) n g(K) ∈ int(Kδ) 6= ∅ (see [46]). Moreover, for all invertible affine maps T , one has n+1(cid:17)n insures that (T (K))δ = T (Kδ) . (12) To prove that K → Kδ is continuous from Kn to Kn, we need some notation. For u ∈ Sn−1, we define aδ,K(u) to be unique real number such that Then voln ({x ∈ K : hx, ui ≥ aδ,K(u)}) = δ voln(K). Kδ = \u∈Sn−1 {x ∈ K : hx, ui ≤ aδ,K(u)} . (13) Lemma 4. Let K ∈ Kn, u ∈ Sn−1, 0 < δ < (cid:16) n Let a ∈ R satisfy R +∞ f (t)dt = δR +∞ −∞ f (a) ≥ δ a n+1(cid:17)n and f (t) = {x ∈ K : hx, ui = t}. f (t)dt. Then one has n−1 n max t∈R f (t). Proof. By the Brunn-Minkowski theorem (see [7, 43]), f M = f (m) = maxt∈R f (t). 1 n−1 is concave on {f > 0}. Put We suppose first that m < a. Let g be the affine function on R such that gn−1(m) = f (m) and gn−1(a) = f (a). As f n−1 is concave on {f 6= 0}, one has gn−1 ≤ f on [m, a] and gn−1 ≥ f on {f 6= 0}\ [m, a]. Thus there exists c ≤ m and d ≥ a, such that g(c) > 0, g(d) > 0, 1 Z d c gn−1(t)dt =Z +∞ −∞ f (t)dt and Z d a gn−1(t)dt =Z +∞ a f (t)dt. 8 (c) ≥ M = f (m). Moreover, by (a) = f (a). We replace now g1 with a new function g2 that is affine 1 1 Let g1 = g1[c,d]. Since g is non increasing on [c, d], gn−1 construction, gn−1 on its support [c′, d′], c′ ≤ a ≤ d′, and satisfies gn−1 Z d′ (c′) ≥ gn−1 (t)dt =Z +∞ (a) = f (a) and clearly gn−1 One still has gn−1 computation gives Z d′ gn−1 2 gn−1 2 f (t)dt −∞ and c′ a 2 2 2 (a) = f (a), gn−1 (d′) = 0, 2 (t)dt =Z +∞ a f (t)dt. 1 (c) ≥ M . Now, an easy f (a) = gn−1 2 (a) = δ n−1 n , gn−1 2 (c′) ≥ δ n−1 n M. We suppose next that m ≥ a. The same reasoning, with 1 − δ instead of δ, gives n−1 n M . Since 0 < δ < 1 2 , the statement follows. (cid:3) f (a) ≥ (1 − δ) 2 . Let 2 and η > 0. There there exists ε > 0 (depending only on r, R, n, δ) such that, Proposition 4. Let 0 < r ≤ R < ∞ and let K ∈ Kn satisfy, rBn 0 < δ < 1 whenever a convex body L satisfies dH (K, L) ≤ ε, one has for every u ∈ Sn−1 2 ⊆ K ⊆ RBn aδ,K(u) − η ≤ aδ,L(u) ≤ aδ,K(u) + η. Proof. Let ρ > 0. With the hypothesis on K, we may choose ε > 0 small enough such that whenever dH (K, L) ≤ ε, then (1 − ρ)K ⊆ L ⊆ (1 + ρ)K. Fix u ∈ Sn−1 and define fK(t) = voln−1 ({x ∈ K : hx, ui = t}) and fL(t) = voln−1 ({x ∈ L : hx, ui = t}) . Then aδ,K := aδ,K(u) and aδ,L := aδ,L(u) satisfy Z +∞ aδ,K fK(t)dt = δK and Z +∞ aδ,L fL(t)dt = δL. Let θ > 0. For ρ > 0 small enough one has, K∆L ≤ ((1 + ρ)n − (1 − ρ)n)K ≤ θ. For such a ρ one has also ZR fK(t) − fL(t)dt ≤ ZR so that voln−1 ({x ∈ K∆L : hx, ui = t}) dt = K∆L ≤ θ, +Z +∞ aδ,L fK(t) − fL(t)dt ≤ 2θ. (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Z[aδ,K ,aδ,L] For α > 0 given, let θ = αK2 . Then fK(t)dt(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Z +∞ ≤(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) aδ,K fK(t)dt −Z +∞ aδ,L fL(t)dt(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) fK(t)dt − δK(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 9 Z +∞ aδ,L (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ α δK. 1 n−1 K on For some β ∈ [−α, α], one has hence that aδ,L = a(1+β)δ,K. Concavity of f {fK 6= 0} implies that Z[aδ,K ,aδ,L] ≥ aδ,K − aδ(1+β),K min(cid:0)fK(aδ,K), fK(a(1+β)δ,K)(cid:1) . fK(t)dt(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) min(cid:18)fK(aδ,K), fK(a(1+α)δ(cid:19) ≥(cid:18) min(1 + β, 1)δ(cid:19) If M = maxt∈R fK(t), we get by Lemma 4, n−1 n n−1 n M. ≥(cid:0)(1 − α)δ(cid:1) Since K ⊆ RBn 2 , we estimate M from above by M ≤ γn = Rn−1Bn−1 2 , which is an upper bound independent of u. It follows that if α > 0 is small enough, then (cid:3) aδ,K − aδ,L ≤ 2 θ γn (cid:0)(1 − α)δ(cid:1)− n−1 n ≤ η. The next proposition shows that the map K 7→ Kδ as defined in (11), is an affine invariant set mapping. , the mapping K 7→ Kδ is is an affine invariant set n+1(cid:1)n Proposition 5. For 0 < δ <(cid:0) n mapping from Kn to Kn . n+1(cid:1)n Proof. We take 0 < δ < (cid:0) n so that int(Kδ) 6= ∅ and g(K) ∈ int(Kδ) . It is clear that K → Kδ is an affine invariant mapping and it is clear that g(K) ∈ Kδ. We now fix a body K ∈ Kn and we verify the continuity of the mapping K → Kδ at K. We may suppose that 0 is the center of mass of K. For some 0 < r ≤ R < ∞, one has rBn 2 ⊆ Kδ ⊆ K ⊆ RBn 2 . By the choice of δ, aδ,K(u) > 0 for every u ∈ Sn−1, where aδ,K(u) is as in (13). Let η, η′ > 0 satisfy η′ ≤ ηr ≤ η minu∈Sn−1 aδ,K(u). We use the notation of the preceding proposition to find ε > 0 such that for any L with dH (K, L) ≤ ε, one has aδ,K(u) − η′ ≤ aδ,L(u) ≤ aδ,K(u) + η′, (1 − η)aδ,K(u) ≤ aδ,L(u) ≤ (1 + η)aδ,K(u), or whence (1 − η)Kδ ⊆ Lδ ⊆ (1 + η)Kδ. 2 , it then follows that, given ρ > 0, for η > 0 small enough, one Since rBn 2 ⊆ Kδ ⊆ RBn has dH (Lδ, Kδ) ≤ ρ. (cid:3) As a corollary, we obtain new affine invariant points. 10 n+1(cid:1)n Corollary 1. Let 0 < δ < (cid:0) n and let p : Kn → Rn be an affine invariant point. Then K → p(Kδ) is also an affine invariant point. In particular, for the centroid g, K 7→ g(K \ Kδ) is an affine invariant point. Proof. Affine invariance follows from Remark (i) after Definition 2 and continuity from Proposition 5. The second statement follows now from the trivial identity g(K) = Kδ K g(Kδ) + K \ Kδ K g(K \ Kδ), which gives g(K \ Kδ) = K K \ Kδ g(K) − Kδ K \ Kδ g(Kδ), as an affine combination of continuous affine invariant points. (cid:3) The next lemma is key for many of the proofs that will follow. Lemma 5. Let m ≥ n + 1 and for 1 ≤ i ≤ m, let vi ∈ Rn be the vertices of a polytope P in Kn. For all ε > 0 there exists z ∈ P with kv1 − zk ≤ ε and 0 < r ≤ ε such that 2 (z, r) ⊂ P and if K = conv (Bn Bn 2 (z, r), v2, . . . , vm), then K satisfies (i) K ⊆ P , v2, . . . , vm are extreme points of K and dH (K, P ) ≤ ε. (ii) For sufficiently small δ, kv1 − g (K \ Kδ)k ≤ 2ε Proof. There exists a hyperplane H that striclty separates v1 and {v2, . . . , vm}, such that for all x ∈ H−∩ P we have that kx− v1k < ε. Let z ∈ int(H−)∩ int(P ). Then there exists 0 < r ≤ ε such that Bn (i) By construction of K, v2 . . . , vm are extreme points of K. Also, for all x ∈ K ∩ H−, one has kx − v1k < ε. Therefore K ⊆ P ⊆ K + εBn (ii) We have for δ > 0, 2 (z, r) ⊆ H− ∩ P . Let K = conv (Bn 2 and thus dH (K, P ) ≤ ε. 2 (z, r), v2, . . . , vm) . g (K \ Kδ) = (K ∩ H +) \ Kδ + (K ∩ H−) \ Kδ K \ Kδ K \ Kδ g(cid:0)(cid:0)K ∩ H +(cid:1) \ Kδ(cid:1) g(cid:0)(cid:0)K ∩ H−(cid:1) \ Kδ(cid:1) . Since g ((K ∩ H−) \ Kδ) ∈ int (K) ∩ int (H−), one has (14) Observe that ∂K contains a cap of ∂B(z, r), so that kv1 − g(cid:0)(cid:0)K ∩ H−(cid:1) \ Kδ(cid:1)k ≤ ε. C =Z∂K 1 κ n+1 K dµK > 0. By Theorem 4, one has for δ sufficiently small, K \ Kδ = K − Kδ ≥ C 2cn 2 n+1 . (δK) (15) 11 Let R = max{kxk : x ∈ P}. As the Gauss curvature is equal to 0 everywhere on the boundary ∂ (K ∩ H +), again by Theorem 4, one has for sufficiently small δ, As we get K ∩ H + −(cid:12)(cid:12)(cid:12)(cid:12) cn 2 n+1 (δK) (K ∩ H +) δK K∩H+ (cid:12)(cid:12)(cid:12)(cid:12) Cε 4R . ≤ (cid:0)K ∩ H +(cid:1) \ Kδ ⊆(cid:0)K ∩ H +(cid:1) \(cid:0)K ∩ H +(cid:1) δK K∩H + 2 (cid:12)(cid:12)(cid:0)K ∩ H +(cid:1) \ Kδ(cid:12)(cid:12) ≤ Cε 4Rcn 2 n+1 . (δK) , (16) (17) It folllows from (15) and (16) that for δ small enough one has We get thus from (14) and (17) (K ∩ H +) \ Kδ K \ Kδ ε R . ≤ K \ Kδ kg (K \ Kδ) − g(cid:0)(cid:0)K ∩ H−(cid:1) \ Kδ(cid:1)k = (K ∩ H +) \ Kδ ≤ (K ∩ H +) \ Kδ K \ Kδ (R + R) ≤ ε ≤ ε 2R kg(cid:0)(cid:0)K ∩ H +(cid:1) \ Kδ(cid:1) − g(cid:0)(cid:0)K ∩ H−(cid:1) \ Kδ(cid:1)k (cid:0)kg(cid:0)(cid:0)K ∩ H−(cid:1) \ Kδ(cid:1)k + kg(cid:0)(cid:0)K ∩ H +(cid:1) \ Kδ(cid:1)k(cid:1) Altogether, kv1−g (K \ Kδ)k ≤ kv1−g(cid:0)(cid:0)K ∩ H−(cid:1) \ Kδ(cid:1)k+kg (K \ Kδ)−g(cid:0)(cid:0)K ∩ H−(cid:1) \ Kδ(cid:1)k ≤ 2ε. (cid:3) 3.2 Proof of Theorem 1: Pn is infinite dimensional. Here, we answer in the negative Grunbaum's question whether there exists a finite basis for Pn, i.e. affine invariant points pi ∈ Pn, 1 ≤ i ≤ l, such that every p ∈ Pn can be written as l l p = Xi=1 αipi, with αi ∈ R and αi = 1. Xi=1 Recall that Pn is an affine subspace of C(Kn, Rn), the continuous functions on Kn with values in Rn and that we denote by V Pn the subspace parallel to Pn. Thus, with the centroid g, The dimension of Pn is the dimension of V Pn. We introduce a norm on V Pn, V Pn = Pn − g. kvk = sup 2 ⊆K⊆nBn 2 kv(K)k, K∈Kn,Bn for v ∈ V Pn. 12 (18) (19) Observe that the set {K ∈ Kn : Bn 2 } is a compact subset of (Kn, dH ). Therefore (19) is well defined and it is a norm: v = p − g 6= 0 implies that there is C with v(C) 6= 0. By John's theorem (e.g., [?]), there is an affine, invertible map T with Bn 2 ⊆ T (C) ⊆ nBn 2 ⊆ K ⊆ nBn 2 . Thus, v(T (C)) = (p − g)(T (C)) = p(T (C)) − g(T (C)) = T (p(C)) − T (g(C)). Since T = S + x0, where S is a linear map, v(T (C)) = S(p(C) − g(C)) = S((p − g)(C)) 6= 0. Hence kvk ≥ kv(T (C))k > 0. For the proof of Theorem 1 and Theorem 3, we will make use of the following theorem by Schutt and Werner [46]. There, µK is the usual surface measure on ∂K and for x ∈ ∂K, κ(x) is the generalized Gauss curvature at x, which is defined µK almost everywhere. Theorem 4. [46] Let K be a convex body in Rn. Then, if cn = 2(cid:16)Bn−1 n+1 (cid:17) , one has n+1 2 cn lim δ→0 K − Kδ (δK) n+1 2 =Z∂K 1 n+1 (x) dµK(x). κ Proof of Theorem 1. We show that the closed unit ball of V Pn is not compact. For K ∈ Kn and δ > 0, let Kδ be the convex floating body of K. Let g be the centroid and let gδ : Kn → Rn be the affine invariant point given by gδ(K) = g(K \ Kδ). The set of vectors {vδ = gδ − g: δ > 0} is bounded. gδ(K) ∈ K Indeed, since g(K) ∈ K and kvδk ≤ , j ≥ 1, does not have a convergent subsequence: We show that for all (kg(K)k + kgδ(K)k) ≤ 2n. sup 2 ⊆K⊆nBn K∈Kn,Bn 2 2 such that j The sequence v 1 N there are ℓ ≥ m ≥ N and K ∈ Kn with Bn (K) − v 1 v 1 m ℓ As K we choose the union of the cylinder D = [−1, 1] × Bn−1 ball, 2 where, with e1 = (1, 0, . . . , 0) ∈ Rn, K(h) = D ∪ C(h), and a cap of a Euclidean (20) C(h) =(cid:18) h2 + 2h − 1 2h e1 + 1 + h2 2h Bn 2(cid:19) ∩ {x = (x1, . . . , xn) ∈ Rn : x1 ≥ 1}. 2 ⊆ K ⊆ nBn . 1 10 (K)(cid:13)(cid:13)(cid:13) ≥ (cid:13)(cid:13)(cid:13) As h → 0, K(h) → D and, by Corollary 1, g 1 h0 > 0 such that g(cid:16)K(h) \ (K(h)) 1 kg 1 m m (K(h))k =(cid:13)(cid:13)(cid:13) m(cid:17)(cid:13)(cid:13)(cid:13) ≤ 13 (K(h)) → g 1 m (D) = 0. Thus there exists 1 10 , for all h ≤ h0. Now we show that we can choose ℓ sufficiently big so that kg 1 ℓ (K(h))k ≥ 1 2 . (21) We apply the same argument as in the proof of Lemma 5. Let H be the hyperplane such that K(h) ∩ H− = C(h) and K(h) ∩ H + = D. Then as in (14), (K(h)) = (cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12) g 1 ℓ Since we get D \ K(h) 1 K(h) \ K(h) 1 g(cid:16)D \ K(h) 1 C(h) \ K(h) 1 K(h) \ K(h) 1 ℓ(cid:17) . g(cid:16)C(h) \ K(h) 1 ℓ(cid:17) + (cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12) ℓ(cid:12)(cid:12)(cid:12) ℓ(cid:12)(cid:12)(cid:12) ℓ(cid:12)(cid:12)(cid:12) ℓ(cid:12)(cid:12)(cid:12) g(D \ K(h) 1 ℓ ) ∈ D and g(C(h) \ K(h) 1 ℓ ) ∈ C(h) )k ≤ Therefore, by triangle inequality, kg(D \ K(h) 1 ℓ √2 and kg(C(h) \ K(h) 1 ℓ k ≥ 1. kg 1 ℓ C(h) \ K(h) 1 K(h) \ K(h) 1 ℓ(cid:12)(cid:12)(cid:12) ℓ(cid:12)(cid:12)(cid:12) By Theorem 4, we get as in (15), for ℓ large enough, with α(h) =R∂K(h) κ ℓK(h)(cid:1) (K(h))k ≥ (cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12) ℓK(h)(cid:1) ℓ(cid:12)(cid:12)(cid:12) ℓ(cid:12)(cid:12)(cid:12) Z∂K(h) D \ K(h) 1 K(h) \ K(h) 1 n+1 K(h)dµK(h) = (cid:0) 1 √2 (cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12) K(h) \ K(h) 1 2cn − . 2 2 n+1 (22) 1 n+1 K(h)dµK(h), α(h). Also by Theorem 4, we get as in (16), (cid:12)(cid:12)(cid:12) 1 κ n+1 2cn ℓ(cid:12)(cid:12)(cid:12) ≥ (cid:0) 1 ℓ(cid:12)(cid:12)(cid:12) ℓ(cid:12)(cid:12)(cid:12) ≤ =(cid:12)(cid:12)(cid:12)(cid:0)K(h) ∩ H +(cid:1) \ K(h) 1 D \ K(h) 1 (cid:12)(cid:12)(cid:12) Now we finish the proof as in Lemma 5. ✷ ε cn α(h) (cid:18) 1 ℓK(h)(cid:19) 2 n+1 . 3.3 Proof of Theorem 2. It was also asked by Grunbaum [13] if for every K ∈ Kn, Pn(K) = Fn(K), where Fn(K) = {x ∈ Rn : T x = x, for all affine T with T K = K}. Observe that it is clear that Pn(K) ⊆ Fn(K). We will prove that Pn(K) = Fn(K), if Pn(K) is (n − 1)- dimensional. To do so, we, again, first need to define new affine invariant set mappings. Actually, in the proof of Theorem 2 we show that the group of isometries of K equals {In, S} = {T : Rn → Rn affine one to one, T K = K}, where S is reflection about a hyperplane, i.e. S : Rn → Rn is bijective and there is a hyperplane H and a direction ξ /∈ H such that S(h + tξ) = h − tξ for all h ∈ H. 14 Lemma 6. (i) Let p ∈ Pn and let g be the centroid. For 0 < ε < 1, define Ap,ε(K) =(cid:26)x ∈ K(cid:12)(cid:12)(cid:12)(cid:12) hx, p((K − g(K))◦)i ≥ (1 − ε) sup y∈Khy, p((K − g(K))◦)i(cid:27) . Then Ap,ε : Kn → Kn is an affine invariant set map. (ii) Let p ∈ Pn and let q ∈ Pn be proper. Then Aq,p,ε : Kn → Kn given by Aq,p,ε(K) = {x ∈ K hx, p((K − g(K))◦)i ≥ (1 − ε)hq(K), p((K − g(K))◦)i} is an affine invariant set map. Observe that 0 ∈ Pn((K − g(K))◦) since 0 is the Santal´o point of (K − g(K))◦. Therefore, Pn((K − g(K))◦) is a subspace of Rn. Proof. Let T be an invertible, affine map and T = S + a its decomposition in a linear map S and a translation a. Then for any convex body C that contains 0 in its interior, Moreover, (S(C))◦ = S∗−1(C◦). p((T (K) − g(T (K)))◦) = p((S(K − g(K)))◦) = p(S∗−1((K − g(K))◦)) = S∗−1(p((K − g(K))◦)). Since S∗−1∗ = S−1, Ap,ε(T (K)) = =(x ∈ T (K)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) =(x ∈ T (K)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) hx, p((T (K) − g(T (K)))◦)i ≥ (1 − ε) hS−1x, p((K − g(K))◦)i ≥ (1 − ε) sup sup y∈T (K)hy, p((T (K) − g(T (K)))◦)i) y∈T (K)hS−1y, p((K − g(K))◦)i) and one verifies easily that Ap,ε(T (K)) = T (Ap,ε(K)). Please note that Ap,ε(K) is convex, compact and nonempty. ✷ Lemma 7. Let K ∈ Kn and let P : Rn → Rn be the orthogonal projection onto Pn((K− g(K))◦). Then the restriction of P to the subspace Pn(K − g(K)) is an isomorphismn between Pn(K − g(K)) and Pn((K − g(K))◦). In particular, dim(Pn(K − g(K))) = dim(Pn((K − g(K)))◦) Proof. On the hyperplane Pn((K − g(K))◦), P (K − g(K)) has an interior point. This holds because otherwise, by Fubini, voln(K) = 0. Let k = dim(Pn((K−g(K)))◦). We choose u1 ∈ Pn((K−g(K))◦). Then g(Au1,ε1 ) is a proper affine invariant point. Now we choose u2 ∈ Pn((K − g(K))◦) that is orthogonal to P (g(Au1,ε1 )). Then P (g(Au1,ε1 )) and P (g(Au2,ε2 )) are linearly independent. 15 Eventually, P (g(Au1,ε1)), . . . , P (g(Auk ,εk )) are linearly independent, and therefore g(Au1,ε1 ), . . . , g(Auk ,εk ) are linearly independent. Therefore, dim(Pn((K − g(K)))◦) ≤ dim(Pn(K − g(K))). Now we interchange the roles of Pn(K − g(K)) and Pn((K − g(K))◦) and get the inverse inequality. Let Q denote the restriction of P to the subspace Pn(K−g(K)). g(Au1,ε1), . . . , g(Auk ,εk ) is a basis of Pn(K − g(K)) and P (g(Au1,ε1 )), . . . , P (g(Auk ,εk )) is a basis of Pn((K − g(K)))◦. Q is a bijection between the two bases, thus Q is an isomorphism. ✷ Lemma 8. Let K ∈ Kn. Then for every point x from the relative interior of K ∩ Pn(K) there is a proper affine invariant point q with q(K) = x. Proof. We use the same notation as in Lemma 7 and its proof. We may assume that g(K) = 0. Suppose that there is an interior point x of Pn(K) ∩ K in the hyperplane Pn(K) for which there is no proper affine invariant point q with q(K) = x. The set {p(K)p is a proper affine invariant point} is convex. P : Rn → Rn is the orthogonal projection onto Pn(K◦). Then P (Pn(K)∩ K) is a convex set in the hyperplane Pn(K◦). Since P is an isomorphism between the hyperplanes Pn(K ) and Pn(K◦) we have P (x) /∈ P ({p(K)p is a proper affine invariant point}). Moreover, P (x) is an interior point of P (Pn(K) ∩ K). By the Hahn-Banach theorem there is u ∈ Pn(K◦) such that for all proper affine invariant points p we have On the other hand, there is an affine invariant point q with q(K◦) = u. Then g ◦ Au,hu,xi is a proper affine invariant point with hu, xi ≥ hu, P (p(K))i. which is a contradiction. ✷ hu, xi < hu, g ◦ Aq,hu,xii, Lemma 9. Let K ∈ Kn and suppose that dim(Pn(K)) = n − 1. Then S : Rn → Rn with for all y ∈ Pn(K − g(K)) and x ∈ Pn((K − g(K))◦)⊥ is a linear map such that S(y + x) = y − x S(K − g(K)) = K − g(K). 16 Proof. By Lemma 7, the orthogonal projection onto Pn((K − g(K))◦) restricted to Pn(K − g(K)) is an isomorphism. Therefore, Rn = Pn(K − g(K)) ⊕ Pn((K − g(K))◦)⊥. By Lemma 8 for every y ∈ Pn(K − g(K)) ∩ int(K) there is a proper affine invariant point q with y = q(K). Let u1, . . . , un−1 be an orthonormal basis in Pn((K − g(K))◦). The map Aε : Kn → Kn defined by Aε(K) = n−1 \i=1 {x ∈ Khq(K), uii − ε ≤ hx, uii ≤ hq(K), uii + ε} is an affine invariant set map. As q(K) is an interior point of Aε(K), Aε(K) ∈ Kn. Moreover, lim ε→0 Aε(K) = K ∩ (q(K) + Pn((K − g(K))◦)⊥) in the Hausdorff metric. g◦ Aε is a proper affine invariant point. Since all affine invariant points are elements of Pn(K) lim ε→0 (g ◦ Aε)(K) = q(K). On the other hand, q(K) is the midpoint of K ∩ (q(K) + Pn((K − g(K))◦)⊥). ✷ Proof of Theorem 2. Theorem 2 now follows immediately from Lemma 9. Indeed, Lemma 9 provides a map T = S − S(g(K)) + g(K) with T (K) = K and such that for all z ∈ Pn(K) and for all x ∈ Pn((K − g(K))◦)⊥, T (z + x) = z − x. Consequently, if w /∈ Pn(K), then T (w) 6= w, which means that the complement of Pn(K) is contained in the complement of Fn(K). Remark. As a byproduct of the preceding results, it can be proved that if K ∈ Kn satisfies Pn(K) = Rn and if Sn(K) = {A(K) : A ∈ Sn}, then Sn(K) is dense in Kn. It might be conjectured that for general K ∈ Kn, Sn(K) is dense in {C ∈ Kn : Fn(C) ⊆ Fn(K)}. 3.4 Proof of Theorem 3. In this subsection we show that the set of all K such that Pn(K) = Rn, is dense in Kn and consequently the set of all K such that Pn(K) = Fn(K) is dense in Kn. A further corollary is that, for every k ∈ N, 0 ≤ k ≤ n, there exists a convex body Qk such that P(Qk) is a k-dimensional affine subspace of Rn. It is relatively easy to construct examples of convex bodies K in the plane such that Pn(K) = R2. To do so in higher dimensions is more involved and we present a construction in the proof of Theorem 3 below. First, we will briefly mention two examples in the plane. Example 1. Let S be a regular simplex in the plane and let J (S) be the ellipsoid of maximal area inscribed in S. We show in the section below that the center j(S) of J (S) 17 2 , the Euclidean ball centered is an affine invariant point. We can assume that J (S) = B2 at 0 with radius 1. Then e.g. S = conv(cid:0)(−1,−√3), (−1,√3), (2, 0)(cid:1). Let 0 < λ < 1 be such that H((1 + λ)e1, e1) ∩ int(S) 6= ∅ and consider the convex body S1 = S ∩ H +((1 + λ)e1, e1) obtained from S by cutting of a cap from S. Then still j(S1) = 0 but the center of gravity has moved to the left of 0. Next, let γ > 0 be such that H((1 + γ)u, u)∩ int(S1) 6= ∅, where u = (−1,√3) and consider the convex body S2 = S1 ∩ H +((1 + γu), u) obtained from S1 by cutting of a cap from S1. Then still j(S2) = 0 but the center of gravity g(S2) of S2 has moved and it is different from the Santal´o point s(S2) of S2. j(S2), g(S2) and s(S2) are three affinely independent points of R2, hence span R2. 2 Example 2. Let S be the equilateral triangle in the plane centered at 0 of Example 1 with vertices a = (2, 0), b = (−1,√3) and c = (−1,−√3). Then, as noted in Example 1, B2 2 is the John ellipse J (S) of S. Let b1, c1 be two points on the segments [a, b] and [a, c], such that the segment [b1, c1] does not intersect B2 2 is still the John ellipse of the quadrangle conv (b, b1, c1, c). Now the Lowner ellipse L(T ) of the triangle T = conv (b1, b, c) is centered at 1 3 (b1 + b + c) 6= 0, if b1 6= a. L(T ) intersects the segment 2 and thus c′ → a. [a, c] at c and at some point c′. When b1 → a, one has L(T ) → 2B2 2, and thus for some c′′ ∈ [a, c], So we may choose b1 such that [b1, c′] does not meet B2 [b1, c1] does not meet B2 2 for any c1 ∈ [c′, c′′]. Finally, let P (c1) be the quadrangle P (c1) = conv (b, b1, c1, c), with c1 ∈ [c′, c′′]. Since b1, b, c and c1 are the vertices of P (c1) and c1 ∈ L(T ), L(T ) is also the Lowner ellipsoid L(P (c1))of P (c1). Altogether, 2 . Then B2 The John ellipse of P (c1) is B2 2 which is centered at 0, so that the affine invariant point j(P (c1)) = 0. The Lowner ellipse of P (c1) is centered at 1 3 (b1 + b + c), so that the affine invariant point l((P (c1)) = 1 3 (b1 + b + c) 6= 0. An easy computation shows that the centroid of P (c1) moves on an hyperbola when c1 varies in [c′, c′′]. So, in general, these three points are not on line. (cid:3) Proof of Theorem 3. The set of n-dimensional polytopes is dense in (Kn, dH ). Let P be a polytope and let η > 0 be given. Then it is enough to show that there exists a convex body Q with dH (P, Q) < η and such that Pn(Q) = Rn. We describe the idea of the proof. For a properly constructed convex body Q we will construct ∆i ∈ Pn, 1 ≤ i ≤ n + 1, in such a way that the ∆i(Q) are affinely independent. The construction of such a Q is done inductively: we first construct Q1 very near P and such that ∆1(Q1) is near an extreme point v1 of Q1. Then we construct Q2 very near Q1 and P and such that ∆1(Q2) is near the extreme point v1 of Q2 and ∆2(Q2) is near an extreme point v2 6= v1 of Q2. Let P = conv (v1, . . . , vm) be a polytope with non-empty interior and with m vertices, m ≥ n + 1. We pick n + 1 affinely independent vertices of P . We can assume that these are v1, . . . , vn+1. Let 0 < η1 < η n+2 be given. By Lemma 5, there exists z1 ∈ P , kv1 − z1k ≤ η1, and 0 < r1 ≤ η1 such that Bn 2 (z1, r1) ⊆ P and such that Q1 = conv (Bn 2 (z1, r1), v2, . . . , vm) 18 has v2, . . . , vm as extreme points, and for sufficiently small δ1, dH (Q1, P ) ≤ η1, We let ε1 < η1 and choose an ε1-net Pε1 on ∂ (Bn 2 (z1, r1)) and put kv1 − g (Q1 \ (Q1)δ1 )k ≤ 2η1. P1 = conv (Pε1, v2, v3, . . . , vm) . (23) (24) and ε > 0, there exists γ(K, δ, ε) such that if Then P1 ⊆ Q1 ⊆ P and dH (P1, Q1) ≤ ε1 < η1. By Corollary 1, for a given K ∈ Kn, for n+1(cid:1)n a given 0 < δ <(cid:0) n As dH (P1, Q1) ≤ ε1, we get that then kg(K \ Kδ) − g(L \ Lδ)k < ε. (25) dH (K, L) < γ(K, δ, ε), for L ∈ Kn, kg(Q1 \ (Q1)δ1 ) − g(P1 \ (P1)δ1 )k < η1, if we choose in addition ε1 such that ε1 < γ(Q1, δ1, η1). Thus, together with (24), kv1 − g(P1 \ (P1)δ1 )k ≤ 3 η1. (26) Observe that v2, . . . , vm are extreme points of P1. Now we apply Lemma 5 to P1. Let η2 < min{ε1, γ(P1, δ1, η1)}. By Lemma 5 there exists z2 ∈ P1, kv2 − z2k ≤ η2, and 0 < r2 ≤ η2 such that Bn 2 (z2, r2) ⊂ P1 and such that Q2 = conv (Pε1, Bn 2 (z2, r2), v3, . . . , vm) has v3, . . . , vm as extreme points, and for sufficiently small δ2, dH (Q2, P1) ≤ η2, kv2 − g (Q2 \ (Q2)δ2 )k ≤ 2η2. (27) (28) As kv1 − z1k ≤ η1 and kv2 − z2k ≤ η2, we have that dH (Q2, P ) ≤ η1. Moreover, as dH (Q2, P1) ≤ η2 < γ(P1, δ1, η1), we get by (25) with ε = η1 and by (26) that kv1 − g(Q2 \ (Q2)δ1 )k ≤ kv1 − g(P1 \ (P1)δ1 )k+kg(P1 \ (P1)δ1 ) − g(Q2 \ (Q2)δ1 )k ≤ 4η1. Now we let ε2 < min{η2, γ(Q2, δ1, η1)}, choose an ε2-net Pε2 on ∂ (Bn 2 (z2, r2)) and put P2 = conv (Pε1 ,Pε2 , v3, . . . , vm) . Then P2 ⊆ Q2 ⊆ P and dH (P2, Q2) ≤ ε2. By (25), with ε = η2, and if we choose in addition ε2 < η(Q2, δ2, η2), we get and thus, together with (28), kg(Q2 \ (Q2)δ2 ) − g(P2 \ (P2)δ2 )k < η2 kv2 − g(P2 \ (P2)δ2 )k ≤ 3 η2. (29) 19 Please note that v3, . . . , vm are extreme points of P2. Now we apply Lemma 5 to P2. Let η3 < min{ε2, γ(P2, δ2, η2)}. By Lemma 5 there exists z3 ∈ P2, kv3 − z3k ≤ η3, and 0 < r3 ≤ η3 such that Bn 2 (z3, r3) ⊂ P2 and such that Q3 = conv (Pε1 ,Pε1 , Bn 2 (z3, r3), v4, . . . , vm) has v4, . . . , vm as extreme points, and for sufficiently small δ3, dH (Q3, P2) ≤ η3, kv3 − g (Q3 \ (Q3)δ3 )k ≤ 2 η3. (30) (31) As kv1 − z1k ≤ η1, kv2 − z2k ≤ η2 and kv3 − z3k ≤ η3 we have that dH (Q3, P ) ≤ η1. Moreover, as dH (Q3, P2) ≤ η3 < γ(P2, δ2, η2), we get by (25) with ε = η2 and (29) that kv2 − g(Q3 \ (Q3)δ2 )k ≤ kv2 − g(P2 \ (P2)δ2 )k+kg(P2 \ (P2)δ2 ) − g(Q3 \ (Q3)δ2 )k ≤ 4η2. As dH (Q2, Q3) ≤ ε2 < γ(Q2, δ1, η1), it follows from (25) with ε = η1 that kg(Q2 \ (Q2)δ1 ) − g(Q3 \ (Q3)δ1 )k ≤ η1. By (30), it also follows from (25) with ε = η1 that kg(P1 \ (P1)δ1 ) − g(Q2 \ (Q2)δ1 )k ≤ η1. This, together with (26) gives kv1 − g(Q3 \ (Q3)δ1 )k ≤ kv1 − g(P1 \ (P1)δ1 )k + kg(P1 \ (P1)δ1 ) − g(Q2 \ (Q2)δ1 )k + kg(Q2 \ (Q2)δ1 ) − g(Q3 \ (Q3)δ1 )k ≤ 5η1. We continue to obtain Q = Qn+1 and affine invariant points ∆i = g(Q \ Qδi), 1 ≤ i ≤ n + 1, such that for all i, kvi − ∆i(Q)k ≤ (n + 2)η1 < η. As for 1 ≤ i ≤ n + 1, the vi are affinely independant, so are the ∆i. It remains to show that On = {K ∈ Kn : Pn(K) = Rn} is open in (Kn, dH ). Observe that K ∈ On if and only if for some p1 . . . , pn+1 ∈ Pn (depending on K), voln(conv(cid:0)p1(K) . . . , pn+1(K)(cid:1) > 0. (cid:3) Since L → vol(conv(cid:0)p1(L) . . . , pn+1(L)(cid:1) is continuous on Kn, it follows that On is open. Corollary 2. For every k ∈ N, 0 ≤ k ≤ n, there exists a convex body Qk such that P(Qk) is a k-dimensional affine subspace of Rn. Proof. For k = 0, we take a centrally symmetric body. For k = n, we take the body Q of Theorem 3. For 1 ≤ k ≤ n − 1, we take the intermediate bodies Qk constructed in the proof of Theorem 3. (cid:3) 20 References [1] W. Blaschke, Vorlesungen uber Differentialgeometrie II, Affine Differentialgeometrie. Springer Verlag, Berlin, (1923). [2] K.J. Boroczky, Stability of the Blaschke-Santal´o and the affine isoperimetric inequality Adv. in Math. 225 (2010), 1914 -- 1928. [3] K.J. Boroczky, E. Lutwak, D. Yang, and G. Zhang, The logarithmic Minkowski problem, to appear in Journal of AMS. [4] J. Bourgain and V. D. Milman, New volume ratio properties for convex symmetric bodies in Rn , Invent. Math. 88 (1987), 319 -- 340. [5] S. Campi and P. Gronchi, The Lp-Busemann-Petty centroid inequality, Adv. in Math. 167 (2002), 128 -- 141. [6] R. J. Gardner, A positive answer to the Busemann-Petty problem in three dimensions, Ann. of Math. (2) 140 (1994), 435-47. [7] R.J. Gardner Geometric tomography. Second edition. Encyclopedia of Mathematics and its Applications, 58. Cambridge University Press, Cambridge (2006). [8] R. J. Gardner, The dual Brunn-Minkowski theory for bounded Borel sets: Dual affine quermassintegrals and inequalities, Adv. Math. 216 (2007), 358-386. [9] R. J. Gardner, A. Koldobsky, and T. Schlumprecht, An analytical solution to the Busemann-Petty problem on sections of convex bodies, Ann. of Math. (2) 149 (1999), 691-703. [10] R. J. Gardner and G. Zhang, Affine inequalities and radial mean bodies. Amer. J. Math. 120, no.3, (1998), 505-528. [11] Y. Gordon, M. Meyer and S. Reisner, Zonoids with minimal volume product -- a new proof, Proc. Amer. Math. Soc. 104 (1988), 273 -- 276. [12] E. Grinberg and G. Zhang, Convolutions, transforms, and convex bodies, Proc. London Math. Soc. (3) 78 (1999), 77 -- 115. [13] B. Grunbaum, Measures of symmetry for convex sets, Proc. Sympos. Pure Math. 7, (1963), 233 -- 270. [14] C. Haberl, Blaschke valuations, Amer. J. Math., 133, (2011), 717 -- 751. [15] C. Haberl and F. Schuster, General Lp affine isoperimetric inequalities. J. Differential Geometry 83 (2009), 1-26. [16] C. Haberl, E. Lutwak, D. Yang and G. Zhang, The even Orlicz Minkowski problem, Adv. Math. 224 (2010), 2485-2510. [17] D. Klain, Star valuations and dual mixed volumes, Adv. Math. 121 (1996), 80-101. [18] D. Klain, Invariant valuations on star-shaped sets, Adv. Math. 125 (1997), 95-113. [19] G. Kuperberg, From the Mahler Conjecture to Gauss Linking Integrals, Geometric And Functional Analysis 18 (2008), 870 -- 892. [20] M. Ludwig, Ellipsoids and matrix valued valuations, Duke Math. J. 119 (2003), 159-188. [21] M. Ludwig, Minkowski areas and valuations, J. Differential Geometry, 86, (2010), 133 -- 162. 21 [22] M. Ludwig and M. Reitzner, A Characterization of Affine Surface Area, Adv. in Math. 147 (1999), 138-172. [23] M. Ludwig and M. Reitzner, A classification of SL(n) invariant valuations. Annals of Math. 172 (2010), 1223-1271. [24] E. Lutwak, The Brunn-Minkowski-Firey theory I : Mixed volumes and the Minkowski problem, J. Differential Geom. 38 (1993), 131 -- 150. [25] E. Lutwak, The Brunn-Minkowski-Firey theory II : Affine and geominimal surface areas, Adv. Math. 118 (1996), 244-294. [26] E. Lutwak and G. Zhang, Blaschke-Santal´o inequalities, J. Differential Geom. 47, (1997), 1 -- 16. [27] E. Lutwak, D. Yang and G. Zhang, Lp affine isoperimetric inequalities, J. Differential Geom. 56 (2000), 111 -- 132. [28] E. Lutwak, D. Yang and G. Zhang, A new ellipsoid associated with convex bodies, Duke Math. J. 104 (2000), 375 -- 390. [29] E. Lutwak, D. Yang and G. Zhang, Sharp Affine Lp Sobolev inequalities, J. Differential Geom. 62 (2002), 17 -- 38. [30] E. Lutwak, D. Yang and G. Zhang, The Cramer -- Rao inequality for star bodies, Duke Math. J. 112 (2002), 59-81. [31] E. Lutwak, D. Yang and G. Zhang, Volume inequalities for subspaces of Lp, J. Differ- ential Geom. 68 (2004), 159 -- 184. [32] M. Meyer and E. Werner, The Santalo-regions of a convex body, Transactions of the AMS 350 (1998), 4569 -- 4591. [33] M. Meyer and E. Werner, On the p-affine surface area. Adv. Math. 152 (2000), 288 -- 313. [34] M. Meyer, C. Schutt and E. Werner, New affine measures of symmetry for convex bodies, Adv. Math. 228, (2011), 2920 -- 2942. [35] F. Nazarov, The Hormander proof of the Bourgain-Milman theorem, preprint, 2009. [36] F. Nazarov, F. Petrov, D. Ryabogin and A. Zvavitch, A remark on the Mahler conjecture: local minimality of the unit cube, Duke Mathematical J. 154, (2010), 419 -- 430. [37] G. Paouris, Concentration of mass on convex bodies, Geometric and Functional Analysis 16, (2006), 1021 -- 1049. [38] C. M. Petty, Isoperimetric problems, Proc. Conf. Convexty and Combinatorial Geometry Univ. Oklahoma 1971, University of Oklahoma, (1972), 26 -- 41. [39] S. Reisner, Zonoids with minimal volume-product, Math. Z. 192 (1986), 339 -- 346. [40] S. Reisner, Minimal volume product in Banach spaces with a 1-unconditional basis, J. London Math. Soc.36 (1987), 126 -- 136. [41] S. Reisner, C. Schutt and E. Werner, Mahler's conjecture and curvature International Mathematics Research Notices, DOI 10.1093/imrn/rnr003 (2011) [42] J. Saint-Raymond, Sur le volume des corps convexes sym´etriques, S´eminaire d'Initiation `a l'Analyse, 1980-1981, Universit´e PARIS VI, Paris 1981. 22 [43] R. Schneider, Convex Bodies: The Brunn-Minkowski Theory, Encyclopedia of Mathe- matics and its Applications 44, Cambridge University Press, Cambridge (1993). [44] F. Schuster, Crofton measures and Minkowski valuations, Duke Math. J. 154 (2010), 1 -- 30. [45] F. Schuster and M. Weberndorfer, Volume Inequalities for Asymmetric Wulff Shapes, J. Differential Geom., in press. [46] C. Schutt and E. Werner, The convex floating body, Math. Scand. 66, (1990), 275 -- 290. [47] C. Schutt and E. Werner The convex floating body of almost polygonal bodies, Geom. Dedic. 44, (1992), 169 -- 188. [48] C. Schutt and E. Werner, Polytopes with Vertices Chosen Randomly from the Boundary of a Convex Body, Geom. Aspects of Funct. Analysis, Lecture Notes in Math. 1807, (2003), 241 -- 422. [49] C. Schutt and E. Werner, Surface bodies and p-affine surface area. Adv. Math. 187 (2004), 98-145. [50] J. E. Spingarn, An Inequality for Sections and Projections of a Convex Set, Proc. Amer. Math. Soc., 118, No. 4, (1993), 1219 -- 1224. [51] A. Stancu, The Discrete Planar L0-Minkowski Problem. Adv. Math. 167 (2002), 160-174. [52] A. Stancu, On the number of solutions to the discrete two-dimensional L0-Minkowski problem. Adv. Math. 180 (2003), 290-323. [53] E. Werner and D. Ye, New Lp affine isoperimetric inequalities, Adv. Math. 218 (2008), no. 3, 762-780. [54] E. Werner and D. Ye, Inequalities for mixed p-affine surface area, Math. Ann. 347 (2010), 703-737. [55] G. Zhang, Restricted chord projection and affine inequalities, Geom. Dedicata, 39 (1991), 213 -- 222. [56] G. Zhang, Intersection bodies and Busemann-Petty inequalities in R4, Annals of Math. 140 (1994), 331-346. [57] G. Zhang, A positive answer to the Busemann-Petty problem in four dimensions, Annals of. Math. 149 (1999), 535-543. [58] G. Zhang, New Affine Isoperimetric Inequalities, ICCM 2007, Vol. II, 239-267. Mathieu Meyer Universit´e de Paris Est - Marne-la-Vall´ee Equipe d'Analyse et de Math´ematiques Appliqu´ees Cit´e Descartes - 5, bd Descartes Champs-sur-Marne 77454 Marne-la-Vall´ee, France [email protected] Carsten Schutt Christian Albrechts Universitat Mathematisches Seminar 23 24098 Kiel, Germany [email protected] Elisabeth Werner Department of Mathematics Case Western Reserve University Cleveland, Ohio 44106, U. S. A. [email protected] Universit´e de Lille 1 UFR de Math´ematiques 59655 Villeneuve d'Ascq, France 24
1411.7292	3	1411	2016-02-27T09:52:54	A convenient notion of compact set for generalized functions	[ "math.FA" ]	We introduce the notion of functionally compact sets into the theory of nonlinear generalized functions in the sense of Colombeau. The motivation behind our construction is to transfer, as far as possible, properties enjoyed by standard smooth functions on compact sets into the framework of generalized functions. Based on this concept, we introduce spaces of compactly supported generalized smooth functions that are close analogues to the test function spaces of distribution theory. We then develop the topological and functional analytic foundations of these spaces.	math.FA	math	A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS PAOLO GIORDANO AND MICHAEL KUNZINGER Abstract. We introduce the notion of functionally compact sets into the theory of nonlinear generalized functions in the sense of Colombeau. The motivation behind our construction is to transfer, as far as possible, properties enjoyed by standard smooth functions on compact sets into the framework of generalized functions. Based on this concept, we introduce spaces of compactly supported generalized smooth functions that are close analogues to the test function spaces of distribution theory. We then develop the topological and functional analytic foundations of these spaces. 1. Introduction A main advantage of nonlinear generalized functions in the sense of Colombeau as compared to Schwartz distributions is the fact that they can be viewed as set- theoretic maps on domains consisting of generalized points. This change of perspec- tive allows to develop several branches of the theory in close analogy to classical analysis, and thereby has become increasingly important in recent years (cf., e.g., [26, 4, 2, 11, 12, 27, 3, 16, 6, 18]). In particular, appropriate topologies on spaces of nonlinear generalized functions, the so called sharp topologies (see below for the definition), have been introduced in [28, 29] and have since been studied by many authors. Apart from their central position in the structure theory of Colombeau algebras, they also supply the foundation for applications in the theory of nonlinear partial differential equations (e.g., for a suitable concept of well-posedness). From the point of view of analysis, a key notion underlying many existence re- sults is that of compactness. It turns out, however, that sharply compact subsets of generalized points display certain unwanted properties: e.g., no infinite subset of Rn is sharply compact since the trace of the sharp topology on subsets of Rn is 2010 Mathematics Subject Classification. 46F30,46A13,13J99. Key words and phrases. Functionally compact sets, Colombeau generalized functions, generalized smooth functions, locally convex modules. P. Giordano has been supported by grant P25116 and P25311 of the Austrian Science Fund FWF. M. Kunzinger has been supported by grants P23714 and P25326 of the Austrian Science Fund FWF. 1 A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 2 discrete. In fact, this is a necessary consequence of the set eR of generalized num- bers containing actual infinitesimals, hence seems unavoidable also in alternative approaches, cf. [17, Prop. 2.1] and [18, Thm. 25]. The importance of a convenient notion of compactness for nonlinear general- ized functions has been recognized by several authors, most recently in [1]. The approach we take in the present paper is to introduce an appropriate concept of compactly supported generalized function, and then to study spaces consisting of such functions, which, in analogy to the test function space D(Ω) in distribution theory, we denote by GD(U ). The domain U here is a set of generalized points. Based on Garetto's theory of locally convex eC-modules [11, 12, 14] we then de- velop the topological and functional analytic foundations of these spaces. We find that they are indeed close analogues of the classical spaces of test functions in that they are countable strict inductive limits of complete metric spaces GDK(U ) (ana- logues of DK(Ω) in distribution theory) satisfying properties paralleling those of the classical strict (LF)-spaces D(Ω). The plan of the paper is as follows: In the remainder of this introduction we fix some basic notions used throughout this work. Section 2 introduces what we call functionally compact sets, based on work by Oberguggenberger and Vernaeve in [27]. Building on this, in Section 3 we define compactly supported generalized smooth functions (GSF), as well as the corresponding spaces GD(U ) and GDK(U ). We also show that every Colombeau generalized function f ∈ Gs(Ω) (in particular, every Schwartz distribution) defines a compactly supported GSF ¯f : eR −→ eR that coincides with f on eΩc. In order to obtain appropriate topologies on these spaces, the basic properties of classical norms, yet take values in eR, thereby generalizing we define so-called generalized norms in Section 4. These are maps that share a standard alternative description of the sharp topology on generalized numbers (cf. [4, 17]). In Sections 5 and 6 these generalized norms are employed to endow the spaces GDK(U ) with metric topologies. In particular, in Section 5.1 we study connections between non-Archimedean properties and Hausdorff topological vector spaces of generalized functions, proving an impossibility theorem: there does not exist a Hausdorff topological vector subspace of the Colombeau special algebra which contains the Dirac delta and even a single trace of an open set of the sharp topology. The completeness of the spaces GDK(U ) is established in Section 7. In the final Section 8 we derive the fundamental functional analytic properties of the space GD(U ). 1.1. Basic notions. Our main references for Colombeau's theory are [8, 9, 25, 22]. The special Colombeau algebra Gs(Ω) over an open subset Ω of Rn is defined as the quotient E s M (Ω)/N s(Ω), where (setting I := (0, 1] and noting that in the naturals A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 3 N = {0, 1, 2, 3 . . .} we include zero.) M (Ω) := {(uε) ∈ C∞(Ω)I ∀K ⋐ Ω∀α ∈ Nn ∃N ∈ N : sup x∈K ∂αuε(x) = O(ε−N )} E s N s(Ω) := {(uε) ∈ C∞(Ω)I ∀K ⋐ Ω∀α ∈ Nn ∀m ∈ N : sup x∈K ∂αuε(x) = O(εm)}. M (Ω) are called moderate, those of N s(Ω) are called negligible. Nets Elements of E s in E s M (Ω) are written as (uε), and u = [uε] denotes the corresponding equivalence class in Gs(Ω). For (uε) ∈ N s(Ω) we also write (uε) ∼ 0. We will abbreviate 'Colombeau generalized function' by CGF. Gs(−) is a fine sheaf of differential al- gebras and there exist sheaf embeddings (based on smoothing via convolution) of the space of Schwartz distributions D′ into Gs (cf. [22]). Given Ω ⊆ Rn open, the space of generalized points in Ω is eΩ = ΩM / ∼, where ΩM = {(xε) ∈ ΩI ∃N ∈ N : xε = O(ε−N )} is called the set of moderate nets and (xε) ∼ (yε) if xε − yε = O(εm) for every m ∈ N. In the particular case Ω = R we obtain the ring of Colombeau generalized numbers (CGN) eR = RM / ∼ (and analogously for eC), which can also be written as eR = RM /N s, where N s is the set of all negligible nets of real numbers (xε) ∈ RI , i.e. such that (xε) ∼ 0. eR is an ordered ring with respect to its natural order relation: x ≤ y iff there are representatives (xε) and (yε) such that xε ≤ yε for ε sufficiently small. We point out that, in the present work, the notion x > y does not mean x ≥ y and x 6= y. Rather, it is to be understood as x − y ≥ 0 and x − y invertible. By [22, 1.2.38] and [24, Prop. 3.2] we have: Lemma 1. Let x ∈ eR. Then the following are equivalent: (i) (ii) For each representative (xε) of x there exists some ε0 and some m such that x > 0. xε > εm for all ε < ε0. (iii) For each representative (xε) of x there exists some ε0 such that xε > 0 for all ε < ε0. We shall use the notation dεm := [εm] ∈ eR for any m ∈ R. Hence x > 0 is equivalent to x ≥ dεm for some m > 0. If P(ε) is a property of ε ∈ I, we will also sometimes use the notation ∀0ε : P(ε) to denote ∃ε0 ∈ I ∀ε ∈ (0, ε0] : P(ε). The space of compactly supported generalized points eΩc is defined by Ωc/ ∼, where Ωc := {(xε) ∈ ΩI ∃K ⋐ Ω∀0ε : xε ∈ K} and ∼ is the same equivalence relation as in the case of eΩ. Concerning intervals, we use the following notations: [a, b] := {x ∈ eR a ≤ x ≤ b}, [a, b]R := [a, b] ∩ R. Also, for x, y ∈ eRn we write x ≈ y if x − y is infinitesimal, i.e. if x − y ≤ r for all r ∈ R>0. As already indicated above, the natural topology for Colombeau-type spaces is the so-called sharp topology ([7, 29, 28, 4, 5, 23, 16]). This topology is generated A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 4 ([2, 3, 17]). For Euclidean balls, we will write B E by balls Bρ(x) = {y ∈ eRn y − x < ρ}, where − is the natural extension of the Euclidean norm to eRn, [xε] := [xε] ∈ eR, and ρ ∈ eR>0 is strictly positive ρ (x) = {y ∈ Rn y − x < ρ}. On the other hand, the so-called Fermat-topology on eRn (see [17, 18]) is generated by the balls Br(x) for x ∈ eRn and r ∈ R>0. Originally, the sharp topology was introduced using an ultrametric as follows: The map v : RM −→ (−∞,∞] v((uε)) := sup{b ∈ R uε = O(εb)}. gives a pseudovaluation on eR. Then setting − e : eR → [0,∞), ue := exp(−v(u)) provides a translation-invariant complete ultrametric ds : eR ×eR −→ R+ ds(u, v) := u − ve on eR, which induces the sharp topology on eR. Garetto in [11, 12] extended the above construction to arbitrary locally convex spaces by functorially assigning a space of CGF GE to any given locally convex space E. In this approach, the seminorms of E are used to define pseudovaluations which induce a generalized locally convex topology on the eC-module GE, again called sharp topology. In the present paper, we will exclusively work with eR-modules. We note, however, that all our constructions trivially carry over to the eC-case. For any S ⊆ I, eS denotes the equivalence class in eR of the characteristic function of S (cf. [4, 30]). Any eS is an idempotent, and eS + eSc = 1. Also, eS 6= 0 if and only if 0 ∈ S. For any subset A of eRn, its interleaving (cf. [27]) is defined as eSj aj m ∈ N,{S1, . . . , Sm} a partition of I, aj ∈ A interl(A) :=   If (Aε) is a net of subsets of Rn then the internal set ([27, 31]) generated by (Aε) mXj=1 . is and the strongly internal set ([18]) generated by (Aε) is [Aε] =n[xε] ∈ eRn xε ∈ Aε for ε smallo , hAεi :=n[xε] ∈ eRn xε ∈ε Aεo . Here, xε ∈ε Aε means that xε ∈ Aε for ε small and that the same property holds for any representative of [xε]. The net (Aε) is called sharply bounded if there exists some N ∈ R>0 such that for ε sufficiently small we have supx∈Aε x ≤ ε−N . Equivalently, we have that (Aε) is sharply bounded if there exists ρ ∈ eR>0 such that [Aε] ⊆ Bρ(0). A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 5 f : X −→ Y is a generalized smooth function (GSF) Finally, given X ⊆ eRn and Y ⊆ eRd, then (see [18]) if there exists a net uε ∈ C∞(Ωε, Rd) defining f in the sense that X ⊆ hΩεi, f ([xε]) = [uε(xε)] ∈ Y and (∂αuε(xε)) ∈ Rd M for all x = [xε] ∈ X and all α ∈ Nn. The space of GSF from X to Y is denoted by GC∞(X, Y ) (in contrast to [18], where the notation eG(X, Y ) was used). GSF are a natural generalization of CGF to general domains. In particular, for any Ω ⊆ Rn open, GC∞(eΩc) ≃ Gs(Ω). GSF on subsets of eRn, endowed with the sharp topology, form a sub-category of the category of topological spaces. In particular, they can be composed unrestrictedly. 2. A new notion of compact subset for nonlinear generalized functions Even though the intervals [a, b] ⊆ eR, a, b ∈ R, are neither compact in the sharp nor in the Fermat topology (see [18, Thm. 25]), analogously to the case of smooth functions, a GSF satisfies an extreme value theorem on such sets. In fact, we have: Proposition 2. Let f ∈ GC∞(X,eR) be a generalized smooth function defined on the subset X of eRn. Let ∅ 6= K = [Kε] ⊆ X be an internal set generated by a sharply bounded net (Kε) of compact sets Kε ⋐ Rn , then ∃m, M ∈ K ∀x ∈ K : f (m) ≤ f (x) ≤ f (M ). (2.1) Proof. By [18, Lem. 28], f can be represented by a net uε ∈ C∞(Rn, Rd). Since K 6= ∅, for ε sufficiently small, say for ε ∈ (0, ε0], Kε is non-empty and, by assumption, it is also compact. For all ε ∈ (0, ε0] we have ∃mε, Mε ∈ Kε ∀x ∈ Kε : uε(mε) ≤ uε(x) ≤ uε(Mε). Since the net (Kε) is sharply bounded, both the nets (mε) and (Mε) are moderate. Therefore m = [mε], M = [Mε] ∈ K ⊆ X. Take any x ∈ [Kε], then there exists a representative (xε) such that xε ∈ Kε for ε small. Therefore f (m) = [uε(mε)] ≤ [uε(xε)] = f (x) ≤ f (M ). (cid:3) We shall use the assumptions on K and (Kε) given in this theorem to introduce a new notion of "compact subset" which behaves better than the usual classical notion of compactness in the sharp topology. Definition 3. A subset K of eRn is called functionally compact, denoted by K ⋐f eRn, if there exists a net (Kε) such that (i) K = [Kε] ⊆ eRn (ii) (iii) ∀ε ∈ I : Kε ⋐ Rn (Kε) is sharply bounded A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 6 If, in addition, K ⊆ U ⊆ eRn then we write K ⋐f U . Finally, we write [Kε] ⋐f U if (ii), (iii) and [Kε] ⊆ U hold. We note that in (iii) it suffices to ask that Kε be closed since it is bounded by (ii), at least for ε small. In fact, we have: Lemma 4. A subset K of eRn is functionally compact if and only if it is internal and sharply bounded. Proof. By [27, Lemma 2.4 and Cor. 2.2], every sharply bounded internal set K has a sharply bounded representative (Kε) consisting of closed (hence compact) subsets of Rn. (cid:3) We motivate the name functionally compact subset by anticipating that on this type of subsets, GSF have properties very similar to those that ordinary smooth functions have on standard compact sets. Remark 5. (i) By [27, Prop. 2.3], any internal set K = [Kε] is closed in the sharp topology. it is not closed. (ii) compact. In particular, [0, 1] = ^[0, 1]R = [[0, 1]R] is functionally compact. In particular, the open interval (0, 1) ⊆ eR is not functionally compact since If H ⋐ Rn is a non-empty ordinary compact set, then eH = [H] is functionally (iii) The empty set ∅ =e∅ ⋐f eR. (iv) By Lemma 4, eRn is not functionally compact since it is not sharply bounded. (v) The set of compactly supported points eRc is not functionally compact because the GSF f (x) = x does not satisfy the conclusion (2.1) of Prop. 2. We start the study of functionally compact sets by proving suitable generaliza- tions of theorems from classical analysis. Theorem 6. Let K ⊆ X ⊆ eRn, f ∈ GC∞(X,eRd). Then K ⋐f eRn implies f (K) ⋐f eRd. Proof. Let (Kε) be as in Def. 3 and let the GSF f be defined by the net uε ∈ C∞(Rn, Rd). Let us first prove that f (K) = [uε(Kε)]. In fact, y ∈ f (K) = f ([Kε]) is equivalent to ∃(xε) ∈ Rn M ∀0ε : xε ∈ Kε and y = [uε(xε)]. (2.2) This necessary entails y ∈ [uε(Kε)]. Vice versa, if y ∈ [uε(Kε)], then there exists (yε) ∈ Rd M such that yε ∈ uε(Kε) for ε small. Hence, for each of these ε there also exists xε ∈ Kε such that yε = uε(xε), which implies y = [uε(xε)], i.e. (2.2) since (Kε) is sharply bounded. Clearly, uε(Kε) ⋐ Rd, so it remains to prove that the A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 7 net (uε(Kε)) is sharply bounded. If ∀ε0 ∃ε ≤ ε0 : Kε = ∅, then [Kε] = K = ∅, so f (K) = ∅ and the conclusion is trivial. Otherwise, assume that Kε 6= ∅ for ε ≤ ε0 and proceed by contradiction assuming that ∀k ∈ N∃(εkn)n ↓ 0 ∀n∃ykn ∈ uεkn (Kεkn ) : ykn > ε−k kn . (2.3) We can write ykn = uεkn (xkn) for some xkn ∈ Kεkn . Next, set ε0 := ε00 and for k , εk−1(cid:1) and set εk := εknk . Take any ¯xε ∈ Kε k > 0 pick nk such that εknk < min(cid:0) 1 for each ε ≤ ε0, and set xε := xknk if ε = εk and xε := ¯xε otherwise. Then xε ∈ Kε for ε ≤ ε0, so x = [xε] ∈ K ⊆ X and (uε(xε)) ∈ Rd M by the definition of GSF, which contradicts (2.3). (cid:3) As a corollary of this theorem and Rem. (5).(ii) we get a = [ε−N ], b = [ε−M ] with M > N . Corollary 7. If a, b ∈ eR and a ≤ b, then [a, b] ⋐f eR. Let us note that a, b ∈ eR can also be infinite, e.g. a = [−ε−N ], b = [ε−M ] or Lemma 8. Let K, H ⋐f eRn, then we have: (i) K ∪ H ⊆ interl(K ∪ H) ⋐f eRn If K ∪ H is internal, then it is functionally compact If K ∩ H is internal, then it is functionally compact. (ii) (iii) Proof. (i) follows from [27, Prop. 2.8] which implies K ∪ H ⊆ interl(K ∪ H) = [Kε ∪ Hε], where the nets (Kε) and (Hε) satisfy Def. 3. Property (ii) follows from [27, Lemma 2.7] which implies that if the union of internal sets is internal, then it is equal to its interleaving. Property (iii) is a consequence of Lemma 4. (cid:3) Lemma 4 is the following: If H ⊆ K ⋐f eRn, then also H is sharply bounded. So, another consequence of Corollary 9. Let H ⊆ K ⋐f eRn, then H internal implies H ⋐f eRn. Finally, in the following result we consider the product of functionally compact sets: Proposition 10. Let K ⋐f eRn and H ⋐f eRd, then K × H ⋐f eRn+d. In particular, if ai ≤ bi for i = 1, . . . , n, then Qn Proof. From [27, Prop. 2.13] if follows that K × H is internal, in fact for K = [Kε], H = [Hε], K × H = [Kε × Hε]. From this representation it immediately follows that H × K is sharply bounded as well, so Lemma 4 gives the claim. i=1[ai, bi] ⋐f eRn. (cid:3) A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 8 3. Compactly supported generalized smooth functions Our main goal in this section is to define and study an analogue within GSF of the space DK(Ω) of smooth functions supported in a fixed compact set K ⋐ Ω. In order to define this space, we first try to define the concept of support of a GSF. Clearly, if ϕ ∈ D[−a,a]R(R), one would expect that ϕ should have compact support also if we think of ϕ as a GSF. In fact, supp(ϕ) should be contained in [−a, a] ⋐f eR. Already this basic requirement implies that if f ∈ GC∞(X, Y ), the natural definition S(f ) := X \[nBρ(x) ∩ X x ∈ X, ρ ∈ eR>0, fBρ(x)∩X = 0o doesn't fit with our intuition. Indeed, if we take the aforementioned ϕ so that ϕ(0) = 1, and S ⊆ (0, 1] such that 0 ∈ ¯S and 0 ∈ Sc, then ϕ(eS dε−1) = eSc 6= 0 and eS dε−1 ∈ S(ϕ) \ [−a, a]. This motivates the following Definition 11. Let X ⊆ eRn, Y ⊆ eRd and f ∈ GC∞(X, Y ), then where here (−) denotes the relative closure in X with respect to the sharp topology. Using this concept, we have supp(ϕ) ⊆ [−a, a]. In fact, ϕ(x) = [ϕ(xε)] > 0 supp(f ) := {x ∈ X f (x) > 0}, implies ϕ(xε) > εq for some q ∈ R>0 and for ε small, and hence xε ∈ [−a, a]R. Remark 12. (i) (ii) In the setting of Colombeau algebras, one usually defines the support of some f ∈ Gs(Ω) as a subset of Ω ⊆ Rn, i.e., as a set of classical points, namely as suppGs(f ) := Ω \S(cid:8)Br(x) ∩ Ω x ∈ Ω, r ∈ R>0, fBr (x)∩Ω = 0(cid:9), where the last equality has to be understood in Gs(Br(x) ∩ Ω). Using X = eΩc as the natural domain of any f ∈ Gs(Ω) (cf. [18, Thm. 37]), it is then immediate that supp(f ) ∩ Ω ⊆ suppGs(f ). Let u ∈ D′(Ω) be a Schwartz distribution and denote by ι : D′(Ω) → GC∞(eΩc, R) a standard embedding via convolution. Then supp(ι(u)) ∩ Ω ⊆ (iii) Assume that the embedding ι : D′(Ω) → GC∞(eΩc, R) has been defined by using a mollifier ρ ∈ S(Rn) which is identically equal to 1 in the ball B E p (0), p ∈ R>0. Then δ(x) = dε−n for each x ∈ Bp·dε(0) and hence Bp·dε(0) ⊆ supp(ι(δ)), whereas supp(ι(δ)) ∩ Rn = {0}. In general, supp(f ) is not an internal set because it is not generally closed supp(u), as follows from (i) and the fact that ι is a sheaf-morphism. (iv) by finite interleaving (see [27, Lem. 2.7]). Consider e.g. X with only near standard points and f ∈ GC∞(X,eR) which is strictly positive on two disjoint intervals. However, if X itself is closed under finite interleaving then so is supp(f ). A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 9 If (uε) defines f ∈ GC∞(X, Y ), the internal set [supp(uε)] is not intrinsically defined since it depends on the defining net (uε). Consider, e.g., uε(x) := ϕ(x) + ε1/ε > 0 where ϕ ∈ C∞(R, R In our further analysis we will repeatedly make use of the following notion: ≥0). Definition 13. For A ⊆ eRn we call the set the strong exterior of A. ext(A) := {x ∈ eRn ∀a ∈ A : x − a > 0} This set can also be described in the following way: Lemma 14. If A ⊆ eRn, then ext(A) = {x ∈ eRn ∀S ⊆ I : eS 6= 0 ⇒ xeS 6∈ AeS}. Proof. ⊆: Let x = [xε] ∈ ext(A) and suppose that there exists some S ⊆ I with eS 6= 0 and some a = [aε] ∈ A such that xeS = aeS. Then there exists some q > 0 such that xε − aε > εq for ε small. However, xeS = aeS implies that xε − aε = O(εq+1) for ε → 0, ε ∈ S, a contradiction. ⊇: If there exists some a = [aε] ∈ A such that x − a 6> 0 then there is a k for all k. Letting S := {εk k ∈ N} implies sequence εk ↓ 0 with xεk − aεk < εk xeS = aeS. (cid:3) For non-trivial internal sets we have the following characterization of the strong exterior: Lemma 15. Let ∅ 6= [Kε] = K ⊆ eRn. Then ext(K) = hK c εi. Proof. Suppose first that x ∈ hK c εi and let S ⊆ I with eS 6= 0. Suppose that there existed some a ∈ K with xeS = aeS. Since a ∈ K there exists a representative (aε) of a with aε ∈ Kε for all ε. Then xeS = aeS implies that there exists a representative (xε) of x and a sequence εk ↓ 0 in S with xεk = aεk ∈ Kεk for all k. But this contradicts the fact that x ∈ hK c εi. Conversely, if x 6∈ hK c εi then there exists a representative (xε) of x and a sequence εm ↓ 0 with xεm ∈ Kεm for all m. Since K 6= ∅, there exists some w = [wε] ∈ K with wε ∈ Kε for all ε. Now let aε :=  and set S := {εm m ∈ N}. Then a = [aε] ∈ [Kε] and xeS = aeS by construction. Thus xeS ∈ KeS, and so x 6∈ ext(K). if ε = εm otherwise xεm wε (cid:3) As an immediate conclusion we obtain: A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 10 Corollary 16. Let K = [Kε] = [Lε] 6= ∅. Then hK c εi = hLc εi. The next result relates the support of a GSF to the exterior of certain internal sets. To formulate it concisely, we introduce the following notations: Denote by Kf the set of all internal ∅ 6= K ⊆ eRn with ext(K) 6= ∅ and such that there exists a net uε ∈ C∞(Rn, Rd) that defines f and such that [uε(xε)] = 0 for all [xε] ∈ ext(K). Also, denote by Hf the set of all the internal sets of the form K = [supp(uε)] ⊆ eRn for some net uε ∈ C∞(Rn, Rd) that defines f and such that both K and ext(K) are non empty. Then we have: Lemma 17. Let X ⊆ eRn, Y ⊆ eRd and f ∈ GC∞(X, Y ). Then supp(f ) ⊆ X ∩ \K∈Kf K ⊆ X ∩ \K∈Hf Proof. Since X ∩TK∈Kf first inclusion in (3.1), it suffices to prove that K. (3.1) K is a sharply closed subset of X, in order to show the {x ∈ X f (x) > 0} ⊆ X ∩ \K∈Kf K. Let x ∈ X be such that f (x) > 0, so that ∃r ∈ R>0 : f (x) > dεr. (3.2) Let K = [Kε] ∈ Kf , and assume, by contradiction, that x = [xε] /∈ K. We first prove that (xεk ) ⊆ K c εk , where r comes from (3.2). In fact, suppose to the contrary that ∃q ∈ N>r ∃(εk)k∈N ↓ 0 ∀k ∈ N : B E εq k (3.3) ∀q ∈ N>r ∃εq ∀ε ≤ εq ∃y(q) ε ∈ B E εq (xε) : y(q) We may assume that (εq)q∈N ↓ 0. Setting yε := y(q) x = [yε] ∈ K, which contradicts x /∈ K. ε By assumption ∃z = [zε] ∈ ext(K), and hence ε ∈ Kε. for ε ∈ (εq+1, εq], we have ∃s ∈ N>q ∀0ε : d(zε, Kε) > εs, where q comes from (3.3). Using (εk)k∈N from (3.3), we set xε := xεk if ε = εk and xε := zε otherwise. Then x := [xε] ∈ ext(K). But K ∈ Kf , so there exists a net uε ∈ C∞(Rn, Rd) that defines f and such that [uε(xε)] = 0. In particu- lar, uεk (xεk ) = uεk (xεk ) = O(ε2s k ) as k → +∞, which contradicts f (x) = [uε(xε)] > dεr > dεs. Turning now to the second inclusion, assume that x ∈ X\TK∈Hf exists a net (uε) that defines f , with ∅ 6= K := [supp(uε)] ⊆ eRn and ext(K) 6= ∅, K, so that there A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 11 but such that x /∈ K. We want to show that x /∈ TK ′∈Kf K′. But for all [yε] ∈ ext(K) = hsupp(uε)ci, we have yε /∈ supp(uε) for ε small. Hence, uε(yε) = 0 for these ε, and this yields [uε(yε)] = 0. This proves that K ∈ Kf , but x /∈ K. (cid:3) Remark 18. (i) Using methods from Nonstandard Analysis, one can prove the converse of the K. first inclusion in (3.1) in the following case: If X = eRn and the sharp interior of nx ∈ eRn f (x) = 0o is non-empty, then supp(f ) = \K∈Kf K and choose some z in the interior of ny ∈ eRn To see this, let x ∈ TK∈Kf f (y) = 0} and some q > 0 such that Bdεq (z) ⊆ny ∈ eRn f (y) = 0o. Given a representative (vε) of f , let (χε) be a moderate net of smooth functions such that χε vanishes on B E εq+1 (zε). Then setting uε := χεvε gives a new representative of f such that [{y ∈ Rn uε(y) = 0}] has non-empty sharp interior. Let mε ∈ N, mε → ∞ as ε → 0. Then set Kε := {y ∈ Rn uε(y) ≥ εmε} and K := [Kε]. It follows that ∅ 6= ext(K) ⊆ {y ∈ Rn f (y) = 0}, so x ∈ K. Fix any k ∈ N. Now using nonstandard notation, letting ρ := [ε], u := [uε], and fixing a representative of x, which we temporarily simply denote by x, the above in particular implies εq+2 (zε) and is identically equal to 1 on Rn \ B E ∗N ∞ ⊆ {m ∈ ∗N ∃yk ∈ ∗Rn : yk − x ≤ ρk ∧ u(yk) ≥ ρm}. By the underspill principle, therefore, there also exists some m ∈ N and some yk ∈ ∗Rn such that yk−x ≤ ρk and u(yk) ≥ ρm. Taking equivalence classes of these nets, it follows that there exist yk ∈ eRn such that yk − x ≤ dεk and f (yk) > 0. Since yk → x in the sharp topology, this implies that x ∈ supp(f ). (ii) The assumption ext(K) 6= ∅ in the definition of Kf is essential. To illustrate this, take ϕ as defined before Def. 11, pick any sequence εk ↓ 0 and set Kε := {0} for ε = εk, and Kε = Rn otherwise. Then ext(K) = ∅, and obviously supp(ϕ) 6⊆ K. (iii) The question of whether the reverse of the last inclusion in (3.1) also holds remains open. In the following section we shall see that even though the notion of support intro- duced above may not be entirely satisfactory, there nevertheless is a very convenient notion of being compactly supported for GSF. A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 12 3.1. The spaces GDK(U, Y ) and GD(U, Y ). A frequently used idea to solve prob- lems like the previous ones comes by considering a family of GSF having "good rep- resentatives", i.e. possessing a defining net (uε) that conforms to our intuition and includes the examples we have in mind. In the following, we denote by (u1, . . . , ud) the components of a function u which takes values, e.g., in Rd. Definition 19. Let ∅ 6= K ⋐f U ⊆ eRn and Y ⊆ eRd, then we say that f is a GSF compactly supported in K, and we write f ∈ GDK(U, Y ) if f ∈ GC∞(U, Y ) and there exists a net (uε) such that: (uε) defines f , where uε ∈ C∞(Rn, Rd) for all ε. (i) ∀α ∈ Nn ∀[xε] ∈ ext(K) : (ii) [∂αuε(xε)] = 0. Moreover, we set GD(U, Y ) := [∅6=K⋐fU GDK(U, Y ). We will simply use the symbols GDK(U ) and GD(U ) if Y = eR. Remark 20. (i) (ii) Lemma 17 implies that if f ∈ GDK(U, Y ), then supp(f ) ⊆ K because K ∈ Kf . The converse implication for an arbitrary subset U remains an open problem. For the case U = eRn, and if supp(f ) is not empty, see Thm. 27 It is clear that, in general, another net defining f : U −→ eRd will not nec- essarily satisfy (ii) of Def. 19 because such a net is not bound to have any below. particular behavior outside of U . (iii) Set Kε := B E 1 (0) and Lε := Kε \ B E εi in (ii) instead of the simpler [K c ε -mesh of points for Kε = B E e−1/ε(0), then for the Hausdorff distance of Kε and Lε we obtain dH(Kε, Lε) = e− 1 ε . By [27, Cor. 2.10], it fol- lows that [Kε] = [Lε]. If we consider a net of smooth functions such that 1/2(0) = 1, uεR2\Kε = 0, uε ≥ 0, then supy∈R2\Kε ∂αuε(y) = 0 but uεBE supy∈R2\Lε ∂αuε(y) = 1. This motivates the use of the strongly internal set hK c ε ]. Analogously, we can consider as Lε an e− 1 1 (0). Lε may also contain points in K c ε, but so that d(x, Kε) = e− 1 ε . This example shows clearly that we need to be sufficiently "far" from ∂Kε to be sure that [∂αuε(xε)] = 0, i.e. at points x = [xε] such that [d(xε, Kε)] > 0, as stated in (ii). Since ext(K) is indepen- dent of the choice of representative of K by Cor. 16, so is Def. 19. This is essential to prove the completeness of the spaces GDK(U ) and GD(U ). The following result will turn out to be useful when proving results by contradic- tion in several instances below. It permits to restrict the analysis to only two cases: A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 13 points in K or in ext(K). To state it more clearly, we say that a generalized point [yε] joins points of the sequence (yk)k∈N at (εk)k∈N if ∀N ∈ N∃k ≥ N : yεk = yk. Lemma 21. Let K = [Kε] ⋐f eRn, (εk)k∈N a sequence in (0, 1] which strictly decreases to 0, and (yk)k∈N a sequence in Rn. For each k ∈ N, let xk ∈ Kεk be such that d(yk, xk) = d(yk, Kεk ). Then either (i) or (i) ∃[yε] ∈ ext(K) : [yε] joins points of the sequence (yk)k∈N at (εk)k∈N ∃[¯yε], [¯xε] joining points of the sequences (yk)k∈N and (xk)k∈N (respectively) at (εk)k∈N such that [¯yε] = [¯xε] and ¯xε ∈ Kε ∀ε. Proof. We can always pick a point eε ∈ K c ε so that e := [eε] ∈ ext(K); in fact, since (Kε) is sharply bounded, we can find eε ∈ Rn \ Kε so that d(eε, Kε) > 1 and (eε) is moderate. We can also take a point iε ∈ Kε for each ε because, without loss of generality, we can assume that Kε 6= ∅ for all ε. The first alternative in the statement is realized if ∃b ∈ R>0 ∃N ∈ N∀k ≥ N : yk − xk > εb k. (3.4) Set yε := yk if ε = εk and yε := eε otherwise. Then, if ε = εk, by (3.4) we have d(yε, Kε) = yk − xk > εb; otherwise d(yε, Kε) = d(eε, Kε) > 1 ≥ εb. Therefore, [yε] ∈ hK c εi = ext(K) and, of course, [yε] joins points of the sequence (yk)k∈N at (εk)k∈N. Vice versa, if ∀h ∈ N∃kh > h : ykh − xkh ≤ εh kh, (3.5) then we can set ¯yε := ykh, ¯xε := xkh if ε = εkh and ¯yε := ¯xε := iε otherwise. Then, if ε = εkh , by (3.5) we have ¯yε − ¯xε = ykh − xkh ≤ εh = εh; otherwise ¯yε − ¯xε = iε − iε = 0, and (i) follows. (cid:3) kh We use the above result to guarantee that the maximum values of any partial derivative ∂αf are attained on K and not outside, as precisely stated in the following Lemma 22. Let (uε) and K = [Kε] satisfy Def. 19, then Proof. By contradiction, assume that ∀α ∈ Nn ∀i = 0, . . . , n : (cid:20) sup y∈Rn ∂αui ε(y)(cid:21) 6=(cid:20) sup ∃a ∈ R>0 ∃εk ց 0 ∀k ∈ N : (cid:12)(cid:12)(cid:12)(cid:12) sup (cid:20) sup y∈Rn ∂αui x∈Kε ∂αui y∈Rn vεk (y) − sup For simplicity of notation, set vε := ∂αui ε. Inequality (3.7) means (3.6) (3.7) ε(y)(cid:21) =(cid:20) sup ε(x)(cid:21) . x∈Kε ∂αui ε(x)(cid:21) . x∈Kε vεk (x)(cid:12)(cid:12)(cid:12)(cid:12) > εa k. A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 14 Thus ∀k ∈ N∃yk ∈ Rn : εa k + sup x∈Kεk vεk (x) < vεk (yk) . (3.8) We can hence apply Lemma 21. In the first case (i) we have y := [yε] ∈ ext(K) so that [vε(yε)] = 0 by Def. 19 (ii). Since [yε] joins points of (yk)k∈N at (εk)k∈N, from (3.8) and [vε(yε)] = 0 we get εa k ≤ εa k + sup x∈Kεk vεk (x) < vεk (yk) = vεk (yεk ) < εa+1 k , for k sufficiently big, which gives a contradiction in the first case. In the second case, i.e., (i) of Lemma 21, we have ¯y := [¯yε] = ¯x := [¯xε] ∈ K ⊆ U . Therefore [vε(¯yε)] = [vε(¯xε)] by the definition of GSF. Since [¯yε] and [¯xε] join points of (yk)k∈N and (xk)k∈N, respectively, at (εk)k∈N, for k sufficiently big we have εa k + sup x∈Kεk vεk (x) < vεk (yk) ≤ vεk (yk) − vεk (xk) + vεk (xk) ≤ εa+1 k + sup x∈Kεk vεk (x) , leading to a contradiction also in the second case. (cid:3) The previous result will be essential to prove that any compactly supported GSF can be extended to the whole of eRn, and to define an eR-valued norm of f that does not depend on K. Using Lemma 22, we can prove that any derivative of a compactly supported GSF is globally bounded in an appropriate sense: Lemma 23. Let the net (uε) satisfies Def. 19, then β ≤ α ⇒(cid:20) sup ∀α ∈ Nn ∃C ∈ eR∀β ∈ Nn : y∈Rn ∂βuε(y)(cid:21) ≤ C. Proof. By the extreme value property Prop. 2 we can set Ciβ :=(cid:2)supy∈Kε ∂βui where β ∈ Nn, β ≤ α, and i = 1, . . . , n. Set C := 1 + √n maxβ≤α C > Ciβ and property (3.6) yields(cid:2)supy∈Rn ∂βui ε(y)(cid:3) =(cid:2)supx∈Kε ∂βui and hence (cid:2)supy∈Rn ∂βuε(y)(cid:3) < C. entire eRn: Theorem 24. Let ∅ 6= K ⋐f U ⊆ eRn and f ∈ GDK(U,eRd) be defined by (uε) which satisfies Def. 19. Then (uε) defines a GSF of the type eRn −→ eRd and there exists one and only one ¯f ∈ GDK(eRn,eRd) such that: Moreover, compactly supported GSF can be extended in a unique way to the Ciβ . Then ε(y)(cid:3), ε(x)(cid:3) = Ciβ (cid:3) 1≤i≤n (i) (ii) ¯fK = fK ¯fext(K) = 0. A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 15 Moreover, this ¯f satisfies (iii) (i) If U is sharply open and α ∈ Nn, then ∂α ¯fU = ∂αf ¯feRn can be identified with a Colombeau generalized function. c Proof. The existence part follows by showing that for all α ∈ Nn and all [xε] ∈ eRn M . This follows since Lemma 23 yields that for any α ∈ Nn we have (∂αuε(xε)) ∈ Rd there exists some (Cαε) ∈ Rn M such that supx∈Rn ∂αuε(x) ≤ Cαε for all ε small. To prove uniqueness, let g ∈ GDK(eRn,eRd) be such that gK = ¯fK = f and gext(K) = ¯fext(K) = 0. By contradiction, assume that ¯f (y) = [uε(yε)] 6= g(y) = [vε(yε)], for some y = [yε] ∈ eRn, where (vε) defines g. Thus ∃a ∈ R>0 ∃εk ց 0 ∀k ∈ N : uεk (yεk ) − vεk (yεk ) > εa k. (3.9) By Lemma 21, this leaves two possibilities. In the first one, there exists a point z = [zε] ∈ ext(K) which joins points of the sequence (yεk )k∈N at (εk)k∈N. Therefore g(z) = [vε(zε)] = ¯f (z) = [uε(zε)] = 0, which gives a contradiction at ε = εk when compared with (3.9). In the second one, there exists a point ¯z = [¯zε] ∈ K joining points of the sequence (yεk )k∈N at (εk)k∈N. Once again, we have g(¯z) = [vε(¯zε)] = ¯f (¯z) = [uε(¯zε)] = f (z), in contradiction to (3.9). Furthermore, [18, Thm. 31] implies claim (iii). Finally, (i) follows from the (cid:3) We also have this simple but useful result: Thm. 24 opens the possibility to restrict our attention to compactly supported isomorphism GC∞(eRc,eRd) ≃ Gs(R)d. Lemma 25. Let ∅ 6= K ⋐f U ⊆ eRn, U be a sharply open set, f ∈ GDK(U,eRd) and α ∈ Nn. Then ∂αf ∈ GDK(U,eRd). GSF whose domain is the entire eRn, as stated in the following Theorem 26. For ∅ 6= K ⋐f eRn and Y ⊆ eRd set GDg(K, Y ) :=nf ∈ GC∞(eRn, Y ) fext(K) = 0o , where the g superscript means globally defined. Let K ⊆ U ⊆ eRn, then (3.10) (i) (ii) If f ∈ GDg(K, Y ), then fU ∈ GDK(U, Y ). If f ∈ GDK(U,eRd), then ∃! ¯f ∈ GDg(K,eRd) : ¯fK = fK. (iii) GDg(K, Y ) = GDK(eRn, Y ). Proof. (i): Let (uε) be any net that defines f , so that uε ∈ C∞(Rn, Rd). Clearly fU ∈ GC∞(U, Y ), so for all α ∈ Nn and x = [xε] ∈ ext(K) it remains to show that [∂αuε(xε)] = 0. By Lem. 15, we get ext(K) = hK c εi, which is a sharply open set. So x ∈ ext(K) yields Br(x) ⊆ ext(K) for some r ∈ eR>0, and hence fBr(x) = 0. Thereby ∂αf (x) = [∂αuε(xε)] = 0. A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 16 (ii): This is exactly Thm. 24. (iii): This follows directly from (i) by setting U = eRn, and by Def. 19 which yields fext(K) = 0. For the extension of property (ii) to arbitrary codomains Y ⊆ eRd (provided that U is strongly internal) see Thm. 62 below. (cid:3) Note explicitly that in (3.10), instead of the more technical properties of Def. 19, we have a concise and simpler pointwise condition. Notwithstanding this and several other positive aspects of definition (3.10) (see the second part of Lem. 17 and the following Thm. 27), in the present work, we prefer not to change Def. 19 in favor of (3.10): On the one hand, Def. 19 is nearer to the classical definition of DK(Ω), where the domain is Ω ⊆ Rn; on the other hand, for applications of these notions to geometry, locally defined functions are a more natural setting. We can also summarize these results by saying that compactly supported GSF have a good notion of being compactly supported, and globally defined compactly supported GSF are in addition well-behaved with respect to the concept of support introduced above. This is also clearly confirmed by the following Theorem 27. Let Y ⊆ eRd, and f ∈ GC∞(eRn, Y ) such that supp(f ) 6= ∅. Then f ∈ GDK(eRn, Y ) if and only if supp(f ) ⊆ K. Proof. One implication is immediate from Remark 20.(i). Conversely, if supp(f ) ⊆ K, by Thm. 26.(iii) we have to show that fext(K) = 0. Let x ∈ ext(K) but assume that f (x) 6= 0. Pick any y ∈ ny′ ∈ eRn f (y′) > 0o, which is non-empty because supp(f ) 6= ∅. Let (uε) be a net that defines f and let y = [yε]. Since f (x) 6= 0, there exists S ⊆ I such that eS 6= 0 and f (x)eS > 0. Then setting z := xeS + yeSc, we have f (z) > 0 and hence z ∈ supp(f ) ⊆ K. But then xeS = zeS ∈ KeS which, by Lem. 14, yields x /∈ ext(K), a contradiction. 3.2. Examples of compactly supported GSF. A first class of examples comes by considering each ϕ ∈ DK(Ω), K ⋐ Ω ⊆ Rn. Indeed, it suffices to set Kε := K and uε(x) := ϕ(x) if x ∈ Ω and uε(x) := 0 otherwise to have that ϕ ∈ GD eK(eΩc,eR), where we recall that eK = [K]. Therefore DK (Ω) ⊆ GD eK(eΩc,eR). Moreover, since any given CGF can be defined by a net (uε) of maps with sharply (cid:3) bounded compact supports, we have the following result: Theorem 28. Let Ω be an open subset of Rn and J = [Jε] ∈ eR be a CGN such that limε→0+ Jε = +∞. Set Kε := {x ∈ Ω x ≤ Jε} and K := [Kε]. Then for all f ∈ GC∞(eΩc,eRd) there exists ¯f ∈ GDK(eRn,eRd) such that ¯f eΩc Proof. Set Uε := {x ∈ Ω x < 1 2 Jε} so that Uε ⊆ Kε for ε small. Let χε ∈ C∞(Ω, R) be such that χUε = 1 and supp(χε) ⊆ Kε. Let f ∈ GC∞(eΩc,eRd) be = f . A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 17 represented by (vε), with vε ∈ C∞(Rn, Rd), and set uε := χε · vε. Then each uε is compactly supported and any x = [xε] ∈ eΩc satisfies xε ∈ Uε for ε small because limε→0+ Jε = +∞. Therefore ¯f := [uε(−)] ∈ GDK(eRn,eRd), and if xε ∈ Uε then uε(xε) = vε(xε), so ¯f eΩc = f . (cid:3) This theorem gives an infinity of non-trivial examples of compactly supported GSF. Moreover, even though ¯f depends on the fixed infinite number J ∈ eR, every such ¯f contains all the information of the original CGF f because ¯f eΩc Finally, the constant function f (x) = 1 for all x ∈ eR is not compactly supported. holds. Then choosing r large enough that dε−r ∈ eR \ K we arrive at f ( dε−r) = In fact, by contradiction, assume that f admits (uε) and (Kε) such that Def. 19 [uε(ε−r)] = 0. = f . 4. Generalized norms on GDK and GD (i) As a first step to topologizing the spaces GDK and GD we prove: Theorem 29. Let ∅ 6= K ⋐f U ⊆ eRn, then (ii) For all non-empty H ⋐f U , the inclusion K ⊆ H implies GDK(U,eRd) ⊆ GDK(U,eRd) is an eR-module GDH (U,eRd). (ii): Take f ∈ GDK(U,eRd). Since K ⊆ H, by [27, Prop. 2.8] for each repre- sentative (Kε) of K we get the existence of a representative (Hε) of H such that Kε ⊆ Hε for all ε. Therefore, ext(K) = hK c εi = ext(H) and hence the conclusion follows. Proof. (i) is immediate from Def. 19. εi ⊇ hH c (cid:3) From the extreme value property, Prop. 2, it is natural to expect that the fol- lowing CGN could serve as generalized eR-valued norms. Definition 30. Let ∅ 6= K ⋐f U ⊆ eRn, where U is a sharply open set. Let m ∈ N and f ∈ GDK(U,eRd). Then kfkm,K := max α≤m 1≤i≤d max(cid:0)(cid:12)(cid:12)∂αf i(Mαi)(cid:12)(cid:12) ,(cid:12)(cid:12)∂αf i(mαi)(cid:12)(cid:12)(cid:1) ∈ eR, where mαi, Mαi ∈ K satisfy ∀x ∈ K : ∂αf i(mαi) ≤ ∂αf i(x) ≤ ∂αf i(Mαi). The following result permits to calculate the (generalized) norm kfkm,K using any net (vε) that defines f . In case the net (vε) satisfies Def. 19, it also permits to prove that this norm does not depend on K, as is the case for any ordinary compactly supported smooth function. A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 18 Then we have: Even though kfkm,K ∈ eR, using an innocuous abuse of language, in the following we will simply call kfkm,K a norm. Proposition 31. Under the assumptions of Def. 30, let the set K = [Kε] ⋐f eRn. ε(x)(cid:12)(cid:12)(cid:21) If the net (vε) defines f , then kfkm,K =(cid:20)maxα≤m   . x∈Rn(cid:12)(cid:12)∂αui ε(x)(cid:12)(cid:12) If (uε) defines f and (uε) satisfies Def. 19, then supx∈Kε(cid:12)(cid:12)∂αvi kfkm,K =  max α≤m 1≤i≤d 1≤i≤d (4.1) sup (ii) (i) Proof. In proving (i) we will also prove that the norm kfkm,K is well-defined, i.e. it does not depend on the particular choice of points mαi, Mαi as in Def. 30. As in the proof of Prop. 2, we get the existence of ¯mαiε, ¯Mαiε ∈ Kε such that Hence (cid:12)(cid:12)∂αvi ε( ¯mαiε) ≤ ∂αvi ε( ¯mαiε)(cid:12)(cid:12) ,(cid:12)(cid:12)∂αvi max(cid:0)(cid:12)(cid:12)∂αvi ε(x) ≤ ∂αvi ε( ¯Mαiε)(cid:12)(cid:12)(cid:1). Thus ε( ¯mαiε)(cid:12)(cid:12) ,(cid:12)(cid:12)∂αvi ε( ¯Mαiε). sup max α≤m 1≤i≤d ∀x ∈ Kε : ∂αvi ε(x)(cid:12)(cid:12) ≤ max(cid:0)(cid:12)(cid:12)∂αvi x∈Kε(cid:12)(cid:12)∂αvi But ¯mαiε, ¯Mαiε ∈ Kε, so x∈Kε(cid:12)(cid:12)∂αvi ε(x)(cid:12)(cid:12) ε(x)(cid:12)(cid:12) ≤ max   =  max   max α≤m 1≤i≤d α≤m 1≤i≤d sup ε( ¯Mαiε)(cid:12)(cid:12)(cid:1) . ε( ¯Mαiε)(cid:12)(cid:12)(cid:1) max(cid:0)(cid:12)(cid:12)∂αvi  = max(cid:0)(cid:12)(cid:12)∂αf i( ¯Mαi)(cid:12)(cid:12) ,(cid:12)(cid:12)∂αf i( ¯mαi)(cid:12)(cid:12)(cid:1) . ε( ¯mαiε)(cid:12)(cid:12) ,(cid:12)(cid:12)∂αvi α≤m 1≤i≤d = max α≤m 1≤i≤d (ii): By Lemma 22, we have that From this, both the fact that the norm kfkm,K is well-defined and claim (i) follow.   max x∈Kε(cid:12)(cid:12)∂αui ε(x)(cid:12)(cid:12)   = max α≤m 1≤i≤d α≤m 1≤i≤d sup (cid:20) sup x∈Kε(cid:12)(cid:12)∂αui (cid:20) sup x∈Rn(cid:12)(cid:12)∂αui ε(x)(cid:12)(cid:12)(cid:21) = ε(x)(cid:12)(cid:12)(cid:21) =  max α≤m 1≤i≤d = max α≤m 1≤i≤d sup ε(x)(cid:12)(cid:12) x∈Rn(cid:12)(cid:12)∂αui   . Corollary 32. Let ∅ 6= K ⋐f U ⊆ eRn, where U is a sharply open set. Let ∅ 6= H ⋐f U and m ∈ N. If f ∈ GDK(U,eRd)∩GDH (U,eRd), then kfkm,K = kfkm,H =: kfkm. Proof. The right hand side of (4.1) does not depend on K. (cid:3) Another consequence of Prop. 31 is the following: (cid:3) A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 19 Corollary 33. Let U ⊆ eRn, f ∈ GD(U,eRd) and ¯f ∈ GD(eRn,eRd) be the extension of f defined in Thm. 24. Then for all m ∈ N, kfkm = k ¯fkm. Our use of the term "norm" is justified by the following (i) (ii) Proposition 34. Let ∅ 6= K ⋐f U ⊆ eRn, where U is a sharply open set. Let f , g ∈ GDK(U,eRd) and m ∈ N. Then kfkm ≥ 0 kfkm = 0 if and only if f = 0 (iii) ∀c ∈ eR : kc · fkm = c · kfkm kf + gkm ≤ kfkm + kgkm. kf · gkm ≤ cm · kfkm · kgkm for some cm ∈ R>0. (iv) (v) Proof. (i), (iii) and (iv) follow directly from Prop. 31, as does (v), using the Leibniz rule. The 'if'-part of property (ii) follows from (4.1). (cid:3) U : ε ), d(kε, U c (4.2) (cid:3) ∀K, H ⋐f U : interl(H ∪ K) ⊆ U. We now prove that also the space GD(U,eRd) is an eR-module, at least for certain Proposition 35. Let U ⊆ eRn be a non empty sharply open set. Assume that Then GD(U,eRd) is an eR-module. Proof. Since in Thm. 29 we already proved that GDK(U,eRd) is closed with respect to products by scalars, we only need to prove that GD(U,eRd) is closed with respect to sum. Let f ∈ GDK(U,eRd), g ∈ GDH (U,eRd) and let (uε), (vε) satisfy Def. 19 for f and g, respectively. Lemma 8 and [27, Prop. 2.8] imply that interl(H ∪ K) = [Hε ∪ Kε] ⋐f eRn. But ext(H ∪ K) = hH c ε ∩ K c εi = ext(H) ∩ ext(K). Therefore, ∂α(uε + vε) is zero on ext(H ∪ K). By our assumption (4.2) ∅ 6= interl(H ∪ K) ⋐f U , so that f + g ∈ GD(U,eRd). Proposition 36. Let U ⊆ eRn be a non-empty sharply open set. If U is eR-convex Proof. Assume that U is eR-convex, i.e. xh + (1 − x)k ∈ U for all h, k ∈ U and all x ∈ [0, 1]. Then for all H, K ⊆ U (even if we do not assume them to be functionally compact), and all y ∈ interl(H ∪ K), we can write y = eS · h + eSc · k for some S ⊆ I and h ∈ H, k ∈ K. Thus y = eS · h + (1− eS)· k ∈ U since eS ∈ [0, 1] and h, k ∈ U . Now, assume that U is a strongly internal set, i.e. for some net (Uε) of subsets of Rn, we have U = hUεi. We continue to use the notations for y as above. Since ε ) > εq for some q ∈ R>0 and h, k ∈ U , [18, Thm. 8] entails that d(hε, U c In the following result, we give two general sufficient conditions for (4.2) to hold. or U is a strongly internal set, then (4.2) holds. εi = hH c εi ∩ hK c A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 20 1 n , ε small, where h = [hε], k = [kε]. But y = eS · h + eSc · k, so for all ε small, if ε ∈ S then yε = hε and if ε /∈ S then yε = kε. In any case, d(yε, U c ε ) > εq, hence y ∈ hUεi = U . (cid:3) Example 37. If U = (−1, 1) ∪ (2, 4) ⊆ eR, then U is a sharply open set, but it 2(cid:3)R(cid:3) does not satisfy condition (4.2) of Prop. (35): let H :=(cid:2)(cid:2)− 1 and xε := 0 if ε ∈ h 1 n+1(cid:17) if n ∈ N>0 is even, and xε := 3 otherwise. Then (R) ⊆ GDH (eRc), x := [xε] ∈ interl(H ∪ K) but x /∈ U . Moreover, let ϕ ∈ D[− 1 2 , 1 (R) ⊆ GDK(eRc) be positive non-trivial functions. Then, as we showed ψ ∈ D[ 5 2 , 7 in the proof of Prop. 35, the GSF ϕ + ψ ∈ GC∞(U,eR) is compactly supported in 2(cid:3)R ∪(cid:2) 5 2(cid:3)R(cid:3) = interl(H ∪ K) 6⊆ U . Finally, let f := ϕU and g := ψU , so (cid:2)(cid:2)− 1 that f ∈ GDH (U,eR) and g ∈ GDK(U,eR). Rem. 20.(i) yields that f +g /∈ GDJ (U,eR) 2(cid:1)R and M of for all J ⋐f U : Otherwise, taking suitable sub-intervals L of (cid:0)− 1 (cid:0) 5 2(cid:1)R where f resp. g do not vanish, we would have [L∪M ] ⊆ supp(f +g) ⊆ J ⊆ U 2 , 7 (here supp(f + g) is the support as in Def. 11). But the inclusion [L ∪ M ] ⊆ U is impossible -- a counterexample can be constructed similar to the above x. 2(cid:3)R(cid:3), K :=(cid:2)(cid:2) 5 2 , 7 2 , 1 2 , 1 2 , 1 2 , 7 2 ]R 2 ]R This example shows that an assumption like (4.2) is necessary to have the closure of the space GD(U,eR) with respect to sum. 5. Topological structure on GDK Using our eR-valued norms, it is now natural to define Definition 38. Let ∅ 6= K ⋐f U ⊆ eRn, where U is a sharply open set. Let f ∈ GDK(U,eRd), m ∈ N, ρ ∈ eR>0, then ρ (f ) := ng ∈ GDK(U,eRd) kf − gkm < ρo. In case any confusion could If V ⊆ GDK(U,eRd), then we say that V is a sharply open set if Bm arise, we will use the more precise symbol Bm ρ (f, K) := Bm ρ (f ). (ii) (i) Moreover, we say that V is Fermat open if ∀v ∈ V ∃m ∈ N∃ρ ∈ eR>0 : Bm ∀v ∈ V ∃m ∈ N∃r ∈ R>0 : Bm ρ (v) ⊆ V. r (v) ⊆ V. As in [18, Thm. 2] it follows that sharply open sets as well as Fermat open sets form topologies on GDK(U,eRd). On the other hand, it is also natural to view the space GDK(U,eRd) inside Garetto's theory [11, 12] of eR-locally convex algebras. In this section, we will realize this A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 21 comparison, proving that the space GDK(U,eR) is a Fr´echet eR-module and a topo- logical algebra. For this purpose, we will only consider the sharp topology. In- deed, as we will see below, the Fermat topology is less interesting in this con- text since it doesn't permit to prove the continuity of the product by scalars developed in [11, 12], is that the domain U contains generalized points. empty sharply open set. The main problem in performing this comparison, which (r, f ) ∈ eR × GDK(U,eRd) 7→ r · f ∈ GDK(U,eRd). In the following, we will always assume that ∅ 6= K ⋐f U ⊆ eRn, where U is a non- doesn't permit to view our space GDK(U,eR) as a particular case of the theory Using the valuation v on eR, it is natural to introduce the following notions: Definition 39. Let m ∈ N and f ∈ GD(U,eR), then: vm(f ) := v(kfkm) ∈ R (i) (ii) Pm(f ) := e−vm(f ). From the properties of the valuation v and of the e-norm − e = e−v(−) on eR Proposition 40. For each m ∈ N, we have: (i) (see [4]), the following result directly follows. vm : GD(U ) −→ R ∪ {+∞} is a valuation, i.e. for all f , g ∈ GD(U ): • vm(0) = +∞ • vm(λ · f ) ≥ v(λ) + vm(f ) ∀λ ∈ eR • vm( dεa · f ) = v( dεa) + vm(f ) = a + vm(f ) ∀a ∈ R • vm(f + g) ≥ min [vm(f ), vm(g)]. (ii) Pm : GD(U ) −→ R is an ultra-pseudo-norm, i.e. for all f , g ∈ GD(U ): • Pm(f ) = 0 if and only if f = 0 • Pm(λ · f ) ≤ λe · Pm(f ) ∀λ ∈ eR • Pm( dεa · f ) = dεae · Pm(f ) = e−a · Pm(f ) ∀a ∈ R • Pm(f + g) ≤ max [Pm(f ),Pm(g)]. The following result states that to define the sharp topology, instead of the above ultra-pseudo-norms we can equivalently use the countable family of generalized norms (kfkm)m∈N. Theorem 41. (i) Sum and product in GDK(U ) are continuous in the sharp topology. Therefore, GDK(U ) is a topological eR-algebra. subspace {(f, g) ∈ GDK(U ) × GDK(U ) ∀ m ∈ N : kfkm, kgkm < ∞}. (ii) The product in GDK(U ) is continuous in the Fermat topology only on the (iii) For f ∈ GDK(U ), set Cm r (f ) := {g ∈ GDK(U ) Pm(f − g) < r} (r ∈ R>0, m ∈ N). A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 22 Then for each q, s ∈ R>0 we have: (i) (ii) If q ≤ − log r, then Cm r (f ) ⊆ Bm If q ≥ − log s and s < r, then Bm dεq (f ) dεq (f ) ⊆ Cm r (f ). (iv) The sharp topology on GDK(U ) is the coarsest topology such that each Pm is continuous. (v) GDK(U ) is a separated locally convex topological eR-module. Proof. (i), (ii): Continuity of the sum in the sharp (and in the Fermat) topology follows directly from the triangle inequality Prop. 34.(iv). The continuity of the product at (f0, g0) and property (ii) follow from Prop. 34 (v) via kf · g − f0 · g0km ≤ cm(kf − f0km · kg − g0km + kf − f0km · kg0km+ + kf0km · kg − g0km). (iii): Let us first assume q ≤ − log r and g ∈ Cm r (f ), so that Pm(f − g) < r and vm(f − g) > − log r. This implies ∃b > − log r : max α≤m i≤d sup x∈Kε(cid:12)(cid:12)∂αui ε(x) − ∂αvi ε(x)(cid:12)(cid:12) = O(εb), (5.1) where (uε) and (vε) define f and g, respectively. Property (5.1) yields the existence of some M > 0 such that for ε sufficiently small we obtain sup x∈Kε(cid:12)(cid:12)∂αui max α≤m i≤d ε(x) − ∂αvi ε(x)(cid:12)(cid:12) + εb ≤ (M + 1) · εb < ε− log r ≤ εq. Therefore kf − gkm + dεb ≤ dεq, so kf − gkm < dεq. Now, let us assume q ≥ − log s, s < r, and g ∈ Bm Therefore, ∀0ε : max α≤m i≤d sup x∈Kε(cid:12)(cid:12)∂αui ε(x) − ∂αvi Taking the − e-norm we get (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) sup x∈Kε(cid:12)(cid:12)∂αui ε(x)(cid:12)(cid:12) ε(x) − ∂αvi max α≤m i≤d that is Pm(f − g) < r as claimed. (iv) follows directly from (iii) and Prop. 40.(ii). (v) follows from Prop. 34.(ii) and [11, Prop. 1.11]. dεq (f ), so that kf − gkm < dεq. ε(x)(cid:12)(cid:12) < εq ≤ ε− log s. (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)e ≤ elog s = s < r, (cid:3) 5.1. Generalized functions and non-Archimedean properties. In this sec- tion, we want to clarify some relationships between the classical notion of convexity, We have seen that balls Bm the notion of eR-convexity of [11] and the use of eR-valued norms. GDK(U ); they are convex in the usual sense, i.e. if f , g ∈ Bm ρ (0), ρ ∈ eR>0, define a neighborhood system of 0 for ρ (0), t ∈ [0, 1] (in A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 23 particular if t ∈ [0, 1]R), then ktf + (1 − t)gkm ≤ tkfkm + (1 − t)kgkm < tρ + (1 − t)ρ = ρ. ρ (0) is also balanced: if t ∈ eR, t ≤ 1, then t·Bm Moreover, each ball Bm ρ (0). However, this space is not a classical locally convex topological vector space over the field R because of two reasons: (i) the product by scalars is not continuous with respect to the Euclidean topology on R, (ii) Lemma 23 implies that the property ρ (0) ⊆ Bm ∀f ∈ GDK(U )∃t ∈ eR : f ∈ t · Bm ρ (0) proof of Thm. 41 (ii), this is a necessary consequence of the existence of generalized holds for t ∈ eR but it cannot be extended to t ∈ R. As we have seen in the functions with infinite eR-valued k − km-norm. On the other hand, even though the sets Cm r (0) are defined using R only, i.e. without mentioning any non-Archimedean property, they satisfy ∀f ∈ Cm r (0)∀λ ∈ R : λ · f ∈ Cm r (0), (5.2) and this is possible only because they are infinitesimal sets. In fact, we have seen in Thm. 41 that Cm r (0) ⊆ Bm dεq (0) for q ≤ − log r. More generally, a set A ⊆ GDK(U ) can be eR-balanced (see [11]), i.e. λA ⊆ A ∀λ ∈ eR : λe ≤ 1, and at the same time can be thought of as "small" only in case A consists infini- tesimal points. For example, the ball Bm 1 (0) is not. In fact, we have dεb (0) is eR-balanced, but Bm Lemma 42. Suppose that A ⊆ GDK(U ) and m ∈ N are such that A + A ⊆ A ∃r ∈ R>0 : A ⊆ Bm r (0). (5.3) r (0) for all n ∈ N 6=0. Therefore, kukm < r Then every element u ∈ A has infinitesimal norm: kukm ≈ 0. Proof. In fact, (5.3) implies n·u ∈ A ⊆ Bm for all n ∈ N 6=0, which proves our claim. Let us note that condition (5.3) holds both for A which is eR-balanced or eR- naturally induced to consider a topology on eR which contains infinitesimal neigh- convex. These remarks permit to show that in dealing with generalized functions, we are borhoods (hence inducing the discrete topology on R, see [17]). This is due to the coexistence of a continuous product by scalars and of an infinite element in GDK(U ), as stated in the following general result. In a possible interpretation of n (cid:3) A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 24 topology τ , and < eR as the strict order relation < of Lemma 1. its statement, we can think of R as R with a topology τ , eR as eR with the sharp Theorem 43. Let (R, +R,·R, <R, τ ) and (eR, + eR,· eR, < eR, τ ) be Hausdorff topological ordered rings such that (R, +,·, <) is a substructure of (R, +R,·R, <R), which in turn is a substructure of (eR, + eR,· eR, < eR) and such that Let (G, +G,·G, σ) be a Hausdorff topological R-module, and −G : G −→ eR, −R : R −→ R be maps such that r ·G gG = rR · eR gG for all r ∈ R and all g ∈ G. As- sume that any τ -neighborhood of 0 ∈ R contains a ball BR η (0) := {s ∈ R sR <R η} for some η ∈ R, η >R 0, and that there exists some ρ ∈ eR with ρ > eR 0 such that ∀r ∈ R ∀s ∈ R : r <eR s =⇒ r <R s. ρ (0) := {g ∈ G gG < eR ρ} is σ-open. Finally, assume that gG > eR ρ · eR M. gG is invertible , ∀M ∈ R>0 : ∃g ∈ G : the ball BG (5.4) (5.5) Then the induced topology τ ∩ R is discrete. Proof. Since G is a topological R-module, the product by scalars is τ×σ-continuous, and r ·G g = 0, (5.6) lim r→0 r∈R where g ∈ G comes from assumption (5.5). By hypothesis, the ball BG and every τ -neighborhood of r = 0 ∈ R contains some ball BR entails that there exists some η >R 0 such that ρ (0) ∈ σ η (0). Therefore (5.6) ∀r ∈ R : rR <R η =⇒ ρ > eR r ·G gG = rR · eR gG. (5.7) For each s ∈ R>0 take M ∈ R>0 such that 1 from (5.5). For all r ∈ R such that rR <R η, we have < s < eR rR < eR ρ gG 1 M M < s, so that gG > eR ρ · eR M > eR 0 because gG is invertible in eR. Therefore rR < eR s and hence rR <R s by (5.4). This means that r is infinitesimal in the ring R, i.e. the ball BR ρ (0) is contained in the monad of 0 (see e.g. [17] for the notion of monad) and so also every ball BR η (¯r) is contained in the monad of ¯r ∈ R. Therefore, [17, Prop. 2.1] implies the conclusion. (cid:3) We can therefore say that if we want to find a space G of generalized functions which is an ordinary Hausdorff topological vector space on R, then we cannot define the topologies τ and σ using seminorms valued in a non-Archimedean (see (5.5)) extension of R. This results confirms Rem. 43 of [10]. As a consequence, we have the following impossibility result: A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 25 (G, +G,·G) is a linear subspace of GDK(U ) for some ∅ 6= K ⋐f U ⊆ eRn Corollary 44. There does not exist any real Hausdorff topological vector space (G, +G,·G, σ) such that: (i) (ii) G contains the Dirac delta δ ∈ G (iii) ∃m ∈ N∃ρ ∈ eR>0 : ρ < 1 , Bm In particular, the Colombeau algebra Gs(Ω) does not contain any real Hausdorff topological vector subspace G such that some Bm ρ (0) ∩ G ∈ σ ρ (0) ∩ G is open and δ ∈ G. Proof. By contradiction, in Thm. 43, set R := R with the usual Euclidean topology τ , eR := eR with the sharp topology; set gG := kgkm, where m ∈ N comes from (iii) and we used the inclusion (i); set rR := r the usual absolute value in R. Note also that Bm ρ (0) ∩ G = {g ∈ G kgkm < ρ} = {g ∈ G gG < eR ρ}. If δ ∈ G, then δG = kδkm is infinite and invertible in eR, so that Thm. 43 implies that the Euclidean topology would be discrete. The second part of the claim follows from Thm. 28. (cid:3) We can summarize Cor. 44 by saying that a real Hausdorff topological vector space structure G for a space of generalized functions cannot contain even a single trace Bm ρ (0) ∩ G, ρ < 1, of a sharply open ball. This result does not contradict [1, Prop. 4], where it is stated that the sharp topology induces on bounded sets of the real locally convex space Ga(Ω) ⊆ Gs(Ω) a topology which is finer than the topology σa on Ga(Ω). On the other hand, Cor. 44 implies that Bm ρ (0)∩Ga(Ω) /∈ σa for all ρ ∈ eR>0, ρ < 1. 6. Metric structure on GDK In this section, we want to use [11, Thm. 1.14] to prove metrizability of GDK(U ). However, we will apply this result using an explicit and simple countable base of neighborhoods of the origin which consists of eR-absorbent and absolutely eR-convex The idea is to consider only points of balls f ∈ Bm ρ (0) whose norm kfkm is infinitely smaller than ρ. To formally express this idea, we introduce the following sets. In this way, we will arrive at a simpler metric. Definition 45. Let ρ ∈ eR>0, m ∈ N and g ∈ GDK(U ). Then ρ · kf − gkm ≈ 0(cid:27) . ρ (g) :=(cid:26)f ∈ GDK(U ) U m 1 In case any confusion might arise, we will use the more precise symbol U m U m ρ (g). ρ (g, K) := Proposition 46. (i) U m ρ (0) is eR-absorbent and absolutely eR-convex. A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 26 (ii) Both the system nU m ρ (v) v ∈ GDK(U ), m ∈ N, ρ ∈ eR>0, ρ ≈ 0o , and the system {U n dεn (v) v ∈ GDK(U ), n ∈ N>0} generate the sharp topology on GDK(U ). i.e. u dεb ∈ U m Proof. To see that U m ρ (0) is eR-absorbent, by [11, Def. 1.1] we have to show that dεb·ρ ≈ 0. But ρ is strictly positive, so ρ ≥ dεp for some p ∈ R. Moreover, kukm ∈ eR, so kukm ≤ dεq for some q ∈ R. Therefore ∀u ∈ GDK(U )∃a ∈ R∀b ∈ R ρ (0), which is equivalent to (cid:13)(cid:13) u ≤a : u ∈ dεb · U m dεb(cid:13)(cid:13)m · 1 kukm dεb · ρ ≤ dεq−b−p, ρ = kukm ρ (0), and we have dεq−b−p ≈ 0 if and only if b < q − p. This proves (6.1). (6.1) ρ (0) is balanced, assume λ ∈ eR with λe ≤ 1. Then v(λ) ≥ 0, ρ (0) then To prove that U m so λ ≤ c for some c ∈ R>0. Therefore, if u ∈ U m = λkukm kλukm ρ ρ ≈ 0, 1 ρ (0). ρ (0), we have In order to prove (ii), we note that U m so λu ∈ U m Finally, we show eR-convexity: for all a, b ∈ R ρ · k dεa · u + dεb · vkm ≤ dεa · kukm ρ (v) ⊆ Bm ≈ 0 implies ku−vkm ρ (v) ⊆ U m√ρ(v) because ku − vkm < ρ implies 1√ρ·ku−vkm ≤ √ρ ≈ 0. Finally, if ρ ≈ 0, every U m ρ (v) and w ∈ Bm ρ ·ku− vkm ≈ 0. The proof for the second system in (ii) follows by observing that given ρ > 0, ρ ≈ 0, there exists q ∈ N such that ρ ≥ dεq, and setting n := max(m, q) we have U n ρ (v). (cid:3) ≥0 and all u, v ∈ U m + dεb · kvkm ρ ≈ 0. ρ (v) because ku−vkm ρ (v) is sharply open: if u ∈ U m < 1. Also, if ρ ≈ 0, then Bm ρ ·kw− vkm ≤ 1 ρ ·kw− ukm + 1 dεn (v) ⊆ U m ρ2 (u), then 1 ρ ρ ρ From [11, Thm. 1.14], we have that GDK(U ) is metrizable with metric d2(u, v) = +∞Xn=1 2−n · minnPU n dεn (0)(u − v), 1o . (6.2) Concerning (6.2) we recall (see [11]) that if A ⊆ GDK(U ) is eR-absorbent, then, for all u ∈ GDK(U ), we define VA(u) := sup(cid:8)b ∈ R u ∈ dεb · A(cid:9) A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 27 The following result gives a metric which is equivalent to (6.2) but is defined by a simpler formula. PA(u) := e−VA(u). (6.3) Proposition 47. Set An := U n dεn (0) for n ∈ N>0, and let u ∈ GDK(U ). Then (i) VAn (u) = vn(u) − n (ii) The map de(u, v) = +∞Xn=1 emin[n−vn(u−v),0]−n. is a metric on GDK(U ) that is equivalent to d2. Proof. Concerning (i), we note that u ∈ dεb · An ⇒ ⇒ b ≤ vn(u) − n, u dεb ∈ U n dεn (0) ⇒ kukn dεb+n ≈ 0 ⇒ v(cid:0) dε−b−n · kukn(cid:1) ≥ 0 so VAn (u) ≤ vn(u) − n. Conversely, if we had VAn (u) < vn(u) − n, we could pick VAn (u) < b < vn(u) − n. Then as above it would follow that kukn dεb+n ≈ 0, contradicting the definition of VAn (u). This proves (i). In order to prove (ii), we use (i) in (6.2): PAn(u − v) = e−vn(u−v)+n and d2(u, v) = = ≥ 2−n · minne−vn(u−v)+n, 1o = 2−n · emin[n−vn(u−v),0] ≥ emin[n−vn(u−v),0]−n = de(u, v). +∞Xn=1 +∞Xn=1 +∞Xn=1 But we also have e−n ≥ 2−n−1 for all n, so that de(u, v) ≥ 1 2 d2(u, v). Using Prop. 40, it is easily checked that de is a metric, which we have just proved to be equivalent to d2. (cid:3) 7. Completeness of GDK In order to prove the completeness of GDK(U ), we generalize the proof of [11, Prop. 3.4] (based in turn on [28]) to the present context. Theorem 48. The space GDK(U ) with the sharp topology is complete. Proof. By Prop. 47 GDK(U ) with the sharp topology is metrizable. Hence, it suffices to consider a Cauchy sequence (un)n∈N in this topology, i.e., ∀q ∈ R>0 ∀i ∈ N∃N ∈ N∀m, n ≥ N : kun − umki < dεq. A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 28 Setting i = q = k ∈ N>0, this implies the existence of a strictly increasing sequence (nk)k∈N in N such that kunk+1 − unkkk < dεk. Hence picking any representatives (unε) of un as in Def. 19 we have (cid:20)max α≤k sup x∈Rn(cid:12)(cid:12)∂αunk+1,ε(x) − ∂αunk,ε(x)(cid:12)(cid:12)(cid:21) <(cid:2)εk(cid:3) ∀k ∈ N>0. By Lemma 1 this yields that for each k ∈ N>0 there exists an εk such that εk ց 0 and ∀ε ∈ (0, εk) : max α≤k sup x∈Rn(cid:12)(cid:12)∂αunk+1,ε(x) − ∂αunk,ε(x)(cid:12)(cid:12) < εk. Now set (7.1) (7.2) hkε :=  unk+1,ε − unk,ε ∈ C∞(Rn, R) 0 ∈ C∞(Rn, R) uε := un0ε + ∞Xk=0 hkε ∀ε ∈ I. if ε ∈ (0, εk) if ε ∈ [εk, 1) Since εk ց 0, for all ε ∈ I we have ε /∈ (0, εk) for all k ≥ ¯k, with ¯k sufficiently big. Therefore, uε = un¯k+1,ε ∈ C∞(Rn, R). In order to prove that (uε) defines a GSF of the type U → eR, take [xε] ∈ U and α ∈ N. We claim that (∂αuε(xε)) ∈ RM . Now if p ∈ N satisfies α ≤ p, then for any x ∈ Rn we have ∞Xk=p+1 ∂αuε(x) ≤(cid:12)(cid:12)∂αunp+1,ε(x)(cid:12)(cid:12) + ∂αhkε(x) . From (7.1) and (7.2) we get that ∂αhkε(x) ≤ εk for all k ≥ p + 1, x ∈ Rn and all ε ∈ (0, 1]. Hence for ε ∈ (0, 1], α ≤ p and all x ∈ Rn we obtain ∂αuε(x) ≤(cid:12)(cid:12)∂αunp+1,ε(x)(cid:12)(cid:12) + εp+1 1 − ε . (7.3) Inserting x = xε and noting that (∂αunp+1,ε(xε)) ∈ RM proves our claim. Moreover, since all (unε) satisfy Def. 19, we also conclude from (7.3) that for any α and any [xε] ∈ ext(K) we have [∂αuε(xε)] = 0, and hence the GSF [uε(−)]U ∈ GDK(U ). Finally, ku− unpki < dεp−1 for all p ∈ N>1 and all i ≤ p. This yields that (unk )k tends to u in the sharp topology, and hence so does (un). (cid:3) 8. The space GD as inductive limit of GDK In this section, we always assume that U ⊆ eRn is a non-empty strongly internal set. By Prop. 35 and Prop. 36, this entails that GD(U ) is an eR-module. In order to define a natural topology on GD(U ) we will employ [11, Thm. 1.18], which we restate here for the reader's convenience: Theorem 49. Let G be an eR-module, (Gγ)γ∈Γ be a family of locally convex topolog- ical eR-modules and, for each γ ∈ Γ, let iγ : Gγ −→ G be an eR-linear map. Assume A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 29 that G = span [γ∈Γ iγ (Gγ)  γ (V ) is a −→ Gγ. (see [11] and Def. (6.3)) is the finest and let V ∈ V if and only if V ⊆ G, V is absolutely eR-convex and i−1 neighborhood of 0 in Gγ for all γ ∈ Γ. Then each V ∈ V is eR-absorbent and the topology τ induced by the gauges {PV }V ∈V eR-locally convex topology on G such that iγ is continuous for all γ ∈ Γ. Endowed with this topology, G is called the inductive limit (colimit) of the spaces (Gγ)γ∈Γ and we write G = lim Since GD(U ) =S∅6=K⋐fU GDK(U ), we may therefore equip it with the inductive limit topology with respect to the inclusions ιK : GDK(U ) ֒→ GD(U ). We call the resulting eR-locally convex topology the sharp topology on GD(U ). Hence Henceforth we will denote the sharp topology on GDK(U ) by σK(U ) (or, for short, by σK). Also, the inductive limit topology on GD(U ) will be denoted by σ(U ) (or by σ). Setting −→ GDK(U ) (∅ 6= K ⋐f U ). GD(U ) = lim dεn (0) :=(cid:26)f ∈ GD(U ) kfkn dεn ≈ 0(cid:27) , U n where n ∈ N>0, we obtain an eR-absorbent and absolutely eR-convex subset of GD(U ) such that i−1 dεn (0, K) ∈ σK. Therefore, these sets generate a coarser topology than the sharp topology σ. The proof is identical to that of Prop. 46 (i). dεn (0) ∩ GDK(U ) = U n dεn (0)) = U n K (U n From the (co-)universal property of inductive limits ([11, Prop. 1.19]) we imme- diately conclude: Proposition 50. Let H be a locally convex topological eR-module. For each non- empty K ⋐f U , let TK : GDK(U ) −→ H be an eR-linear and continuous map. Assume that TK(f ) = TH (f ) if f ∈ GDK(U )∩GDH (U ). Then there exists one and only one map T : GD(U ) −→ H which is eR-linear and continuous and such that T ◦ ιK = TK for all non-empty K ⋐f U . The work [11] includes a detailed analysis of countable inductive limits G = −→ Gn (n ∈ N), where (Gn)n∈N is increasing, G =Sn∈N Gn, and where the topology lim on Gn is that induced by Gn+1. Such inductive limits are called strict. As in the case of classical function spaces like D(Ω) for Ω open in Rn, the importance of strict inductive limits in the theory of eR-locally convex models ultimately stems from the possibility of covering every open set Ω ⊆ Rn by a countable increasing family of compact sets. The key point in the structure theory of strict inductive limits as above is that a countable family (Gn)n∈N permits to define recursively a A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 30 family of neighborhoods of 0. Using the latter, one can prove that the topology on G induces on each Gn its given topology. We will show below that similar properties hold for GD(U ). We shall see that the assumption of U being strongly internal and sharply open are essential for this task. To begin with, we prove a strengthening of Thm. 29: Proposition 51. Let H, K ⋐f U be non-empty sets, with H ⊆ K. Then GDH (U ) is a topological subspace of GDK(U ), i.e., σK(U )GDH (U) = σH (U ). Proof. In this proof we will use the more precise notation Bm Def. 38). ρ (u, K) for balls (see Let V ∈ σK. We claim that V ∩GDH (U ) ∈ σH . For each u ∈ V ∩GDH (U ) there exist m ∈ N and ρ ∈ eR>0 such that Bm ρ (u, K) because GDH (U ) ⊆ GDK(U ) and because the norm k − km doesn't depend on H, K. Therefore, Bm ρ (u, K) ⊆ V . But Bm ρ (u, H) ⊆ Bm ρ (u, H) ⊆ V ∩ GDH (U ), which proves our claim. ρ (u, K) u ∈ W , m ∈ N , ρ ∈ eR>0 , Bm Conversely, if W ∈ σH , then we set V :=[nBm and we claim that W = V ∩GDH (U ). In fact, since W ∈ σH , for all u ∈ W we have ρ (u, H) ⊆ W for some ρ and m. By (8.1) Bm Bm ρ (u, K) ⊆ V , and so u ∈ V ∩GDH (U ). Vice versa, if u ∈ V ∩ GDH (U ), then u ∈ Bm ρ (v, K) for some v ∈ W , m, ρ, such that Bm ρ (u, H) ⊆ Wo ∈ σK, ρ (v, H) ⊆ W . So ku − vkm < ρ and hence u ∈ Bm ρ (v, H) ⊆ W . (8.1) (cid:3) We now show that the space GD(U ) can be seen as a strict inductive limit of a countable increasing family of subspaces GDK(U ). Indeed, since U is strongly internal, we can write U = hUεi for some net (Uε) of subsets of Rn. Since U is non-empty, by [18, Thm. 8], fixing any x = [xε] ∈ U we obtain: ε ) > εN , xε ≤ ε−N . ∃N ∈ N∀0ε : d(xε, U c (8.2) With N as in (8.2), we define ε ) ≥ εj , x ≤ ε−j(cid:9) Kjε : =(cid:8)x ∈ Rn d(x, U c Kj : = [Kjε] ∀j ∈ N ≥N . Using this notation, we have: Theorem 52. If N ∈ N satisfies (8.2) for some x ∈ U , then ∅ 6= Kj ⋐f U for all j ∈ N (i) (ii) Kj ⊆ Kj+1 for all j ∈ N ≥N (iii) U =Sj≥N Kj (iv) For every ∅ 6= K ⋐f U = hUεi there exists some j ≥ N such that K ⊆ Kj. ≥N A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 31 (v) GD(U ) is the strict inductive limit of the family GDKj (U ), j ≥ N : GD(U ) = lim −→ GDKj (U ) (j ≥ N ). Proof. (i), (ii): It follows immediately from the definition that each (Kjε) is com- pact and that (Kjε)ε is sharply bounded, so Kj ⋐f U . Moreover, KN 6= ∅ by (8.2), hence Kj 6= ∅ follows by (ii), which again is immediate from the definition. (iii): If x = [xε] ∈ U = hUεi, then d(xε, U c ε ) > εj1 for ε small and for some j1 ∈ N>0. Since (xε) is moderate, xε ≤ ε−j2 for ε small and some j2 ∈ N. Setting j := max(j1, j2, N ) we hence have that x ∈ Kj. In order to prove (iv), we need the following strengthening of [18, Thm. 11]: Lemma 53. Let H = [Hε] ⋐f V = hVεi, then ∃j ∈ N∀[xε] ∈ [Hε]∀0ε : d(xε, V c ε ) ≥ εj. (8.3) Proof of Lemma 53. Equation (8.3) expresses that for all representatives (xε) ∈ Rn M , if xε ∈ Hε for ε small, then ∀0ε : d(xε, V c By contradiction, assume that ε ) ≥ εj. M : (cid:0)∀0ε : xjε ∈ Hε(cid:1) , ∃(εjk)k ↓ 0 ∀k : d(xjεjk , V c ∀j ∈ N∃(xjε) ∈ Rn jk. By recursively applying this condition, we get that for all j ∈ N there exists a moderate (xjε) and some εj ∈ (0, 1] such that xjε ∈ Hε for ε ≤ εj and εjk ) < εj d(xjεjk , V c εjk ) < εj jk, (8.4) where (εjk)k ↓ 0. Since (εjk)k ↓ 0, without loss of generality we can assume to have defined recursively (εj)j so that (εj)j ↓ 0 and εj > εjkj > εj+1 for some subsequence (kj)j ↑ +∞. Set xε := xjε ∈ Hε for ε ∈ (εj+1, εj], so that xεjkj for all j. Then (xε) ∈ Rn M since H is sharply bounded and x := [xε] ∈ H ⊆ hVεi, which entails = xjεjkj (8.5) Therefore, for j ∈ N sufficiently big, (8.5) holds at ε = εjkj < 1 and j > q. Thus d(xjεjkj ∃q ∈ R>0 ∀0ε : d(xε, V c , which contradicts (8.4). ) = d(xεjkj ε ) > εq. ) > εq jkj (cid:3) , V c εjkj , V c εjkj > εj jkj Continuing the proof of (iv), if K ⋐f U is non-empty, by applying Lemma 53 we obtain ∃j1 ∈ N∀[xε] ∈ K ∀0ε : d(xε, U c ε ) ≥ εj1 . On the other hand, sharp boundedness of (Kε), where K = [Kε], implies ∃j2 ∈ N∀[xε] ∈ K ∀0ε : xε ≤ ε−j2 . A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 32 The converse inclusion follows directly from (i). Therefore, for j := max(j1, j2, N ), we get K ⊆ Kj, hence (iv), and thereby also GDK(U ) ⊆ GDKj (U ) (using Thm. 29). It follows that GD(U ) ⊆ Sj≥N GDKj (U ). It remains to prove that the topology σ on GD(U ) coincides with the inductive eR-locally convex topology generated by (cid:0)GDKj (U )(cid:1)j≥N . Let us denote the latter topology by σ′. We have σ ⊆ σ′ by definition of the inductive topology and by Thm. 52.(i). To see that, conversely, σ′ ⊆ σ we show that for every K ⋐f U the inclusion (GDK(U ), σK) ֒→ (GD(U ), σ′) is continuous (see Thm. 49). Now given any K ⋐f U , by what we have proved above there exists some j ≥ N with K ⊆ Kj. But then Prop. 51 implies the continuity of (GDK(U ), σK) ֒→ (GDKj (U ), σKj ) ֒→ (GD(U ), σ′) and thereby our claim. (cid:3) As a consequence of this result, we can now prove a series of corollaries by applying the general theorems of [11] concerning countable strict inductive limits. Corollary 54. If ∅ 6= K ⋐f U = hUεi, then GDK(U ) is a topological subspace of GD(U ), i.e., σ(U )GDK (U) = σK (U ). Proof. According to Thm. 52 (iv) we may pick j ≥ N such that K ⊆ Kj. By [11, Prop. 1.21], GDKj (U ) carries the trace topology of GD(U ). Since, in turn, GDK(U ) is a topological subspace of GDKj (U ) by Prop. 51, the claim follows. Corollary 55. GD(U ) is separated. Proof. This follows from Thm. 41 (v) and [11, Cor. 1.24]. (cid:3) (cid:3) Lemma 56. If ∅ 6= H, K ⋐f U and H ⊆ K, then GDH (U ) is closed in GDK(U ). Proof. This is immediate from Prop. 51 and Thm. 48. (cid:3) Corollary 57. If ∅ 6= K ⋐f U , then GDK(U ) is closed in GD(U ). Proof. This follows from Thm. 48, Cor. 55, and Cor. 54. (cid:3) Corollary 58. Let B ⊆ GD(U ), then B is bounded in GD(U ) if and only if there exists a non-empty K ⋐f U such that B is bounded in GDK(U ). Proof. ⇒: This follows from [11, Thm. 1.26] and Lemma 56. ⇐: If B is bounded in GDK(U ), then [11, Lem. 1.27] yields ∀(un)n ∈ BN ∀(λn)n ∈ eRN : λn → 0 in eR =⇒ λnun → 0 in GDK(U ). Pick j ≥ N such that K ⊆ Kj. Since generalized norms k − km do not depend on K, Kj, condition (8.6) holds also in GDKj (U ) ⊇ GDK(U ). Therefore, from [11, (8.6) A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 33 Lem. 1.27] we get that B is bounded in GDKj (U ) and [11, Thm. 1.26] yields that B is bounded in GD(U ). (cid:3) A similar proof applies to this corollary, which is a consequence of [11, Cor. 1.29]: Corollary 59. Let (un)n ∈ GD(U )N, then un → 0 in GD(U ) if and only if there exists a non-empty K ⋐f U such that un ∈ GDK(U ) and un → 0 in GDK(U ). Finally, from [11, Thm. 1.32], Lemma 56 and Thm. 48 we obtain: Corollary 60. GD(U ) with the sharp topology is complete. Using Lemma 53, we can also show that any compactly supported generalized assuming that U is a strongly internal set. a net (uε) of smooth functions which are compactly supported in an arbitrarily smooth function f ∈ GDK(U, Y ) on a sharply open set U ⊆ eRn is defined by εa(0)i of K = [Kε]. We recall that in this section we are small extension hKε + B E Theorem 61. Let ∅ 6= K ⋐f U . Let Y ⊆ eRd, f ∈ GDK(U, Y ) and K = [Kε] ⋐f eRn. Let j ∈ N be as in (8.3) and a ∈ R such that a ≥ j. Then there exist nets (uε), (Hε) such that: εa (0) for all ε. [Hε] ⋐f U (uε) defines f and uε ∈ DHε(Rn, Rd) for all ε (i) (ii) (iii) Hε ⊆ Kε + B E Proof. Let U = hUεi, where each Uε ⊆ Rn is an open set (cf. [18, Cor. 9]), and let (vε) satisfy Def. 19 for f and K = [Kε]. By [18, Thm. 11], we can assume Kε ⊆ Uε εa/2(0), and denote by χLε the characteristic function for all ε. Let Lε := Kε + B E 1 (0)) have unit integral and set ψε := (εa/2)−nψ(2x/εa). Then of Lε. Let ψ ∈ D(B E M (Rn), and (χLε ∗ ψεa )Kε = 1. Set uε := (χLε ∗ ψε) · vε. Then (χLε ∗ ψε) (ψε) ∈ E s defines a GSF of the type eRn −→ eR and hence (uε) defines a GSF of the type U −→ eRd. Moreover, Hε := supp(uε) ⊆ supp(χLε ∗ ψε) ⊆ Kε + B E a ≥ j, we have [Hε] ⊆ hUεi = U . It remains to prove that f (x) = [uε(xε)] for all x = [xε] ∈ U . Suppose this was not the case. Then there would exist some y = [yε] ∈ U , some b > 0 and a sequence εk ց 0 such that εa (0). Since uεk (yεk ) − vεk (yεk ) ≥ εb (8.7) for all k ∈ N. By Lemma 21, we may without loss of generality assume that either yε ∈ Kε for all ε or that y ∈ ext(K). In the first case, vε(yε) = uε(yε) for all ε, contradicting (8.7). In the second case, k Since (vε(yε))) is negligible, we again arrive at a contradiction to (8.7). (cid:3) vε(yε) − uε(yε) ≤ 2vε(yε). A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 34 Using this result, we can now prove the extension of property (ii) of Thm. 26 to Assume that we have an operator I : D(Rn) −→ R with the property that if (uε) and (vε) define f ∈ GD(U ), where uε, vε ∈ D(Rn), then [I(uε)] = [I(vε)] ∈ eR. Then Thm. 61 permits to extend I to the whole of GD(U ). arbitrary codomains Y ⊆ eRd: Theorem 62. Let ∅ 6= K ⋐f U , Y ⊆ eRd and f ∈ GDK(U, Y ), then ∃! ¯f ∈ GDg(K, Y ) : ¯fK = fK. Proof. We only have to prove that ¯f (x) ∈ Y for all x ∈ eRn. Let (uε) and (Hε) as in Thm. 61. By Thm. 24, we have that (uε) also defines ¯f . For each ε pick any point hε ∈ ∂Hε and set ¯xε := xε if xε ∈ Hε, and ¯xε := hε otherwise. Therefore ¯x := [¯xε] ∈ [Hε] ⊆ U . Then, if xε /∈ Hε, uε(xε) − uε(¯xε) = uε(xε) − uε(hε) = 0 because uε ∈ DHε (Rn, Rd). Thus ¯f (x) = f (¯x) ∈ Y . (cid:3) 9. Conclusions and Further developments The notion of functionally compact set we introduced in the present work permits to show that compactly supported GSF are close analogues of classical compactly supported smooth functions. In particular, their functional analytic properties parallel those of the test function space of distribution theory. At the same time, for suitable K, the space GDK(eRn) contains extensions to all CGF Gs(Ω) and hence also all Schwartz distributions. The theory developed here opens the door to addressing several central topics in the theory of nonlinear generalized functions from a new angle. As indicated after Thm. 61, a direct generalization of the integral of compactly supported functions to compactly supported GSF is feasible. An immediate application of this lies in a theory of integration for GSF that we hope will allow to harmonize the Schwartz view of generalized functions as functionals with that prevalent in Colombeau theory of considering generalized functions as pointwise maps. Our approach will take inspiration from Garetto's very fruitful duality theory of locally convex eC-modules [12, 15, 13]. A further natural development of the present article goes in the direction of a generalization to suitable types of asymptotic gauges (see [20, 19]) and hence to the full and the diffeomorphism invariant Colombeau algebras. for ODE with a GSF right hand side, or a Hahn-Banach theorem for functionals mit to generalize results from classical analysis, like a Picard-Lindelof theorem Moreover, one can ask whether eR-valued generalized norms in GDK(U ) per- I : GDK(U ) −→ eR defined by diffeologically smooth functionals (see [21]) of the type Iε : DKε(Uε) −→ R, analogously to the way a GSF is defined by a net of smooth functions. A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 35 Acknowledgment: We would like to thank H. Vernaeve for helpful discussions and the anonymous referee for several suggestions that have led to considerable improvements in Sec. 3. References [1] Aragona, J., Colombeau, J.F., Juriaans, S.O., Locally convex topological algebras of general- ized functions: compactness and nuclearity in a nonlinear context, Trans. Amer. Math. Soc., to appear. [2] Aragona, J., Fernandez, R., Juriaans, S. O., A Discontinuous Colombeau Differential Calcu- lus. Monatsh. Math. 144 (2005), 13 -- 29. [3] Aragona, J., Fernandez, R., Juriaans, S.O., Natural topologies on Colombeau algebras, Topol. Methods Nonlinear Anal. 34 (2009), no. 1, 161 -- 180. [4] Aragona, J., Juriaans, S. O., Some structural properties of the topological ring of Colombeau's generalized numbers, Comm. Algebra 29 (2001), no. 5, 2201 -- 2230. [5] Aragona, J., Juriaans, S.O., Oliveira, O.R.B., Scarpal´ezos, D., Algebraic and geometric theory of the topological ring of Colombeau generalized functions, Proc. Edinb. Math. Soc. (2) 51 (2008), no. 3, 545564. [6] Aragona, J., Fernandez, R., Juriaans, S.O., Oberguggenberger, M., Differential calculus and integration of generalized functions over membranes. Monatsh. Math. 166 (2012), 1 -- 18. [7] Biagioni, H.A., A Nonlinear Theory of Generalized Functions, Lecture Notes in Mathematics 1421, Springer, Berlin, 1990. [8] Colombeau, J.F., New generalized functions and multiplication of distributions. North- Holland, Amsterdam, 1984. [9] Colombeau, J.F., Elementary introduction to new generalized functions. North-Holland, Am- sterdam, 1985. [10] Delcroix, A., Hasler, M.F., Pilipovic, S., Valmorin, V., Sequence spaces with exponent weights. Realizations of Colombeau type algebras, Dissertationes Mathematicae 447, 1 -- 73 (2007). [11] Garetto, C., Topological structures in Colombeau algebras: Topological C-modules and du- ality theory. Acta Appl. Math. 88, no. 1, 81 -- 123 (2005). [12] Garetto, C., Topological structures in Colombeau algebras: investigation of the duals of Gc(Ω), G(Ω) and GS (Rn), Monatsh. Math. 146 (2005), no. 3, 203 -- 226. [13] Garetto, C., Fundamental solutions in the Colombeau framework: applications to solvability and regularity theory, Acta. Appl. Math. 102 (2008), 281 -- 318. [14] Garetto, C., Closed graph and open mapping theorems for topological eC-modules and appli- cations, Math. Nachr. 282 (2009), no. 8, 1159 -- 1188. [15] Garetto, C., Hormann, G., On duality theory and pseudodifferential techniques for Colom- beau algebras: generalized delta functionals, kernels and wave front sets, Bull. Acad. Serbe Cl. Sci. (2006), 31, 115 -- 136. [16] Garetto, C., Vernaeve, H., Hilbert eC-modules: structural properties and applications to variational problems, Trans. Amer. Math. Soc. 363 (2011), no. 4, 2047 -- 2090. [17] Giordano, P., Kunzinger, M., New topologies on Colombeau generalized numbers and the Fermat-Reyes theorem, J. Math. Anal. and Appl. 399 (2013) 229 -- 238. [18] Giordano, P., Kunzinger, M., Vernaeve, H., Strongly internal sets and generalized smooth functions, J. Math. Anal. Appl. 422 (2015) 56 -- 71. A CONVENIENT NOTION OF COMPACT SET FOR GENERALIZED FUNCTIONS 36 [19] Giordano, P., Luperi Baglini, L., Asymptotic gauges: Generalization of Colombeau type algebras, Math. Nachr. 1 -- 28 (2015). [20] Giordano, P., Nigsch, E., Unifying order structures for Colombeau algebras, Math. Nachr. Vol- ume 288, Issue 11-12, pages 1286 -- 1302, (2015). [21] Giordano P., Wu E., Categorical frameworks for generalized functions, Arab. J. Math., Vol- ume 4, Issue 4, 301 -- 328, 2015. [22] Grosser, M., Kunzinger, M., Oberguggenberger, M., Steinbauer, R., Geometric theory of generalized functions, Kluwer, Dordrecht, 2001. [23] Mayerhofer, E., Spherical completeness of the non-Archimedean ring of Colombeau general- ized numbers, Bull. Inst. Math. Acad. Sin. (N.S.) 2 (2007), no. 3, 769 -- 783. [24] Mayerhofer, E., On Lorentz geometry in algebras of generalized functions, Proc. Roy. Soc. Ed- inburgh Sect. A 138 (2008), no. 4, 843 -- 871. [25] Oberguggenberger, M. Multiplication of Distributions and Applications to Partial Differential Equations, volume 259 of Pitman Research Notes in Mathematics. Longman, Harlow, 1992. [26] Oberguggenberger, M., Kunzinger, M., Characterization of Colombeau generalized functions by their pointvalues, Math. Nachr. 203 (1999), 147 -- 157. [27] Oberguggenberger, M., Vernaeve, H., Internal sets and internal functions in Colombeau the- ory, J. Math. Anal. Appl. 341 (2008) 649 -- 659. [28] Scarpal´ezos, D., Some remarks on functoriality of Colombeau's construction; topological and microlocal aspects and applications, Int. Transf. Spec. Fct. 1998, Vol. 6, no. 1 -- 4, 295 -- 307. [29] Scarpal´ezos, D., Colombeau's generalized functions: topological structures; microlocal prop- erties. A simplified point of view, I. Bull. Cl. Sci. Math. Nat. Sci. Math. no. 25 (2000), 89 -- 114. [30] Vernaeve, H., Ideals in the ring of Colombeau generalized numbers, Comm. Algebra (2010) 38(6):2199 -- 2228. [31] Vernaeve, H. Nonstandard principles for generalized functions, J. Math. Anal. Appl., Volume 384, Issue 2, 2011, 536 -- 548. University of Vienna, Austria E-mail address: [email protected] University of Vienna, Austria E-mail address: [email protected]
1204.5654	4	1204	2012-06-13T14:18:25	Oscillating convolution operators on the Heisenberg group	[ "math.FA" ]	In this paper, we consider oscillating convolution operotors on the Heisenberg group $H^n_a$ with respect to the norm $\rho(x,t) = \rho_1(b x, b t)$ with $\rho_1(x,t)= (\|x\|^4 + t^2)^{1/4}$. We obtain $L^2$ boundedness properties using the oscillatory integral estimates for degenerate phases in the Euclidean setting. Our result contains an improvement of the Lyall's result for the cases $\frac{a^2}{b^2} \geq C_{\beta}$.	math.FA	math	OPTIMAL CONDITIONS FOR L2 BOUNDEDNESS OF STRONGLY SINGULAR CONVOLUTION OPERATORS ON THE HEISENBERG GROUP WOOCHEOL CHOI Abstract. Strongly singular convolution operators TKα,β with the kernel Kα,β on the Heisen- berg group Hn a were introduced in [10]. For case a2 < Cβ , Laghi and Lyall [8] obtained the sharp range for α, β for which the operator TKα,β are bounded on L2(Hn a ). They used the classical L2 −L2 boundedness theorem for oscillatory integral operators with non-degenerate phases. But, if a2 ≥ Cβ , the phase function related with the operators are no more non-degenerate. However, in this paper, we obtain the sharp range for α, β for the case a2 ≥ Cβ . To carry out this case, we show that the canonical graph related with the phases satisfy folding type conditions and utilize the recent developed theory for degenerate oscillatory integral operators (see [5], [11]). Our setting is on the Heisenberg group Hn a (a ∈ R∗) which have base manifold R2n+1 with the group law 1. Introduction where J is the 2n × 2n matrix (x, t) · (y, s) = (x + y, s + t − 2axtJy) 0 −In In 0! (In is the n × n identity matrix). This group has the following dilation law λ · (x, t) = (λx, λ2t). For each kernel K, associated convolution operators are defined by TKf (x, t) := K ∗ f (x, t) =ZHn K(cid:0)(x, t) · (y, s)−1(cid:1) f (y, s)dydx. We say that the operator TK is bounded on Lp(Hn) if there exist a C > 0 such that kTKf kp ≤ Ckf kp, for all f ∈ C∞ 0 (Hn) And a natural quasi-norm on the Heisenberg group is defined by ρ(x, t) = (x4 + t2)1/4. This quasi-norm satisfies ρ(λ · (x, t)) = λρ(x, t). For this quasi-norm, we define the strongly singular kernels for each α > 0 • Kα,β(x, t) = ρ(x, t)−(2n+2+α)eiρ(x,t)−β χ(ρ(x, t)), β > 0, where χ is a smootho bump function in a small neighborhood of the origin. We let the operator TKα,β be the convolution operators on the Heisenberg group Hn a with the kernel Kα,β. This operator was firstly introduced in [10] and there, the necessary condition is referred as α ≤ (n + 1 2 )β and it was shown that TKα,β is bounded if α ≤ nβ. To show this, the group fourier transform was utilized and a lengthy calculation for estimating some oscillatory integrals was needed to obtain the 2000 Mathematics Subject Classification. Primary. 1 2 WOOCHEOL CHOI results. But, Laghi and Lyall [8] showed that we can get sharp results in the restricted case a2 < Cβ for some Cβ > 0 only using the Hormander's L2 − L2 boundedness theorems for non-degenerate oscillatory integral operators [1]. In this paper, we consider the cases a2 ≥ Cβ and obtain sharp conditions using the recent theory for oscillatory integral operators with degenerate phases (See section 2 for the details). The theory for the degenerate oscillatory integral operators have been developed deeply, but the use of the theorem has been restricted to the X-ray transform. Strongly singular convolution operators were studied originally in Rn. Such operators represent some oscillating multipliers and operators of this type were first studied, using Fourier transform techniques, in the Euclidean with ρ(x) = x by Hirschman [7] in the case d = 1 and then in higher dimensions by Wainger [15], Fefferman [3], and Fefferman and Stein [4]. On the other hand, a similar kind of convolution operators with the kernel with α, β > 0 also have been studied in [12],[13]. This kernel has no sinularity near the zero, but it has relatively small decaying property at infinity. Note that the case β = 1 corresponds to the kernel of Bochner-Riesz means. For β 6= 1, the Lp − Lq estimates including the hardy space estimates were well understood. The difference between two cases comes from the fact that the phase kernel x − yβ is degenerate only if β = 1. We now consider an analogue problem on the Heisenberg groups with the following kernels 1 xn−α eixβ • Lα,β(x, t) = ρ(x, t)−(2n+2−α)eiρ(x,t)β χ(ρ(x, t)−1), β > 0. We denote the gruop convolution operators TLα,β with the kernel Lα,β. In literatures, operators like TKα,β are called as strongly singular operators and TLα,β as oscillating convolution operators. In this paper, we shall find the optimal ranges of α and β where the convolution operators associated with Kα,β ,Lα,β are bounded on L2(Hn). For a2 b2 ≥ Cβ, the phases are no more non-degenerate. So, we need to deal with oscillatory inte- gral operators with degenerate phases. Theory for this kind of operators have been studied largely with considering various conditions on phase functions to give a different decaying properties, See [6]. We will use results in [5],[11]. To utilize such theory, we should carefully investigate the folding type for our phases. Interestingly, we have different folding types according to the values a, b and β. Before stating our results, we recall the former results in [8], [10]. Theorems [8],[10]. (1) TKα,β is bounded on L2(Hn) if α ≤ nβ. (2) If 0 < a2 < Cβ , then TKα,β is bounded on L2(Hn) if and only if α ≤ (n + 1/2)β. (Cβ = β+2 2 (2β + 5 +p(2β + 5)2 − 9)) We obtain sharp results on the L2 → L2 boundedness of TKα,β for the cases a2 ≥ Cβ. Namely, we obtain the following results: Theorem 1.1. If a2 > Cβ , then TKα,β is bounded on L2(Hn a2 = Cβ, then TKα,β is bounded on L2(Hn a ) if and only if α ≤ (n + 1 4 )β. a ) if and only if α ≤ (n + 1 3 )β. If For the operators TLα,β , we also have the sharp ranges for the L2 → L2 boundedness except the cases β = −1 or −2. Theorem 1.2. TLα,β is bounded on L2 if and only if (i) 0 < β < 1 : For a2 < Cβ, α ≤ (n + 1 2 )β, for a2 = Cβ, α ≤ (n + 1 4 )β, and for a2 > Cβ, α ≤ (n + 1 3 )β. OPTIMAL BOUNDS FOR OSCILLATING CONVOLUTION OPERATORS ON THE HEISENBERG GROUP 3 (ii) 1 < β < 2 : α ≤ (n + 1 (iii) 2 < β : α ≤ (n + 1 2 )β. 3 )β. Remark 1.3. For the cases β = −1 or β = −2, we can also obtain the sharp results for some values α, b where the phase becomes non-degenerate or of folding type 2. But, in these cases, higher order folding type conditions than 3 appear for some α, b and the present theory for degenerate oscillatory integral estimates does not cover these cases. The theory have been established optimally only for the one or two folding cases ([5], [11]). Remark 1.4. In Rn, the oscillating kernel is of the form x−γeixβ In this case, the different behavior for the phases xβ according to the value a is characterized only by the with β 6= 0. two cases where β 6= 1 or β = 1. Precisely, we have det(cid:16) ∂2 det(cid:16) ∂2 ∂x∂y x − yβ(cid:17) 6= 0 for x 6= y, but ∂x∂y x − y(cid:17) = 0 for every x 6= y and this is the hardest case, which is correspond to the Bochner-Riesz means operators. We see that for our cases, β = 1 and β = 2 present the lowest degeneracy oscillatory operators (see 4.1) and it also hard to establish from the lack of the theory of oscillatory integral estimates for lower folding type cases. This paper is organized as follows. In section 2, we proceed an usual process to decompose the kernel dyadically and reduce our problem to a local oscillatory integral estimates. In section 3, we recall some essential things for our oscillatory integral operators with degenerate phase functions. Finally, in section 4, we study some geometry of the canonical relation and projection maps related with the phase functions. Then, we conclude the proof of the main theorems. 2. Dyadic decompostion and Localization We use a standard argument to reduce our problems to some oscillatory integral estimates on R2n+1. For euclidean space, the argument is well explained in [14]. Though the same mechanism is easily adapted, we shall prove it because our setting is on a group. We split the kernel Kα,β and Lα,β as ∞ Kα,β(x, t) = Lα,β(x, t) = η(2jρ(x, t))Kα,β(x, t) (2.1) η(2−jρ(x, t))Lα,β(x, t) ∞ Xj=0 Xj=0 0 (R) is a bump function supported in [1/2, 2]. and satisfiesP∞ where η ∈ C∞ r ≤ 1. We let K j For notational convenience, we omit the index α and β from now. We let Tjf = K j α,β = η(2j(ρ(x, t)))Kα,β (x, t) and Tjf = Kj ∗ f and Lj α,β ∗ f and Sjf = Lj α,β ∗ f . Then, j=0 η(2jr) = 1 for all 0 < α,β = η(2−jρ(x, t))Lα,β (x, t). Lemma 2.1. For each N ∈ N, there exists a constant CN such that kT ∗ kS∗ j Tj ′ kL2→L2 + kTjT ∗ j Sj ′ kL2→L2 + kSjS∗ j ′ kL2→L2 ≤ CN 2−max{j,j ′}N j ′ kL2→L2 ≤ CN 2−max{j,j ′}N (2.2) holds when j − j′ ≥ cβ for some sufficiently large constant cβ > 0. Proof. The proof follows using the integration parts technique in usual ways. See [10] where the proof for Tj is given. (cid:3) 4 WOOCHEOL CHOI By the Cotlar-Stein Lemma, we only need to show that holds uniformly for j ∈ N with some constant C > 0. kTjkL2→L2 + kSjkL2→L2 ≤ C Let K j α,β(x, t) = K j Lj α,β(x, t) = Lj α,β(2−j · (x, t)) = η(ρ(x, t))2j(Q+α)ρ(x, t)−Q−αei2jβ ρ(x,t)−β α,β(2−j · (x, t)) = η(ρ(x, t))2−j(Q−α)ρ(x, t)−Q+αei2jβ ρ(x,t)β Let fj(x, t) = f (2j · (x, t)). Then K j ∗ f (2−j · (x, t)) = 2−jQ( K j ∗ fj)(x, t). kK j α,β ∗ f (x, t)kL2 = 2−jQ/2kKα,β ∗ f (2−j · (x, t))kL2 ≤ 2−jQ/2 · 2−jQk K j ≤ 2−jQ/2 · 2−jQk K j = 2−jQk K j α,βkL2→L2kf kL2. α,β ∗ f (2j · ())(x, t)kL2 α,βkL2→L2kf (2−j·)kL2 , . (2.3) (2.4) (2.5) (2.6) Similarly, we also have kLj α,β ∗ f kL2 ≤ 2jQk Lj α,βkL2→L2kf kL2. So, it suffice to prove suitable boundedness for Tj and Sj. Now, we further modify our operator locally using the fact that the kernels of Tj and Sj are supported in {(x, t) : ρ(x, t) ≤ 2}. To do a and each B(gk, 4) contains only dn's other gl members in G. this, we find a set of point G = {gk : k ∈ N} such that Sk∈N B(gk, 2) = Hn We can split f =P∞ j f (x, t) =Z K j α,β(cid:0)(x, t) · (y, s)−1(cid:1) · η(cid:0)ρ(cid:0)(x, t) · g−1 k=1 fk with each fk supported in B(gk, 2). And we let T l,k k (cid:1)(cid:1) η(cid:0)ρ(cid:0)(x, t) · g−1 l (cid:1)(cid:1) f (y, s)dyds Then, L2(B(gk,2)) flk2 L2(B(gk,2)) k Tj ∗ f k2 L2(Hn a ) ≤ ≤ = . . ∞ k Tj ∗ f k2 ∞ Xl=1 ∞ ∞ k Tj ∗ Xk=1 Xk=1 Xk=1 Xk=1 Xl:ρ(gl·g−1 ∞ sup ρ(gl·g−1 k )≤2 k Tj ∗ X{l:ρ(gl·g−1 k )≤2} flk2 L2((B(gk,2))) k T l,k j kL2→L2kflk2 L2 k )≤2 k T l,k j kL2→L2kf k2 L2. (2.7) (2.8) (2.9) (2.10) Because det (Dx,t ((x, t) · g)) = 1 for all g ∈ Hn a , T l,k j f ((x, t) · gk) =Z K j(cid:0)(x, t) · (y, s)−1(cid:1) η(ρ(x, t))η(ρ(y, s) · (gk · g−1 ) . 1. If we let ψ ((x, t), (y, s)) = η(ρ(x, t))η(ρ((y, s)·(gk ·g−1 l ))f ((y, s) · gk)dyds. ))) and substitute l Notice that ρ(gk ·g−1 f as f (x) := f (x · gk). l OPTIMAL BOUNDS FOR OSCILLATING CONVOLUTION OPERATORS ON THE HEISENBERG GROUP 5 k )≤2 k T l,k Pj,k,ρ(gl·g−1 j k will be obtained if we prove kAjkL2→L2 . 1 for Aj f (x, t) =Z K j α,β(cid:0)(x, t) · (y, s)−1(cid:1) ψ ((x, t), (y, s)) f (y, s)dyds with a compactly supported smooth function ψ. Since K j α,β(x, t) = η(ρ(x, t))2j(Q+α)ρ(x, t)−Q−αei2jβ ρ(x,t)−β α,β(x, t) = η(ρ(x, t))2−j(Q+α)ρ(x, t)−Q+αei2jβ ρ(x,t)β Lj The matters are reduced to show that kTAj kL2→L2 . 2jQ and kTBj kL2→L2 . 2−jQ Let Aj(x, t) = 2jαµ(x, t)ei2jβ ρ(x,t)−β Bj(x, t) = 2−jαµ(x, t)ei2jβ ρ(x,t)β (2.11) (2.12) where with µ is a smooth funtion supported on ρ(x, t) ∼ 1. We define the operators LAj , LBj by LAj f (x, t) =Z Aj(cid:0)(x, t) · (y, s)−1(cid:1) ψ ((x, t), (y, s)) f (y, s)dyds LBj f (x, t) =Z Bj(cid:0)(x, t) · (y, s)−1(cid:1) ψ ((x, t), (y, s)) f (y, s)dyds. To prove Theorem 1.1 and Theorem 1.2, it suffices to establish the following two theorems. (2.13) Theorem 2.2. The following inequalities hold uniformly for j. If a2 > Cβ, If a2 = Cβ, kLAj kL2→L2 . 2j(α−(n+ 1 kLAj kL2→L2 . 2j(α−(n+ 1 3 )β), 4 )β). Theorem 2.3. The following inequalities hold uniformly for j. (i) − 1 < β < 0, If a2 < Cβ, If a2 = Cβ, (ii) − 2 < β < −1, (iii) β < −2, kLBj kL2→L2 . 2j(α−(n+ 1 kLBj kL2→L2 . 2j(α−(n+ 1 2 )β), 4 )β). kLBj kL2→L2 . 2j(α−(n+ 1 2 )β). kLBj kL2→L2 . 2j(α−(n+ 1 3 )β). In the next section, we shall breifly review on the theory related to the operators LAj and LBj . And in section 4, we will expoit some required geometry related with the phase function ρ(x, t)β to conclude the proof of Theorem 2.2 and Theorem 2.3. 3. L2 theory for oscillatory integral operators In this section, we review on some L2 → L2 theory for oscillatory integral operators. The form of operators we are concern is given as T φ λ f (x) =ZRn eiλφ(x,y)a(x, y)f (y)dy where φ, a ∈ C∞(Rn × Rn) and a has a compact support. We firstly state the fundamental theorem of Hormander [1]. Theorem 3.1. Suppose that the phase function φ satisfies det(cid:16) ∂2φ Then we have the following inequality ∂xi∂yj(cid:17) 6= 0 on the support of a. where the implicit constant is independent of λ. kT φ λ kL2→L2 . λ− n 2 , λ ≥ 1 6 WOOCHEOL CHOI We say that φ is non-degenerate if it satisfies the assumption of the above theorem. And we equals to zero. This say φ is degenerate if there is some point (x0, y0) where det(cid:16) ∂2 theorem gives the sharp decaying of kT φ λ kL2→L2 in terms of λ. But, the phase functions of our operators (see ) can become degenerate according to the values of a and β. For a degenerate φ, the optimal number κφ for which the inequality kTλkL2→L2 . λ−κφ holds would be less than n 2 . The number κφ's are related to the folding type conditions of the phase φ (See below). For phases whose folding degrees are ≤ 3, the sharp numbers κφ were obtained in [5], [11]. We shall use the results. In fact, the results for folding types ≤ 3 in [5] is the best known results and there are no results for folding types > 3 except the very restricted result in [2]. Finding completely the numbers κφ corresponding to all degenerate phases φ seems a widely open problem. ∂xi∂yj(cid:17)(cid:12)(cid:12)(cid:12)(x0,y0) It is well-known that the decaying property is strongly related th the geometry of the canonical relation Cφ = {(x, φx(x, y), y, −φy(x, y)) ⊂ T ∗(Rn x ) × T ∗(Rn y ) (3.1) To describe the geometry of Cφ, we need the following definition Definition 3.2. Let M1,M2 be smooth manifolds of dimension n, and f : M1 → M2 is a smooth map of corank ≤ 1. We let the singular set S = {P ∈ M1 : rank(Df ) < n at P }. Then we say that f has a k− type fold at P0 for P0 ∈ S if (1) rank(Df )P0 = n − 1, (2) det(Df ) vanishes of k order in the null direction at P0. Here, the null direction means the unique direction vector v such that (Dvf )P0 = 0. Now we consider the two projection maps πL : CΦ → T ∗(Rn x ) and πR : CΦ → T ∗(Rn y ). (3.2) Then the following theorem is proved in [11] for one folds cases and in [5] for two folds cases. Theorem 3.3. Suppose that the projection maps πL and πR have one fold singularities, then kTλf kL2(Rn) ≤ Cλ− (n−1) 2 − 1 3 kf kL2(Rn). If the projection maps φL and πR have two hold singularities, kTλf kL2(Rn) ≤ Cλ− (n−1) 2 − 1 4 kf kL2(Rn). 4. Geometry of the Canonical relation maps In this section, we study the projection maps (3.2) associated with the phase function of our operators in Theorem 2.2 and Theorem 2.3 to use Theorem 3.3. Recall that ρ(x, t) = (x4 + t2)1/4 and the phase function φ of the integral operators LAj and LBj is To write the group law explicitly, we write x = (x1, x2) and y = (y1, y2) with x1, x2, y1, y2 ∈ Rn. Let Φ(x, t) = ρ(x, t)−β for simplicity, then we have φ(x, t, y, s) = ρ−β(cid:0)(x, t) · (y, s)−1(cid:1) . φ(x, t, y, s) = Φ(cid:0)x1 − y1, x2 − y2, t − s − 2a(x1y2 − x2y1)(cid:1) (4.1) OPTIMAL BOUNDS FOR OSCILLATING CONVOLUTION OPERATORS ON THE HEISENBERG GROUP 7 Let us use the derivatives ∂x, ∂t to denote the position derivatives of Φ (i,e. ∂xj Φ = Φxj , ∂tΦ = Φt and ∂xj (xjΦ) = xj Φxj ). Firstly, to determine the phase function Φ is non-degenerate or not, we should calculate the determinant of the matrix So, we now compute the matrix H. In (4.1), by the chain-rule, we have H =(cid:18) ∂2φ(x, t, y, s) ∂(y,s)∂(x,t) (cid:19) . ∂ ∂xj ∂ ∂xj+n φ(x, t, y, s) = (∂xj + 2ayn+j∂t)Φ((x,t)·(y,s)−1) (4.2) φ(x, t, y, s) = (∂xj+n − 2ayj∂t)Φ((x,t)·(y,s)−1) Let A(y) = I 0 2aJy 1 ! Using a similar calculation, we see that H(x, t, y, s) = A(x) (∂i∂jΦ) A(y)t + 2a∂tΦ J 0 0 = A(x)"(∂i∂jΦ) + 2a∂tΦ J 0 0! A(y)t = J 0 0 0 where the second equality holds because A(x) J L(x, t, y, s) = "(∂i∂jΦ) + 2a∂tΦ J 0 0 A(y)t 0!(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)((x,t)·(y,s)−1) 0!#(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)((x,t)·(y,s)−1) 0!#(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)((x,t)·(y,s)−1) 0!. We let 0 0 . (4.3) (4.4) (4.5) Thus, to study the matrix H, it is enough to analyze the matrix L. Morerover, we have det(A(x)) = det(A(y)) = 1, which implies det(H(x, t, y, s)) = det(L(x, t, y, s)). In [8], the determinant of the matrix H was calculated directly without using this observation. Here we calculate it by obtaining the determinant of L. We now calculate the hessian matrix of Φ. For 1 ≤ i, j ≤ 2n, ∂jΦ(x, t) = − ∂tΦ(x, t) = − β 4 β 4 (x4 + t2)− β 4 −1(4xjx2) (x4 + t2)− β 4 −1(2t) (4.6) and ∂i∂jΦ(x, t) = β(x4 + t2)− β ∂i∂tΦ(x, t) = β(β + 4)(x4 + t2)− β 4 −2x2xi · 4 −2(cid:2)(β + 4)x4 − 2(x4 + t2)(cid:3) xixj − β(x4 + t2)− β 4 −1δij x2 (4.7) (4.8) t 2 t Φ(x, t) = β(β + 4)(x4 + t2)− β ∂2 4 −2 t 2 · t 2 − β(x4 + t2)− β 4 −1 1 2 . 8 WOOCHEOL CHOI Let D = (x2x, t 2 )t, then the above computations show that (∂i∂jΦ) + 2a∂tΦ J 0 0 0!(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(x,t) = β(β + 4)(x4 + t2)− β 4 −2D · Dt− β(x4 + t2)− β 4 −1 x2I + atJ + 2x · xt 0 = −β(x4 + t2)− β 4 −1(E + R) To write the matrices simply, we let B = x2I + atJ, K = x · xt and E = B + 2K 0 2! and R = − 0 1 (β + 4) x4 + t2 D · Dt. (4.9) 0 1 2! (4.10) (4.11) Then, in (4.5) and (4.9), we have L(x, t, y, s) = [−β(x4 + t2)− β 4 −1(E + R)](x,t)=(x,t)·(y,s)−1 (4.12) Now we are ready to obtain the following Lemma Lemma 4.1. We have det H(x, t, y, s) = F ((x, t) · (y, s)−1) where F (x, t) = ca,β(x4 + a2t2)m1 (x4 + t2)m2f (x, t) for some m1, m2, ca,β ∈ R and f (x, t) = 2(β + 1)x8 + (3(β + 2) − 2a2)x4t2 + (β + 2)a2t4. Proof. From (4.3) and (4.12), it is enought to show that det[−β(x4 + t2)− β 4 −1(E + R)] = F (x, t). Moreover, from the form of given F in the theorem, we only need to compute det(E + R). We recall the form E + R = B + 2K 0 2! − 0 1 (β + 4) x4 + t2 D · Dt. For notational convenience, we always use case-letter fi to denote the i′th row of matrix capital F . Notice that D · Dt is of rank 1 and will exploit the following convention det(P + Q) = det(P ) + det m Xj=1 (4.13) p1 ... pj−1 qj pj+1 ... pm     for any m × m matrix P and Q with rank Q = 1. Recall B = x2 + atJ and K = x · xt, then direct calculations show that det(B) = (x4 + a2t2)n (4.14) OPTIMAL BOUNDS FOR OSCILLATING CONVOLUTION OPERATORS ON THE HEISENBERG GROUP 9 and = n Xj=1 xj det n Xj=1 b1 ... bj−1 kj bj+1 ... b2n     + n Xj=1 xj+n det     (4.15) xj(x2xj + xn+j at)(x4 + a2t2)n−1 + xj+n (x2xj+n − xjat)(x4 + a2t2)n−1(4.16) b1 ... bj+n−1 kj+n bj+n+1 ... b2n n Xj=1 (4.17) (4.18) (4.19) (4.20) = (x4 + a2t2)n−1x4 Thus, from (4.4), (4.14) and (4.15), we have det(B + 2K) = (x4 + a2t2)n + 2x4(x4 + a2t2)n−1 = (x4 + a2t2)n−1(3x4 + a2t2) Using (4.4) again, we have det(E + R) = det(E) + 1 2 2n Xj=1 det From (4.18) =: S1 + S2 + S3 e1 ... ej−1 rj ej+1 ... e2n     + det  e1 ... e2n r2n+1   S1 = det B + 2K 0 2! = 0 1 1 2 det(B + 2K) = 1 2 (x4 + a2t2)n−1(3x4 + a2t2) Since rank K = 1, we see that det e1 ... ej−1 rj ej+1 ... e2n     = det b1 + 2k1 ... bj−1 + 2kj−1 −(β+4)x4 x4+t2 kj bj+1 + 2kj+1 ... b2n + 2k2n     = − (β + 4)x4 x4 + t2 det b1 ... bj−1 kj bj+1 ... b2n     . So, S2 = − 1 2(cid:18) (β + 4)x4 x4 + t2 (cid:19) x4(x4 + a2t2)n−1 10 Finally, WOOCHEOL CHOI S3 = det E + 2K ∗ 0 − β+4 x4+t2 4! = − t2 β + 4 x4 + t2 t2 4 = − β + 4 x4 + t2 t2 4 (x4 + a2t2)n−1(3x4 + a2t2) det(B + 2K) (4.21) Adding all these terms, we have det (∂i∂jΦ)(x, t) + 2a∂tΦ(x, t) J 0 0!! = p(x4 + a2t2)q(x4 + t2)f (x, t) 0 where p(r) = cprm1 , q(r) = rm2 for some m1, m2, cp ∈ R and f (x, t) = 2(β + 1)x8 + (3(β + 2) − 2a2)x4t2 + (β + 2)a2t4. Now, we should determine when the determinant of H(x, t, y, s) can be zero for some values (x, t, y, s) with ρ(cid:0)(x, t) · (y, s)−1(cid:1) ∼ 1. Furthermore, to determin the folding type in the degenerate case, it will be decisive to know the shape of factorization. We take it in the following theorem Lemma 4.2. For some nonzero constants γ, c, c1, c2, c3 with c1 6= c2 and c3 > 0 which are deter- mined by β and a, (cid:3) • case 1 : β ∈ (−1, 0) ∪ (0, ∞) · For a2 · For a2 · For a2 b2 < Cβ , f (x, t) > 0. b2 = Cβ , f (x, t) = γ(x2 − ct2)2. b2 > Cβ , f (x, t) = γ(x2 − c1t)(x2 + c1t)(x2 − c2t)(x2 + c2t). • case 2 : β ∈ (−2, −1) · f (x, t) = γ(x2 − c1t)(x2 + c1t)(x4 + c3t2). • case 3 : β ∈ (∞, −2) · f (x, t) < 0. Proof. Firstly, we see that f (x, t) > 0 for 3(β + 2) − 2a2 > 0. And f (x, t) > 0 also holds if the dicriminant ∆ = 4a4 − 4(β + 2)(2β + 5)a2 + 9(β + 2)2 is negative. This range is equal to C− β < a2 < C+ β where But, C± β = β + 2 2 (cid:16)2β + 5 ±p(2β + 5)2 − 9(cid:17) . C− β = (β + 2) 2 (2β + 5 −p(2β + 5)2 − 9) = < (β + 2) 2 (β + 2) 2 So, we compress two conditions as f (x, t) > 0 for a2 < C+ β . (2β + 5 −p(2β + 2)(2β + 8) (2β + 5 −p(2β + 2)2) = 2 3(β + 2) (4.22) (cid:3) OPTIMAL BOUNDS FOR OSCILLATING CONVOLUTION OPERATORS ON THE HEISENBERG GROUP 11 For degenerate cases, we need to analyze the canonical relation (3.1) associated with our phase function Φ and the projection maps (3.2) CΦ = {(cid:0)(x, t), Φ(x,t), (y, s), −Φ(y,s)(cid:1)} ⊂ T ∗(R2n+1) × T ∗(R2n+1) πL : CΦ → T ∗(R2n+1) and πR : CΦ → T ∗(R2n+1). We will check the condition (1) and (2) of Definition 3.2 to prove the following theorem Theorem 4.3. On the hypersurface S, (1) If β ∈ (−2, −1) or β ∈ (−1, 0) ∪ (0, ∞) and a2 b2 > Cβ, the projection maps πL and πR are both of folding type 1. (2) If β ∈ (−1, 0) ∪ (0, ∞) and a2 b2 = Cβ, the projection maps πL and πR are both of folding type 2. Let We need the following lemma to show that rank of πL and πR drop by 1 simply. S = {(x, t, y, s) : det H(x, t, y, s) = 0}. Lemma 4.4. Let L1(x, t, y, s) be the first (2n) × (2n)matrix of L(x, t, y, s) and suppose that (x, t, y, s) is contained in S. Then, provided β 6= −4. det L1(x, t, y, s) 6= 0 Proof. For simplicity, let (z, w) := (x, t) · (y, s)−1. Except the nonzero common facts, we only need to check the determinant of is nonzero for (z, w) 6= (0, 0). It can be calculated in the same way using (??) and (4.18) so that M (z, w) =(cid:16)z2I + awJ + 2z · zt − (β + 4) z4+w2 x · zt(cid:17) det(M (z, w)) = (z4 + a2w2)n + (z4 + a2w2)n−1z4(cid:18)2 − (β + 4) z4 = (z4 + a2w2)n−1 z4 + w2 z4 z4 + w2(cid:19) (cid:2)−(β + 1)z8 + (a2 + 3)z4w2 + a2w4(cid:3) (4.23) But, (z, w) is on S and satisfies 2(β + 1)z8 + (3(β + 2) − 2a2)z4w2 + (β + 2)a2w4 = 0 (4.24) From above two equalites, we have (3(β + 2) + 6)z4w2 + (β + 4)a2w4 = 0 So det(M (z, w)) = (z4 + a2w2)n−1 z4 + w2 w2 2 (β + 4)(cid:2)3z4 + a2w2(cid:3) If w = 0, then z becomes zero in (4.24). Because (z, w) 6= (0, 0), w should be nonzero. Thus det(M (z, w)) 6= 0. (cid:3) 12 WOOCHEOL CHOI proof of Theorem 4.3. We will only consider the case (1) in the theorem. The other case can be derived similrary. It is only remained to show that at the hypersurface S, we have the folding type conditions as in the theorem. Recall that S = {(x, t, y, s) ∈ R(2(2n+1) F(cid:0)(x, t) · (y, s)−1(cid:1) = F (cid:0)x − y, s − t + 2axtJy(cid:1) = 0, = {(x, t, y, s) ∈ R2(2n+1) f(cid:0)x − y, s − t + 2axtJy(cid:1) = 0, ρ(cid:0)(x, t) · (y, s)−1(cid:1) ∼ 1}. From Theorem 4.2, we have ρ(cid:0)(x, t) · (y, s)−1(cid:1) ∼ 1} (4.25) f (x, t) = γ(x2 − c1t)(x2 + c1t)(x2 − c2t)(x2 + c2t). We need to show that at each point P0 ∈ S, det(Df ) vanishes of 1 order in each null directions of dπL and dπR at P0. Let P0 = (x0, t0, y0, s0) and we may assume that P0 is contained in S1 =: {(x, t, y, s) ∈ R2(2n+1) x − y2 − c1(s − t + 2axtJy) = 0}. We may identifiy the manifold CΦ = {(cid:0)(x, t), Φ(x,t), (y, s), −Φ(y,s)(cid:1)} with an open set in R(2n+1) × R(2n+1) by the diffeomorphsim φ : R(2n+1) × R(2n+1) → S given by Let the null direction vL in R2(2n+1) of dπL at P0. It means that φ(x, t, y, s) =(cid:0)(x, t), Φ(x,t), (y, s), −Φ(y,s)(cid:1) . I ∂2Φ ∂(x,t)∂(x,t) 0 ∂2Φ ∂(y,s)∂(x,t)! vt = 0 Thus, v is of the form v = (0, 0, z, w) with w ∈ R2n, s ∈ R and (w, s) satisfies ∂2Φ ∂(y,s)∂(x,t) zt w! = 0 To verify det H(x, t, y, s) vanishes of order 1 in the direction vL, it suffice to show that vL is not orthogonal to the gradient vector vg of det H(x, t, y, s) at P0. From a direct calculation, we get the gradient vector vg as D(x,t),(y,s)Φ(cid:0)(x, t) · (y, s)−1(cid:1)(cid:12)(cid:12)p =(cid:0)2(x − y) − 2acβ,1aJy, −cβ,1, −2(x − y) − 2acβ,1xtJ, cβ,1(cid:1) Suppose that vL and vg are orthogonal. We are going to find a contradiction. It means that −2(x − y) · z − 2acβ,1xtJ · z + cβ,1w = 0 From (4.3), we have ( ∂2 ∂xi∂yj Φ) = A(y)"(∂i∂jΦ) − 2a∂tρ J 0 0!# A(x)t · zt w! 0 and A(x)t · zt w! = 0 . . . 0 0 0 1 0 0 · · · −2ax1 · · · · · · 0 0 0 · · · . . . · · · −2axn 1 0 0 ... 0 1   z1 z2 ... z2n w     1 0 0 0   2axn+1 (4.26) =(cid:16)z1, z2, · · · , z2n, 2a(xn+1z1 + · · · + x2nzn − x1zn+1 − · · · − xnz2n) + w(cid:17)t OPTIMAL BOUNDS FOR OSCILLATING CONVOLUTION OPERATORS ON THE HEISENBERG GROUP 13 On the other hand, from the orthogonal assumption, we have It means that so Observe that z ·(cid:0)−2(x − y) − 2cβ,1axtJ(cid:1) + w · cβ,1 = 0 2a(xn+1z1 + · · · − xnz2n) + w = 2(x − y) · z cβ,1 A(x)t · zt z2, · · · , z2n, 2(x−y)·z cβ,1 (cid:17)t w! =(cid:16)z1,   x4x2 1 ... x4xnx1 x2 t 2 x1 "(∂i∂jΦ) − 2a∂tρ J 0 0!# = (β + 4) 0 · · · . . . · · · · · · x4x1xn ... x4x2 n x2 t 2 xn t 2 t 2 x2x1 ... x2xn t2 4   − (x4 + t2) J 0 0 1 2! . Here, x → x − y and t → t − s + 2axtJy = x−y2 cβ,1 following should hold . So, if we calculate with the bottom row, the (β + 4)"x − y2 · 1 2 x − y2 cβ,1 Rearranging it, we obtain (x − y) · z + x − y4 c2 β,1 · 2 cβ,1 (x − y) · z# − 1 2 (x − y4 + x − y4 c2 β,1 ) 2 cβ,1 (x − y) · z = 0 " β + 2 2cβ,1 + 1 c3 β,1# x − y4 (x − y) · z = 0. And Thus we have (x − y) · z = 0. So A(x)t · z and H1(x, t, y, s) ·(cid:16)(z1, z2, · · · , z2n)(cid:17)t w! =(cid:16)z1, z2, · · · , z2n, 0(cid:17)t We can also have the same conclusion for dπR with the same argument. (cid:3) = 0. But this is a contradiction because det H1 6= 0. proof of Theorem 2.2 and Theorem 2.3. The theorems follow immediately when we use Theorem 3.3 based on the folding types established in Theorem 4.3. (cid:3) References [1] L. Hormander, Oscillatory integrals and multipliers on F Lp, Ark. Math. 11, (1973), 1-11. [2] Cuccagna, Scipio L2L2 estimates for averaging operators along curves with two-sided kk-fold singularities. Duke Math. J. 89 (1997), no. 2, 203-216. [3] C. Fefferman, Inequalities for strongly singular convolution operators, Acta Math., 124 (1970), pp. 9-36. [4] C. Fefferman and E. M. Stein, H p spaces of several variables, Acta Math., 129 (1972), pp. 137-193. [5] A. Greenleaf and A. Seeger, Oscillatory integral operators with low-order degeneracies. Duke Math. J. 112 (2002), no. 3, 397-420. [6] , Oscillatory and Fourier integral operators with degenerate canonical relations. Proceedings of the 6th International Conference on Harmonic Analysis and Partial Differential Equations (El Escorial, 2000). Publ. Mat. 2002, Vol. Extra, 93-141. [7] I. I. Hirschman, Multiplier Transforms I, Duke Math. J., 26 (1956), pp. 222242. [8] N. Laghi and N Lyall, Strongly singular integral operators associated to different quasi-norms on the Heisenberg group. Math. Res. Lett. 14 (2007), 825-838. 14 WOOCHEOL CHOI [9] , Strongly singular Radon transforms on the Heisenberg group and folding singularities Pacific J. Math. 233 (2007), 403-415. [10] N. Lyall, Strongly singular convolution operators on the Heisenberg group. Trans. Amer. Math. Soc. 359 (2007), 4467-4488 . [11] Y. Pan and C.D. Sogge, Oscillatory integrals associated to folding canonical relations, Colloq. Math. 60/61, (1990), 413-419. [12] Y. Pan and G. Sampson, The complete (Lp, Lp) mapping properties for a class of oscillatory integrals. J. Fourier Anal. Appl. 4 (1998), 93-103. [13] P. Sjolin, Convolution with oscillating kernels. Indiana Univ. Math. J. 30 (1981), no. 1, 47-55. [14] E. M. Stein, harmonic Analysis: Real-Variable Methods, Orthogonality, and Oscillatory integrals, Princeton Unive. Press, Princeton, 1993. [15] S. Wainger, Special trigonometric series in k-dimensions, Memoirs of the AMS 59, (1965), American Math. Society. School of Mathematical Sciences, Seoul National University, Seoul 151-747, Korea E-mail address: [email protected]
1501.03267	2	1501	2016-04-21T07:15:44	Operator Lipschitz functions on Banach spaces	[ "math.FA", "math.OA" ]	Let $X$, $Y$ be Banach spaces and let $\mathcal{L}(X,Y)$ be the space of bounded linear operators from $X$ to $Y$. We develop the theory of double operator integrals on $\mathcal{L}(X,Y)$ and apply this theory to obtain commutator estimates of the form $\\|f(B)S-Sf(A)\\|_{\mathcal{L}(X,Y)}\leq \textrm{const} \\|BS-SA\\|_{\mathcal{L}(X,Y)}$ for a large class of functions $f$, where $A\in\mathcal{L}(X)$, $B\in \mathcal{L}(Y)$ are scalar type operators and $S\in \mathcal{L}(X,Y)$. In particular, we establish this estimate for $f(t):=\|t\|$ and for diagonalizable operators on $X=\ell_{p}$ and $Y=\ell_{q}$, for $p<q$ and $p=q=1$, and for $X=Y=\mathrm{c}_{0}$. We also obtain results for $p\geq q$. We also study the estimate above in the setting of Banach ideals in $\mathcal{L}(X,Y)$. The commutator estimates we derive hold for diagonalizable matrices with a constant independent of the size of the matrix.	math.FA	math	OPERATOR LIPSCHITZ FUNCTIONS ON BANACH SPACES JAN ROZENDAAL, FEDOR SUKOCHEV, AND ANNA TOMSKOVA ABSTRACT. Let X, Y be Banach spaces and let L(X, Y) be the space of bounded linear operators from X to Y. We develop the theory of double operator integrals on L(X, Y) and apply this theory to obtain commutator estimates of the form k f (B)S − S f (A)kL(X,Y) ≤ const kBS − SAkL(X,Y) for a large class of functions f , where A ∈ L(X), B ∈ L(Y) are scalar type opera- tors and S ∈ L(X, Y). In particular, we establish this estimate for f (t) := t and for diagonalizable operators on X = ℓ p and Y = ℓq, for p < q and p = q = 1, and for X = Y = c0. We also obtain results for p ≥ q. We study the estimate above in the setting of Banach ideals in L(X, Y). The commutator estimates we derive hold for diagonalizable matrices with a constant independent of the size of the matrix. 1. INTRODUCTION Let X be a Banach space and let L(X) be the space of all bounded linear op- erators on X. Let A, B ∈ L(X) be scalar type operators (see Definition 3.1 below) : sp(A) ∪ sp(B) → C be a bounded Borel function, where sp(A) on X. Let f (resp. sp(B)) is the spectrum of the operator A (resp. B). We are interested in Lip- schitz type estimates (1.1) where k · kL(X) is the uniform operator norm on the space L(X), and more gener- ally in commutator estimates k f (B) − f (A)kL(X) ≤ const kB − AkL(X) , k f (B)S − S f (A)kL(X,Y) ≤ const kBS − SAkL(X,Y) (1.2) for Banach spaces X and Y, scalar type operators A ∈ L(X) and B ∈ L(Y), and S ∈ L(X, Y). This problem is well known in the special case where X = Y is a separable Hilbert space, such as ℓ2, and A and B are normal operators on X. In this paper we study such estimates in the Banach space setting, and specifically for X = ℓp and Y = ℓq with p, q ∈ [1, ∞]. space H the estimate In the special case where A, B are self-adjoint bounded operators on a Hilbert k f (B) − f (A)kL(H) ≤ const kB − AkL(H) (1.3) was established by Peller [24, 26] (see also [10]) for f : R → R in the Besov class B1 ∞,1(R) (for the definition of B1 ∞,1(R) see Section 3.3). This result extended a long line of results from [6]- [7], in which the theory of double operator integration Date: October 16, 2018. 2010 Mathematics Subject Classification. Primary 47A55, 47A56; secondary 47B47. The first-named author is supported by NWO-grant 613.000.908 "Applications of Transference Principles". 1 2 JAN ROZENDAAL, FEDOR SUKOCHEV, AND ANNA TOMSKOVA was developed to study the difference f (B) − f (A) (see also [8]). This theory was revised and extended in various directions, including the Banach space setting, in [11]. However, until now the results in the general setting were much weaker than in the Hilbert space setting. In this paper we show that for scalar type operators on Banach spaces one can obtain results matching those on Hilbert spaces. ∞,1(R). In Corollary 4.9 below we prove that (1.1) holds when A, B ∈ L(X) are scalar type operators with real spectrum and f ∈ B1 It is immediate from the definition of a scalar type operator that every normal operator on H is of scalar type. Therefore, Corollary 4.9 extends (1.3) to the Banach space setting. More generally, (1.2) holds for f ∈ B1 ∞,1(R) and for all S ∈ L(X, Y) (see Corollary 4.8.) ∞,1(R) and the results mentioned above do not apply. Moreover, the techniques which we used to obtain (1.1) for f ∈ B1 ∞,1(R) cannot be applied to the absolute value function (see Remark 8.4). However, the absolute value function is very important in the theory of matrix analysis and perturbation theory (see [5, Sections VII.5 and X.2]). If f is the absolute value function then f /∈ B1 In the case where H is an infinite-dimensional Hilbert space, it was proved by Kato [17] that the function t 7→ t, t ∈ R does not satisfy (1.3). An earlier example of McIntosh [22] showed the failure in general of the commutator estimate (1.2) for this function, in the case X = Y = H. Later, it was proved by Davies [9] that for 1 ≤ p ≤ ∞ and the Schatten von-Neumann ideal Sp with the corresponding norm k · kS p, the estimate kB − AkS p ≤ constkB − AkS p holds for all A, B ∈ Sp if and only if 1 < p < ∞. Commutator estimates for the absolute value function and different Banach ideals in L(H) have also been studied in [14]. The proofs in [9, 11, 14] are based on Macaev's celebrated theorem (see [16]) or on the UMD-property of the reflexive Schatten von-Neumann ideals. However, the spaces L(X, Y) are not UMD-spaces, and therefore the techniques used in [9, 11, 14] do not apply to them. To study (1.2) for X = ℓp and Y = ℓq, we use completely different methods from those of [9, 11, 14]. Instead, we use the theory of Schur multipliers on the space L(ℓp, ℓq) developed by Bennett [3], [4]. Let p, q ∈ [1, ∞] with p < q, p = q = 1 or p = q = ∞. In Section 6 we show (see Theorem 6.7) that, for diagonalizable operators (for the definition see Section 5) A ∈ L(ℓp) and B ∈ L(ℓq), and for the absolute value function f , k f (B)S − S f (A)kL(ℓp,ℓq) ≤ const kBS − SAkL(ℓp,ℓq) (1.4) holds for all S ∈ L(ℓp, ℓq) (where ℓ∞ should be replaced by c0). In particular, (1.5) k f (B) − f (A)kL(ℓ1) ≤ const kB − AkL(ℓ1) and (1.6) k f (B) − f (A)kL(c0) ≤ const kB − AkL(c0) for diagonalizable operators on ℓ1 respectively c0. Therefore we show that, even though (1.4) fails for p = q = 2, and in particular (1.1) fails for X = ℓ2 and f the absolute value function, one can obtain commutator estimates for p < q and Lipschitz estimates (1.1) for X = ℓ1 or X = c0. We also obtain results for p ≥ q. In particular, for p = q = 2 we prove (see Corollary 6.15) that for each ǫ ∈ (0, 1] there exists a constant C ≥ 0 such that OPERATOR LIPSCHITZ FUNCTIONS ON BANACH SPACES 3 the following holds. Let A, B ∈ L(ℓ2) be compact self-adjoint operators, and let U, V ∈ L(ℓ2) be unitaries such that λjPj U AU−1 = VBV−1 = ∞ ∑ j=1 µjPj, ∞ ∑ j=1 and j=1 and {µj}∞ where {λj}∞ j ∈ N, are the basis projections corresponding to the standard basis of ℓ2. Then (1.7) j=1 are sequences of real numbers and the Pj ∈ L(ℓ2), for kB − AkL(ℓ2) ≤ C min(cid:16)kV(B − A)U−1kL(ℓ2,ℓ2−ǫ), kV(B − A)U−1kL(ℓ2+ǫ,ℓ2)(cid:17), where we let the right-hand side equal infinity if V(B − A)U−1 /∈ L(ℓ2, ℓ2−ǫ) ∪ L(ℓ2+ǫ, ℓ2). We note that the constants which appear in our results depend on the spectral constants of A and B from Section 3.1, and in (1.4), (1.5) and (1.6) on the diagonal- izability constants of A and B from (5.5). These quantities are independent of the norms of A and B, and to obtain constants which do not depend on A and B in any way one merely has to restrict to operators with a sufficiently bounded spec- tral or diagonalizability constant. This is already done implicitly on Hilbert spaces by considering normal operators, for which the quantities involved are equal to 1. For example, in (1.7) the constant C does not depend on A or B in any way. Our results therefore truly extend the known estimates on Hilbert spaces, the main dif- ference between Hilbert spaces and general Banach spaces being that on Hilbert spaces one has a large and easily identifiable class of operators with spectral con- stant 1 or which are diagonalizable by an isometry. We study the commutator estimate in (1.2) in the more general form k f (B)S − S f (A)kI ≤ const kBS − SAkI , (1.8) where I is an operator ideal in L(X, Y). For example, we prove in Corollary 4.8 that (1.3) holds for a general Banach ideal I in L(X) with the strong convex com- pactness property (for definitions see Section 3.2), with respect to the norm k·kI . We also present (see Theorem 7.3) an example of a Banach ideal (I, k · kI ) in L(lp∗, ℓp), for 1 < p < ∞ and 1 p∗ = 1, namely the ideal of p-summing opera- tors, such that any Lipschitz function f (in particular, the absolute value function) satisfies (1.8). p + 1 In the final section we apply our results to finite-dimensional spaces, and obtain commutator estimates for diagonalizable matrices. Any diagonalizable matrix is a scalar type operator, hence estimates (1.4)-(1.8) hold for diagonalizable matrices A and B with a constant independent of the size of the matrix. 2. NOTATION AND TERMINOLOGY All vector spaces are over the complex number field. Throughout, X and Y de- note Banach spaces, the space of bounded linear operators from X to Y is L(X, Y), and L(X) := L(X, X). We identify the algebraic tensor product X∗⊗ Y with the space of finite rank operators in L(X, Y) via (x∗⊗ y)(x) := hx∗, xiy for x ∈ X, x∗ ∈ X∗ and y ∈ Y. The spectrum of A ∈ L(X) is sp(A), and by IX ∈ L(X) we denote the identity operator on X. Throughout the text we use the abbreviations SOT and WOT for the strong and weak operator topology, respectively. 4 JAN ROZENDAAL, FEDOR SUKOCHEV, AND ANNA TOMSKOVA The Borel σ-algebra on a Borel measurable subset σ ⊆ C will be denoted by Bσ, and B := BC. For measurable spaces (Ω1, Σ1) and (Ω2, Σ2) we denote by Σ1 ⊗ Σ2 the σ-algebra on Ω1 × Ω2 generated by all measurable rectangles σ1 × σ2 with σ1 ∈ Σ1 and σ2 ∈ Σ2. If (Ω, Σ) is a measurable space then B(Ω, Σ) is the space of all bounded Σ-measurable complex-valued functions on Ω, a Banach algebra with the supremum norm k fkB(Ω,Σ) := sup ω∈Ω f (ω) ( f ∈ B(Ω, Σ)). We simply write B(Ω) := B(Ω, Σ) and k fk∞ := k fkB(Ω,Σ) when little confusion can arise. If µ is a complex Borel measure on a measurable space (Ω, Σ) and X is a Banach space, then a function f : Ω → X is µ-measurable if there exists a sequence of X- valued simple functions converging to f µ-almost everywhere. For Banach spaces : Ω → L(X, Y), we say that f is strongly measurable if X and Y and a function f ω 7→ f (ω)x is a µ-measurable mapping Ω → Y for each x ∈ X. If µ is a positive measure on a measurable space (Ω, Σ) and f : Ω → [0, ∞] is a function, we let ZΩ f (ω) dµ(ω) := infZΩ g(ω) dµ(ω) ∈ [0, ∞], The H older conjugate of p ∈ [1, ∞] is denoted by p∗ and is defined by 1 where the infimum is taken over all measurable g : Ω → [0, ∞] such that g(ω) ≥ f (ω) for ω ∈ Ω. p∗ = 1. The indicator function of a subset σ of a set Ω is denoted by 1σ. We will often identify functions defined on σ with their extensions to Ω by setting them equal to zero off σ. p + 1 3. PRELIMINARIES 3.1. Scalar type operators. In this section we summarize some of the basics of scalar type operators, as taken from [15]. Let X be a Banach space. A spectral measure on X is a map E : B → L(X) such that the following hold: • E(∅) = 0 and E(C) = IX; • E(σ1 ∩ σ2) = E(σ1)E(σ2) for all σ1, σ2 ∈ B; • E(σ1 ∪ σ2) = E(σ1) + E(σ2) − E(σ1)E(σ2) for all σ1, σ2 ∈ B; • E is σ-additive in the strong operator topology. Note that these conditions imply that E is projection-valued. Moreover, by [15, Corollary XV.2.4] there exists a constant K such that (σ ∈ B). kE(σ)kL(X) ≤ K (3.1) An operator A ∈ L(X) is a spectral operator if there exists a spectral measure E on X such that AE(σ) = E(σ)A and sp(A, E(σ)X) ⊆ σ for all σ ∈ B, where sp(A, E(σ)X) denotes the spectrum of A in the space E(σ)X. For a spectral opera- tor A, we let ν(A) denote the minimal constant K occurring in (3.1) and call ν(A) the spectral constant of A. This is well-defined since the spectral measure E asso- ciated with A is unique, cf. [15, Corollary XV.3.8]. Moreover, E is supported on sp(A) in the sense that E(sp(A)) = IX [15, Corollary XV.3.5]. Hence we can define an integral with respect to E of bounded Borel measurable functions on sp(A), as This definition is independent of the representation of f , and follows. For f = ∑n Borel for 1 ≤ j ≤ n, we let (3.2) (cid:13)(cid:13)(cid:13)(cid:13)Zsp(A) f dE(cid:13)(cid:13)(cid:13)(cid:13)L(X) αjE(σj). = n ∑ j=1 f dE := Zsp(A) kxkX=kx∗kX∗ =1(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) sup j sup αj ≤ sup ≤ 4 k fkB(sp(A)) ≤ 4ν(A)k fkB(sp(A)) , n ∑ j=1 αjx∗E(σj)x(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) kxkX=kx∗kX∗ =1kx∗E(·)xkvar sup kxkX=kx∗kX∗ =1 sup σ⊆sp(A)x∗E(σ)x OPERATOR LIPSCHITZ FUNCTIONS ON BANACH SPACES 5 j=1 αj1σj a finite simple function with αj ∈ C and σj ⊆ sp(A) where kx∗E(·)xkvar is the variation norm of the measure x∗E(·)x. Since the simple functions lie dense in B(sp(A)), for general f ∈ B(sp(A)) we can define Zsp(A) f dE := lim n→∞Zsp(A) fn dE ∈ L(X) if { fn}∞ n=1 ⊆ B(sp(A)) is a sequence of simple functions with k fn − fk∞ → 0 as n → ∞. This definition is independent of the choice of approximating sequence and (3.3) It is straightforward to check that (cid:13)(cid:13)(cid:13)(cid:13)Zsp(A) f dE(cid:13)(cid:13)(cid:13)(cid:13)L(X) ≤ 4ν(A)k fkB(sp(A)) . Zsp(A) (α f + g) dE = αZsp(A) f dE +Zsp(A) f g dE =(cid:18)Zsp(A) f dE(cid:19)(cid:18)Zsp(A) g dE(cid:19) Zsp(A) g dE, for all α ∈ C and simple f , g ∈ B(sp(A)), and approximation extends these iden- tities to general f , g ∈ B(sp(A)). Moreover,Rsp(A) 1 dE = E(sp(A)) = IX. Hence the map f 7→ Rsp(A) f dE is a continuous morphism B(sp(A)) → L(X) of unital is bounded on sp(A) andRsp(A) λ dE(λ) ∈ L(X) is well defined. Banach algebras. Since the spectrum of A is compact, the identity function λ 7→ λ Definition 3.1. A spectral operator A ∈ L(X) with spectral measure E is a scalar type operator if A =Zsp(A) λ dE(λ). The class of scalar type operators on X is denoted by Ls(X). For A ∈ Ls(X) with spectral measure E and f ∈ B(sp(A)) we define (3.4) f (A) :=Zsp(A) f dE. 6 JAN ROZENDAAL, FEDOR SUKOCHEV, AND ANNA TOMSKOVA Then, as remarked above, f 7→ f (A) is a continuous morphism B(sp(A)) → L(X) of unital Banach algebras with norm bounded by 4ν(A). Note also that (3.5) hx∗, f (A)xi =Zsp(A) f (λ) dhx∗, E(λ)xi for all f ∈ B(sp(A)), x ∈ X and x∗ ∈ X∗. Indeed, for simple functions this follows from (3.2), and by taking limits one obtains (3.5) for general f ∈ B(sp(A)). Finally, we note that a normal operator A on a Hilbert space H is a scalar type operator with ν(A) = 1, and in this case (3.3) improves to (3.6) (cid:13)(cid:13)(cid:13)(cid:13)Zsp(A) f dE(cid:13)(cid:13)(cid:13)(cid:13)L(H) ≤ k fkB(sp(A)) , as known from the Borel functional calculus for normal operators. 3.2. Spaces of operators. In this section we discuss some properties of spaces of operators that we will need later on. < ǫ. First we provide a lemma about approximation by finite rank operators. Recall that a Banach space X has the bounded approximation property if there exists M ≥ 1 such that, for each K ⊆ X compact and ǫ > 0, there exists S ∈ X∗⊗ X with kSkL(X) ≤ M and supx∈K kSx − xkX Lemma 3.2. Let X and Y be Banach spaces such that X is separable and either X or Y has the bounded approximation property. Then each T ∈ L(X, Y) is the SOT-limit of a norm bounded sequence of finite rank operators. Proof. Fix T ∈ L(X, Y). By [21, Proposition 1.e.14] there exists a norm bounded net(cid:8)Tj(cid:9)j∈J ⊆ X∗⊗ Y having T as its SOT-limit. It is straightfroward to see that the strong operator topology is metrizable on bounded subsets of L(X, Y) by d(S1, S2) := 2−k kS1xk − S2xkkY (S1, S2 ∈ L(X, Y)), ∞ ∑ k=1 (cid:3) where {xk}k∈N ⊆ X is a countable subset that is dense in the unit ball of X. Hence there exists a subsequence of(cid:8)Tj(cid:9)j∈J with T as its SOT-limit. Let X and Y be Banach spaces and let Z be a Banach space which is continu- ously embedded in L(X, Y). Following [36] (in the case where Z is a subspace of L(X, Y)), we say that Z has the strong convex compactness property if the follow- ing holds. For any finite measure space (Ω, Σ, µ) and any strongly measurable bounded f : Ω → Z, the operator T ∈ L(X, Y) defined by (x ∈ X), (3.7) f (ω)x dµ(ω) belongs to Z with kTkZ ≤ RΩk f (ω)kZ dµ(ω). By the Pettis Measurability Theo- rem, any separable Z has this property. Indeed, if Z is separable then combining Propositions 1.9 and 1.10 in [35] shows that any strongly measurable f : Ω → Z is µ-measurable as a map to Z. If f is bounded as well, then (3.7) defines an element of Z with Tx :=ZΩ kTkZ ≤ZΩ k f (ω)kZ dµ(ω). OPERATOR LIPSCHITZ FUNCTIONS ON BANACH SPACES 7 It is shown in [36] and [31] that the compact and weakly compact operators have the strong convex compactness property, but not all subspaces of L(X, Y) do. Moreover, if N is a semifinite von Neumann algebra on a separable Hilbert space H, with faithful normal semifinite trace τ, and F is a rearrangement invariant Ba- nach function space with the Fatou property, then E = N ∩ F (N, τ) has the strong convex compactness property [2, Lemma 3.5]. Now let (Ω, µ) be a finite measure space and let f Lemma 3.3. Let X and Y be separable Banach spaces and Z a Banach space continuously embedded in L(X, Y). If BZ := {z ∈ Z kzkZ ≤ 1} is SOT-closed in L(X, Y), then Z has the strong convex compactness property. Proof. The proof follows that of [2, Lemma 3.5]. First we show that BZ is a Polish space in the strong operator topology. As in the proof of Lemma 3.2, bounded subsets of L(X, Y) are SOT-metrizable. The finite rank operators are SOT-dense in L(X, Y), hence L(X, Y) is SOT-separable. Therefore BZ is SOT-separable and metrizable. By assumption, BZ is complete. : Ω → Z be bounded and strongly measurable. Without loss of generality, we may assume that f (Ω) ⊆ BZ and that µ is a probability measure. For each y∗ ∈ Y∗ and x ∈ X, the mapping BZ → [0, ∞), T 7→ hy∗, Txi is continuous. The collection of all these mappings, for y∗ ∈ Y∗ and x ∈ X, separates the points of BZ. Moreover, ω 7→ hy∗, f (ω)xi is a measurable mapping Ω → [0, ∞) for each y∗ ∈ Y∗ and x ∈ X. By [35, Propo- sitions 1.9 and 1.10], f is the µ-almost everywhere SOT-limit of a sequence of BZ- k=1. Let Tn := RΩ fn dµ ∈ BZ. By the dominated valued simple functions { fk}∞ convergence theorem, Tn(x) → T(x) :=RΩ f (ω)x dµ(ω) as n → ∞, for all x ∈ X. By assumption, T ∈ BZ. Now let g : Ω → [0, ∞) be measurable such that 1 ≥ g(ω) ≥ k f (ω)kZ for ω ∈ Ω, and define h(ω) := f (ω) RΩ g(η) dµ(η)dµ(ω) for ω ∈ Ω. By what we have shown above, x 7→ RΩ h(ω)x dν(ω) defines an element of BZ. Since Tx =RΩ f (ω)x dµ(ω) =RΩ g(ω) dµ(ω)RΩ h(ω)x dν(ω), we obtain kTkZ ≤ RΩ g(ω) dµ(ω), as remained to be shown. Remark 3.4. Note that the converse implication does not hold. Indeed, if X is a Hilbert space (or more generally, a Banach space with the metric approximation property) then the finite rank operators of norm less than or equal to 1 are SOT- dense in the unit ball of L(X). Therefore the compact operators of norm less than or equal to 1 are not SOT-closed in L(X) if X is infinite-dimensional. However, by [36, Theorem 1.3], the space of compact operators on X has the strong convex compactness property. (cid:3) g(ω) and dν(ω) := g(ω) embedded in L(X, Y). We say that (I, k · kI ) is a Banach ideal in L(X, Y) if Let X and Y be Banach spaces and I a Banach space which is continuously • For all R ∈ L(Y), S ∈ I and T ∈ L(X), RST ∈ I with kRSTkI ≤ • X∗⊗ Y ⊆ I with kx∗⊗ ykI = kx∗kX∗kykY for all x∗ ∈ X∗ and y ∈ Y. kRkL(Y) kSkI kTkL(X); By Lemma 3.3 and [12, Proposition 17.21] (using that the SOT and WOT closures of a convex set coincide), for separable X and Y, any maximal Banach ideal (for the definition see e.g. [27]) in L(X, Y) has the strong convex compactness property. This includes a large class of operator ideals, such as the ideal of absolutely p- summing operators, the ideal of integral operators, etc. [12, p. 203]. 8 JAN ROZENDAAL, FEDOR SUKOCHEV, AND ANNA TOMSKOVA 3.3. Algebras of functions. In this section we discuss some algebras of functions that will be essential in later sections. Let σ1, σ2 ⊆ C be Borel measurable subsets and let A(σ1 × σ2) be the class of Borel functions ϕ : σ1 × σ2 → C such that (3.8) ϕ(λ1, λ2) =ZΩ a1(λ1, ω)a2(λ2, ω) dµ(ω) for all (λ1, λ2) ∈ σ1 × σ2, where (Ω, Σ, µ) is a finite measure space and a1 ∈ B(σ1 × Ω, Bσ1 ⊗ Σ), a2 ∈ B(σ2 × Ω, Bσ2 ⊗ Σ). For ϕ ∈ A(σ1 × σ2) let kϕkA(σ1×σ2) := infZΩ ka1(·, ω)kB(σ1) ka2(·, ω)kB(σ2) dµ(ω), where the infimum runs over all possible representations1 in (3.8) (it is straightfor- ward to show that the map ω 7→ ka1(·, ω)kB(σ1) ka2(·, ω)kB(σ2) is measurable). Lemma 3.5. For all σ1, σ2 ⊆ C measurable, A(σ1 × σ2) is a unital Banach algebra which is contractively included in B(σ1 × σ2). Proof. That A(σ1 × σ2) is a vector space is straightforward, and that it is normed algebra is shown in [28, Lemma 3] for σ1 = σ2 = R (the proof in our setting is identical). The completeness of A(σ1 × σ2) follows by showing that an absolutely convergent series of elements in A(σ1 × σ2) converges in A(σ1 × σ2). This is done by considering a direct sum of the measure spaces involved. (cid:3) We now state sufficient conditions for a function to belong to A. The first will be used in the proof of Proposition 5.6. Let W1,2(R) be the space of all g ∈ L2(R) with weak derivative g′ ∈ L2(R), endowed with the norm kgkW1,2(R) := kgkL2(R) + kg′kL2(R) for g ∈ W1,2(R). Lemma 3.6. (3.9) [28, Theorem 9] Let g ∈ W1,2(R) and let ψg(λ1, λ2) :=( g(log( λ1 λ2 0 )) if λ1, λ2 > 0 otherwise . Then ψg ∈ A(R2) with(cid:13)(cid:13)ψg(cid:13)(cid:13)A(R2) ≤ √2kgkW1,2(R). The second condition involves the Besov space B1 ∞,1(R). Following [25], let {ψk}k∈Z be a sequence of Schwartz functions on R such that, for each k ∈ Z, the Fourier transform F ψk of ψk is supported on [2k−1, 2k+1] and F ψk+1(x) = F ψk(2x) for all x > 0, and such that ∑∞ k=−∞ F ψk(x) = 1 for all x > 0. Let ψ∗k be defined by F ψ∗k (x) = F ψk(−x) for k ∈ Z and x ∈ R. If f is a distribution on R such that {2k k f ∗ ψkkL∞(R)}k∈Z ∈ ℓ1(Z) and {2k(cid:13)(cid:13) f ∗ ψ∗k(cid:13)(cid:13)L∞(R)}k∈Z ∈ ℓ1(Z), then f admits a representation (3.10) f = ∑ k∈Z f ∗ ψk + ∑ k∈Z f ∗ ψ∗k + P, where P is a polynomial. 1This might seem problematic from a set-theoretic viewpoint. The problem can be fixed by choosing an equivalence class of such a representation for each r ∈ R which can occur in the infimum. OPERATOR LIPSCHITZ FUNCTIONS ON BANACH SPACES 9 We let the homogeneous Besov space B1 ∞,1(R) consist of all distributions as above for which P = 0. Then B1 ∞,1(R) is a Banach space when equipped with the norm ∞,1(R) := k fk B1 For f ∈ B1 ∞ ∑ k=−∞ 2k k f ∗ ψkkL∞(R) + ∞,1(R) define ψ f : C2 → C by ψ f (λ1, λ2) :=( f (λ2)− f (λ1) f ′(λ1) λ2−λ1 ∞ ∑ k=−∞ 2k k f ∗ ψ∗kkL∞(R) ( f ∈ B1 ∞,1(R)). if (λ1, λ2) ∈ R2 and λ1 6= λ2 if λ1 = λ2 ∈ R . Lemma 3.7. There exists a constant C ≥ 0 such that ψ f ∈ A(R2) for each f ∈ B1 ∞,1(R), with(cid:13)(cid:13)(cid:13)ψ f(cid:13)(cid:13)(cid:13)A(R2) ≤ C k fk B1 Proof. Let f ∈ B1 has a representation ∞,1(R). ∞,1(R). In [25, Theorem 2] (see also [26, p. 535]) it is shown that ψ f ψ f (λ1, λ2) =ZΩ a1(λ1, ω)a2(λ2, ω) dµ(ω) for (λ1, λ2) ∈ R2, where (Ω, µ) is a measure space and a1 and a2 are measurable functions on R × Ω such that ZΩ ka1(·, ω)k∞ ka2(·, ω)k∞ dµ(ω) ≤ C k fk B1 ∞,1(R) for some constant C ≥ 0 independent of f . The desired conclusion now follows by replacing a1(λ1, ω) and a2(λ2, ω) by a1(λ1,ω) , and dµ(ω) ka1(·,ω)k∞ by ka1(·, ω)k∞ ka2(·, ω)k∞ dµ(ω). (cid:3) respectively a2(λ2,ω) ka2(·,ω)k∞ In [25, Theorem 3] Peller also states a condition on a function f on R which is necessary in order for ϕ f ∈ A(R2) to hold, and this condition is only slightly weaker than f ∈ B1 ∞,1(R). 4. DOUBLE OPERATOR INTEGRALS AND LIPSCHITZ ESTIMATES 4.1. Double operator integrals. Fix Banach spaces X and Y, scalar type opera- tors A ∈ Ls(X) and B ∈ Ls(Y) with spectral measures E respectively F, and ϕ ∈ A(sp(A) × sp(B)). Let a representation (3.8) for ϕ be given, with corre- sponding (Ω, µ) and a1 ∈ B(sp(A) × Ω), a2 ∈ B(sp(B) × Ω). For ω ∈ Ω, let a1(A, ω) := a1(·, ω)(A) ∈ L(X) and a2(B, ω) := a2(·, ω)(B) ∈ L(Y) be defined by the functional calculus for A respectively B from Section 3.1. Lemma 4.1. Let S ∈ L(X, Y) have separable range. Then, for each x ∈ X, ω 7→ a2(B, ω)Sa1(A, ω)x is a weakly measurable map Ω → Y. Proof. Fix x ∈ X. If a1 = 1σ for some σ ⊆ sp(A) × Ω then it is straightforward to show that hx∗, a1(A, ·)xi is measurable for each x∗ ∈ X∗. As S has separable range, Sa1(A, ·)x is µ-measurable. If a2 is an indicator function as well, the same argument shows that a2(B, ·)y is weakly measurable for each y ∈ Y. General argu- ments, approximating Sa1(A, ·)x by simple functions, show that a2(B, ·)Sa1(A, ·)x is weakly measurable. By linearity this extends to simple a1 and a2, and for gen- eral a1 and a2 let { fk}k∈N, {gk}k∈N be sequences of simple functions such that 10 JAN ROZENDAAL, FEDOR SUKOCHEV, AND ANNA TOMSKOVA a1 = limk→∞ fk and a2 = limk→∞ gk uniformly. Then a1(A, ω) = limk→∞ fk(A) and a2(B, ω) = limk→∞ gk(B) in the operator norm, for each ω ∈ Ω. The desired measurability now follows. (cid:3) Now suppose that Y is separable, that I is a Banach ideal in L(X, Y) and let S ∈ L(X, Y). By (3.3), (4.1) ka2(B, ω)Sa1(A, ω)kI ≤ 16 ν(A)ν(B)kSkI ka1(·, ω)kB(sp(A)) ka2(·, ω)kB(sp(B)) for w ∈ Ω. Since I is continuously embedded in L(X, Y), by the Pettis Measura- bility Theorem, Lemma 4.1 and (4.1) we can define the double operator integral (4.2) T A,B ϕ (S)x :=ZΩ a2(B, ω)Sa1(A, ω)x dµ(ω) ∈ Y (x ∈ X). Throughout, we will use Tϕ for T A,B the context. ϕ when the operators A and B are clear from Proposition 4.2. Let X and Y be separable Banach spaces such that X or Y has the bounded approximation property, and let A ∈ Ls(X), B ∈ Ls(Y), and ϕ ∈ A(sp(A) × sp(B)). Let I be a Banach ideal in L(X, Y) with the strong convex compactness prop- erty. Then (4.2) defines an operator T A,B ϕ ∈ L(I) which is independent of the choice of representation of ϕ in (3.8), with (4.3) (cid:13)(cid:13)(cid:13)T A,B ϕ (cid:13)(cid:13)(cid:13)L(I ) ≤ 16 ν(A)ν(B)kϕkA(sp(A)×sp(B)) . Proof. By (4.1) and the strong convex compactness property, Tϕ(S) ∈ L(I) for all S ∈ I, with (cid:13)(cid:13)Tϕ(S)(cid:13)(cid:13)I ≤ 16 ν(A)ν(B)kSkIZΩ ka1(·, ω)kB(sp(A)) ka2(·, ω)kB(sp(B)) dµ(ω). Clearly Tϕ is linear, hence the result follows if we establish that Tϕ is independent of the representation of ϕ. For this it suffices to let ϕ ≡ 0. Now, first consider S = x∗⊗ y for x∗ ∈ X∗ and y ∈ Y, and let x ∈ X, y∗ ∈ Y∗ and w ∈ Ω. Recall that E and F are the spectral measures of A and B, respectively. Then hy∗, a2(B, ω)Sa1(A, ω)xi =Zsp(B) =Zsp(B) =Zsp(B)Zsp(A) a1(λ, ω)a2(η, ω)dhx∗, E(λ)xidhy∗, F(η)yi by (3.5). Now Fubini's Theorem and the assumption on ϕ yield Tϕ(S) = 0. By linearity, Tϕ(S) = 0 for all S ∈ X∗ ⊗ Y. By Lemma 3.2, a general S ∈ I is the SOT-limit of a bounded (in L(X, Y)) sequence {Sn}n∈N ⊆ X∗ ⊗ Y. The dominated convergence theorem shows that Tϕ(S)x = limn→∞ Tϕ(Sn)x = 0 for all x ∈ X, which implies that Tϕ is independent of the representation of ϕ and concludes the proof. (cid:3) a2(η, ω) dhy∗, F(η)Sa1(A, ω)xi a2(η, ω)hx∗, a1(A, ω)xi dhy∗, F(η)yi OPERATOR LIPSCHITZ FUNCTIONS ON BANACH SPACES 11 Note that, if A and B are normal operators on separable Hilbert spaces X and Y, then (4.3) improves to (4.4) (cid:13)(cid:13)(cid:13)T A,B ϕ (cid:13)(cid:13)(cid:13)L(I ) ≤ kϕkA(sp(A)×sp(B)) , ϕ by appealing to (3.6) instead of (3.3) in (4.1). Remark 4.3. Let H be an infinite-dimensional separable Hilbert space and S2 the ideal of Hilbert-Schmidt operators on H. There is a natural definition (see [8]) of a double operator integral T A,B ∈ L(S2) for all ϕ ∈ B(C2) and normal op- erators A, B ∈ L(H), such that T A,B if ϕ ∈ A(sp(A) × sp(B)). One could wonder whether Proposition 4.2 can be extended to a larger class of func- tions than A(sp(A) × sp(B)) using an extension of the definition of T A,B in (4.2) which coincides with T A,B on S2. But it follows from [24, Theorem 1] that T A,B extends to a bounded operator on I = L(H) if and only if ϕ ∈ A(sp(A) × sp(B)). Hence Proposition 4.2 cannot be extended to a larger function class than A(sp(A) × sp(B)) in general. However, for specific Banach ideals, e.g. ideals with the UMD property, results have been obtained for larger classes of func- tions [11, 29]. = T A,B ϕ ϕ ϕ ϕ ϕ Remark 4.4. The assumption in Proposition 4.2 that X or Y has the bounded ap- proximation property is only used, via Lemma 3.2, to ensure that each S ∈ I is the SOT-limit of a bounded (in L(X, Y)) net of finite-rank operators. Clearly this is true for general Banach spaces X and Y if I is the closure in L(X, Y) of X∗⊗ Y. In [20] the authors consider an assumption on X and I, called condition c∗λ, which guarantees that each S ∈ I is the SOT-limit of a bounded net of finite-rank oper- ators. It is shown in [20] that this condition is strictly weaker than the bounded approximation property, for certain non-trivial ideals. In the results throughout the paper where we assume that X has the bounded approximation property, one may assume instead that X has satisfies condition c∗λ for I for some λ ≥ 1. 4.2. Commutator and Lipschitz estimates. Let p1 :, p2 : C2 → C be the coordinate projections p1(λ1, λ2) := λ1, p2(λ1, λ2) := λ2 for (λ1, λ2) ∈ C2. Note that f ◦ p1, f ◦ p2 ∈ A(σ1 × σ2) for all σ1, σ2 ⊆ C Borel and f ∈ B(σ1 ∪ σ2). For A and B selfadjoint operators on a Hilbert space and I a non-commutative Lp-space, the following lemma is [28, Lemma 3]. Lemma 4.5. Under the assumptions of Proposition 4.2, the following hold: ϕ algebras. (1) The map ϕ 7→ T A,B is a morphism A(sp(A) × sp(B)) → L(I) of unital Banach (2) Let f ∈ B(sp(A) ∪ sp(B)) and S ∈ L(X, Y). Then Tf ◦p1(S) = S f (A) and Tf ◦p2(S) = f (B)S. In particular, Tp1(S) = SA and Tp2(S) = BS. Proof. The structure of the proof is the same as that of [28, Lemma 3]. Linearity in (1) is straightforward. Fix ϕ1, ϕ2 ∈ A(sp(A) × sp(B)) with representations as in (3.8), with corresponding measure spaces (Ωj, µj) and bounded Borel functions a1,j ∈ B(sp(A) × Ωj) and a2,j ∈ B(sp(B) × Ωj) for j ∈ {1, 2}. Then ϕ := ϕ1 ϕ2 also has a representation as in (3.8), with Ω = Ω1 × Ω2, µ = µ1 × µ2 the product measure and a1 = a1,1a1,2, a2 = a2,1a2,2. By multiplicativity of the functional 12 JAN ROZENDAAL, FEDOR SUKOCHEV, AND ANNA TOMSKOVA calculus for A, a1(A, (ω1, ω2)) =(cid:0)a1,1(·, ω1)a1,2(·, ω2)(cid:1)(A) = a1,1(A, ω1)a1,2(A, ω2) for all (ω1, ω2) ∈ Ω, and similarly for a2(B, (ω1, ω2)). Applying this to (4.2) yields Tϕ(S)x =ZΩ =ZΩ1 = Tϕ1 (Tϕ2(S))x a2(B, ω)Sa1(A, ω)x dµ(ω) a2,1(B, ω1)Tϕ2 (S)a1,1(A, ω1)x dµ1(ω1) for all S ∈ I and x ∈ X, which proves (1). Part (2) follows from (4.2) and the fact that Tϕ is independent of the representation of ϕ. (cid:3) For f : sp(A) ∪ sp(B) → C define (4.5) ϕ f (λ1, λ2) := f (λ2) − f (λ1) λ2 − λ1 for (λ1, λ2) ∈ sp(A) × sp(B) with λ1 6= λ2. Theorem 4.6. Let X and Y be separable Banach spaces such that X or Y has the bounded approximation property, and let I be a Banach ideal in L(X, Y) with the strong convex compactness property. Let A ∈ Ls(X) and B ∈ Ls(Y), and let f ∈ B(sp(A) ∪ sp(B)) be such that ϕ f extends to an element of A(sp(A) × sp(B)). Then (4.6) for all S ∈ L(X, Y) such that BS − SA ∈ I. In particular, if X = Y and B − A ∈ I, k f (B)S − S f (A)kI ≤ 16 ν(A)ν(B)(cid:13)(cid:13)ϕ f(cid:13)(cid:13)A(sp(A)×sp(B)) kBS − SAkI k f (B) − f (A)kI ≤ 16 ν(A)ν(B)(cid:13)(cid:13)ϕ f(cid:13)(cid:13)A(sp(A)×sp(B)) kB − AkI . Proof. Note that (p2 − p1)ϕ f = f ◦ p2 − f ◦ p1. By Lemma 4.5, f (B)S − S f (A) = Tf ◦p2(S) − Tf ◦p1(S) = T(p2−p1)ϕ f (S) = Tp2 ϕ f (S) − Tp1 ϕ f (S) = Tϕ f (Tp2(S) − Tp1 (S)) = Tϕ f (BS − SA) for each S ∈ I. Proposition 4.2 now concludes the proof. (cid:3) Letting X and Y be Hilbert spaces and A and B normal operators, we gener- alize results from [8, 28] to all Banach ideals with the strong convex compactness property. As mentioned in Section 3.2, this includes all separable ideals and the so-called maximal operator ideals, which in turn is a large class of ideals contain- ing the absolutely (p, q)-summing operators, the integral operators, and more [12, p. 203]. Note that, for normal operators, we can improve the estimate in (4.6) by appealing to (4.4) instead of (4.3). Corollary 4.7. Let A ∈ L(X) and B ∈ L(Y) be normal operators on separable Hilbert spaces X and Y. Let I be a Banach ideal in L(X, Y) with the strong convex compact- ness property, and let f ∈ B(sp(A) ∪ sp(B)) be such that ϕ f extends to an element of A(sp(A) × sp(B)). Then k f (B)S − S f (A)kI ≤(cid:13)(cid:13)ϕ f(cid:13)(cid:13)A(sp(A)×sp(B)) kBS − SAkI OPERATOR LIPSCHITZ FUNCTIONS ON BANACH SPACES 13 for all S ∈ L(X, Y) such that BS − SA ∈ I. In particular, if X = Y and B − A ∈ I, k f (B) − f (A)kI ≤(cid:13)(cid:13)ϕ f(cid:13)(cid:13)A(sp(A)×sp(B)) kB − AkI . Combining Theorem 4.6 with Lemma 3.7 yields the following, a generalization of [25, Theorem 4]. Corollary 4.8. There exists a universal constant C ≥ 0 such that the following holds. Let X and Y be separable Banach spaces such that X or Y has the bounded approximation property, and let I be a Banach ideal in L(X, Y) with the strong convex compactness property. Let f ∈ B1 ∞,1(R), and let A ∈ Ls(X) and B ∈ Ls(Y) be such that sp(A) ∪ sp(B) ⊆ R. Then (4.7) k f (B)S − S f (A)kI ≤ Cν(A)ν(B)k fk B1 ∞,1(R)kBS − SAkI for all S ∈ L(X, Y) such that BS − SA ∈ I. In particular, if X = Y and B − A ∈ I, k f (B) − f (A)kI ≤ Cν(A)ν(B)k fk B1 ∞,1(R)kB − AkI . In the case where the Banach ideal I is the space L(X, Y) of all bounded opera- tors from X to Y, we obtain the following corollary. Corollary 4.9. There exists a universal constant C ≥ 0 such that the following holds. Let X and Y be separable Banach spaces such that either X or Y has the bounded approximation property. Let f ∈ B1 ∞,1(R), and let A, B ∈ Ls(X) be such that sp(A) ∪ sp(B) ⊆ R. Then (4.8) k f (B)S − S f (A)kL(X,Y) ≤ Cν(A)ν(B)k fk B1 ∞,1(R)kBS − SAkL(X,Y) for all S ∈ L(X, Y). In particular, if X = Y then k f (B) − f (A)kL(X) ≤ Cν(A)ν(B)k fk B1 ∞,1(R)kB − AkL(X) . Remark 4.10. Corollaries 4.8 and 4.9 yield global estimates, in the sense that (4.7) and (4.8) hold for all scalar type operators A and B with real spectrum, and the constant in the estimate depends on A and B only through their spectral constants ν(A) and ν(B). Local estimates follow if f ∈ B(R) is contained in the Besov class locally. More precisely, given scalar type operators A ∈ Ls(X) and B ∈ Ls(Y) with real spectrum, suppose there exists g ∈ B1 ∞,1(R) with g(s) = f (s) for all s ∈ sp(A) ∪ sp(B). Then (with notation as in Corollary 4.8) (4.9) k f (B)S − S f (A)kI ≤ Cν(A)ν(B) kgk B1 ∞,1(R)kBS − SAkI for all S ∈ L(X, Y) such that BS − SA ∈ I. This follows directly from Theorem 4.6. 5. SPACES WITH AN UNCONDITIONAL BASIS In this section we prove some results for specific scalar type operators, namely operators which are diagonalizable with respect to an unconditional Schauder ba- sis. These results will be used in later sections. In this section we assume all spaces to be infinite-dimensional, but the results and proofs carry over directly to finite- dimensional spaces. This will be used in Section 8.1. 14 JAN ROZENDAAL, FEDOR SUKOCHEV, AND ANNA TOMSKOVA 5.1. Diagonalizable operators. Let X be a Banach space with an unconditional Schauder basis {ej}∞ j=1 ⊆ X. For j ∈ N, let Pj ∈ L(X) be the projection given by Pj(x) := xjej for all x = ∑∞ Assumption 5.1. For simplicity, we assume in this section that(cid:13)(cid:13)(cid:13)∑j∈N Pj(cid:13)(cid:13)(cid:13)L(X) = 1 for all non-empty N ⊆ N. This condition is satisfied in the examples we consider in later sections, and simplifies the presentation. For general bases one merely gets additional constants in the results. k=1 xkek ∈ X. An operator A ∈ L(X) is diagonalizable (with respect to {ej}∞ j=1) if there exists j=1 ∈ ℓ∞ of complex numbers such that U ∈ L(X) invertible and a sequence {λj}∞ λjPjx U AU−1x = (5.1) ∞ ∑ j=1 (x ∈ X), where the series converges since {ek}∞ this case A is a scalar type operator, with point spectrum equal to {λj}∞ (cid:8)λj(cid:9)∞ k=1 is unconditional, cf. [32, Lemma 16.1]. In j=1, sp(A) = j=1 and spectral measure E given by E(σ) = ∑ λj∈σ U−1PjU (5.2) for σ ⊆ C Borel. The set of all diagonalizable operators on X is denoted by Ld(X). We do not explicitly mention the basis {ej}∞ j=1 with respect to which an operator is diagonalizable, since this basis will be fixed throughout. Often we write A ∈ j=1, U) in order to identify the operator U and the sequence {λj}∞ Ld(X, {λj}∞ from above. For A ∈ Ld(X, {λj}∞ j=1, U) and f ∈ B(C), it follows from (3.4) that j=1 (5.3) f (A) = U−1(cid:16) ∞ ∑ j=1 f (λj)Pj(cid:17)U. Since any Banach space with a Schauder basis is separable and has the bounded approximation property, we can apply the results from the previous section to diagonalizable operators, and we obtain for instance the following. Corollary 5.2. There exists a universal constant C ≥ 0 such that the following holds. Let X and Y be Banach spaces with unconditional Schauder bases, and let I be a Banach ideal in L(X, Y) with the strong convex compactness property. Let f ∈ B1 ∞,1(R), and let A ∈ Ld(X) and B ∈ Ld(Y) be such that sp(A) ∪ sp(B) ⊆ R. Then k f (B)S − S f (A)kI ≤ Cν(A)ν(B)k fk B1 ∞,1(R)kBS − SAkI for all S ∈ L(X, Y) such that BS − SA ∈ I. In particular, if X = Y and B − A ∈ I, k f (B) − f (A)kI ≤ Cν(A)ν(B)k fk B1 ∞,1(R)kB − AkI . Since this result does not apply to the absolute value function (and neither does the more general Theorem 4.6), and because of the importance of the absolute value function, we now study Lipschitz estimates for more general functions. Let Y be a Banach space with an unconditional Schauder basis { fk}∞ let the projections Qk ∈ L(Y) be given by Qk(y) := yk fk for all y = ∑∞ k=1 ⊆ Y, and l=1 yl fl ∈ Y OPERATOR LIPSCHITZ FUNCTIONS ON BANACH SPACES 15 and k ∈ N. Let λ = {λj}∞ j=1 and µ = {µk}∞ and let ϕ : C2 → C. For n ∈ N, define Tλ,µ ϕ,n ∈ L(L(X, Y)) by k=1 be sequences of complex numbers, (5.4) Tλ,µ ϕ,n (S) := ϕ f (λj, µk)QkSPj (S ∈ L(X, Y)). n ∑ j,k=1 Note that Tλ,µ ϕ,n ∈ L(I) for each Banach ideal I in L(X, Y). Let f ∈ B(C) and extend the divided difference from (4.5), given by ϕ f (λ1, λ2) := f (λ2) − f (λ1) λ2 − λ1 for (λ1, λ2) ∈ C2 with λ1 6= λ2, to a function ϕ f : C2 → C. Lemma 5.3. Let X and Y be Banach spaces with unconditional Schauder bases, and let I be a Banach ideal in L(X, Y). Let λ = {λj}∞ k=1 be sequences of complex numbers, and let A ∈ Ld(X, λ, U), B ∈ Ld(Y, µ, V), f ∈ B(C) and n ∈ N. Then j=1 and µ = {µk}∞ k f (B)Sn − Sn f (A)kI ≤ kUkL(X)kV−1kL(Y)(cid:13)(cid:13)(cid:13)Tλ,µ ϕ f ,n(V(BS − SA)U−1)(cid:13)(cid:13)(cid:13)I j,k=1 V−1QkVSU−1PjU. for all S ∈ I such that BS − SA ∈ I, where Sn := ∑n Proof. Let S ∈ I be such that BS − SA ∈ I. For the duration of the proof write Pj := U−1PjU ∈ L(X) and Qk := V−1QkV ∈ L(Y) for j, k ∈ N. By (5.3), and using that PjPk = 0 and QjQk = 0 for j 6= k, ∑ i,l=1 f (B)Sn − Sn f (A) = ∑ i,l=1 ∞ ∑ j=1 f (µk)Qk(cid:16) n QlSPi(cid:17) − f (λj)(cid:16) n QlSPi(cid:17)Pj ∞ ∑ k=1 n ∑ j,k=1 n ∑ j,k=1 n ∑ j,k=1 n ∑ j,k=1 = = = = ( f (µk) − f (λj))QkSPj f (µk) − f (λj) ∑ µk6=λj µk − λj (µkQkSPj − λjQkSPj) ϕ f (λj, µk)Qk(cid:16)(cid:16) ∞ ∑ l=1 µl Ql(cid:17)S − S(cid:16) ∞ ∑ i=1 λi Pi(cid:17)(cid:17)Pj ϕ f (λj, µk)Qk(BS − SA)Pj (V(BS − SA)U−1)U. = V−1T A,B ϕ f where we have used that BS − SA ∈ Mn. Now use the ideal property of I to conclude the proof. (cid:3) (5.5) For a sequence λ of complex numbers and A ∈ Ld(X, λ, U ), define KA := infnkUkL(X)kU−1kL(X)(cid:12)(cid:12)(cid:12) A ∈ Ld(X, λ, U)o . We will call KA the diagonalizability constant of A. Using the unconditionality of the Schauder basis of X and Assumption 5.1, one can show that KA does not depend on the specific ordering of the sequence λ. Since the sequence λ is, up to ordering, 16 JAN ROZENDAAL, FEDOR SUKOCHEV, AND ANNA TOMSKOVA uniquely determined by A (it is the point spectrum of A), KA only depends on A. Moreover, by Assumption 5.1 and (5.2), kE(σ)kL(X) ≤(cid:13)(cid:13)U−1(cid:13)(cid:13)L(X) kUkL(X) for all σ ⊆ C Borel and U ∈ L(X) such that A ∈ Ld(X, λ, U), where E is the spectral measure of A. Hence (5.6) ν(A) ≤ KA, where ν(A) is the spectral constant of A from Section 3.1. We now derive commutator estimates for A and B in the operator norm, under a boundedness assumption which will be verified for specific X and Y in later sections. Proposition 5.4. Let X and Y be Banach spaces with unconditional Schauder bases, A ∈ Ld(X, λ, U), B ∈ Ld(Y, µ, V) and f ∈ B(C). Suppose that < ∞. (5.7) C := sup Then n∈N(cid:13)(cid:13)(cid:13)Tλ,µ ϕ f ,n(cid:13)(cid:13)(cid:13)L(L(X,Y)) k f (B)S − S f (A)kL(X,Y) ≤ CKAKB kBS − SAkL(X,Y) for all S ∈ L(X, Y). Proof. Let S ∈ L(X, Y) and let Sn ∈ L(X, Y) be as in Lemma 5.3 for n ∈ N. It is straightforward to show that, for each x ∈ X, Snx → Sx as n → ∞. Hence f (B)Snx − Sn f (A)x → f (B)Sx − S f (A)x as n → ∞, for each x ∈ X. Lemma 5.3 and (5.7) now yield k f (B)S − S f (A)kL(X,Y) ≤ lim sup n→∞ k f (B)Sn − Sn f (A)kL(X,Y) ≤ CkUkkV−1kkV(BS − SA)U−1kL(X,Y) ≤ CkUkkU−1kkVkkV−1k kBS − SAkL(X,Y). Taking the infimum over U and V concludes the proof. (cid:3) Remark 5.5. Proposition 5.4 also holds for more general Banach ideals in L(X, Y). Indeed, let I be a Banach ideal in L(X, Y) with the property that, if {Sm}∞ m=1 ⊆ I is an I-bounded sequence which SOT-converges to some S ∈ L(X, Y) as m → ∞, then S ∈ I with kSkI ≤ lim supm→∞ kSmkI . With notation as in Proposition 5.4, if C := sup n∈N(cid:13)(cid:13)(cid:13)Tλ,µ ϕ f ,n(cid:13)(cid:13)(cid:13)L(I ) < ∞ then k f (B)S − S f (A)kI ≤ CKAKB kBS − SAkI for all S ∈ L(X, Y) such that BS − SA ∈ I. This follows directly from the proof of Proposition 5.4. OPERATOR LIPSCHITZ FUNCTIONS ON BANACH SPACES 17 5.2. Estimates for the absolute value function. It is known that Lipschitz esti- mates for the absolute value function are related to estimates for so-called trian- gular truncation operators. For example, in [18] and [14] it was shown that the boundedness of the standard triangular truncation on many operator spaces is equivalent to Lipschitz estimates for the absolute value function. We prove that, in our setting, triangular truncation operators are also related to Lipschitz esti- mates for f the absolute value function. We do so by relating the assumption in (5.7) to so-called triangular truncation operators associated to sequences. We will then bound the norms of these operators in later sections for specific X and Y. Let λ = {λj}∞ j=1 and µ = {µk}∞ k=1 be sequences of real numbers, and let X and Y be as before. For n ∈ N define Tλ,µ ∆,n ∈ L(L(X, Y)) by ∑ µk≤λj (S ∈ L(X, Y)). Tλ,µ ∆,n (S) := n ∑ j,k=1 QkSPj (5.8) We call T A,B ∆ the triangular truncation associated to λ and µ. For f (t) := t for t ∈ R, define ϕ f : C2 → C by ϕ f (λ1, λ2) :=( λ1−λ2 λ1−λ2 1 (5.9) if λ1 6= λ2 otherwise . The following result relates the norm of Tλ,µ ϕ f ,n to that of Tλ,µ ∆,n . Proposition 5.6. There exists a universal constant C ≥ 0 such that the following holds. Let X and Y be Banach spaces with unconditional Schauder bases and let I be a Banach ideal in L(X, Y) with the strong convex compactness property. Let λ and µ be bounded sequences of real numbers. Let f (t) := t for t ∈ R. Then (cid:13)(cid:13)(cid:13)Tλ,µ ϕ f ,n(S)(cid:13)(cid:13)(cid:13)I ≤ C(cid:16)kSkI + kTλ,µ ∆,n (S)kI(cid:17) for all n ∈ N and S ∈ I. In particular, if supn∈N kTλ,µ ∆,n (S)kL(L(X,Y)) holds. Proof. Let n ∈ N and S ∈ I, and write λ = {λj}∞ k=1. Throughout the proof we will only consider λj and µk for 1 ≤ j, k ≤ n, but to simplify the presentation we will not mention this explicitly from here on. We can decompose Tλ,µ ϕ f ,n(S) as j=1 and µ = {µk}∞ < ∞ then (5.7) Tλ,µ ϕ f ,n(S) = ∑ λk,µk≥0 QkSPj − ∑ µk <0<λj µk + λj ∑ λj <0<µk µk − λj QkSPj − ∑ µk + λj µk − λj QkSPj+ λk,µk≤0QkSPj + ∑ λk,µk=0QkSPj. Note that some of these terms may be zero. By the ideal property of I and As- sumption 5.1, (5.10) (cid:13)(cid:13)(cid:13) ∑ λj,µk≥0QkSPj(cid:13)(cid:13)(cid:13)I ≤(cid:13)(cid:13)(cid:13) ∑ µk≥0 Qk(cid:13)(cid:13)(cid:13)L(Y) kSkI(cid:13)(cid:13)(cid:13) ∑ λj≥0Pj(cid:13)(cid:13)(cid:13)L(X) ≤ kSkI . are each bounded by kSkI . 18 JAN ROZENDAAL, FEDOR SUKOCHEV, AND ANNA TOMSKOVA To bound the other terms it is sufficient to show that Similarly,(cid:13)(cid:13)(cid:13)∑λk,µk≤0 QkSPj(cid:13)(cid:13)(cid:13)I and(cid:13)(cid:13)(cid:13)∑λk,µk=0 QkSPj(cid:13)(cid:13)(cid:13)I µk + λj QkSPj(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)I ≤ C′(cid:16) kSkI +(cid:13)(cid:13)(cid:13)Tλ,µ µk − λj ∆,n (S)(cid:13)(cid:13)(cid:13)I(cid:17) ∑ λj,µk >0 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) for some universal constant C′ ≥ 0. Indeed, replacing λ by −λ and µ by −µ then yields the desired conclusion. Let Φ(S) := ∑ λj,µk >0 µk − λj µk + λj QkSPj, and define g ∈ W1,2(R) by g(t) := 2 et+1 for t ∈ R. Note that Φ(S) is equal to ∑ 0<µk≤λj(cid:16)g(cid:16)log λj µk(cid:17) − 1(cid:17) QkSPj + ∑ 0<λj<µk(cid:16)1 − g(cid:16)log λj µk(cid:17)(cid:17) QkSPj. Now let ψg : C2 → C be as in (3.9), and let A := ∑∞ k=1 µkQk ∈ L(Y). Let T A,B ∑∞ ψg be as in (4.2). One can check that (Tλ,µ Φ(S) =T A,B ψg ∆,n (S)Pj+ ∆,n (S)) − ∑ >0Qk(S − Tλ,µ ∑ λj,µk >0QkTλ,µ ∆,n (S))Pj − T A,B λj,µk ψg (S − Tλ,µ ∆,n (S)). j=1 λjPj ∈ L(X) and B := Any Banach space with a Schauder basis is separable and has the bounded ap- proximation property, hence Lemma 3.6 and Proposition 4.2 yield By (5.6), ν(A) = ν(B) = 1. Similarly, . ψg ψg (cid:13)(cid:13)(cid:13)T A,B (cid:13)(cid:13)(cid:13)T A,B (Tλ,µ (S − Tλ,µ ∆,n (S))(cid:13)(cid:13)(cid:13)I ≤ 16√2 ν(A)ν(B)kgkW1,2(R)(cid:13)(cid:13)(cid:13)Tλ,µ ∆,n (S))(cid:13)(cid:13)(cid:13)I ∆,n (S))(cid:13)(cid:13)(cid:13)I ≤ 16√2 kgkW1,2(R)(cid:16)(cid:13)(cid:13)(cid:13)S(cid:13)(cid:13)(cid:13)I +(cid:13)(cid:13)(cid:13)Tλ,µ ∆,n (S))(cid:13)(cid:13)(cid:13)I(cid:17). +(cid:13)(cid:13)(cid:13) ∑ ∆,n (S)(cid:13)(cid:13)(cid:13)I +(cid:13)(cid:13)(cid:13)Tλ,µ ∆,n (S))Pj(cid:13)(cid:13)(cid:13) ≤ 2(cid:13)(cid:13)(cid:13)S(cid:13)(cid:13)(cid:13)I +(cid:13)(cid:13)(cid:13)Tλ,µ ∆,n (S)(cid:13)(cid:13)(cid:13)I(cid:17), ∆,n (S)Pj(cid:13)(cid:13)(cid:13)I kΦ(S)kI ≤(cid:16)2 + 32√2 kgkW1,2(R)(cid:17)(cid:16)(cid:13)(cid:13)(cid:13)S(cid:13)(cid:13)(cid:13)I >0Qk(S − Tλ,µ λj,µk >0 QkTλ,µ . (cid:3) (cid:13)(cid:13)(cid:13) ∑ λj,µk as desired. By the same arguments as in (5.10), Combining all these estimates yields 6. THE ABSOLUTE VALUE FUNCTION ON L(ℓp, ℓq) In this section we study the absolute value function on the space L(ℓp, ℓq). We show that the absolute value function is operator Lipschitz on L(ℓp, ℓq) for p < q < ∞, on L(ℓ1) and on L(c0). We also obtain results for p ≥ q. The key idea of the proof is entirely different from the techniques used in [9], [11], [14] and [18], which are based on a special geometric property of the reflexive Schatten von Neumann ideals (the UMD-property), a property which L(ℓp, ℓq) OPERATOR LIPSCHITZ FUNCTIONS ON BANACH SPACES 19 does not have. Instead, we prove our results by relating estimates for the operators from (5.8) to the standard triangular truncation operator, defined in (6.1) below. For this we use the theory of Schur multipliers on L(ℓp, ℓq) developed in [4]. We then appeal to results from [3] about the boundedness of the standard triangular truncation on L(ℓp, ℓq). 6.1. Schur multipliers. For p ∈ [1, ∞), let {ej}∞ j=1 be the standard Schauder basis of ℓp, with the corresponding projections Pj(x) := xjej for x = ∑∞ k=1 xkek and j ∈ N. We consider this basis and the corresponding projections on all ℓp-spaces simultaneously, for simplicity of notation. Note that Assumption 5.1 is satisfied for this basis. For q ∈ [1, ∞], any operator S ∈ L(ℓp, ℓq) can be represented by an infinite matrix S = {sjk}∞ j,k=1, where sjk := (S(ek), ej) for j, k ∈ N. For an infinite j,k=1 the product M ∗ S := {mjksjk} is the Schur product of the matrix M = {mjk}∞ matrices M and S. The matrix M is a Schur multiplier if the mapping S 7→ M ∗ S is a bounded operator on L(ℓp, ℓq). Throughout, we identify Schur multipliers with their corresponding operators. The notion of a Schur multiplier is a discrete version of a double operator in- tegral (for details see e.g. [30, 33]). Schur multipliers on the space L(ℓp, ℓq) are also called (p, q)-multipliers. We denote by M(p, q) the Banach space of (p, q)- multipliers with the norm kMk(p,q) := supn(cid:13)(cid:13)M ∗ S(cid:13)(cid:13)L(ℓp,ℓq)(cid:12)(cid:12)(cid:12)kSkL(ℓp,ℓq) ≤ 1o . Remark 6.1. We will also consider (p, q)-multipliers M for p = ∞ and q ∈ [1, ∞]. Any operator S ∈ L(c0, ℓq) corresponds to an infinite matrix S = {sjk}∞ j,k=1, and M is said to be a (∞, q)-multiplier if the mapping S 7→ M ∗ S is a bounded operator on L(c0, ℓq), and we define the Banach space M(∞, q) in the obvious way. Below we will often not explicitly distinguish the case p = ∞ from 1 ≤ p < ∞, in order to keep the presentation simple. The reader should keep in mind that for p = ∞ the space ℓp should be replaced by c0. Remark 6.2. It is straightforward to see that kMk(p,q) ≥ supj,k∈Nmj,k for all p, q ∈ [1, ∞] and M ∈ M(p, q). For p, q ∈ [1, ∞] and S ∈ L(ℓp, ℓq), define T∆(S) := ∑ (6.1) k≤j PkSPj, which is a well-defined element of L(ℓr, ℓs) for suitable r, s ∈ [1, ∞], by Proposition 6.3 below. The operator T∆ is the (standard) triangular truncation. This operator can be identified with the following Schur multiplier. Let T′∆ = {t′jk}∞ j,k=1 be a matrix given by t′jk = 1 for k ≤ j and t′jk = 0 otherwise. It is clear that T∆ extends to a bounded linear operator on L(ℓp, ℓq) if and only if T′∆ is a (p, q)-multiplier. For n ∈ N and r, s ∈ [1, ∞] we will consider the operators T∆,n ∈ L(L(ℓp, ℓq), L(ℓr, ℓs)), given by T∆,n(S) := ∑ 1≤k≤j≤nPkSPj (S ∈ L(ℓp, ℓq). The dependence of the (p, q)-norm of T∆ on the indices p and q was determined in [3] and [19] (see also [34]), and is as follows. 20 JAN ROZENDAAL, FEDOR SUKOCHEV, AND ANNA TOMSKOVA Proposition 6.3. Let p, q ∈ [1, ∞]. Then the following statements hold. (i) [3, Theorem 5.1] If p < q, 1 = p = q or p = q = ∞, then T∆ ∈ M(p, q). (ii) [19, Proposition 1.2] If 1 6= p ≥ q 6= ∞, then there is a constant C > 0 such that kT∆,nkL(L(ℓp,ℓq)) ≥ C ln n for all n ∈ N. (iii) [3, Theorem 5.2] If 1 6= p ≥ q 6= ∞, then for each s > q and r < p, and T∆ : L(ℓp, ℓq) → L(ℓr, ℓq) T∆ : L(ℓp, ℓq) → L(ℓp, ℓs) are bounded. bounded mapping L(ℓp, ℓq) → L(ℓr, ℓs), with j,k=1, let eM = {emjk}∞ We will also need the following result, a generalization of [4, Theorem 4.1]. For j,k=1 be obtained from M by repeating the a matrix M = {mjk}∞ first column, i.e., emj1 = mj1 and emjk = mj(k−1) for j ∈ N and k ≥ 2. Proposition 6.4. Let p, q, r, s ∈ [1, ∞] with r ≤ p. Let M = {mjk}∞ S 7→ M ∗ S is a bounded mapping L(ℓp, ℓq) → L(ℓr, ℓs). Then S 7→ eM ∗ S is also a In particular, if M ∈ M(p, q) then eM ∈ M(p, q) with kMk(p,q) =(cid:13)(cid:13)(cid:13)eM(cid:13)(cid:13)(cid:13)(p,q) kMkL(L(ℓp,ℓq),L(ℓr,ℓs)) =(cid:13)(cid:13)(cid:13)eM(cid:13)(cid:13)(cid:13)L(L(ℓp,ℓq),L(ℓr,ℓs)) Proof. The proof is almost identical to that of [4, Theorem 4.1], and the condition r ≤ p is used to ensure that x1p + x2p ≤ (x1r + x2r)p/r for all x1, x2 ∈ C (with the obvious modification for p = ∞ or r = ∞). j,k=1 be such that . . (cid:3) Remark 6.5. By considering the transpose M′ of a matrix M, and using that M′ : L(ℓq∗, ℓp∗ ) → L(ℓs∗, ℓr∗) with (cid:13)(cid:13)M′(cid:13)(cid:13)L(L(ℓ q∗,ℓ p∗ ), L(ℓ s∗,ℓ r∗ )) = kMkL(L(ℓp,ℓq), L(ℓr,ℓs)) , as is straightforward to check, one obtains from Proposition 6.4 that for q∗ ≥ s∗, that is, for s ≤ q, row repetitions do not alter the k·kL(L(ℓp,ℓq), L(ℓr,ℓs))-norm of a matrix. Moreover, since kSkL(ℓp,ℓq) is invariant under permutations of the columns and rows of S ∈ L(ℓp, ℓq), rearrangement of the rows and columns of M ∈ L(L(ℓp, ℓq), L(ℓr, ℓs)) does not change the norm kMkL(L(ℓp,ℓq), L(ℓr,ℓs)). The following lemma is crucial to our main results. Lemma 6.6. Let p, q, r, s ∈ [1, ∞] with r ≤ p and s ≤ q. Let λ = {λj}∞ µ = {µk}∞ k=1 be sequences of real numbers. Then j=1 and (cid:13)(cid:13)(cid:13)Tλ,µ ∆,n(cid:13)(cid:13)(cid:13)L(L(ℓp,ℓq), L(ℓr,ℓs)) ≤ kT∆,nkL(L(ℓp,ℓq), L(ℓr,ℓs)) for all n ∈ N. OPERATOR LIPSCHITZ FUNCTIONS ON BANACH SPACES 21 ∆,n (S) = M ∗ S for all S ∈ L(ℓp, ℓq), where M = {mjk}∞ Proof. Note that Tλ,µ j,k=1 is such that mjk = 1 if 1 ≤ j, k ≤ n and µk ≤ λj, and mjk = 0 otherwise. It suffices to prove that kMkL(L(ℓp,ℓq), L(ℓr,ℓs)) ≤ kT kL(L(ℓp,ℓq), L(ℓr,ℓs)). Assume that M is non- zero, otherwise the statement is trivial. By Remark 6.5, rearrangement of the rows and columns of M does not change its norm. Hence, we may assume that {λj}n j=1 and {µk}n k=1 are decreasing. Now M has the property that, if mjk = 1, then mil = 1 for all i ≤ j and k ≤ l ≤ m2. By Proposition 6.4 and Remark 6.5, we may omit repeated rows and columns of M, and doing this repeatedly reduces M to T∆,N for some 1 ≤ N ≤ n. Noting that kT∆,NkL(L(ℓp,ℓq), L(ℓr,ℓs)) ≤ kT∆,nkL(L(ℓp,ℓq), L(ℓr,ℓs)) concludes the proof. (cid:3) 6.2. The cases p < q, p = q = 1, p = q = ∞. We now combine the theory from the previous sections to deduce our main results. Throughout this section let f (t) := t for t ∈ R. Theorem 6.7. Let p, q ∈ [1, ∞] with p < q, p = q = 1 or p = q = ∞. Then there exists a constant C ≥ 0 such that the following holds (where ℓ∞ should be replaced by c0). Let A ∈ Ld(ℓp) and B ∈ Ld(ℓq) have real spectrum. Then k f (B)S − S f (A)kL(ℓp,ℓq) ≤ CKAKB kBS − SAkL(ℓp,ℓq) for all S ∈ L(ℓp, ℓq). Proof. Simply combine Propositions 5.4 and 5.6 and Lemma 6.6 with Proposition 6.3 (i). (cid:3) We single out the specific case in Theorem 6.7 where p = q = 1 or p = q = ∞ and S is the identity operator. Corollary 6.8. There exists a universal constant C ≥ 0 such that k f (B) − f (A)kL(ℓ1) ≤ CKAKB kB − AkL(ℓ1) for all A, B ∈ Ld(ℓ1) with real spectrum. Corollary 6.9. There exists a universal constant C ≥ 0 such that k f (B) − f (A)kL(c0) ≤ CKAKB kB − AkL(c0) for all A, B ∈ Ld(c0) with real spectrum. In the case of Theorem 6.7 where p = 2 or q = 2, we can apply our results to compact self-adjoint operators. By the spectral theorem, any compact self-adjoint operator A ∈ L(ℓ2) has an orthonormal basis of eigenvectors, and therefore A ∈ Ld(ℓ2, λ, U) for some sequence λ of real numbers and an isometry U ∈ L(ℓ2). Thus Theorem 6.7 yields the following corollaries. Corollary 6.10. Let p ∈ [1, 2). Then there exists a constant C ≥ 0 such that the following holds. Let A ∈ Ld(ℓp) and let B ∈ L(ℓ2) be compact and self-adjoint. Then k f (B)S − S f (A)kL(ℓp,ℓ2) ≤ CKA kBS − SAkL(ℓp,ℓ2) for all S ∈ L(ℓp, ℓ2). 22 JAN ROZENDAAL, FEDOR SUKOCHEV, AND ANNA TOMSKOVA Corollary 6.11. Let q ∈ (2, ∞]. Then there exists a constant C ≥ 0 such that the following holds (where ℓ∞ should be replaced by c0). Let A ∈ L(ℓ2) be compact and self-adjoint, and let B ∈ Ld(ℓq). Then k f (B)S − S f (A)kL(ℓ2,ℓq) ≤ CKB kBS − SAkL(ℓ2,ℓq) for all S ∈ L(ℓ2, ℓq). Remark 6.12. Corollaries 6.8 and 6.9 show that the absolute value function is op- erator Lipschitz on ℓ1 and c0, in the following sense. For fixed M ≥ 1, there exists a constant C ≥ 0 such that k f (B) − f (A)k ≤ C kB − Ak for all diagonalizable operators A and B such that KA, KB ≤ M, and C is indepen- dent of A and B. For p < q a similar statement holds. Restricting B ∈ Ld(ℓq) and f (B) ∈ L(ℓq) to operators from ℓp to ℓq, and letting S be the inclusion mapping ℓp ֒→ ℓq in Theorem 6.7, one can suggestively write k f (B) − f (A)kL(ℓp,ℓq) ≤ C kB − AkL(ℓp,ℓq) , for all A ∈ Ld(ℓp) and B ∈ Ld(ℓq) with KA, KB ≤ M. This also applies to Corol- laries 6.10 and 6.11. 6.3. The case p ≥ q. We now examine the absolute value function f on L(ℓp, ℓq) for p ≥ q, and obtain the following result. Proposition 6.13. Let p, q ∈ (1, ∞] with p ≥ q. Then for each s < q there exists a constant C ≥ 0 such that the following holds (where ℓ∞ should be replaced by c0). Let A ∈ Ld(ℓp, λ, U) and B ∈ Ld(ℓq, µ, V) have real spectrum, and let S ∈ L(ℓp, ℓq) be such that V(BS − SA)U−1 ∈ L(ℓp, ℓs). Then k f (B)S − S f (A)kL(ℓp,ℓq) ≤ CkUkL(ℓp)kV−1kL(ℓq)kV(BS − SA)U−1kL(ℓp,ℓs). In particular, if p = q and V(B − A)U−1 ∈ L(ℓp, ℓs), then k f (B) − f (A)kL(ℓp) ≤ CkUkL(ℓp)kV−1kL(ℓp)kV(B − A)U−1kL(ℓp,ℓs). Proof. Let R := V(BS − SA)U−1. With notation as in Lemma 5.3, for each n ∈ N. Proposition 5.6, Lemma 6.6 (with p = r and with q and s in- terchanged) and Proposition 6.3 (iii) (with q and s interchanged) yield a constant C′ ≥ 0 such that k f (B)Sn − Sn f (A)kL(ℓp,ℓq) ≤ kUkL(ℓp)kV−1kL(ℓq)(cid:13)(cid:13)(cid:13)Tλ,µ (cid:13)(cid:13)(cid:13)Tλ,µ ϕ f ,n(R)(cid:13)(cid:13)(cid:13)L(ℓp,ℓq) ≤ C′(cid:16)kRkL(ℓp,ℓq) + kRkL(ℓp,ℓs)(cid:17) . ϕ f ,n(R)(cid:13)(cid:13)(cid:13)L(ℓp,ℓq) Since L(ℓp, ℓs) ֒→ L(ℓp, ℓq) contractively, k f (B)Sn − Sn f (A)kL(ℓp,ℓq) ≤ CkUkL(ℓp)kV−1kL(ℓq)kV(BS − SA)U−1kL(ℓp,ℓs) for all n ∈ N, where C = 2C′. Finally, as in the proof of Proposition 5.4, one lets n tend to infinity to conclude the proof. (cid:3) OPERATOR LIPSCHITZ FUNCTIONS ON BANACH SPACES 23 In the same way, appealing to the second part of Proposition 6.3 (iii), one de- duces the following result. Proposition 6.14. Let p, q ∈ [1, ∞) with p ≥ q. Then for each r > p there exists a constant C ≥ 0 such that the following holds (where ℓ∞ should be replaced by c0). Let A ∈ Ld(ℓp, λ, U) and B ∈ Ld(ℓq, µ, V) have real spectrum, and let S ∈ L(ℓp, ℓq) be such that V(BS − SA)U−1 ∈ L(ℓr, ℓq). Then k f (B)S − S f (A)kL(ℓp,ℓq) ≤ CkUkL(ℓp)kV−1kL(ℓq)kV(BS − SA)U−1kL(ℓr,ℓq). In particular, if p = q and V(B − A)U−1 ∈ L(ℓr, ℓq), then k f (B) − f (A)kL(ℓp) ≤ CkUkL(ℓp)kV−1kL(ℓp)kV(B − A)U−1kL(ℓr,ℓq). We single out the case where p = q = 2. Here we write f (A) = A for a normal operator A ∈ L(ℓ2), since then f (A) is equal to A := √A∗ A. Note also that the following result applies in particular to compact self-adjoint operators. For simplicity of the presentation we only consider ǫ ∈ (0, 1], it should be clear how the result extends to other ǫ > 0. Corollary 6.15. For each ǫ ∈ (0, 1] there exists a constant C ≥ 0 such that the follow- ing holds. Let A ∈ Ld(ℓ2, λ, U) and B ∈ Ld(ℓ2, µ, V) be self-adjoint, with U and V unitaries, and let S ∈ L(ℓ2). If V(BS − SA)U−1 ∈ L(ℓ2, ℓ2−ǫ), then kBS − SAkL(ℓ2) ≤ CkV(BS − SA)U−1kL(ℓ2,ℓ2−ǫ) and if V(BS − SA)U−1 ∈ L(ℓ2+ǫ, ℓ2) then kBS − SAkL(ℓ2) ≤ CkV(BS − SA)U−1kL(ℓ2+ǫ,ℓ2). In particular, if V(B − A)U−1 ∈ L(ℓ2, ℓ2−ǫ), then kB − AkL(ℓ2) ≤ CkV(B − A)U−1kL(ℓ2,ℓ2−ǫ) and if V(B − A)U−1 ∈ L(ℓ2+ǫ, ℓ2), then kB − AkL(ℓ2) ≤ CkV(B − A)U−1kL(ℓ2+ǫ,ℓ2). 7. LIPSCHITZ ESTIMATES ON THE IDEAL OF p-SUMMING OPERATORS Let H be a separable infinite-dimensional Hilbert space. It was shown in [1] that a matrix M = {mjk}∞ j,k=1 is a Schur multiplier on the Hilbert-Schmidt class S2 ⊂ L(H) if and only if supj,k mjk < ∞. By [23], S2 coincides with the Banach ideal Πp(H) of all p-summing operators (see the definition below) for all 1 ≤ p < ∞. Hence a matrix M = {mjk}∞ j,k=1 is a Schur multiplier on Πp(H) if and only if supj,k mjk < ∞. In Lemma 7.2 we show that the same statement is true for the Banach ideal Πp(ℓp∗, ℓp) in L(ℓp∗, ℓp), for all 1 < p < ∞. As a corollary we obtain operator Lipschitz estimates on Πp(ℓp∗, ℓp) for each Lipschitz function f on C. Let X and Y be Banach spaces and 1 ≤ p < ∞. An operator S : X → Y is p-absolutely summing if there exists a constant C such that for each n ∈ N and each collection {xj}n j=1 ⊆ X, (7.1) ∑ Y(cid:17) 1 (cid:16) n j=1(cid:13)(cid:13)S(xj)(cid:13)(cid:13)p p ≤ C sup kx∗kX∗≤1(cid:16) n ∑ j=1hx∗, xjip(cid:17) 1 p . 24 JAN ROZENDAAL, FEDOR SUKOCHEV, AND ANNA TOMSKOVA Below we consider p-absolutely summing operators from ℓp∗ to ℓp. We first The smallest such constant is denoted by πp, and Πp(X, Y) is the space of p- absolutely summing operators from X to Y. We let Πp(X) := Πp(X, X). By Propo- sitions 2.3, 2.4 and 2.6 in [13], (Πp(X, Y), πp(·)) is a Banach ideal in L(X, Y). present the following result. Lemma 7.1. Let 1 < p < ∞ and S = {sjk}∞ ∞ ∑ j,k=1. Then S ∈ Πp(ℓp∗, ℓp) if and only if < ∞. p cp :=(cid:16) ∞ ∑ j=1 k=1sjkp(cid:17) 1 In this case, πp(S) = cp. Proof. In [13, Example 2.11] it is shown that, if cp < ∞, then S ∈ Πp(ℓp∗, ℓp) with πp(S) ≤ cp. For the converse, let n ∈ N and let xj := ej ∈ ℓp∗ for 1 ≤ j ≤ n. By (7.1) (with X = ℓp∗ and Y = ℓp), (cid:16) n ∑ k=1 ∞ ∑ j=1sjkp(cid:17) 1 p ≤ πp(S). Letting n tend to infinity concludes the proof. (cid:3) For the following corollary of Lemma 7.1, recall that a matrix M is said to be a Schur multiplier on a subspace I ⊆ L(ℓp, ℓq) if S 7→ M ∗ S is a bounded map on I. Recall also the definition of the standard triangular truncation T∆ from (6.1). Corollary 7.2. Let p ∈ (1, ∞) and let M = {mjk}∞ multiplier on Πp(ℓp∗, ℓp) if and only if supj,k mjk < ∞. In this case, j,k=1 be a matrix. Then M is a Schur p∗ ,ℓp)) = sup In particular, T∆ ∈ L(Πp(ℓp∗, ℓp)) with kT∆kL(Πp(ℓ kMkL(Πp(ℓ j,k mjk. p∗,ℓp)) = 1. less, T∆ is bounded on the ideal Πp(ℓp∗, ℓp) ⊂ L(ℓp∗, ℓp) for all p ∈ (1, ∞). Observe that T∆ /∈ L(L(ℓp∗, ℓp)) for p∗ ≥ p, by Proposition 6.3 (ii). Neverthe- For a Lipschitz function f : C → C, write k fkLip := sup z1,z2∈C z16=z2 Moreover, let ϕ f : C2 → C be given by f (z1) − f (z2) . z1 − z2 ϕ f (λ1, λ2) :=( λ1−λ2 λ1−λ2 0 if λ1 6= λ2 otherwise . We now prove our main result concerning commutator estimates on Πp(ℓp∗, ℓp). : C → C be Theorem 7.3. Let p ∈ (1, ∞), A ∈ Ld(ℓp∗ ) and B ∈ Ld(ℓp). Let f Lipschitz. Then (7.2) for all S ∈ L(ℓp∗, ℓp) such that BS − SA ∈ Πp(ℓp∗, ℓp). πp( f (B)S − S f (A)) ≤ KAKB k fkLip πp(BS − SA) OPERATOR LIPSCHITZ FUNCTIONS ON BANACH SPACES 25 j=1 and µ = {µk}∞ It follows directly from (7.1) that, if {Sm}∞ Proof. Let λ = {λj}∞ k=1 be sequences such that A ∈ Ld(ℓp∗, λ, U) and B ∈ Ld(ℓp, µ, V) for certain U ∈ L(ℓp∗) and V ∈ L(ℓp). m=1 ⊆ Πp(ℓp∗, ℓp) is a πp-bounded se- quence which SOT-converges to some S ∈ L(X, Y), then S ∈ Πp(ℓp∗, ℓp) with πp(S) ≤ lim supm→∞ πp(Sm). Hence, by Remark 5.5, it suffices to prove that supn∈N kTλ,µ p∗,ℓp)) ≤ k fkLip, where ϕ f ,nkL(Πp(ℓ Tλ,µ ϕ f ,n(S) = n ∑ j,k=1 ϕ f (λj, µk)PkSPj (S ∈ Πp(ℓp∗, ℓp)) for n ∈ N. Fix n ∈ N and note that Tλ,µ M = {mjk}∞ mjk = 0 otherwise. Then ϕ f ,n(S) = M ∗ S for S ∈ Πp(ℓp∗, ℓp), where j,k=1 is the matrix given by mjk = ϕ f (λj, µk) for 1 ≤ j, k ≤ n, and sup j,k mjk ≤ sup j,k ϕ f (λj, µk) ≤ k fkLip . Corollary 7.2 now concludes the proof. q + 1 Problem 7.4. Let p, q, r ∈ [1, ∞] be such that q ≥ 2 and 1 2 . Then the Schatten class Sr coincides with the Banach ideal Πp,q(ℓ2) of (p, q)-summing operators on ℓ2 (for the definition of (p, q)-summing operators see [13]). Hence, by [14], the standard triangular truncation is bounded on Πp,q(ℓ2) for these indices. For which r1, r2 ∈ [1, ∞] is the standard triangular truncation bounded on Πp,q(ℓr1, ℓr2 )? Are there other non-trivial ideals I in L(ℓp, ℓq) such that the standard triangular truncation is bounded on I? As shown in Theorem 7.3 (and the rest of this paper), answers to these questions are linked to commutator estimates for diagonalizable operators. 2 and 1 q − 1 p − 1 r = 1 < 1 p (cid:3) 8. MATRIX ESTIMATES In this section we apply the theory developed in Sections 4, 5 and 6 to finite- dimensional spaces, and in doing so we obtain Lipschitz estimates which are in- dependent of the dimension of the underlying space. The results from Section 7, in particular Theorem 7.3, also yield dimension-independent estimates on finite- dimensional spaces, but we leave the straightforward derivation of these results to the reader. 8.1. Finite-dimensional spaces. Let n ∈ N and let X be an n-dimensional Banach space with basis {e1, . . . , en} ⊂ X. For 1 ≤ k ≤ n, let Pk ∈ L(X) be the projec- tion given by Pk(x1e1 + . . . + xnen) := xkek for (x1, . . . , xn) ∈ Cn. Recall that an operator A ∈ L(X) is diagonalizable if there exists U ∈ L(X) invertible such that (8.1) U AU−1 = n ∑ k=1 λkPk for some (λ1, . . . , λn) ∈ Cn. In this case we write A ∈ Ld(X, {λj}n also the definition of spectral and scalar type operators from Section 3.1. Lemma 8.1. Each A ∈ L(X) is a spectral operator. Furthermore, A is a scalar type operator if and only if A is diagonalizable. If A ∈ Ld(X, {λj}n j=1, U) then the spectral j=1, U). Recall 26 JAN ROZENDAAL, FEDOR SUKOCHEV, AND ANNA TOMSKOVA measure E of A is given by E(U) = 0 if U ∩ sp(A) = ∅, and E({λ}) = ∑λj=λ U−1PjU for λ ∈ sp(A). Proof. It was already remarked in Section 5 that any diagonalizable operator is a scalar type operator, with spectral measure as specified. By [15, Theorem XV.4.5], an operator T ∈ L(Y) on an arbitrary Banach space Y is a spectral operator if and only if T = S + N for a scalar type operator S ∈ L(Y) and a generalized nilpotent operator N ∈ L(Y) such that NS = SN, and this decomposition is unique. Com- bining this with the Jordan decomposition for matrices yields that each A ∈ L(X) is a spectral operator. If A is a scalar type operator, then the Jordan decomposition yields a commuting diagonalizable S and a nilpotent N such that A = S + N. By the uniqueness of such a decomposition [15, Theorem XV.4.5], N = 0 and A = S is diagonalizable. (cid:3) Let A ∈ Ld(X, {λj}n j=1, U). As in (5.3), f (A) = U−1(cid:16) n ∑ k=1 f (λk)Pk(cid:17)U. L(X); Let Y be a finite-dimensional Banach space. A norm k·k on L(X, Y) is symmetric if • kRSTk ≤ kRkL(Y) kSk kTkL(X) for all R ∈ L(Y), S ∈ L(X, Y) and T ∈ • kx∗⊗ yk = kx∗kX∗ kykY for all x∗ ∈ X∗ and y ∈ Y. Clearly (L(X, Y), k·k) is a Banach ideal in the sense of Section 3.2 if and only if k·k is symmetric. The following result extends inequalities which were known for self-adjoint operators on finite-dimensional Hilbert spaces and unitarily invariant norms (see e.g. [18] and [5, Chapter X]), to diagonalizable operators on finite-dimensional Ba- nach spaces and symmetric norms. Note that, for general finite-dimensional Ba- nach spaces X and Y, a symmetric norm on L(X, Y) is not unitarily invariant. Let A := A(C × C) be as in Section 3.3, and for f ∈ B(C) let ϕ f (λ1, λ2) := f (λ2)− f (λ1) for (λ1, λ2) ∈ C2 with λ1 6= λ2, as in (4.5). Recall the definition of the spectral constants ν(A) and ν(B) from Section 3.1. The following is a direct corollary of Theorem 4.6, since L(X, Y) has the strong convex compactness property. Theorem 8.2. Let f ∈ B(C) be such that ϕ f extends to an element of A. Let X and Y be finite-dimensional Banach spaces, k·k a symmetric norm on L(X, Y), and let A ∈ Ld(X) and B ∈ Ld(Y). Then k f (B)S − S f (A)k ≤ 16 ν(A)ν(B)(cid:13)(cid:13)ϕ f(cid:13)(cid:13)A kBS − SAk k f (B) − f (A)k ≤ 16 ν(A)ν(B)(cid:13)(cid:13)ϕ f(cid:13)(cid:13)A kB − Ak . for all S ∈ L(X, Y). In particular, if X = Y, (8.2) Corollary 8.3. There exists a universal constant C ≥ 0 such that the following holds. Let X and Y be finite-dimensional Banach spaces and k·k a symmetric norm on L(X, Y). Let f ∈ B1 ∞,1(R), and let A ∈ Ld(X) and B ∈ Ld(Y) be such that sp(A) ∪ sp(B) ⊆ R. Then λ2−λ1 k f (B)S − S f (A)k ≤ C ν(A)ν(B) k fk B1 ∞,1(R)kBS − SAk OPERATOR LIPSCHITZ FUNCTIONS ON BANACH SPACES 27 for all S ∈ L(X, Y). In particular, if X = Y, k f (B) − f (A)k ≤ C ν(A)ν(B)k fk B1 ∞,1(R)kB − Ak . Remark 8.4. Let σ1, σ2 ⊂ C be finite sets. Then any ϕ : σ1 × σ2 → C belongs to A(σ1 × σ2). Indeed, one can find a representation as in (3.8) by letting Ω be finite and solving a system of linear equations. Therefore Theorem 4.6 yields an estimate k f (B)S − S f (A)k ≤ 16 ν(A)ν(B)(cid:13)(cid:13)ϕ f(cid:13)(cid:13)A(sp(A)×sp(B)) kBS − SAk as in (4.6) for all f ∈ B(C). This might lead one to think that the restriction in Theorem 8.2 that ϕ f extends to an element of A is not really necessary. However, for general f ∈ B(C) the norm(cid:13)(cid:13)ϕ f(cid:13)(cid:13)A(sp(A)×sp(B)) may blow up as the number of points in sp(A) and sp(B) grows to infinity. Indeed, as remarked before, letting f ∈ B(C) be the absolute value function and considering the operator norm, a dimension-independent estimate as in (8.2) does not hold for all self-adjoint oper- ators on all finite-dimensional Hilbert spaces [5, (X.25)]. Hence ϕ f does not extend to an element of A, and one cannot expect to obtain Theorem 8.2 or Corollary 8.3 for all bounded Borel functions on C. 8.2. The absolute value function. We now apply the results for the absolute value function from previous sections to finite-dimensional spaces. Throughout this sec- tion, let f (t) := t for t ∈ R. First note that Lemma 5.3 and Proposition 5.6 relate commutator estimates to estimates for triangular truncation operators for general symmetric norms on ma- trix spaces. Theorem 6.7 immediately yields the following result. It shows that, although the Lipschitz estimate k f (B) − f (A)kL(ℓn q ) ≤ const kB − AkL(ℓn p,ℓn p,ℓn q ) does not hold with a constant independent of the dimension n for f the absolute value function, p = q = 2 and all self-adjoint operators on ℓn 2 , one can neverthe- less obtain such estimates for p < q, p = q = 1 or p = q = ∞ by considering diagonalizable operators A and B for which KA, KB ≤ M, for some fixed M ≥ 1. For A a diagonalizable operator, recall the definition of KA from (5.5). Theorem 8.5. Let p, q ∈ [1, ∞] with p < q, p = q = 1 or p = q = ∞. Then there exists a constant C ≥ 0 such that the following holds. Let n ∈ N and let A ∈ Ld(ℓn p) and B ∈ Ld(ℓn q ) have real spectrum. Then k f (B)S − S f (A)kL(ℓn q ) ≤ CKAKB kBS − SAkL(ℓn p,ℓn p,ℓn q ) p, ℓn q ). In particular, k f (B) − f (A)kL(ℓn q ) ≤ CKAKB kB − AkL(ℓn p,ℓn p,ℓn q ) . for all S ∈ L(ℓn For n ∈ N and p ∈ [1, ∞) let ℓn k(x1, . . . , xn)kp :=(cid:16) n ∑ and let ℓn ∞ be Cn with the norm k(x1, . . . , xn)k∞ := max p denote Cn with the p-norm j=1xjp(cid:17)1/p 1≤j≤nxj (cid:0)(x1, . . . , xn) ∈ Cn(cid:1), (cid:0)(x1, . . . , xn) ∈ Cn(cid:1). 28 JAN ROZENDAAL, FEDOR SUKOCHEV, AND ANNA TOMSKOVA Of course Corollaries 6.10 and 6.11 imply that for p = 2 or q = 2 and A or B self-adjoint, the estimates in Theorem 8.5 simplify. For p ≥ q, Propositions 6.13 and 6.14 yield dimension-independent estimates. We state the estimates which follow from Proposition 6.13, the analogous result that follows from Proposition 6.14 should be obvious. Proposition 8.6. Let p, q ∈ (1, ∞] with p ≥ q. For each s < q there exists a constant C ≥ 0 such that the following holds. Let n ∈ N, and let A ∈ Ld(ℓn p, λ, U) and B ∈ Ld(ℓn q , µ, V) have real spectrum. Then k f (B)S − S f (A)kL(ℓn q ) ≤ CkUkL(ℓn p,ℓn p)kV−1kL(ℓn q )kV(BS − SA)UkL(ℓn p,ℓn s ) for all S ∈ L(ℓn p, ℓn q ). In particular, k f (B) − f (A)kL(ℓn q ) ≤ CkUkL(ℓn p,ℓn p)kV−1kL(ℓn q )kV(B − A)UkL(ℓn p,ℓn s ). In the case p = q = 2, Corollary 6.15 implies the following. Again we only consider ǫ ∈ (0, 1], for simplicity, but the result extends in an obvious manner to other ǫ > 0. We write f (A) = A = √A∗ A for a normal operator A on ℓn Corollary 8.7. For each ǫ ∈ (0, 1] there exists a constant C ≥ 0 such that the following holds. Let n ∈ N and let A ∈ Ld(ℓn 2 , µ, V) be self-adjoint operators, with U and V unitaries. Then 2 , λ, U) and B ∈ Ld(ℓn 2 . 2 ) ≤ C min(kV(B − A)U−1kL(ℓn 2−ǫ), kV(B − A)U−1kL(ℓn 2 ,ℓn (8.3) kB − AkL(ℓn We note that (8.3) in turn yields, for instance, the following estimate: 2 ) min(kU−1kL(ℓn 2−ǫ), kVkL(ℓn 2 ,ℓn 2 ) ≤ CkB − AkL(ℓn kB − AkL(ℓn 2+ǫ,ℓn 2 )). 2+ǫ,ℓn 2 )). Acknowledgements. Most of the work on this article was performed during a re- search visit of the first-named author to the University of New South Wales. He would like to thank UNSW and its faculty, in particular Michael Cowling, for their wonderful hospitality. We are indebted to many colleagues in Delft and at UNSW for their suggestions and ideas. In particular, we would like to mention Ian Doust, Ben de Pagter, Denis Potapov and Dmitriy Zanin. We also thank Albrecht Pietsch for useful comments and references concerning p-summing operators. REFERENCES [1] J. Arazy. Certain Schur-Hadamard multipliers in the spaces Cp. Proc. Amer. Math. Soc., 86(1):59 -- 64, 1982. [2] N. A. Azamov, A. L. Carey, P. G. Dodds, and F. A. Sukochev. Operator integrals, spectral shift, and spectral flow. Canad. J. Math., 61(2):241 -- 263, 2009. [3] G. Bennett. Unconditional convergence and almost everywhere convergence. Z. Wahrschein- lichkeitstheorie und Verw. Gebiete, 34(2):135 -- 155, 1976. [4] G. Bennett. Schur multipliers. Duke Math. J., 44(3):603 -- 639, 1977. [5] R. Bhatia. Matrix analysis, volume 169 of Graduate Texts in Mathematics. Springer-Verlag, New York, 1997. [6] M. S. Birman and M. Z. Solomjak. Double Stieltjes operator integrals. In Probl. Math. Phys., No. I, Spectral Theory and Wave Processes (Russian), pages 33 -- 67. Izdat. Leningrad. Univ., Leningrad, 1966. [7] M. S. Birman and M. Z. Solomjak. Double Stieltjes operator integrals. III. In Problems of mathematical physics, No. 6 (Russian), pages 27 -- 53. Izdat. Leningrad. Univ., Leningrad, 1973. [8] M. Sh. Birman and M. Solomyak. Double operator integrals in a Hilbert space. Integral Equations Operator Theory, 47(2):131 -- 168, 2003. OPERATOR LIPSCHITZ FUNCTIONS ON BANACH SPACES 29 [9] E. B. Davies. Lipschitz continuity of functions of operators in the Schatten classes. J. London Math. Soc. (2), 37(1):148 -- 157, 1988. [10] B. de Pagter and F. A. Sukochev. Differentiation of operator functions in non-commutative L p- spaces. J. Funct. Anal., 212(1):28 -- 75, 2004. [11] B. de Pagter, H. Witvliet, and F. A. Sukochev. Double operator integrals. J. Funct. Anal., 192(1):52 -- 111, 2002. [12] A. Defant and K. Floret. Tensor norms and operator ideals, volume 176 of North-Holland Mathematics Studies. North-Holland Publishing Co., Amsterdam, 1993. [13] J. Diestel, H. Jarchow, and A. Tonge. Absolutely summing operators, volume 43 of Cambridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge, 1995. [14] P. G. Dodds, T. K. Dodds, B. de Pagter, and F. A. Sukochev. Lipschitz continuity of the absolute value and Riesz projections in symmetric operator spaces. J. Funct. Anal., 148(1):28 -- 69, 1997. [15] N. Dunford and J. T. Schwartz. Linear operators. Part III: Spectral operators. Interscience Publishers [John Wiley & Sons, Inc.], New York-London-Sydney, 1971. With the assistance of William G. Bade and Robert G. Bartle, Pure and Applied Mathematics, Vol. VII. [16] I. C. Gohberg and M. G. Kreın. Theory and applications of Volterra operators in Hilbert space. Translated from the Russian by A. Feinstein. Translations of Mathematical Monographs, Vol. 24. American Mathematical Society, Providence, R.I., 1970. [17] T. Kato. Continuity of the map S 7→ S for linear operators. Proc. Japan Acad., 49:157 -- 160, 1973. [18] H. Kosaki. Unitarily invariant norms under which the map A → A is Lipschitz continuous. Publ. [19] S. Kwapie ´n and A. Pełczy ´nski. The main triangle projection in matrix spaces and its applications. Res. Inst. Math. Sci., 28(2):299 -- 313, 1992. Studia Math., 34:43 -- 68, 1970. [20] S. Lassalle and P. Turco. The weak bounded approximation property for A. Online at [21] J. Lindenstrauss and L. Tzafriri. Classical Banach spaces. I. Springer-Verlag, Berlin, 1977. Sequence http://arxiv.org/abs/1410.5670, 2014. spaces, Ergebnisse der Mathematik und ihrer Grenzgebiete, Vol. 92. [22] A. McIntosh. Counterexample to a question on commutators. Proc. Amer. Math. Soc., 29:337 -- 340, 1971. [23] A. Pełczy ´nski. A characterization of Hilbert-Schmidt operators. Studia Math., 28:355 -- 360, 1966/1967. [24] V. V. Peller. Hankel operators in the theory of perturbations of unitary and selfadjoint operators. Funktsional. Anal. i Prilozhen., 19(2):37 -- 51, 96, 1985. [25] V. V. Peller. Hankel operators in the perturbation theory of unbounded selfadjoint operators. In Analysis and partial differential equations, volume 122 of Lecture Notes in Pure and Appl. Math., pages 529 -- 544. Dekker, New York, 1990. [26] V. V. Peller. Multiple operator integrals and higher operator derivatives. J. Funct. Anal., 233(2):515 -- 544, 2006. [27] A. Pietsch. Operator ideals, volume 20 of North-Holland Mathematical Library. North-Holland Pub- lishing Co., Amsterdam-New York, 1980. Translated from German by the author. [28] D. Potapov and F. Sukochev. Double operator integrals and submajorization. Math. Model. Nat. Phenom., 5(4):317 -- 339, 2010. [29] D. Potapov and F. Sukochev. Operator-Lipschitz functions in Schatten-von Neumann classes. Acta Math., 207(2):375 -- 389, 2011. [30] D. Potapov, F. Sukochev, and A. Tomskova. On the Arazy conjecture concerning Schur multipliers on Schatten ideals. Adv. Math., 268:404 -- 422, 2015. [31] G. Schl uchtermann. On weakly compact operators. Math. Ann., 292(2):263 -- 266, 1992. [32] Ivan Singer. Bases in Banach spaces. I. Springer-Verlag, New York-Berlin, 1970. Die Grundlehren der mathematischen Wissenschaften, Band 154. [33] A. Skripka, F. Sukochev, and A. Tomskova. Multilinear operator integrals: Development and ap- plications. preprint, 2015. [34] F. Sukochev and A. Tomskova. (E, F)-Schur multipliers and applications. Studia Math., 216(2):111 -- 129, 2013. [35] N. N. Vakhania, V. I. Tarieladze, and S. A. Chobanyan. Probability distributions on Banach spaces, volume 14 of Mathematics and its Applications (Soviet Series). D. Reidel Publishing Co., Dordrecht, 1987. Translated from the Russian and with a preface by Wojbor A. Woyczynski. [36] J. Voigt. On the convex compactness property for the strong operator topology. Note Mat., 12:259 -- 269, 1992. 30 JAN ROZENDAAL, FEDOR SUKOCHEV, AND ANNA TOMSKOVA DELFT INSTITUTE OF APPLIED MATHEMATICS, MEKELWEG 4, 2628CD DELFT, THE NETHER- LANDS E-mail address: [email protected] SCHOOL OF MATHEMATICS & STATISTICS, UNIVERSITY OF NSW, KENSINGTON NSW 2052 AUS- TRALIA E-mail address: [email protected] E-mail address: [email protected]
1701.07311	1	1701	2017-01-25T13:55:27	Simultaneous universality	[ "math.FA", "math.DS" ]	In this paper, the notion of simultaneous universality is introduced, concerning operators having orbits that simultaneously approximate any given vector. This notion is related to the well known concepts of universality and disjoint universality. Several criteria are provided, and several applications to specific operators or sequences of operators are performed, mainly in the setting of sequence spaces or spaces of holomorphic functions.	math.FA	math	SIMULTANEOUS UNIVERSALITY L. BERNAL-GONZ ´ALEZ AND A. JUNG Abstract. In this paper, the notion of simultaneous universality is introduced, concerning operators having orbits that simultaneously approximate any given vector. This notion is related to the well known concepts of universality and disjoint universality. Several criteria are provided, and several applications to specific operators or sequences of operators are performed, mainly in the setting of sequence spaces or spaces of holomorphic functions. 1. Introduction In this paper, we are concerned with the phenomenon of simultaneous appro- ximation by the action of several operators or, more generally, by the action of several sequences of mappings. When the existence of a dense orbit under an operator is proved, we are speaking about universality or hypercyclicity, see below. In many situations, it is possible to show the existence of one vector whose orbits under two or more operators approximate any given vector. Pushing the question quite further, we wonder under what conditions such approximation takes place by using a common subsequence. This, together with its connection with other kinds of joint universality, will make up the main aim of the present manuscript. Next, we fix some related notation and terminology to be used in this work. For a good account of concepts, results and history concerning hypercyclicity, the reader is referred to the books [2, 21]. By N, N0, R, C, D, B(a, r), B(a, r) (a ∈ C, r > 0) we denote, respectively, the set of positive integers, the set N ∪ {0}, the real line, the complex plane, the open unit disk {z ∈ C : z < 1}, the open disk with center a and radius r, and the corresponding closed disk. Let X, Y be two Hausdorff topological spaces, and Tn : X → Y (n = 1, 2, . . . ) be a sequence of continuous mappings. Recall that (Tn) is said to be universal whenever there is some (Tn)-orbit which is dense in Y , that is, there exists an element x0 ∈ X -- called universal for (Tn) -- such that {Tnx0 : n ∈ N} = Y. Note that Y must be separable. We denote by U((Tn)) the set of universal ele- ments for (Tn). When X = Y and T : X → X is a continuous self-mapping, 2010 Mathematics Subject Classification. 30E10, 47B33, 47A16, 47B38. Key words and phrases. hypercyclic operator, composition operator, disjoint universality, si- multaneous universality. 1 2 BERNAL AND JUNG then T is called universal provided that the sequence (T n) of iterates of T (i.e., T 1 = T , T 2 = T ◦ T , T 3 = T ◦ T 2, and so on) is universal, in which case the set U((T n)) of universal elements will be denoted by U(T ). A sequence Tn : X → Y (n = 1, 2, . . . ) of continuous mappings is said to be densely universal if U((Tn)) is dense in X. Birkhoff's transitivity theorem asserts that, if X is a Baire space (in particular, if X is completely metrizable) and Y is second-countable (in particular, if X is metrizable and separable), then (Tn) is densely universal if and only if (Tn) is transitive (that is, given nonempty open sets U ⊂ X, V ⊂ Y , there is N ∈ N with TN (U) ∩ V 6= ∅); if this is the case, then U((Tn)) is residual (in fact, a dense Gδ subset) in X. If X lacks isolated points and T : X → X is universal, then U(T ) is dense in X (so residual if X is, in addition, completely metrizable). In the case in which X and Y are topological vector spaces over K (= R or C) and (Tn) ⊂ L(X, Y ) := {linear continuous mappings X → Y }, the words hypercyclic and universal are synonymous, although hypercyclic is mostly used, as well as the alternative notation HC((Tn)) := U((Tn)) (and HC(T ) := U(T ) for T ∈ L(X) := L(X, X) = {operators on X}). In particular, we have if X and Y are F-spaces with Y separable, then HC((Tn)) (HC(T ), with X separable, resp.) is residual in X as soon as (Tn) is transitive (as soon as T is hypercyclic, resp.). Recall that an F-space is a completely metrizable topological vector space. Assume now that X, Y are topological spaces, with X a Baire space and Y second-countable, and that Sn : X → Y and Tn : X → Y (n ∈ N) are densely universal sequences. Since U((Sn)), U((Tn)) are dense Gδ subsets of X, we have that U((Sn)) ∩ U((Tn)) is also dense, so non-empty. Hence there is a common hypercyclic element x ∈ X. So, for a given point y ∈ Y , there are sequences {n1 < n2 < · · · } and {m1 < m2 < · · · } in N such that Snj x → y and Tmj x → y as j → ∞. Then the following question arises naturally: Under what conditions on (Sn) and (Tn) one can guarantee the exis- tence of an element x ∈ X such that, for any given y ∈ Y , there is one sequence {n1 < n2 < · · · } ⊂ N such that Snj x −→ y ←− Tnj x as j → ∞? Of course, a similar question can be posed for finitely many sequences and for finitely many single operators on X, just by considering the sequences of their iterates in the latter case. With this in mind, the new concept of simultaneous universality will be introduced in the next section, and compared to other related notions existing in the literature, such as those of disjoint hypercyclicity and the weakly mixing property. Several sufficient conditions for simultaneous universal- ity/hypercyclicity will be provided in Section 3. Examples of finite families of simultaneous hypercyclic operators will be furnished in sections 4 -- 6, starting with SIMULTANEOUS UNIVERSALITY 3 multiples of an operator and ending up in the frameworks of sequence spaces and of spaces of analytic functions on complex domains. 2. Simultaneously universal sequences Let us define the new concept that is the matter of this paper. If p ∈ N and Y is a nonempty set, then by ∆(Y p) we denote the diagonal of Y p = Y × · · · × Y (p times), that is, the subset ∆(Y p) = {(y, y, . . . , y) : y ∈ Y }. If Y is a topological space, then Y p is assumed to be endowed with the product topology. Definition 2.1. Let p ∈ N and X, Y be Hausdorff topological spaces. Assume that, for each j ∈ {1, . . . , p}, Tj,n : X → Y (n ∈ N) is a sequence of continuous mappings. Consider the sequence [T1,n, . . . , Tp,n] : x ∈ X 7−→ (T1,nx, . . . , Tp,nx) ∈ Y p (n ∈ N). Let also T1, . . . , Tp : X → X be continuous mappings. (a) We say that the sequences (T1,n), . . . , (Tp,n) are simultaneously universal (or s-universal ) whenever there exists an element x0 ∈ X -- called s-universal for (T1,n), . . . , (Tp,n) -- satisfying {[T1,n, . . . , Tp,n]x0 : n ∈ N} ⊃ ∆(Y p). The set of such s-universal elements will be denoted by s-U((T1,n), . . . , (Tp,n)). (b) The sequences (T1,n), . . . , (Tp,n) are said to be densely simultaneously uni- versal if the set s-U((T1,n), . . . , (Tp,n)) is dense in X. And they are called hereditarily simultaneously universal (hereditarily densely simultaneously universal, resp.) if, for every strictly increasing sequence (nk) ⊂ N, the sequences (T1,nk), . . . , (Tp,nk) are s-universal (densely s-universal, resp.). (c) The mappings T1, . . . , Tp are called s-universal (densely s-universal, here- ditarily s-universal, hereditarily densely s-universal, resp.) if the sequences (T n p ) are s-universal (densely s-universal, hereditarily s-universal, hereditarily densely s-universal, resp.). The set s-U((T n p )) of cor- responding s-universal elements will be denoted by s-U(T1, . . . , Tp). 1 ), . . . , (T n 1 ), . . . , (T n Remarks 2.2. 1. If Y is first-countable (in particular, if Y is metrizable), then the s-simultaneous universality of (Tj,n)n∈N (1 ≤ j ≤ p) means the existence of some x0 ∈ X enjoying the property that, for every y ∈ Y , there is a (strictly increasing) sequence (nk) ⊂ N such that Tj,nkx0 → y as k → ∞ (j = 1, . . . , p). 2. In [18, Kapitel 1] the notion of relative universality on a closed subset of the arrival space is introduced under very general assumptions. In the present paper we study a special case of this situation (note that ∆(Y p) is closed in Y p since Y p is Hausdorff) under more specific hypotheses. 3. According to the introduction, if X, Y are topological vector spaces and Tj,n, Tj ∈ L(X, Y ) (j = 1, . . . , p; n ∈ N), then we use the expressions "s-hypercyclic", 4 BERNAL AND JUNG "densely s-hypercyclic" and "hereditarily densely s-hypercyclic" rather than "s- universal", "densely s-universal" and "hereditarily densely s-universal", respec- tively. s-U((T1,n), . . . , (Tp,n)) and s-HC(T1, . . . , Tp) := s-U(T1, . . . , Tp) in this case. 4. For a single operator T , hypercyclicity (hereditary hypercyclicity, resp.) equivalent to dense hypercyclicity (hereditary dense hypercyclicity, resp.). s-HC((T1,n), . . . , (Tp,n)) denote In addition, we will := is 5. The property of simultaneous universality of (T1,n), . . . , (Tp,n) is weaker than the property that the sequence ([T1,n, . . . , Tp,n]) is subspace-universal for ∆(Y p), meaning that the set {[T1,n, . . . , Tp,n]x0 : n ∈ N} ∩ ∆(Y p) is dense in ∆(Y p) for some x0 ∈ X (see e.g. [1,22,24] for results on subspace-hypercyclicity/universality). Before going on, we want to compare s-universality to other related concepts defined in the literature. In 2007, B`es, Peris and the first author ([11],[4]) intro- duced the notion of disjoint (or d-) universality (sometimes called d-hypercyclicity in the mentioned references). Under the same assumptions and terminology as in Definition 2.1, the sequences (T1,n), . . . , (Tp,n) are said to be d-universal whenever the sequence [T1,n . . . , Tp,n] : X → Y p (n ∈ N) is universal, that is, whenever there exists some x0 ∈ X such that the joint orbit {(T1,nx0, . . . , Tp,nx0) : n ∈ N} is dense in Y p. As a matter of fact, d-universality should not be confused with the universality of the sequence T1,n ⊕ · · · ⊕ Tp,n : (x1, . . . , xp) ∈ X p 7−→ (T1,nx1, . . . , Tp,nxp) ∈ Y p. Trivially, disjoint universality of (T1,n), . . . , (Tp,n) implies universality of the last sequence as well as simultaneous universality of (T1,n), . . . , (Tp,n). Also, trivially, s-universality implies the universality of each sequence (Tj,n)n∈N (j = 1, . . . , p) (in particular, Y must be separable). But no other implications among these properties hold, even considering only p = 2 and sequences of iterates of single operators. The following examples illustrate this situation: 1. Assume that T is a hypercyclic operator on a topological vector space. Then the operators T, T are s-hypercyclic but not d-hypercyclic. 2. In 1969, S. Rolewicz [26] proved that if c ∈ K has modulus > 1 and B is the backward shift (xn) ∈ ℓ2 7→ (xn+1) ∈ ℓ2, then the operator cB is hypercyclic. In particular, the operators T = 2B and S = 4B = 2T are hypercyclic, but T, S are clearly not s-hypercyclic. 3. Since each of the operators T, S of the latter example is mixing (see the definition at the beginning of the next section, regarding the sequences of iterates; see also [21, p. 46]), the operator T ⊕ S is hypercyclic, but T, S are not s-hypercyclic. 4. De la Rosa and Read [15] were able to construct a Banach space X and an operator T ∈ L(X) such that T is hypercyclic (hence T, T are s-hypercyclic) but T is not weakly mixing on X, meaning that T ⊕ T is not hypercyclic on X 2. SIMULTANEOUS UNIVERSALITY 5 While d-hypercyclic operators must be substantially different, s-hypercyclicity allows more similarity. For instance, an operator can never be d-hypercyclic with a scalar multiple of itself (see [11, p. 299]). Nevertheless, s-hypercyclicity is possible in concrete situations. This will be analyzed in Section 4. Sections 5 and 6 are devoted to more specific operators, namely backward shifts and operators on spaces of analytic functions. We close this section by establishing, under appropriate assumptions, the exis- tence of large vector subspaces consisting, except for zero, of s-hypercyclic vectors. Theorem 2.3. (a) Let X be a topological vector space and Tj ∈ L(X) (j = 1, . . . , p). If T1, . . . , Tp are s-hypercyclic and at least one of them commutes with the others, then s-HC(T1, . . . , Tp) contains, except for 0, a dense lin- ear subspace of X. (b) Let X and Y be two topological vector spaces such that Y is metrizable. Assume that (Tj,n) ⊂ L(X, Y ) (j = 1, . . . , p) are hereditarily s-hypercyclic sequences. Then s-HC((T1,n), . . . , (Tp,n)) contains, except for 0, an infinite dimensional vector subspace of X. (c) Let X and Y be two metrizable separable topological vector spaces. Assume that (Tj,n) ⊂ L(X, Y ) (j = 1, . . . , p) are hereditarily densely s-hypercyclic sequences. Then s-HC((T1,n), . . . , (Tp,n)) contains, except for 0, a dense linear subspace of X. Proof. (a) By hypothesis, there is i ∈ {1, . . . , p} such that TiTj = TjTi (j = 1, . . . , p). Therefore P (Ti)Tj = TjP (Ti) for all j and every polynomial P with coefficients in K. Let P denote the set of such polynomials. Of course, the operator Ti is hypercyclic. From a result by Wengenroth [29], the operator P (Ti) has dense range as soon as P ∈ P \ {0}. Pick any x0 ∈ s-HC(T1, . . . , Tp). Let us define M := {P (Ti)x0 : P ∈ P \ {0}}. Then M is a linear subspace of X. It is dense because M contains the orbit {T n i x0 : n ∈ N}, that is dense in X as x0 ∈ HC(Ti). It remains to show that M \ {0} ⊂ s-HC(T1, . . . , Tp). To this end, fix u ∈ M \ {0}. Then there is P ∈ P \ {0} such that u = P (Ti)x0. It must be proved that Z ⊃ ∆(X p), 1 u, . . . , T n p u) : n ∈ N} = {(P (Ti)T n where Z := {(T n p x0) : n ∈ N}, where the last equality follows from commutativity. We know that ∆(X p) ⊂ {(T n p x0) : n ∈ N}, ϕ := P (Ti) and Φ : X p → X p be the mapping defined as Φ(x1, . . . , xp) := (ϕ(x1), . . . , ϕ(xp)). Then, as ϕ is continuous, we get p x0) : n ∈ N}. Let A := {(T n 1 x0, . . . , P (Ti)T n 1 x0, . . . , T n 1 x0, . . . , T n Z = Φ(A) ⊃ Φ(A) ⊃ Φ(∆(X p)) = {(ϕ(x), . . . , ϕ(x)) : x ∈ X}, so Z ⊃ {(ϕ(x), . . . , ϕ(x)) : x ∈ X}. Given y ∈ X and a neighborhood U of (y, y, . . . , y), there exists a neighborhood V of y such that U ⊃ V p. Since ϕ 6 BERNAL AND JUNG has dense range, one can find x ∈ X with ϕ(x) ∈ V . Then (ϕ(x), . . . , ϕ(x)) ∈ In other words, (y, . . . , y) ∈ {(ϕ(x), . . . , ϕ(x)) : x ∈ X}, so (y, . . . , y) ∈ Z. U. Consequently, Z ⊃ ∆(X p), as required. (b) -- (c). By mimicking the proofs of Theorems 1 -- 2 of [3] (in which the results are given for a single sequence (Tn)), we can construct recursively a sequence (xN )N ∈N ⊂ X and a family {(q(N, k))k∈N : N ∈ N0} of strictly increasing subsequences of N satisfying, for all N ∈ N, the following conditions: xN ∈ GN ∩ s-HC((T1,q(N −1,k), . . . , (Tp,q(N −1,k))) and Tj,q(l,k)xN → 0 as k → ∞ for all l ≥ N and all j ∈ {1, . . . , p}, where G0 := X and GN := X \ span {x1, . . . , xN −1} (N ∈ N) if the assumptions of (b) hold, while {GN }N ∈N denotes any fixed open basis of X if the assumptions of (c) hold. Then M := span {xN : N ∈ N} is the sought-after vector subspace. The details are left as an exercise. (cid:3) 3. s-Universality criteria A number of workable sufficient conditions will be useful to detect s-universality. Recall that a sequence of continuous mappings Tn : X → Y (n ∈ N) is called mix- ing provided that,given nonempty open sets U ⊂ X, V ⊂ Y , there is N ∈ N such that Tn(U) ∩ V 6= ∅ for all n ≥ N. The corresponding notion of simultaneous mixing property arises naturally, as well as the one of simultaneous transitivity. Note that Tn(U) ∩ V 6= ∅ is equivalent to U ∩ T −1 n (V ) 6= ∅. Definition 3.1. Let p ∈ N and X, Y be Hausdorff topological spaces. Assume that, for each j ∈ {1, . . . , p}, Tj,n : X → Y (n ∈ N) is a sequence of continuous mappings. Let also T1, . . . , Tp : X → X be continuous mappings. We say that: (a) The sequences (T1,n), . . . , (Tp,n) are simultaneously transitive (or s-transitive) provided that, for every pair of nonempty open sets U ⊂ X, V ⊂ Y , there j=1 T −1 j,N (V ) 6= ∅. is N ∈ N such that U ∩Tp is N ∈ N such that U ∩Tp (b) The sequences (T1,n), . . . , (Tp,n) are simultaneously mixing (or s-mixing) provided that, for every pair of nonempty open sets U ⊂ X, V ⊂ Y , there j=1 T −1 j,n (V ) 6= ∅ for all n ≥ N. (c) The mappings T1, . . . , Tp are simultaneously transitive (simultaneously mix- p ) are s-transitive (s-mixing, 1 ), . . . , (T n ing, resp.) whenever the sequences (T n resp.). j=1 T −1 j=1 T −1 Remark 3.2. Corresponding concepts of d-transitivity and d-mixing were intro- j,n (Vj) (Vj nonempty open subsets of Y , j = 1, . . . , p) j,n (V ). Also, most criteria given in this section have their counterparts for the related d-properties as provided in [4] and [11]. A thorough study of d-mixing operators is provided in [8]. duced in [11], where Tp appears instead of Tp Note that, contrary to the one-sequence case, the facts U ∩Tp j=1 Tj,N (U) ∩ V 6= ∅ are not equivalent. Observe also that Tp and Tp j,N (V ) 6= ∅ j,n (V ) = j=1 T −1 j=1 T −1 SIMULTANEOUS UNIVERSALITY 7 [T1,n, . . . , Tp,n]−1(V p). From the definitions, it is easy to check that the sequences (T1,n), . . . , (Tp,n) are s-mixing if and only if, for every strictly increasing sequence (nk) in N, the sequences (T1,nk), . . . , (Tp,nk) are s-transitive. The following propo- sition provides what can be called the Birkhoff s-transitivity theorem. Proposition 3.3. Under the same assumptions and terminology as in Definition 3.1, let us suppose, in addition, that X is Baire and Y is second-countable. Then we have: (i) The sequences (T1,n), . . . , (Tp,n) are s-transitive if and only if they are densely s-universal. If this is the case, then the set s-U((T1,n), . . . , (Tp,n)) is residual in X. (ii) The sequences (T1,n), . . . , (Tp,n) are s-mixing if and only if, for every strictly increasing sequence (nk) ⊂ N, the sequences (T1,nk), . . . , (Tp,nk) are densely s-universal. Proof. Part (ii) is an immediate consequence of (i). Let us prove (i). Fix a countable open basis (Vm) of Y , as well as a point x0 ∈ X. Then x0 ∈ s- U((T1,n), . . . , (Tp,n)) if and only if, given a nonempty open set V ⊂ Y , there is j,n (V ). Since each V contains some Vm and each Vm is a nonempty subset of Y , the last property is the n ∈ N with [T1,n, . . . , Tp,n]x0 ∈ V p, that is, x0 ∈Sn∈NTp same as x0 ∈Tm∈NSn∈NTp j,n (Vm), which shows that j=1 T −1 j=1 T −1 s-U((T1,n), . . . , (Tp,n)) = \m∈N [n∈N p\j=1 T −1 j,n (Vm). (1) j=1 T −1 Since the Tj,n's are continuous, each setTp are s-transitive then every set Sn∈NTp j=1 T −1 j,n (Vm) is open. If (T1,n), . . . , (Tp,n) j,n (Vm) (m ∈ N) is (open and) dense. Hence their (countable) intersection, which equals s-U((T1,n), . . . , (Tp,n)) by (1), is a dense Gδ subset (so residual) in X because X is Baire. Conversely, assume that the set of s-universal elements is dense in X and fix a nonempty open subset V of j,n (Vm) j,n (V ) is also dense. But this means j,N (V ) 6= ∅ (cid:3) Y . Then there is m ∈ N with V ⊃ Vm. It follows from (1) thatSn∈NTp is dense in X, so the bigger set Sn∈NTp that, given a nonempty set U ⊂ X, there is N ∈ N such that U ∩Tp or, in other words, the sequences (T1,n), . . . , (Tp,n) are s-transitive. j=1 T −1 j=1 T −1 j=1 T −1 In the linear case, we state the following set of sufficient conditions, that are inspired by the results contained in [19, Sect. 1c] and the references cited in it. Theorem 3.4. Let X and Y be topological vector spaces such that X is Baire and Y is metrizable and separable, and let (Tj,n)n∈N (j = 1, . . . , p) be sequences in L(X, Y ). Assume that there are respective dense subsets X0 of X and Y0 of Y satisfying at least one of the following conditions: 8 BERNAL AND JUNG (A) For every pair of vectors x ∈ X0, y ∈ Y0, there exist sequences (nk) ⊂ N and (xk) ⊂ X with xk → 0, Tj,nkx → 0 and Tj,nkxk → y (j = 1, . . . , p) as k → ∞. (B) For every x ∈ X0, the sequences (Tj,nx)n∈N (j = 1, . . . , p) converge in Y to a common limit and, for every y ∈ Y0, there exist sequences (nk) ⊂ N and (xk) ⊂ X with xk → 0 and Tj,nkxk → y (j = 1, . . . , p) as k → ∞. (C) For every x ∈ X0, there exists a sequence (nk) ⊂ N such that the sequences (Tj,nkx)k∈N (j = 1, . . . , p) converge in Y to a common limit and, for every y ∈ Y0, there exists a sequence (xn) ⊂ X such that xn → 0 and Tj,nxn → y (j = 1, . . . , p) as n → ∞. Then (Tj,n)n∈N (j = 1, . . . , p) are densely s-hypercyclic. Proof. According to Proposition 3.3, we should show that (Tj,n)n∈N (j = 1, . . . , p) are s-transitive. With this aim, fix a pair of nonempty open sets U ⊂ X, V ⊂ Y . We should exhibit an N ∈ N such that U ∩Tp j=1 T −1 j,N (V ) 6= ∅. Assume first that (A) holds. By density, there are x ∈ X0 and y ∈ Y0 such that x ∈ U and y ∈ V . Define A := U − x and B := V − y. Then A and B are open neighborhoods of 0 in X and Y respectively. Take a 0-neighborhood C ⊂ Y satisfying C + C ⊂ B. Consider the sequences (nk) and (xk) provided by (A). Then there is k ∈ N such that xk ∈ A, Tj,nkx ∈ C and Tj,nkxk ∈ y + C (j = 1, . . . , p). Let u := x + xk and N := nk. We get u ∈ x + A = U and Tj,N u = Tj,N x + Tj,N xk ∈ C + y + C ⊂ y + B = V (j = 1, . . . , p), so that u ∈ U ∩Tp j=1 T −1 j,N (V ). Suppose now that (B) holds. By density, there is x ∈ X0 such that x ∈ U. Define A := U − x, a neighborhood of 0. By hypothesis, there is z ∈ Y such that Tj,n → z as n → ∞ (j = 1, . . . , p). Since Y0 is dense in Y , there is y ∈ Y0 with y ∈ z + V . Let B := V − y + z, a neighborhood of 0 in Y . Take a 0-neighborhood C ⊂ Y satisfying C + C ⊂ B. We have that Tj,n ∈ z + C (j = 1, . . . , p) for n ≥ n0, say. Consider the sequences (nk) and (xk) provided by (B) for the vector y − z, so that xk → 0 and Tj,nkxk → y − z (j = 1, . . . , p) as k → ∞. Choose k ∈ N so large that nk ≥ n0, xk ∈ A and Tj,nkxk ∈ y − z + C (j = 1, . . . , p). Let u := x + xk and N := nk. Then u ∈ x + A = U and, for every j = 1, . . . , p, Tj,N u = Tj,N x + Tj,N xk ∈ z + C + y − z + C = y + C + C ⊂ y + B = V, so that u ∈ U ∩Tp similar and left as an exercise. j=1 T −1 j,N (V ), as required. Under assumption (C), the proof is (cid:3) Two of the most popular criteria of hypercyclicity are the so-called blow-up/col- lapse criterion and the hypercyclicity criterion (see [2, 20, 21]). Now, we can obtain their respective s-versions. Proposition 3.5. [s-Blow-up/Collapse Criterion] Let X be a Baire metrizable separable topological vector space, and let (Tj,n)n∈N (j = 1, . . . , p) be sequences SIMULTANEOUS UNIVERSALITY 9 in L(X). Suppose that, for every nonempty open subsets U, V of X and every 0-neighborhood W ⊂ X there is N ∈ N such that W ∩ p\j=1 T −1 j,N (V ) 6= ∅ 6= U ∩ p\j=1 T −1 j,N (W ). Then (Tj,n)n∈N (j = 1, . . . , p) are densely s-hypercyclic. Proof. Fix a pair of nonempty open sets U, V ⊂ X. Choose vectors x ∈ U, y ∈ V . It suffices to exhibit sequences sequences (nk) ⊂ N and (xk) ⊂ X with xk → x and Tj,nkxk → y (j = 1, . . . , p), because this would entail the existence of some (V ). In other words, the sequences (Tj,n)n∈N (j = 1, . . . , p) would be s-transitive, hence densely s-hypercyclic by Proposition 3.3. k ∈ N such that xk ∈ U and Tj,nkxk ∈ V (j = 1, . . . , p), so xk ∈ U ∩Tp j=1 T −1 j,nk With this aim, choose a fundamental decreasing sequence (Wk) of 0-neighbor- hoods. Then (Uk) := (x + Wk) and (Vk) := (y + Wk) are fundamental decreasing sequences of x-neighborhoods and y-neighborhoods, respectively. By hypothesis, for each k ∈ N, there are nk ∈ N and points x′ k ∈ Wk ∩ k. Then xk → x k → x). Finally, k ∈ Vk and (cid:3) (Vk) and x′′ as k → ∞ because x′ Tj,nkxk = Tj,nkx′ Tj,nkx′′ k → y + 0 = y (j = 1, . . . , p) because Tj,nkx′ j=1 T −1 j,nk k → 0) and x′′ k ∈ Uk ∩Tp k ∈ Wk (so x′ k + Tj,nkx′′ k ∈ Wk for all k ∈ N. j=1 T −1 j,nk Tp k and x′′ (Wk). Let xk := x′ k such that x′ k + x′′ k ∈ Uk (so x′′ Recall that the convex hull conv(A) of a subset A of a vector space X is the least convex subset of X containing A. Definition 3.6. Let X be a Baire metrizable separable locally convex space, (nk) ⊂ N be a strictly increasing sequence and Tj ∈ L(X) (j = 1, . . . , p). We say that T1, . . . , Tp satisfy the s-hypercyclicity criterion with respect to (nk) if there are subsets X0 ⊂ X, W0 ⊂ X p such that X0 is dense in X and as well as mappings Rk : W0 → X (k ∈ N) such that W0 ⊃ ∆(X p) j → 0 pointwise on X0 as k → ∞ (j = 1, . . . , p), (i) T nk (ii) Rk → 0 pointwise on W0 as k → ∞ and (iii) For every w = (w1, . . . , wp) ∈ W0 and every j ∈ {1, . . . , p} there is yj ∈ conv({w1, . . . , wp}) such that T nk j Rkw → yj as k → ∞. Theorem 3.7. [s-Hypercyclicity Criterion] Let X be a Baire metrizable separa- ble locally convex space and Tj ∈ L(X) (j = 1, . . . , p). If T1, . . . , Tp satisfy the s-hypercyclicity criterion with respect to some (nk) ⊂ N, then (T nk p ) are s-mixing. In particular, T1, . . . , Tp are densely s-hypercyclic. 1 ), . . . , (T nk Proof. Let U, V ⊂ X be nonempty open sets. Then there are x0 ∈ U ∩ X0 and y0 ∈ V . By local convexity, there is a convex open set eV with y0 ∈ eV ⊂ V . As 10 BERNAL AND JUNG (y0, . . . , y0) ∈ ∆(X p) ⊂ W0, one can find w = (w1, · · · , wp) ∈ W0 such that wj ∈ eV for all j = 1, . . . , p. Put zk := x0 + Rkw (k ∈ N). Then, due to (ii), zk → x0 + 0 = x0 ∈ U as k → ∞. Moreover, for every j ∈ {1, . . . , p} we get thanks to (i) and (iii) that T nk j zk = T nk j x0 + T nk j Rkw −→ 0 + yj = yj as k → ∞, there is k0 ∈ N such that, for all k ≥ k0, we have zk ∈ U and T nk (V ) 6= ∅, as required. where, for each j, yj ∈ conv({w1, . . . , wp}) ⊂ conv(eV ) = eV ⊂ V . Consequently, (j = 1, . . . , p) or, in other words, zk ∈ U ∩Tp j zk ∈ V (cid:3) Remarks 3.8. 1. Examples of spaces X satisfying the assumptions of Theorem 3.7 are the Fr´echet spaces, that is, the locally convex F-spaces. If local convexity is dropped from the assumptions, then the conclusion still holds if we replace (iii) by the (stronger) condition: (iii') T nk j Rkw → wj as k → ∞ for every w = (w1, . . . , wp) ∈ W0 and every j ∈ {1, . . . , p}. j=1 T −nk j 2. In [11, Proposition 2.6] the following d-hypercyclicity criterion was proved, where X is a Fr´echet space, (nk) ⊂ N is a strictly increasing sequence and Tj ∈ L(X) (j = 1, . . . , p): Assume that there exist dense subsets X0, X1, . . . , Xp ⊂ X and mappings Sk,j : Xj → X (k ∈ N; 1 ≤ j ≤ p) satisfying T nk j → 0 (k → ∞) pointwise on X0, Sk,j → 0 (k → ∞) pointwise on Xj, and T nk j Sk,l → δj,lidXl (k → ∞) pointwise on Xl (1 ≤ j, l ≤ p). Then (T nk p ) are d-mixing (see [11, Definition 2.1]). In particular, by [11, Proposition 2.3], T1, . . . , Tp are densely d-hypercyclic. 1 ), . . . , (T nk Now, we can obtain a disjoint hypercyclicity criterion under weaker assump- tions. Namely, let us assume that there are dense subsets X0 ⊂ X, W0 ⊂ X p and mappings Rk : W0 → X (k ∈ N) satisfying (i) -- (ii) of Definition 3.6 together with (iii') of the preceding remark (it is easy to check that these assumptions are weaker than those of the d-hypercyclicity criterion in [11]). Then (T nk p ) are d- mixing. Indeed, let U, V1, . . . , Vp ⊂ X be nonempty open sets. By density, there are x0 ∈ U ∩ X0 and w = (w1, · · · , wp) ∈ W0 ∩ (V1 × · · · × Vp). Let zk := x0 + Rkw (k ∈ N). Then zk → x0 +0 = x0 ∈ U as k → ∞. Moreover, for every j ∈ {1, . . . , p} we get T nk j Rkw −→ 0 + wj = wj as k → ∞. Consequently, there is k0 ∈ N such that, for all k ≥ k0, we have zk ∈ U and T nk j zk ∈ Vj (j = 1, . . . , p), 1 ), . . . , (T nk j zk = T nk j x0 + T nk (Vj) 6= ∅, which is the d-mixing property. that is, zk ∈ U ∩Tp j=1 T −nk j 3. Several sets of conditions on T1, . . . , Tp such that these operators satisfy the s-hypercyclicity criterion with respect to a strictly increasing sequence (nk) ⊂ N are -- as it is easy to check -- the following: SIMULTANEOUS UNIVERSALITY 11 T nk l=1 Sk,l → idY0 pointwise on Y0 (j = 1, . . . , p). (a) There are dense subsets X0, Y0 ⊂ X and mappings Sk,j : Y0 → X (k ∈ j=1 Sk,j → 0 pointwise on Y0 and N; 1 ≤ j ≤ p) such that (i) holds, Pp j Pp N; 1 ≤ j ≤ p) such that (i) holds, Pp (b) There are subsets X0, X1, . . . , Xp ⊂ X in such a way that X0 is dense in X and X1 × · · · × Xp ⊃ ∆(X p) as well as mappings Sk,j : Xj → X (k ∈ j=1 Sk,jxj → 0 for all (x1, . . . , xp) ∈ l=1 Sk,lxl) → xj for all (x1, . . . , xp) ∈ X1 ×· · ·×Xp X1 ×· · ·×Xp, and T nk and all j = 1, . . . , p. j (Pp In view of (b), we see that if T1, . . . , Tp satisfy the d-hypercyclicity criterion with respect to (nk), then they also satisfy the s-hypercyclicity criterion with respect to (nk). B`es and Peris [10] have proved that satisfaction of the hypercyclicity criterion, hereditary hypercyclicity and transitivity of self-sums are equivalent (see also [5]). Moreover, they established a similar result for d-hypercyclicity [11, Theorem 2.7]. Now, we prove that a corresponding statement also holds for s-hypercyclicity, with the d-hypercyclicity criterion replaced by the s-hypercyclicity criterion (Theorem 3.7), so showing that the latter is rather natural. Proposition 3.9. Let X be a separable Fr´echet space and Tj ∈ L(X) (j = 1, . . . , p). Consider the following statements: (a) T1, . . . , Tp satisfy the s-hypercyclicity criterion. (b) (T nk (c) ⊕m (d) T1 ⊕ T1, . . . Tp ⊕ Tp are s-transitive on X 2. 1 ), . . . , (T nk k=1T1, · · · ⊕m p ) are hereditarily densely s-hypercyclic for some (nk) ⊂ N. k=1 Tp are s-transitive on X m for all m ∈ N. Then we have: (A) (a), (b) and (c) are equivalent. (B) If there exists i ∈ {1, . . . , p} such that TiTj = TjTi for all j ∈ {1, . . . , p}, then (a), (b), (c) are equivalent to (d). Proof. In the proof of (A), we follow closely the proof of Theorem 2.7 in [11], while the proof of (B) runs similar as the proof of Theorem 2.3, (3) ⇒ (1), in [10]. (A) (a) ⇒ (b): T1, . . . , Tp satisfy the s-hypercyclicity criterion with respect to some (nk) ⊂ N, so that they also satisfy it for any subsequence (mk) of (nk). By Theorem 3.7, (T mk ) are s-mixing and therefore densely s-hypercyclic. ), . . . , (T mk 1 p (b) ⇒ (c): Let m ∈ N be fixed and let ∅ 6= Ul, Vl ⊂ X open (l = 1, . . . , m). It suffices to show that there exists N ∈ N such that Ul ∩ p\j=1 T −N j (Vl) 6= ∅ for all l = 1, . . . , m. (1) Since (T mk p the sequences (T nk ), . . . , (T mk 1 ) are densely s-hypercyclic for each subsequence (mk) of (nk), p ) are s-mixing (cf. Proposition 3.3(ii)). Hence, for 1 ), . . . , (T nk 12 BERNAL AND JUNG each l ∈ {1, . . . , m}, there exists k0(l) ∈ N such that Ul ∩ Tp (Vl) 6= ∅ for all k ≥ k0(l). Then (1) is satisfied simply by choosing N := max{k0(1), . . . , k0(m)}. (c) ⇒ (a): Due to the assumption, we have: j=1 T −nk j (∗) For every m ∈ N and every 2m-tuple U1, . . . , Um, V1, . . . , Vm of nonempty open subsets of X there is N ∈ N arbitrarily large such that (1) holds. Let (An)n∈N, (Bn)n∈N be bases of nonempty sets of the topology of X. For n ∈ N, we write Wn := B(0, 1/n) (open d-balls, d being a translation-invariant distance generating the topology of X) and An,0 := An, Bn,0 := Bn. j=1 T −n1 j j=1 T −n1 j to (∗) (with m = 2), there is n1 ∈ N such that B1 ∩Tp W1 ∩ Tp Choose a nonempty open set A1,1 with diam(A1,1) < 1/2 and A1,1 ⊂ A1. Due (W1) 6= ∅ and (A1,1) 6= ∅. Thus, there exist a nonempty open set B1,1 with diam(B1,1) < 1/2, B1,1 ⊂ B1 and T n1 j (B1,1) ⊂ W1 for all j = 1, . . . , p, as well as a point w1,1 ∈ W1 with T n1 j w1,1 ∈ A1,1 for all j = 1, . . . , p. Now, for i = 1, 2, choose Ai,3−i open, nonempty, such that diam(Ai,3−i) < 1/3, Ai,3−i ⊂ Ai,2−i and A1,2 ∩ A2,1 = ∅. Due to (∗) (with m = 4), there is n2 ∈ N with n2 > n1 such that (Ai,3−i) 6= ∅ for i = 1, 2. Thus, there exist nonempty open sets Bi,3−i with diam(Bi,3−i) < 1/3, Bi,3−i ⊂ Bi,2−i and T n2 j (Bi,3−i) ⊂ W2 (i = 1, 2) for all j = 1, . . . , p as well as points wi,3−i ∈ W2 (i = 1, 2) with T n2 (W2) 6= ∅ and W2 ∩Tp j wi,3−i ∈ Ai,3−i (i = 1, 2) for all j = 1, . . . , p. Bi,2−i ∩Tp j=1 T −n2 j=1 T −n2 j j Continuing this process inductively, by using (∗) with m = 2k in step k, we ob- tain a strictly increasing sequence (nk) ⊂ N, nonempty open sets Ai,k+1−i, Bi,k+1−i with diam(Ai,k+1−i) < 1 k+1 and points wi,k+1−i ∈ Wk (1 ≤ i ≤ k; k ∈ N) such that k+1, diam(Bi,k+1−i) < 1 (i) Ai,k+1−i ⊂ Ai,k−i, Bi,k+1−i ⊂ Bi,k−i for all 1 ≤ i ≤ k, k ∈ N. (ii) For each k ∈ N, the sets Ai,k+1−i, 1 ≤ i ≤ k, are pairwise disjoint. (iii) T nk (iv) T nk j (Bi,k+1−i) ⊂ Wk (k ∈ N; 1 ≤ i ≤ k; 1 ≤ j ≤ p), and j wi,k+1−i ∈ Ai,k+1−i (k ∈ N; 1 ≤ i ≤ k; 1 ≤ j ≤ p). For each fixed i ∈ N, the sequences of closed sets (Ai,r)r∈N and (Bi,r)r∈N are decreasing (due to (i)) with diam(Ai,r), diam(Bi,r) < 1 r+i. The completeness of X implies the existence of points ai, bi ∈ X (i ∈ N) such that Tr∈N Ai,r = {ai} and Tr∈N Bi,r = {bi}. Put X0 := {bi : i ∈ N} ⊂ X and W0 := {ai : i ∈ N}p ⊂ X p. As ai ∈ Ai,1 ⊂ Ai,0 = Ai and bi ∈ Bi,1 ⊂ Bi,0 = Bi for all i ∈ N, we obtain that X0 is dense in X and W0 is dense in X p. Due to (ii), we have that ai 6= ak whenever i 6= k (indeed, if i < k, say, then ai ∈ Ai,k+1−i and ak ∈ Ak,1, but Ai,k+1−i ∩ Ak,1 = ∅). Hence, SIMULTANEOUS UNIVERSALITY 13 for each k ∈ N, the function Rk : W0 → X given by Rk(ai1, . . . , aip) =  is well defined. Altogether, we have: pXl=1 1 p 0 wil,k+1−il if k ≥ max l=1,...,p il otherwise pPp pPp • For all j = 1, . . . , p, all i ∈ N and all k ≥ i one has, due to (iii), that j → 0 (k → ∞) pointwise j (Bi,k+1−i) ⊂ Wk = B(0, 1/k), so T nk T nk j bi ∈ T nk on X0 for every j = 1, . . . , p. • For every (ai1, . . . , aip) ∈ W0 and every k ≥ maxl=1,...,p il, one has l=1 wil,k+1−il → 0 (k → ∞), because wil,k+1−il ∈ Rk(ai1, . . . , aip) = 1 Wk = B(0, 1/k). Therefore Rk → 0 (k → ∞) pointwise on W0. • For all j = 1, . . . , p, all (ai1, . . . , aip) ∈ W0 and all k ≥ maxl=1,...,p il, we get T nk j Rk(ai1, . . . , aip) = 1 j wil,k+1−il ∈ Ail,k+1−il and the sequence of sets Ail,k+1−il (k ∈ N) collapses to the singleton {ail} as k → ∞ for each l, we get T nk l=1 ail ∈ conv({ai1, . . . , aip}) as k → ∞. j wil,k+1−il. Since T nk j Rk(ai1, . . . , aip) → 1 pPp l=1 T nk Thus, T1, . . . , Tp satisfy the s-hypercyclicity criterion with respect to (nk). The proof of (A) is finished. (B) Obviously, (c) always implies (d). Assume now that (d) holds and that some Ti commutes with all Tj's. Our goal is to prove that (a) is satisfied. Let us fix any vector (x0, y0) ∈ s-HC(T1 ⊕ T1, . . . , Tp ⊕ Tp). We claim that, for each m ∈ N, the vector (x0, T m i y0) is also s-hypercyclic for T1 ⊕ T1, . . . , Tp ⊕ Tp. Indeed, as Ti is hypercyclic, it has dense range, from which one obtains, inductively, i (X) is dense in X. Put A := X × T m that every set T m i (X), so that A is dense in X 2. Given (u, v) ∈ A there is w ∈ X such that v = T m i w. By s-hypercyclicity, there exists (nk) ⊂ N such that T nk j y0 → w (k → ∞) for all j = 1, . . . , p. Hence, for all j, T nk j x0 → u and, by commutativity together with continuity of T m j (T m i y0) = T m i w = v (k → ∞). i i y0) : n ∈ N} ⊃ ∆(Ap). Since Therefore Σ := {[(T1 ⊕ T1)n, . . . , (Tp ⊕ Tp)n](x0, T m ∆(Ap) ⊃ ∆((X 2)p) and Σ is closed, we get Σ ⊃ ∆((X 2)p), which proves the claim. j x0 → u and T nk j y0) −→ T m , we get T nk i (T nk In particular, as y0 is hypercyclic for Ti, for each nonempty open set U ⊂ X there exists some u ∈ U such that (x0, u) is s-hypercyclic for T1 ⊕ T1, . . . , Tp ⊕ Tp. Thus, fixing a decreasing basis (Uk) of neighborhoods of 0 and using induction, we can find for each k ∈ N some uk ∈ Uk and nk ∈ N with nk > nk−1 (where n0 := 0) such that (α) T nk (β) T nk j x0 ∈ Uk for all j = 1, . . . , p and j uk ∈ x0 + Uk for all j = 1, . . . , p. 14 BERNAL AND JUNG i x0 : n ∈ N} and W0 := X p We define X0 := {T n 0 . Note that X0 is dense in X as x0 is Ti-hypercyclic, so W0 is dense in X p (hence W0 ⊃ ∆(X p)). Now, observe that no orbit of any hypercyclic vector can be finite, that is, T m i x0 if m 6= n. Thus, for each k ∈ N, the mapping i x0 6= T n Rk : (T m1 i x0, . . . , T mp i x0) ∈ W0 7−→ 1 p · pXl=1 T ml i uk ∈ X (1) is well defined. We have: (i) For every j = 1, . . . , p and every m ∈ N, T nk j x0) → T m i 0 = 0 (k → ∞), where commutativity and continuity of Ti together with property (α) have been used. This shows that T nk j → 0 pointwise on X0 for all j = 1, . . . , p. i x0) = T m i (T nk j (T m (ii) From the continuity of each T ml it follows that T ml from (1) that Rk → 0 pointwise on W0. and the fact that uk ∈ Uk (hence uk → 0), i uk → 0 (k → ∞) for every l = 1, . . . , p. Then one derives i (iii) For every j = 1, . . . , p and every (m1, . . . , mp) ∈ Np, it follows from (1) that T nk j Rk(T m1 i x0, . . . , T mp i x0) = 1 p · pXl=1 T nk j T ml i uk = 1 p · pXl=1 T ml i T nk j uk −→ 1 p · pXl=1 T ml i x0 ∈ conv({T m1 i x0, . . . , T mp i x0}) as k → ∞, because of (β) (which implies T nk commutativity and continuity of each T ml . i j uk → x0) together with the This tells us that T1, . . . , Tp satisfy the s-hypercyclicity criterion, as required. (cid:3) We raise here the question whether (d) is equivalent to (a)-(b)-(c) without as- suming any commutativity. 4. Scalar multiples of an operator We start by studying s-hypercyclicity of scalar multiples of one operator. We have already pointed out that there is no chance of d-hypercyclicity in this case. Recall that an operator T on a topological vector space X is called hereditarily hypercyclic whenever (T nk) is universal for every strictly increasing sequence (nk) ⊂ N. It is well known -- and easy to see -- that, if X is an F-space and T ∈ L(X), then T is hereditarily hypercyclic if and only if T is mixing. Proposition 4.1. Let X be a topological vector space, p ∈ N, c1, . . . , cp ∈ K and T, T1, . . . , Tp ∈ L(X). We have: SIMULTANEOUS UNIVERSALITY 15 (a) Assume that X is metrizable and locally convex. If T, c1T, . . . , cpT are s-hypercyclic then the cj's are unimodular, that is, c1 = · · · = cp = 1. (b) Suppose that X is metrizable. If T ∈ L(X) is hereditarily hypercyclic and the scalars cj are unimodular, then T, c1T, . . . , cpT are densely s- hypercyclic. Proof. (a) Assume that T, c1T, . . . , cpT are s-hypercyclic, and fix j ∈ {1, . . . , p}. Let c := cj. Then T, cT are s-hypercyclic, so there is x0 ∈ s-HC(T, cT ). Since X is metrizable, we can find a sequence (nk) ⊂ N such that T nkx0 → x0 and cnkT nkx0 → x0 as k → ∞. Of course, x0 6= 0. But X is locally convex, so its topology is defined by a separating family of seminorms. Therefore there is a continuous seminorm q on X such that q(x0) > 0. Consider the sequence of vectors uk := (cnk − 1)T nkx0 (k ∈ N). q(T nk x0) → 0 On the one hand, we have uk = cnkT nkx0 − T nkx0 → x0 − x0 = 0, so q(uk) → 0 by the continuity of q. On the other hand, we get q(uk) = cnk − 1q(T nkx0), hence cnk − 1 = q(uk) q(x0) = 0. Therefore cnk → 1 as k → ∞, which implies c = 1, that is, cj = 1, as required. (b) The result is trivial if K = R (for cj = ±1, so (cjT )2n = T 2n for all n and all j = 1, . . . , p). The complex case K = C is more delicate. Recall that a subset E ⊂ T := {z = 1} is said to be a Dirichlet set provided that there is a strictly increasing sequence (nk) ⊂ N such that supz∈E znk − 1 → 0 as k → ∞. It is well-known that every finite subset of T is Dirichlet (see [13, Theorem 8.138(a)]). In particular, there exists (nk) ⊂ N strictly increasing such that cnk j → 1 (j = 1, . . . , p). According to the hypothesis, we may take x0 ∈ HC((T nk)). Given x ∈ X, there is a subsequence (mk) ⊂ (nk) such that T mkx0 → x and, of course, cmk j → 1 (j = 1, . . . , p) as k → ∞. Therefore, we obtain (cjT )mk x0 → x for all j = 1, . . . , p and hence HC((T nk)) ⊂ s-HC(T, c1T, . . . , cpT ). But HC((T nk)) is dense, so s-HC(T, c1T, . . . , cpT ) also is. (cid:3) Remarks 4.2. 1. In part (b) of the last proposition, hereditary hypercyclicity is needed in order to obtain common subsequences (nk) to perform approximations. If this is not claimed, then, by a result due to Le´on and Muller, any unimodular multiple of a hypercyclic operator on any topological vector space is always hyper- cyclic, even with the same set of hypercyclic vectors (see [23] and [21, pp. 339 -- 340]). 2. It is known that the d-mixing property of T1, . . . , Tp implies that c1T1, . . . , cpTp are also d-mixing for all unimodular scalars c1, . . . , cp (cf. [7, Remark 24(i)]). How- ever, the corresponding result in case of s-mixing operators does not hold in gen- eral. Indeed, for a mixing operator T , the pair T, T is clearly s-mixing, but T, −T are not s-mixing any more. To see this, assume that T, −T are s-mixing. Then Proposition 3.3(ii) would imply that (T nk), ((−T )nk) are densely s-universal for each strictly increasing sequence (nk) in N -- but s-universality of the sequences 16 BERNAL AND JUNG (T 2k+1) and ((−T )2k+1) = (−T 2k+1) is clearly not possible. In connection with this, it is stated in [7, Remark 24(ii)] and actually proved in [28, Proposition 4.9] that in case of unimodular scalars c1, . . . , cp every d-hypercyclic vector x0 for T1, . . . , Tp is also d-hypercyclic for c1T1, . . . , cpTp. The proof uses crucially the fact that such a vector x0 satisfies (x0, . . . , x0) ∈ HC(T1 ⊕ · · · ⊕ Tp). Thus, it cannot be adapted for s-hypercyclicity. Hence, we pose the question: Does the equality s-HC(T1, . . . , Tp) = s-HC(c1T1, . . . , cpTp) hold? 3. Concerning again part (b) and regarding its proof, we may obtain a much stronger result in the case K = C and X a Banach space. Recall that a nonempty subset E ⊂ C is said to be perfect if it is closed and each point of E is an accumulation point of E. In particular, every perfect set is un- countable. It is well known (see [13, Theorem 8.138(b)]) that there are perfect Dirichlet subsets of T. We have that if E ⊂ T is a perfect Dirichlet set and T ∈ L(X) is mixing, then the uncountable family of rotations {cT : c ∈ E ∪ {1}} is densely uniformly s-hypercyclic, in the sense that there is a dense set of vectors x0 ∈ X satisfying the following: for every y ∈ X there is (nk) ⊂ N such that limk→∞ supc∈E∪{1} k(cT )nkx0 − yk = 0. Indeed, we can take a sequence (mk) ⊂ N such that supc∈E∪{1} cmk − 1 = supc∈E cmk − 1 → 0 as k → ∞. As T is mixing, the set HC((T mk)) is dense. If x0 ∈ HC((T mk)), then there is a subse- quence (nk) ⊂ (mk) with T mk x0 → y. The conclusion follows from the inequality k(cT )nkx0 − yk ≤ kcnk(T nkx0 − y)k + k(cnk − 1)yk. 4. Proposition 4.1 furnishes examples of pairs of operators -- on spaces of sequences or of holomorphic functions (see sections 5 -- 6) -- that are s-hypercyclic but not d- hypercyclic: the multiples 2B, −2B of the backward shift B on ℓq (1 ≤ q < ∞) or c0; D, −D on H(C) (Df := f ′); Cϕ, −Cϕ on H(G), where Cϕf := f ◦ ϕ, G ⊂ C is a simply connected domain and ϕ is a run-away automorphism of G. 5. Backward shifts and s-hypercyclicity In this section, we consider the sequence spaces c0 and ℓq (1 ≤ q < ∞) over K = R or C. If a = (an)n∈N is a bounded sequence in K \ {0}, then Ba will denote the weighted backward shift Ba : (x0, x1, x2, . . . ) ∈ X 7→ (a1x1, a2x2, . . . ) ∈ X on X = c0 or ℓq. The unweighted backward shift B is B = Ba, where a = (1, 1, 1, . . . ). Salas characterized the hypercyclicity of Ba in terms of the weight sequence a. B`es and Peris [11, Theorem 4.1] did the same for the d-hypercyclicity of different powers of Ba. This characterization happens to hold also for s- hypercyclicity. Proposition 5.1. Let X = c0 or ℓq (1 ≤ q < ∞), p ≥ 2 and let r1, . . . , rp ∈ N with r1 < r2 < · · · < rp be given. For each l ∈ {1, . . . , p}, let al = (al,n)n∈N be a weight sequence. Then the following are equivalent: SIMULTANEOUS UNIVERSALITY 17 a1, . . . , Brp a1, . . . , Brp al,j+1 ··· al,j+rlm ap are d-hypercyclic. ap are s-hypercyclic. (i) Br1 (ii) Br1 (iii) For every M > 0 and every k ∈ N there is m ∈ N satisfying, for each j ∈ {0, 1, . . . , k}, that al,j+1 · · · al,j+rlm > M (1 ≤ l ≤ p) and as,j+(rl−rs)m+1 ··· as,j+rlm > M (1 ≤ s < l ≤ p). a1, . . . , Brp a1, . . . , Brp ap satisfy the d-hypercyclicity criterion. ap satisfy the s-hypercyclicity criterion. (iv) Br1 (v) Br1 Proof. The equivalence of (i), (iii) and (iv) is proved in [11, Theorem 4.1]. That (i) implies (ii) is trivial. Moreover, (ii) ⇒ (iii) is proved in fact in the proof of "(a) ⇒ (b)" of the same reference, since only the simultaneous approximation of one vector (namely e0 + · · · + eq) is used. Finally, we clearly have (iv) ⇒(v) ⇒ (ii). (cid:3) Remarks 5.2. 1. An analogous result about equivalence of d- and s-hypercyclicity also works for powers of weighted bilateral shifts (see Theorem 4.7 of [11] and its proof). a, . . . , Bp a ⊕ · · · ⊕ Bp 2. Corollary 4.4 in [11] also works with just s-universality, as it is a consequence of Theorem 4.1 there. In particular, we have that Ba, B2 a are s-hypercyclic on X if and only if Ba ⊕ B2 a is hypercyclic on X p. B`es, Martin and Peris [7, p. 855] constructed an operator T := Ba on ℓ2 such that T is hypercyclic but T ⊕ T 2 is not hypercyclic on ℓ2 ⊕ ℓ2, so that T, T 2 is not d-hypercyclic on ℓ2. Then we obtain that T, T 2 are even not s-hypercyclic. According to [21, Theorem 4.8], the mentioned T = Ba is not mixing. In [8, Sect. 3], a mixing operator T ∈ L(ℓ2) for which T, T 2 are not d-mixing is exhibited. But the existence of a mixing T on a separable Banach space such that T, T 2 are not d-hypercyclic is unknown so far [8, Question 3.7]. A more delicate question arises when r1 ≤ r2 ≤ · · · ≤ rp. In [11, Corol- lary 4.2], the following is proved for weighted powers of the unweighted backward shift: if p ≥ 2 and rl ∈ N, λl ∈ K (1 ≤ l ≤ p) with r1 ≤ r2 ≤ · · · ≤ rp, then λ1Br1, . . . , λpBrp are d-hypercyclic if and only if r1 < r2 < · · · < rp and 1 < λ1 < λ2 < · · · < λp. The following result shows that s-hypercyclicity is possible under slightly weaker assumptions. Proposition 5.3. Let p ≥ 2, and let rl ∈ N, λl ∈ K (1 ≤ l ≤ p) with r1 ≤ r2 ≤ · · · ≤ rp. Let A denote the set A := {j ∈ {1, . . . , p − 1} : rj = rj+1} and consider the conditions (i) 1 < λj for all j ∈ {1, . . . , p}, (ii) λj < λj+1 for all j ∈ {1, . . . , p − 1} \ A, (iii) λj = λj+1 for all j ∈ A. Then λ1Br1, . . . , λpBrp are s-hypercyclic on X = c0 or ℓq (1 ≤ q < ∞) if and only if (i),(ii) and (iii) hold. 18 BERNAL AND JUNG Proof. First, suppose that conditions (i),(ii) and (iii) hold. We write {1, . . . , p}\A = {t1, . . . , td}, with d ∈ N and t1 < · · · < td. As the set {λi/λj : i, j ∈ {1, . . . , p} with λi = λj} ⊂ T is finite, it is a Dirichlet set. Hence there exists a strictly increasing sequence (nk) ⊂ N such that → 1 (k → ∞) for all i, j ∈ {1, . . . , p} with λi = λj. (1) (cid:18) λi λj(cid:19)nk Consider the set X0 of finite sequences, that is, X0 := c00 = {x = (xn) ∈ X : exists n0 = n0(x) ∈ N such that xn = 0 for all n ≥ n0}. Then X0 is dense in X. If we set W0 := ∆(X p 0 ) ⊃ ∆(X p) because X0 is dense in X. Now, we set Tj := λjBrj (j = 1, . . . , p). Define, for each k ∈ N, the mapping Rk : W0 → X as follows. If x = (x1, x2, . . . , xN , 0, 0, 0, . . . ) ∈ X0 and w = (x, x, . . . , x), then 0 ) ⊂ X p, then W0 = ∆(X p (0, 0, 0, . . . ) Rkw =(cid:26) (01, u1, 02, u2, . . . , 0N , uN , 0, 0, 0, . . . ) xN(cid:1) (l ≥ 1). We have: x1, . . . , 1 nk λ tl if l ≥ 2 and ul :=(cid:0) 1 λ nk tl if nk ≥ N if nk < N, where 01 := (0, 0, . . . , 0) [rt1nk times], 0l := (0, 0, . . . , 0) [(rtl − rtl−1)nk − N times] (a) For each j ∈ {1, . . . , p} and each x = (x1, x2, . . . , xN , 0, 0, 0, . . . ) ∈ X0, T nk j x = 0 as soon as rjnk > N, so T nk j → 0 (k → ∞) pointwise on X0. (b) For every w = (x, . . . , x) ∈ W0 as before, the definition of Rk together with (i) yields Rkw → 0 as k → ∞. (c) Fix w = (x, . . . , x) ∈ W0, where x = (x1, x2, . . . , xN , 0, 0, 0, . . . ). For every j ∈ {1, . . . , p} there is exactly one l ∈ {1, . . . , d} such that λj = λtl, due to (ii) and (iii). Finally, if nk ≥ N, we have T nk j Rkw =(cid:18)(cid:0) λj λtl(cid:1)nk x1,(cid:0) λj (cid:0) λj λtl+1(cid:1)nk x1,(cid:0) λj λtd(cid:1)nk x1,(cid:0) λj (cid:0) λj λtl(cid:1)nkx2, . . . ,(cid:0) λj λtl+1(cid:1)nkx2, . . . ,(cid:0) λj λtd(cid:1)nkx2, . . . ,(cid:0) λj λtl(cid:1)nk xN , 0, 0, . . . , 0, λtl+1(cid:1)nkxN , 0, 0, . . . , 0, . . . , λtd(cid:1)nk xN , 0, 0, 0, 0, . . .(cid:19). It follows from (ii) that ( λj λts all ν ∈ {1, . . . , N}, while (1) entails that ( λj λtl ν ∈ {1, . . . , N}. Consequently, T nk )nkxν → 0 as k → ∞ for all s ∈ {l+1, . . . , d} and )nkxν → xν as k → ∞ for all j Rkw → (x1, x2, . . . , xN , 0, 0, 0, . . . ) = x. An application of the s-hypercyclicity criterion (see also Remark 3.8.1) concludes the first part of the proof. Now, suppose that λ1Br1, . . . , λpBrp are s-hypercyclic. Since hypercyclic opera- tors on normed spaces have norm larger than 1, we obtain 1 < kλjBrj k = λjkBrj k = λj SIMULTANEOUS UNIVERSALITY 19 for all j = 1, . . . , p (cf. the proof of Corollary 4.2 in [11]), i.e. condition (i) holds. For each j ∈ {1, . . . , p − 1}\A, we have rj < rj+1. Hence, as λjBrj , λj+1Brj+1 are s-hypercyclic, Proposition 5.1, (ii) ⇒ (iii), and the same approach as in the proof of Corollary 4.2 in [11] yield λj < λj+1, i.e. condition (ii) holds. Finally, for each j ∈ A, we have rj = rj+1. Hence, the s-hypercyclicity of λjBrj , λj+1Brj+1 = λj+1 λj · λjBrj implies λj+1/λj = 1 (see Proposition 4.1(a)) and thus λj = λj+1, i.e. condition (iii) holds. (cid:3) For instance, the operators 2B, 3B2, −3B2, being not d-hypercyclic, are s- hypercyclic. Further study of d-hypercyclicity of weighted unilateral and bilateral backward shifts can be found in [9]. 6. s-hypercyclicity in spaces of holomorphic functions Let G ⊂ C be a domain, that is, a nonempty connected open subset of C. We endow the space H(G) of all holomorphic (or analytic) functions G → C with the topology of uniform convergence on compacta, so that H(G) becomes a separable Fr´echet space. In this section we are concerned with s-hypercyclicity of finite sets of operators on H(G) (or on subspaces of it) for certain domains G. Recall that if X is a topological vector space and T ∈ L(X), then T is said to be supercyclic provided that there exists some x0 ∈ X whose projective orbit {λT nx0 : n ∈ N, λ ∈ K} is dense in X. If T1, . . . , Tp ∈ L(X), they are called d- supercyclic (see [7]) if there is x0 ∈ X such that {λ[T n p ]x0 : n ∈ N, λ ∈ K} is dense in X p. Consistently, we say that T1, . . . , Tp are s-supercyclic whenever {λ[T n p ]x0 : n ∈ N, λ ∈ K} ⊃ ∆(X p). 1 , . . . , T n 1 , . . . , T n Let LF T (D) denote the family of all linear fractional transformations ϕ(z) = az+b cz+d of the complex plane such that ϕ(D) ⊂ D. The subfamily Aut(D) of automor- phisms of D consists of all onto members of LF T (D). See e.g. [27, Chapter 1] for terminology related to these families. If ν ∈ R, then Sν denotes the weighted Hardy space Sν = {f (z) = Pn≥0 anzn ∈ H(D) : kf k := (Pn≥0 an2(n + 1)2ν)1/2 < ∞}. Each Sν is a Hilbert space, and the choices ν = −1/2, 0, 1/2 correspond, respecti- vely, to the classical Bergman, Hardy and Dirichlet spaces. Thanks to the results in [7], we obtain without effort the next two assertions. Proposition 6.1. Let ϕ1, . . . , ϕp ∈ LF T (D) pairwise distinct. Then the following are equivalent: (a) Cϕ1, . . . , Cϕp are s-supercyclic on H(D). (b) µ1Cϕ1, . . . , µpCϕp are s-mixing on H(D) for all nonzero scalars µ1, . . . , µp. (c) Cϕ1, . . . , Cϕp are d-supercyclic on H(D). (d) µ1Cϕ1, . . . , µpCϕp are d-mixing on H(D) for all nonzero scalars µ1, . . . , µp. 20 BERNAL AND JUNG (e) ϕ1 . . . , ϕp have no fixed point in D, and satisfy that if any two ϕl, ϕj have j(α) < 1 is not possible. the same attractive fixed point α, then ϕ′ l(α) = ϕ′ Proof. The equivalence of (c), (d) and (e) is proved in [7, Theorem 4]. The im- plications (d) ⇒ (b) ⇒ (a) are trivial. Finally, (a) ⇒ (e) is proved in fact in the proof of Theorem 4 in [7]. Indeed, it is used there a result (Lemma 14 in [7]) as- serting that if ϕ1, ϕ2 ∈ LF T (D) are hyperbolic and share an attractive fixed point α with ϕ′ 2(α), then Cϕ1, Cϕ2 are not d-supercyclic on H(D). But a closer look at its proof shows that Cϕ1, Cϕ2 are in fact even not s-supercyclic; indeed, via contradiction, only one function g is assumed to be simultaneously approximated by projective orbits. (cid:3) 1(α) = ϕ′ Proposition 6.2. Let ϕ1, . . . , ϕp ∈ LF T (D) pairwise distinct and let ν < 1/2. Then the following are equivalent: (a) Cϕ1, . . . , Cϕp are s-supercyclic on Sν. (b) Cϕ1, . . . , Cϕp are s-mixing on Sν. (c) Cϕ1, . . . , Cϕp are d-supercyclic on Sν. (d) Cϕ1, . . . , Cϕp are d-mixing on Sν. (e) Each ϕl is a parabolic automorphism or a hyperbolic map without fixed points in D, and there are no two ϕl, ϕj having a common fixed point α such that ϕ′ l(α) = ϕ′ j(α) < 1. Proof. The equivalence of (c), (d) and (e) is proved in [7, Theorem 3]. The impli- cations (d) ⇒ (b) ⇒ (a) are trivial. As for (a) ⇒ (e), observe that in the proof of Theorem 3 in [7], only the supercyclicity of each Cϕl is necessary for the first assertion in (e) and that the Comparison Principle [7, Proposition 8] -- that also works for s-supercyclicity -- implies that Cϕ1, . . . , Cϕp are s-supercyclic on H(D). Now, the second assertion of (e) follows from Proposition 6.1. (cid:3) Remarks 6.3. 1. Recall that if X is an F-space and T ∈ L(X) is invertible and hypercyclic, then T −1 is also hypercyclic. Analogously as in Example 22 in [7], by combining the preceding two propositions, we obtain that there are hyperbolic ϕ1, ϕ2 ∈ Aut(D) such that Cϕ1, Cϕ2 are d-hypercyclic (so s-hypercyclic) = (Cϕ2)−1 on H 2(D) (the Hardy space) and on H(D), and Cϕ−1 are even not s-supercyclic on H 2(D) or H(D) (note that ϕ−1 are also 1 hyperbolic). Hence, in general, the d-hypercyclicity of T1, . . . , Tp does not imply the s-hypercyclicity of T −1 if T1, . . . , Tp are invertible. Moreover, finitely many composition operators generated by non-elliptic automorphisms of D may be not s-hypercyclic on H(D) or on H 2(D). = (Cϕ1)−1, Cϕ−1 and ϕ−1 2 1 , . . . , T −1 p 1 2 2. Further study of d-hypercyclicity of composition operators, this time on weighted Bergman spaces on D, is performed in [30]. In 1929 Birkhoff [12] proved that the translation operator τa (a ∈ C \ {0}) given by (τaf )(z) = f (z + a) is hypercyclic on the space H(C) of entire functions. It is SIMULTANEOUS UNIVERSALITY 21 proved in [4, Prop. 5.5] and [11, Theorem 3.1] that if a1, . . . , ap are pairwise distinct nonzero complex numbers, then τa1, . . . , τap are d-hypercyclic. Trivially, we obtain: if a1, . . . , ap ∈ C \ {0}, then τa1, . . . , τap are s-hypercyclic. As the next proposition shows, we may obtain a slight extension to weighted translation operators. Proposition 6.4. Let p ≥ 2, and let a1, . . . , ap, λ1, . . . , λp ∈ C \ {0} such that λj = λl for all j, l ∈ {1, . . . , p} with aj = al. Then there is a sequence (nk) ⊂ N such that the sequences (λ1τa1)nk, . . . , (λpτap)nk are s-mixing. In particular, the operators λ1τa1, . . . , λpτap are densely s-hypercyclic on H(C). Proof. Select a finite sequence {j(1) < j(2) · · · < j(q)} ⊂ {1, . . . , p} satisfy- ing that, if bl := aj(l) (l = 1, . . . , q), then the bl's are pairwise distinct and {a1, . . . , ap} = {b1, . . . , bq}. Let µl := λj(l). Consider the operators Tj := λjτaj (j = 1, . . . , p) and Sl := Tj(l) = µlτbl (l = 1, . . . , q). Let us prove that S1, . . . , Sq are s-mixing. In fact, by following the approach of the proof of [11, Theorem 3.1], we can prove that they are even d-mixing. To this end, and taking into account that the sets V (h, r, ε) := {f ∈ H(C) : f (z) − h(z) < ε for all z ∈ B(0, r)} (h ∈ H(C), ε > 0, r > 0), form a basis for the topology of H(C), it is enough to prove that, for given h, g1, . . . , gq ∈ H(C) and ε, r > 0, there is n0 ∈ N such that, for every n ≥ n0, there exists an entire function f with 2r f (z) − h(z) < ε and (Sn l f )(z) − gl(z) < ε bi−bl + max1≤l≤q (z ∈ B(0, r), l = 1, . . . , q). (1) 2r Select n0 ∈ N with n0 > maxi6=l bl . Then, for each n ≥ n0, the disks B(0, r), B(nb1, r), . . . , B(nbq, r) are pairwise disjoint. Pick s > r such that the disks B(0, s), B(nb1, s), . . . , B(nbq, s) are still pairwise disjoint. Let K := B(0, r) ∪ B(nb1, r) ∪ · · · ∪ B(nbq, r) and Ω := B(0, s) ∪ B(nb1, s) ∪ · · · ∪ B(nbq, s). Note that Ω is an open set, Ω ⊃ K and K is a compact subset having connected complement. Consider the function F : Ω → C defined by F (z) = h(z) if z ∈ B(0, s) and F (z) := µ−n l gl(z−nbl) if z ∈ B(nbl, s) (1 ≤ l ≤ q). Then F ∈ H(Ω). From Runge's approximation theorem (see e.g. [16]), it follows that there exists a polynomial f (so f ∈ H(C)) such that f (z) − F (z) < ε/(1 + µn l ) for all z ∈ K. But this implies that f (z) − h(z) < ε on B(0, r) and µn l f (z) − gl(z − nbl) < ε on B(nbl, r). Since the last inequality is equivalent to µn l f (z + nbl) − gl(z) < ε on B(0, r), (1) is obtained. As the set D := {λj/λl : j, l ∈ {1, . . . , p} with aj = al} ⊂ T is finite, it is a Dirichlet set. Then there is a strictly increasing sequence (nk) ⊂ N such that ξnk → 1 as k → ∞, for all ξ ∈ D. Fix a subsequence (mk) of (nk). Since S1, . . . , Sq are s-mixing, the set in )) )). For each ν ∈ {1, . . . , p} there is a unique l = l(ν) ∈ (see Proposition 3.3). 1 ), . . . , (Smk 1 ), . . . , (Smk s-HC((Smk s-HC((Smk is dense q q Fix f 22 BERNAL AND JUNG ν → 1, hence ξmk {1, . . . , q} such that aν = bl, so that λν = µl. Observe that ξν := λν/µl ∈ D. Then ξnk ν → 1 (k → ∞) for all ν ∈ {1, . . . , p}. Given g ∈ H(C), we can find a subsequence (pk) of (mk) with Spk l(ν)f → g (k → ∞) uniformly on compacta for every ν ∈ {1, . . . , p}. Since ξpk ν → 1 for all ν, we obtain that T pk ν f = ξpk l(ν)f −→ 1 · g = g (k → ∞) uniformly on compacta for every ν = 1, . . . , p. Therefore f ∈ s-HC((T mk )), which shows that this set is dense. By Proposition 3.3, the sequences (T nk p ) are s-mixing, as re- quired. (cid:3) 1 ), . . . , (T nk ν Spk ), . . . , (T mk p 1 Another important collection of operators on H(C) is that of differentiation operators. Consider the derivative operator D : f ∈ H(C) 7→ f ′ ∈ H(C). Its It is shown in hypercyclicity on H(C) was proved by MacLane in 1952 [25]. [11, Prop. 3.3] that if p ≥ 2, r1, . . . , rp ∈ N with r1 < · · · < rp and λ1, . . . , λp ∈ C \ {0}, then λ1Dr1, . . . , λpDrp are d-mixing, so densely d-hypercyclic. Concerning s-hypercyclicity, the following proposition shows that somewhat softer assumptions are allowed, although, similarly to the last proposition, we have not been able to obtain the s-mixing property for the whole sequences. Proposition 6.5. Let r1 ≤ · · · ≤ rp be positive integers and λ1, . . . , λp ∈ C \ {0}, where p ≥ 2. Suppose that λj = λl for all j, l ∈ {1, . . . , p} with rj = rl. Then there is a sequence (nk) ⊂ N such that the sequences (λ1Dr1)nk, . . . , (λpDrp)nk are s-mixing. In particular, the operators λ1Dr1, . . . , λpDrp are densely s-hypercyclic on H(C). Proof. As the set {λj/λl : j, l ∈ {1, . . . , p} with rj = rl} ⊂ T is finite, it is a Dirichlet set. Then there is a strictly increasing sequence (nk) ⊂ N such that (λj/λl)nk → 1 as k → ∞, for all j, l ∈ {1, . . . , p} with rj = rl. Put X0 := {polynomials} = span{zm : m ∈ N0} and W0 := ∆(X p 0 ). Then X0 is dense in X := H(C) and W0 = ∆(X p 0 ) ⊃ ∆(X p). Let Tj := λjDrj (1 ≤ j ≤ p). For each k ∈ N, define the map Rk : W0 → X via Rk(zm, . . . , zm) := pXl=1 1 τ (l) · 1 λnk l · zm+rlnk (m + 1)(m + 2) · · · (m + rlnk) , where τ (l) := card {i ∈ {1, . . . , p} : ri = rl} (1 ≤ l ≤ p). Then Rk is extended to the whole W0 by linearity. We have: (i) T nk j zm = 0 as soon as nkrj > m, so T nk {1, . . . , p} and all m ≥ 0. Therefore, by linearity, T nk for all j ∈ {1, . . . , p}. j zm → 0 as k → ∞ for all j ∈ j → 0 (k → ∞) on X0 SIMULTANEOUS UNIVERSALITY 23 (ii) Fix m ∈ N0 and a compact set K ⊂ C. There is M ∈ (0, +∞) with K ⊂ B(0, M). Given k ∈ N, we obtain Rk(zm, . . . , zm) ≤ sup z∈K ≤ ≤ pXl=1 pXl=1 pXl=1 1 τ (l) · 1 λnk l · M m+rlnk (m + 1)(m + 2) · · · (m + rlnk) 1 M m+rlnk/λnk l τ (l) (m + 1)(m + 2) · · · (m + nk) M m τ (l) (M rl/λl)nk nk! → 0 (k → ∞) Hence, by linearity, Rk → 0 (k → ∞) pointwise on W0. (iii) Fix m ∈ N0, j ∈ {1, . . . , p} and k ∈ N with nk > m. Let us compute the action of T nk j Rk on each (zm, . . . , zm). This yields three sums, the first of them corresponding to those l ∈ {1, . . . , p} with rl < rj, that equals 0. Therefore T nk j Rk(zm, . . . , zm) = 0 + pXl=1 rl=rj 1 τ (l) λl(cid:1)nk · zm ·(cid:0) λj zm+(rl−rj )nk (m + 1)(m + 2) · · · (m + (rl − rj)nk) + pXl=1 rl>rj 1 τ (l) −→ 1 τ (j) ·(cid:0) λj λl(cid:1)nk · pXl=1 rl=rj · zm · 1 + 0 = zm (k → ∞) uniformly on compacta in C, because τ (j) = τ (l) and ( λj λl (j, l) with rj = rl. By linearity again, we get T nk j = 1, . . . , p and all (w, . . . , w) ∈ W0. )nk → 1 for all j Rk(w, . . . , w) → w for all The conclusion now follows from Theorem 3.7 (or from Remark 3.8.1). (cid:3) For instance, the operators 5D, D2, −D2, eiD2, 1 10D3, −3D4 are s-hypercyclic, but clearly not d-hypercyclic. An extension unifying both Birkhoff's and MacLane's theorems takes place by considering convolution operators on H(C), that is, operators commuting with n=0 anzn ∈ H(C). Then Φ is said to be of exponential type provided that there are positive constants A, B such that Φ(z) ≤ A exp(Bz) for all z ∈ C. Then its associated differential operator n=0 anf (n) (f ∈ H(C)) defines an ope- rator on H(C). Moreover, an operator T ∈ L(H(C)) is of convolution if and only if T = Φ(D) for some entire function Φ of exponential type. Note that D and τa all translations τa. Let Φ(z) = P∞ Φ(D) = P∞ n=0 anDn given by Φ(D)f = P∞ 24 BERNAL AND JUNG are special cases (take Φ(z) ≡ z and Φ(z) ≡ eaz, resp.). Godefroy and Shapiro [17] proved in 1991 that any nonscalar convolution operator is hypercyclic. If G is any domain in C, then Φ(D) is also an operator on H(G) whenever Φ is of subexponential type, that is, for given ε > 0 there is a constant A > 0 such that Φ(z) ≤ A exp(εz) for all z ∈ C. We have that also Φ(D) is hypercyclic on H(G) provided that G is simply connected (i.e. its complement with respect to the one-point compactification C∞ of C is connected) and Φ is not constant. For s-hypercyclicity, we present the following assertion, with which we put an end to this introductory paper on s-universality. Proposition 6.6. Assume that G ⊂ C is a simply connected domain and that Φ1, . . . , Φp are entire functions of subexponential type (or just of exponential type if G = C). Assume also that the set is nonempty and that each set U0 :=(cid:8)λ ∈ C : max Ui :=(cid:8)λ ∈ C : Φi(λ) > 1 and max 1≤j≤p Φj(λ) < 1(cid:9) Φj(λ) ≤ Φi(λ)(cid:9) (1 ≤ i ≤ p) 1≤j≤p has nonempty interior U 0 satisfy Φi(λ) = Φj(λ) for some λ ∈ U 0 i . Suppose, in addition, that whenever i, j ∈ {1, . . . , p} i , there exists ζ ∈ T with Φj = ζ · Φi. Then there is a sequence (nk) ⊂ N such that the sequences (Φ1(D))nk, . . . . . . , (Φp(D))nk are s-mixing. In particular, the operators Φ1(D), . . . , Φp(D) are densely s-hypercyclic on H(C). Proof. We write eλ := exp(·λ)G for λ ∈ C. It is easy to see that the functions eλ are linearly independent. Denote Vi := U 0 i (1 ≤ i ≤ p). As U0, V1, . . . , Vp are open and nonempty, we obtain that X0 := span{eλ : λ ∈ U0} is dense in X := H(G) (because G is simply connected: use Runge's approximation theorem together with the fact that span{exp(·λ) : λ ∈ U0} is dense in H(C); see e.g. [17, Sect. 5]). i=1 span{eλ : λ ∈ Vi} is dense in X p. Hence W0 :=Qp As A := {ζ ∈ T : exist l, j ∈ {1, . . . , p} with Φj = ζΦl} ⊂ T is finite, it is a Dirichlet set; hence there is a strictly increasing sequence (nk) ⊂ N such that ζ nk → 1 for all ζ ∈ A. For each i ∈ {1, . . . , p}, we put Ti := Φi(D)H(G), Ei := {j ∈ {1, . . . , p} : exists ζ ∈ T with Φj = ζΦi} and τ (i) := card(Ei). Notice that if i ∈ Ej, then Ei = Ej (just use that T is a multiplicative group), hence τ (i) = τ (j). Given i ∈ {1, . . . , p} and vi ∈ span{eλ : λ ∈ Vi}, there are uniquely determined scalars ci,1, . . . , ci,J(i) ∈ l=1 ci,leλi,l. For k ∈ N C and pairwise distinct λi,1, . . . , λi,J(i) ∈ Vi such that vi =PJ(i) we define Rk : W0 → X as Rkw := pXi=1 1 τ (i) · J(i)Xl=1 ci,l Φi(λi,l)nk · eλi,l, (1) SIMULTANEOUS UNIVERSALITY 25 where w = (v1, . . . , vp) ∈ W0 and the vi's are as above. We have: (i) If λ ∈ U0 and j ∈ {1, . . . , p}, then T nk j eλ = Φj(λ)nkeλ → 0 as k → ∞, because Φj(λ) < 1. By linearity, we get T nk j → 0 on X0. l=1 ci,leλi,l, as above. Since Φi(λi,l) > 1, we get Φi(λi,l)nk → +∞ as k → ∞, for each i ∈ {1, . . . , p} and each l = 1, . . . , J(i). From (1) one derives that Rkw → 0. (ii) Let w = (v1, . . . , vp) ∈ W0, so that vi = PJ(i) (iii) Again, let w = (v1, . . . , vp) ∈ W0, with vi = PJ(i) {1, . . . , p} and k ∈ N. We compute l=1 ci,leλi,l. Fix j ∈ T nk j Rkw = = pXi=1 pXi=1 1 τ (i) 1 τ (i) · · J(i)Xl=1 J(i)Xl=1 ci,l Φi(λi,l)nk · T nk j eλi,l ci,l ·(cid:18)Φj(λi,l) Φi(λi,l)(cid:19)nk eλi,l = Ak + Bk, where Ak (Bk, resp.) denotes the part of the preceding sum corresponding to those i ∈ Ej (i 6∈ Ej, resp.). If i ∈ Ej, there is ζ = ζi,j ∈ A such that = ζ nk → 1 as k → ∞. Note that τ (i) = τ (j) if i ∈ Ej. Therefore, on the one hand, Φj = ζ ·Φi, so that(cid:16) Φj(λi,l) Φi(λi,l)(cid:17)nk J(i)Xl=1 pXi=1 Ak → ci,l · eλi,l = 1 τ (i) · i∈Ej 1 τ (j) · pXi=1 i∈Ej J(i)Xl=1 ci,l · eλi,l = 1 τ (j) · pXi=1 i∈Ej vi. On the other hand, if i 6∈ Ej, we have that Φj(λi,l)/Φi(λi,l) < 1 for all l = 1, . . . , J(i) (indeed, as λi,l ∈ Vi, we have Φj(λi,l) ≤ Φi(λi,l); if we assume Φj(λi,l) = Φi(λi,l), then there would exist ζ ∈ T with Φj = ζ · Φi, which would yield i ∈ Ej, a contradiction). Hence (cid:0) Φj (λi,l) Φi(λi,l)(cid:1)nk → 0, so Bk → 0. This entails T nk j Rkw = Ak + Bk → 1 τ (j) · pXi=1 i∈Ej vi (k → ∞), and the last vector belongs to conv({v1, . . . , vp}) since in the last sum there are exactly τ (j) summands. The conclusion follows, once again, from the s-hypercyclicity criterion (Theorem 3.7). (cid:3) Remark 6.7. Proposition 3.4 in [11] (see also [4, Theorem 5.3]) asserts that if U0 and Wi := {λ ∈ C : Φi(λ) > 1 and maxj6=i Φj(λ) < Φi(λ)} (1 ≤ i ≤ p) are nonempty, then Φ1(D), . . . , Φp(D) are d-mixing. If these assumptions are satisfied, then the assumptions of Proposition 6.6 are also satisfied. Note that Proposition 6.6 includes the case Φ1 = Φ, Φj = cjΦ with cj = 1 (j = 2, . . . , p). 26 BERNAL AND JUNG This paper does not intend to be exhaustive. Of course, many more sets of operators or of sequences of operators may be analyzed under the point of view of s-universality/s-hypercyclicity. For instance, consider a compact set K ⊂ C and the Banach space (A(K), k · k∞) of continuous functions K → C that are holomorphic on K 0. Let TK,n : f ∈ H(D) 7→ (Snf )K ∈ A(K) (n ∈ N), where Snf denotes the nth partial sum of the Taylor series of f around the origin. Assume that K ⊂ C\D and that K has connected complement. Then Costakis and Tsirivas [14, Sect. 3] have recently shown that, given any two strictly increasing sequences (nk), (mk) in N, the sequences (TK,nk) and (TK,mk) are -- by using our terminology -- s-universal. Even more, they have shown that \ (cid:8)s-U((TK,nk), (TK,mk)) : K ⊂ C \ D compact, C \ K connected(cid:9) is a residual subset of H(D). Acknowledgements. The first author has been partially supported by Plan Andaluz de Investigaci´on de la Junta de Andaluc´ıa FQM-127 Grant P08-FQM- 03543 and by MEC Grant MTM2015-65242-C2-1-P. The second author has been supported by DFG-Forschungsstipendium JU 3067/1-1. References [1] N. Bamerni, V. Kadets, and A. Kılı¸cman, Hypercyclic operators are subspace hypercyclic, J. Math. Anal. Appl. 435 (2016), 1812 -- 1815. [2] F. Bayart and E. Matheron, Dynamics of Linear Operators, Cambridge University Press, 2009. Cambridge Tracts in Mathematics. [3] L. Bernal-Gonz´alez, Densely hereditarily hypercyclic sequences and large hypercyclic mani- folds, Proc. Amer. Math. Soc. 127 (1999), 3279 -- 3285. [4] L. Bernal-Gonz´alez, Disjoint hypercyclic operators, Studia Math. 182 (2007), 113 -- 131. [5] L. Bernal-Gonz´alez and K.-G. Grosse-Erdmann, The Hypercyclicity Criterion for sequences of operators, Studia Math. 157 (2003), no. 1, 17 -- 32. [6] J. B`es and O. Martin, Compositional disjoint hypercyclicity equals disjoint supercyclicity, Houston J. Math. 38 (2012), 1149 -- 1163. [7] J. B`es, O. Martin, and A. Peris, Disjoint hypercyclic linear fractional composition operators, J. Math. Anal. Appl. 381 (2011), 843 -- 856. [8] J. B`es, O. Martin, A. Peris, and S. Shkarin, Disjoint mixing operators, J. Funct. Anal. 263 (2012), 1283 -- 1322. [9] J. B`es, O. Martin, and R. Sanders, Weighted shifts and disjoint hypercyclicity, J. Operator Theory 72 (2014), 15 -- 40. [10] J. B`es and A. Peris, Hereditarily hypercyclic operators, J. Funct. Anal. 167 (1999), no. 1, 94 -- 112. [11] [12] G. D. Birkhoff, D´emonstration d'un th´eor`eme ´el´ementaire sur les fonctions enti`eres, C. R. , Disjointness in Hypercyclicity, J. Math. Anal. Appl. 336 (2007), 297 -- 315. Acad. Sci. Paris 189 (1929), 473 -- 475. [13] L. Bukovsk´y, The structure of the real line, Springer Basel, Warsaw, 2011. SIMULTANEOUS UNIVERSALITY 27 [14] G. Costakis and N. Tsirivas, Doubly universal Taylor series, J. Approx. Theory 180 (2014), 21 -- 31. [15] M. De La Rosa and C. Read, A hypercyclic operator whose direct sum T ⊕T is not hypercyclic, J. Operator Theory 61 (2009), 369 -- 380. [16] D. Gaier, Lectures on complex approximation, Birkhauser, Basel-London-Stuttgart, 1987. [17] G. Godefroy and J. H. Shapiro, Operators with dense, invariant, cyclic vectors manifolds, J. Funct. Anal. 98 (1991), no. 2, 229 -- 269. [18] K.-G. Grosse-Erdmann, Holomorphe Monster und universelle Funktionen, Vol. 176, Mit. Math. Sem. Giessen, 1987. [19] [20] , Universal families and hypercyclic operators, Bull. Amer. Math. Soc. 36 (1999), no. 3, 345 -- 381. , Recent developments in hypercyclicity, Rev. R. Acad. Cienc. Ser. A Mat. 97 (2003), no. 3, 273 -- 286. [21] K.-G. Grosse-Erdmann and A. Peris, Linear Chaos, Springer, London, 2011. [22] C. M. Le, On subspace-hypercyclic operators, Proc. Amer. Math. Soc. 139 (2011), 2847 -- 2852. [23] F. Le´on-Saavedra and V. Muller, Rotations of hypercyclic and supercyclic operators, In- tegr. equ. oper. theory 50 (2004), 385 -- 391. [24] B. F. Madore and R. A. Mart´ınez-Avendano, Subspace hypercyclicity, J. Math. Anal. Appl. 373 (2011), 502 -- 511. [25] G. R. MacLane, Sequences of derivatives and normal families, J. Anal. Math. 2 (1952), no. 1, 72 -- 87. [26] S. Rolewicz, On orbits of elements, Studia Math. 32 (1969), 17 -- 22. [27] J. H. Shapiro, Composition Operators and Classical Function Theory, Universitext, Springer- Verlag, New York, 1993. [28] S. Shkarin, Universal elements for non-linear operators and their applications, J. Math. Anal. Appl. 348 (2008), 193 -- 210. [29] J. Wengenroth, Hypercyclic operators on nonlocally convex spaces, Proc. Amer. Math. Soc. 131 (2003), no. 6, 1759 -- 1761. [30] L. Zhang and Z. H. Zhou, Dynamics of composition operators on weighted Bergman spaces, Indag. Math. (N.S.) 27 (2016), no. 1, 406 -- 418. Departamento de An´alisis Matem´atico Facultad de Matem´aticas, Universidad de Sevilla Apdo. 1160, Avenida Reina Mercedes 41080 Sevilla, Spain. E-mail address: [email protected] Fachbereich IV Mathematik Universitat Trier D-54286 Trier, Germany E-mail address: [email protected]
1904.03088	1	1904	2019-04-05T14:32:16	Elliptic problems and holomorphic functions in Banach spaces	[ "math.FA" ]	In the first part we show that a vector-valued almost separably valued function $f$ is holomorphic (harmonic) if and only if it is dominated by an $L^1_\mathrm{loc}$ function and there exists a separating set $W\subset X'$ such that $\langle f,x'\rangle$ is holomorphic (harmonic) for all $x'\in W$. This improves a known result which requires $f$ to be locally bounded. In the second part we consider classical results in the $L^p$ theory for elliptic differential operators of second order. In the vector-valued setting these results are shown to be equivalent to the UMD property.	math.FA	math	ELLIPTIC PROBLEMS AND HOLOMORPHIC FUNCTIONS IN BANACH SPACES WOLFGANG ARENDT, MANUEL BERNHARD, AND MARCEL KREUTER Abstract. In the first part we show that a vector-valued almost separably valued function f is holomorphic (harmonic) if and only if it is dominated by an L1 loc function and there exists a separating set W ⊂ X ′ such that hf, x′i is holomorphic (harmonic) for all x′ ∈ W . This improves a known result which requires f to be locally bounded. In the second part we consider classical results in the Lp theory for elliptic differential operators of second order. In the vector-valued setting these results are shown to be equivalent to the UMD property. 1. Introduction Let f : Ω → X, where Ω is an open subset of C (or Rd) and X is a complex (real) Banach space. The function f is called holomorphic (harmonic) if it is complex differentiable (twice partially differentiable with ∆f = 0). The first part of this article is concerned with a cri- terion for vector-valued holomorphy (harmonicity). The function f is called weakly holomorphic (weakly harmonic) if x′ ◦ f is holomorphic (harmonic) for all x′ ∈ X ′. We say that f is very weakly holomorphic (very weakly harmonic) if there exists a separating subset W ⊂ X ′ such that x′ ◦ f is holomorphic (harmonic) for all x′ ∈ W . It was shown in [11] (see also [12]) that a vector-valued function f is holomorphic if and only if it is locally bounded and very weakly holomorphic. This answered a question posted in [17] ten years earlier. A very short proof was given in [4]. In [2] it was shown that a similar approach yields the analogous result for harmonic functions. The first part of this paper is concerned with an improvement of these results. It is known, that very weak holomorphy alone is not sufficient [4, Theorem 1.5]. However, we will show that the boundedness assumption can be weakened. We say that a set F of functions from Ω to X is locally L1- bounded if there exists a function g ∈ L1 loc(Ω, R) such that kf (ξ)kX ≤ g(ξ) holds for almost all ξ ∈ Ω and for all f ∈ F . A single function Date: April 8, 2019. 2010 Mathematics Subject Classification. Primary: 35J25; Secondary: 46B20, 31C05, 30A99. Key words and phrases. Vector-valued holomorphic functions, harmonic func- tions, elliptic equations, regularity, UMD spaces. 1 ELLIPTIC PROBLEMS AND HOLOMORPHIC FUNCTIONS 2 f : Ω → X is called locally L1-bounded if {f } is locally L1-bounded. A net {fi}i∈I is called locally L1-bounded if {fi, i ∈ I} is locally L1- bounded. Our result is the following. Theorem 1.1. Let X be a Banach space and Ω ⊂ C (Ω ⊂ Rd) be open. A function f : Ω → X is holomorphic (harmonic) if and only if it is locally L1-bounded and very weakly holomorphic (very weakly harmonic). We give two proofs of this result, one of which is very short, but is only valid if X is separable, and one that follows the approach in [4] and [2]. The first proof will also yield a shortcut proof for the vector-valued version of Weyl's lemma. This result will be needed in the second part of the paper. The set L(D(Ω, R), X) of X-valued distributions on Ω will be denoted by D′(Ω, X). Theorem 1.2 (Weyl). Let Ω ⊂ Rd be open, X be a Banach space and let f ∈ L1 loc(Ω, X) such that ∆f = 0 in D′(Ω, X). Then f has a harmonic representative; that is, there exists f ∗ ∈ C ∞(Ω, X) such that ∆f ∗ = 0 and f = f ∗ almost everywhere. Recall that ∆f = 0 in the sense of distributions means thatR f ∆ϕ = 0 for all test functions ϕ ∈ C ∞ c (Ω, R). In view of the formulation of Theorem 1.1 we want to remark that this is equivalent to saying that there exists a separating set W ⊂ X ′ such that ∆(x′◦f ) = 0 in D′(Ω, R) for all x′ ∈ W . In the second part of the paper, Sections 4 - 6, we investigate some classical elliptic problems. Again, we ask whether the solutions with values in a Banach space have the same regularity as in the scalar case. Using a result by Geiss, Montgomery and Saksman [8] on homogeneous vector-valued multipliers we prove our main technical tool, Thereom 4.3, on regularity properties of Newtonian potentials. This result will be used to determine the domain of the Laplacian on Lp(Rd, X). One of our main results shows that on a bounded domain Ω of class C 1,1 the following classical property characterizes UMD-spaces: Given f ∈ Lp(Ω, X) there exists a unique u ∈ W 1,p 0 (Ω, X) ∩ W 2,p(Ω, X) solving ∆u = f . More general elliptic operators are also considered in Section 6. Parts of this work are contained in the third author's thesis [15]. 2. Harmonic and Holomorphic Functions -- The Separable Case In this section we will give a proof of Theorem 1.1 in the case of a separable Banach space. ELLIPTIC PROBLEMS AND HOLOMORPHIC FUNCTIONS 3 Lemma 2.1. Let Ω ⊂ Rd be open and let f ∈ L1 loc(Ω, R) such that ∆f = 0 in D′(Ω, R). Furthermore let ρr be a mollifier supported in B(0, r), r > 0, consisting of radial functions. Then ρr ∗ f = f almost everywhere in Ωr := {ξ ∈ Ω, dist(ξ, ∂Ω) > r}. Proof. Since ∆f = 0 distributionally there exists a harmonic represen- tative f ∗ of f by Weyl's Lemma in the real-valued case [5, Chapter II, §3, Proposition 1]. Since f = f ∗ almost everywhere it follows that ρr ∗ f = ρr ∗ f ∗ everywhere in Ω. Now by [7, Chapter 2, Proof of Theorem 6] we have that ρr ∗ f ∗ = f ∗ on Ωr from which the claim follows. (cid:3) Proof of Theorem 1.1 if X is separable. (a) We start with the case that f is very weakly harmonic. Then f is very weakly measurable, and it follows from the Krein-Smulyan theorem (c.f. Section 3) that f is mea- surable -- for a full proof we refer the reader to [1, Theorem 1.2] or [14, Theorem 1.1.20]. Since f is locally L1-bounded it follows that f ∈ L1 loc(Ω, X). Let ρr be a mollifier supported in B(0, r). Then the function fr := ρr ∗ f is well-defined and smooth in Ωr. Lemma 2.1 shows that hfr, x′i = ρr ∗ hf, x′i = hf, x′i in Ωr for every x′ in the sepa- rating set W ⊂ X ′ for which hf, x′i is harmonic. Since W is separating it follows that fr = f in Ωr. In particular: f is smooth and hence -- using again that W is separating -- ∆f = 0. (b) Now we come to the case where f is very weakly holomorphic. Analogously to the harmonic case one sees that f is locally integrable. Let z0 ∈ Ω and let r0 > 0 such that B(z0, r0) ⊂⊂ Ω. Since f is inte- grable on B(z0, r0) it follows from Fubini's theorem that f is integrable on the sphere S(z0, r) for almost all r ≤ r0. Choose such an r and define u(z) := 1 2πiZw−z0=r f (w) z − w dw for all z ∈ B(z0, r). As in the scalar case one shows that u defines a holomorphic function. Cauchy's integral formula shows that hu(z), x′i = hf (z), x′i for all x′ ∈ W and all z ∈ B(z0, r). Since W is separating it follows that u = f and hence f is holomorphic. (cid:3) The approach used for the case where f is very weakly harmonic also yields the Proof of Theorem 1.2. Let ρr be a mollifier supported in B(0, r) ⊂⊂ Ω. By assumption the function fr := ρr ∗ f is well-defined. Since f is measurable we may assume that X is separable. In this case there exists a countable separating set W ⊂ X ′ [14, Proposition B.1.10]. Lemma 2.1 shows that for every x′ ∈ W there exists a negligible set ELLIPTIC PROBLEMS AND HOLOMORPHIC FUNCTIONS 4 Nx′ such that hfr, x′i = hf, x′i in Ωr\Nx′. Since W is countable the set N := Sx′∈W Nx′ is negligible. Furthermore W separates X and hence fr = f almost everywhere in Ωr. For every x′ ∈ W the function hf, x′i has a harmonic representative by Weyl's Lemma in the real-valued case [5, Chapter II, §3, Proposition 1]. Since hfr, x′i is a continuous representative of hf, x′i in Ωr it follows that hfr, x′i is the harmonic representative of hf, x′i in Ωr. Since W is separating it follows from Theorem 1.1 that fr is harmonic in Ωr. The claim now follows by taking a sequence rn → 0 and defining f ∗(ξ) := frn(ξ), where ξ ∈ Ωrn. Then f ∗ : Ω → X is well-defined, harmonic, and conincides with f almost everywhere. (cid:3) We want to give a holomorphic version of Theorem 1.2 using the distributional Cauchy-Riemann equations D1u = D2v D2u = −D2v. For vector-valued functions -- which do not have a real or imaginary part -- we make sense of these equations by saying that a function f : Ω → X from an open subset of C into a complex Banach space satisfies the Cauchy-Riemann equations very weakly distributionally if there exists a separating set W ⊂ X ′ such that the functions u := Rehf, x′i and v := Imhf, x′i satisfy the Cauchy-Riemann equations distributionally for every x′ ∈ W . The following lemma is known, but we present a proof using our results. Lemma 2.2 ([10, Theorem 9]). Let Ω ⊂ C be open and let f ∈ L1 loc(Ω, C) such that f satisfies the Cauchy-Riemann equations distri- butionally. Then f has a holomorphic representative. Proof. It follows from the Cauchy-Riemann equations that the func- tions u := Re f and v := Im f are harmonic in D′(Ω, R). Lemma 2.1 shows that for a radially symmetric mollifier ρr supported in B(0, r) we have fr := ρr ∗ f = f almost everywhere in Ωr. Since fr is continuously partially differentiable it satisfies the Cauchy-Riemann equations in the classical sense in Ωr and thus is holomorphic. We may now define the representative of f analogously to the proof of Theorem 1.2. (cid:3) Theorem 2.3. Let Ω ⊂ C be open and let X be a complex Banach space. Suppose f ∈ L1 loc(Ω, X) satisfies the Cauchy-Riemann equations very weakly distributionally. Then f has a holomorphic representative. Proof. Since f is measurable we may assume that X is separable. By [14, Theorem B.1.11] we may assume that W is countable. Let ρr be a radially symmetric mollifier supported in B(0, r) and define fr := ρr ∗f . Let x′ ∈ W . By Lemma 2.2 we know that hf, x′i has a holomorphic ELLIPTIC PROBLEMS AND HOLOMORPHIC FUNCTIONS 5 representative and the proof tells us that in Ωr this representative is given by ρr ∗ hf, x′i = hfr, x′i. Since W is countable, it follows that fr is a representative of f in Ωr. Furthermore, Theorem 1.1 shows that fr is holomorphic in Ωr. The representative is then defined as above. (cid:3) 3. Harmonic and Holomorphic Functions -- The General Case As announced before, we will now give a proof of Theorem 1.1 which is valid also in non-seperable spaces. We use arguments of [4] and [2] but add a new idea to get around with the L1 loc-hypothesis only. We gather some results which we will need for the proof. By σd−1 we denote the (d − 1)-dimensional Hausdorff measure in Rd and by ωd the Hausdorff measure of the unit sphere Sd−1 ⊂ Rd. Lemma 3.1 ([2, Lemma 2.3 and Theorem 5.2], also c.f. [13]). (a) Let Ω ⊂ C be open and let X be a Banach space. A function f : Ω → X is holomorphic if and only if it is weakly holomorphic. In this case, f satisfies Cauchy's integral formula f (z) = 1 2πiZw−z0=r0 f (w) w − z dw for all z0 ∈ Ω and r0 > 0 such that z ∈ B(z0, r0) ⊂⊂ Ω. (b) Let Ω ⊂ Rd be open and let X be a Banach space. A function f : Ω → X is harmonic if and only if it is weakly harmonic. In this case, f satisfies Poisson's integral formula f (ξ) = 1 ωdr0 ZSd−1(ξ0,r0) r2 0 − ξ − ξ02 ξ − sd f (s) dσd−1(s) for all ξ0 ∈ Ω and r0 > 0 such that ξ ∈ B(ξ0, r0) ⊂⊂ Ω. Proposition 3.2. Let X be a Banach space and let {fi}i∈I be a locally L1-bounded net of X-valued holomorphic (harmonic) functions on the open set Ω ⊂ C (Ω ⊂ Rd). Assume that f := limi∈I fi exists pointwise in Ω. Then f is a holomorphic (harmonic) function and f = limi∈I fi uniformly on compact sets. Proof. We start with the case of holomorphic functions. Let z0 ∈ Ω. By Fubini's theorem the net {fi}i∈I is locally L1-bounded on the set {w ∈ Ω, w − z0 = r0} for almost all r0 > 0. Fix such an r0 > 0 and denote by Γ the set {w ∈ Ω, w − z0 = r0}. Cauchy's integral formula yields kfi(z1) − fi(z2)kX ≤ ≤ 1 2πZΓ C 2π kfi(w)kX(cid:12)(cid:12)(cid:12)(cid:12) z1 − z2 (w − z1)(w − z2)(cid:12)(cid:12)(cid:12)(cid:12) kfikL1(Γ,X)z1 − z2 dw ELLIPTIC PROBLEMS AND HOLOMORPHIC FUNCTIONS 6 for all z1, z2 ∈ B(z0, r0), some constant C = C(dist(z1, Γ), dist(z2, Γ)) and all i ∈ I. Since the net {fi}i∈I is locally L1-bounded this shows that {fi}i∈I is equicontinuous on compact subsets. Hence f = limi∈I fi exists uniformly on compact sets and it follows that f satisfies Cauchy's integral formula and is thus holomorphic. The case of harmonic func- tions is treated analogously using Poisson's integral formula. (cid:3) Theorem 3.3 (Krein-Smulyan [16, Corollary 2.7.12]). Let X be a Ba- nach space and let Y ⊂ X ′ be a subspace. Then Y is closed in the weak-∗ topology if and only if Y ∩ BX ′ is weakly-∗ closed, where BX ′ denotes the closed unit ball in X ′. Proof of Theorem 1.1 for general Banach spaces. Consider the space Y := {x′ ∈ X ′, x′ ◦ f is holomorphic (harmonic)}. Since W ⊂ Y it follows that Y is weak-∗ dense in X ′. It remains to show that Y is closed in the weak-∗ topology since then the result follows from Lemma 3.1. By the Krein-Smulyan theorem it suffices to show that Y ∩ BX ′ is weakly-∗ closed for every r > 0. Let {x′ i}i∈I be a net in Y ∩ BX ′ such that x′ i ⇀∗ x′ ∈ BX ′. The net formed by the i ◦ f is locally L1-bounded and converges pointwise to functions fi := x′ x′◦f . By Proposition 3.2 it follows that x′◦f is holomorphic (harmonic) and hence x′ ∈ Y . (cid:3) Vitali's convergence theorem is usually stated for bounded sequences of holomorphic functions. We apply our results to show that it also holds for locally L1-bounded sequences. Let Ω be an open and con- nected set in C (or Rd). A subset N ⊂ Ω is called a set of uniqueness for holomorphic (harmonic) functions if every holomorphic (harmonic) function which vanishes on N also vanishes on Ω. It is well known that any infinite set contained in a compact subset ofΩ is a set of uniqueness for holomorphic functions. This does not hold for harmonic functions. On the other hand, if the closure of N ⊂ Ω has non-empty interior, then N is a set of uniqueness for harmonic functions. Theorem 3.4 (Vitali). Let X be a Banach space and let fn be a locally L1-bounded sequence of X-valued holomorphic (harmonic) functions. Suppose that N ⊂ Ω is a set of uniqueness for holomorphic (harmonic) functions such that limn→∞ fn exists pointwise on N. Then limn→∞ fn exists uniformly on compact sets and defines a holomorphic (harmonic) function. Proof. The function F : Ω → ℓ∞(N, X) z 7→ (fn(z))n∈N is holomorphic (harmonic) by Theorem 1.1. Let c(N, X) ⊂ ℓ∞(N, X) be the closed subspace of all convergent sequences and denote by q the ELLIPTIC PROBLEMS AND HOLOMORPHIC FUNCTIONS 7 quotient map ℓ∞(N, X) → ℓ∞(N, X)/c(N, X). Then q ◦ F is holomor- phic (harmonic) and vanishes on N. Since N is a set of uniqueness we have q ◦ F = 0, that is, F (z) is convergent for every z ∈ Ω. The claim now follows from Proposition 3.2. (cid:3) 4. Newtonian Potentials With this section we start the second part of this paper on elliptic Lp theory in Banach spaces. Our results about harmonic functions from Section 3 will play a role in Section 6. In the remainder of the paper X denotes a real Banach space. We recall some facts about the Newtonian potential which can be proved analogously to the real-valued case, see [9, Section 4.2] and [5, Chapter II, §3]. The fundamental solution for the Laplace equation is given by Rd\{0} ∋ ξ 7→ Φ(ξ) =( 1 2π log ξ, d(2−d)λ(B(0,1)) ξ2−d, 1 if d = 2 if d > 2. For f ∈ L1(Rd, X) with compact support we define the Newtonian potential of f via Φ ∗ f. The Newtonian potential of f is an element of L1 loc(Rd, X) and satisfies Poisson's equation ∆(Φ ∗ f ) = f in D′(Rd, X). Furthermore, if f is compactly supported and Holder continuous, the Newtonian potential of f is in C 2(Rd, X) and satisfies ∆(Φ ∗ f ) = f in the classical sense, cf. [9, Section 4.2]. In this section we will show that certain classical Lp estimates for the Newtonian potential on domains imply the UMD property of X. For an overview concerning the UMD property we refer the reader to [14, Chapter 5]. The base for our results is the following multiplier theorem. We denote by MLp(Rd, X) the space of all scalar-valued Lp(Rd, X) multipliers, see [14, Definition 5.3.3]. Theorem 4.1 ([8, Theorem 3.1]). Let d ≥ 2 and let X be a Ba- nach space. Let m ∈ C ∞(Rd\{0}, R) be even, not constant and 0- homogeneous, that is, m(λξ) = m(ξ) for all ξ ∈ Rd\{0} and λ > 0. Suppose that m ∈ MLp(Rd, X) for some 1 < p < ∞. Then X has the UMD property Corollary 4.2. Let j, k ∈ {1, . . . , d}. If the second-order Riesz trans- form RjRk (associated with the multiplier − ξiξj ξ2 ) is bounded in Lp(Rd, X) for some 1 < p < ∞, then X has the UMD property. ELLIPTIC PROBLEMS AND HOLOMORPHIC FUNCTIONS 8 Using this corollary we may prove the main result of this section -- a further characterization of the UMD property which will be useful in the next section. Theorem 4.3. Let d ≥ 2, Ω ⊂ Rd be open and non-empty and let 1 < p < ∞. Suppose that there exists a constant C > 0 and j, k ∈ {1, . . . , d} such that the estimate kDjk(Φ ∗ f )kLp(Ω,X) ≤ Ckf kLp(Ω,X) holds for all f ∈ C ∞ Proof. Since Φ ∈ L1 C ∞ c (Rd, X). Moreover c (Ω, X). Then X has the UMD property. loc(Rd, R), one has Φ ∗ f ∈ C ∞(Rd, X) for all f ∈ ∆(Φ ∗ f ) = f in the classical sense. For f : Rd → X and λ > 0 define the dilation fλ(x) := f (λx) of f . Consider the operator T : C ∞ c (Rd, X) → C ∞(Rd, X) given by It is remarkable that T commutes with dilation, that is, T f = Djk(Φ ∗ f ). (T f )λ = T fλ for all f ∈ C ∞ c (Rd, X). To see this we first note that Φ(λ−1·) = λd−2Φ(·) + cd(λ) where cd(λ) = 0 if d > 2 and c2(λ) is a constant. Consequently, for f ∈ C ∞ c (Rd, X), λ > 0 and ξ ∈ Rd we have (Φ ∗ fλ)(ξ) =Z Φ(ξ − η)f (λη) dη ω λ(cid:17) f (ω) dω = λ−dZ Φ(cid:16)ξ − = λ−dZ Φ(λ−1(λξ − ω)f (ω) dω = λ−2Z Φ(λξ − ω)f (ω) dω + λ−dcd(λ)Z f (ω) dω = λ−2(Φ ∗ f )(λξ) + λ−dcd(λ)Z f (ω) dω. Consequently, since the second term does not depend on ξ, (1) Djk(Φ ∗ fλ) = (Djk(Φ ∗ f ))λ. Next we note that for each measurable function g : Rd → X and λ > 0 we have (2) d p kgλkLp(Rd,X) = kgkLp(Rd,X). λ ELLIPTIC PROBLEMS AND HOLOMORPHIC FUNCTIONS 9 Using Corollary 4.2 it remains to show that RjRk is bounded on Lp(Rd, X). Since Ω is open we may assume that it contains (−1, 1)d. Let f ∈ C ∞ c (Rd, X). Since ∆(Φ ∗ f ) = f it follows that RjRkf = Djk(Φ ∗ f ) = T f . Choose λ > 0 such that supp f ⊂ (−λ, λ)d. Thus fλ ∈ C ∞ c (Ω, X) and we have RjRkf = Djk(Φ ∗ f ) by (1). Using (1) and (2) as well as the assumption we obtain kRjRkf kLp((−λ,λ)d,X) = λ d p k(RjRkf )λkLp((−1,1)d,X) ≤ λ = λ d p k(Djk(Φ ∗ f ))λkLp(Ω,X) d p kDjk(Φ ∗ fλ)kLp(Ω,X) d p CkfλkLp(Ω,X) ≤ λ = Ckf kLp(Rd,X). Letting λ → ∞, this shows that RjRk is bounded. (cid:3) 5. The Domain of the Laplacian on Lp(Rd, X) Let X be a Banach space and 1 < p < ∞. The operator ∆p is defined as the distributional Laplacian with maximal domain in Lp(Rd, X), that is, D(∆p) := {f ∈ Lp(Rd, X), ∆f ∈ Lp(Rd, X)} ∆pf := ∆f. It is not difficult to see that ∆p is the generator of the Gaussian semigroup, see Proposition 5.5 below. If X has the UMD property, the following estimate is known. Proposition 5.1 ([14, Proposition 5.5.4]). Let X be a Banach space that has the UMD property and let 1 ≤ p < ∞. Suppose that u ∈ Lp(Rd, X) with ∆u ∈ Lp(Rd, X). Then for j, k ∈ {1, . . . , d} we have Djku ∈ Lp(Rd, X) and there exists a constant C ≥ 0 such that kDjkukLp(Rd,X) ≤ Ck∆ukLp(Rd,X). In fact: Djku is given by the second-order Riesz transform RjRk∆u. Using this we now show Proposition 5.2. Let X be a Banach space which has the UMD prop- erty and let 1 < p < ∞. Then D(∆p) = W 2,p(Rd, X). We will need the following lemmata for the proof. Lemma 5.3 ([14, Lemma 5.5.5]). The space C ∞ ∆p (1 ≤ p < ∞). c (Rd, X) is a core for ELLIPTIC PROBLEMS AND HOLOMORPHIC FUNCTIONS 10 Lemma 5.4 (Interpolation). Let 1 ≤ p < ∞ and let Ω ⊂ Rd be open. For every ε > 0 there exists a constant Cε > 0 such that k∇ukLp(Ω,X d) ≤ εkD2ukLp(Ω,X d×d) + CεkukLp(Ω,X), for every u ∈ W 2,p is also valid in W 2,p(Ω, X). 0 (Ω, X). If Ω has a C 1,1 boundary such an inequality Proof. This can be proved analogously to the real-valued case [9, Theo- rems 7.27 and 7.28]. Note that the more elegant proof [9, Exercise 7.19] using the Rellich-Kondrachov theorem does not work in this case since the compact embeddings obviously cannot hold in infinite dimensional spaces. (cid:3) Proof of Proposition 5.2. The inclusion " ⊇ " is clear. Now let f ∈ D(∆p) and let ϕn ∈ C ∞ c (Rd, X) such that ϕn → f and ∆ϕn → ∆f . By the estimates in Proposition 5.1 and Lemma 5.4 there exists a constant C > 0 such that kϕnkW 2,p(Rd,X) ≤ C(kϕnkLp(Rd,X) + k∆ϕnkLp(Rd,X)), for all n ∈ N. This shows that ϕn is Cauchy in W 2,p(Rd, X) and hence f ∈ W 2,p(Rd, X). (cid:3) We now want to show the converse of Proposition 5.2. We will need Proposition 5.5. Let 1 ≤ p < ∞. The operator ∆p is the generator of the strongly continuous Gaussian semigroup G on Lp(Rd, X) given by (G(t)f )(ξ) := (4πt)− d 2 ZRd f (ξ − η) exp(cid:18)− η2 4t (cid:19) dη, where t > 0, ξ ∈ Rd and f ∈ Lp(Rd, X). Proof. The assertion is well-known if X = R [3, Example 3.7.6]. Testing with x′ ∈ X ′ it follows immedeately that G is a semigroup. The strong continuity of G is also well-known [3, Lemma 1.3.3]. Let A be the generator of G and let ∆R p be the operator ∆p for X = R. Consider the space D := span{f ⊗ x, f ∈ D(∆R p ), x ∈ X}. Since D is dense in Lp(Rd, X) and invariant under the semigroup G it follows that D is a core for A [6, Proposition I.1.7]. Obviously, D ⊂ D(∆p) and ∆p coincides with A on D. Since ∆p is closed, it follows that A ⊂ ∆p. To show the inclusion A ⊃ ∆p note that λ ∈ ρ(A) for λ > 0. It remains to show that λ − ∆p is injective. But this follows immediately from the real-valued case. (cid:3) Theorem 5.6. The Banach space X has the UMD property if and only if D(∆p) = W 2,p(Rd, X) ELLIPTIC PROBLEMS AND HOLOMORPHIC FUNCTIONS 11 for some, equivalently all, 1 < p < ∞. Proof. It remains to show the "if" part. Since ∆p generates a C0 semi- group there exists µ > 0 such that µ ∈ ρ(∆p). By assumption we have kDjDkf kLp(Rd,X) ≤ kf kW 2,p(Rd,X) ≤ Ckµf − ∆f kLp(Rd,X), where C = kR(µ, ∆p)kL(Lp(Rd,X),W 2,p(Rd,X)) and 1 ≤ j, k ≤ d. This holds in particular for f ∈ C ∞ c (Rd, X). Taking the Fourier transform on both sides yields that the function m(ξ) := −4π2ξjξk µ + 4π2ξ (ξ ∈ Rd) is in MLp(Rd, X). We now use a scaling argument with the same notation as in the proof of Theorem 4.3. The transformation formula shows that F ±1fλ = λ−d(F ±1f )λ−1 for every f ∈ C ∞ multiplier mλ satisfies c (Rd, X). Hence the operator Tmλ assiociated with the Tmλf = F −1(mλF f ) = F −1((m(F f )λ−1)λ) = F −1((mλdF fλ)λ) = (F −1(mF fλ))λ−1 = (Tmfλ)λ−1, from which we can estimate kTmλf kLp(Rd,X) = k(Tmfλ)λ−1kLp(Rd,X) = λ d p kTmfλkLp(Rd,X) d p kTmkL(Lp(Rd,X))kfλkLp(Rd,X) ≤ λ = kTmkL(Lp(Rd,X))kf kLp(Rd,X). By symmetry we obtain kTmλkL(Lp(Rd,X)) = kTmkL(Lp(Rd,X)). Note that mλ → − ξjξk ξ2 =: m∞ pointwise and that mλ ≤ 1 for all λ > 0. By the dominated con- vergence theorem we have Tmf → Tm∞f . Fatou's lemma shows that m∞ ∈ MLp(Rd, X) and hence the claim follows from Theorem 4.1. (cid:3) ELLIPTIC PROBLEMS AND HOLOMORPHIC FUNCTIONS 12 6. Elliptic operators on domains In the last section we showed that the Laplacian on Lp(Rd, X) has the maximal regularity domain W 2,p(Rd, X) if and only if X is a UMD space. Our aim in this section is to show the analogous result for the Dirichlet Laplacian on a bounded domain Ω of class C 1,1. In fact, we also consider more general operators. Let L be an elliptic operator in non-divergence form given by L := aijDij + biDi + c, where aij, bi, c ∈ L∞(Ω, R) and a = (aij)ij is a symmetric matrix satis- fying aij(·)ξiξj ≥ λξ2 almost everywhere in Ω for some fixed λ > 0 and all ξ ∈ Rd. In this section we consider the Dirichlet problem (Lu = f u − ϕ ∈ W 1,p 0 (Ω, X), where f ∈ Lp(Ω, X) and ϕ ∈ W 2,p(Ω, X) are given. We will show that the existence of a unique solution is equivalent to the UMD property. We first start with the sufficiency of the UMD property. For L we have the following Lp estimate. Theorem 6.1. Let Ω ⊂ Rd be open and bounded with a C 1,1 boundary and let L be an elliptic operator as above. Furthermore assume that a ∈ C(Ω, Rd×d). Let X be a Banach space which has the UMD property and let 1 < p < ∞. Then there exists a constant C > 0 such that kukW 2,p(Ω,X) ≤ C(kukLp(Ω,X) + kLukLp(Ω,X)) for all u ∈ W 2,p(Ω, X) ∩ W 1,p 0 (Ω, X). Proof. Proceed as in the proof of [9, Theorem 9.13] proving the esti- mates for the Laplacian [9, Theorem 9.9] using Proposition 5.1 and also using the interpolation estimate in Lemma 5.4. (cid:3) To show existence we will need an estimate which does not depend on kukLp(Ω,X). As in Lemma 5.4, we cannot prove this estimate analo- gously to the real-valued case [9, Lemma 9.17] since this proof uses the Rellich-Kondrachov theorem. We gather some information about the real-valued case. Theorem 6.2. (a) Let Ω ⊂ Rd be open and bounded with a C 1,1- boundary and let L be an elliptic operator with a ∈ C(Ω, Rd×d) and c ≤ 0. Then for every data f ∈ Lp(Ω, R) and ϕ ∈ W 2,p(Ω, R) with 1 < p < ∞ there exists a unique u ∈ W 2,p(Ω, R) satisfying Lu = f and u − ϕ ∈ W 1,p 0 (Ω, R). ELLIPTIC PROBLEMS AND HOLOMORPHIC FUNCTIONS 13 (b) In the setting of (a) let ϕ = 0 and define T : Lp(Ω, R) → W 2,p(Ω, R) ∩ W 1,p 0 (Ω, R) via f 7→ u. Then −T is a positive operator, that is, T f ≤ 0 whenever f ≥ 0. Proof. (a) is the assertion of [9, Theorem 9.15]. For the proof of (b) we first let f ∈ C ∞ c (Ω, R)+. Then f ∈ Ld(Ω, R) and hence by uniqueness the solution u := T f is an element of W 2,d(Ω, R) ∩ W 1,d 0 (Ω, R). Fur- thermore it is continuous up to the boundary by Morrey's embedding theorem. Since Ω has a C 1,1 boundary this implies that u∂Ω = 0 in the classical sense. Suppose that u(ξ) > 0 for some ξ ∈ Ω. Then u has a nonnegative maximum in Ω. This contradicts the maximum principle [9, Theorem 9.6]. Now let f ≥ 0 be arbitrary. There exist nonnegative functions fn ∈ C ∞ c (Ω, R) such that fn → f in Lp(Ω, R). By the first step we know that the solution un := T fn is non-positive. The estimate in [9, Lemma 9.17] shows that un is Cauchy and hence convergent in W 2,p(Ω, R) ∩ W 1,p (cid:3) 0 (Ω, R). The uniqueness of the solution shows that u ≤ 0. Proposition 6.3. In the setting of Theorem 6.1 let c ≤ 0. Then we have the estimate kukW 2,p(Ω,X) ≤ CkLukLp(Ω,X), for some C > 0. Proof. Let T be the operator in Theorem 6.2 (b) considered as a bounded operator Lp(Ω, R) → Lp(Ω, R). The operator T can be linearly ex- tended to finite sums of tensors of the form f ⊗x with f ∈ Lp(Ω, R) and x ∈ X. Since −T is a positive operator there exists a unique bounded operator T with the same norm as T mapping Lp(Ω, X) → Lp(Ω, X) which coincides with T on finite sums of tensors [14, Theorem 2.1.3]. Note that T Lu = u and thus kukLp(Ω,X) ≤ k T kL(Lp(Ω,X))kLukLp(Ω,X). Combined with the estimate in Theorem 6.1 this yields the result. (cid:3) We are now in a position to prove the existence and uniqueness of strong solutions for the Poisson problem with Dirichlet boundary data. Theorem 6.4. Let Ω ⊂ Rd be open and bounded with a C 1,1 boundary and let L be an elliptic operator with a ∈ C(cid:0)Ω, Rd×d(cid:1) and c ≤ 0. Furthermore let X be a space which has the UMD property. Then for every data f ∈ Lp(Ω, X) and ϕ ∈ W 2,p(Ω, X) with 1 < p < ∞ there exists a unique u ∈ W 2,p(Ω, X) solving Lu = f such that u − ϕ ∈ W 1,p 0 (Ω, X). Proof. By subtracting Lϕ from f one sees that it is enough to consider k=1 fk ⊗ xk be a simple function with fk ∈ Lp(Ω, R) and xk ∈ X. For the data fk the real-valued Theorem the case ϕ = 0. Let first f = Pn ELLIPTIC PROBLEMS AND HOLOMORPHIC FUNCTIONS 14 6.2 yields existence of a solution uk ∈ W 2,p(Ω, R) ∩ W 1,p the function 0 (Ω, R). Thus n u := uk ⊗ xk Xk=1 is a solution for f . For general f ∈ Lp(Ω, X) there exists a sequence fn of finite sums of tensors which converges to f in Lp(Ω, X). Let un be the solution for fn. Then by Proposition 6.3 we have kun − umkW 2,p(Ω,X) ≤ Ckfn − fmkLp(Ω,X). Hence un → u in W 2,p(Ω, X)∩W 1,p follows from the Lp estimates. 0 (Ω, X) and Lu = f . The uniqueness (cid:3) The existence theorem has the following converse which gives again a characterization of the UMD property by a regularity property of the Poisson equation on domains. Corollary 6.5. Let 1 < p < ∞, Ω ⊂ Rd be an open and bounded set with a C 1,1 boundary and let X be a Banach space. The following are equivalent: (i) For every f ∈ Lp(Ω, X) there exists a unique u ∈ W 2,p(Ω, X) ∩ W 1,p 0 (Ω, X) satisfying ∆u = f . (ii) X has the UMD property. Proof. It remains to show the implication (i) ⇒ (ii). Let ω ⊂⊂ Ω be nonempty with a C 1,1 boundary. For f ∈ Lp(ω, X) denote by f the extension to Rd by 0. Recall that w := Φ ∗ f ∈ L1 loc(Rd, X) satisfies ∆w = f in D′(Ω, X). Let u be the solution for f according to (i). Then u −w solves ∆(u − w) = 0. Theorem 1.2 shows that u − w has a harmonic representative which is in particular in C 2(Ω, X). Since u ∈ W 2,p(ω, X) we also have w = u − (u − w) ∈ W 2,p(ω, X). Hence the Newtonian potential defines a mapping from Lp(ω, X) into W 2,p(ω, X). We claim that the graph of this mapping is closed. Let fn → f ∈ Lp(ω, X) such that Φ ∗ fn → w ∈ W 2,p(ω, X). Then for every x′ ∈ X ′ the functions hfn, x′i, hf, x′i and hw, x′i satisfy the analogue. R has the UMD property and thus Theorem 5.1 shows that hw, x′i = hΦ ∗ f, x′i. Choosing x′ from a countable separating subset of X ′ [14, Proposition B.1.10] yields the claim. Now the closed graph theorem shows the existence of a constant C ≥ 0 such that kΦ ∗ f kW 2,p(ω,X) ≤ Ckf kLp(ω,X) for all f ∈ Lp(ω, X). Finally, Theorem 4.3 shows that X has the UMD property. (cid:3) We want to relate Corollary 6.5 to the generator of the Dirichlet Laplacian. At first we establish an abstract result. Lemma 6.6. Let Ω ⊂ Rd be an open set, 1 ≤ p < ∞ and let T be a positive strongly continuous semigroup on Lp(Ω, R) with generator ELLIPTIC PROBLEMS AND HOLOMORPHIC FUNCTIONS 15 A. Let X be a Banach space. Then there exists a unique strongly continuous semigroup T on LP (Ω, X) satisfying x′◦ T (t)f = T (t)(x′◦f ) for all f ∈ Lp(Ω, X) and all x′ ∈ X ′. Denote by A its generator. Let f, g ∈ Lp(Ω, X). Then f ∈ D( A) and Af = g if and only if x′ ◦ f ∈ D(A) and A(x′ ◦ f ) = x′ ◦ g for all x′ ∈ X ′. Proof. By [14, Theorem 2.1.3] there is a unique bounded operator Tt on Lp(Ω, X) such that Tt(f ⊗ x) = Ttf ⊗ x for all f ∈ Lp(Ω, R) and all x ∈ X. This is the same as saying that x′ ◦ Ttf = Tt(x′ ◦ f ) for all f ∈ Lp(Ω, X) and x′ ∈ X ′. It is obvious from the first property that T := ( Tt)t≥0 is a strongly continuous semigroup on Lp(Ω, X). For f, g ∈ Lp(Ω, X) one has f ∈ D( A) and Af = g if and only if Tsg ds = Ttf − f for all t > 0. Using this and the corresponding 0 assertion for A the last claim follows from the fact that the integral commutes with functionals. (cid:3) R t Now let Ω ⊂ Rd be open and bounded with a C 1,1 boundary and let 1 ≤ p < ∞. Then the operator A given by D(A) := W 1,p Au := ∆u 0 (Ω, R) ∩ W 2,p(Ω, R) generates a positive strongly continuous semigroup T on Lp(Ω, R). Consider the induced semigroup T on Lp(Ω, X), where X is a Banach space, and denote by A its generator. Then clearly 0 (Ω, X) ∩ W 2,p(Ω, X) ⊂ D( A) W 1,p by the preceding lemma. The identity does not hold in general: Corollary 6.7. Let 1 < p < ∞ and let T, A, T and A be as above. Then W 1,p 0 (Ω, X) ∩ W 2,p(Ω, X) = D( A) if and only if X has the UMD property. Proof. By [14, Theorem 2.1.3] the norms of T and T coincide. Since kTtk ≤ Me−εt for all t > 0 the same estimate holds for T . Thus A is invertible. Now Corollary 6.5 yields the claim. (cid:3) References [1] W. Arendt. Approximation of degenerate semigroups. Taiwanese J. Math., 5(2):279 -- 295, 2001. [2] W. Arendt. Vector-valued holomorphic and harmonic functions. Concr. Oper., 3:68 -- 76, 2016. [3] W. Arendt, C. J. K. Batty, M. Hieber, and F. Neubrander. Vector-valued Laplace transforms and Cauchy problems, volume 96 of Monographs in Math- ematics. Birkhauser/Springer Basel AG, Basel, second edition, 2011. [4] W. Arendt and N. Nikolski. Vector-valued holomorphic functions revisited. Math. Z., 234(4):777 -- 805, 2000. ELLIPTIC PROBLEMS AND HOLOMORPHIC FUNCTIONS 16 [5] R. Dautray and J.-L. Lions. Mathematical analysis and numerical methods for science and technology. Vol. 1. Springer-Verlag, Berlin, 1990. [6] K.-J. Engel and R. Nagel. One-parameter semigroups for linear evolution equa- tions, volume 194 of Graduate Texts in Mathematics. Springer-Verlag, New York, 2000. With contributions by S. Brendle, M. Campiti, T. Hahn, G. Meta- fune, G. Nickel, D. Pallara, C. Perazzoli, A. Rhandi, S. Romanelli and R. Schnaubelt. [7] L. C. Evans. Partial differential equations, volume 19 of Graduate Studies in Mathematics. American Mathematical Society, Providence, RI, second edition, 2010. [8] S. Geiss, S. Montgomery-Smith, and E. Saksman. On singular integral and martingale transforms. Trans. Amer. Math. Soc., 362(2):553 -- 575, 2010. [9] D. Gilbarg and N. S. Trudinger. Elliptic partial differential equations of second order. Classics in Mathematics. Springer-Verlag, Berlin, 2001. Reprint of the 1998 edition. [10] J. D. Gray and S. A. Morris. When is a function that satisfies the Cauchy- Riemann equations analytic? Amer. Math. Monthly, 85(4):246 -- 256, 1978. [11] K.-G. Grosse-Erdmann. The Borel-Okada Theorem Revisited. Habilitationss- chrift, FernUniversitat Hagen, 1992. [12] K.-G. Grosse-Erdmann. A weak criterion for vector-valued holomorphy. Math. Proc. Cambridge Philos. Soc., 136(2):399 -- 411, 2004. [13] A. Grothendieck. Sur certains espaces de fonctions holomorphes. I. J. Reine Angew. Math., 192:35 -- 64, 1953. [14] T. Hytonen, J. van Neerven, M. Veraar, and L. Weis. Analysis in Banach spaces. Vol. I. Martingales and Littlewood-Paley theory, volume 63 of Ergeb- nisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics [Results in Mathematics and Related Areas. 3rd Se- ries. A Series of Modern Surveys in Mathematics]. Springer, Cham, 2016. [15] M. Kreuter, Vector-valued elliptic boundary value problems on rough do- mains, Open Access Repositorium der Universitat Ulm, Dissertation, http://dx.doi.org/10.18725/OPARU-11852, 2019 [16] R. E. Megginson. An introduction to Banach space theory, volume 183 of Grad- uate Texts in Mathematics. Springer-Verlag, New York, 1998. [17] V. Wrobel. Analytic functions into Banach spaces and a new characterization for isomorphic embeddings. Proc. Amer. Math. Soc., 85(4):539 -- 543, 1982. Wolfgang Arendt, Institute of Applied Analysis, Ulm University, 89069 Ulm, Germany E-mail address: [email protected] Manuel Bernhard, Institute of Applied Analysis, Ulm University, 89069 Ulm, Germany E-mail address: [email protected] Marcel Kreuter, Institute of Applied Analysis, Ulm University, 89069 Ulm, Germany E-mail address: [email protected]
1306.3356	3	1306	2013-10-09T15:05:36	Microlocal Properties of Bisingular Operators	[ "math.FA", "math.AP" ]	We study the microlocal properties of bisingular operators, a class of operators on the product of two compact manifolds. We define a wave front set for such operators, and analyse its properties. We compare our wave front set with the $SG$ wave front set, a global wave front set which shares with it formal similarities.	math.FA	math	MICROLOCAL PROPERTIES OF BISINGULAR OPERATORS MASSIMO BORSERO AND REN´E SCHULZ Abstract. We study the microlocal properties of bisingular operators, a class of operators on the product of two compact manifolds. We define a wave front set for such operators, and analyse its properties. We compare our wave front set with the SG wave front set, a global wave front set which shares with it formal similarities. Introduction Bisingular operators were originally introduced by L. Rodino in [Rod75] as a class of operators on a product of two compact manifolds Ω1 × Ω2, defined as linear and continuous operators A = Op(a) whose symbol satisfies, in local product-type coordinates, the estimate Dα1 ξ1 Dα2 ξ2 Dβ1 x1 Dβ2 x2 a(x1, x2, ξ1, ξ2) ≤ Cα1,α2,β1,β2hξ1im1−α1hξ2im2−α2. A simple and fundamental example of a bisingular operator is the tensor product A1 ⊗ A2 of two pseudodifferential operators, with symbols in the Hormander class, Ai ∈ Lmi(Ωi), i = 1, 2, while more complex examples include the vector-tensor product A1 ⊠A2 studied in [Rod75], and the double Cauchy integral operator studied in [NR06]. To each symbols of a bisingular operator we can associate two maps σ1 : Ω1 × Rn1 → Lm2(Ω2) σ2 : Ω2 × Rn2 → Lm1(Ω1), and with these maps the bisingular calculus takes the form of a calculus with vector valued symbols. General vector valued calculi have been deeply studied, for example, by B. V. Fedosov, B.-W. Schulze and N. N. Tarkhanov in [FST98]. In this paper we deal with bisingular operators whose symbols follow Hormander-type estimates (see e.g. [Hor85]), however a global version of bisingular calculus was defined by U. Battisti, T. Gramchev, S. Pilipovi´c and L. Rodino in [BGPR13]. In particular, we only study operators on compact manifolds, given explicitly in local coordinates. We also note that 'product-type' operators calculi, similar to bisingular calculus, were intro- duced by V. S. Pilidi [Pil73], R. V. Duducava [Dud79a], [Dud79b], and, more recently, by R. Melrose and F. Rochon [MR06]. Moreover, multisingular cal- culi were considered by V. S. Pilidi [Pil71] and L. Rodino [Rod80]. Applications of bisingular calculus include Index Theory, see e.g. [NR06], Analitic Number Theory, see e.g. [Bat12], and Geometric Analysis, see 2010 Mathematics Subject Classification. 35S05, 35A18, 35A27. Key words and phrases. Bisingular operators, Microlocal analysis, Wave front set. 1 2 M. BORSERO AND R. SCHULZ e.g. [GH13]. The aim of this paper is to study the microlocal properties of bisingular operators. In order to do this, we define a suitable wave front set for such operators, called the Bi-wave front set, which is the union of three compo- nents WFbi(u) = WF1 bi(u) ∪ WF2 bi(u) ∪ WF12 bi (u), u ∈ D′(Ω1 × Ω2). This definition is formulated using the calculus only, and, roughly speaking, has this interpretation: WF1 bi takes care of the singulari- ties which, in the wave front space, lie in the axis ξ2 = 0, WF2 bi takes care of the singularities which lie in the axis ξ1 = 0, and WF12 bi of the remaining singularities, which include all the classical ones. Our wave front set is re- lated to the classical Hormander wave front set WFcl (see [Hor83]) via the following inclusion WFclu ∩ (Ω1 × Ω2 × (Rn1 \ {0}) × (Rn2 \ {0})) ⊂ WF12 bi (u). Moreover, we have a global regularity result WFbi(u) = ∅ ⇔ u ∈ C∞(Ω1 × Ω2), which implies that the bi-wave front set encompasses all the singularities. Our main result is the following: Proposition 0.1. Let C be a bisingular operator, u ∈ D′(Ω1 × Ω2). Then WFbi(Cu) ⊂ WFbi(u). This Proposition shows the our wave front set is microlocal with respect to bisingular operators. Then, we define an appropriate notion of characteristic set for bisingular operators, given again as a union of three components, Charbi(C) := Char1 bi(C) ∪ Char2 bi(C) ∪ Char12 bi (C), and with this notion we can get a microellipticity result for the 1- and 2- components of the bi-wave front set Proposition 0.2. Let C be a bisingular operator, u ∈ D′(Ω1 × Ω2). Then WFi bi(u) ⊆ Chari bi(C) ∪ WFi bi(Cu), i = 1, 2. We note strong formal similarities between the bisingular calculus and the so called SG-calculus, introduced on Rn by H. O. Cordes [Cor95] and C. Parenti [Par72], see also R. Melrose [Mel95], Y. Egorov and B.-W. Schulze [ES97], and E. Schrohe [Sch87]. For this reason, our wave front set has for- mal connections to and similar features as the global SG-wave front set, or S -wave front set, introduced by S. Coriasco and L. Maniccia [CM03], see also R. Melrose [Mel94] for a geometric scattering version. The paper is organized as follows. In Section 1 we fix some notation and briefly review the bisingular calculus. In Section 2, following the ideas in [Hor83] and [GS94], we study the mapping properties of bisingular oper- ators and their microlocal properties with respect to the classical wave front set WFcl. In Section 3 we define the bi-wave front set and state the main results concercing microlocality and microellipticity of bisingular operators. MICROLOCAL PROPERTIES OF BISINGULAR OPERATORS 3 In Sections 4 we compare the bisingular calculus and the SG calculus, fo- cusing on the relations and differences between the bi-wave front set and the SG wave front set. Acknowledgements. We are grateful to Profs. D. Bahns, U. Battisti, S. Coriasco, L. Rodino, I. Witt for valuable advice and constructive criticism. This work was supported by the German Research Foundation (Deutsche Forschungsgemeinschaft, DFG) through the Institutional Strategy of the University of Gottingen, in particular through the research training group GRK 1493 and the Courant Research Center "Higher Order Structures in Mathematics". The second author is also grateful for the support received by the Studiens- tiftung des Deutschen Volkes and the German Academic Exchange Service (DAAD), as part of this collaboration was funded within the framework of a "DAAD Doktorandenstipendium". 1. Preliminaries 1.1. Introduction to bisingular calculus. In this section we will recall the main definitions and properties of bisingular symbols and bisingular operators with homogeneous principal symbols. For the Hormander pseu- dodifferential operators, whose tensor products provide the model example of bisingular operators, we use the notations from [Hor85]. Ωi, i = 1, 2, denotes an open domain of Rni. Definition 1.1. Sm1,m2(Ω1, Ω2) is the set of C∞(Ω1 × Ω2 × Rn1 × Rn2) functions such that, for all multiindex αi, βi and for all compact subsets Ki ⊂⊂ Ωi, i = 1, 2, there exists a constant Cα1,α2,β1,β2,K1,K2 > 0 such that Dα1 x2 a(x1, x2, ξ1, ξ2) ≤ Cα1,α2,β1,β2,K1,K2hξ1im1−α1hξ2im2−α2, for all xi ∈ Ki, ξi ∈ Rni. As usual, hξi := (1 + ξ2) 2 . An element of Sm1,m2(Ω1, Ω2) is called a symbol. Definition 1.2. A linear operator A : C∞ a bisingular operator if it can be written in the form 0 (Rn1+n2) → C∞(Rn1+n2) is called Dα2 ξ2 Dβ1 x1 Dβ2 ξ1 1 A(u)(x1, x2) = Op(a)[u](x1, x2) 1 (2π)n1+n2 ZRn1ZRn2 = ei(x1·ξ1+x2·ξ2) a(x1, x2, ξ1, ξ2)u(ξ1, ξ2) dξ1dξ2, where a ∈ Sm1,m2(Ω1, Ω2) and u denotes the Fourier transform of u. Lm1,m2(Ω1, Ω2) denotes the set of all bisingular operators with symbol in Sm1,m2(Ω1, Ω2). Moreover, we set S∞,∞(Ω1, Ω2) := [m1,m2 S−∞,−∞(Ω1, Ω2) := \m1,m2 Sm1,m2(Ω1, Ω2) Sm1,m2(Ω1, Ω2) and by L∞,∞(Ω1, Ω2), L−∞,−∞(Ω1, Ω2) the corresponding class of operators. The operators in L−∞,−∞(Ω1, Ω2) are called smoothing operators. 4 M. BORSERO AND R. SCHULZ We associate to every a ∈ Sm1,m2(Ω1, Ω2) the two maps A1 : Ω1 × Rn1 → Lm2(Ω2) (x1, ξ1) 7→ a(x1, x2, ξ1, D2), A2 : Ω2 × Rn2 → Lm1(Ω1) (x2, ξ2) 7→ a(x1, x2, D1, ξ2), and for a ∈ Sm1,m2(Ω1, Ω2), b ∈ Sp1,p2(Ω1, Ω2) we set (for fixed x1, x2 re- spectively) a ◦1 b(x1, x2, ξ1, D2) := (A1 ◦ B1)(x1, x2, ξ1, D2) ∈ Lm2+p2(Ω2) a ◦2 b(x1, x2, D1, ξ2) := (A2 ◦ B2)(x1, x2, D1, ξ2) ∈ Lm1+p1(Ω1). Definition 1.3. Let a ∈ Sm1,m2(Ω1, Ω2). Then a has a homogeneous prin- cipal symbol if i) there exists am1;· ∈ Sm1,m2(Ω1, Ω2) such that am1;·(x1, x2, tξ1, ξ2) = tm1am1;·(x1, x2, ξ1, ξ2), ∀x1, x2, ξ2, ∀ξ1 > 1, t > 0, a − ψ1(ξ1)am1;· ∈ Sm1−1,m2, where ψ1 is an 0-excision function. Moreover, am1;·(x1, x2, ξ1, D2) ∈ Lm2 cl (Ω2), so, being a classical symbol on Ω2, it admits an asymptotic expansion with respect to the ξ2 variable. ii) there exists a·;m2 ∈ Sm1,m2(Ω1, Ω2) such that a·;m2(x1, x2, ξ1, tξ2) = tm2a·;m2(x1, x2, ξ1, ξ2), ∀x1, x2, ξ1, ∀ξ2 > 1, t > 0, a − ψ2(ξ2)a·;m2 ∈ Sm1,m2−1, ψ2 as ψ1 above. Moreover, a·;m2(x1, x2, D1, ξ2) ∈ Lm1 cl (Ω1), so, being a classical symbol on Ω1, it admits an asymptotic expansion with respect to the ξ1 variable. iii) the symbols am1;· and a·;m2 have the same leading term, so there exists am1;m2 such that am1;· − ψ2(ξ2)am1;m2 ∈ Sm1,m2−1(Ω1, Ω2), a·;m2 − ψ1(ξ1)am1;m2 ∈ Sm1−1,m2(Ω1, Ω2), and a − ψ1am1;· − ψ2a·;m2 + ψ1ψ2am1;m2 ∈ Sm1−1,m2−1(Ω1, Ω2). The symbols which admit a full bi-homogeneous expansion in ξ1 and ξ2 are called classical symbols, their class is denoted by Sm1,m2 (Ω1, Ω2), and the corresponding operator class by Lm1,m2 (Ω1, Ω2). cl cl MICROLOCAL PROPERTIES OF BISINGULAR OPERATORS 5 The previous Definition implies that, given A ∈ Lm1,m2 define maps σ1, σ2, σ12 in this way cl (Ω1, Ω2), we can σ1(A) : T ∗Ω1 \ 0 → Lm2 cl (Ω2) (x1, ξ1) 7→ am1;·(x1, x2, ξ1, D2), σ2(A) : T ∗Ω2 \ 0 → Lm1 cl (Ω1) (x2, ξ2) 7→ a·;m2(x1, x2, D1, ξ2), σ12(A) : (T ∗Ω1 \ 0) × (T ∗Ω2 \ 0) → C (x1, x2, ξ1, ξ2) 7→ am1;m2(x1, x2, ξ1, ξ2), such that, denoting by σ(P )(x, ξ) the principal symbol of a pseudodifferential operator P , we have σ(σ1(A)(x1, ξ1))(x2, ξ2) = σ(σ2(A)(x2, ξ2))(x1, ξ1) = σ12(A)(x1, x2, ξ1, ξ2) = am1;m2(x1, x2, ξ1, ξ2). We call the couple (σ1(A), σ2(A)) the principal symbol of A. In the following, we only consider bisingular operators on the product of two compact manifolds Ω1, Ω2. They are defined as above in local coordinates. For those, there exists a notion of ellipticity, called bi-ellipticity. For more details, see [Rod75]. Definition 1.4. Let A ∈ Lm1,m2 cl (Ω1, Ω2). We say that A is bi-elliptic if i) σ12(A)(v1, v2) 6= 0 for all (v1, v2) ∈ (T ∗Ω1 \ 0) × (T ∗Ω2 \ 0); ii) σ1(A)(v1) is exactly invertible as an operator in Lm2 cl (Ω2) for all iii) σ2(A)(v2) is exactly invertible as an operator in Lm1 cl (Ω1) for all v1 ∈ T ∗Ω1 \ 0, with inverse in L−m2 v2 ∈ T ∗Ω2 \ 0, with inverse in L−m1 cl cl (Ω2); (Ω1). To elaborate on this definition, we give some examples. Example 1.5. Consider the differential operator cβ1,β2(x1, x2)Dβ1 1 Dβ2 2 , A = Xβ1≤m1 with C∞ coefficients. In this case (1) β2≤m2 (2) σ1(A)(x1, ξ1) = Xβ1=m1 σ2(A)(x2, ξ2) = Xβ1≤m1 β2≤m2 β2=m2 cβ1,β2(x1, x2)ξβ1 1 Dβ2 2 cβ1,β2(x1, x2)Dβ1 1 ξβ2 2 . A full bi-homogeneous expansion is given by σi,j(A)(x1, x2, ξ1, ξ2) = Xβ1=i cβ1,β2(x1, x2)ξβ1 1 ξβ2 2 . condition σ12(A) = The bi-ellipticity σm1,m2(A)(v1, v2) 6= 0 for all (v1, v2) ∈ (T ∗Ω1 \ 0) × (T ∗Ω2 \ 0) and given by of A is the β2=j 6 M. BORSERO AND R. SCHULZ the invertibility of the two maps (1) and (2). We may give a global meaning to A on a product of compact manifolds Ω1× Ω2 by taking, for example, Ωj = Tj, the nj-dimensional torus, j = 1, 2, and xj angular coordinates on Tj. With this in mind, it is possible to study some model cases of operators of the form A ⊗ B. In the following Table 1.1 we mean by ΨDO the classical pseudodifferential operators on Ω1 × Ω2, and by ΨDO-order and ΨDO-ell. their order and ellipticity, respectively. Operator I ⊗ I −∆1 ⊗ I + I ⊗ (−∆2) −∆1 ⊗ (−∆2) −∆1 ⊗ (−∆2 + I) ΨDO-order ΨDO-ell. Bi-order Bi-ell. √ × × × √ √ (0, 0) (2, 2) (2, 2) (2, 2) (2, 2) (−2,−2) √ √ × × × 0 2 4 4 4 (−∆1 + I) ⊗ (−∆2 + I) (−∆1 + I)−1 ⊗ (−∆2 + I)−1 not a ΨDO Table 1. Some model cases of bisingular operators of tensor product type Theorem 1.6. Let A ∈ Lm1,m2 B ∈ L−m1,−m2 (Ω1, Ω2) such that cl cl (Ω1, Ω2) be bi-elliptic. Then there exists AB = I + K1 BA = I + K2, where I is the identity map and K1, K2 are smoothing operators. Moreover the principal symbol of B is (σ1(A)−1, σ2(A)−1). From now on we will assume, for simplicity, that symbols of bisingular op- erators have compact support in the x1, x2 variables. Theorem 1.7. Let a ∈ Sm1,m2(Ω1, Ω2), b ∈ Sp1,p2(Ω1, Ω2). Then AB ∈ Lm1+p1,m2+p2(Ω1, Ω2), and its symbol c(x1, x2, ξ1, ξ2) has the asymptotic ex- pansion c ∼ ∞Xj=0 cm1+p1−j,m2+p2−j where cm1+p1−j,m2+p2−j = c1 m1+p1−j−1,m2+p2−j + c2 + c12 m1+p1−j,m2+p2−j m1+p1−j,m2+p2−j−1 MICROLOCAL PROPERTIES OF BISINGULAR OPERATORS 7 and c1 c2 c12     x2 b  x2 b  m1+p1−j−1,m2+p2−j = Xα2=j 1 α2! ∂α2 ξ2 a ◦1 Dα2 x2 b − Xα1≤j 1 α1! ∂α1 ξ1 ∂α2 ξ2 a Dα1 x1 Dα2 m1+p1−j,m2+p2−j−1 = Xα1=j 1 α1! ∂α1 ξ1 a ◦2 Dα1 x1 b − Xα2≤j 1 α2! ∂α1 ξ1 ∂α2 ξ2 a Dα1 x1 Dα2 m1+p1−j,m2+p2−j 1 = Xα1=α2=j α1!α2! ∂α1 ξ1 ∂α2 ξ2 a Dα1 x1 Dα2 x2 b Corollary 1.7.1. Let a ∈ Sm1,m2(Ω1, Ω2), b ∈ Sp1,p2(Ω1, Ω2). Then the com- mutator [A, B] := AB − BA belongs to Lm1+p1−1,m2+p2 + Lm1+p1,m2+p2−1. Proof. By Theorem 1.7 we have as leading order terms (j = 0): c1 m1+p1−1,m2+p2 = a ◦1 b − b ◦1 a c2 m1+p1,m2+p2−1 = a ◦2 b − b ◦2 a c12 m1+p1,m2+p2 = 0. Then, expanding c1 and c2 according to the definition of ◦j, j = 1, 2, we get cj = i{a, b}j + terms of order (mj + pj − 2), where with {a, b}j we denote the Poisson bracket of a and b in the j- argument. Therefore, the leading terms (up to order (m1+p1−2, m2+p2−2)) of the expansion of c can be written as c = 0 + i({a, b}1 + {a, b}2) ∈ Sm1+p1−1,m2+p2 + Sm1+p1,m2+p2−1. (cid:3) Remark 1.8. This behaviour under commutators is indeed something pe- culiar about bisingular calculus. It has the consequence that we can not use a lot of the common "commutator tricks" in the proofs to obtain microlocal properties. Example 1.9. For a better understanding of this phenomenon, consider the model case of a tensor product where A = A1 ⊗ A2 ∈ Lm1,m2, B = B1 ⊗ B2 ∈ Lp1,p2. Then [A, B] = [A1 ⊗ A2, B1 ⊗ B2] = A1B1 ⊗ A2B2 − B1A1 ⊗ B2A2 {z = [A1, B1] ⊗ A2B2 } Lm1+p1−1,m2+p2 + B1A1 ⊗ [A2, B2] } Lm1+p1,m2+p2−1 {z . To close this section, we note that all pseudodifferential operators of order zero or lower, in particular those corresponding to cut-offs and excision functions, are bisingular operators. Lemma 1.9.1. S0(Ω1 × Ω2) ⊂ S0,0(Ω1, Ω2). 8 M. BORSERO AND R. SCHULZ Proof. Let a ∈ S0(Ω1×Ω2). Then for all pair of multiindex α = (α1, α2), β = (β1, β2) we have Dβ1 x1 Dβ2 x2 a(x1, x2, ξ1, ξ2) = Dα ξ Dβ x a(x, ξ) ≺ hξi−α Dα1 ξ1 Dα2 ξ2 ≺ hξi−α1hξi−α2 ≺ hξ1i−α1hξ2i−α2, that is a ∈ S0,0(Ω1, Ω2). (cid:3) 2. Classical microlocal analysis of bisingular operators Let us first recall the notion of classical wave front set, as introduced by Hormander [Hor83]. Definition 2.1. Let Ω ⊂ Rn open, u ∈ D′(Ω). The distribution u is microlo- cally C∞ near (x0, ξ0) ∈ T∗Ω\ 0 if one of the following equivalent conditions is satisfied: (1) There exists a pseudodifferential operator A ∈ L0(Ω), non- characteristic at (x0, ξ0), such that Au ∈ C∞(Ω), (2) There exists a cut-off φ ∈ C∞ 0 (Ω) with φ ≡ 1 in an open set contain- ing x0 such that there exists a conic open set Γ ⊂ Rn \ 0 containing ξ0 and constants CN , R > 0 such that ∀ N ∈ N cφu(ξ) ≤ CNhξi−N ∀ξ ∈ Γ, ξ > R The classical wave front set of u ∈ D′(Ω), that we denote by WFcl(u), is the complement of the set of points where u is microlocally C∞. It is now interesting to compare, for given operators A, the sets WFcl(Au) and WFcl(u). An operator is microlocal, if WFcl(Au) ⊂ WFcl(u). 2.1. Mapping properties of bisingular operators. In this Section we shall estimate the classical wave front set WFcl(Au) for a linear operator A and a distribution u in terms of the Schwartz kernel KA of A, defined as follows: Definition 2.2. To each operator A : D(Ω1) → D′(Ω2) we can uniquely associate a distribution KA, called the Schwartz kernel of A, such that for all f ∈ D(Ω1), g ∈ D(Ω2) we have hAf, gi = hKA, f ⊗ gi. The Schwartz kernel of a bisingular operator with symbol a is then defined by the oscillatory integral (3) KA(x1, x2, y1, y2) =ZRn1+n2 ei(x1−y1,x2−y2)·(ξ1,ξ2)a(x1, x2, ξ1, ξ2)d−ξ1d−ξ2. The Schwartz kernel Theorem states the following smoothing property: Proposition 2.3. A linear map D′(Ω1 × Ω2) → D′(Ω1 × Ω2) is actually a mapping to D(Ω1 × Ω2), i.e. smoothing, if and only if its distributional kernel is in D((Ω1 × Ω2)×2). Therefore pseudodifferential operators with symbols in S−∞ and bisingular operators with symbol in S−∞,−∞ can be seen to be smoothing. MICROLOCAL PROPERTIES OF BISINGULAR OPERATORS 9 As the prototype of a bisingular operator is the tensor product of two pseu- dodifferential operators, it makes sense also to define what is meant by an operator that is smoothing in one set of variables only. Definition 2.4. We define the space C∞(Ω1,D′(Ω2)) as all such u ∈ D′(Ω1× Ω2) such that for each f ∈ D(Ω2), the distribution D(Ω1) ∋ u(g ⊗ ·) : g 7→ u(g ⊗ f ) is actually a smooth function. Correspondingly, we can define C∞(Ω2,D′(Ω1)). We can now list the mapping properties of bisingular operators on these spaces, following the ideas in [Tr`e67]. Lemma 2.4.1. Bisingular operators map the spaces C∞(Ω1,D′(Ω2)) and C∞(Ω2,D′(Ω1)) into themselves. Let a ∈ S−∞,m, then the bisingular operator Op(a) maps D′(Ω1 × Ω2) to C∞(Ω1,D′(Ω2)) and C∞(Ω2,D′(Ω1)) to C∞(Ω1 × Ω2). let a ∈ Sm,−∞, then the bisingular operator Op(a) maps Accordingly, D′(Ω1 × Ω2) to C∞(Ω2,D′(Ω1)) and C∞(Ω1,D′(Ω2)) to C∞(Ω1 × Ω2). The following Lemma (see e.g. [GS94]) indicates how the singularities of a distribution transform under the action of a linear operator in terms of the singularities of its kernel. Lemma 2.4.2. Let K ⊂ D′(Ω1 × Ω2), and denote by the same letter the corresponding operator K : C∞ WF′(K) := {(x1, x2, ξ1,−ξ2) ∈ T ∗(Ω1 × Ω2) \ 0; (x1, x2, ξ1, ξ2) ∈ WFcl(K)} WF′ WF′ Then if u ∈ E ′(Ω2) and WFcl(u) ∩ WF′ Ω1(K) := {(x1, ξ1) ∈ T ∗Ω1 \ 0; ∃y ∈ Ω2 with (x1, y, ξ1, 0) ∈ WF′(K)} Ω2(K) := {(x2, ξ2) ∈ T ∗Ω2 \ 0; ∃x ∈ Ω1 with (x, x2, 0, ξ2) ∈ WF′(K)}. 0 (Ω2) → D′(Ω1). Set Ω2(K) = ∅, we have WFcl(Ku) ⊂ WF′(K)(WFcl(u)) ∪ WF′ Ω1(K), where we considered WF′(K) as a relation T ∗Ω2 → T ∗Ω1. 2.2. Classical microlocality properties of bisingular operators. From the previous Lemma it is easy to derive that all pseudodifferential operators are microlocal. In fact, if K is the kernel of a pseudodifferential operator A on Ω = Ω1 = Ω2, then WF′ Ω2(K), WF′(K) = {(x, x, ξ,−ξ)}, hence from Lemma 2.4.2 we get WFcl(Au) ⊂ WFcl(u). We now study the microlocal properties of bisingular operators. Example 2.5. Consider Ω1 = Ω2 = R. We further pick positive φ, ψ ∈ 0 (R). Now define the pseudodifferential operator Tφ on C∞ C∞ Ω1(K) = ∅ = WF′ 0 (R) by Tφ(f ) := φ ∗ f. Then the operator A := Tφ ⊗ I is a tensor product of two pseudodifferential operators and thus a bisingular operator on R2. Now consider the distribution u = ψ⊗ δ. It has the following wave front set: WFcl(u) = {(x1, 0, 0, ξ2) x1 ∈ supp(ψ), ξ2 ∈ R \ 0} . Then it is easy to see that WFcl(Au) =(cid:8)(x1, 0, 0, ξ2) x1 ∈(cid:0) supp(ψ) + supp(φ)(cid:1), ξ2 ∈ R \ 0(cid:9) . 10 M. BORSERO AND R. SCHULZ The example can be similarly given in local coordinates on a product of two compact manifolds. We restricted ourselves to the Euclidean space for the sake of comprehensibility. The previous example shows that, in general, bisingular operators do not have the microlocal property. As a motivation, we start with the model case of a tensor product of two pseudodifferential operators. For that we use the well-known estimate for the wave front set of a tensor product of distributions (cf. e.g. [Hor83]): Lemma 2.5.1. Let u ∈ D′(Ω1), v ∈ D′(Ω2). Then WFcl(u ⊗ v) ⊂ WFcl(u) × WFcl(v) ∪ (supp(u) × {0}) × WFcl(v) ∪ WFcl(v) × (supp(v) × {0}). This can be used to estimate the wave front relation for a bisingular oper- ator given as the tensor product of two pseudodifferential operators. As a matter of fact this extends by direct calculation using standard techniques of integral regularization (see, e.g. [Hor83], [Shu01], [GS94]) to Theorem 2.6. Let A ∈ Lm1,m2(Ω1, Ω2). Then we can estimate the wave front set of the corresponding kernel as follows: WFcl(KA) ⊂ {(x1, x2, x1, x2, ξ1, ξ2,−ξ1,−ξ2)xi ∈ Ωi, ξi ∈ Rni \ 0} ∪ {(x1, x2, x1, y2, ξ1, 0,−ξ1, 0)xi, yi ∈ Ωi, ξ1 ∈ Rn1 \ 0} ∪ {(x1, x2, y1, x2, 0, ξ2, 0,−ξ2)xi, yi ∈ Ωi, ξ2 ∈ Rn2 \ 0}. Corollary 2.6.1. Let A ∈ Lm1,m2(Ω1, Ω2), u ∈ E ′(Ω1 × Ω2). Then WFcl(Au) ⊂ WFcl(u) ∪ {(x1, x2, 0, ξ2);∃y1 ∈ Ω1 : (y1, x2, 0, ξ2) ∈ WFcl(u)} ∪ {(x1, x2, ξ1, 0);∃y2 ∈ Ω2 : (x1, y2, ξ1, 0) ∈ WFcl(u)}. (cid:3) Proof. Use Theorem 2.6 and Lemma 2.4.2. Based on this observation we find that bisingular operators are microlo- cal with respect to a modified version of the classical wave front set. It is obtained by dropping the information about the precise location of the sin- gularities with the covariable ξ1 = 0 or ξ2 = 0 in the corresponding variable. Proposition 2.7. Let A ∈ Lm1,m2(Ω1, Ω2), u ∈ E ′(Ω1 × Ω2). Then WFcl(Au) ⊂ ]WFcl(u) := {(x1, x2, ξ1, ξ2) : (x1, x2, ξ1, ξ2) ∈ WFcl(u);ξ1ξ2 6= 0} ∪ {(x1, x2, ξ1, 0);∃y2 ∈ Ω2 : (x1, y2, ξ1, 0) ∈ WFcl(u)} ∪ {(x1, x2, 0, ξ2);∃y1 ∈ Ω1 : (y1, x2, 0, ξ2) ∈ WFcl(u)}. Proposition 2.8. Let A ∈ Lm1,m2 cl (Ω1, Ω2) be bi-elliptic. Then ]WFcl(Au) = ]WFcl(u). MICROLOCAL PROPERTIES OF BISINGULAR OPERATORS 11 Proof. We already know that ]WFcl(Au) ⊂ ]WFcl(u). Now, A ∈ Lm1,m2 that BA − I = K ∈ L−∞,−∞(Ω1, Ω2). Thus (Ω1, Ω2), therefore there exists B ∈ L−m1,−m2 cl cl ]WFcl(u) := ]WFcl((BA − K)u) ⊂ ]WFcl(BAu) ∪ ]WFcl(Ku) (Ω1, Ω2) such ⊂ ]WFcl(Au) ∪ ∅ = ]WFcl(Au). (cid:3) This notion encourages to study microlocal properties of bisingular operators with respect to this modified notion of gWF(u). However, all of the previous results have been obtained by the study of distribution kernels. It is far more desirable to have a notion of singularities in terms of the actual calculus. This will be provided in the next section. 3. A wave front set in terms of bisingular operators 3.1. Microlocal properties of bisingular operators. While being the description that naturally arises when analysing the kernels of bisingular op- erators, the notion of wave front set used in the previous section has several drawbacks: it is defined in terms of the Hormander wave front set, so in order to calculate it one first has to find that set and then "forget informa- tion". Also, it is not defined in terms of the bisingular calculus, but indeed with respect to the pseudodifferential one. In the following we establish a second notion that does not have these draw- backs. In fact it turns out to be quite similar to the notion introduced in [CM03] for the SG-calculus. This is not surprising, as there is a strong similarity in the formulas arising in both calculi. From now on, all the pseudodifferential operators will be assumed as prop- erly supported, and all the bisingular operators to be classical. Definition 3.1. Let u ∈ D′(Ω1 × Ω2). We define WFbi(u) ⊂ Ω1 × Ω2 × (Rn1+n2 \ 0) as WFbi(u) = WF1 bi(u) ∪ WF2 bi(u) ∪ WF12 bi (u) where • (x1, x2, ξ1, 0) is not in WF1 characteristic at (x1, ξ1) such that bi(u) if there exists an A ∈ L0 cl(Ω1) non- (A ⊗ I)u ∈ C∞(Ω1,D′(Ω2)). • (x1, x2, 0, ξ2) is not in WF2 characteristic at (x2, ξ2) such that bi(u) if there exists an A ∈ L0 cl(Ω2) non- (I ⊗ A)u ∈ C∞(Ω2,D′(Ω1)). • (x1, x2, ξ1, ξ2), ξ1ξ2 6= 0, is not in WF12 L0 cl(Ωi), non-characteristic at (xi, ξi), i = 1, 2, such that bi (u) if there exists a Ai ∈ (A1 ⊗ A2)u ∈ C∞(Ω1 × Ω2) (A1 ⊗ I)u ∈ C∞(Ω1,D′(Ω2)), (I ⊗ A2)u ∈ C∞(Ω2,D′(Ω1)). (4) (5) (6) (7) (8) 12 M. BORSERO AND R. SCHULZ Remark 3.2. Note that the conditions (7) and (8) do not follow from (6): Take u ∈ D′(R × R), u = δ(x − 1)δ(y + 1) and ψ smooth such that ψ ≡ 1 for x > 1/2 and ψ ≡ 0 for x ≤ 0. Then (ψ(x) ⊗ ψ(y))u = 0, as (1,−1) /∈ supp(ψ ⊗ ψ). However, for g ∈ D(R) with g(−1) 6= 0 we have (ψ(x) ⊗ I)u(. ⊗ g) = δ(x − 1)g(−1), which is not smooth. Remark 3.3. For a distribution of the form u = u1 ⊗ u2, we have that WF12 bi (u) = WFcl(u1) × WFcl(u2). In fact we have the following inclusion result: Lemma 3.3.1. If a point (x1, x2, ξ1, ξ2), ξ1ξ2 6= 0, is not in WF12 then it is not in WFcl(u). bi (u), Proof. The proof is a variant of [Hor91], Proposition 2.8. By definition there exists A := A1 ⊗ A2 ∈ L0,0(Ω1 × Ω2), with σ12(A)(x1, x2, ξ1, ξ2) 6= 0, such that Au ∈ C∞(Ω1 × Ω2). Now take a (ΨDO symbol) ψ(x1, x2, ξ1, ξ2) ∈ L0 cl(Ω1 × Ω2) such that • on the support of ψ we have hξ1i . hξ2i . hξ1i • ψ is non-characteristic at (x1, x2, ξ1, ξ2). Then the (bi-singular) operator product B := ψ(x1, x2, D1, D2) ◦ A(x1, x2, D1, D2) yields a pseudodifferential operator of order 0, plus a smoothing remainder, by virtue of the above inequality on the support of ψ and the symbol expand in Theorem 1.7. It has the following properties: the sense of ΨDOs) at (x1, x2, ξ1, ξ2) and of zero order. • its principal symbol is ψ · σ12(A), and thus is non-characteristic (in • Bu = ψ(x1, x2, D1, D2)A(x1, x2, D1, D2)u ∈ C∞ This proves the claim. (cid:3) Remark 3.4. This lemma asserts that in the conic region where both co- variables are non-vanishing we can pass from bisingular to pseudodifferential calculus by multiplying by a ΨDO. This has the consequence that the two operator classes have similar microlocal properties (with respect to the clas- sical wave front set) in that region. 1, y, ξ0 The following Lemma gives a similar interpretation to the remaining compo- nents, illustrating the loss of localization of singularities already encountered in the previous section. Lemma 3.4.1. Let u ∈ E ′(Ω1 × Ω2), (x0 all y ∈ Ω2 we have (x0 (x0 bi(u). Similarly, if for all x ∈ Ω1 we have (x, x0 we have (x, x0 Proof. We prove the claim for WF1 there exist a cut-off φ ∈ C∞ non-vanishing in a (conic) neighbourhood (x0 such that If for 1, 0) /∈ WFcl(u) then for all y ∈ Ω2 we have 2) /∈ WFcl(u) then for all x ∈ Ω1 1, 0) /∈ WFcl(u). Then 0 (Ω1×Ω2) and a conic localizer ψy ∈ C∞(Rn1+n2), 1, 0) respectively, 1, 0) /∈ WF1 2, 0, ξ0 bi. Take (x0 1, y, ξ0 1, ξ0 1) ∈ Ω1 × (Rn1 \ 0). 2 ) /∈ WF2 bi(u). 1, y) and (ξ0 1, y, ξ0 2, 0, ξ0 ψy(ξ1, ξ2)F(x1,x2)7→(ξ1,ξ2){φ(x1, x2)u} ∈ S (Rn1+n2). MICROLOCAL PROPERTIES OF BISINGULAR OPERATORS 13 As this holds true for any y, and due to compactness of the support of u, there exists a cut-off φ1 ∈ C∞ 0 (Ω1) such that for some conic localizer ψ ∈ C∞(Rn1+n2) ψ(ξ1, ξ2)F(x1,x2)7→(ξ1,ξ2){φ(x1)u} ∈ S (Rn1+n2). This means we can find a conic localizer ψ1 ∈ C∞(Rn1), non-vanishing 1, such that ψ(ξ1)F(x1,x2)7→(ξ1,ξ2){φ(x1)u} ∈ C∞(Rn1+n2), rapidly around ξ0 decaying with respect to the first variable ξ1 for fixed ξ2, and polynomially bounded everywhere, by the Paley-Wiener-Schwartz Theorem. We define A ∈ L0 cl(Ω1) as the operator Av(y1) = F −1 ξ17→y1 (ψ(ξ1)Fx17→ξ1{φ(x1)u}. By the assumptions on φ1 and ψ1, A is non-characteristic in the sense of pseudodifferential operators at (x0 1). But for any f ∈ D(Ω2) we have ξ17→x1hψ(ξ1)F(y1,y2)7→(ξ1,ξ2){φ(y1)u}, bfi is a smooth that [(A⊗I)u](f )(x1) = F −1 function, which means (A ⊗ I)u ∈ C∞(Ω1,D′(Ω2)). Proposition 3.5. Let u ∈ E ′(Ω1 × Ω2). Then 1, ξ0 (cid:3) WFbi(u) = ∅ ⇔ u ∈ C∞(Ω1 × Ω2). bi(u) = ∅ we can find for each (x1, x2, ξ1, ξ2) an A ∈ L0 Proof. Assume WFbi(u) = ∅. Then, by virtue of Lemma 3.3.1, we have WFcl(u) ∩ {(x1, x2, ξ1, ξ2) : ξ1ξ2 6= 0} = ∅. Thus bu is rapidly decaying on any ray R · (ξ1, ξ2) where ξ1ξ2 6= 0. As also WF1 cl(Ω1) non- characteristic at (x1, ξ1) such that, by Lemma 2.4.1, for any B ∈ Lm,−∞(Ω1× Ω2) we have B(I ⊗ A)u ∈ C∞(Ω1 × Ω2). By compactness and a parametrix construction we can thus conclude that for all B ∈ Lm,−∞(Ω1 × Ω2) we get Bu ∈ C∞(Ω1 × Ω2). Now pick φ ∈ C∞ 0 (Ω1 × Ω2) with φ ≡ 1 on the support of u, and define b(x1, x2, ξ1, ξ2) = φ(x1, x2)f (ξ2), with f ∈ S (Rd2). Then S (Rd1+d2) ∋ F(Bu) = F(cid:0)b(x1, x2, D1, D2)u(cid:1) = (1 ⊗ f )bu. As f was arbitrary and rapidly decaying, this means that bu must already be rapidly decaying in the first variable. Repeating the argument for the second variable proves the assertion. (cid:3) Using a parametrix construction, we get by the same arguments: Proposition 3.6. Let u ∈ E ′(Ω1 × Ω2). Then WF1 WF2 bi(u) = ∅ ⇔ u ∈ C∞(Ω1,D′(Ω2)), bi(u) = ∅ ⇔ u ∈ C∞(Ω2,D′(Ω1)). Remark 3.7. Note that u ∈ C∞(Ω1,D′(Ω2))TC∞(Ω2,D′(Ω1)) does not imply that u ∈ C∞(Ω1 × Ω2). Following [Tr`e67], a counterexample is, for instance, δ(x1−x2). The additional regularity needed such that u ∈ C∞(Ω1× Ω2) is therefore, by Proposition 3.5, encoded in WF12 bi (u). The bisingular wave front set admits the following properties: Proposition 3.8 (Properties of WFbi). Let u, v ∈ D′(Ω1 × Ω2), f ∈ C∞(Ω1 × Ω2). 14 M. BORSERO AND R. SCHULZ jointly. • WFbi a closed set and is conic with respect to both covariables • Let Ω1 = Ω2 = Ω. Define for f ∈ C∞(Ω× Ω) Au(x1, x2) = u(x2, x1). Then we can define the pull-back A∗ as an endomorphism on D′(Ω × Ω) by duality and we have (x1, x2, ξ1, ξ2) ∈ WFbi(Au) ⇔ (x2, x1, ξ2, ξ1) ∈ WFbi(u). • WFbi(u + v) ⊂ WFbi(u) ∪ WFbi(v); WFbi(f u) ⊂ WFbi(u). bi(u). bi(Cu) ⊂ WF1 Remark 3.9. These properties are quite similar to the ones the SG-wave front set of [CM03] admits. This is not very surprising, as the bisingular calculus is formally very similar in its definition to the SG-calculus, through which the SG-wave front set is introduced. We will explore this connection in Section 4. Lemma 3.9.1. Let C ∈ Lm1,m2(Ω1 × Ω2), u ∈ D′(Ω1 × Ω2). Then we have WF1 Proof. Let (x1, x2, ξ1, 0) /∈ WF1 bi(u). By definition there exists a non-char (at (x1, ξ1)) A ∈ L0 cl(Ω1) such that (A⊗ I)u ∈ C∞(Ω1,D′(Ω2). In particular, by Lemma 2.4.1 we have ∀B ∈ Lm,−∞(Ω1 × Ω2) that B(A ⊗ I)u is smooth. By the standard pseudodifferential calculus we can thus find an A′ ∈ L0(Ω1) such that A + A′ is elliptic in the sense of pseudodifferential operators and such that the symbol of A′ vanishes on a conic neighbourhood Γ of (x1, ξ1). We thus have a parametrix P ∈ L0(Ω1) such that P (A + A′) = I − R with R ∈ L−∞(Ω1). Take H ∈ L0(Ω1) such that H is non-characteristic at (x1, ξ1) and such that the symbol of H vanishes outside a proper subcone of Γ. Then we have: (H ⊗ I)Cu = (H ⊗ I)C(cid:0)(P (A + A′) + R) ⊗ I(cid:1)u = (H ⊗ I)C(P ⊗ I)(A ⊗ I)u + (H ⊗ I)C(P A′ ⊗ I)u+ + (H ⊗ I)C(R ⊗ I)u ∈ C∞(Ω1,D′(Ω2)). The first summand is in C∞(Ω1,D′(Ω2)) due to the definition of A, the second as the symbol expansion given in Theorem 1.7, using the support properties of the symbols of H and A′, gives an operator in L−∞,0. The third one is in C∞(Ω1,D′(Ω2)) as R ∈ L−∞,0 is already a smoothing operator in the first variable. This proves the claim. (cid:3) Lemma 3.9.2. Let C ∈ Lm1,m2(Ω1 × Ω2), u ∈ D′(Ω1 × Ω2). Then we have WF12 Proof. Let (x1, x2, ξ1, ξ2) /∈ WF12 exist Ai ∈ L0 cl(Ωi), non-char at (xi, ξi), i = 1, 2, such that bi (u). Then, by definition, we know there bi (Cu) ⊂ WF12 bi (u). (A1 ⊗ A2)u ∈ C∞(Ω1 × Ω2), (A1 ⊗ I)u ∈ C∞(Ω1,D′(Ω2)), (I ⊗ A2)u ∈ C∞(Ω2,D′(Ω1)). By the standard pseudodifferential calculus we can thus find A′ such that Ai + A′ such that the symbol of A′ i ∈ L0(Ωi) i is elliptic in the sense of pseudodifferential operators and i vanishes on a conic neighborhood Γi of (xi, ξi). MICROLOCAL PROPERTIES OF BISINGULAR OPERATORS 15 1) ⊗ (A2 + A′ We then have two parametrices Pi ∈ L0(Ωi) such that (P1 ⊗ P2)(cid:0)(A1 + A′ 2)(cid:1) = I ⊗ I − R1 ⊗ I − I ⊗ R2 − R1 ⊗ R2 {z } with Ri ∈ L−∞(Ωi). Now pick Hi ∈ L0(Ωi) such that Hi is non-characteristic at (xi, ξi) and such that the symbol of Hi vanishes outside a proper subcone of Γi. Then, recalling that by the standard pseudodifferential calculus if two operators have disjoint support their product is a smoothing operator, and using Lemma 2.4.1, :=R , (H1 ⊗ H2)Cu = 2)u ∈C∞by eq (6) {z {z ∈C∞by (8) and by the support of H1,A′ 1 (H1 ⊗ H2)C(cid:18)(P1 ⊗ P2)(cid:0)(A1 + A′ 1) ⊗ (A2 + A′ = (H1 ⊗ H2)C(P1 ⊗ P2)(A1 ⊗ A2)u } + (H1 ⊗ H2)C(P1 ⊗ P2)(A1 ⊗ A′ } 1,A′ 2 + (H1 ⊗ H2)C(R1 ⊗ I)u + (H1 ⊗ H2)C(I ⊗ R2)u + (H1 ⊗ H2)CRu } 2)(cid:1) + R1 ⊗ I + I ⊗ R2 + R(cid:19)u 1 ⊗ A2)u } } + (H1 ⊗ H2)C(P1 ⊗ P2)(A′ + (H1 ⊗ H2)C(P1 ⊗ P2)(A′ = (H1 ⊗ H2)C(R1 ⊗ I)u + (H1 ⊗ H2)C(I ⊗ R2)u modC∞. Now, without loss of generality, we proceed with the calculations only for the term (H1 ⊗ H2)C(R1 ⊗ I)u. We have ∈C∞by (7) and by the support of H2,A′ 2 {z {z ∈C∞by the support of H1,H2,A′ ∈C∞because R∈L−∞,−∞ 1 ⊗ A′ 2)u {z 2) + R2))u (H1 ⊗ H2)C(R1 ⊗ I)u = (H1 ⊗ H2)C(R1 ⊗ (P2(A2 + A′ = (H1 ⊗ H2)C(R1 ⊗ P2)(I ⊗ A2)u } + (H1 ⊗ H2)C(R1 ⊗ P2)(I ⊗ A′ } ∈ C∞, + (H1 ⊗ H2)C(R1 ⊗ I)(I ⊗ R2)u } {z {z {z ∈C∞by the support of H2,A′ 2 ∈C∞because R∈L−∞,−∞ ∈C∞by (8) 2)u therefore (H1 ⊗ H2)Cu ∈ C∞. With similar computations, one can check that (H1 ⊗ I)Cu ∈ C∞(Ω1,D′(Ω2)) (I ⊗ H2)Cu ∈ C∞(Ω2,D′(Ω1)) and this proves the claim. (cid:3) The preceding Lemmas lead to the following proposition, which asserts that this definition of wave front set is indeed suitable for the calculus of bisin- gular operators: Proposition 3.10 (Microlocality of bisingular operators). Let C ∈ Lm1,m2(Ω1 × Ω2), u ∈ D′(Ω1 × Ω2). Then we have WFbi(Cu) ⊂ WFbi(u). 16 M. BORSERO AND R. SCHULZ 3.2. Microelliptic properties of bisingular operators. From the pre- vious proposition, we can conclude that bielliptic operators preserve the bi-wave front set: Corollary 3.10.1. Let A ∈ Lm1,m2(Ω1, Ω2) be bi-elliptic. WFbi(Au) = WFbi(u). Then Proof. One inclusion follows directly form Proposition 3.10. The other fol- lows proceeding like in Proposition 2.8. (cid:3) Next we study the microellipticity properties of bisingular operators. For that we need a suitable definition of a characteristic set. As in Definition 1.4, Ω1 and Ω2 are considered as compact manifolds. Definition 3.11. Let B ∈ Lm1,m2(Ω1, Ω2) and v0 = (x1, ξ1) ∈ Ω1×(Rn1\0). We say that B is not 1-characteristic at V := π−1 2 v0 := {(x1, y, ξ1, 0) : y ∈ Ω2} if (1) for all v ∈ V there exists an open conic neighbourhood Θ of v such (2) σ1(B) ∈ Lm2 (Ω2) in an open that σ12 6= 0 on Θ \ (R+v), conic neighbourhood Γ of v0. cl (Ω2) is invertible with inverse in L−m2 cl bi(B) be the set of all V such that B is 1-characteristic at V . Let Char1 We define Char2 exchanging the roles of σ1(B) and σ2(B). Finally, we define Char12 0, where σ12(z) = 0. Set bi(B) accordingly for W := π−1 2 w0, w0 ∈ Ω2 × (Rn2 \ 0), by bi (B) as the set of points z = (x1, x2, ξ1, ξ2), ξ1ξ2 6= Charbi(B) := Char1 bi(B) ∪ Char2 bi(B) ∪ Char12 bi (B). Remark 3.12. B is bi-elliptic iff Charbi(B) = ∅. Remark 3.13. With this notion Definition 3.1 can also be expressed in the form cl(Ω1) Charcl(A)×(Ω2×(Rn2 \0)) (A⊗I)u∈C∞(Ω1,D′(Ω2)) cl(Ω2) (Ω1×(Rn1 \0))×Charcl(A) (I⊗A)u∈C∞(Ω2,D′(Ω1)) WF1 bi(u) = WF2 bi(u) = WF12 bi (u) = \A∈L0 \A∈L0 \Ai∈L0 cl(Ω1) Char1 Char2 bi(A ⊗ I), {z } bi(I ⊗ A), } {z bi (A1 ⊗ A2). Char12 (A1⊗A2)u∈C∞ (A1⊗I)u∈C∞(Ω1,D′(Ω2)) (I⊗A2)u∈C∞(Ω2,D′(Ω1)) With the definition of Charbi we can review in the following Table 3.2 the model cases of Table 1.1, setting Rn 0 = Rn1 0 × Rn2 0 . Lemma 3.13.1. Let C ∈ Lm1,m2(Ω1, Ω2) be such that bi(C) ∩ (Γ × Ω2 × {0}) = ∅. 0 = Rn \ 0, Ω = Ω1 × Ω2 and Rn12 Char1 MICROLOCAL PROPERTIES OF BISINGULAR OPERATORS 17 Operator I ⊗ I −∆1 ⊗ I + I ⊗ (−∆2) −∆1 ⊗ (−∆2) −∆1 ⊗ (−∆2 + I) (−∆1 + I) ⊗ (−∆2 + I) (−∆1 + I)−1 ⊗ (−∆2 + I)−1 Char1 bi Char2 bi Char12 bi 0 Ω × Rn12 0 ∅ ∅ 0 × {0} Ω × {0} × Rn2 Ω × Rn1 Ω × Rn1 0 × {0} Ω × {0} × Rn2 Ω × Rn1 0 × {0} ∅ ∅ ∅ ∅ ∅ 0 ∅ ∅ ∅ ∅ ∅ Table 2. Charbi for model cases of bisingular operators Let a ∈ S0(Ω1) have support in a closed cone Γ and be non-char (in the sense of ΨDO) in Γ0. Then there exists a H ∈ L−m1,−m2(Ω1, Ω2) such that HC = A ⊗ I − R, where R ∈ L−∞,0. Proof. The requirements on the support of a mean precisely that C is el- liptic with respect to a in the sense of [Cor95], Theorem 2.3.3. Therefore, by the classical calculus of pseudodifferential operators, we can find a sym- bol e ∈ S−m1,−m2 such that for all fixed (x2, ξ2) the operator E(x2, ξ2) = e(x1, x2, D1, ξ2) is a local parametrix with respect to a namely, E(x2, ξ2) σ2(C)(x2, ξ2) = R(x2, ξ2) + (A ⊗ 1)(x2, ξ2), where R(x2, D2) ∈ L−∞,0(Ω1, Ω2) and E(x2, D2) ∈ L−m1,−m2(Ω1, Ω2). Now define H as the operator with principal symbol h = ψ1(x1, ξ1)am1;·em1;· + ψ2(x2, ξ2)a·;m2c−1 ·;m2 − ψ1(x1, ξ1)ψ2(x2, ξ2)am1;m2c−1 m1;m2. Using the calculus and Theorem 1.7, it is straightforward that H matches the requirements. (cid:3) Lemma 3.13.2. Let C ∈ Lm1,m2(Ω1 × Ω2), u ∈ D′(Ω1 × Ω2). Then we have WF1 bi(u) ⊆ Char1 bi(C) ∪ WF1 bi(Cu). Proof. Let (x1, x2, ξ1, 0) /∈ Char1 bi(Cu). Then there exists A ∈ L0(Ω1) such that we have (A ⊗ I)HCu ∈ C∞(Ω1,D′(Ω2)), with H as in Lemma 3.13.1, due to microlocality (Proposition 3.10) of H. Then we have bi(C) ∪ WF1 (A2 ⊗ I)u = (A ⊗ I)(R − HC)u u − = (A ⊗ I)R } {z ∈ L−∞,0 (A ⊗ I)HCu } {z ∈ C∞(Ω1,D′(Ω2)) by assumption on A ∈ C∞. Using Lemma 2.4.1 on the first summand, this proves the claim. (cid:3) Using the previous Lemma, we conclude that Proposition 3.14 (Microellipticity of bisingular operators with re- spect to the 1 and 2 components of WFbi). Let C ∈ Lm1,m2(Ω1, Ω2), u ∈ D′(Ω1 × Ω2). Then WFi bi(u) ⊆ Chari bi(C) ∪ WFi bi(Cu), 18 i = 1, 2. M. BORSERO AND R. SCHULZ bi -component. This can be seen by the following example: We do, however, not obtain full microellipticity, i.e. with respect to the WF12 Example 3.15. Consider C = −∆ ⊗ (−∆), u = δ ⊗ 1 + 1 ⊗ δ. Then Char12 bi (Cu) = ∅. But take any A1 non-characteristic at (0, ξ1), ξ1 6= 0. Then (A1 ⊗ I)u = (A1δ) ⊗ 1 which is never in C∞(Ω1,D′(Ω2)), and this means that WF12 bi (u) is non- empty, because the (7)- and, by a similar argument, (8)-requirements fail to hold. bi (C) = ∅ and Cu = 0 ∈ C∞(Ω1 × Ω2), i.e. WF12 Remark 3.16. The counterexample to full microellipticity could be cir- cumvented by imposing stronger invertibility conditions in the definition of Char12 bi . This would, however, break the characterization of the wave front set of Remark 3.13, and lead to a loss of local information, while yielding no interesting cases not already covered by Corollary 3.10.1. With our definition we obtain, however, the following Lemma, which can be regarded as a microellipticity result for the (6)-part of Definition 3.1, for operators given by a tensor product. Lemma 3.16.1. Let Ci ∈ Lmi(Ωi), u ∈ D′(Ω1 × Ω2). If (x1, x2, ξ1, ξ2) /∈ Char12 bi (C1 ⊗ C2) ∪ WF12 bi ((C1 ⊗ C2)u) there exist operators Hi ∈ L0(Ωi), non-characteristic at (xi, ξi), such that (H1 ⊗ H2)u ∈ C∞(Ω1 × Ω2). Proof. By the standard pseudodifferential calculus we can pick Bi ∈ L−mi(Ωi) non-characteristic at (xi, ξi). Then the product BiCi ∈ L0 is non-characteristic at (xi, ξi). Proposition 3.10 and the definition of the bi- wave front set guarantee us the existence of Ai ∈ L0, non-characteristic at (xi, ξi), such that (A1 ⊗ A2)(B1C1 ⊗ B2C2)u ∈ C∞(Ω1 × Ω2). Therefore Hi := AiBiCi fulfils the claim. (cid:3) 4. Comparison with SG calculus In this section we will compare bisingular calculus with SG calculus. SG calculus is a global calculus obtained from the classical calculus by treating the variables and covariables equivalently by imposing on the symbols similar estimates as in bisingular calculus. These a priori formal similarities lead to interesting similarities in the calculus and in the analysis of singularities. However, the two calculi also differ in important aspects, as we will point out throughout the section. Definition 4.1. A function a(x, ξ) ∈ C∞(R2n) is called a SG symbol belong- ing to SGµ,m(Rn) := SGµ,m iff for every α, β ∈ Zn + there exists a constant Cα,β > 0 such that x Dβ Dα ξ a(x, ξ) ≤ Cα,βhξiµ−βhxim−α for every x, ξ ∈ Rn. A SG pseudodifferential operator is an operator of the form Au(x) :=Z eix·ξ a(x, ξ)u(ξ) d−ξ, u ∈ S , MICROLOCAL PROPERTIES OF BISINGULAR OPERATORS 19 and the class of operators with symbols in SGµ,m is denoted by LGµ,m. A symbol a ∈ SGm1,m2 is called SG classical if it admits a homogeneous ex- pansion with respect to ξ, for ξ >> 1, a homogeneous expansion in x, for x >> 1, and the two expansions satisfy certain compatibility conditions, we refer here to [Cor95, ES97] for a precise definition of classical symbols. We limit our attention (in this context) to classical operators, i.e. such that their symbols are SG classical. As usual one proceeds to develop a calculus for these operators. As every classical SG operator A is also classical pseu- dodifferential operator, it admits a principal symbol σψ(A), homogeneous in the first variable. In addition, by exchanging the roles of the variables and covariables one obtains a symbol σe(A), homogeneous in the second vari- able. The two satisfy a compatibility condition, i.e. that there is a third, bihomogeneous symbol σψe(A), the leading term of the corresponding ex- pansions of the ψ and e-symbols. The principal homogeneous symbol of the operator is then the couple (σψ(A), σe(A)) which gives rise to the principal symbol Symp(x, ξ) = φψ(ξ)σψ(A) + φe(x)σe(A) − φψ(x)φe(ξ)σψe(A), where φ∗ are 0-excision functions. So far, this is very similar to the bisingular calculus, but the expansion formula for the symbol of a product is in fact a lot simpler. The operator compositions arising there are not present, and the composition formula is just the one corresponding to the c12-term in Theorem 1.7. This leads to a definition of ellipticity close to our notion of 12-ellipticity, as no such thing as full invertibility of the symbols as operators is needed in the parametrix construction: Definition 4.2. A symbol a ∈ SGµ,m is SG-elliptic iff there exist constants R, C1, C2 > 0 such that C1hξiµhxim ≤ a(x, ξ) ≤ C2hξiµhxim when x + ξ ≥ R. With this notion of ellipticity, we have the Fredholm property, i.e. an SG- elliptic operator admits a parametrix in the calculus. Another important aspect to note is that in SG calculus we have Proposition 4.3. Let p ∈ SGm,µ(Rn), q ∈ SGr,ν(Rn). Then the commuta- tor [P, Q] := P Q − QP belongs to LGm+r−1,µ+ν−1(Rn). while in the bisingular setting the analogon in fact does not hold true, see Corollary 1.7.1 and Example 1.9 which illustrates that the difference stems from an interaction of the Ω1 and Ω2-parts of the operators which is not present in the SG case, where variables and covariables are independent. The notion of SG calculus can be used to introduce a global analysis of singularities. It turns out that this exhibits interesting similarities with the above analysis of bisingular operators. In the following, we refer to [Cor95], [CM03]; see also [CJT13a], [CJT13b]. First, we introduce SG characteristic sets and the SG wave front set. 20 M. BORSERO AND R. SCHULZ Definition 4.4 (SG characteristic sets). Let A ∈ LGm1,m2. Define the SG characteristic set of A as CharSG(A) = Charψ SG(A) ∪ Chare SG(A) ∪ Charψe SG(A), where Charψ Chare Charψe SG(A) = {(x, ξ) ∈ Rn × (Rn \ 0) : σψ(A)(x, ξ) = 0}, SG(A) = {(x, ξ) ∈ (Rn \ 0) × Rn : σe(A)(x, ξ) = 0}, SG(A) = {(x, ξ) ∈ (Rn \ 0) × (Rn \ 0) : σψe(A)(x, ξ) = 0}. Definition 4.5 (SG wave front set). Let u ∈ S ′(Rn). Define the SG wave front set of u as WFSG(u) = WFψ SG(u) ∪ WFψe SG(u), where SG(u) ∪ WFe SG(u) = \A∈LG0,0(Rn) SG(u) = \A∈LG0,0(Rn) SG(u) = \A∈LG0,0(Rn) Au∈S (Rn) Au∈S (Rn) Charψ SG(A), Chare SG(A), Charψe SG(A). WFψ WFe WFψe This notion exhibits the following properties: Au∈S (Rn) Proposition 4.6 (Properties of the SG wave front set). Let u, v ∈ S ′(Rn), f ∈ S (Rn). Then: (1) WFSG(u) is a closed set and WFψ variable x, WFe with respect to both of them independently, SG(u) is conical with respect to the SG(u) with respect to the covariable ξ and WFSG(u) (2) (x, ξ) ∈ WFSG(u) ⇔ (ξ,−x) ∈ WFSG(Fu) (Fourier Symmetry), (3) WFSG(u + v) ⊆ WFSG(u) ∪ WFSG(v); WFSG(f u) ⊆ WFSG(u), (4) WFSG(u) = ∅ ⇔ u ∈ S (Rn) (Global regularity). Already we see the similarities, but also apparent differences, with the bisin- gular notion: • Fourier transformation (i.e. exchange of variables and covariables) corresponds to the exchange of variables x1 and x2 in Proposition 3.8. • The conical properties of the individual components of the wave front sets correspond to the homogeneity properties of the corresponding principal symbol part. • The global (i.e.) S -regularity is of course due to the fact that SG- calculus imposes bounds on the variables also. Moreover, SG operators satisfy SG microlocalty and microellipticity, i.e. (just as in the bisingular case): MICROLOCAL PROPERTIES OF BISINGULAR OPERATORS 21 Proposition 4.7 (SG Microlocality and SG Microellipticity). Let u ∈ S ′(Rn) and C ∈ LGm,µ. Then we have the double inclusion WFSG(Cu) ⊆ WFSG(u) ⊆ CharSG(C) ∪ WFSG(Cu). This again stems from an individual double inclusion with respect to each wave front set component. A capital difference lies, however, in the structure of the wave front set. For the components of the SG wave front set have WFSG(u) ⊂(cid:0)Rn × (Rn \ 0)(cid:1) ∪(cid:0)(Rn \ 0) × Rn(cid:1) ∪(cid:0)(Rn \ 0) × (Rn \ 0)(cid:1). Here, the(cid:0)Rn×(Rn\0)(cid:1) ∋ (x, ξ) component corresponds exactly to singular- ities at finite arguments x with propagation direction ξ, the (cid:0)Rn × (Rn \ 0)(cid:1) (cid:0)(Rn \ 0) × (Rn \ 0)(cid:1) component corresponds to high oscillations present at component yields the same interpretation in the Fourier transformed space (growth singularities of u become differential singularities of u) and the infinite arguments. In the bisingular case, new phenomena are present. The 12-component has the classical interpretation, in the sense that it includes all the 'classical' sin- gularities (see Lemma 3.3.1), but the other components lose some amount of localization. This is reflected in the structure of the (1- and 2-components) of the bi-wave front set. In fact, WFbi(u) ⊂ (Ω1 × Ω2) × (Rn1+n2 \ 0) = (Ω1 × Ω2) ×(cid:0)(Rn1 \ 0) × {0}(cid:1) ⊔ (Ω1 × Ω2) ×(cid:0){0} × (Rn2 \ 0)(cid:1) ⊔ (Ω1 × Ω2) ×(cid:0)(Rn1 \ 0) × (Rn2 \ 0)(cid:1), where if e.g. for one x1 we have (x1, x2, 0, ξ2) present in the wave front set, all (y, x2, 0, ξ2) are present in the wave front set. This is due to the fact that bi-ellipticity involves true invertibility, i.e. a non-local requirement. Another difference arises as follows: the 1 and 2-component can be under- stood as the boundary faces of the 12-component, whereas in the SG case the ψe-component is interpreted as the corner of the wave front space where the e- and ψ-component meet, i.e. the roles as boundaries are interchanged, see Figure 1. This has the following consequence: SG(u) and WFψe SG(u) = (R \ 0) × (R \ 0). Example 4.8. Consider the one dimensional case. Following here Example in [CM03], there exists a distribution u(x) = eix2/2, x ∈ R, such that 2.7. WFψ SG(u) = ∅ = WFe However, there cannot exist a distribution v ∈ E ′(Ω1×Ω2), Ω1, Ω2 ⊂ R, such bi(v) = ∅ = WF2 that WF1 bi (v) = Ω1 × Ω2 × (R \ 0) × (R \ 0). This is because WF12 bi (v) = Ω1 × Ω2 × (R \ 0) × (R \ 0) is an open set, and WFbi(v) has to be closed. Similarly v = δ ⊗ 1 has WF1 bi (v) = ∅ = WF2 SG(u) = R × (R \ 0) but WFψe bi(v) = ({0} × Ω2) × ((R \ 0) × {0}), WF12 bi(v), but there cannot exist a distribution u such that WFe SG(u) = ∅, again due to closedness. bi(v) and WF12 [Bat12] U. Battisti, Weyl asymptotics of bisingular operators and Dirichlet divisor problems, Math. Z. (2012), no. 272, 1365 -- 1381. References 22 M. BORSERO AND R. SCHULZ ξ WFψ SG WFψe SG ξ2 WF2 bi WF12 bi WFe SG x WF1 bi ξ1 WFSG WFbi Figure 1. A schematic comparison of the components of WFSG and WFbi [BGPR13] U. Battisti, T. Gramchev, S. Pilipovi´c, and L. Rodino, Globally Bisingular El- liptic Operators, Operator Theory, Pseudo-Differential Equations, and Math- ematical Physics, Operator Theory: Advances and Applications, vol. 228, Birkhauser, Basel, 2013. [CJT13a] S. Coriasco, K. Johansson, and J. Toft, Global wave-front sets of Banach, Fr´echet and Modulation space types, and pseudo-differential operators, J. of Diff. Eq. (2013). [CJT13b] [CM03] [Cor95] , Local wave-front sets of Banach and Fr´echet types, and pseudo- differential operators, Monatshefte fur Mathematik 169 (2013), no. 3-4, 285 -- 316. S. Coriasco and L. Maniccia, Wave Front Set at Infinity and Hyperbolic Lin- ear Operators with Multiple Characteristics, Ann. Global Anal. Geom. (2003), no. 24, 375 -- 400. H. O. Cordes, The Technique of Pseudodifferential Operators, Cambridge Uni- versity Press, 1995. [Dud79a] R. V. Duducava, On the index of bisingular integral operators. I., Math. Nachr. (1979), no. 91, 431 -- 460. [Dud79b] , On the index of bisingular integral operators. II., Math. Nachr. (1979), [ES97] [FST98] [GH13] [GS94] [Hor83] [Hor85] [Hor91] no. 92, 289 -- 307. Y. V. Egorov and B.-W. Schulze, Pseudo-differential operators, singularities, applications, Operator Theory: Advances and Applications, vol. 91, Birkhauser Verlagl, Basel, 1997. B.V. Fedosov, B.-W. Schulze, and N.N. Tarkhanov, On the index of elliptic operators on a wedge, J. Funct. Anal. (1998), no. 156, 164 -- 209. D. Grieser Laplacian on Q-rank 1 Locally Symmetric Spaces, Preprint (2013), http://arxiv.org/abs/1212.3459v2. A. Grigis and J. Sjostrand, Microlocal Analysis for Differential Operators, An Introduction, Cambridge University Press, 1994. L. Hormander, The Analysis of Linear Partial Differential Operators, vol. I, Springer-Verlag, 1983. and E. Hunsicker, A Parametrix Construction for the in , The Analysis of Linear Partial Differential Operators, vol. III, Springer-Verlag, 1985. , Quadratic hyperbolic operators, Microlocal Analysis and Applications (L. Cattabriga and L. Rodino, eds.), Lecture Notes in Mathematics, vol. 1495, Springer Berlin Heidelberg, 1991, pp. 118 -- 160. MICROLOCAL PROPERTIES OF BISINGULAR OPERATORS 23 [Mel94] [Mel95] [MR06] [NR06] [Par72] [Pil71] [Pil73] [Rod75] R. Melrose, Spectral and scattering theory for the laplacian on asymptotically euclidian spaces, 1994. , Geometric scattering theory. stanford lectures, Cambridge University Press, 1995. R. Melrose and F. Rochon, Index in K-theory for families of fibred cusp oper- ators, K-theory 37 (2006), pp.25 -- 104. F. Nicola and L. Rodino, Residues and Index for Bisingular Operators, In: C ∗-algebras and elliptic theory, Trends Math., Birkhauser (2006), pp.187 -- 202. C. Parenti, Operatori pseudodifferenziali in Rn e applicazioni, Ann. Mat. Pura Appl. 93 (1972), 359 -- 389. V. S. Pilidi, Multidimensional bisingular operators, Dokl. Akad. Nauk SSSR (1971), no. 201, 787 -- 789. , Computation of the index of a bisingular operator., Funkcional. Anal. i Prilozen. (1973), no. 7, 93 -- 94. L. Rodino, A class of pseudo differential operators on the product of two man- ifolds and applications, Ann. Scuola Norm. Sup. Pisa Cl. Sci. (1975), no. 4, 287 -- 302. [Rod80] , Polysingular integral operators, Ann. Mat. Pura Appl. (1980), no. 124, [Sch87] 59 -- 106. E. Schrohe, Spaces of weighted symbols and weighted sobolev spaces on man- ifolds, Pseudo-Differential Operators, Lecture Notes in Math., vol. 1256, Springer, Berlin, 1987, pp. 360 -- 377. [Shu01] M. A. Shubin, Pseudodifferential Operators and Spectral Theory, second ed., [Tr`e67] Springer-Verlag, 2001. F. Tr`eves, Topological Vector Spaces, Distributions, and Kernels, Academic Press, 1967. Dipartimento di Matematica "Giuseppe Peano", Universit`a degli Studi di Torino, Via Carlo Alberto 10, 10123 Torino (TO), Italy. E-mail address: [email protected] Mathematisches Institut, Georg-August Universitat Gottingen, Bunsen- strasse 3 -- 5, D -- 37073 Gottingen, Germany E-mail address: [email protected]
1809.02593	1	1809	2018-09-07T17:37:21	Cyclic row contractions and rigidity of invariant subspaces	[ "math.FA", "math.OA" ]	It is known that pure row contractions with one-dimensional defect spaces can be classified up to unitary equivalence by compressions of the standard $d$-shift acting on the full Fock space. Upon settling for a softer relation than unitary equivalence, we relax the defect condition and simply require the row contraction to admit a cyclic vector. We show that cyclic pure row contractions can be "transformed" (in a precise technical sense) into compressions of the standard $d$-shift. Cyclic decompositions of the underlying Hilbert spaces are the natural tool to extend this fact to higher multiplicities. We show that such decompositions face multivariate obstacles of an algebraic nature. Nevertheless, some decompositions are obtained for nilpotent commuting row contractions by analyzing function theoretic rigidity properties of their invariant subspaces.	math.FA	math	CYCLIC ROW CONTRACTIONS AND RIGIDITY OF INVARIANT SUBSPACES RAPHAEL CLOU ATRE AND EDWARD J. TIMKO Abstract. It is known that pure row contractions with one-dimensional de- fect spaces can be classified up to unitary equivalence by compressions of the standard d-shift acting on the full Fock space. Upon settling for a softer rela- tion than unitary equivalence, we relax the defect condition and simply require the row contraction to admit a cyclic vector. We show that cyclic pure row contractions can be "transformed" (in a precise technical sense) into compres- sions of the standard d-shift. Cyclic decompositions of the underlying Hilbert spaces are the natural tool to extend this fact to higher multiplicities. We show that such decompositions face multivariate obstacles of an algebraic nature. Nevertheless, some decompositions are obtained for nilpotent commuting row contractions by analyzing function theoretic rigidity properties of their invari- ant subspaces. 1. Introduction The structure of Hilbert space contractions has long been recognized as deeply connected to that of concrete multiplication operators on spaces of analytic func- tions on the unit disc. The prototypical result in that direction is the fact that pure contractions may be viewed as compressions to co-invariant subspaces of unilateral shifts. This crucial fact was fleshed out significantly in the original edition of [37], and the resulting theory has been influential since then. A major theme in modern operator theory is the investigation of multivariate phenomena, prompted by the discovery of curious behaviour [30],[38]. The multi- variate counterpart to the Sz.-Nagy -- Foias theory was systematically developed by Popescu in a series of papers, starting with [32],[33]. The setting in this case is that of row contractions, with no commutation relation assumed between the individual operators. The non-commutative multivariate theory runs parallel to the univariate one, with most classical results finding a close analogue. The key insight into the commutative world was provided by Arveson in [5] who identified the appropriate space of analytic functions on the ball to serve as a replacement for the classical Hardy space. This space is now known as the Drury-Arveson space and is the centerpiece of many problems in function theory due its universal property [1]. It should also be pointed out that the commutative world can in fact be recovered ef- ficiently from the non-commutative one upon taking appropriate quotients, as laid out in [21],[35]. Going back to the foundational single-variable result, it is known that if T is a Hilbert space contraction with one-dimensional defect space (in the sense that I − T T ∗ has rank one), then it is unitarily equivalent to a compression of the 2010 Mathematics Subject Classification. 47A13, 47A20, 47A45 (47B32, 46L52). The first author was partially supported by an NSERC Discovery Grant. 1 2 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO standard unilateral shift on the Hardy space of the disc. In this case, the connection with function theory is especially transparent, and tools developed in the setting of the Hardy space can be brought to bear quite successfully to understand the structure of T . There are corresponding results in the multivariate setting [32],[29] that connect the study of row contractions to complex function theory in several variables. Unfortunately, the condition on the defect space being one-dimensional is quite restrictive, and it is not stable under natural perturbations such as scaling or simi- larity. It is thus desirable to replace it by a more flexible condition, such as cyclicity. The tradeoff, then, is that we cannot expect to describe T up to unitary equiva- lence with compressions of the standard shift anymore. This relaxation idea is the basis for the very successful Jordan model theory of constrained contractions (also known as C0-contractions) developed in [8]. The crucial result is that a cyclic pure contraction can be "transformed" into a compression of the standard shift. The multivariate analogue of this relaxation procedure seems to have been overlooked, and providing it is one of the main motivation behind this project. It should be noted here that there is a well-developed model theory for arbitrary row contrac- tions that takes place in spaces of vector-valued holomorphic functions [37, Theorem VI.2.3],[12],[31],[35]. However, the focus there is different than ours, and the two strands of the theory do not overlap beyond the case of one-dimensional defect spaces. Of course, it is desirable to deal with contractions T which are not necessarily cyclic as well. The natural approach for doing so is to decompose the underlying Hilbert space into cyclic subspaces that are invariant for T , and then apply the cyclic result on each piece. In [8], this is called the splitting principle, and allows the cyclic theory to be leveraged in the general case of arbitrary multiplicity and to culminate in a powerful classification theorem that mirrors the Jordan canonical form for matrices. Inspired by this success, at the onset of this project another one of our objectives was to obtain such cyclic decompositions in the multivariate context. Surprisingly, this ambition was met with serious obstacles, as we unraveled multivariate obstructions of a seemingly purely algebraic nature. This is an inter- esting discovery, and shows that if a satisfactory model theory is to emerge in this context, it will have to rely on entirely new ideas and tools. At least superficially, this is reminiscent of what prompted the work in [7]. Although cyclic decompositions appear elusive in general, we obtain some lim- ited partial results in this direction based on considerations of independent interest on function theoretic rigidity properties of invariant subspaces. More precisely, if M is an invariant subspace for a row contraction T , we are interested in describ- ing situations where M is completely determined by the kernel of the functional calculus associated to the restriction TM. We also explore the corresponding ques- tion for co-invariant subspaces. Such inquiries generate significant interest in the setting of Hilbert modules on reproducing kernel Hilbert spaces; see [14] and the references therein. Our work on this topic initiates a multivariate exploration of equivalence classes of invariant subspaces for constrained contractions, as studied in [11],[9],[10],[28],[15],[16]. We now describe the organization of the paper and state our main results. To do so, we require the following notation. For a pure commuting row contraction T , we denote by Annc(T ) the set of multipliers ϕ on the Drury-Arveson space CYCLIC ROW CONTRACTIONS AND RIGIDITY OF INVARIANT SUBSPACES 3 with the property that ϕ(T ) = 0. More detail can be found in Section 2, which contains background material and gathers many preliminary results that are used throughout. In Section 3, we investigate the class of commuting row contractions T that can be transformed into the standard commuting d-shift. Our analysis is based on the existence of cyclic vectors enjoying certain additional properties (Theorem 3.1). One of the main results of the paper is the following (see Corollary 3.7). Theorem 1.1. Let T = (T1, . . . , Td) be a pure, commuting, cyclic row contraction on some Hilbert space. Let M denote the compression of the standard commuting d-shift associated to Annc(T ). Then, we have that M is a quasi-affine transform of T . Interestingly, despite the commuting nature of the statement, our proof hinges crucially on the non-commuting dilation results of [32]. In Section 4, we turn to the question of rigidity of invariant and co-invariant subspaces. For general non-commuting row contractions, this problem is known to encounter significant difficulties, so we focus on the commuting setting. We illustrate that the rigidity witnessed in the univariate context can fail in several variables (Example 2). Nevertheless, some amount of rigidity can be established in various cases. In the special case of nilpotent row contractions, our results (Theorems 4.1 and 4.5) imply the following. Theorem 1.2. Let T = (T1, . . . , Td) be a nilpotent commuting row contraction on some Hilbert space H. The following statements hold. (1) Assume that T ∗ is cyclic. Let M, N ⊂ H be invariant subspaces for T such (2) Assume that T is cyclic. Let M ⊂ H be an invariant subspace for T such that Annc(TM) = Annc(TN). Then, M = N. that Annc(TM) = Annc(T ). Then, M = H. It is relevant to mention here that it is not generally true that T ∗ is cyclic if T is a cyclic commuting nilpotent row contraction (see Example 5). Regarding co-invariant subspaces, we mention here the finite-dimensional version of Theorem 4.3. Theorem 1.3. Let T = (T1, . . . , Td) be a cyclic pure commuting row contraction on some finite-dimensional Hilbert space H. Let M, N ⊂ H be co-invariant subspaces for T such that Annc(PMTM) = Annc(PNTN). Then, M = N. Finally, in Section 5 we examine the existence of invariant decompositions. If T = (T1, . . . , Td) is a d-tuple of operators on some Hilbert space H, a pair (M, N) of non-trivial invariant subspaces for T1, . . . , Td is an invariant decomposition for T if M ∩ N = {0} and M + N = H. Building on the results of Section 4, Theorem 5.1 establishes the existence of invariant decompositions for nilpotent commuting row contractions under some additional conditions. In another direction, Example 3 (an adaptation of an idea from [25]) exhibits a non-cyclic nilpotent commuting row contraction that admits no invariant decomposition with the property that TM is cyclic. This is surprising, as such cyclic decompositions always exist for single constrained contractions. We shed some light on these difficulties by con- sider separating vectors. These are known to be abundant for single constrained contractions, yet we show that they can fail to exist in several variables, even for 4 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO nilpotent commuting row contractions on finite-dimensional spaces (Example 5). On the other hand, we have the following (Theorem 5.3). Theorem 1.4. Let T = (T1, . . . , Td) be a nilpotent commuting row contraction on some Hilbert space H. Set A = C[x1, . . . , xd]/(C[x1, . . . , xd] ∩ Annc(T )). Then, for any positive integer s ≥ dimA, the ampliation T (s) has a dense set of separating vectors in H(s). One of the recurring themes of the paper is that nilpotent commuting row con- tractions exhibit rich and intricate behaviours that are somehow unique to the multivariate world. Indeed, as we will see throughout, such row contractions give rise to multivariate counter-examples to classical univariate theorems of interest. This is the motivation behind our focusing on the nilpotent case in several of our main results above. While some of our proofs can doubtlessly be adapted to the wider class of commuting row contractions satisfying other types of polynomial relations, we do not to pursue such generalizations in this paper. 2. Preliminaries 2.1. The Fock spaces, the Drury-Arveson space and some associated op- erator algebras. Throughout the paper, d is a fixed positive integer. Let F+ d denote the free semigroup on the symbols {1, . . . , d}. Let F2 d denote the full Fock space over Cd, that is F2 d = CΩ ⊕ (Cd)⊗n ∞Mn=1 where Ω is some distinguished unit vector. Fix an orthonormal basis {e1, . . . , ed} of Cd. If w = k1k2 . . . ks ∈ F+ d , then we set ew = ek1 ⊗ ek2 ⊗ . . . ⊗ eks ∈ F2 d. The set {ew : w ∈ F+ Lk : F2 d → F2 d } is an orthonormal basis of F2 d denote the left creation operator acting as d. For each 1 ≤ k ≤ d, we let Lkv = ek ⊗ v, v ∈ F2 d. It is readily verified that the operators L1, . . . , Ld are isometries with pairwise orthogonal ranges, so that dXk=1 LkL∗ k ≤ I. The weak-∗ closed unital subalgebra of B(F2 d) generated by L1, . . . , Ld is denoted by Ld, and is usually referred to as the non-commutative analytic Toeplitz algebra [33],[20],[21]. Likewise, the norm closed unital subalgebra of B(F2 d) generated by L1, . . . , Ld is denote by Ad, and is usually called the non-commutative disc algebra [33]. The subspace Fs d ⊂ F2 d spanned by the symmetric elementary tensors is the symmetric Fock space, and it is co-invariant for L1, . . . , Ld. It was recognized by Arveson [5] that Fs d can be identified with a space of analytic functions on the open unit ball Bd ⊂ Cd. Nowadays, this space H 2 d is called the Drury-Arveson space. CYCLIC ROW CONTRACTIONS AND RIGIDITY OF INVARIANT SUBSPACES 5 For each 1 ≤ k ≤ d, we denote by xk : Cd → C the coordinate function, so that xk(z) = zk for z = (z1, . . . , zd) ∈ Cd. We write N for the set of non-negative integers. We use the standard multi-index notation: for The Drury-Arveson space H 2 with powers series expansion d is the Hilbert space of analytic functions f on Bd we set and satisfying α = (α1, . . . , αd) ∈ Nd α = α1 + ··· + αd, α! = α1!··· αd! 1 ··· xαd xα = xα1 d . f = Xα∈Nd cαxα α! α1!··· αd! Xα∈Nd α1!··· αd!(cid:19)1/2 α! (cid:18) cα2 < ∞. xα, α ∈ Nd. An orthonormal basis for H 2 d is given by the weighted monomials See [5] or [21] for more details. Another useful interpretation of H 2 reproducing kernel Hilbert space on Bd (see [2]) with kernel given by the formula d is that it is the k(z, w) = 1 , z, w ∈ Bd. 1 − hz, wi We denote by Md the multiplier algebra of H 2 there is a bounded linear operator Mϕ ∈ B(H 2 Mϕf = ϕf, d . Associated to a multiplier ϕ ∈ Md d ) defined as f ∈ H 2 d . Identifying each multiplier in Md with its associated multiplication operator, we may view Md as a weak-∗ closed unital subalgebra of B(H 2 d ). In particular, Md is an operator algebra. Under this identification, the multiplier norm of ϕ ∈ Md is simply the operator norm of Mϕ. Every polynomial is a multiplier, and we denote by Ad the norm closure of the polynomials in Md. We will require the following basic fact about multipliers. Theorem 2.1. Let ϕ ∈ Md. Fix n ∈ N and w ∈ Bd. For each α ∈ Nd with α = n, there exists a multiplier ϕα ∈ Md with the property that (x − w)αϕα. ∂αϕ ∂xα (w) (x − w)α α! ϕ = Xα<n + Xα=n Proof. Apply [24, Cor. 4.2] inductively. (cid:3) Next, we record useful identifications which are well known. The details of the proof appear to be difficult to track down explicitly, so we provide them for the reader's convenience. Theorem 2.2. The following statements hold. 6 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO (1) There is a unitary operator U : H 2 LkFs U Mxk U ∗ = PFs d d → Fs , d d with the property that 1 ≤ k ≤ d. (2) Let Wd ⊂ Ld denote the weak-∗ closure of the commutator ideal of Ld.Then, there is a unital completely isometric and weak-∗ homeomorphic algebra isomorphism such that Ψ : Ld/Wd → Md Ψ(Lk + Wd) = U ∗(PFs d LkFs d )U = Mxk for every 1 ≤ k ≤ d. there is a unital completely isometric algebra isomorphism (3) Let Cd ⊂ Ad denote the norm closure of the commutator ideal of Ad. Then, such that Φ : Ad/Cd → Ad Φ(Lk + Cd) = Mxk , 1 ≤ k ≤ d. Proof. (1) The existence of U follows from the discussion in [5, Section 1]. (2) This follows from (1) and [21, Corollary 2.3]. (3) This argument is based on the idea behind the proof of [19, Corollary 8.3]. Let q : Ad → Ad/Cd denote the quotient map. By the Blecher -- Ruan -- Sinclair characterization of operator algebras [13], there is a Hilbert space H and a unital completely isometric homomorphism π : Ad/Cd → B(H). Then, (π(q(L1)), . . . , π(q(Ld))) is a commuting row contraction on H. In particular, invoking [5, Theorem 6.2] for instance, we obtain a unital completely contractive homomorphism such that σ(Mxk ) = π(q(Lk)) for every 1 ≤ k ≤ d. Next, let ι : Ad/Cd → Ld/Wd be defined as σ : Ad → π(Ad/Cd) It is readily checked that Cd is the norm closure of the ideal generated by ι(T + Cd) = T + Wd, T ∈ Ad. {LjLk − LkLj : 1 ≤ j, k ≤ d} inside of Ad. Furthemore, it follows from [20, Proposition 2.4] that Wd is the weak-∗ closure of the ideal generated by {LjLk − LkLj : 1 ≤ j, k ≤ d} inside of Ld. Hence, Cd ⊂ Wd so that ι is a well-defined unital completely contrac- tive homomorphism. We note that (Ψ ◦ ι ◦ π−1 ◦ σ)(Mxk ) = Mxk for every 1 ≤ k ≤ d, whence Ψ ◦ ι ◦ π−1 ◦ σ is simply the inclusion of Ad inside Md. This forces π−1 ◦ σ : Ad → Ad/Cd to be completely isometric and in particular surjective, since its image contains Lk + Cd for every 1 ≤ k ≤ d. Take Φ = π−1 ◦ σ. (cid:3) CYCLIC ROW CONTRACTIONS AND RIGIDITY OF INVARIANT SUBSPACES 7 When d = 1, the space H 2 1 coincides with the usual Hardy space on the open unit disc D. In this case, the operator algebras A1 and M1 can be identified with the disc algebra A(D) and the algebra of bounded analytic functions H ∞(D), respectively. It is important to note that when d > 1, however, the quantities kMϕk and kϕk∞ = sup z∈Bd ϕ(z) are in general not comparable [5],[21]. This makes the function theory in H 2 d and Md less transparent than it is in the algebra of bounded holomorphic functions H ∞(Bd). 2.2. Dilations to d-shifts. We recall the main features of the multivariate gener- alization of the classical Sz.-Nagy -- Foias model theory from [37], starting with the non-commuting setting. define the standard d-shift to be the d-tuple of operators L = (L1, L2, . . . , Ld) acting on F2 d. More generally, a d-tuple V = (V1, . . . , Vd) of operators on some Hilbert space H is called a d-shift if there is another Hilbert space K and a unitary operator U : H → F2 U VkU ∗ = Lk ⊗ IK, d ⊗ K such that 1 ≤ k ≤ d. The cardinal number dim K is then called the multiplicity of the d-shift. Interest- ingly, the multiplicity of a d-shift can be interpreted in a different way as the next theorem shows. Recall that a subset Γ ⊂ H is cyclic for a d-tuple T = (T1, . . . , Td) if Here, given a word w = i1i2 . . . is ∈ F+ d , we use the notation H = span{Twγ : w ∈ F+ d , γ ∈ Γ}. Tw = Ti1 Ti2 . . . Tis. Theorem 2.3. Let V = (V1, . . . .Vd) be a d-shift. Then, the multiplicity of V coincides with the least cardinality of a cyclic set for V . Proof. This follows from [35, Corollary 1.10]. (cid:3) The importance of d-shifts lies in the fact that they can be used to model gen- eral d-tuples of operators satisfying some natural necessary conditions, as we now illustrate. A d-tuple T = (T1, . . . , Td) of operators on some Hilbert space H is said to be a row contraction if the row operator defined as [T1, . . . , Td]ξ = ξ = (ξ1, . . . , ξd) ∈ H(d) [T1, . . . , Td] : H(d) → H dXk=1 H(d) = H ⊕ H ⊕ . . . ⊕ H. Tkξk, is contractive. Here, we use the standard notation Equivalently, T is a row contraction if and only if dXk=1 TkT ∗ k ≤ I. Furthermore, we define the defect space of T to be lim kT ∗ n→∞ Xw=n ∆T = I − wξk2 = 0, ξ ∈ H. k! H, TkT ∗ dXk=1 8 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO The row contraction T is said to be pure if and the defect of T to be dT = dim ∆T . A straightforward calculation reveals that d-shifts are pure row contractions. The following result shows that in some sense they are the universal example. Theorem 2.4. Let T = (T1, . . . , Td) be a d-tuple of operators on some Hilbert space H. Then, the following statements are equivalent. (i) The d-tuple T is a pure row contraction. (ii) There is another Hilbert space K containing H along with a d-shift V = (V1, . . . , Vd) on K such that In this case, the multiplicity of the d-shift V is dT . T ∗ k = V ∗ k H, 1 ≤ k ≤ d. Proof. This is [32, Theorems 2.1 and 2.8]. (cid:3) Of particular interest for us in this paper are row contractions T = (T1, . . . , Td) with the property that T1, . . . , Td commute with each other. In this case, we say that T is a commuting row contraction. While Theorem 2.4 applies just as well to commuting row contractions, it fails to encode the commutativity. In particular, one may want to replace the d-shift therein by some "commutative" version which acts on the symmetric Fock space (or Drury-Arveson space), as opposed to the full Fock space. To make this precise, recall that for each 1 ≤ k ≤ d the operator Mxk ∈ B(H 2 d ) denotes the multiplication operator by the coordinate function xk. A standard calculation reveals that the commuting d-tuple Mx = (Mx1, . . . , Mxd) is a pure row contraction. We will refer to this commuting row contraction as the standard commuting d-shift. More generally, a commuting d-tuple Z = (Z1, . . . , Zd) of operators on some Hilbert space H is called a commuting d-shift if there is another Hilbert space K and a unitary operator U : H → H 2 d ⊗ K such that U ZkU ∗ = Mxk ⊗ IK, 1 ≤ k ≤ d. The cardinal number dim K is then called the multiplicity of the commuting d-shift. Just like their non-commuting counterparts, commuting d-shifts occupy a universal role among pure commuting row contractions. Theorem 2.5. Let T = (T1, . . . , Td) be a d-tuple of operators on some Hilbert space H. Then, the following statements are equivalent. (i) The d-tuple T is a pure commuting row contraction. (ii) There is another Hilbert space K containing H along with a commuting d- shift Z = (Z1, . . . , Zd) on K such that In this case, the multiplicity of the commuting d-shift Z is dT . k = Z ∗ T ∗ kH, 1 ≤ k ≤ d. CYCLIC ROW CONTRACTIONS AND RIGIDITY OF INVARIANT SUBSPACES Proof. This is [5, Theorem 4.5] (alternatively, see [35, Theorem 2.1]). 9 (cid:3) 2.3. Functional calculus and constrained row contractions. In [33], Popescu extends the classical von Neumann inequality to the setting of row contractions and defines an appropriate functional calculus, as follows. Theorem 2.6. Let T = (T1, . . . , Td) be a row contraction on some Hilbert space H. Then, there is a unital completely contractive homomorphism αT : Ad → B(H) such that αT (Lk) = Tk for every 1 ≤ k ≤ d. If T is pure, then αT extends to a unital completely contractive and weak-∗ continuous homomorphism βT : Ld → B(H). Proof. Combine [33, Theorems 3.6 and 3.9] with [34, Theorem 2.1]. (cid:3) There is a version of this result adapted to the commuting setting, and the resulting homomorphisms are naturally compatible. Corollary 2.7. Let T = (T1, . . . , Td) be a commuting row contraction on some Hilbert space H. Let Cd ⊂ Ad and Wd ⊂ Ld denote respectively the norm closure and the weak-∗ closure of the commutator ideals. Let q : Ad → Ad/Cd and qw : Ld → Ld/Wd be the corresponding quotient maps. Let Φ : Ad/Cd → Ad and Ψ : Ld/Wd → Md be the canonical isomorphisms from Theorem 2.2. Then, the following statements hold. T : Ad/Cd → B(H) such that αT = α′ T ◦ Φ−1, and (1) follows. T : Ld/Wd → B(H) such that βT = β′ (cid:3) (1) There is a unital completely contractive homomorphism cαT : Ad → B(H) such that cαT (Mxk ) = Tk for every 1 ≤ k ≤ d. Moreover, we have αT = cαT ◦ Φ ◦ q. (2) If T is pure, then cαT extends to a unital completely contractive and weak-∗ continuous homomorphism cβT : Md → B(H). Moreover, we have βT = cβT ◦ Ψ ◦ qw. Proof. Since T is commuting, we see that Cd ⊂ ker αT , so there is a unital com- pletely contractive homomorphism α′ T ◦ q. Put For (ii), we note that Wd ⊂ ker βT , so there is a unital completely contractive homomorphism β′ T ◦ qw. A standard compact- ness argument using the Krein-Smulyan theorem shows that β′ T is in fact weak-∗ T ◦ Ψ−1. cαT = α′ continuous. Statement (2) follows upon setting cβT = β′ We typically make no explicit mention of the homomorphisms αT ,cαT , βT and cβT defining the functional calculi, and for an appropriate element θ we simply write The existence of a weak-∗ continuous extension to Md of the Ad -- functional calculus for a commuting row contraction does not require T to be pure. Indeed, the class of so-called absolutely continuous commuting row contractions is larger [17],[18]. Nevertheless, for the purposes of this work we will restrict our attention to the pure case. Let T = (T1, . . . , Td) be a pure row contraction. Then, the non-commutative annihilator of T is the weak-∗ closed two-sided ideal of Ld given by Annnc(T ) = {θ ∈ Ld : θ(T ) = 0}. We say that T is constrained if Annnc(T ) is non-trivial. If T happens to be com- muting, then its commutative annihilator is the weak-∗ closed ideal of Md given θ(T ). 10 by RAPHA EL CLOU ATRE AND EDWARD J. TIMKO It follows from Corollary 2.7 that Annc(T ) = {θ ∈ Md : θ(T ) = 0}. so in particular (1) Annnc(T ) = (Ψ ◦ qw)−1(Annc(T )) Annc(T ) = (Ψ ◦ qw)(Annnc(T )). In the classical case where d = 1, a constrained contraction T is usually said to be of class C0 [8]. We feel our choice of terminology is a bit more descriptive, and it is consistent with its use in [35]. We also mention that an absolutely continuous contraction for which the associated H ∞ -- functional calculus has non-trivial kernel is automatically pure, and thus constrained. To see this, combine the proof of [8, Lemma II.1.12] with [17, Theorem 3.2]. For commuting row contractions, the situation is more complicated: see [17, Theorem 5.4] and the discussion that follows it. Let b ⊂ Ld be a weak-∗ closed two-sided ideal. We let d] = span{θv : θ ∈ b, v ∈ F2 d} [bF2 and Then, Fb is co-invariant for Ld + L′ Ld. Define Fb = F2 d ⊖ [bF2 d]. d, where L′ d ⊂ B(F2 d) denotes the commutant of Lb = (PFb L1Fb , . . . , PFb L1Fb ). It is readily verified that Lb is a pure row contraction, and it is a consequence of the following result that Annnc(Lb) = b. Thus, the row contraction Lb provides a basic example of a constrained row contraction. Lemma 2.8. Let K ⊂ F2 d be a closed subspace which is co-invariant for Ld + L′ d. Let b = Annnc(PKL1K, . . . , PKLdK). Then, b is the unique weak-∗ closed two-sided ideal of Ld with the property that K = Fb. Proof. It follows from [20, Theorem 2.1] that there is a weak-∗ closed two-sided ideal c ⊂ Ld such that F2 d], whence K = Fc. It only remains to show uniqueness. Assume that a ⊂ Ld is a weak-∗ closed two-sided ideal such that d ⊖ K = [aF2 F2 d]. Since b was chosen to be the annihilator of (PKL1K, . . . , PKL1K), we see that PKθK = 0 for every θ ∈ b. Using that K is co-invariant for Ld we obtain d ⊂ F2 θF2 d ⊖ K for every θ ∈ b. We conclude that d ⊖ K = [cF2 [bF2 d] ⊂ F2 d ⊖ K = [aF2 d]. An application of [20, Theorem 2.1] yields b ⊂ a. On the other hand, it is immediate that a annihilates (PKL1K, . . . , PKLdK), whence a ⊂ b. (cid:3) We mention here that the condition that K be co-invariant for L′ d in the previous Indeed, by [27, Theorem 3.7] there is a proper lemma cannot simply removed. closed subspace K ⊂ F2 d which is co-invariant for Ld and which satisfies {θ ∈ Ld : θΩ ⊂ K⊥} = {0} so in particular we see that {θ ∈ Ld : θF2 d ⊂ K⊥} = {0}. CYCLIC ROW CONTRACTIONS AND RIGIDITY OF INVARIANT SUBSPACES 11 Thus Annnc(PKL1K, . . . , PKL1K) = {0}, yet K 6= F2 d. ideal and let We now examine the commuting case. Let J ⊂ Md be a weak-∗ closed two-sided HJ = H 2 d ⊖ [J H 2 d ] In this setting, the analogue of Lemma 2.8 holds without additional conditions on the co-invariant subspace. Lemma 2.9. Let K ⊂ H 2 d be a closed subspace which is co-invariant for Md. Let J = Annc(PKMx1K, . . . , PKMxdK). Then, J is the unique weak-∗ closed ideal of Md with the property that K = HJ . Proof. Repeat the argument used in the proof of Lemma 2.8, invoking [23, Theorem 2.4] rather than [20, Theorem 2.1]. (cid:3) Define a pure commuting row contraction as We record an important property of MJ . MJ = (PHJ Mx1HJ , . . . , PHJ MxdHJ ). Lemma 2.10. Let J ⊂ Md be a weak-∗ closed ideal. Let b = (Ψ ◦ qw)−1(J ) ⊂ Ld. Then, MJ is unitarily equivalent to Lb and Annc(MJ ) = J . Proof. By virtue of Lemma 2.9 we see that Annc(MJ ) = J . Next, it follows from Theorem 2.2 that UJ U ∗ = {PFs d XFs d : (Ψ ◦ qw)(X) ∈ J } = {PFs d XFs d : X ∈ b} whence [bF2 d]. Clearly, we see that Wd ⊂ b. Invoking [20, Proposition 2.4], we find Xv : X ∈ b, v ∈ Fs d ] = span{PFs d} = PFs U [J H 2 d d (Fs d)⊥ = [WdF2 d] ⊂ [bF2 d] and thus In particular, we find [bF2 d] = U [J H 2 d ] ⊕ (Fs d)⊥. Fb = [bF2 d]⊥ = Fs d ⊖ U [J H 2 d ] = U (H 2 d ⊖ [J H 2 d ]) = U HJ . We conclude that MJ is unitarily equivalent to Lb via the operator U . (cid:3) 2.4. Quasi-affine transforms. Let T = (T1, . . . , Td) and T ′ = (T ′ d) be two d-tuples of operators acting on Hilbert spaces H and H′ respectively. We say T ′ is a quasi-affine transform of T and write T ′ ≺ T , if there exists an injective bounded linear operator X : H′ → H with dense range such that XT ′ k = TkX for 1 ≤ k ≤ d. The relationship T ′ ≺ T is a common weakening of similarity that we will play a prominent role in our work. 1, . . . , T ′ 12 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO Lemma 2.11. Let H and H′ be Hilbert spaces and let X : H′ → H be a bounded linear operator. Let T ∈ B(H) and T ′ ∈ B(H′) such that XT ′ = T X. Then, the subspace H′ ⊖ ker X is co-invariant for T ′ and In particular, if X has dense range, then XPH′⊖ker X T ′H′⊖ker X = T XH′⊖ker X . PH′⊖ker X T ′H′⊖ker X ≺ T. Proof. Note that whence T ′ ker X ⊂ ker X. We conclude that H′ ⊖ ker X is co-invariant for T ′. Furthermore, we have XT ′ ker X = T X ker X = {0} XPH′⊖ker X T ′H′⊖ker X = XT ′H′⊖ker X = T XH′⊖ker X . Finally, since XH′⊖ker X : H′ ⊖ ker X → H is injective and has the same range as X, we conclude that whenever X has dense range. PH′⊖ker X T ′H′⊖ker X ≺ T (cid:3) Annihilators are well behaved under quasi-affine transforms, and we record the following easy fact. Lemma 2.12. Let T = (T1, . . . , Td) and T ′ = (T ′ d) be pure row contractions on some Hilbert spaces H and H′. Let X : H′ → H be a bounded linear operator such that XT ′ k = TkX for every 1 ≤ k ≤ d. Then, following statements hold. 1, . . . , T ′ (1) If X is injective, then Annnc(T ) ⊂ Annnc(T ′). (2) If X has dense range, then Annnc(T ′) ⊂ Annnc(T ). (3) If X is injective and has dense range, then Annnc(T ) = Annnc(T ′). Proof. This easily follows from the fact that for every θ ∈ Ld we have Xθ(T ′) = θ(T )X. (cid:3) 3. Transforming the standard d-shift In order to understand general pure row contractions, it is natural to turn to The- orem 2.4. This result reduces the problem to the study of compressions of d-shifts to co-invariant subspaces. Such a strategy is appealing, as the non-commutative function theory of Ld that was significantly developed in [20],[21],[22] could then be leveraged, much in the way that the function theory of H ∞(D) has helped single operator theory blossom [37]. Unfortunately, in general the d-shift from Theorem 2.4 is of high multiplicity, and thus the link with Ld is obscured. In this section, to circumvent this difficulty we investigate the possibility for d-shifts of multiplicity one to be quasi-affine transforms of pure row contractions, taking inspiration from and generalizing a known result for single operators [8, Theorem I.3.7]. We first exhibit a necessary and sufficient condition in the form of the existence of certain cyclic vectors, the images of which satisfy some weak form of orthogonality. Recall that given vectors ξ, η ∈ H, we denote by ξ⊗ η ∈ B(H) the rank-one operator defined as Theorem 3.1. Let T = (T1, . . . , Td) be a row contraction on some Hilbert space H and let C > 0. The following statements are equivalent. (ξ ⊗ η)h = hh, ηiξ, h ∈ H. CYCLIC ROW CONTRACTIONS AND RIGIDITY OF INVARIANT SUBSPACES 13 (i) There is a cyclic vector ξ ∈ H for T that satisfies Xw∈F+ d Twξ ⊗ Twξ ≤ C2I. (ii) There exists an operator X : F2 and XLk = TkX for 1 ≤ k ≤ d. injective operator X : V → H with dense range such that kXk ≤ C and d → H with dense range such that kXk ≤ C d which is co-invariant for L and an (iii) There exists a closed subspace V ⊂ F2 XPVLkV = TkX for 1 ≤ k ≤ d. Proof. Assume first that there exists a closed subspace V ⊂ F2 d which is co-invariant for L and an injective operator X : V → H with dense range such that kXk ≤ C and XPVLkV = TkX for k = 1, . . . , d. Let ξ = XPVΩ. Since Ω is obviously cyclic for L and V is co- invariant, it is readily seen that PVΩ is cyclic for (PVL1V, . . . , PVLdV). For a word w ∈ F+ d , we compute Twξ = TwXPVΩ = XPVLwPVΩ and thus conclude that ξ is cyclic for T , since X has dense range. In addition, for w ∈ F+ d and h ∈ H we see that hX ∗h, LwΩi = hX ∗h, PVLwΩi = hX ∗h, PVLwPVΩi = hh, XPVLwPVΩi = hh, Twξi, and since {LwΩ}w∈F+ d is an orthonormal basis for F2 d, we infer that hX ∗h, LwΩi2 = Xw∈F+ d h(Twξ ⊗ Twξ)h, hi hh, Twξi2 Twξ ⊗ Twξ ≤ C2I. d kX ∗hk2 = Xw∈F+ = Xw∈F+ Xw∈F+ d d whence (2) This shows that (iii) implies (i). Assume next that there is a cyclic vector ξ ∈ H for T satisfying (2). Then, given h ∈ H, we see that h(Twξ ⊗ Twξ)h, hi ≤ C2khk2. We conclude that the vector Xw∈F+ d hh, Twξi2 = Xw∈F+ Xw∈F+ d d hh, TwξiLwΩ 14 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO belongs to F2 operator Y : H → F2 d by d and has norm at most Ckhk. Hence, we may define a bounded linear hh, TwξiLwΩ, h ∈ H. d Y h = Xw∈F+ k h, LuΩi =Xw∈F+ d Plainly kY k ≤ C. Moreover, given 1 ≤ k ≤ d, a word u ∈ F+ we find k h, TwξiLwΩ, LuΩ+ = hT ∗ hT ∗ hY T ∗ k h, Tuξi d , and a vector h ∈ H and = hh, TkTuξi kY h, LuΩi = hY h, LkLuΩi =Xw∈F+ d hL∗ hh, TwξiLwΩ, LkLuΩ+ = hh, TkTuξi. We conclude that Y T ∗ Y h = 0 then k = L∗ kY for every 1 ≤ k ≤ d. Furthermore, we note that if 0 = hY h, LwΩi = hh, Twξi for every w ∈ F+ Thus, Y is injective. Choosing X = Y ∗ shows that (i) implies (ii). Finally, the fact that (ii) implies (iii) follows from Lemma 2.11. d , which implies that h = 0 since ξ is assumed to be cyclic for T . (cid:3) It is now a trivial matter to extract the desired necessary and sufficient condition. Corollary 3.2. Let T = (T1, . . . , Td) be a row contraction on some Hilbert space H. Then, the following statements are equivalent. (i) There is a cyclic vector ξ ∈ H for T such that the operatorPw∈F+ Twξ⊗Twξ d which is co-invariant for L and such (ii) There exists a closed subspace V ⊂ F2 is bounded. d that PVLV ≺ T. Proof. This follows immediately from Theorem 3.1. (cid:3) The property above that Pw∈F+ Twξ ⊗ Twξ be bounded can be seen as a re- laxation of the condition that ξ be a wandering vector, in the sense that the set is orthogonal. It will turn out below that this weak orthogonality con- {Twξ}w∈Fd dition is automatically satisfied for pure row contractions that admit any cyclic vector (see Corollary 3.6). + d Given a d-tuple T = (T1, . . . , Td) on some Hilbert space H, we defined its mul- tiplicity to be the least cardinality µT of a cyclic set for T . We emphasize that in the special case of d-shifts, this definition of multiplicity agrees with the one we gave in Section 2. The following simple observation will be use implicitly. Let T ′ be a d-tuple of operators on some other Hilbert space H′ for which there exists an operator X : H′ → H with dense range such that XT ′ k = TkX for k = 1, . . . , d. If Γ ⊂ H′ is cyclic for T ′, then a straightforward verification reveals that XΓ ⊂ H is cyclic for T . In particular, this shows that µT ≤ µT ′ . In preparation for our main result, we need a technical tool. CYCLIC ROW CONTRACTIONS AND RIGIDITY OF INVARIANT SUBSPACES 15 Lemma 3.3. Let T = (T1, . . . , Td) be a pure row contraction on some Hilbert space H. Then, there is a d-shift V = (V1, . . . , Vd) acting on some Hilbert space K with multiplicity µT and a contractive operator X : K → H with dense range such that TkX = XVk for 1 ≤ k ≤ d. Proof. By Theorem 2.4, there is a Hilbert space K0 containing H along with a d-shift S = (S1, . . . , Sd) on K0 with µS = dT and such that 1 ≤ k ≤ d. T ∗ k = S∗ kH, By assumption, there is a subset Γ ⊂ H of cardinality µT such that H = span{Twγ : w ∈ F+ d , γ ∈ Γ}. Let d , γ ∈ Γ} K = span{Swγ : w ∈ F+ w = Xw=n PKSwPKS∗ and put Vk = SkK for k = 1, . . . , d. For each n ∈ N, we note that wPK ≤ PK Xw=n which shows at once that V is pure, and that Pd k ≤ I. By the non- commutative Wold -- von Neumann decomposition [32, Theorem 1.3], we infer that V is a d-shift. w PK Xw=n k=1 VkV ∗ VwV ∗ SwS∗ Next, define X = PHK : K → H. For k = 1, . . . , d, using that H and K are respectively co-invariant and invariant for S we compute TkX = PHSkPHK = PHSkK = PHPKSkK = XSkK = XVk. Hence, for w ∈ F+ d and γ ∈ Γ we find XSwγ = XVwγ = TwXγ = Twγ which implies X K = H, so that X has dense range. In particular, we must have that µT ≤ µV . On the other hand, by definition of K we see that Γ is a cyclic set for V , so that µV ≤ card(Γ) = µT . (cid:3) We now arrive at the last technical step before the main result of the section. Roughly speaking, it says that up to a quasi-affine transform, the multiplicity of a pure row contraction can be made to agree with its defect. Lemma 3.4. Let T = (T1, . . . , Td) be a pure row contraction on some Hilbert space H. Then, there is a pure row contraction T ′ = (T ′ d) on some other Hilbert space H′ such that µT ′ = dT ′ = µT , along with a contractive injective operator X : H′ → H with dense range such that TkX = XT ′ If T is commuting, then so is T ′. k for 1 ≤ k ≤ d. 1, . . . , T ′ Proof. By Lemma 3.3, there a d-shift V = (V1, . . . , Vd) with µV = µT acting on some Hilbert space K and a contractive operator Y : K → H with dense range such that TkY = Y Vk for every 1 ≤ k ≤ d. Put and for each 1 ≤ k ≤ d let H′ = K ⊖ ker Y and X = Y H′ T ′ k = PH′ VkH′ . 16 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO Clearly, X is contractive, injective, and it has dense range. Moreover, it follows from Lemma 2.11 that TkX = XT ′ (3) Since H′ is co-invariant subspace for V , we see that T ′∗ It follows that T ′ is a pure row contraction. Using that k for 1 ≤ k ≤ d. In particular µT ≤ µT ′. w = V ∗ wH′ for every w ∈ F+ d . I − dXk=1 ′∗ k = PH′ I − T ′ kT k!H′ VkV ∗ dXk=1 we find dT ′ ≤ dV . But V is a d-shift, so dV = µV by Theorem 2.3 and we infer (4) On the other hand, since T ′ is pure we may apply Theorem 2.4 to find a Hilbert space K′ containing H′ and a d-shift S = (S1, . . . , Sd) with dT ′ = µS and such that ′∗ kH′ for every 1 ≤ k ≤ d. Choose a subset Γ ⊂ K′ with cardinality dT ′ that k = S∗ T is cyclic for S. For w ∈ F+ d and γ ∈ Γ, we may use that H′ is co-invariant for S to find dT ′ ≤ µT . T ′ wPH′γ = PH′SwPH′γ = PH′Swγ. Consequently, we have that span{T ′ wPH′ γ : w ∈ F+ d , γ ∈ Γ} = PH′(cid:16)span{Swγ : w ∈ F+ = PH′ K′ = H′ d , γ ∈ Γ}(cid:17) so that PH′ Γ is cyclic for T ′. Hence, (5) µT ′ ≤ dT ′ . Combining inequalities (3), (4) and (5) yields µT ′ = dT ′ = µT as desired. Finally, if T is commuting, then Annnc(T ) contains the commutator ideal, so (cid:3) that T ′ is also commuting by Lemma 2.12. We remark here that the preceding two proofs are faithful adaptations of the single-variable ones found in [8, Lemma I.3.5 and Theorem I.3.7]. Interestingly however, this approach forces us through the non-commuting multivariate world, even if we start with a commuting row contraction. To see why, we note that there are examples of cyclic invariant subspaces for Mx for which the restriction has infinite defect. Take for instance the cyclic invariant subspace M ⊂ H 2 d generated by x1, so that M = Mα∈Nd,α1≥1 Cxα. Then M clearly has infinite co-dimension, so we may invoke [6, Theorem F] to conclude that the restriction MxM has infinite defect. This underlines potential difficulties in establishing the equality dV = µV , which is used crucially in the proof of Lemma 3.4. Fortunately, with our approach this equality follows from Theorem 2.3 since V is known to be a d-shift, being the restriction of a d-shift to an invariant subspace. For commuting d-shifts, the behaviour of such restrictions is more subtle. Assume for instance that the restriction of the standard commuting d- shift to some cyclic proper invariant subspace is another commuting d-shift. Since it must be cyclic, this second commuting d-shift must be of multiplicity one, and thus CYCLIC ROW CONTRACTIONS AND RIGIDITY OF INVARIANT SUBSPACES 17 unitarily equivalent to the standard commuting d-shift. This is however impossible, by [26, Corollary 5.5]. Although it is plausible that a completely different approach could circumvent these issues in the commuting setting, the preceding remarks show that our current proofs have an inexorable non-commutative aspect built into them. The following is one of our main results. Theorem 3.5. Let T = (T1, . . . , Td) be a pure cyclic row contraction on some Hilbert space H. Then there is a contractive operator X : F2 d → H with dense range such that XLk = TkX for k = 1, . . . , d. d → H with dense If T is commuting, then there is a contractive operator X ′ : H 2 range such that X ′Mxk = TkX ′ for k = 1, . . . , d. Proof. By Lemma 3.4, we see that there is a pure row contraction T ′ = (T ′ 1, . . . , T ′ d) on some other Hilbert space H′ such that dT ′ = 1, along with a contractive injective operator Y : H′ → H with dense range such that TkY = Y T ′ k for 1 ≤ k ≤ d. Up to unitary equivalence, we may assume by Theorem 2.4 that H′ ⊂ F2 kH′ for every k = 1, . . . , d. Define X : F2 d → H as X = Y PH′ . For 1 ≤ k ≤ d, we calculate using that H′ is co-invariant for L that ′∗ k = L∗ d and T XLk = Y PH′ Lk = Y PH′ LkPH′ kPH′ = TkY PH′ = Y T ′ = TkX. Moreover, we note that X F2 d = Y PH′ F2 d = Y H′ = H so that X has dense range, which establishes the first statement. For the second, we assume in addition that T is commuting. Then, 0 = (TjTk − TkTj)X = X(LjLk − LkLj) and thus ker X contains (LjLk − LkLj)F2 ker X contains (Fs d → Fs 2.2, there is a unitary operator U : H 2 U Mxk U ∗ = PFs LkFs d for every 1 ≤ j, k ≤ d. We infer that d)⊥ by [20, Proposition 2.4], whence X = XPFs . By Theorem d with the property that 1 ≤ k ≤ d. , d d d We set X ′ = XU , which still has dense range. Using that Fs we obtain d is co-invariant for L X ′Mxk = XU Mxk = XPFs d LkU = XLkU = TkX ′ for k = 1, . . . , d. (cid:3) Comparing the previous result with Theorem 3.1, we obtain the following con- sequences. Corollary 3.6. Let T = (T1, . . . , Td) be a pure cyclic row contraction on some Hilbert space H. Then, there is a vector ξ ∈ H that is cyclic for T with the additional Twξ ⊗ Twξ ≤ I. Moreover, the following statements hold. property thatPw∈F+ (1) If kTwξk = 1 for every w ∈ F+ (2) If d > 1 and T is commuting, then kTwξk < 1 for some w ∈ F+ d . d , then T must be unitarily equivalent to the standard d-shift. d 18 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO Proof. The existence of a vector ξ with the announced property follows by com- bining Theorems 3.1 and 3.5. For (1), assume that kTwξk = 1 for each w ∈ F+ d . The condition that Pw∈F+ d } to be an orthonormal set. But since ξ is cyclic, the set {Twξ : w ∈ F+ d } must in fact be an orthonormal basis for H. It is now straightforward to construct a unitary equiva- lence between T and the standard d-shift L. Finally, (2) is trivial consequence of (1) since the standard d-shift is non-commuting. (cid:3) Twξ ⊗ Twξ ≤ I then forces {Twξ : w ∈ F+ d We close this section by refining Theorem 3.5 to obtain another one of our main results. See Subsection 2.3 for the definition of the d-tuple MJ . Corollary 3.7. Let T = (T1, . . . , Td) be a pure cyclic commuting row contraction on some Hilbert space H. Let J = Annc(T ). Then MJ is a quasi-affine transform of T . Proof. By Theorem 3.5, there is a contractive operator X : H 2 d → H with dense range such that XMxk = TkX for k = 1, . . . , d. In turn, invoking Lemma 2.11, we see that there is a closed subspace K ⊂ H 2 d which is co-invariant for Mx and such that It follows from Lemma 2.12 that (PKMx1K, . . . , PKMxdK) ≺ T. whence Annnc(T ) = Annnc(PKMx1K, . . . , PKMxdK) Annc(T ) = Annc(PKMx1K, . . . , PKMxdK) by virtue of Equation (1). By Lemma 2.9, we conclude that K = HJ so that MJ ≺ T . (cid:3) Given a pure cyclic row contraction T = (T1, . . . , Td) with b = Annnc(T ), we do not claim that Lb ≺ T . The difficulty in achieving this is connected with the discussion following Lemma 2.8. This is also related to determining whether a given co-invariant subspace coincides with the so-called "maximal b-constrained piece" of L as defined in [35]. We will not address these issues further in this paper. To summarize, the basic outcome of this section is a classification of some pure row contractions by means of compressions of the standard d-shift. By allowing for quasi-affine transforms we capture the behaviour of a class of objects that is more flexible than those with defect equal to one (which is required in Theorem 2.4). To illustrate this flexibility, we note for instance that if T is a cyclic pure row contraction with defect one, then the scaled row contraction rT has defect equal to dim H for 0 < r < 1, even though it is still pure and cyclic. Furthermore, the cyclicity condition is preserved under similarity. 4. Rigidity of invariant subspaces In this section, we examine the rigidity of invariant or co-invariant subspaces for pure row contractions. More precisely, assume that T = (T1, . . . , Td) is a pure row contraction on some Hilbert space H, and let M ⊂ H be a closed subspace which is invariant for T . Throughout, we let TM denote the restricted d-tuple (T1M, . . . , TdM) CYCLIC ROW CONTRACTIONS AND RIGIDITY OF INVARIANT SUBSPACES 19 and PM⊥ TM⊥ denote the compressed d-tuple (PM⊥ T1M⊥, . . . , PM⊥ TdM⊥ ). We aim to determine the extent to which the annihilators Annnc(TM) or Annnc(PM⊥ TM⊥) determine M. We begin by revisiting a pathological example that we encountered before. Example 1. By [27, Theorem 3.7], there is a proper closed subspace K ⊂ F2 is co-invariant for Ld and which satisfies {θ ∈ Ld : θF2 d ⊂ K⊥} = {0}. d which In particular, we see that yet K 6= F2 d. Annnc(PKLK) = {0} = Annnc(L) (cid:3) In view of this difficulty, we will focus our attention for the remainder of the paper on the commuting setting. We start by considering a relatively simple case. A commuting d-tuple (N1, . . . , Nd) is said to be nilpotent if for each 1 ≤ k ≤ d there is mk ∈ N such that N mk Theorem 4.1. Let N = (N1, . . . , Nd) be a cyclic nilpotent commuting row con- traction on some Hilbert space H. Let M ⊂ H be an invariant subspace for N such that Annc(NM) = Annc(N ). Then, M = H. Proof. Let ξ ∈ H be a cyclic vector for N . We find k = 0. Since N is nilpotent, this implies that H is finite-dimensional and H = span{N αξ : α ∈ Nd}. H = {p(N )ξ : p ∈ C[x1, . . . , xd]}. The subset J ⊂ C[z1, . . . , zd] defined as is an ideal with the property that J = {p ∈ C[z1, . . . , zd] : p(N )ξ ∈ M} For convenience, we set and M = {p(N )ξ : p ∈ J }. A = Annc(NM) ∩ C[x1, . . . , xd] Observe that a polynomial p satisfies p(N ) = 0 if and only if p(N )ξ = 0. Thus, B = Annc(N ) ∩ C[x1, . . . , xd]. A = {f ∈ C[x1, . . . , xd] : fJ ⊂ B}. Clearly, the set {α ∈ Nd : xα /∈ B} is finite and contains (0, . . . , 0). Choose an element β of that set with maximal length. Then, we have xkxβ ∈ B for every 1 ≤ k ≤ d. Assume now that M is a proper subspace of H, whence ξ /∈ M. Let q be a polynomial with non-zero constant term. Then, there is another polynomial r such that r(N )q(N ) = I, so in particular ξ = r(N )q(N )ξ. We conclude that 20 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO q /∈ J , which shows J consists of polynomials with zero constant term. Hence, xβJ ⊂ B, and therefore xβ ∈ A. This shows that A 6= B and in particular Annc(NM) 6= Annc(N ). (cid:3) In the case of a single cyclic constrained contraction, the conclusion of the pre- vious theorem always holds [8, Theorem III.2.13]. Unfortunately, the general mul- tivariate situation is more complicated. In fact, the aforementioned theorem does not even extend to the commuting bivariate case, as the following example shows. Example 2. Let J be the weak-∗ closed ideal of M2 generated by x2. Then, the space HJ = H 2 2 that depends only on the variable x1 and in particular we have 2 ] consists of those functions in H 2 2 ⊖ [J H 2 The pure commuting pair PHJ Mx2HJ = 0. MJ = (PHJ Mx1HJ , 0) is constrained, and by Lemma 2.10 we have Annc(MJ ) = J . Note also that 1 ∈ HJ is a cyclic vector for MJ . Let U : H 2 1 → HJ be the linear operator defined as (U f )(z1, z2) = f (z1), (z1, z2) ∈ B2 for f ∈ H 2 1 . It is readily verified that U is unitary. Observe now that U ∗(PHJ Mx1HJ )U is the usual isometric 1-shift. Let θ ∈ M1 be a non-constant inner function and let M = U (θH 2 1 ) ⊂ HJ . Then, M is a proper closed subspace of HJ which is invariant for MJ . If we let R1 = PHJ Mx1M, then we see that (R1, 0) is the restriction of MJ to M. Further, if we let U ′ = U MθU ∗ then we see that U ′ : HJ → M is unitary with Thus R is unitarily equivalent to MJ , and in particular it is a pure cyclic commuting row contraction with U ′(PHJ Mx1HJ )U ′∗ = R1. Annnc(R) = Annnc(MJ ) and Annc(R) = Annc(MJ ), even though M 6= HJ . (cid:3) We note that the zero set of the commutative annihilator has dimension one in the preceding example. At present, we know of no multivariate counterexample to [8, Theorem III.2.13] where this zero set has dimension zero. Theorem 4.1 may be evidence towards the relevance of having a zero-dimensional zero set, but at the time of this writing we do not know for sure. In light of Example 2, we then seek natural sufficient conditions for multivariate rigidity of invariant subspaces. The next development will gain insight by first considering co-invariant subspaces; it is based on the following simple fact. Lemma 4.2. Let M, N ⊂ H 2 Then, we have M = N if and only if Annc(PMMxM) = Annc(PNMxN). Proof. This follows immediately from Lemma 2.9. d be closed subspaces that are co-invariant for Md. (cid:3) CYCLIC ROW CONTRACTIONS AND RIGIDITY OF INVARIANT SUBSPACES 21 Before proceeding, we introduce some notation. Let S = (S1, . . . , Sd) be a d-tuple of operators on some Hilbert space H and let S′ = (S′ d) be another d-tuple of operators on some other Hilbert space H′. Denote by Q(S′, S) the collection of injective bounded linear operators X : H′ → H with dense range such that XS′ k = SkX for every k = 1, . . . , d. Recall that if T is a cyclic pure commuting row contraction and J = Annc(T ), then Q(MJ , T ) is non-empty by Corollary 3.7. ing row contractions. We now prove a rigidity result for co-invariant subspaces of cyclic pure commut- 1, . . . , S′ Theorem 4.3. Let T = (T1, . . . , Td) be a cyclic pure commuting row contraction on some Hilbert space H. Put J = Annc(T ). Let M, N ⊂ H be closed co-invariant subspaces for T such that Annc(PMTM) = Annc(PNTN). Then, X ∗M = X ∗N for every X ∈ Q(MJ , T ). In particular, M = N if X ∗M and X ∗N are closed for some X ∈ Q(MJ , T ). Proof. Fix X ∈ Q(MJ , T ). Define M′ = X ∗M and N′ = X ∗N, so that M′ and N′ are co-invariant for MJ , and hence for Mx. Let Z = PMXM′ : M′ → M. We have Z ∗ = X ∗M, so Z ∗ is injective with dense range and xM′ ). Z ∗ ∈ Q(T ∗M, M ∗ Thus, Likewise, the operator Y = PNXN′ is injective with dense range, and Z ∈ Q(PM′ MxM′ , PMTM). Y ∈ Q(PN′ MxN′ , PNTN). We conclude from Lemma 2.12 and from the assumption on M and N that Annc(PM′ MxM′ ) = Annc(PN′ MxN′). By virtue of Lemma 4.2, we obtain M′ = N′, that is X ∗M = X ∗N. If X ∗M and X ∗N are closed, then X ∗M = X ∗N and we conclude that M = N since X ∗ is injective. (cid:3) The following consequence is noteworthy as it applies in particular to the nilpo- tent setting, thus complementing Theorem 4.1. Corollary 4.4. Let T = (T1, . . . , Td) be a cyclic pure commuting row contraction on some finite-dimensional Hilbert space H. Let M, N ⊂ H be co-invariant subspaces for T such that Then, M = N. Annc(PMTM) = Annc(PNTN). Proof. Finite-dimensional subspaces are closed, so this follows at once from Theo- rem 4.3. (cid:3) We now return to invariant subspaces and identify a sufficient condition for rigidity. Theorem 4.5. Let T = (T1, . . . , Td) be a pure commuting row contraction on some Hilbert space H. Let J ⊂ Md be a weak-∗ closed ideal and let X ∈ Q(M ∗ J , T ). Let M, N ⊂ H be closed invariant subspaces for T such that Annc(TM) = Annc(TN). If M and N are contained in the range of X, then M = N. 22 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO Proof. If f ∈ X −1M, then for every 1 ≤ k ≤ d we obtain X(M ∗ xkHJ )f = TkXf ∈ TkM ⊂ M. Thus X −1M is invariant for M ∗ that X −1N is also invariant for M ∗ x . Next, let J and hence for M ∗ x . An identical argument shows Z = XX −1M : X −1M → M Y = XX −1N : X −1N → N. Because X is injective, it follows that Z and Y are injective as well. Moreover, Z and Y are surjective since M, N are contained in the range of X. For k = 1, . . . , d, we note that xkX −1M) and (TkN)Y = Y (M ∗ xkX −1N). d be the conjugate linear, anti-unitary operator such that Jxα = xα (TkM)Z = Z(M ∗ d → H 2 Let J : H 2 for every α ∈ Nd. Then, we note that (TkM)Z = Z((JMxk J)∗X −1M) and (TkN)Y = Y ((JMxk J)∗X −1N) for every k = 1, . . . , d. A standard approximation procedure yields that JMdJ = Md and ϕ(TM)Z = Z((JMϕJ)∗X −1M) and ϕ(TN)Y = Y ((JMϕJ)∗X −1N) for every ϕ ∈ Md. Observe then that ϕ ∈ Annc(TM) if and only if PX −1M(JMϕJ)X −1M = 0 which in turn is equivalent to JMϕJ ∈ Annc(PX −1MMxX −1M). Thus, we find Likewise, we infer that Annc(PX −1MMxX −1M) = J Annc(TM)J. By assumption, we may therefore write Annc(PX −1NMxX −1N) = J Annc(TN)J. Annc(PX −1MMxX −1M) = Annc(PX −1NMxX −1N). In view of Lemma 4.2, we then know that X −1M = X −1N. Applying X on both sides yields M = N. (cid:3) In general, it is not clear how to determine that Q(M ∗ J , T ) is non-empty in order to apply the previous result. In some cases however, the existence of a cyclic vector can be exploited. Corollary 4.6. Let T = (T1, . . . , Td) be a nilpotent commuting row contraction on some Hilbert space H. Assume that T ∗ is cyclic. Let M, N ⊂ H be invariant subspaces for T such that Annc(TM) = Annc(TN). Then, M = N. Proof. The commuting d-tuple T ∗ is also nilpotent. Applying [36, Theorem 3.8] with k = m = 1, Q = {0} and f (Z) = Z1 + . . . + Zd (in the notation of that paper), we see that there is an invertible operator Y ∈ B(H) such that the d-tuple Y T ∗Y −1 = (Y T ∗ 1 Y −1, . . . , Y T ∗ d Y −1) CYCLIC ROW CONTRACTIONS AND RIGIDITY OF INVARIANT SUBSPACES 23 is a nilpotent cyclic commuting row contraction. In particular, Y T ∗Y −1 is pure, so if we put J = Annc(Y T ∗Y −1), then by virtue of Corollary 3.7, we see that there is X ∈ Q(MJ , Y T ∗Y −1). Next, observe that H is finite-dimensional since the nilpotent d-tuple T ∗ is cyclic, whence X is invertible and (X −1Y )∗ ∈ Q(M ∗ J , T ). Theorem 4.5 then yields M = N. (cid:3) It is not clear that Theorem 4.1 could be derived from the previous result, as it is not typically true that the adjoint of a cyclic nilpotent row contraction is still cyclic; see Example 5 below. 5. Invariant decompositions and separating vectors We saw in Corollary 3.7 that pure commuting row contractions can be effectively classified using compressions of the standard commuting d-shift, provided that they admit a cyclic vector. Inspired by the successful univariate theory developed in [8], a natural subsequent step in this classification program would involve moving past the setting of cyclicity using the following device. Let T = (T1, . . . , Td) be a d-tuple of operators on some Hilbert space H and let M, N ⊂ H be non-trivial invariant subspaces for T . We say that the pair (M, N) is an invariant decomposition for T if M ∩ N = {0} and M + N = H. In the special case where TM is cyclic, we say that (M, N) is a cyclic decomposition for T . In the single-variable case, non-cyclic constrained contractions always admit cyclic decompositions [8, Theorem III.3.1]. We start this section with an exam- ple illustrating that cyclic decompositions are more elusive in several variables. For this purpose, we introduce one more notion. An invariant decomposition (M, N) will be said to be topological if the stronger condition M + N = H is satisfied. The following is an adaptation of [25, Theorem 2.3] Example 3. Let {en : n ∈ N} denote the canonical orthonormal basis of ℓ2(N). Let H = ℓ2(N) ⊕ ℓ2(N). For each n ∈ N, define vectors ξn, ηn ∈ H as ξn = en ⊕ 0, ηn = 0 ⊕ en. Then, {ξn, ηn : n ∈ N} is an orthonormal basis for H. We may define bounded linear operators T1, T2 on H as T1 = ηn+1 ⊗ ξn, T2 = ηn ⊗ ξn where the series converge in the strong operator topology of B(H). We find 1 √2 1 √2 ∞Xn=1 ∞Xn=1 ∞Xn=1 ∞Xn=1 2T1T ∗ 1 = ηn+1 ⊗ ηn+1 2T2T ∗ 2 = ηn ⊗ ηn. and 24 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO 1 and 2T2T ∗ Thus, 2T1T ∗ In particular, we see that T1 and T2 are scalar multiples of partial isometries, and thus have closed range. Moreover, 2 are orthogonal projections. T1T ∗ 1 + T2T ∗ 2 ≤ I so that T = (T1, T2) is a row contraction on H. Since ξn is orthogonal to ηm for every n, m ∈ N, we see that T 2 1 = T1T2 = T2T1 = T 2 2 = 0. Thus, T is a nilpotent commuting row contraction. Using Theorem 2.1 it is readily seen that Annc(T ) is the weak-∗ closed ideal of M2 generated by {x2 1, x1x2, x2 2}. We infer that the range of T β is closed for all β ∈ N2. Assume that M, N ⊂ H are closed invariant subspaces for T with M ∩ N = {0} and H = M + N. Observe that and Furthermore, we have that T1H = ⊕∞ n=2 Cηn T2H = ⊕∞ n=1 Cηn. T1H = T1M + T1N, T2H = T2M + T2N and T1M ⊂ M, T2M ⊂ M, T1N ⊂ N, T2N ⊂ N. Using that M ∩ N = {0} and that T1H ⊂ T2H, we infer that T1M ⊂ T2M and T1N ⊂ T2N. Since T1H is a co-dimension one closed subspace of T2H, we infer that either T1M = T2M or T1N = T2N. By symmetry, we may suppose the former. Given a subspace K ⊂ H, we set and νξ(K) = inf{n ∈ N : ξn /∈ K⊥} νη(K) = inf{n ∈ N : ηn /∈ K⊥}. Note that νξ(K) is infinite whenever ξn ∈ K⊥ for every n ∈ N. Using that T ∗ ξn and T ∗ 2 ηn = ξn for every n ∈ N, we see that 1 ηn+1 = νη(T2K) = νξ(K) and νη(T1K) = νξ(K) + 1. Recall that T1M = T2M, thus νξ(M) = νη(T2M) = νη(T1M) = νξ(M) + 1 which forces νξ(M) = +∞. Therefore, ξn ∈ M⊥ for every n ∈ N. Consequently, we find M ⊂ ⊕∞ n=1 Cηn = T2H. We infer that M ⊂ T2M which further implies that T2M = M. Finally, using that T 2 2 = 0 we obtain M = {0}. This shows that T does not admit a topological invariant decomposition. If (M, N) is a cyclic decomposition for T , then using that T is nilpotent and that TM is cyclic we conclude that M must be finite-dimensional. It is well-known then that H = M + N, so that (M, N) would be a topological invariant decomposition for T . We thus conclude that T admits no cyclic decomposition either. (cid:3) CYCLIC ROW CONTRACTIONS AND RIGIDITY OF INVARIANT SUBSPACES 25 We mention here a classical result of Apostol and Stampfli [4, Theorem 5] which implies that if T is a nilpotent operator on a Hilbert space H with the property that T nH is closed for every n ∈ N, then T admits a topological cyclic decomposition. The previous example shows that this theorem fails in several variables. In spite of Example 3, by leveraging the work done previously on rigidity of invariant subspaces we identify a sufficient condition that allows for an invariant decomposition. Theorem 5.1. Let T = (T1, . . . , Td) be a nilpotent commuting row contraction on some Hilbert space H. Let M ⊂ H be an invariant subspace for T such that (TM)∗ has a cyclic vector and Annc(TM) = Annc(T ). Then, there is a subspace N ⊂ H that is invariant for T such that M ∩ N = {0} and H = M + N. Proof. By assumption, there is a vector ξ ∈ M which is cyclic for (TM)∗. Let Put S = PKTK and observe that Annc(T ) ⊂ Annc(S). Define X : M → K as X = PKM. Using that K is coinvariant for T and M is invariant for T , we obtain K = span{T ∗αξ : α ∈ Nd}. SkX = PKTkPKM = PKTkM = PKPMTkM = XTkM for every 1 ≤ k ≤ d, so that for α ∈ Nd we have X ∗T ∗αK = X ∗S∗α = (TM)∗αX ∗. X ∗ξ = PMξ = ξ Furthemore, whence X ∗T ∗αξ = (TM)∗αX ∗ξ = (TM)∗αξ for every α ∈ Nd. By choice of ξ being cyclic for (TM)∗, we conclude that X ∗ has dense range and thus X is injective. Now, XM ⊂ K is easily seen to be invariant for S and so it follows from Lemma 2.12 that (SkXM)X = XTkM, 1 ≤ k ≤ d By assumption, we then infer that Annc(SXM) ⊂ Annc(TM). and thus We conclude that Annc(SXM) ⊂ Annc(T ) Annc(SXM) ⊂ Annc(S). Notice now that S∗ is cyclic by the construction of K. We may thus invoke Corollary 4.6 to find XM = K, so that X ∗ is injective. Finally, define N = K⊥. We find Annc(SXM) = Annc(S). and M ∩ N = M ∩ K⊥ = ker X = {0} M + N = (M⊥ ∩ N⊥)⊥ = (M⊥ ∩ K)⊥ = (ker X ∗)⊥ = H. (cid:3) 26 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO One shortcoming of the previous theorem is that the existence of the decompo- sition hinges on the adjoint of TM being cyclic. This condition is a bit mysterious. Interestingly, in the case of a single constrained contraction, this condition is in fact equivalent to TM being cyclic by [8, Theorem III.2.3]. Unfortunately, this convenient equivalence does not hold in the multivariate world, as we will see in Example 5. Given these multivariate obstacles, and in light of the univariate mechanism [8, Theorem III.3.1] for producing cyclic decompositions, it seems worthwhile to investigate more precisely where this mechanism fails in several variables. The rest of the section will be devoted to this goal, through the lens of the following notion. Let T = (T1, . . . , Td) be a constrained commuting row contraction on some Hilbert space H. A subset Σ ⊂ H is said to be separating for T if, given any θ ∈ Md, the condition Σ ⊂ ker θ(T ) forces θ(T ) = 0. This means that Σ is a separating set for the operator algebra {θ(T ) : θ ∈ Md}. When Σ consists of only one element ξ ∈ H, we say that ξ is a separating vector for H. In the classical one-variable setting, separating vectors are called maximal [8], but we feel that in our context our choice of terminology is a bit more descriptive. We start with a concrete example where a separating vector can be constructed explicitly. Example 4. Let T = (T1, . . . , Td) be a pure commuting row contraction on some Hilbert space H. Assume that Annc(T ) is the weak-∗ closed ideal of Md generated by {xn1 d } for some fixed positive integers n1, . . . , nd. Set 1 , . . . , xnd Ω = {α ∈ Nd : xα /∈ Annc(T )} and put γ = (n1 − 1, . . . , nd − 1). Since γ ∈ Ω, we may choose a non-zero vector ξ ∈ H\ ker(T γ). We claim that ξ is separating for T . To see this, let ϕ ∈ Md with ϕ(T ) 6= 0. An application of Theorem 2.1 with n > (n1 − 1) + (n2 − 1) + . . . + (nd − 1) shows that there is a polynomial p such that ϕ−p ∈ Annc(T ), whence ϕ(T ) = p(T ). Furthermore, it is clear that p can be chosen to be of the form p =Xα∈Ω cαxα. Because p(T ) 6= 0, the set S = {α ∈ Ω : cα 6= 0} is non-empty. Choose µ ∈ S with minimal length. It is easily seen that there is β ∈ Nd such that µ + β = γ and α + β /∈ Ω for every α ∈ S \ {µ}. Thus T βp(T ) = cµT γ. Since ξ /∈ ker T γ and cµ 6= 0, it follows that T βp(T )ξ 6= 0 and in particular p(T )ξ 6= 0. Therefore, ϕ(T )ξ 6= 0, and we conclude that ξ is separating for T . (cid:3) Unfortunately, separating vectors do not always exist, even for commuting nilpo- tent pairs on finite-dimensional spaces. CYCLIC ROW CONTRACTIONS AND RIGIDITY OF INVARIANT SUBSPACES 27 Example 5. With respect to the usual orthonormal basis, we define linear opera- tors on C3 as T1 = T2 = 1 √3 0 1 0 0 0 0 −1 0 0  , 1 √3 0 0 0 0 0 0 1 0 0 . Easy computations show that T 2 1 = T1T2 = T2T1 = T 2 2 = 0 and that T1T ∗ 1 + T2T ∗ 2 = 0 0 0 0 0 1 0 0 0 . Hence, T = (T1, T2) is a nilpotent commuting row contraction. Let ϕ ∈ M2. An ap- plication of Theorem 2.1 yields constants a, b, c ∈ C and multipliers ϕ12, ϕ11, ϕ22 ∈ M2 such that ϕ = a + bx1 + cx2 + ϕ11x2 1 + ϕ12x1x2 + ϕ22x2 2. Then, we find ϕ(T1, T2) = aI + bT1 + cT2 = a 0 b/√3 a (c − b)/√3 0 a 0 0  . Thus, ϕ(T1, T2) = 0 if and only if a = b = c = 0. This shows that Annc(T ) is the weak-∗ closed ideal of M2 generated by {x2 C3, and consider the polynomial p defined as We claim that no vector v ∈ C3 is separating for T . Indeed, write v = (v1, v2, v3) ∈ 1, x1x2, x2 2}. p =(v3x1 + (v3 − v1)x2 x1 if v3 6= 0 or v1 6= 0, otherwise. Then p(T )v = 0 yet p(T ) 6= 0, which means that v is not separating. This es- tablishes the claim that T has no separating vector, and thus no cyclic vector either. Note however that the vector (0, 1, 0) is easily verified to be cyclic for the pair T ∗ = (T ∗ 2 ). Anticipating Theorem 5.3 below, we also remark that the set {(1, 0, 0), (0, 0, 1)} is separating for T . (cid:3) 1 , T ∗ Interestingly, separating vectors for single constrained contractions are known to exist in abundance [8, Theorem II.3.6]. Inspection of the proofs of the results leading up to [8, Theorem III.3.1] reveals that separating vectors play a crucial role there. Thus, the difficulties associated to the construction of cyclic decompositions in the multivariate context may be explained, in part, by the scarcity of separating vectors brought to light in Example 5. Another consequence of Example 5 is that nilpotent commuting d-tuples are not as nicely behaved as their single variable counterparts. For instance, by [3, Theorem 1] it is known that if T is a nilpotent contraction, then there is a collection {Tλ : λ ∈ Λ} of cyclic nilpotent contractions such that ⊕λ∈ΛTλ ≺ T. Trivially, a cyclic vector is necessarily separating, so the following lemma implies that such an operator T must admit a separating vector. In light of the previous 28 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO example, we conclude that a direct analogue of [3, Theorem 1] cannot hold in several variables. Lemma 5.2. The following statements hold. 1, . . . , T ′ (1) Let T = (T1, . . . , Td) and T ′ = (T ′ d) be constrained commuting row contractions on some Hilbert spaces H and H′. Assume that T ′ ≺ T . If T ′ has a separating vector, then so does T . (2) For each n ∈ N, let Tn be a constrained commuting row contraction on some Hilbert space Hn. Assume that Tn has a separating vector for every n ∈ N. Then, ⊕∞ n=1Tn has a separating vector as well. Proof. For (1), let X : H′ → H be an injective operator with dense range such that XT ′ k = TkX for every 1 ≤ k ≤ n. Let ξ ∈ H′ be a separating vector for T ′ and let θ ∈ Md. Assume that θ(T )Xξ = 0. Then 0 = θ(T )Xξ = Xθ(T ′)ξ and we infer that θ(T ′)ξ = 0 since X is injective. Using that ξ is separating for T ′, we thus find that θ(T ′) = 0. Next, Lemma 2.12 and Equation (1) imply that so that θ(T ) = 0. Hence Xξ is separating for T . Annc(T ) = Annc(T ′) Turning to (2), for each n ∈ N we let ξn ∈ Hn be a separating vector for Tn with norm 1. Define ξ ∈ ⊕∞ n=1Hn as It is readily verified that if θ ∈ Md, then ξ =(cid:18)ξ1, 1 2 ξ2, . . . , 1 2n ξn, . . .(cid:19) . θ(T ) = ⊕∞ n=1θ(Tn) and θ(T )ξ =(cid:18)θ(T1)ξ1, 1 2 θ(T2)ξ2, . . . , 1 2n θ(Tn)ξn, . . .(cid:19) . The desired statement readily follows from these observations. Our final result below offers a counterpoint to Example 5 and shows that, at least for nilpotent commuting row contractions, the existence of separating vectors can be insured provided that we allow for finite ampliations. Given a d-tuple of operators T = (T1, . . . , Td) on some Hilbert space H and a positive integer n ∈ N, we define (cid:3) which acts on the Hilbert space T (n) = T ⊕ T ⊕ . . . ⊕ T H(n) = H ⊕ H ⊕ . . . ⊕ H. Theorem 5.3. Let N = (N1, . . . , Nd) be a nilpotent commuting row contraction on some Hilbert space H. Set A = C[x1, . . . , xd]/(C[x1, . . . , xd] ∩ Annc(N )). Then, for any positive integer s ≥ dimA the ampliation N (s) has a dense set of separating vectors in H(s). CYCLIC ROW CONTRACTIONS AND RIGIDITY OF INVARIANT SUBSPACES 29 Proof. Since N is nilpotent, there are positive integers m1, . . . , md such that N mk 0 for every 1 ≤ k ≤ d. An application of Theorem 2.1 with k = n > (m1 − 1) + (m2 − 1) + . . . + (md − 1) shows that for any ϕ ∈ Md, there is a polynomial p such that ϕ(N ) = p(N ). For this reason, a vector (ξ1, . . . , ξs) ∈ H(s) is separating for N (s) if whenever p is a polynomial, the equalities are equivalent to p(N ) = 0. rather than general multipliers. In other words, it suffices to consider polynomials p(N )ξ1 = ··· = p(N )ξs = 0 Note that A is a finite-dimensional vector space, and put δ = dimA. For each 1 ≤ i ≤ δ, choose a non-empty open subset Ui ⊂ H. Given a polynomial p, we denote by [p] its image in the quotient A. For every h ∈ H, set Ah = {[p] ∈ A : p(T )h = 0}. We claim that there are vectors ξ1 ∈ U1, . . . , ξδ ∈ Uδ such that ∩δ i=1Aξi = {0}. Assume otherwise. For convenience, set Aξ0 = A. Fix a non-zero element [p1] ∈ Aξ0 . Then, p1(T ) 6= 0 so we may choose ξ1 ∈ U1 \ ker p1(T ). Thus, [p1] /∈ Aξ1 and in particular Aξ1 is a proper subspace of Aξ0 . Suppose that for 1 ≤ i ≤ δ − 1 we have constructed ξ1 ∈ U1, . . . , ξi ∈ Ui j=0Aξj is a proper subspace of ∩i−1 with the property that ∩i j=0Aξj . By our standing assumption, we see that ∩i j=0Aξj is non-zero, so there is a non-zero element [pi+1] ∈ ∩i j=0Aξj . Correspondingly, there is a vector ξi+1 ∈ Ui+1 \ ker pi+1(T ). We see that [pi+1] /∈ Aξi+1 , whence ∩i+1 j=0Aξj is a non-zero proper subspace of ∩i j=0Aξj . By induction, we obtain vectors ξ1 ∈ U1, . . . , ξδ ∈ Uδ with the property that for every j=0Aξj is a non-zero proper subspace of ∩i−1 1 ≤ i ≤ δ, we have that ∩i j=0Aξj . This forces dimA ≥ δ + 1, contrary to the choice of δ. The claim is established. Now, let p be a polynomial such that p(T )ξi = 0 for every 1 ≤ i ≤ δ. Then, [p] ∈ ∩δ i=0Aξi and therefore [p] = 0. This shows that U1× . . .× Uδ ⊂ H(δ) contains a separating vector for T (δ). Since U1, . . . , Uδ were arbitrary non-empty open subsets of H, we conclude that the set Σ ⊂ H(δ) of separating vectors for T (δ) is dense in H(δ). Finally, for s > δ it is readily verified that every vector in Σ ⊕ H(s−δ) is separating for T (s), so the result follows. (cid:3) We record the following simple reformulation. Corollary 5.4. Let N = (N1, . . . , Nd) be a nilpotent commuting row contraction on some Hilbert space H. Then, there is a finite set of cardinality at most δ which is separating for T , where δ is the dimension of C[x1, . . . , xd]/(C[x1, . . . , xd] ∩ Annc(N )). Proof. This is an immediate consequence of Theorem 5.3. (cid:3) Comparing with Example 4, it is apparent that the upper bound on the car- dinality of the separating set in the previous corollary is far from sharp. In fact, the argument used in Example 4 can be refined and extended to provide a better estimate. Because we do not have a concrete use for it, we only state the result and leave the details of the proof to the interested reader. Recall that we denote by [p] the image of a polynomial p in the quotient A. 30 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO Let N = (N1, . . . , Nd) be a nilpotent commuting row contraction on some Hilbert space H. Let Ωe = {α ∈ Nd : T α 6= 0 and T αTk = 0 for every 1 ≤ k ≤ d}. If {[xα] : α ∈ Ωe} is a linearly independent subset of A, then T admits a separating set of cardinality at most card Ωe. The reader will note that in Example 4, we have card Ωe = 1, while for Example 5 we find card Ωe = 2. References [1] Jim Agler and John E. McCarthy, Complete Nevanlinna-Pick kernels, J. Funct. Anal. 175 (2000), no. 1, 111 -- 124. MR1774853 [2] , Pick interpolation and Hilbert function spaces, Graduate Studies in Mathematics, vol. 44, American Mathematical Society, Providence, RI, 2002. MR1882259 (2003b:47001) [3] C. Apostol, R. G. Douglas, and C. Foia¸s, Quasi-similar models for nilpotent operators, Trans. Amer. Math. Soc. 224 (1976), no. 2, 407 -- 415. MR0425651 [4] Constantin Apostol and Joseph Stampfli, On derivation ranges, Indiana Univ. Math. J. 25 (1976), no. 9, 857 -- 869. MR0412890 [5] William Arveson, Subalgebras of C ∗-algebras. III. Multivariable operator theory, Acta Math. 181 (1998), no. 2, 159 -- 228. MR1668582 (2000e:47013) [6] [7] , The curvature invariant of a Hilbert module over C[z1, · · · , zd], J. Reine Angew. Math. 522 (2000), 173 -- 236. MR1758582 (2003a:47013) , The free cover of a row contraction, Doc. Math. 9 (2004), 137 -- 161 (electronic). MR2054985 (2005e:47013) [8] Hari Bercovici, Operator theory and arithmetic in H∞, Mathematical Surveys and Mono- graphs, vol. 26, American Mathematical Society, Providence, RI, 1988. MR954383 (90e:47001) [9] , The quasisimilarity orbits of invariant subspaces, J. Funct. Anal. 95 (1991), no. 2, 344 -- 363. MR1092130 (92e:47017) [10] Hari Bercovici and Thomas Smotzer, Quasisimilarity of invariant subspaces for uniform Jordan operators of infinite multiplicity, J. Funct. Anal. 140 (1996), no. 1, 87 -- 99. MR1404575 [11] Hari Bercovici and Allen Tannenbaum, The invariant subspaces of a uniform Jordan operator, J. Math. Anal. Appl. 156 (1991), no. 1, 220 -- 230. MR1102607 [12] T. Bhattacharyya, J. Eschmeier, and J. Sarkar, Characteristic function of a pure commuting contractive tuple, Integral Equations Operator Theory 53 (2005), no. 1, 23 -- 32. MR2183594 (2006m:47012) [13] David P. Blecher, Zhong-Jin Ruan, and Allan M. Sinclair, A characterization of operator algebras, J. Funct. Anal. 89 (1990), no. 1, 188 -- 201. MR1040962 [14] Xiaoman Chen and Kunyu Guo, Analytic Hilbert modules, Chapman & Hall/CRC Research Notes in Mathematics, vol. 433, Chapman & Hall/CRC, Boca Raton, FL, 2003. MR1988884 [15] Raphael Clouatre, Quasisimilarity of invariant subspaces for C0 operators with multiplicity two, J. Operator Theory 70 (2013), no. 2, 495 -- 511. MR3138367 [16] , Quasiaffine orbits of invariant subspaces for uniform Jordan operators, J. Funct. Anal. 266 (2014), no. 7, 4101 -- 4114. MR3170203 [17] Raphael Clouatre and Kenneth R. Davidson, Absolute continuity for commuting row con- tractions, J. Funct. Anal. 271 (2016), no. 3, 620 -- 641. MR3506960 [18] , Duality, convexity and peak interpolation in the Drury-Arveson space, Adv. Math. 295 (2016), 90 -- 149. MR3488033 [19] Raphael Clouatre and Michael Hartz, Multiplier algebras of complete Nevanlinna-Pick spaces: dilations, boundary representations and hyperrigidity, J. Funct. Anal. 274 (2018), no. 6, 1690 -- 1738. MR3758546 [20] Kenneth R. Davidson and David R. Pitts, The algebraic structure of non-commutative ana- lytic Toeplitz algebras, Math. Ann. 311 (1998), no. 2, 275 -- 303. MR1625750 (2001c:47082) [21] , Nevanlinna-Pick interpolation for non-commutative analytic Toeplitz algebras, In- tegral Equations Operator Theory 31 (1998), no. 3, 321 -- 337. MR1627901 (2000g:47016) CYCLIC ROW CONTRACTIONS AND RIGIDITY OF INVARIANT SUBSPACES 31 [22] , Invariant subspaces and hyper-reflexivity for free semigroup algebras, Proc. London Math. Soc. (3) 78 (1999), no. 2, 401 -- 430. MR1665248 (2000k:47005) [23] Kenneth R. Davidson, Christopher Ramsey, and Orr Moshe Shalit, Operator algebras for analytic varieties, Trans. Amer. Math. Soc. 367 (2015), no. 2, 1121 -- 1150. MR3280039 [24] Jim Gleason, Stefan Richter, and Carl Sundberg, On the index of invariant subspaces in spaces of analytic functions of several complex variables, J. Reine Angew. Math. 587 (2005), 49 -- 76. MR2186975 [25] Phillip Griffith, On the decomposition of modules and generalized left uniserial rings, Math. Ann. 184 (1969/1970), 300 -- 308. MR0257136 [26] Kunyu Guo, Junyun Hu, and Xianmin Xu, Toeplitz algebras, subnormal tuples and rigidity on reproducing C[z1, . . . , zd]-modules, J. Funct. Anal. 210 (2004), no. 1, 214 -- 247. MR2052120 (2005a:47007) [27] David W. Kribs, Factoring in non-commutative analytic Toeplitz algebras, J. Operator The- ory 45 (2001), no. 1, 175 -- 193. MR1823067 [28] Wing Suet Li and Vladim´ır Muller, Invariant subspaces of nilpotent operators and LR- sequences, Integral Equations Operator Theory 34 (1999), no. 2, 197 -- 226. MR1694708 (2001d:47013) [29] V. Muller and F.-H. Vasilescu, Standard models for some commuting multioperators, Proc. Amer. Math. Soc. 117 (1993), no. 4, 979 -- 989. MR1112498 (93e:47016) [30] Stephen Parrott, Unitary dilations for commuting contractions, Pacific J. Math. 34 (1970), 481 -- 490. MR0268710 (42 #3607) [31] Gelu Popescu, Characteristic functions for infinite sequences of noncommuting operators, J. [32] [33] [34] [35] [36] Operator Theory 22 (1989), no. 1, 51 -- 71. MR1026074 (91m:47012) , Isometric dilations for infinite sequences of noncommuting operators, Trans. Amer. Math. Soc. 316 (1989), no. 2, 523 -- 536. MR972704 (90c:47006) , von Neumann inequality for (B(H)n )1, Math. Scand. 68 (1991), no. 2, 292 -- 304. MR1129595 (92k:47073) , Non-commutative disc algebras and their representations, Proc. Amer. Math. Soc. 124 (1996), no. 7, 2137 -- 2148. MR1343719 (96k:47077) , Operator theory on noncommutative varieties, Indiana Univ. Math. J. 55 (2006), no. 2, 389 -- 442. MR2225440 (2007m:47008) , Similarity problems in noncommutative polydomains, J. Funct. Anal. 267 (2014), no. 11, 4446 -- 4498. MR3269883 [37] B´ela Sz.-Nagy, Ciprian Foias, Hari Bercovici, and L´aszl´o K´erchy, Harmonic analysis of op- erators on Hilbert space, enlarged, Universitext, Springer, New York, 2010. MR2760647 (2012b:47001) [38] N. Th. Varopoulos, On an inequality of von Neumann and an application of the metric theory of tensor products to operators theory, J. Functional Analysis 16 (1974), 83 -- 100. MR0355642 (50 #8116) Department of Mathematics, University of Manitoba, Winnipeg, Manitoba, Canada R3T 2N2 E-mail address: [email protected] E-mail address: [email protected]
1209.2574	1	1209	2012-09-12T11:35:32	On Iyengar-Type Inequalities via Quasi-Convexity and Quasi-Concavity	[ "math.FA" ]	In this paper, we obtain some new estimations of Iyengar-type inequality in which quasi-convex(quasi-concave) functions are involved. These estimations are improvements of some recently obtained estimations. Some error estimations for the trapezoidal formula are given. Applications for special means are also provided.	math.FA	math	ON IYENGAR-TYPE INEQUALITIES VIA QUASI-CONVEXITY AND QUASI-CONCAVITY M. EMIN OZDEMIR⋆ Abstract. In this paper, we obtain some new estimations of Iyengar-type inequality in which quasi-convex(quasi-concave) functions are involved. These estimations are improvements of some recently obtained estimations. Some error estimations for the trapezoidal formula are given. Applications for special means are also provided. 1. Short Historical Background and Introduction If it is necessary to bound one quantity by another, the classical inequalities are very useful for this purpose. This first book called " Inequalities " written by Hardy, Littlewood and Polya at cambridge University Press in 1934 represents the first effort to systemize a rapidly expanding domain. In this sense, the second important book " Classical and New Inequalities in Analysis " is written by D.S. Mitrinovi´c, J.E.Pe´caric and A.M. Fink. The third book " Analytic Inequalities " written by D.S. Mitrinovi´c, and the other book " Means and Their Inequalities " written by Bullen, D.S. Mitrinovi´c, D.S. Vasic, P.M. Today inequalities play a significant role for the development in all fields of Mathematics. They have applications in a variety of applied Mathematics. For example, convex functions are tractable in optimization because local optimality guarantees global optimality. In recent years a number of authors have discov- ered new integral inequalities for convex, s−convex functions, logarithmic con- vex functions, h−convex functions, quasi-convex functions, m−convex functions, (α, m)−convex functions, co-ordinated convex functions, and Godunova-Levin func- tion, P −function. On November 22, 1881, Hermite (1822-1901) sent a letter to the Journal Math- esis. This letter was published in Mathesis 3 (1883,p.82) and in this letter an inequality presented which is well-known in the literature as Hermite-Hadamard integral inequality : (1.1) f(cid:18) a + b 2 (cid:19) ≤ 1 b − aZ b a f (x) dx ≤ f (a) + f (b) 2 , where f : I ⊆ R → R is a convex function on the interval I of a real numbers and a, b ∈ I with a < b. If the function f is concave, the inequality in (1.1) is reversed. That is f(cid:18) a + b 2 (cid:19) ≥ 1 b − aZ b a f (x) dx ≥ f (a) + f (b) 2 . 2000 Mathematics Subject Classification. Primary 26D15. Key words and phrases. Weighted Holder Inequality, Holder Inequality, Power-mean Inequal- ity, Differentiable Function, quasi-convex Function. ⋆Corresponding Author. 1 2 M. EMIN OZDEMIR⋆ For recent results, generalizations and new inequalities related to the inequality (1.1) see ([7]-[17]). Then left hand side of Hermite-Hadamard inequality (LHH) can also be esti- mated by the inequality of Iyengar. f (a) + f (b) 2 (1.2) where − f (x) dx ≤ M (b − a) 4 − [f (b) − f (a)]2 4M (b − a) In [3], Daniel Alexandru Ion proved the following inequalities of Iyengar type for differentiable quasi−convex functions: (1.3) f (a) + f (b) 2 (b − a) 4 (sup {f ′ (a) , f ′ (b)}) where f : [a, b] → R is differentiable function on (a, b) , and f ′ is quasi−convex on [a, b] with a < b. (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) and (1.4) ≤ f (x) − f (y) x − y ; x 6= y(cid:27) (cid:12)(cid:12)(cid:12)(cid:12) a 1 b − aZ b M = sup(cid:26)(cid:12)(cid:12)(cid:12)(cid:12) b − aZ b 1 − a ≤ f (x) dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) b − aZ b p (cid:16)supnf ′ (a) − 1 a 1 f (a) + f (b) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 2 (b − a) 2 (p + 1) f (x) dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) p p−1 , f ′ (b) p−1 p p p−1o(cid:17) where f : [a, b] → R is differentiable function on (a, b) , and f ′ p p−1 is quasi−convex on [a, b] with a < b. We give some necessary definitions and mathematical preliminaries for quasi−convex functions which are used throughout this paper. Definition 1. (see [1]) A function f : [a, b] → R is said to be quasi−convex on [a, b] if (1.5) f (λx + (1 − λ) y) ≤ max {f (x) , f (y)} , holds for all x, y ∈ [a, b] and λ ∈ [0, 1] .For additional results on quasi−convexity, see [2]. Clearly, any convex function is quasi−convex function. Furthermore, there exists quasi−convex functions which are not convex. See [3] : 1, t2, t ∈ [−2, −1] t ∈ (−1, 2] g (t) =  is not a convex function on [−2, 2] ,but it is a quasi−convex function on [−2, 2] .If we choose g : [−2, 2] → R, g (−2) = 1, g (2) = 4 and for α = 1 2 , a = −2, b = 0, we get g (αa + (1 − α) b) = g (−1) = 1 and αg (a) + (1 − α) g (b) = 1 2 g (−2) + 1 2 g (0) = 1 2 .Thus it is not convex but it is quasi−convex function for all α ∈ [0, 1] , g (−α2 + (1 − α) 2) ≤ max {g (−2) , g (2)} = max {1, 4} = 4. The main purpose of this paper is to point out new estimations of the inequality in (1.2) , but now for the class of quasi−convex functions. In order to prove our main results we need the following lemma (see [4]). ON IYENGAR-TYPE INEQUALITIES 3 Lemma 1. Let f : I ⊂ R → R be a twice differentiable mapping on I ◦, a, b ∈ I with a < b and f ′′ be integrable on [a, b]. Then the following equality holds: f (a) + f (b) 2 − 1 b − aZ b a f (x)dx = (b − a)2 2 Z 1 0 t (1 − t) f ′′ (ta + (1 − t)b) dt. The main results of this paper are given by the following theorems. 2. The Results Theorem 1. Let f : I ◦ ⊂ [0, ∞) → R, be a twice differentiable mapping on I ◦, such that f ′′ ∈ L[a, b], a, b ∈ I with a < b. If f ′′q is quasi−convex on [a, b] for q > 1, then the following inequality holds: (2.1) ≤ q a 2 2 1 q−1 − (b − a)2 f (a) + f (b) f (x)dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) , f ′′(b)q(cid:9)(cid:1) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) b − aZ b (cid:18) q − 1 2q − p − 1(cid:19) ×(cid:0)max(cid:8)f ′′(a)q β (x , y) =Z 1 tx−1 (1 − t)y−1 dt, 1 q 0 (β (p + 1, q + 1)) 1 q x, y > 0. where 1 p + 1 q = 1 and β ( , ) is Euler Beta Function: Theorem 2. Let f : I ◦ ⊂ [0, ∞) → R, be a twice differentiable mapping on I ◦, such that f ′′ ∈ L[a, b], a, b ∈ I with a < b. If f ′′q is quasi−convex on [a, b] for q ≥ 1, then the following inequality holds: Theorem 3. With the assumptions of Theorem 1, we obtain another (2.2) ≤ f (a) + f (b) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 2 (b − a)2 (cid:18) q a 2 1 q−1 − f (x)dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) b − aZ b (q + 1) (q + 2)(cid:19) (cid:0)max(cid:8)f ′′(a)q f (x)dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) p (cid:0)max(cid:8)f ′′(a)q b − aZ b (β (2, p + 1)) − 1 a 1 3. proof of main results 4 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) f (a) + f (b) 2 (b − a)2 21+ 1 q ≤ 1 q . , f ′′(b)q(cid:9)(cid:1) 1 q . , f ′′(b)q(cid:9)(cid:1) Proof of Theorem 1: Using Lemma 1 and the well known Holder's inequality for q > 1, ≤ (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) f (a) + f (b) 2 − (b − a)2 2 (cid:18)Z 1 0 t a 1 b − aZ b q−1 dt(cid:19) f (x) dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) q (cid:20)Z 1 q−1 q−p 0 tp (1 − t)q f ′′ (ta + (1 − t) b)q 1 q , dt(cid:21) 4 M. EMIN OZDEMIR⋆ where 1 p + 1 q = 1. On the other hand, since f ′′q is quasi−convex on [a, b], we know that for any t ∈ [0, 1] Therefore, we obtain f ′′ (ta + (1 − t) b)q ≤ max(cid:8)f ′′(a)q , f ′′(b)q(cid:9) . f (a) + f (b) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 2 (b − a)2 2 (b − a)2 2 (b − a)2 2 q−1 a 0 t 1 q−p − b − aZ b (cid:18)Z 1 q−1 dt(cid:19) (cid:18)Z 1 q−1 dt(cid:19) 2q − p − 1(cid:19) (cid:18) q − 1 q−p t 0 q q−1 f (x) dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) q (cid:20)Z 1 q (cid:20)Z 1 tp (1 − t)q(cid:0)max(cid:8)f ′′(a)q 0 0 q−1 tp (1 − t)q f ′′ (ta + (1 − t) b)q (β (p + 1, q + 1)) 1 q (cid:0)max(cid:8)f ′′(a)q ≤ ≤ = 1 q dt(cid:21) 1 q 1 q , , f ′′(b)q(cid:9)(cid:1) dt(cid:21) , f ′′(b)q(cid:9)(cid:1) which completes the proof. Corollary 1. In Theorem 1, if we choose M = Supx∈(a,b) f ′′(x) < ∞, we get (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) f (a) + f (b) 2 1 − b − aZ b 2q − p − 1(cid:19) M(cid:18) q − 1 a ≤ (b − a)2 2 f (x) dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) q−1 q (β (p + 1, q + 1)) 1 q . Proof of Theorem 2: From Lemma 1 and the well known power-mean in- equality we obtain t (1 − t) f ′′ (ta + (1 − t) b) dt f (a) + f (b) 2 − (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (b − a)2 2 (b − a)2 2 (b − a)2 2 (b − a)2 2 (b − a)2 4 a 0 0 1 1− 1 b − aZ b f (x) dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) q (cid:18)Z 1 q (cid:18)Z 1 (q + 1)(q + 2)(cid:19) Z 1 (cid:18)Z 1 tdt(cid:19) (cid:18)Z 1 tdt(cid:19) 2(cid:19)1− 1 q (cid:18) (cid:18) 1 (q + 1)(q + 2)(cid:19) (cid:18) 1− 1 2 1 1 q 0 0 0 ≤ ≤ ≤ = = 1 q dt(cid:19) t (1 − t)q f ′′ (ta + (1 − t) b)q 1 q t (1 − t)q(cid:0)max(cid:8)f ′′(a)q (cid:0)max(cid:8)f ′′(a)q , f ′′(b)q(cid:9)(cid:1) (cid:0)max(cid:8)f ′′(a)q 1 q , f ′′(b)q(cid:9)(cid:1) dt(cid:19) , f ′′(b)q(cid:9)(cid:1) 1 q 1 q . The proof of Theorem 2 is completed. Corollary 2. Under the assumptions of Theorem 2, ON IYENGAR-TYPE INEQUALITIES 5 have 2 Case i: Since limq→∞(cid:16) <(cid:18) (q+1)(q+2)(cid:17) (q + 1)(q + 2)(cid:19) 1 3 2 1 q = 1 and limq→1+(cid:16) 1 q 2 (q+1)(q+2)(cid:17) = 1 3 , we 1 q < 1, q ∈ [1, ∞). Therefore, f (a) + f (b) 2 − 1 b − aZ b a (3.1) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) In (3.1), ≤ (b − a)2 4 (cid:0)max(cid:8)f ′′(a)q , f ′′(b)q(cid:9)(cid:1) 1 q . f (x) dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) • if f ′′q is decreasing, we get f (a) + f (b) 2 − 1 b − aZ b a • if f ′′q is increasing , we get f (a) + f (b) 2 − 1 b − aZ b a (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ (b − a)2 4 f ′′(a) , ≤ (b − a)2 4 f ′′(b) . f (x) dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) f (x) dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Case ii: If we choose M = Supx∈(a,b) f ′′(x) < ∞ in (2.2), then the inequality in (2.1) is better than the inequality in (2.2). Proof of Theorem 3: From Lemma 1 with properties of modulus we get (3.2) ≤ (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) f (a) + f (b) 2 − 1 b − aZ b a (b − a)2 2 Z 1 0 f (x) dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) t (1 − t) f ′′ (ta + (1 − t) b) dt. Now, if we use the following weighted version of Holder's inequality [5, p. 117]: (3.3) ZI (cid:12)(cid:12)(cid:12)(cid:12) f (s)g(s)h(s)ds(cid:12)(cid:12)(cid:12)(cid:12) ≤(cid:18)ZI f (s)p h(s)ds(cid:19) 1 p (cid:18)ZI 1 q g(s)q h(s)ds(cid:19) for p > 1, p−1 + q−1 = 1, h is nonnegative on I and provided all the other integrals exist and are finite. 6 M. EMIN OZDEMIR⋆ If we rewrite the inequality (3.2) with respect to (3.3) with f ′′q is quasi−convex on [a, b] for all t ∈ [0, 1], we get f (a) + f (b) 2 − 1 b − aZ b a f (x) dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) t (1 − t) f ′′ (ta + (1 − t) b) dt (1 − t) f ′′ (ta + (1 − t) b) tdt (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 0 Z 1 Z 1 (cid:18)Z 1 0 0 (b − a)2 2 (b − a)2 2 (b − a)2 2 (b − a)2 2 (1 − t)p tdt(cid:19) 1 p (cid:18)Z 1 0 f ′′ (ta + (1 − t) b)q 1 q tdt(cid:19) (β (2, p + 1)) 1 p max(cid:8)f ′′(a)q , f ′′(b)q(cid:9) 2 1 q ! . The proof of Theorem 3 is completed. Corollary 3. In Theorem 3, if we choose M = Supx∈(a,b) f ′′(x) < ∞, we get f (a) + f (b) 2 − 1 b − aZ b a (b − a)2 21+ 1 q M (β (2, p + 1)) 1 p . Remark 1. From Theorems 1-3, we get ≤ f (x) dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) f (x) dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) b − aZ b 1 a − q−1 q (β (p + 1, q + 1)) 1 ≤ min {v1, v2, v3} ≤ = ≤ = (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 2 f (a) + f (b) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:18) q − 1 2q − p − 1(cid:19) (cid:18) (b − a)2 4 where v1 = (b − a)2 2 v2 = and q−1 q 2 (q + 1) (q + 2)(cid:19) 1 q , , f ′′(b)q(cid:9)(cid:1) q ×(cid:0)max(cid:8)f ′′(a)q , f ′′(b)q(cid:9)(cid:1) (cid:0)max(cid:8)f ′′(a)q 1 q v3 = (b − a)2 21+ 1 q (β (2, p + 1)) 1 p (cid:0)max(cid:8)f ′′(a)q , f ′′(b)q(cid:9)(cid:1) 1 q . 4. Error Estimates for the Trapezoidal Rule Let d be a partition a = x0 < x1 < x2 < ... < xn = b of the interval [a, b] and consider the quadrature formula (4.1) where Z b a f (x)dx = Ti(f, d) + Ei(f, d), i = 1, 2, ..., n − 1 T1(f, d) = n−1 Xi=0 f (xi) + f (xi+1) 2 (xi+1 − xi) ON IYENGAR-TYPE INEQUALITIES 7 for the Trapezoidal version and n−1 T2(f, d) = f(cid:18) xi + xi+1 2 (cid:19) (xi+1 − xi) Xi=0 for the Midpoint formula and Ei(f, d) denotes the associated approximation errors. Proposition 1. Suppose that all the assumptions of Theorem 1 are satisfied for every division d of [a, b], we have E(f, d) ≤ q−1 (β (p + 1, q + 1)) q 1 n−1 2q − p − 1(cid:19) 2(cid:18) q − 1 (xi+1 − xi)3(cid:0)max(cid:8)f ′′(xi)q Xi=0 × 1 q , f ′′(xi+1)q(cid:9)(cid:1) 1 q . Proof. Appliying Theorem 1 on the subinterval (xi+1, xi) , i = 1, 2, ..., n − 1 of the partition and by using the quasi−convexity of f ′′q , we obtain Hence in (4.1), we have a (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Z b f (x)dx − T (f, d)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ f (x)dx(cid:12)(cid:12)(cid:12)(cid:12) (β (p + 1, q + 1)) 1 q 1 2 xi 1 2 q−1 q − q . n−1 (xi+1 − xi)2 f (xi) + f (xi+1) (xi+1 − xi)Z xi+1 (cid:18) q − 1 2q − p − 1(cid:19) , f ′′(xi+1)q(cid:9)(cid:1) (cid:12)(cid:12)(cid:12)(cid:12) ×(cid:0)max(cid:8)f ′′(xi)q = (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Xi=0(cid:26)Z xi+1 Z xi+1 Xi=0(cid:12)(cid:12)(cid:12)(cid:12) 2q − p − 1(cid:19) 2(cid:18) q − 1 (xi+1 − xi)3(cid:0)max(cid:8)f ′′(xi)q Xi=0 f (x)dx − f (x)dx − n−1 n−1 × ≤ ≤ q−1 q 1 2 2 xi xi (β (p + 1, q + 1)) f (xi) + f (xi+1) f (xi) + f (xi+1) (xi+1 − xi)(cid:27)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (xi+1 − xi)(cid:12)(cid:12)(cid:12)(cid:12) , f ′′(xi+1)q(cid:9)(cid:1) q . 1 q 1 (cid:3) Proposition 2. Suppose that all the assumptions of Theorem 2 are satisfied for every division d of [a, b], we have E(f, d) ≤ 1 q 2 1 n−1 4(cid:18) (q + 1)(q + 2)(cid:19) (xi+1 − xi)3(cid:0)max(cid:8)f ′′(xi)q Xi=0 × , f ′′(xi+1)q(cid:9)(cid:1) 1 q . Proof. The proof is immediate follows from Theorem 2 and by applying a similar argument to the Proposition 1. (cid:3) 8 M. EMIN OZDEMIR⋆ Proposition 3. Suppose that all the assumptions of Theorem 3 are satisfied for every division d of [a, b], we have (β (1, p + 1)) 1 p E(f, d) ≤ 1 2 × (xi+1 − xi)2 max(cid:8)f ′′(a)q , f ′′(b)q(cid:9) 2 ! 1 q . n−1 Xi=0 Proof. The proof is immediate follows from Theorem 3 and by applying a similar argument to the Proposition 1. (cid:3) 5. Applications to special means Let us consider the special means for real numbers a, b (a 6= b). We take 1. Arithmetic mean: A(a, b) = a + b 2 , a, b ∈ R. 2. Logarithmic mean: L(a, b) = a − b ln a − ln b 3. Generalized log-mean: Ln(a, b) =(cid:20) bn+1 − an+1 (n + 1)(b − a)(cid:21) a 6= b , a, b 6= 0, a, b ∈ R. , 1 n , n ∈ Z\ {−1, 0} , a, b ∈ R, a 6= b. Proposition 4. Let a, b ∈ R, a < b and n ∈ N, n ≥ 2. Then we have A(an, bn) − Ln n(a, b) ≤ n(n − 1) (b − a)2 2 ×(cid:18) q − 1 2q − p − 1(cid:19) q−1 q (β (p + 1, q + 1)) 1 q (cid:16)maxna(n−2)q , b(n−2)qo(cid:17) 1 q . Proof. The assertion follows from Theorem 1 applied to the quasi−convex mapping f (x) = xn, x ∈ R. (cid:3) Proposition 5. Let a, b ∈ R, a < b and n ∈ N, n ≥ 2. Then we have A(an, bn) − Ln n(a, b) ≤ n(n − 1) (b − a)2 4 1 ×(cid:18) 2 (q + 1)(q + 2)(cid:19) q (cid:16)maxna(n−2)q , b(n−2)qo(cid:17) 1 q . Proof. The assertion follows from Theorem 2 applied to the quasi−convex mapping f (x) = xn, x ∈ R. (cid:3) Proposition 6. Let a, b ∈ R, a < b and n ∈ N, n ≥ 2. Then we have A(an, bn) − Ln n(a, b) ≤ n(n − 1) (b − a)2 21+ 1 q × (β (2, q + 1)) 1 p (cid:16)maxna(n−2)q , b(n−2)qo(cid:17) 1 q . Proof. The assertion follows from Theorem 3 applied to the quasi−convex mapping f (x) = xn, x ∈ R. (cid:3) ON IYENGAR-TYPE INEQUALITIES 9 References [1] J. Pe´caric, F.Proschan, and Y.L.Tong, Convex Functions, Partial Orderings and Statistical Applications, Academic Press(1992), Inc. [2] J. Ponstein, Seven kinds of convexity, SIAM Review 9, 115-119 (1967) [3] D. A. Ion, Some estimates on the Hermite- Hadamard inequality through Quasi- convex functions, Annals of University of Craiova Math. Comp.Sc. Ser A (2007), 82-87. [4] M. Alomari, M: Darus and type Hadamard quasi−convex, RGMIA Res. Rep.Coll.12(2009) Supplement, Article htpp://www.staff.vu.edu.au/RGMIA/v12(E).asp. functions whose for S.S.Dragomir, second New inequalities derivatives absolute of Hermite- are 17, Online: value [5] S.S.Dragomir, R.P.Agarwal, N.S. Barnett, Inequalities for Beta and Gamma Functions via some classical and new integral inequalities, Journal of Inequalities and Applications, Vol 5, pp.103-165,2000. [6] D.S. Mitrinovi´c, J.E.Pe´caric. A.M. Fink, Classical and new Inequalities in analysis, Kluwer Academic Publishers, 1993, p.106,10,15. [7] U.S. Kirmaci, M. K. Bakula, M. E. Ozdemir and J. Pe´caric, Hadamard type inequalities for s−convex functions, Applied Mathematics and Computation, 193 (2007), p.106, 26-35. [8] M.K. Bakula, M.E. Ozdemir and J. Pe´caric, Hadamard type inequalities for m−convex and (α, m)−convex functions, Journal of Inequalities in Pure and Applied Mathematics, Vol. 9 (2008), Issue 4, article 96. [9] E. Set, M.E. Ozdemir and S.S.Dragomir, On the Hermite-Hadamard inequality and other integral inequalities involving two functions, Journal of Inequalities and Applications, 2010, Article ID 148102,9 pages,doi.10.1155/2010/148102. [10] M.E. Ozdemir, C¸ . Yildiz, A. O. Akdemir, On some new Hadamard type inequalities for Coordinated quasi−convex functions, Submitted. [11] E. Set, M.E. Ozdemir, S.S.Dragomir, On Hadamard type inequalities involving several kinds of convexity, Journal of Inequalities and Applications, 2010, Article ID 1286845,12 pages,doi.10.1155/2010/286845. [12] M. Avci, H. Kavurmaci, M. E. Ozdemir, New inequalities of Hermite-Hadamard type via s−convex functions in the second sense with applications, Appl. Math. Comput., 217 (2011) 5171-5176. [13] M. E. Ozdemir, M. Avci and H. Kavurmaci, Hermite-Hadamard type inequalities via (α, m)−convexity, Comput. Math. Appl., 61 (2011) 2614 -- 2620. [14] M. E. Ozdemir, M. Avci and E. Set, On some inequalities of Hermite-Hadamard type via m−convexity, Appl. Math. Lett., 23 (2010) 1065 1070. [15] H. Kavurmaci, M. Avci and M. E. Ozdemir, New inequalities of Hermite-Hadamard type for convex functions with applications, Journal of Inequalities and Applications 2011, 2011:86. [16] M.E. Ozdemir, A. O. Akdemir, On some Hadamard type inequalities for convex functions on a rectangular box, Journal of non linear Analysis and Application, Volume 2011, Year 2011 Article ID jnaa-00101, 10pages, doi: 10.5899/2011/jnaa-00101, Research Article. [17] A. O. Akdemir, M.E. Ozdemir and S. Varosanec, On some inequalities for h−concave func- tions, Mathematical and Computer Modelling 55 (2012) 746-753. ⋆ATAT URK UNIVERSITY, K.K. EDUCATION FACULTY, DEPARTMENT OF MATH- EMATICS, 25240, CAMPUS, ERZURUM, TURKEY E-mail address: [email protected]
1910.11053	1	1910	2019-10-24T12:30:07	Minimal passive realizations of generalized Schur functions in Pontryagin spaces	[ "math.FA" ]	Passive discrete-time systems in Pontryagin space setting are investigated. In this case the transfer functions of passive systems, or characteristic functions of contractive operator colligations, are generalized Schur functions. The existence of optimal and *-optimal minimal realizations for generalized Schur functions are proved. By using those realizations, a new definition, which covers the case of generalized Schur functions, is given for defects functions. A criterion due to D.Z. Arov and M.A. Nudelman, when all minimal passive realizations of the same Schur function are unitarily similar, is generalized to the class of generalized Schur functions. The approach used here is new; it relies completely on the theory of passive systems.	math.FA	math	Minimal passive realizations of generalized Schur functions in Pontryagin spaces Lassi Lilleberg Abstract. Passive discrete-time systems in Pontryagin space setting are investigated. In this case the transfer functions of passive systems, or characteristic functions of contractive operator colligations, are gener- alized Schur functions. The existence of optimal and ∗-optimal minimal realizations for generalized Schur functions are proved. By using those realizations, a new definition, which covers the case of generalized Schur functions, is given for defects functions. A criterion due to D.Z. Arov and M.A. Nudelman, when all minimal passive realizations of the same Schur function are unitarily similar, is generalized to the class of general- ized Schur functions. The approach used here is new; it relies completely on the theory of passive systems. 1. Introduction An operator colligation Σ = (TΣ; X , U , Y; κ) consists of separable Pontryagin spaces X (the state space), U (the incoming space), and Y (the outgoing space) and the system operator TΣ ∈ L(X ⊕ U , X ⊕ Y), the space of bounded U(cid:19), means the direct operators from X ⊕ U to X ⊕ Y, where X ⊕ U , or (cid:18)X orthogonal sum with respect to the indefinite inner product. The symbol κ is reserved for the finite negative index of the state space. The operator TΣ has the block representation of the form TΣ =(cid:18)A B C D(cid:19) :(cid:18)X U(cid:19) →(cid:18)X Y(cid:19) , (1.1) Department of Mathematics and Statistics, University of Vaasa, P.O. Box 700, 65101 Vaasa, Finland E-mail address: [email protected] . Mathematics Subject Classification (2010). Primary: 47A48; Secondary: 47A56, 47B50, 93B05, 93B07, 93B28. Keywords. Operator colligation, passive system, transfer function, defect functions, generalized Schur class, contractive operator. 2 L. Lilleberg where A ∈ L(X ) (the main operator), B ∈ L(U , X ) (the control operator), C ∈ L(X , Y) (the observation operator), and D ∈ L(U , Y) (the feedthrough operator). If needed, the colligation is written as Σ = (A, B, C, D; X , U , Y; κ). It is always assumed in this paper that U and Y have the same negative index. All notions of continuity and convergence are understood to be with respect to the strong topology, which is induced by any fundamental decomposition of the space in question. The colligation (1.1) will be called as a system since it can be seen as a linear discrete time system of the form (hk+1 = Ahk + Bξk, = Chk + Dξk, σk k ≥ 0, where {hk} ⊂ X , {ξk} ⊂ U and {σk} ⊂ Y. In what follows, the "system" is identified with the operator expression appearing in (1.1). When the system operator TΣ in (1.1) is contractive (isometric, co-isometric, unitary), with respect to the indefinite inner product, the corresponding system is called passive (isometric, co-isometric, conservative). In literature, conservative sys- tems are also called unitary systems. The transfer function of the system (1.1) is defined by θΣ(z) := D + zC(I − zA)−1B, (1.2) whenever I − zA is invertible. Especially, θΣ is defined and holomorphic in a neighbourhood of the origin. The values θΣ(z) are bounded operators from U to Y. Conversely, if θ is an operator valued function holomorphic in a neighbourhood of the origin, and transfer function of the system Σ coinsides with it, then Σ is a realization of θ. In some sources, transfer functions of the systems are also called characteristic functions of operator colligations. Σ of TΣ. That is, Σ∗ = (T ∗ The adjoint or dual of the system Σ is the system Σ∗ such that its system operator is the indefinite adjoint T ∗ Σ; X , Y, U; κ). In this paper, all the adjoints are with respect to the indefinite inner product. For an operator valued function ϕ, the notation ϕ∗(z) is used instead of (ϕ(z))∗ , and the function ϕ#(z) is defined to be ϕ∗(¯z). With this notation, #(z). Since for the transfer function θΣ∗ of Σ∗, it clearly holds θΣ∗ (z) = θΣ contractions between Pontryagin spaces with the same negative index are bi-contractions (cf. eg. [24, Corollary 2.5]), Σ∗ is passive whenever Σ is. In the case where all the spaces are Hilbert spaces, it is well known; see for instance [6, Proposition 8], that the transfer function of the passive system is a Schur function. That is, contractive operator valued function holomorphic in the unit disk D. In the case where U and Y are Hilbert spaces and the state space X is a Pontryagin space, Saprikin showed in [28, Theorem 2.2] that the transfer function of the passive system (1.1) is a generalized Schur function. It will be proved later in Proposition 2.5 that this result holds also in the case when all the spaces are Pontryagin spaces. The generalized Schur class Sκ(U , Y), where U and Y are Pontryagin spaces with the same negative index, is the set of L(U , Y)-valued functions S(z) holomorphic in a neighbourhood (cid:0)hKS(wj, wi)fj, fiiY(cid:1)n i,j=1 , (1.4) Minimal realizations 3 Ω of the origin such that the Schur kernel KS(w, z) = 1 − S(z)S ∗(w) 1 − z ¯w , w, z ∈ Ω, (1.3) has κ negative squares (κ = 0, 1, 2, . . .). This means that for any finite set of points w1, . . . , wn in the domain of holomorphy ρ(S) of S and set of vectors {f1, . . . , fn} ⊂ Y, the Hermitian matrix where h·, ·iY is the indefinite inner product of the space Y, has no more than κ negative eigenvalues, and there exists at least one such matrix that has exactly κ negative eigenvalues. A function S belongs to Sκ(U , Y) if and only if S# κ ∈ S(Y, U); see [1, Theorem 2.5.2]. The class S0(U , Y) coinsides with the ordinary Schur class, and it is written as S(U , Y). The direct connection between the transfer functions of passive systems of the form (1.1) and the generalized Schur functions allows to study the prop- erties of generalized Schur functions by using passive systems, and vice versa. Therefore, a fundamental problem of the subject is, for a given θ ∈ Sκ(U , Y), find a realization Σ of θ with the desired minimality or optimality properties (observable, controllable, simple, minimal, optimal, ∗-optimal); for details, see Theorems 2.7 and 3.5 and Lemma 2.9. The described problem is called a realization problem. In the standard Hilbert space settings, realizations problems, as well as other properties of passive systems, were studied, for instance, by Ando [3], Arov [5, 6], Arov et al. [7, 8, 9, 10], de Branges and Rovnyak [19, 20], Brodskiı [21], Helton [25] and Sz.-Nagy and Foias [30]. The case where the state space is a Pontryagin space while incoming and outgo- ing spaces are still Hilbert spaces, unitary systems were studied, for instance, by Dijksma et al. [22, 23], and passive systems by Saprikin [28], Saprikin and Arov [12], Saprikin et al. [11] and by the author in [27]. The case where all the spaces are Pontryagin spaces, theory of isometric, co-isometric and conservative systems is considered, for instance, in [1, 2, 24]. Especially, in [6], Arov proved the existence of so-called optimal minimal realizations of an ordinary Schur function; for definitions, see Section 3. The proof was based on the existence (right) defect functions. For an ordinary Schur function S(ζ), the (right) defect function ϕ of S is, roughly speaking, the maximal analytic minorant of I − S ∗(ζ)S(ζ). More precicely, this means that for almost every (a.e.) ζ on the unit circle T, it holds ϕ∗(ζ)ϕ(ζ) ≤ I − S ∗(ζ)S(ζ), it holds and for every other operator valued analytic function bϕ with similar property, For the existence of defect functions, see [30, Theorem V.4.2], and for a detailed treatise, see [16, 17, 18]. In [7], Arov et al. constructed (∗-)optimal minimal passive systems in the Hilbert space setting without using defect functions. The construction can be done by taking an appropriate restriction bϕ∗(ζ)bϕ(ζ) ≤ ϕ∗(ζ)ϕ(ζ). 4 L. Lilleberg of some system. In the indefinite setting, if one uses a suitable definition of optimality, a similar method as was used by Arov et al. still produces a (∗- )optimal minimal passive system. In Pontryagin state space case, this was proved by Saprikin [28]. It will be shown in Theorem 3.5 that the same result still holds in the case where all the spaces are Pontryagin spaces. r The study of the class of generalized Schur functions Sκ(U , Y) was con- tinued in [11, 12], in the case where U and Y are Hilbert spaces and the state space is a Pontryagin space. In [12], Saprikin and Arov used the right Kreın- Langer factorization of the form S = SrB−1 for S ∈ Sκ(U , Y), and proved that the existence of the optimal minimal realization of S is equivalent to the existence of the right defect function of Sr. However, they did not define the defect functions for the generalized Schur functions. This was done by the author in [27] by using the Kreın-Langer factorizations. With the defini- tion given therein, the main results of [4] were generalized to the Pontryagin state space setting. The main subjects of [27] include some continuation of the study of products of systems and the stability properties of passive sys- tems, subjects treated earlier by Saprikin et al. in [11]. In the present paper, it will be shown that a concept of defect functions can be defined in the case where all the spaces are Pontryagin spaces. The key idea here is to use optimal minimal passive realizations and conservative embeddings. By using such a definition, it is shown that one can generalize and improve some of the main results from [4], using different proofs than those given in [4] or [27], see Theorem 4.8. Furthermore, in Theorem 4.10, the main results from [9, 10] concerning the criterion when all the minimal realizations of a Schur function are unitarily similar, is generalized to the present indefinite setting. The proof will be carried out entirely by using the theory of passive systems, without applying Hardy space theory or the theory of Hankel operators as in the proof provided in [10]. The paper is organized as follows. In Section 2 basic facts of linear systems, Julia operators, dilations and embeddings are recalled. As a prepa- ration, it is shown in Proposition 2.3 that an arbitrary passive linear system has a conservative dilation. Moreover, Lemma 2.9 gives some usefull rep- resentations and restrictions of passive systems. That lemma will be used extensively later on in this paper. In Section 3, the existence and basic properties of (∗-)optimal minimal realizations are established. The main result of this section is Theorem 3.5. The generalized defect functions are introduced in Section 4. In partic- ularly, Theorem 4.10 in this section can be seen as the main result of the paper. Minimal realizations 5 2. Linear systems, dilations and embeddings Let Σ = (TΣ; X , U , Y; κ) be a linear system as in (1.1). The following sub- spaces X c := span {ran AnB : n = 0, 1, . . .} X o := span {ran A∗nC ∗ : n = 0, 1, . . .} X s := span {ran AnB, ran A∗mC ∗ : n, m = 0, 1, . . .}, (2.1) (2.2) (2.3) are called, respectively, controllable, observable and simple subspaces. The system is said to be controllable (observable, simple) if X c = X (X o = X , X s = X ) and minimal if it is both controllable and observable. When Ω ∋ 0 is some symmetric neighbourhood of the origin, that is, ¯z ∈ Ω when- ever z ∈ Ω, then also X c = span {ran (I − zA)−1B, z ∈ Ω} X o = span {ran (I − zA∗)−1C ∗, z ∈ Ω} X s = span {ran (I − zA)−1B, ran (I − wA∗)−1C ∗, z, w ∈ Ω} (2.4) (2.5) (2.6) The system (1.1) can be expanded to a larger system without changing the transfer function. It can be done by using the so-called Julia operator, see (2.8) below. For a proof of the following theorem and more details about the Julia operators, see [24]. The basic information about the indefinite inner product spaces and their operators can be recalled from [13, 15, 24]. Theorem 2.1. Suppose that X1 and X2 are Pontryagin spaces with the same negative index, and let A : X1 → X2 be a contraction. Then there exist Hilbert spaces DA and DA∗ , linear operators DA : DA → X1, DA∗ : DA∗ → X2 with zero kernels and a linear operator L : DA → DA∗ such that it holds I − A∗A = DAD∗ A, I − AA∗ = DA∗D∗ A∗ , (2.7) and the operator UA :=(cid:18) A DA∗ A −L∗(cid:19) :(cid:18) X1 D∗ DA(cid:19) DA∗(cid:19) →(cid:18) X2 (2.8) is unitary. Moreover, DA, DA∗ and UA are unique up to unitary equivalence. Remark 2.2. The operator DA from Theorem 2.1 is called a defect operator of A. A dilation of a system Σ = (A, B, C, D; X , U , Y; κ) is any system of the form bΣ = (bA, bB, bC, D; bX , U , Y; κ), where bX = D ⊕ X ⊕ D∗, bAD ⊂ D, bA∗D∗ ⊂ D∗, bCD = {0}, bB∗D∗ = {0}. (2.9) 6 L. Lilleberg The spaces D and D∗ are required to be Hilbert spaces. The system operator B1 B (2.10) (2.11) (2.12) D X D∗ 0 0 0 0 D X D∗ B1 B D   A11 A12 A13 A A23 0 A33 A11 A12 A13 A A23 0 A33 Y  , U  → bC =(cid:0)0 C C1(cid:1) . TbΣ of bΣ is of the form   TbΣ = 0   : (cid:0)0 C C1(cid:1) 0 ,  , bB = bA = The system Σ is called a restriction of bΣ. Recall that subspace N of inner product of H. Since X clearly is a regular subspace of bX , there exists the unique orthogonal projection PX from bX to X . Let bA↾X be the restriction of bA to the subspace X . Then, the system Σ can be represented as the Pontryagin space H is regular if it is itself a Pontryagin space with the inherited inner product of h·, ·iH. The subspace N is regular precicely when N ⊥ is regular, where ⊥ refers to orthogonality with respect to the indefinite Σ = (PX bA↾X , PXbB,bC↾X , D; PX bX , U , Y; κ). Dilations and restrictions are denoted by bΣ = dilX →bX Σ, Σ = resbX →XbΣ, mostly without subscripts when the corresponding state spaces are clear. A calculation show that the transfer functions of the original system and its dilation coincide. Moreover, if Σ is passive, then is any retriction of it. The following proposition shows that a passive system has a conservative dilation. For the Hilbert space case, this result is from [5], and for the Pontryagin state space case, see [28]. Proposition 2.3. Let Σ = (A, B, C, D; X , U , Y; κ) be a passive system. Then Proof. Since the system operator T of Σ is a contraction, by Theorem 2.1, there exists a Julia operator there exists a conservative dilation bΣ = (bA, bB,bC, D; bX , U , Y; κ) of Σ. (cid:18) T DT ∗ DT (cid:19) ⇐⇒ T −L∗(cid:19) :(cid:18)X ⊕ U DT ∗ (cid:19) →(cid:18)X ⊕ Y (cid:18)PX DT ∗ PY DT ∗(cid:19) −L∗  :(cid:18)X  →(cid:18)X U(cid:19)DT ∗  Y(cid:19)DT  (cid:18)A B C D(cid:19) (cid:0)D∗ T ↾U(cid:1) with properties introduced in Theorem 2.1. Denote EH(h) = hh, hiH for a vector h in an inner product space H. Then, for x ∈ X , u ∈ U and f ∈ DT ∗ , one has T ↾X D∗ (2.13) D∗ EX (x) + EU (u) + ED T ∗ (f ) = EX (Ax + Bu + PX DT ∗ f ) + EY (Cx + Du + PY DT ∗ f ) (2.14) + EDT (D∗ T ↾X x + D∗ T ↾U u − L∗f ). Minimal realizations 7 Also, for given x′ ∈ X , y ∈ Y and f ′ ∈ DT , there exists an unique triplet x ∈ X , u ∈ U and f ∈ DT ∗ such that A B C D T ↾X D∗ D∗ T ↾U  DT , ℓ2 +(DT ∗ ) := x u PX DT ∗ PY DT ∗ −L∗  f = ∞M1 bX := ℓ2 DT ∗ , x′ y f ′ (2.15) −(DT ) ⊕ X ⊕ ℓ2 +(DT ∗ ), Define ℓ2 −(DT ) := −1M−∞ ℓ2 where L denotes an orthogonal sum of Hilbert spaces. Since ℓ2 +(DT ∗ ) are Hilbert spaces, the space bX clearly is a Pontryagin space with the negative index κ. For u ∈ U and (. . . f−1, x, f1 . . .) ∈ bX , where the underlined element belongs to X , define the operators −(DT ) and T ↾X x − L∗f1, Ax + PX DT ∗ f1, f2, f3, . . .), T ↾U u, Bu, 0, . . .), (. . . , f−1, D∗ bA(. . . , f−2, f−1, x, f1, f2, f3, . . .) := bBu := (. . . , 0, D∗ bC(. . . , f−1, x, f1, . . .) := Cx + PY DT ∗f1. tion of Σ. Moreover, calculations by using the identities (2.14) and (2.15) A calculation shows that the system bΣ := (bA, bB,bC, D; bX , U , Y; κ) is a dila- show that the system operator TbΣ is unitary, and therefore the system bΣ is conservative. (cid:3) It is possible that D = {0} or D∗ = {0} in (2.9). In those cases, the zero space and the corresponding row and column will be left out in (2.10). In particular, if the system Σ with the system operator T as in (1.1) is isometric (co-isometric), then DT = 0 (DT ∗ = 0), and proceeding as in the proof of form (2.10) such that D = {0} (D∗ = {0}). Moreover, for an arbitrary pas- sive system as in (1.1), it is possible to construct an isometric (co-isometric) Proposition 2.3, it is possible to construct a conservative dilation bΣ of the dilation bΣ of the form (2.10) such that D∗ = {0} (D = {0}). There is also an another way to expand the system (1.1), and it is called an embedding. In this expansion, the state space and the main operator will not change. The embedding of the system (1.1) is any system determined by the system operator TeΣ = A eB eC eD! :(cid:18)X eU(cid:19) →(cid:18)X eY(cid:19) ⇐⇒ D21 D22(cid:19) : U ′(cid:19) → Y ′(cid:19) , (cid:0)B B1(cid:1)  X(cid:18) U X(cid:18) Y C1(cid:19) (cid:18) D D12 (cid:18) C A (2.16) 8 L. Lilleberg where U ′ and Y ′ are Hilbert spaces. The transfer function of the embedded system is C1(cid:19) (IX − zA)−1(cid:0)B B1(cid:1) θeΣ(z) =(cid:18) D D12 D21 D22(cid:19) + z(cid:18) C D21 + zC1(IX − zA)−1B D22 + zC1(IX − zA)−1B1(cid:19) =(cid:18) D + zC(IX − zA)−1B =(cid:18) θΣ(z) θ22(z)(cid:19) , D12 + zC(IX − zA)−1B1 θ12(z) θ21(z) (2.17) where θΣ is the transfer function of the original system. The embedded sys- tems will be needed in Section 4. It will be proved in Proposition 2.5 below that the transfer function of any passive system (1.1) is a generalized Schur function with index not larger than the negative index of the state space. For a special case where incoming and outcoming spaces are Hilbert spaces, this result is due to [28, Theorem 2.2]. The proof of the general case follows the lines of Saprikin's proof of the special case. Lemma 2.4. Let Σ = (A, B, C, D; X , U , Y; κ) be a passive system with the transfer function θ. Denote the system operator of Σ as T. If DT =(cid:18)DT,1 DT,2(cid:19) : DT →(cid:18)X U(cid:19) DT ∗ =(cid:18)DT ∗ DT ∗ ,2(cid:19) : DT ∗ →(cid:18)X Y(cid:19) , ,1 (2.18) are defect operators of T and T ∗, respectively, then the identities IY − θ(z)θ∗(w) = (1 − z ¯w)G(z)G∗(w) + ψ(z)ψ∗(w), IU − θ∗(w)θ(z) = (1 − z ¯w)F ∗(w)F (z) + ϕ∗(w)ϕ(z), (2.19) (2.20) with G(z) = C(IX − zA)−1, F (z) = (IX − zA)−1B, ψ(z) = DT ∗ ,2 + zC(IX − zA)−1DT ∗ 1 ϕ(z) = D∗ T,2 + zD∗ T,1(IX − zA)−1B, , (2.21) (2.22) hold for every z and w in a sufficiently small symmetric neighbourhood of the origin. Proof. By applying the results from [1, Theorem 1.2.4] and the identities in (2.7), one deduces that for every z and w in a sufficiently small symmetric Minimal realizations 9 neighbourhood of the origin, it holds IY (cid:19) IY(cid:1)(cid:18)(IX − ¯wA∗)−1C ∗ (cid:19) IY IY − θ(z)θ∗(w) =(cid:0)C(IX − zA)−1 −(cid:0)zC(IX − zA)−1 =(cid:0)C(IX − zA)−1 −(cid:0)zC(IX − zA)−1 −(cid:0)zC(IX − zA)−1 IY(cid:1) T T ∗(cid:18) ¯w(IX − ¯wA∗)−1C ∗ IY(cid:1)(cid:18)(IX − ¯wA∗)−1C ∗ IY(cid:1) (I − DT ∗ D∗ IY(cid:1) ¯w(cid:16)IX − DT ∗ = C(IX − zA)−1(IX − ¯wA∗)−1C ∗ + IY = (1 − z ¯w)C(IX − zA)−1(IX − ¯wA∗)−1C ∗ − ¯wDT ∗ (cid:19) D∗ T ∗ ,1 T ∗)(cid:18) ¯w(IX − ¯wA∗)−1C ∗ IY IY ,1 ,2 (cid:19) ,1(cid:17)(IX − ¯wA∗)−1C ∗ −DT ∗ D∗ T ∗ (IX − ¯wA∗)−1C ∗ + IY −DT ∗ ,1 ,2 ,2! T,2 D∗ D∗ T ∗ + z ¯wC(IX − zA)−1DT ∗ ,1 D∗ T ∗ ,1 (IX − ¯wA∗)−1C ∗ + zC(IX − zA)−1DT ∗ ,1 D∗ T,2 + ¯wDT ∗ ,2 D∗ T ∗ ,1 (IX − ¯wA∗)−1C ∗ + DT ∗ ,2 D∗ T ∗ ,2 = (1 − z ¯w)G(z)G∗(w) + ψ(z)ψ∗(w). Similar calculations show that (2.20) holds also, and the proof is complete. (cid:3) Note that if Σ in Lemma 2.4 is isometric (co-isometric), then DT = 0 (DT ∗ = 0) and therefore ϕ ≡ 0 (ψ ≡ 0). Proposition 2.5. If Σ = (A, B, C, D; X , U , Y; κ) is a passive system, the the transfer function θ of Σ belongs to Sκ′(U , Y), where κ′ ≤ κ. Proof. Denote the system operator of Σ as T. By Lemma 2.4, the kernel Kθ defined as in (1.3) has a representation Kθ(w, z) = G(z)G∗(w) + (1 − z ¯w)−1ψ(z)ψ∗(w), (2.23) where G(z) and ψ(z) are defined as in (2.21). Since the negative index of X is κ and the negative index of the Hilbert space DT ∗ is zero, it follows from [1, Lemma 1.1.1.], that for any finite set of points w1, . . . , wn in the domain of holomorphy of θ and the set of vectors {y1, . . . , yn} ⊂ Y, the Gram matrices (cid:0)hG∗(wj)yj, G∗(wi)yiiX(cid:1)n , i,j=1 (cid:16)hψ∗(wj)yj, ψ∗(wi)yiiD T ∗(cid:17)n , i,j=1 have, respectively, at most κ and zero negative eigenvalues. The kernel (1 − z ¯w)−1 has no negative square, since it is the reproducing kernel of the clas- sical Hardy space H 2(D). The Schur product theorem shows that the ker- nel (1 − z ¯w)−1ψ(z)ψ∗(w) has no negative square. Then it follows from [1, Theorem 1.5.5] that the kernel Kθ has at most κ negative square. That is, θ ∈ Sκ′(U , Y), where κ′ ≤ κ, and the proof is complete. (cid:3) 10 L. Lilleberg Definition 2.6. A passive realization Σ of a genaralized Schur function θ ∈ Sκ(U , Y) is called κ-admissible if the negative index of the state space of Σ coinsides with the negative index κ of θ. In what follows, this paper deals mostly with the κ-admissible realiza- tions. It will turn out that the κ-admissible realizations of θ ∈ Sκ(U , Y) are well behaved is some sense; they have many similar propeties than the standard passive Hilbert space systems. The following realizations theorem is well known, see [1, Theorems 2.2.1, 2.2.2 and 2.3.1]. Theorem 2.7. For a generalized Schur function θ ∈ Sκ(U , Y) there exist re- alizations Σk = (Tk; Xk, U , Y; κ), k = 1, 2, 3, of θ such that (i) Σ1 is observable co-isometric; (ii) Σ2 is controllable isometric; (iii) Σ3 is simple conservative. Conversely, if the system Σ has some of the properties (i) -- (iii), then θΣ ∈ Sκ(U , Y), where κ is the negative index of the state space of Σ. Recall that a Hilbert subspace of the Pontryagin space X is a regular subspace such that its negative index is zero. Conversely, anti-Hilbert sub- space is a regular subspace such that its positive index is zero. When U and Y happens to be Hilbet spaces, the transfer function θ of the passive system Σ = (TΣ; X , U , Y; κ) belongs to class Sκ(U , Y) (with κ = ind−X ) if and only if (X s)⊥ is a Hilbert subspace [27, Lemma 3.2]. In the case when U and Y are Pontryagin spaces with the same negative index, the transfer function θ of the isometric (co-isometric, conservative) system Σ = (TΣ; X , U , Y; κ) belongs to class Sκ(U , Y) if and only if (X c)⊥ ((X o)⊥,(X s)⊥) is a Hilbert subspace [1, Theorem 2.1.2]. For a passive system, one has the following result. Proposition 2.8. For a passive realization Σ = (A, B, C, D; X , U , Y; κ) of θ ∈ Sκ(U , Y), spaces X c, X o and X s are regular and their orthogonal com- plements are Hilbert subspaces. Proof. Let Ω be a symmetric neighbourhood of the origin such that (I−zA)−1 and (I − zA∗)−1 exist for every z ∈ Ω. Represent the kernel Kθ as in (2.23). Since Kθ has κ negative square, a similar argument used in the proof of 2.5 shows that the kernel K1(z, w) = G(z)G∗(w), where G(z) = C(I −zA)−1, has κ negative square. It follows now from [1, Lemma 1.1.1'] that span{ran (I − wA∗)−1C ∗, w ∈ Ω} contains a κ-dimensional maximal anti-Hilbert subspace Xκ. Then, Xκ ⊕ (Xκ)⊥ = X is a fundamental decomposition of X . Especially, (Xκ)⊥ is a Hilbert subspace of X . But (cid:0)span{ran (I − wA∗)−1C ∗, w ∈ Ω}(cid:1)⊥ = (X o)⊥ ⊂ (Xκ)⊥, which implies that (X o)⊥ is a Hilbert subspace, and therefore its orthocom- plement X o is regular. By duality argument, the space X c is a regular subspace and the space (X c)⊥ is a Hilbert subspace. It easily follows from (2.1) -- (2.3) that (X s)⊥ = Minimal realizations 11 (X c)⊥ ∩ (X o)⊥, and therefore (X s)⊥ is also a Hilbert subspace and X s is regular. (cid:3) It follows from the Proposition 2.8 above that the state space X of a κ-admissible realization Σ of θ ∈ Sκ(U , Y) can be decombosed to the control- lable, observable and simple parts. Using this fact, the lemma below, which will be used extensively, can be proved. Lemma 2.9. Let Σ = (A, B, C, D; X , U , Y; κ) be a passive system such that the spaces (X o)⊥, (X c)⊥ and (X s)⊥ are Hilbert subspaces of X . Then the system operator T of Σ has the following representations U D D U U 0 0 (cid:18)(X o)⊥ X o (cid:19) Bo(cid:19) 0 Ao(cid:19) (cid:18)B1 (cid:18)A1 A2   :  T =      (cid:0)0 Co(cid:1) (cid:18)(X c)⊥ Bc(cid:19) A4 Ac(cid:19) (cid:18) 0 X c (cid:19) (cid:18)A3   :  T =      (cid:0)C1 Cc(cid:1) (cid:18)(X s)⊥ Bs(cid:19) X s (cid:19) 0 As(cid:19) (cid:18) 0 (cid:18)A5   :  T =      (cid:0)0 Cs(cid:1)                             12 A′′ 13 A′′ A′′ 23 A′′ 0 33 (cid:0)0 C ′ C ′ 1(cid:1) 12 A′ 13 A′ A′ A′ 0 A′′ 11 A′′ 0 0 A′ 11 A′ 0 0 B′′ 1 B′′ 0 D B′ 1 B′ 0 D     (cid:0)0 C ′′ C ′′ 1(cid:1) D 23 33 : : T = T = Y Y Y (cid:18)(X o)⊥ X o (cid:19) →     (cid:18)(X c)⊥ X c (cid:19) →     (cid:18)(X s)⊥ X s (cid:19) →      →     →          (X o)⊥ PX o X c PX c X o (X c)⊥   U U X o ∩ (X c)⊥ X c ∩ (X o)⊥ X c ∩ (X o)⊥ (2.24) (2.25) (2.26) (2.27)         (2.28) (X o)⊥ PX o X c X o ∩ (X c)⊥ Y PX c X o (X c)⊥ Y The restrictions Σo = (Ao, Bo, Co, D; X o, U , Y; κ) Σc = (Ac, Bc, Cc, D; X c, U , Y; κ) Σs = (As, Bs, Cs, D; X s, U , Y; κ) Σ′ = (A′, B′, C ′, D; PX oX c, U , Y; κ) Σ′′ = (A′′, B′′, C ′′, D; PX cX o, U , Y; κ) (2.29) (2.30) (2.31) (2.32) (2.33) of Σ are passive, and Σo is observable, Σc is controllable, Σs is simple, and Σ′ and Σ′′ are minimal. For any n ∈ N0 and any z in a sufficiently small 12 L. Lilleberg symmetric neighbourhood of the origin, it holds AnB = An c Bc = An s Bs, (I − zA)−1B = (I − zAs)−1Bs = (I − zAc)−1Bc, A∗nC ∗ = A∗ o nC ∗ o = A∗ s (I − zA∗)−1C ∗ = (I − zA∗ s)−1C ∗ nC ∗ s , s = (I − zA∗ c)−1C ∗ c . (2.34) (2.35) (2.36) (2.37) Moreover, if Σ is co-isometric (isometric), then so are Σo and Σs (Σc and Σs). Proof. Since (X o)⊥, (X c)⊥ and (X s)⊥ are Hilbert spaces, the spaces X o, X c and X s are regular subspaces with the negative index κ. It follows from the identities (2.1) -- (2.3) that (X o)⊥, (X s)⊥ are A-invariant, (X c)⊥, (X s)⊥ are A∗-invariant, ran C ∗ ⊂ X o ⊂ X s, ran B ⊂ X c ⊂ X s,  , (2.38) and the representations (2.24) -- (2.26) follow. That is, Σo, Σc and Σs are re- strictions of the passive system Σ, ans therefore they are passive. Let TΣk be the system operator of Σk where k = o, c, s, and let x ∈ X k ⊕ U and x ∈ X k ⊕ Y. Calculation show that TΣk x = T x, T ∗ Σk x = T ∗x, k = c, s, k = o, s. It follows from the equations above that if Σ is co-isometric (isometric), then so are Σo and Σs (Σc and Σs). Suppose x ∈ X o such that CoAn o x = 0 for every n = 0, 1, 2, . . .. Then CAnx =(cid:0)0 Co(cid:1)(cid:18)A1 A2 0 Ao(cid:19)n(cid:18)0 x(cid:19) = C0An 0 x = 0, and the identity (2.2) implies that x ∈ X o ∩ (X o)⊥ = {0}. Thus x = 0, and it can be deduced that Σo is observable. Similar arguments show that Σc is controllable and Σs is simple, the details will be omitted. Let u ∈ U , and n ∈ N0. Then, by (2.25) and (2.26), 0 AnBu =(cid:18)A3 AnBu =(cid:18)A5 A4 Ac(cid:19)n(cid:18) 0 0 As(cid:19)n(cid:18) 0 0 An Bc(cid:19) =(cid:18) 0 Bs(cid:19) =(cid:18) 0 c Bcu(cid:19) = An s Bsu(cid:19) = An An c Bcu s Bsu, and (2.34) holds. By Neumann series, (I − zA)−1B = znAnB ∞Xn=0 Minimal realizations 13 holds for all z in a sufficiently small symmetric neighbourhood of the origin, and (2.35) follows now from (2.34). The equalities (2.36) and (2.37) can be deduced similarly. Since the orthocomplements (X o)⊥ and (X c)⊥ are Hilbert subspaces, it follows from [28, Lemma 3.1] that PX oX c and PX cX o are regular subspaces, and X o ∩ (PX oX c)⊥ = X o ∩ (X c)⊥, X c ∩ (PX cX o)⊥ = X c ∩ (X o)⊥. Since (X o)⊥ ⊂ (PX oX c)⊥, (X c)⊥ ⊂ (PX cX o)⊥ and all the spaces are regular, simple calculations show that (PX oX c)⊥ = (X o)⊥ ⊕ (X o ∩ (PX o X c)⊥), (PX cX o)⊥ = (X c)⊥ ⊕ (X c ∩ (PX cX o)⊥). Therefore, X = PX oX c ⊕ (PX oX c)⊥ = (X o)⊥ ⊕ PX oX c ⊕ (X o ∩ (PX oX c)⊥) = (X o)⊥ ⊕ PX oX c ⊕ (X o ∩ (X c)⊥), and similarly, X = (X c ∩ (X o)⊥) ⊕ PX cX o ⊕ (X c)⊥. Since (X o ∩ (X c)⊥ and X c ∩ (X o)⊥ are also Hilbert spaces, the spaces PX oX c and PX cX o are Pontryagin spaces with the negative index κ. By considering the properties in (2.38), the representations (2.27) and (2.28) follow now easily. That is, Σ′ and Σ′′ are restrictions of Σ, and therefore passive. Denote X ′ := PX oX c. Represent the system operator T of Σ as in (2.27). Then PX ′AnB = PX ′ A′ 11 A′ 0 0 12 A′ 13 A′ A′ 23 A′ 0 33  n B′ 1 B′ 0 = A′nB′ 0 0  = A′nB′, and similarly A′∗nC X ′c = span {ran A′nB′ : n = 0, 1, . . .} = span {ran PX ′AnB : n = 0, 1, . . .} ′∗ = PX ′A∗nC ∗. Therefore, = PX ′span {ran AnB : n = 0, 1, . . .} = PX ′X c = PX ′PX oX c = PX ′X ′ = X ′, and similarly X ′o = PX ′X o = X ′, which implies that Σ′ is minimal. A similar argument shows that Σ′′ is minimal, and the proof is complete. (cid:3) Note that in particular, Lemma 2.9 implies the existence of a minimal passive realization of θ ∈ Sκ(U , Y). Definition 2.10. The restrictions Σo, Σc, Σs, Σ′, and Σ′′ in Lemma 2.9 are called, respectively, the observable, the controllable, the simple (or proper), the first minimal and the second minimal restrictions of Σ. The first minimal and the second minimal restrictions will be considered later in Sections 3 and 4. 14 L. Lilleberg Two realizations Σ1 = (A1, B1, C1, D1; X1, U , Y; κ1), Σ2 = (A2, B2, C2, D2; X2, U , Y; κ2) of the same function θ ∈ Sκ(U , Y) are called unitarily similar if D1 = D2 and there exists a unitary operator U : X1 → X2 such that A1 = U −1A2U, B1 = U −1B2, C1 = C2U. (2.39) In that case, it easily follows that κ1 = κ2. Unitary similarity preserves dy- namical properties of the system and also the spectral properties of the main operator. If two realizations of θ ∈ Sκ(U , Y) both have the same property (i), (ii) or (iii) of Theorem 2.7, then they are unitarily similar [1, Theorem 2.1.3]. The realizations Σ1 and Σ2 above are said to be weakly similar if D1 = D2 and there exists an injective closed densely defined possible unbounded linear operator Z : X1 → X2 with the dense range such that ZA1x = A2Zx, C1x = C2Zx, x ∈ D(Z), and ZB1 = B2, (2.40) where D(Z) is the domain of Z. In Hilbert state space case, a result of Helton [25] and Arov [5] states that two minimal passive realizations of θ ∈ S(U , Y) are weakly similar. However, weak similarity preserves neither dynamical properties of the system nor the spectral properties of its main operator. Helton's and Arov's statement holds also in case where all the spaces are indefinite. This result is stated for reference purposes. Similar argument as Hilbert space case can be applied, definiteness of the inner product play no role. For a proof of special cases, see [14, Theorem 7.1.3], [29, p. 702] and [27, Theorem 2.5]. Note that the realizations are not assumed to be κ-admissible or passive. Proposition 2.11. Two minimal realizations of θ ∈ Sκ(U , Y) are weakly sim- ilar. 3. Optimal minimal systems For κ-admissible realizations of θ ∈ Sκ(U , Y), where U and Y are Pontryagin spaces with the same negative index, one can form the similar theory of optimal minimal passive systems as represented in the standard Hilbert space case in [7] and the Pontryagin state space case in [28]. Techniques, definitions and notations to be used here are similar to what appears in those papers. Denote EX (x) = hx, xiX for a vector x in an inner product space X . The same notation has been used in the proof of Proposition 2.3. Following [7, 12, 28], a passive realization Σ = (A, B, C, D; X , U , Y; κ) of θ ∈ Sκ(U , Y) is called optimal if for any passive realization Σ′ = (A′, B′, C ′, D′; X ′, U , Y; κ) of θ, the inequality n ∈ N0, uk ∈ U , (3.1) EX nXk=0 AkBuk! ≤ EX ′ nXk=0 A′kB′uk! , Minimal realizations 15 holds. On the other hand, the system Σ is called -optimal if it is observable and EX nXk=0 AkBuk! ≥ EX ′ nXk=0 A′kB′uk! , n ∈ N0, uk ∈ U , (3.2) holds for every observable passive realization Σ′ of θ. The requirement for observability must be included for avoiding trivialities, since otherwise every isometric realization of θ would be ∗-optimal; see Lemma 3.3 below and [7, Proposition 3.5 and example on page 144]. In the definition of optimality, the requirement that the considered re- alizations are κ-admissible is essential, as the example below shows. Example 3.1. Let Σ = (A, B, C, D; X , U , Y; κ), Σ′ = (A′, B′, C ′, D′; X ′, U , Y; κ′), where κ < κ′, be passive realization of θ ∈ Sκ(U , Y). Suppose that (3.1) holds. By Lemma 2.9, if (3.1) holds for Σ, it holds also for the controllable restriction Σc = (Ac, Bc, Cc, D′; X c, U , Y; κ) of Σ. For any vector x of the form x = MXn=0 define An c Bcun, {un} ⊂ U , M ∈ N0, Rx = A′nB′un. MXn=0 It is easy to deduce that R is a linear relation. Moreover, since Σc is control- lable by Lemma 2.9, R is densely defined. Since (3.1) holds, R is contractive. It follows now from [1, Theorem 1.4.2] that R can be extended to be every- where defined contractive linear operator. Since ind−X c = κ < κ′ = ind−X ′, it follows from [24, Theorem 2.4] that linear operator from X c to X ′ cannot be contractive, and hence (3.1) cannot hold. It will be shown in Theorem 3.5 below that an optimal (∗-optimal) min- imal realization exists, and it can be constructed by taking the first (second) minimal restriction, introduced in Definition 2.10, of simple conservative re- alizations. More lemmas will be needed before that. Lemma 3.2. Let Σ = (A, B, C, D; X , U , Y; κ) is a passive realization of θ ∈ Sκ(U , Y), and let Σs = (As, Bs, Cs, D; X s, U , Y; κ) be the restriction of Σ to the simple subspace. Then, the first (second) minimal restrictions of Σ and Σs coinside. Proof. Only the proof of the statement conserning about the second minimal restrictions is provided, since the other case is similar. To make notation less cumbersome, write X s = Xp, where p refers to proper part. By Lemma 2.9, 16 L. Lilleberg the equalities (2.34) and (2.36) hold, and it easily follows that X o = X o p , X c = X c p (X o)⊥ = (X s)⊥ ⊕ (X o p )⊥, (X c)⊥ = (X s)⊥ ⊕ (X c p )⊥, where orthogonal complements (X o the space Xp. Therefore, p )⊥ and (X c p )⊥ are taken with respect to PX cX o = PX c p X o p ⊂ X s = Xp, and consequently, PPX c p X o p Ap↾PX c p X o p = PPX c X oA↾X s↾PX c X o = PPX c X o A↾PX c X o, PPX c p X o p Bp = PPX c X oB, Cp↾PX c p X o p = C↾PX c X o , which shows that the second minimal restrictions of Σ and Σs co-inside. (cid:3) To prove the (∗-)optimality of a system, the following lemma is helpful. Lemma 3.3. Let Σ = (A, B, C, D; X , U , Y,κ), Σ′ = (A′, B′, C ′, D; X ′, U , Y; κ), bΣ = (bA, bB,bC, D; bX , U , Y, κ), be realizations of θ ∈ Sκ(U , Y) such that Σ is passive, bΣ is a passive dilation of Σ and Σ′ is the first minimal restriction of bΣ. Then EX ′ nXk=0 A′kB′uk! ≤ EX nXk=0 AkBuk! , n ∈ N0, uk ∈ U . (3.3) Moreover, for any isometric realization Σ1 = (A1, B1, C1, D; X1, U , Y, κ) of θ, it holds n ∈ N0, uk ∈ U . (3.4) 1 B1uk! , Note that Proposition 2.3 quarantees the existence of a passive dilation AkBuk! ≤ EX1 nXk=0 EX nXk=0 bΣ of Σ with the properties discribed above. Proof. SincebΣ is a dilation of Σ, the system operator TbΣ has a representation  bC D! =    , Y TbΣ = bA bB U  → A11 A12 A13 A A23 0 A33  : B1 B 0 D   D X D∗ D X D∗ (3.5) Ak 0 0 (cid:0)0 C1 C(cid:1) Minimal realizations 17 33 X1 X ′ X3 B′ 1 B′ D    X1 X ′ X3 Y can also be represented as A′ 11 A′ 0 0 12 A′ 13 A′ A′ 23 A′ 0   0  : X3 are Hilbert spaces, and X ′ is a Pontryagin space with the negative index κ. Let n ∈ N0 and {uk}n where D and D∗ are Hilbert spaces. On the other hand, by Lemma 2.9, bΣ TbΣ =  , where X1 = (bX o)⊥, X ′ = PbX o bX c and X3 = bX o ∩ (bX c)⊥. The spaces X1 and EX ′ nXk=0 U  → 1(cid:1) (cid:0)0 C ′ C ′ k=0 ⊂ U . Since X3 ⊂ (bX c)⊥, it holds nXk=0 bAkbBuk! = E bX nXk=0 bAkbBuk! − EbX PX1 nXk=0 bAkbBuk! − E bX PX3 = E bX nXk=0 bAkbBuk! − EbX PX1 nXk=0 bAkbBuk! nXk=0 bAkbBuk! . A′kB′uk! = E bX PX ′ With D and D∗ as in (3.5), the identities in (2.9) hold. Therefore, it follows (3.6) A similar calculation as above yields then from the identities (2.1) and (2.2) that D∗ ⊂ (bX c)⊥ and D ⊂ (bX o)⊥ = X1. EX nXk=0 nXk=0 bAkbBuk! . Since X1 is a Hilbert space, the inclusion D ⊂ X1 implies (3.7) AkBuk! = E bX nXk=0 bAkbBuk! − EbX PD EbX PD EX ′ nXk=0 nXk=0 bAkbBuk! ≤ E bX PX1 A′kB′uk! ≤ EX nXk=0 nXk=0 bAkbBuk! . AkBuk! , It follows now from the equations (3.6) and (3.7) that and the inequality (3.3) is proved. (3.7) that Assume that bΣ is isometric. Since D is a Hilbert space, it follows from EX nXk=0 AkBuk! ≤ EbX nXk=0 bAkbBuk! . (3.8) 18 L. Lilleberg 0 U Y c Bc. Therefore By Lemma 2.9, the system operator of bΣ can be represented as (bX c)⊥ (bX c)⊥  → bX c ! D  : bX c !  , Bc(cid:19) A4 Ac(cid:19) (cid:18) 0 TbΣ =(cid:18)A3 (cid:0)C1 Cc(cid:1) where the restriction Σc = (Ac, Bc, Cc, D; bX c, U , Y, κ) is controllable isomet- ric, and for every n = 0, 1, 2, . . . , it holds bAnbB = An kBcuk! . EbX nXk=0 bAkbBuk! = E bX c nXk=0 1 nXk=0 kB1uk! = EX c EX1 nXk=0 kB0uk! = EX c 1 nXk=0 E bX c nXk=0 Similar argument show that if Σc 1 , U , Y, κ) is the restric- tion of the isometric system Σ1 = (A1, B1, C1, D; X1, U , Y, κ) to the control- lable subspace X c 1uk! . 1uk! . 1 is controllable isometric and it holds 1 are unitarily similar, and therefore 1 = (Ac 1, Bc 1, C c 1, D; X c But Σ0 and Σc A1 A0 1 , then Σc Ac (3.9) Ac 1 kBc (3.10) Ac 1 kBc (3.11) (cid:3) By combining (3.8) -- (3.11), the inequality (3.4) follows. Remark 3.4. It follows from the inequality (3.4) of Lemma 3.3 that if there exists an observable isometric realization of θ ∈ Sκ(U , Y), then it is ∗-optimal. In the standard Hilbert space case, results of Arov [6] show that there exist optimal minimal realizations of a Schur function. The construction was based on the existence of the defect functions, see Section 4. Arov et. all provided new geometric proofs of these results in [7]. Saprikin used those new proofs and generalized Arov's results to Pontryagin state space case in [28]. It will be proved next that Arov's results holds in the case when all spaces are Pontryagin spaces. The geometric proofs in [7] can still be applied in the precent setting with few appropriate changes. Theorem 3.5. Let θ ∈ Sκ(U , Y), where U and Y are Pontryagin spaces with the same negative index. Then: (i) The first minimal restriction of a simple conservative realization of θ is optimal minimal; (ii) The minimal passive system Σ∗ is optimal if and only if the dual system Σ is -optimal minimal; (iii) The second minimal restriction of a simple conservative realization of θ is -optimal minimal; (iv) Optimal (-optimal) minimal systems are unique up to unitary similar- ity, and every optimal (*-optimal) minimal realization of θ is the first minimal restriction (second minimal restriction) of some simple conser- vative realization of θ. Minimal realizations 19 Proof. (i) Let Σ′ = (A′, B′, C ′, D; X ′, U , Y; κ) be the first minimal restriction n ∈ N0, uk ∈ U . (3.12) AkBuk! , By Lemma 3.2, it can be assumed that Σ is the first minimal restriction of A′kB′uk! ≤ EX nXk=0 Sκ(U , Y). Let Σ = (A, B, C, D; X , U , Y; κ) be the first minimal restriction of some conservative realization of θ such that its state space has negative index κ. To prove that Σ′ is optimal, Lemma 3.3 shows that it is enough to prove of a simple conservative realization bΣ′ = (bA′, bB′,bC ′, D; bX ′, U , Y; κ) of θ ∈ EX ′ nXk=0 some simple conservative realization bΣ = (bA, bB, bC, D; bX , U , Y; κ) of θ. Since bΣ and bΣ′ are both simple conservative, they are unitarily similar, so there exists a unitary operator U : bX → bX ′ such that bB = U −1bB′, bA = U −1bA′U, bC = bC ′U. ′o = U bX o, bX ′o)⊥ = U (bX o)⊥, ′c = U bX c, (bX Easy calculations shows that bX ′c)⊥ = U (bX c)⊥ and P bX ′c = U PbX o bX c. In particular, (bX ′o bX PX = PPcX o bX c = U −1PP ′ o bX cX ′c U = U −1PX ′U, which implies A = PX bA↾X = U −1PX ′bA′U ↾X = (U ↾X )−1PX ′bA′↾X ′U ↾X = (U ↾X )−1A′U ↾X B = (U ↾X )−1B′, C = C ′U ↾X . It follows that Σ and Σ′ are unitarily similar and the corresponding unitary operator is U0 = U ↾X . Then EX nXk=0 AkBuk! = EX U −1 0 A′kB′uk! = EX ′ nXk=0 nXk=0 A′kB′uk! , Then (3.12) holds, and Σ′ is an optimal minimal system. (ii) Let Σ∗ = (A∗, C ∗, B∗, D∗; X , Y, U; κ) be an optimal minimal passive realization of θ# ∈ Sκ(Y, U). Then Σ = (A, B, C, D; X , U , Y; κ) is a minimal passive realization of θ ∈ Sκ(U , Y). Consider an arbitrary observable passive ′∗ = realization Σ′ = (A′, B′, C ′, D; X ′, U , Y; κ) of θ ∈ Sκ(U , Y). Then Σ ′∗, D∗; X ′, Y, U; κ) is a controllable passive realization of θ#. For (A a vector of the form ′∗, B ′∗, C x′ = nXk=0 ′∗k A ′∗yk, C n ∈ N0, yk ∈ Y, define Sx′ = nXk=0 (A∗)kC ∗yk. 20 L. Lilleberg ′∗ is controllable and Σ∗ is optimal, the domain of S is dense, and it Since Σ holds EX (Sx) = EX nXk=0 (A∗)kC ∗yk! ≤ EX ′ nXk=0 ∗k ′ A ′ C ∗yk! = EX ′(x). That is, S is a contractive linear relation with the dense domain. Then [1, Theorem 1.4.4] shows that the closure of S, which is still denoted as S, is contractive everywhere defined linear operator from X ′ → X . Since X ′ and X are Pontryagin spaces with the same negative index, S ∗ : X → X ′, is contractive as well. The transfer functions of the Σ and Σ′ coincide, and ′∗ = therefore CAmB = C ′A′kB′ for every m ∈ N0. By definition, S(A (A∗)mC ∗, or what is the same thing, C ′A′mS ∗ = CAm, for every m ∈ N0. Then also ′∗)mC C ′A′m+k B′ = CAmAkB = C ′A′m S ∗AkB for m, k ≥ 0. This implies A′kB′ = S ∗AkB since the system Σ′ is observable. It follows now that S ∗ nXk=0 AkBuk! = B′uk! = EX ′ S ∗ nXk=0 A′k B′uk. nXk=0 AkBuk!! ≤ EX nXk=0 AkBuk! , Therefore, EX ′ nXk=0 A′k since S ∗ is contractive. This proves that Σ is ∗-optimal. Suppose then that Σ = (A, B, C, D; X , U , Y; κ) is minimal passive ∗- optimal realization of θ ∈ Sκ(U , Y). Then Σ∗ is a minimal passive realiza- tion of θ# ∈ Sκ(Y, U). To prove the optimality of Σ∗, it suffices to con- ′∗ = sider all the minimal passive realizations of θ#; see Lemma 3.3. Let Σ ′∗, D∗; X ′, Y, U; κ) be a minimal passive realization of θ#. Then (A Σ′ is a minimal passive realization of θ. Since Σ is ∗-optimal, the inequality ′∗, B ′∗, C EX nXk=0 AkBuk! ≥ EX ′ nXk=0 A′kB′uk! , k=0 A′kB′uk for x =Pn holds. Define Kx =Pn k=0 AkBuk. Using similar tech- niques as above, K can be extended to be a contractive operator from X → X ′ such that n ∈ N0, uk ∈ U , K ∗(A ′∗)kC ′∗ = (A∗)kC ∗. Since K ∗ is contractive, EX nXk=0 A∗kC ∗yk! = EX K ∗ ′∗k A C nXk=0 ′∗yk! ≤ EX ′ nXk=0 ′∗k A C ′∗yk! , for {yk} ⊂ Y. This shows that Σ∗ is optimal. (iii) Let Σ be a simple conservative realization of θ ∈ Sκ(U , Y). Then Σ∗ is a simple conservative realization of θ#, and the first minimal restriction Minimal realizations 21 Σ∗′ of Σ∗ is optimal minimal by the part (i). By using the representations (2.27) and (2.28) from Lemma 2.9, it is easy to deduce that the dual system of Σ∗′ is the second minimal restriction Σ′′ of Σ, and it follows from the part (ii) that Σ′′ is ∗-optimal. (iv) Only the proofs of the claims considering optimal minimal realiza- tions will be given, since the claims considering ∗-optimal minimal realizations can be proved analogously. Let Σj = (Aj , Bj, Cj, D; Xj, U , Y; κ), j = 1, 2, be optimal minimal realizations of θ ∈ Sκ(U , Y). In a sufficiently small neigh- bourhood of the origin, the transfer functions θΣ1 and θΣ2 of the systems Σ1 and Σ2 have the Neumann series and they coincide, so C1Ak 2 B2 for k = 0, 1, 2, . . . Since Σ1 is controllable, vectors of the form 1 B1 = C2Ak x = NXk=0 Ak 1 B1uk, uk ∈ U , are dense in X1. Define U x = Ak 2 B2uk. (3.13) NXk=0 Because Σ2 is controllable as well, vectors of the form U x are dense in X2. Since Σ1 and Σ2 both are optimal realizations, EX1(x) = EX1(U x), and therefore U is an isometric linear relation with the dense domain and the dense range. It follows now from [1, 1.4.2] that the closure of U is a unitary operator, which is still denoted as U. Then, trivially B1 = U −1B2. For vector x in (3.13), it holds U A1x = U Ak+1 1 B1uk = Ak+1 2 B2uk = A2U x. NXk=0 NXk=0 It follows that U A1x = A2U x holds in a dense set, and therefore by conti- nuity, everywhere. Thus A1 = U −1A2U. Moreover, for k = 0, 1, 2, . . . , one concludes C1Ak 1 B1 = C2Ak 2 B2 = C2U Ak 1B1. Since spank∈N0 Ak 1 B1 is dense in X1, it must be C1 = C2U. It has been shown that the unitary operator U has all the properties of (2.39), and therefore Σ1 and Σ2 are unitarily similar. Suppose then that Σ = (A, B, C, D; X , U , Y; κ) is an optimal minimal be represented as realization of θ. Let bΣ0 = (bA0, bB0,bC0, D; bX0, U , Y; κ) be some simple conser- vative realization of θ. Lemma 2.9 shows that the system operator of bΣ can  , 0  :  →   = A′ 11 A′ 0 0 12 A′ 13 A′ A′ 23 A′ 0 X1 X ′ X2 Y  X1 X ′ X2 U  B′ 1 B′ D  TbΣ0 33 (cid:0)0 C ′ C ′ 1(cid:1) 22 L. Lilleberg A′ 13 U −1A′   A′ 11 A′ 12U 0 A 0 0 follows from part (i) that Σ′ is optimal minimal, and moreover, as proved above, unitarily similar with Σ. Therefore, there exists a unitary operator U : X → X ′ such that A = U −1A′U, B = U −1B′ and C = C ′U. Define where X1 = (bX o)⊥, X ′ = PbX o bX c and X2 = bX o ∩ (bX c)⊥. Then the system Σ′ = (A′, B′, C ′, D; X ′, U , Y; κ) is the first minimal restriction of bΣ, and it TbΣ = bA bB bC D!  =  , Y   : and let bΣ be the system corresponding the system operator TbΣ. Easy calcu- lations show that bΣ and bΣ0 are unitarily similar and  → I : is the corresponding unitary operator. Therefore bΣ is a simple conservative system. Now bU maps PX oX c to PX ′o X ′c, and bU X ′ = U X ′ = X . It follows that Σ is the first minimal restriction of bΣ. (cid:0)0 C C ′ 1(cid:1) bU =   → 0 I 0 U 0 0 B′ 1 B 0 D X1 X X2 U  23 A′ 33 X1 X X2 0 0 X1 X X2 X1 X ′ X2 (cid:3) 4. Generalized defect functions If U and Y are Hilbert spaces, it is well known that S ∈ S(U , Y) is holomor- phic in the unit disk and it has non-tangential contractive strong limit values almost everywhere (a.e.) on the unit circle T. Therefore, S can be extended to L∞(U , Y) function, that is, the class of weakly measurable a.e. defined and essentially bounded L(U , Y)-valued functions on T. Then it follows from [30, Theorem V.4.2] that there exist a Hilbert space K and an outer function ϕS ∈ S(U , K) such that ϕ∗ S(ζ)ϕS (ζ) ≤ I − S ∗(ζ)S(ζ) (4.1) same property, then a.e. on T, and if a function bϕ ∈ S(U ,bK), where bK is a Hilbert space, has this (4.2) S(ζ)ϕS (ζ) a.e. on T. The function ϕS is called the right defect function of S. For the notions of the outer functions, ∗-outer functions, inner functions and ∗-inner functions, see [30, Chapter V]. From [30, Theorem V.4.2] it is also easy to deduce that there exists a Hilbert space H and a ∗-outer function ψS ∈ S(H, Y) such that bϕ∗(ζ)bϕ(ζ) ≤ ϕ∗ ψS(ζ)ψ∗ S(ζ) ≤ I − S(ζ)S ∗(ζ) (4.3) Minimal realizations 23 a.e. ζ ∈ T and if a Schur function bψ ∈ S(bH, Y) has this same property, then ψS(ζ)ψ∗ (4.4) The function ψS is called the left defect function of S. Both ϕS and ψS are unique up to a unitary constant. S(ζ) ≤ bψ(ζ)bψ∗(ζ). The theory of the defect functions is considered, for instance, in [16, 17, 18]. Various connections of defect functions and passive realizations can be found in [4, 9, 10]. The definition of the defect functions was generalized for functions S ∈ Sκ(U , Y) in [27] by using the Kreın-Langer factorizations and the fact that all functions in Sκ(U , Y) have also contractive strong limit values a.e. on T. If U and Y are Pontryagin spaces such that their negative index is not zero, the defect functions cannot be defined similarly as in the Hilbert space settings, since S ∈ Sκ(U , Y) may not be extendable to the unit circle. In the Hilbert state space case, Arov and Saprikin showed in [12] that for a function S = SrB−1 is the right Kreın-Langer factorization of S, the existence of the optimal minimal realization of S is connected with the existence of the right defect function of Sr. In general, similar connections exist with certain functions constructed by embedded systems, and those function are called defect functions; this is the approach taken here. r ∈ Sκ(U , Y), where SrB−1 r Suppose that Σ = (A, B, C, D; X , U , Y; κ) is a passive realization of θ ∈ Sκ(U , Y). Denote the system operator of Σ by T. Theorem 2.1 shows that there exists a Julia operator of T. By using the same notation as in (2.13), A one can form the Julia embeddingeΣ of the system Σ. Then the corresponding system operator eT is a Julia operator of T, and it is of the form (cid:0)B DT ∗ ,1(cid:1) TeΣ = T,2 −L∗(cid:19) : DT(cid:19) , X(cid:18) Y (cid:18) C T,1(cid:19) (cid:18) D DT ∗ DT ∗ =(cid:18)DT ∗ ,2(cid:19), DT =(cid:18)DT,1 DT,2(cid:19), DT ∗ D∗ DT ∗(cid:19) → X(cid:18) U T ∗ = IX −T T ∗, DT D∗ T = IX −T ∗T, where (4.5) DT ∗ D∗ D∗ ,2 ,1 such that DT and DT ∗ have zero kernels. The transfer function of the Julia embedding is θeΣ(z) =(cid:18) D + zC(I − zA)−1B T,2 + zD∗ D∗ T,1(I − zA)−1B −L∗ + zD∗ DT ∗ ,2 + zC(I − zA)−1DT ∗ T,1(I − zA)−1DT ∗ ,1 ,1(cid:19) (4.6) (4.7) (4.8) =(cid:18)θ(z) ψ(z) ϕ(z) χ(z)(cid:19) . Lemma 2.4 shows that the identities I − θ(z)θ∗(w) = (1 − z ¯w)G(z)G∗(w) + ψ(z)ψ∗(w), I − θ∗(w)θ(z) = (1 − z ¯w)F ∗(w)F (z) + ϕ∗(w)ϕ(z) where G(z) = C(IX − zA)−1, F (z) = (IX − zA)−1B, 24 L. Lilleberg holds for every z and w in a sufficiently small symmetric neighbourhood of the origin. If the negative index of U and Y is denoted as κ1, a similar argument as in the proof of Proposition 2.5 shows that the kernels Kϕ(w, z) and Kψ(w, z) defined as in (1.3) have at most κ + κ1 negative squares. Definition 4.1. Let U and Y be Pontryagin spaces with the same negative index. Let Σ = (A, B, C, D; X , U , Y; κ) be an optimal minimal passive real- as in (4.5). Then the function ϕ in (4.6) is defined to be the right defect function ϕθ of θ. ization of θ ∈ Sκ(U , Y), and let eΣ be the Julia embedding of it, represented sive realization of θ ∈ Sκ(U , Y), and let eΣ be the Julia embedding of it, represented as in (4.5). Then the function ψ in (4.6) is defined to be the left defect function ψθ of θ. Moreover, let Σ = (A, B, C, D; X , U , Y; κ) be a ∗-optimal minimal pas- Remark 4.2. Since optimal (∗-optimal) minimal realizations are unitarily sim- ilar by Theorem 3.5, and Julia operators for contractive operator are essen- tially unique by Theorem 2.1, it can be deduced that the defect functions are essentially uniquely defined by θ ∈ Sκ(U , Y). The definition above is also slightly different from the one given in [27] for functions in the class Sκ(U , Y), where U and Y are Hilbert spaces. The right defect function of θ ∈ Sκ(U , Y) and the left defect function of θ# are closely related to each other. Lemma 4.3. For θ ∈ Sκ(U , Y), it holds ϕ# θ = ψθ# and ψ# θ = ϕθ# Proof. Let Σ = (A, B, C, D; X , U , Y; κ) be an optimal (∗-optimal) minimal realization of θ. Denote the system operator of Σ as T, and the Julia operator TeΣ of T as in (4.5). By Theorem 3.5, the system Σ∗ is ∗-optimal (optimal) minimal, and a calculation shows that T ∗ eΣ the results follow means of (4.6). is the Julia operator of T ∗. Now (cid:3) In the Hilbert space settings, S ∈ S(U , Y) has factorizations of the form S = SiSo = S∗oS∗i, where Si ∈ S(Y, Y) is inner, So ∈ S(U , Y) is outer, S∗o ∈ S(U , Y) is ∗-outer and S∗i ∈ S(Y, Y) is ∗-inner [30, p. 204]. The next proposition shows that for an ordinary Schur function θ ∈ S(U , Y), the outer factor of ϕθ and the ∗- outer factor of ψθ defined above coincide essentially with the usual definition of defect functions. Proposition 4.4. Let θ ∈ Sκ(U , Y), where U and Y are Hilbert spaces. Then ϕ∗ θ(ζ)ϕθ(ζ) ≤ I − θ∗(ζ)θ(ζ) (4.9) Hilbert space and κ′ does not depend on κ, has this same property, then a.e. on T, and if a generalized Schur function bϕ ∈ Sκ′ (U ,bK), where bK is a (4.10) θ(ζ)ϕθ(ζ), bϕ∗(ζ)bϕ(ζ) ≤ ϕ∗ Minimal realizations 25 a.e. on T. If κ = 0, denote the inner and outer factors of ϕθ as ϕθi and ϕθo , respectively. Then, ϕθi is an isometric constant, and if ϕ′ is an outer function with properties (4.1) and (4.2), then it holds U ϕθo = ϕ′, where U is a unitary operator. Moreover, ψθ(ζ)ψ∗ θ (ζ) ≤ I − θ(ζ)θ∗(ζ) a.e. ζ ∈ T and if a generalized Schur function bψ ∈ Sκ′ (bH, Y), where bK is a Hilbert space and κ′ does not depend on κ, has this same property, then ψθ(ζ)ψ∗ θ (ζ) ≤ bψ(ζ)bψ∗(ζ) a.e. ζ ∈ T.If κ = 0, denote the ∗-inner and ∗-outer factors of ψθ as ψθ∗i and ψθ∗o , respectively. Then, ψθ∗i is a co-isometric constant, and if ψ′ is a ∗-outer function with properties (4.3) and (4.4), then it holds ψθ∗ o U ′ = ψ′, where U ′ is a unitary operator. Proof. Let (4.11) be an optimal minimal realization of θ. Denote the system operator of Σ as Σ = (A, B, C, D; X , U , Y; κ) A D∗ B D∗ T, the Julia operator TeΣ of T as in (4.5) and the function ϕ = ϕθ as in (4.6). Since TeΣ is unitary, the operator TΣ′ = (cid:18) C T,1(cid:19) (cid:18) D Σ′ =(cid:18)A, B,(cid:18) C T,2(cid:19) :(cid:18)X DT(cid:19) . U(cid:19) → X(cid:18) Y T,1(cid:19) ,(cid:18) D DT(cid:19) ; κ(cid:19) T,2(cid:19) ; X , U ,(cid:18) Y must be isometric, and therefore the system (4.12) D∗ D∗ for a.e. ζ ∈ T. θ(ζ)(cid:1)(cid:18) θ(ζ) ϕθ(cid:19) . Since Σ′ is an embedding is an isometric realization of the function (cid:18) θ of the minimal system Σ, the system Σ′ is also minimal. It follows from Theorem 2.7 that (cid:18) θ ϕθ(cid:19) ∈ Sκ (U , Y ⊕ DT ) . Since contractive boundary values of generalized Schur functions exist for a.e. ζ ∈ T, it holds ϕθ(ζ)(cid:19) ≤ I ⇐⇒ ϕ∗ (cid:0)θ∗(ζ) ϕ∗ Suppose that a function bϕ ∈ Sκ′(U ,bK), where bK is a Hilbert space, has for a.e. ζ ∈ T. Since the function bϕ has the left Kreın -- Langer factorization of the form bϕ = B−1 bϕ bϕl, where bϕl is an ordinary Schur function, it holds θ(ζ)ϕθ(ζ) ≤ I − θ∗(ζ)θ(ζ) the property for a.e. ζ ∈ T. Then the function (4.13) bϕ∗(ζ)bϕ(ζ) ≤ I − θ∗(ζ)θ(ζ) bϕ∗(ζ)bϕ(ζ) = bϕ∗ l (ζ)bϕl(ζ) bϕl(cid:19) , θ =(cid:18) θ 26 L. Lilleberg belongs to the Schur class Sκ(cid:16)U , Y ⊕ bK(cid:17) , and it has a controllable isometric realization Σ with the system operator That is, A1 B1 D2(cid:19) :(cid:18)X1 T Σ = (cid:18)C1 C2(cid:19) (cid:18)D1 D2(cid:19) + z(cid:18)C1 θ(z) =(cid:18) θ(z) bϕl(z)(cid:19) =(cid:18)D1 =(cid:18)D1 + zC1(I − zA1)−1B1 D2 + zC2(I − zA1)−1B1(cid:19) . U(cid:19) → bK(cid:19) . X1(cid:18)Y C2(cid:19) (I − zA1)−1B1 It follows that Σ1 = (A1, B1, C1, D1; X1, U , Y; κ) (4.14) is a realization of θ, and since Σ is isometric and bK is a Hilbert space, the system Σ1 is passive. Since T Σ is isometric, the defect operator DT Σ of T Σ is zero, and it follows from Lemma 2.4 that 1)−1)(I − zA1)−1B1 (4.15) whenever the expressions are meaningful. By combining the identities (4.8) and (4.15) for optimal minimal realization Σ, one gets 1 (I − zA∗ l (z)bϕl(z) I − θ∗(z)θ(z) = I − θ∗(z)θ(z) −bϕ∗ (cid:0)1 − z2(cid:1)B∗ =(cid:0)1 − z2(cid:1) B∗ 1)−1(I − zA1)−1B1 +bϕ∗ l (z)bϕl(z) =(cid:0)1 − z2(cid:1) B∗(I − zA∗)−1(I − zA)−1B + ϕθ 1 (I − zA∗ for every z in a sufficiently small symmetric neighbourhood Ω of the origin. Since the system Σ is optimal, if follows by using Neumann series that (cid:10)B∗(I − zA∗)−1(I − zA)−1Bu, u(cid:11) = EX(cid:0)(I − zA)−1Bu(cid:1) = EX ∞Xn=0 1 B1uzn! 1)−1(I − zA1)−1B1u, u(cid:11) for every z ∈ Ω and for every u ∈ U . Then it follows from (4.16) that 1 (I − zA∗ An AnBuzn! ∗(z)ϕθ(z) (4.16) ≤ EX1 ∞Xn=0 =(cid:10)B∗ bϕ∗ l (z)bϕl(z) ≤ ϕθ l (ζ)bϕl(ζ) = bϕ∗(ζ)bϕ(ζ) ≤ ϕθ bϕ∗ ∗(z)ϕθ(z), z ∈ Ω. ∗(ζ)ϕθ(ζ) (4.17) By continuity, for a.e. ζ ∈ T. Next suppose that κ = 0. By combining (4.2) and (4.17), it can be deduced that ′∗(ζ)ϕ′(ζ) = ϕθ ϕ ∗(ζ)ϕθ(ζ) = ϕθo ∗(ζ)ϕθi ∗(ζ)ϕθi (ζ)ϕθo (ζ) = ϕθo ∗(ζ)ϕθo (ζ) Minimal realizations 27 for a.e. ζ ∈ T. Then it follows from [30, Proposition V.4.1] that ϕ′ = U ϕθo , where U is a unitary operator. If one puts an outer function bϕl = ϕθo = U −1ϕ′ in (4.13), the construction of an optimal minimal system used in the proof of [6, Theorem 7] shows that the associated system Σ1 in (4.14) is optimal. Since Σ in (4.11) is also optimal, for every z ∈ D, it holds B∗(I − zA∗)−1(I − zA)−1B = B∗ 1 (I − zA∗ 1)−1(I − zA1)−1B1, and then by (4.16) kϕθi(z)ϕθo(z)uk = kϕθo(z)uk for every z ∈ D and every u ∈ U . The outer function ϕθo (z) has a dense range for every z ∈ D [30, Proposition V.2.4]. This implies that ϕθi (z) is an isometry for every z ∈ D, and arguing as in the proof of [30, Proposition V.2.1] one deduces that ϕθi is an isometric constant. The claims involving ϕθ are proved. The claims involving ψθ follow now directly by applying Lemma 4.3. (cid:3) Lemma 4.5. Let Σ0 = (A0, B0, C0, D; X0, U , Y; κ), Σ = (A, B, C, D; X , U , Y; κ) be passive realizations of θ ∈ Sκ(U , Y) such that Σ0 is optimal. If for every z and w in a sufficiently small symmetric neighbourhood Ω of the origin the equality B∗(I − wA∗)−1(I − zA)−1B = B∗ 0 (I − wA∗ 0)−1(I − zA0)−1B0 (4.18) holds, then Σ is optimal. Proof. It follows from Lemma 2.9 that the system operator TΣ of Σ can be represented as in (2.25), the restriction Σc = (Ac, Bc, Cc, D; X c, U , Y; κ) of Σ to the controllable subspace X c is controllable passive, and (2.34) and (2.35) hold. Let x = MXj=1 Aj cBcuj, M ∈ N, {uj}M j=1 ⊂ U . (4.19) For the vectors of the form (4.19), define Rx = MXj=1 Aj 0B0uj. (4.20) Since Σc is controllable, the domain of R is dense. Moreover, Σ0 is optimal, and therefore EX0 (Rx) ≤ EX c (x) . That is, R is contractive, and it follows from [1, Theorem 1.4.2] that the closure of R is everywhere defined contractive linear operator. It is still denoted by R. Since (I − zAc)−1Bc = ∞Xn=0 znAn c Bc, (I − zA0)−1B0 = znAn 0 B0, ∞Xn=0 28 L. Lilleberg holds for every z in a sufficiently small symmetric neighbourhood Ω of the origin, it follows by continuity that for every z ∈ Ω and u ∈ U . Then for all M ∈ N, {zj}M imply now (IX0 − zjA0)−1B0uj, MXj=1 j=1 ⊂ U . Equalities (2.35) and (4.18) j=1 ⊂ Ω, and {uj}M R(cid:0)(I − zAc)−1Bcu(cid:1) = (I − zA0)−1B0u (I − zjAc)−1Bcuj = R MXj=1 (I − zjAc)−1Bcuj EX c MXj=1 MXj=1 MXk=1(cid:10)B∗ MXk=1(cid:10)B∗ MXj=1 (I − zjA0)−1B0uj = EX0 MXj=1 (I − zjAc)−1Bcuj . = EX0R MXj=1 0 (I − zkA∗ c (I − zkA∗ = = c )−1(I − zjAc)−1Bcuj, uk(cid:11)U 0)−1(I − zjA0)−1B0uj, uk(cid:11)U . This implies that R is isometric in span{ran (I − zA1)−1B1, z ∈ Ω}, which is a dense set, since Σ1 is controllable. Since R is bounded, it is now isometric everywhere, and identity (4.20) implies that Σc is optimal. Then it follows from (2.34) that Σ is optimal, and the proof is complete. (cid:3) The main results of [4, Theorem 1.1] were generalized to the Pontryagin state space setting in [27, Theorem 4.4]. By using Definition 4.1, it can be shown that parts of this result, as well as [10, Theorem 1], hold also in the case when all the spaces are indefinite. Moreover, certain parts of [4, Theorem 1.1], [10, Theorem 1] and [27, Theorem 4.4] can be improved. Before stating these results, some lemmas are needed. Lemma 4.6. Let θ ∈ Sκ(U , Y). Then the following statements are equivalent: (i) all κ-admissible minimal passive realizations of θ are unitarily similar; (ii) there exists a minimal passive realization of θ such that it is both optimal and ∗-optimal; (iii) all κ-admissible minimal passive realizations of θ are both optimal and ∗-optimal. Minimal realizations 29 Proof. (i) ⇒ (iii). Suppose (i). Let Σ1 = (A1, B1, C1, D; X1, U , Y; κ), Σ2 = (A2, B2, C2, D; X2, U , Y; κ) be, respectively, minimal passive and optimal (∗-optimal) minimal passive realizations of θ. Let U be the unitary operator from X1 to X2 with the properties described in (2.39). An easy calculation shows that EX2 nXk=0 2 B2uk! = EX1 U Ak Ak 1 B1uk! = EX1 nXk=0 nXk=0 1 B1uk! Ak for every u ∈ U and for every n = 0, 1, 2, . . . which implies that Σ1 is actually optimal (∗-optimal), and therefore (iii) holds. (iii) ⇒ (ii). The claim (iii) trivially implies (ii). (ii) ⇒ (i). Suppose (ii). Let the systems Σ1 = (A1, B1, C1, D; X1, U , Y; κ), Σ2 = (A2, B2, C2, D; X2, U , Y; κ) be minimal passive realizations of θ such that Σ1 is optimal and ∗-optimal. Let Z be the weak similarity mapping from X1 to X2 with the properties described 1 B1uk 2 B2uk. Recall also here the construction of Z in the proof of [27, Theorem 2.5]. Since Σ1 is both optimal and ∗-optimal, in (2.40). It follows from (2.40) that all elements of the formPn 1 B1uk(cid:1) =Pn belongs to the domain of Z, and Z(cid:0)Pn 1 B1uk! = EX1 nXk=0 EX2 nXk=0 nXk=0 2 B2uk! = EX2 Z 1 B1uk! . k=0 Ak k=0 Ak k=0 Ak Ak Ak Ak Then it follows from [1, Theorem 1.4.2] that the operator Z has a unitary extension, and the properties in (2.39) follow by continuity. Therefore Σ1 and Σ2 are unitarily similar. Since unitary similarity clearly is a transitive property, (i) holds, and the proof is complete. (cid:3) Lemma 4.7. If the system Σ = (A, B, C, D; X , U , Y; κ) is an optimal passive realization of θ ∈ Sκ(U , Y), then X c ⊂ X o. Proof. According to Proposition 2.8, the spaces X o and (X o)⊥ are regular subspaces and (X o)⊥ is a Hilbert space. It follows from Lemma 2.9 that the system operator T of Σ can be represented as in (2.24), and the restric- tion Σo = (Ao, Bo, Co, D; X o, U , Y; κ) of Σ to the observable subspace X o is observable passive realization of θ. For n = 0, 1, 2, . . ., it holds An =(cid:18)An 1 0 f (n) An 0 (cid:19) , != P(X o)⊥(cid:16)PN PX o(cid:16)PN n=0 AnBun(cid:17)  . n=0 AnBun(cid:17) where f (n) is an operator depending on n. Then for any N ∈ N0 and any {un}N n=0 ⊂ U , it holds AnBun = PN NXn=0 n=0 (An 1 B1un + f (n)Boun) n=0 An o Boun PN But since Σ is optimal and (X o)⊥ is a Hilbert space, one deduces that AnBun! =E(X o)⊥ P(X o)⊥ NXn=0 P(X o)⊥ NXn=0 AnBun!! + EX o NXn=0 AnBun! = 0. o Boun! . An 30 L. Lilleberg This implies EX NXn=0 That is, span{AnB : n = 0, 1, . . .} ⊂ X o and since X o is closed, also span{AnB : n = 0, 1, . . .} = X c ⊂ X o. (cid:3) The next Theorem contains promised extensions for some results of [4]. In particular, the fact that statements (I)(b), (II)(b) and (III)(b) implies the other statements, respectively, in parts (I), (II) and (III), is new also in the Hilbert space setting. Theorem 4.8. Let θ ∈ Sκ(U , Y), where U and Y are Pontryagin spaces with the same negative index. (I) The following statements are equivalent: (a) ϕθ ≡ 0; (b) all κ-admissible controllable passive realizations of θ are minimal isometric; (c) there exists an observable conservative realization of θ; (d) all simple conservative realization of θ are observable; (e) all observable co-isometric realizations of θ are conservative. (II) The following statements are equivalent: (a) ψθ ≡ 0; (b) all κ-admissible observable passive realization of θ are minimal co- isometric; (c) there exists a controllable conservative realization of θ; (d) all simple conservative realization of θ are controllable; (e) all controllable isometric realizations of θ are conservative. (III) The following statements are equivalent: (a) ϕθ ≡ 0 and ψθ ≡ 0; (b) all κ-admissible simple passive realization of θ are minimal con- servative; (d) there exists a minimal conservative realization of θ. Proof. (I) (a) ⇒ (b). Suppose (a). Let the systems Σ = (A, B, C, D; X , U , Y; κ), Σ0 = (A0, B0, C0, D; X0, U , Y; κ) be, respectively, a controllable passive and an optimal minimal passive real- izations of θ. Represent the Julia embeddings of Σ and Σ0 as in (4.5). Then, (4.8) holds for Σ. Since ϕθ ≡ 0, if follows from the definition of ϕθ that I − θ∗(w)θ(z) = (1 − z ¯w)B∗ 0 (I − ¯wA∗ 0)−1(I − zA0)−1B0 Minimal realizations 31 holds for every z and w in a sufficiently small symmetric neighbourhood Ω of the origin. Since Σ0 is optimal, by considering the Neuman series of (I − zA0)−1B0 and (I − zA0)−1B0, one deduces that B∗ 0 (I − ¯zA∗ 0)−1(I − zA0)−1B0 ≤ B∗(I − ¯zA∗)−1(I − zA)−1B, z ∈ Ω. Then it holds ϕ∗(z)ϕ(z) ≤= 0 for every z ∈ Ω. But since ϕ(z) is an operator whose range belongs to the Hilbert space DT , this implies ϕ(z) = D∗ T,2 + zD∗ T,1(I − zA)−1B = 0, z ∈ Ω. T,2 = 0. Since Σ is controllable, span{(I − zA)−1B; z ∈ Ω} It follows that D∗ is dense in X , and therefore also D∗ T,1 = 0. Then DT = 0, so T is isometric, and Σ is a controllable isometric system. In particular, if Σ is chosen to be minimal passive; for the existence, see Lemma 2.9, the previous argument shows that Σ is a minimal isometric realization of θ. Since all controllable isometric realizations of θ are unitarily similar, they are now also minimal, and (b) holds. (b) ⇒ (c). Suppose (b). Let Σ′ = (A′, B′, C ′, D; X ′, U , Y; κ) be an opti- mal minimal passive realization of θ. The existence of Σ′ follows from The- orem 3.5 (i). By assumption, Σ′ is isometric. It follows from Theorem 3.5 (iv) that Σ′ is the first minimal restriction of the simple conservative system Σ = (A, B, C, D; X , U , Y; κ). By Lemma 2.9, the system operator TΣ of Σ can be represented as in (2.27), where now X ′ = PX oX c. Since the system operator TΣ′ =(cid:18)A′ B′ C ′ D(cid:19) :(cid:18)X ′ U(cid:19) →(cid:18)X ′ Y(cid:19) , of Σ′ is isometric and TΣ is unitary, an easy calculation using the fact that the range space (X o)⊥ is a Hilbert space shows that B′ 12 = 0 in (2.27). But then for every x ∈ (X o)⊥ and every n = 0, 1, 2, . . . , 1 = 0 and A′ That is, (X o)⊥ ⊂ (X c)⊥ and therefore X c ⊂ X o. Since Σ is simple, this implies now X o = X . Then Σ is observable, and (c) holds. (c) ⇒ (a). Suppose (c). Let Σ = (A, B, C, D; X , U , Y; κ) be an observable conservative realization of θ. By Lemma 2.9, Σ can be represented as in (2.27). The first minimal restriction (2.32) of Σ is an optimal minimal realization of θ by Theorem 3.5 (i). But since Σ is observable, X o = X and (X o)⊥ = {0}. It follows that the reprentations (2.25) and (2.27) coinsides. That is, the first minimal restriction Σ′ is just a restriction to the controllable subspace of Σ. By Lemma 2.9, Σ′ is now isometric. Thus if one constructs a Julia operator of TΣ′ as in (2.13), DTΣ′ = 0, and then it follows from the definition of ϕθ and (4.6) that ϕθ ≡ 0, and (a) holds. The equivalences of the statements (c), (d) and (e) follow easily from the facts that all observable co-isometric realizations of θ are unitarily similar, all simple conservative realization of θ are unitarily similar and unitary similarity B∗A∗nx =(cid:0)0 B′∗ 0(cid:1) ∗ A′ 11 0 A′ 13 ∗ 0 A′ 0 ∗ A′ 23 0 0 ∗ A′ 33 ∗ n x 0 0 = 0. 32 L. Lilleberg preserves the structural properties of the system and system operator. The part (I) is proven. (II) The proof is analogous to the proof of the part (I), and the details are omitted. (III) (a) ⇒ (b). Suppose (a). By combining the parts (I) and (II), it fol- lows that all controllable or observable passive realizations of θ are minimal conservative. Consider a simple passive realization Σ = (A.B, C, D; X , U , Y; κ) of θ. It follows from Lemma 2.9 that the contractive system operator T of Σ can be represented as in (2.24), where the restriction Σo in (2.29) is ob- servable passive, and therefore now minimal conservative. Then the system operator of Σo is unitary. Let x ∈ X o. Then, by contractivity of T and unitarity of TΣo Co D(cid:19) :(cid:18)X o TΣo =(cid:18)Ao Bo U (cid:19) →(cid:18)X o Y(cid:19) 0 = E Cox = E (A2x) + E(cid:18)(cid:18)A0x Cox(cid:19)(cid:19) = E (T x) ≤ E(x) = E(TΣo x) = E(cid:18)(cid:18)A0x Cox(cid:19)(cid:19) . A2x Aox 0 x E A1 A2 B1 0 Ao Bo 0 Co D Since A2x ∈ (X o)⊥ and (X o)⊥ is a Hilbert space, it follows that A2 = 0. If one chooses u ∈ U , a similar argument as above shows that B1 = 0. Then for any n ∈ N, it holds 0 AnB =(cid:18)A1 A∗nC ∗ =(cid:18)A∗ 0 Ao(cid:19)n(cid:18) 0 o(cid:19)n(cid:18) 0 0 1 0 A∗ Bo(cid:19) =(cid:18) 0 o(cid:19) =(cid:18) 0 o Bo(cid:19) , o(cid:19) . An A∗n o C ∗ C ∗ This is only possible if (X o)⊥ = 0, since Σ is simple. But then the systems Σ0 and Σ coincide, so the system Σ is minimal conservative, and (b) holds. Now (b) trivially implies (c), and the fact that (c) implies (a) follows (cid:3) by combining the parts (I) and (II). The proof is complete. Remark 4.9. If U and Y are Hilbert spaces, it follows from [27, Lemma 3.2] that simple passive realizations of θ ∈ Sκ(U , Y) are κ-admissible. There- fore, in that case it is not necessary to assume the considered systems to be κ-admissible in Lemma 4.6 and Theorems 4.8 and 4.10, since the other assumptions already guarantee it. However, if U and Y are Pontryagin spaces with the same negative index, it is not known that are all simple passive, or even all minimal passive, realizations of θ ∈ Sκ(U , Y) κ-admissible. If ϕθ ≡ 0 (ψθ ≡ 0), then Theorem 4.8 shows that all κ-admissible mini- mal passive realizations of θ ∈ Sκ(U , Y) are minimal isometric (co-isometric). In particular, they are controllable isometric (observable coisometric), and it follows from Theorem 2.7 that they are unitarily similar. This situation can Minimal realizations 33 occur also when the defect functions do not vanish identically. In what follows, the range of ϕθ and the domain of ψθ will be denoted, respectively, by Dϕθ and Dψθ . In the Hilbert space settings, it is well known [17, 18] that for a stan- dard Schur function θ ∈ S(U , Y), there exists a function χθ ∈ L∞(Dψθ , Dϕθ ) such that the function Θ(ζ) :=(cid:18) θ(ζ) ϕθ(ζ) χθ(ζ)(cid:19) ψθ(ζ) (4.21) has contractive values for a.e. ζ ∈ T. Under certain normalizing conditions for the functions ϕθ and ψθ, the function χθ is unique. In general, χθ may has negative Fourier coefficients and therefore it is not a Schur function. In that case the function Θ in (4.21) is not a Schur function either. However, Arov and Nudelmann showed in [9, 10] that Θ is a Schur function if and only if all minimal passive realizations of θ are unitarily similar. This result will be generalized to the indefinite settings in the following theorem. The proof uses optimal and ∗-optimal realizations as in [9, 10], but it is more elementary. Theorem 4.10. Let θ ∈ Sκ(U , Y), where U and Y are Pontryagin spaces with the same negative index, and let ϕθ and ψθ be defect functions of θ. Then all κ-admissible minimal passive realizations of θ are unitarily similar if and only if there exist an L(Dψθ , Dϕθ )-valued function χθ analytic in a neighbourhood of the origin such that Θ =(cid:18) θ ϕθ χθ(cid:19) ∈ Sκ(cid:18)(cid:18) U Dψθ(cid:19) ,(cid:18) Y Dϕθ(cid:19)(cid:19) ψθ (4.22) Proof. Suppose that all κ-admissible minimal passive realizations of θ ∈ Sκ(U , Y) are unitarily similar. Then it follows from Lemma 4.6 that every κ-admissible minimal passive realization is optimal and ∗-optimal. Take any κ-admissible minimal passive realization Σ of θ and consider its Julia em- bedding as in (4.5). Then the transfer function (4.6) of the Julia embedding belongs to the class Sκ (U ⊕ DT ∗ , Y ⊕ DT ) , and since Σ is both optimal and ∗-optimal, the upper right corner and lower left corner of (4.6) are defect functions of θ. Choose χθ = χ in (4.6), and the necessity is proven. Suppose then that there exists an L(Dψθ , Dϕθ )-valued function χθ such that Θ in (4.22) belongs to the class Sκ (U ⊕ Dψθ , Y ⊕ Dϕθ ) . It suffices to show that there exists minimal passive realization Σ of θ such that it is both optimal and ∗-optimal; see Lemma 4.6. Let be a simple conservative realization of Θ ∈ Sκ (U ⊕ Dψθ , Y ⊕ Dϕθ ) . Then the system operator TΘ of ΣΘ can be represented as ΣΘ = (A, eB,eC, eD; X , U ⊕ Dψθ , Y ⊕ Dϕθ ; κ) Dψθ(cid:19) → Dϕθ(cid:19) . X(cid:18) U X(cid:18) Y D21 D22(cid:19) : (cid:0)B B1(cid:1) TΘ = C1(cid:19) (cid:18) D D12 (cid:18) C A 34 L. Lilleberg In a sufficiently small symmetric neighbourhood Ω of the origin, it holds Θ(z) =(cid:18) θ(z) ψθ(z) ϕθ(z) χθ(z)(cid:19) D21 + zC1 + (I − zA)−1B D22 + zC1(I − zA)−1B1(cid:19) . =(cid:18) D + zC(I − zA)−1B D12 + zC(I − zA)−1B1 The spaces Dϕθ and Dψθ are Hilbert spaces, and therefore it follows that Σ = (A, B, C, D; X , U , Y; κ) is a passive realization of θ. Since ΣΘ is conservative, Lemma 2.4 shows that I − Θ(z)Θ∗(w) I − Θ∗(w)Θ(z) B∗ −θ(z)ϕ∗ IDϕθ − ϕθ(z)ϕ∗ θ (w) θ (w) θ(w)(cid:19) θ(w) − ψθ(z)χ∗ θ(w) θ(w) − χθ(z)χ∗ −ϕθ(z)θ∗(w) − χθ(z)ψ∗ C(I − zA)−1(I − ¯wA∗)−1C ∗ 1 C1(I − zA)−1(I − ¯wA∗)−1C ∗ C1(I − zA)−1(I − ¯wA∗)−1C ∗ =(cid:18)IY − θ(z)θ∗(w) − ψθ(z)ψ∗ = (1 − ¯wz)eC(I − zA)−1(I − ¯wA∗)−1eC ∗ = (1 − ¯wz)(cid:18) C(I − zA)−1(I − ¯wA∗)−1C ∗ θ(w)χθ(z)(cid:19) =(cid:18)IU − θ∗(w)θ(z) − ϕ∗ 1 (I − ¯wA∗)−1(I − zA)−1B1(cid:19) . = (1 − ¯wz)(cid:18)B∗(I − ¯wA∗)−1(I − zA)−1B B∗(I − ¯wA∗)−1(I − zA)−1B1 1 (I − ¯wA∗)−1(I − zA)−1B B∗ − ψ∗ θ (w)ψθ(z) − χ∗ −ψ∗ θ (w)θ(z) − χ∗ θ(w)ϕθ(z) −θ∗(w)ψθ(z) − ϕ∗ θ(w)χθ(z) θ(w)ϕθ(z) IDψθ 1(cid:19) That is, IY − θ(z)θ∗(w) = (1 − ¯wz)C(I − zA)−1(I − ¯wA∗)−1C ∗ + ψθ(z)ψ∗ θ (w), (4.23) IU − θ∗(w)θ(z) = (1 − ¯wz)B∗(I − ¯wA∗)−1(I − zA)−1B + ϕ∗ θ(w)ϕθ(z). (4.24) An easy calculation and Lemma 4.3 show that the equation (4.23) is equiva- lent to IY − θ#∗ (w)θ#(z) = (1 − ¯wz)C(I − ¯wA)−1(I − zA∗)−1C ∗ + ϕθ# ∗(w)ϕθ# (z). (4.25) Let Σ′ = (A′, B′, C ′, D; X ′, U , Y; κ), Σ′′ = (A′′, B′′, C ′′, D; X ′′, U , Y; κ) be, respectively, an optimal minimal and a ∗-optimal minimal realizations of θ. It follows from Theorem 3.5 (ii) that Σ′′∗ = (A′′ ∗, C ′′ ∗, B′′∗, D; X ′′, U , Y; κ) is an optimal minimal realization of θ#. Then, by the definition of ϕθ and ϕθ#, it holds IU − θ∗(w)θ(z) = (1 − wz)B′∗(I − ¯wA′∗)−1(I − zA′)−1B′ + ϕ∗ θ (w)ϕθ(z) IY − θ# ∗ (w)θ#(z) = (1 − wz)C ′′(I − ¯wA′′)−1(I − zA′′∗)−1C ′′∗ + ϕθ# ∗(w)ϕθ# (z). Minimal realizations 35 It follows that B∗(I − ¯wA∗)−1(I − zA)−1B = B′∗(I − ¯wA′∗)−1(I − zA′)−1B′, C(I − ¯wA)−1(I − zA∗)−1C ∗ = C ′′(I − ¯wA′′)−1(I − zA′′∗)−1C ′′∗. By using Lemma 4.5, it can be deduced that Σ and Σ∗ are optimal systems. Then it follows from Lemma 4.7 that X c = X o and therefore X s = X c = X o. By Lemma 2.9, the restriction Σs = (As, Bs, Cs, D; X s, U , Y; κ) of Σ to the s Bs and A∗nC ∗ = simple subspace X s is simple, and it holds AnB = An A∗ s also are optimal systems. More- s over, they are minimal since X s = X c = X o. It follows now from Theorem 3.5 (ii) that Σs is also ∗-optimal, and the proof is complete. (cid:3) s for every n ∈ N0. That is, Σs and Σ∗ nC ∗ I wish to thank Seppo Hassi for helpful discussions while Acknowledgements preparing this paper. References [1] D. Alpay, A. Dijksma, J. Rovnyak, and H. S. V. de Snoo, Schur functions, operator colligations, and Pontryagin spaces, Oper. Theory Adv. Appl., 96, Birkhauser Verlag, Basel-Boston, 1997. [2] D. Alpay, T.Y. Azizov, A. Dijksma and J. Rovnyak, Colligations in Pontrya- gin Spaces with a Symmetric Characteristic Function, Linear Operators and Matrices, Oper. Theory Adv. Appl., 130, Birkhauser, Basel, 2002. [3] T. Ando, De Branges spaces and analytic operator functions, Division of Applied Mathematics, Research Institute of Applied Electricity, Hokkaido University, Sapporo, Japan, 1990. [4] Yu. M. Arlinskiı, S. Hassi and H.S.V de Snoo, Parametrization of contractive block operator matrices and passive discrete-time systems, Complex Anal. Oper. Theory 1 (2007), no. 2, 211 -- 233. [5] D. Z. Arov, Passive linear steady-state dynamical systems, Sibirsk. Mat. Zh. 20 (1979), no. 2, 211 -- 228 (Russian); English transl. in Siberian Math. J. 20 (1979), no. 2, 149 -- 162. [6] D. Z. Arov, Stable dissipative linear stationary dynamical scattering sys- tems, J. Operator Theory 2 (1979), no. 1, 95 -- 126 (Russian); English transl. in Oper. Theory Adv. Appl., 134, Interpolation theory, systems theory and related topics (Tel Aviv/Rehovot, 1999), 99 -- 136, Birkhauser, Basel, 2002. [7] D. Z. Arov, M. A. Kaashoek, and D. P. Pik, Minimal and optimal linear discrete time-invariant dissipative scattering systems. Integr. Equat. Oper. Theory, 29 (1997), 127 -- 154. [8] D. Z. Arov, M. A. Kaashoek, and D. P. Pik, The Kalman-Yakubovich-Popov inequality for discrete time systems of infinite dimension, J. Operator Theory 55 (2006), no. 2, 393 -- 438. [9] D. Z. Arov and M. A. Nudel'man, A criterion for the unitary similarity of minimal passive systems of scattering with a given transfer function, Ukraın. Mat. Zh. 52 (2000), no. 2, 147 -- 156 (Russian); English transl. in Ukrainian Math. J. 52 (2000), no. 2, 161 -- 172. 36 L. Lilleberg [10] D.Z. Arov and M. A. Nudel'man, Conditions for the similarity of all minimal passive realizations of a given transfer function (scattering and resistance matrices), Mat. Sb. 193 (2002), no. 6, 3 -- 24 (Russian); English transl. in Sb. Math. 193 (2002), no. 5-6, 791 -- 810. [11] D. Z. Arov, J. Rovnyak and S. M. Saprikin, Linear passive stationary scat- tering systems with Pontryagin state spaces, Math. Nachr. 279 (2006), no. 13 -- 14, 1396 -- 1424. [12] D. Z. Arov and S. M. Saprikin, Maximal solutions for embedding problem for a generalized Shur function and optimal dissipative scattering systems with Pontryagin state spaces, Methods Funct. Anal. Topology 7 (2001), no. 4, 69 -- 80. [13] T. Ya. Azizov and I. S. Iokhvidov, Foundations of the theory of linear opera- tors in spaces with indefinite metric, Nauka, Moscow, 1986; English transl., John Wiley & Sons Ltd., Chichester, 1989. [14] H. Bart, , I. Z. Gohberg, M. A. Kaashoek and A. C. M. Ran, Factorization of matrix and operator functions: the state space method, Oper. Theory Adv. Appl., 178, Linear Operators and Linear Systems, Birkhauser Verlag, Basel, 2008. [15] J. Bogn´ar, Indefinite inner product spaces, Ergebnisse der Mathematik und ihrer Grenzgebiete, Band 78. Springer-Verlag, New York-Heidelberg, 1974. [16] S. S. Boiko, V. K. Dubovoj, B. Fritzsche and B. Kirstein, Models of con- tractions constructed from the defect function of their characteristic func- tion. Operator theory, system theory and related topics (Beer-Sheva/Rehovot, 1997), 6787, Oper. Theory Adv. Appl., 123, Birkhauser, Basel, 2001. [17] S. S. Boiko, V. K. Dubovoj, Defect functions of holomorphic contractive op- erator functions and the scattering suboperator through the internal chan- nels of a system. Part I, Complex Anal. Oper. Theory 5 (2011), no. 1, 157196. [18] S. S. Boiko, V. K. Dubovoj, A. Ya. Kheifets, Defect functions of holomorphic contractive operator functions and the scattering suboperator through the internal channels of a system: Part II. Complex Anal. Oper. Theory 8 (2014), no. 5, 9911036. [19] L. de Branges and J. Rovnyak, Square Summable Power Series, Holt, Rine- hart and Winston, New-York, 1966. [20] L. de Branges and J. Rovnyak, Appendix on square summable power series, Canonical models in quantum scattering theory, Perturbation Theory and its Applications in Quantum Mechanics (Proc. Adv. Sem. Math. Res. Center, U.S. Army, Theoret. Chem. Inst., Univ. of Wisconsin, Madison, Wis., 1965), pp. 295 -- 392, Wiley, New York, 1966. [21] M.S. Brodskiı, Unitary operator colligations and their characteristic func- tions, Uspekhi Mat. Nauk 33 (1978), no. 4(202), 141 -- 168, 256 (Russian); English transl. in Russian Math. Surveys 33 (1978), no. 4, 159 -- 191. [22] A. Dijksma, H. Langer and H. S. V. de Snoo, Characteristic functions of uni- tary operator colligations in πκ-spaces, Operator theory and systems (Am- sterdam, 1985), 125 -- 194, Oper. Theory Adv. Appl., 19, Birkhauser, Basel, 1986. Minimal realizations 37 [23] A. Dijksma, H. Langer and H. S. V. de Snoo, Unitary colligations in Πκ- spaces, characteristic functions and Straus extensions, Pacific J. Math. 125 (1986), no. 2, 347 -- 362. [24] M. A. Dritschel and J. Rovnyak, Operators on indefinite inner product spaces, Lectures on operator theory and its applications (Waterloo, ON, 1994), 141 -- 232, Fields Inst. Monogr., 3, Amer. Math. Soc., Providence, RI, 1996. [25] J. W. Helton, Discrete time systems, operator models, and scattering theory, J. Functional Analysis 16 (1974), 15 -- 38. [26] M. G. Kreın and H. Langer, Uber die verallgemeinerten Resolventen und die charakteristische Funktion eines isometrischen Operators im Raume Πκ (German), Hilbert space operators and operator algebras (Proc. Internat. Conf., Tihany, 1970), pp. 353 -- 399, Colloq. Math. Soc. J´anos Bolyai, 5. North-Holland, Amsterdam, 1972. [27] L. Lilleberg, Isometric discrete-time systems with Pontryagin state space, Complex Anal. Oper. Theory (2019). https://doi.org/10.1007/s11785-019- 00930-1 [28] S. M. Saprikin, The theory of linear discrete time-invariant dissipative scat- tering systems with state πκ-spaces, Zap. Nauchn. Sem. S.-Peterburg. Ot- del. Mat. Inst. Steklov. (POMI) 282 (2001), Issled. po Lineın. Oper. i Teor. Funkts. 29, 192 -- 215, 281 (Russian); English transl. in J. Math. Sci. (N. Y.) 120 (2004), no. 5, 1752 -- 1765. [29] O. J. Staffans, Well-posed linear systems, Encyclopedia of Mathematics and its Applications, 103, Cambridge University Press, Cambridge, 2005. [30] B. Sz.-Nagy and C. Foias, Harmonic analysis of operators on Hilbert space, North-Holland, New York, 1970.
1510.08801	1	1510	2015-10-29T17:57:31	Riemann integrability versus weak continuity	[ "math.FA" ]	In this paper we focus on the relation between Riemann integrability and weak continuity. A Banach space $X$ is said to have the weak Lebesgue property if every Riemann integrable function from $[0,1]$ into $X$ is weakly continuous almost everywhere. We prove that the weak Lebesgue property is stable under $\ell_1$-sums and obtain new examples of Banach spaces with and without this property. Furthermore, we characterize Dunford-Pettis operators in terms of Riemann integrability and provide a quantitative result about the size of the set of $\tau$-continuous non Riemann integrable functions, with $\tau$ a locally convex topology weaker than the norm topology.	math.FA	math	RIEMANN INTEGRABILITY VERSUS WEAK CONTINUITY GONZALO MART´INEZ-CERVANTES Abstract. In this paper we focus on the relation between Riemann integrabil- ity and weak continuity. A Banach space X is said to have the weak Lebesgue property if every Riemann integrable function from [0, 1] into X is weakly continuous almost everywhere. We prove that the weak Lebesgue property is stable under ℓ1-sums and obtain new examples of Banach spaces with and without this property. Furthermore, we characterize Dunford-Pettis operators in terms of Riemann integrability and provide a quantitative result about the size of the set of τ -continuous non Riemann integrable functions, with τ a locally convex topology weaker than the norm topology. 1. Introduction The study of the relation between Riemann integrability and continuity on Ba- nach spaces started on 1927, when Graves showed in [13] the existence of a vector- valued Riemann integrable function not continuous almost everywhere (a.e. for short). Thus, the following problem arises: Given a Banach space X, determine necessary and sufficient conditions for the Riemann integrability of a function f : [0, 1] → X. A Banach space X for which every Riemann integrable function f : [0, 1] → X is continuous a.e. is said to have the Lebesgue property (LP for short). All classical infinite-dimensional Banach spaces except ℓ1 do not have the LP. For more details on this topic, we refer the reader to [12], [6], [24], [14] and [19]. Regarding weak continuity, Alexiewicz and Orlicz constructed in 1951 a Riemann integrable function which is not weakly continuous a.e. [2]. A Banach space X is said to have the weak Lebesgue property (WLP for short) if every Riemann integrable function f : [0, 1] → X is weakly continuous a.e. This property was introduced in [27]. Every Banach space with separable dual has the WLP and the example of [2] shows that C([0, 1]) does not have the WLP. Other spaces with the WLP, such as L1([0, 1]), can be found in [5] and [28]. In this paper we focus on the relation between Riemann integrability and weak continuity. In Section 2 we present new results on the WLP. In particular, we prove that the James tree space JT does not have the WLP (Theorem 2.3) and we study when ℓp(Γ) and c0(Γ) have the WLP in the nonseparable case (Theorem 2.8). Moreover, we prove that the WLP is stable 2010 Mathematics Subject Classification. 46G10, 28B05, 03E10. Key words and phrases. Riemann integral, Lebesgue property, weak Lebesgue property, Ba- nach space. This work was supported by the research project 19275/PI/14 funded by Fundaci´on S´eneca - Agencia de Ciencia y Tecnolog´ıa de la Regi´on de Murcia within the framework of PCTIRM 2011- 2014. This work was also supported by Ministerio de Econom´ıa y Competitividad and FEDER (project MTM2014-54182-P). 1 2 GONZALO MART´INEZ-CERVANTES under ℓ1-sums (Theorem 2.13) and we apply this result to obtain that C(K)∗ has the WLP for every compact space K in the class M S (Corollary 2.16). Alexiewicz and Orlicz also provided in their paper an example of a weakly con- tinuous non Riemann integrable function. V. Kadets proved in [15] that a Banach space X has the Schur property if and only if every weakly continuous function f : [0, 1] → X is Riemann integrable. Wang and Yang extended this result in [29] to arbitrary locally convex topologies weaker than the norm topology. In the last section of this paper we give an operator theoretic form of these results that, in particular, provides a positive answer to a question posed by Sofi in [26]. Terminology and Preliminaries. All Banach spaces are assumed to be real. In what follows, X ∗ denotes the dual of a Banach space X. The weak and weak∗ topologies of X and X ∗ will be denoted by ω and ω∗ respectively. By an operator we mean a linear continuous mapping between Banach spaces. The Lebesgue measure in R is denoted by µ. The interior of an interval I will be denoted by Int(I). The density character dens(T ) of a topological space T is the minimal cardinality of a dense subset. N a f (t)dt. A partition of the interval [a, b] ⊂ R is a finite collection of non-overlapping closed subintervals covering [a, b]. A tagged partition of the interval [a, b] is a partition {[ti−1, ti] : 1 ≤ i ≤ N} of [a, b] together with a set of points {si : 1 ≤ i ≤ N} that satisfy si ∈ (ti−1, ti) for each i. Let P = {(si, [ti−1, ti]) : 1 ≤ i ≤ N} be a tagged partition of [a, b]. For every function f : [a, b] → X, we denote by f (P) the Riemann Pi=1 (ti − ti−1)f (si). The norm of P is kPk := max{ti − ti−1 : 1 ≤ i ≤ N}. We sum say that a function f : [a, b] → X is Riemann integrable, with integral x ∈ X, if for every ε > 0 there is δ > 0 such that kf (P) − xk < ε for all tagged partitions P of [a, b] with norm less than δ. In this case, we write x =R b Theorem 1.1 ([12]). Let f : [0, 1] → X. The following statements are equivalent: The following criterion will be used for proving the existence of the Riemann integral of certain functions: (1) The function f is Riemann integrable. (2) For each ε > 0 there exists a partition Pε of [0, 1] with kf (P1)− f (P2)k < ε for all tagged partitions P1 and P2 of [0, 1] that have the same intervals as Pε. (3) There is x ∈ X such that for every ε > 0 there exists a partition Pε of [0, 1] such that kf (P)− xk < ε whenever P is a tagged partition of [0, 1] with the same intervals as Pε. We will also be concerned about cardinality. Throughout this paper, c denotes the cardinality of the continuum and cov(M) denotes the smallest cardinal such that there exist cov(M) nowhere dense sets in [0, 1] whose union is the interval [0, 1]. This cardinal coincides with the smallest cardinal such that there exist cov(M) closed sets in [0, 1] with Lebesgue measure zero whose union does not have Lebesgue measure zero (see [4, Theorem 2.6.14]). (εn)∞ A set A ⊂ R is said to be strongly null if for every sequence of positive reals n=1 there exists a sequence of intervals (In)∞ n=1 such that µ(In) < εn for every n ∈ N and A ⊂Sn∈N In. We will be interested in the following result: RIEMANN INTEGRABILITY VERSUS WEAK CONTINUITY 3 Theorem 1.2 ([22]). A set A ⊂ R is strongly null if and only if for every closed set F with Lebesgue measure zero, the set A + F = {a + z : a ∈ A and z ∈ F} has Lebesgue measure zero. We will denote by non(SN ) the smallest cardinal of a non strongly null set. We have that ℵ1 ≤ cov(M) ≤ non(SN ) ≤ c and, under Martin's axiom, and therefore under the Continuum Hypothesis, non(SN ) = cov(M) = c. Furthermore, if b = c then non(SN ) = cov(M). However, there exist models of ZFC satisfying cov(M) < non(SN ). For further references and results on this subject we refer the reader to [3]. 2. The weak Lebesgue property It is known that every Banach space with separable dual has the WLP [28]. Next theorem gives a generalization in terms of cov(M). Theorem 2.1. Every Banach space X such that dens(X ∗) < cov(M) has the WLP. Proof. Let D = {x∗ i }i∈I be a dense subset in X ∗ with I < cov(M) and take f : [0, 1] → X a Riemann integrable function. Then every function x∗ i f is Riemann integrable. Let Ei be the set of points of discontinuity of x∗ i f for every i ∈ I. Each Ei is a countable union of closed sets with measure zero, so E =Si∈I Ei has measure zero since I < cov(M). We claim that f is weakly continuous at every point of Ec. Let x∗ ∈ X ∗ and let M be an upper bound for {kf (t)k : t ∈ [0, 1]}. Fix ε > 0 and t ∈ Ec. Then, there exists x∗ 3M . Since t /∈ Ei, there exists a neighbourhood U of t such that x∗ 3 for every t′ ∈ U . Thus, x∗f (t) − x∗f (t′) ≤ x∗f (t) − x∗ for every t′ ∈ U . Corollary 2.2. Every Banach space with separable dual has the WLP. i f (t′) − x∗f (t′) < ε i ∈ D such that kx∗ i − x∗k < ε i f (t) + x∗ i f (t) − x∗ i f (t′) + x∗ i f (t) − x∗ i f (t′) < ε (cid:3) The space ℓ1 has the WLP because it has the LP. Since every asymptotic ℓ1 space has the LP [19], the space ΛT defined by Odell in [21] is a separable Banach space with nonseparable dual such that it does not contain an isomorphic copy of ℓ1 but it has the WLP (it is asymptotic ℓ1). On the other hand, the James tree space JT (see [1, Section 13.4]) is a separable Banach space with nonseparable dual such that it does not contain an isomorphic copy of ℓ1 and it does not have the WLP: Theorem 2.3. The James tree space does not have the WLP. Proof. We represent the dyadic tree by T = {(n, k) : n = 0, 1, 2, . . . and k = 1, 2, . . . , 2n}. A node (n, k) ∈ T has two immediate successors (n+1, 2k−1) and (n+1, 2k). Then, a segment of T is a finite sequence {p1, . . . , pm} such that pj+1 is an immediate successor of pj for every j = 1, 2, . . . , m − 1. The James tree space JT is the completion of c00(T ) with the norm l kxk = supvuuut Xj=1 2   X(n,k)∈Sj x(n, k)  < ∞, 4 GONZALO MART´INEZ-CERVANTES where the supremum is taken over all l ∈ N and all sets of pairwise disjoint segments S1, S2, . . . , Sl. Let {e(n,k)}(n,k)∈T be the canonical basis of JT , i.e. e(n,k) is the characteristic function of (n, k) ∈ T . Define f : [0, 1] → JT as follows: f (t) =( e(n−1,k) 0 if t = 2k−1 in any other case. 2n with n ∈ N and k = 1, 2, . . . , 2n−1 We claim that f is Riemann integrable. Fix N ∈ N and let {I1, I2, . . . , I2N −1} be a family of closed disjoint intervals of [0, 1] with µ(In) ≤ 1 2N and n 2N ∈ Int(In) for each 1 ≤ n ≤ 2N − 1. X1≤n≤2N −1 Let J1, J2, . . . , J2N be the closed disjoint intervals of [0, 1] determined by Then, µ(Jn) ≤ 1 n for every an ∈ R and every tn ∈ Jn due to the definition of the norm in JT . Thus, any tagged partition PN 1, t2, . . . , t′ with intervals J1, I1, J2, . . . , I2N −1, J2N and points t1, t′ 2N −1, t2N satisfies n=1 a2 Int(In). [0, 1] \ [1≤n≤2N −1 n=1 anf (tn)k ≤qP2N µ(Jn)f (t2n−1)(cid:13)(cid:13)(cid:13)(cid:13) 2N ≤vuut Xn=1 2N and kP2N kf (PN )k ≤(cid:13)(cid:13)(cid:13)(cid:13) Xn=1 ≤vuut Xn=1 µ(Jn)2 + 2N 2N 2N 1 + 2N −1 Xn=1 µ(In) ≤ 1 22N + 1 2N ≤ 2 √2N . Hence, kf (PN )k N→∞−−−−→ 0 and f is Riemann integrable with integral zero. We show that f is not weakly continuous at any irrational point t ∈ [0, 1]. Fix 2nj (cid:17)∞ a irrational point t ∈ [0, 1]. There exists a sequence of dyadic points (cid:16) 2kj −1 converging to t with (nj − 1, kj)∞ j=1 a sequence in T such that (nj+1 − 1, kj+1) is an immediate successor of (nj − 1, kj) for every j ∈ N. Then, P∞ (nj −1,kj ) is a functional in JT ∗, so the sequence f ( 2kj −1 2nj ) = e(nj −1,kj ) is not weakly null and f is not weakly continuous at t. (cid:3) j=1 e∗ j=1 Corollary 2.4 ([2]). C([0, 1]) does not have the WLP. Proof. Since every subspace of a Banach space with the WLP has the WLP and every separable Banach space is isometrically isomorphic to a subspace of C([0, 1]), it follows from the previous theorem and the separability of JT that C([0, 1]) does not have the WLP. (cid:3) Corollary 2.5. Let K be a compact Hausdorff space. (1) If K is metrizable, then C(K) has the WLP if and only if K is countable. (2) If C(K) has the WLP then K is scattered. The converse is not true since c0(c) does not have the WLP (Theorem 2.8) and it is isomorphic to a C(K) space with K scattered. RIEMANN INTEGRABILITY VERSUS WEAK CONTINUITY 5 Proof. If K is a countable compact metric space, then C(K)∗ is separable [10, Theorem 14.24], so C(K) has the WLP (Theorem 2.1). If K is an uncountable compact metric space, then C(K) is isomorphic to C([0, 1]) [1, Theorem 4.4.8], so C(K) does not have the WLP (Corollary 2.4). Finally, if K is not scattered, then C(K) has a subspace isomorphic to C([0, 1]) (see the proof of [10, Theorem 14.26]), so C(K) does not have the WLP. Remark 2.6. Let {Xi}i∈Γ be a family of Banach spaces. Define X := (Li∈Γ Xi)ℓp with 1 < p < ∞ or X := (Li∈Γ Xi)c0 . If f : [0, 1] → X is a bounded function, then its set of points of weak discontinuity is E = Si∈Γ Ei, where each Ei is the set of points of weak discontinuity of fi and fi is the i'th cordinate of f . Thus, the countable ℓp-sum or c0-sum of Banach spaces with the WLP has the WLP. We cannot extend this result to uncountable ℓp-sums or c0-sums even when Xi = R for every i ∈ Γ (Theorem 2.8). (cid:3) Now, we study the WLP for the spaces of the form c0(κ) and ℓp(κ) with κ a cardinal. Theorem 2.7. For any cardinal κ and any 1 < p < ∞, c0(κ) has the WLP if and only if ℓp(κ) has the WLP. α has not Lebesgue measure zero. Proof. If ℓp(κ) does not have the WLP, then there exists a Riemann integrable function f : [0, 1] → ℓp(κ) which is not weakly continuous a.e. If I : ℓp(κ) → c0(κ) is the canonical inclusion, then the function I ◦ f is weakly continuous at a point t ∈ [0, 1] if and only if f is weakly continuous at t by Remark 2.6. Therefore, I ◦ f is not weakly continuous a.e. Since I is an operator, I ◦ f is also Riemann integrable. Thus, c0(κ) does not have the WLP. To prove the other implication, suppose c0(κ) does not have the WLP. Then, there exists a Riemann integrable function f : [0, 1] → c0(κ) which is not weakly continuous a.e. Let fα be the α'th cordinate of f for every α ∈ κ and En α be the set of points where fα has oscillation strictly bigger than 1 n for every n ∈ N. Note that each En α has Lebesgue measure zero. Since f is not weakly continuous a.e., Sα<κ(cid:0)Sn∈N En α(cid:1) has not Lebesgue measure zero, so there exists n ∈ N such that Sα<κ En β(cid:17) for every α ∈ κ \ {0}. The sets Fα are pairwise disjoint. Let χFα : [0, 1] → {0, 1} be the characteristic function of Fα for every α < κ and g : [0, 1] → c0(κ) the function defined by the formula g(t) = Pα<κ χFα (t)eα for every t ∈ [0, 1], where {eα}α<κ is the canonical basis of X. Notice that g is not weakly continuous a.e. since each χFα is not continuous at any point of Fα (because µ(Fα) = 0) and Sα<κ Fα = Sα<κ En α is not Lebesgue null. We claim that g is Riemann integrable. Let ε > 0. Since f is Riemann integrable, there exists a partition Pε of [0, 1] such that kf (P1)− f (P2)k < ε n for all tagged partitions P1 and P2 of [0, 1] that have the same intervals as Pε. For every α < κ and any tagged partitions P1 and P2 of [0, 1] that have the same intervals as Pε, α \(cid:16)Sβ<α En 0 and Fα := En Set F0 := En χFα(P1) − χFα(P2) ≤ N Xi=1 µ(Ii) ≤ nfα(P ′ 1) − fα(P ′ 2) ≤ nkf (P ′ 1) − f (P ′ 2)k < ε µ [Int(Ii)∩Fα6=∅ µ(Ii)eφ(si)(cid:13)(cid:13)(cid:13)(cid:13) kh(P)k =(cid:13)(cid:13)(cid:13)(cid:13)Xsi∈F Ii(cid:19)p(cid:19) µ(cid:18) [φ(si)=α =(cid:18)Xα<κ (1) p (cid:18)Xα<κ µ(cid:18) [φ(si)=α ≤ ε =(cid:13)(cid:13)(cid:13)(cid:13)Xα<κ µ(cid:18) [φ(si)=α Ii(cid:19)(cid:19) p−1 1 p =(cid:18)Xα<κ = Ii(cid:19)eα(cid:13)(cid:13)(cid:13)(cid:13) µ(cid:18) [φ(si)=α µ(cid:18) [φ(si)=α Ii(cid:19)p−1 ≤ ε 1 p p p−1 1 p Ii(cid:19)(cid:19) ≤ (cid:3) 6 GONZALO MART´INEZ-CERVANTES 1 and P ′ for suitable tagged partitions P ′ 2 of [0, 1] with the same intervals as Pε, where I1, I2, . . . , IN are the intervals of Pε whose interior has non-empty intersection with En α. Therefore, g is Riemann integrable. Let h : [0, 1] → ℓp(κ) be the function given by the formula h(t) = Pα<κ χFα(t)eα. Since the sets Fα are pairwise disjoint, the function h is well-defined. Moreover, h is not weakly continuous a.e. because I ◦ h = g. Set F =Sα<κ Fα and φ : F → κ such that φ(t) = α if t ∈ Fα. We claim that h is Riemann integrable with integral zero. Let ε > 0 and Pε = {I1, I2, . . . , IM} be a partition of [0, 1] such that kg(P ′)k < ε for any tagged partition P ′ of [0, 1] with the same intervals as Pε. Notice that (1) Ii! < ε for every α < κ. Thus, for any tagged partition P = {(si, Ii)}M i=1 the following inequalities hold: Therefore, h is Riemann integrable with Riemann integral zero. The LP is separably determined [24]. Nevertheless, it follows from the following theorem that the WLP is not separably determined, since every separable infinite- dimensional subspace of ℓ2(κ) is isomorphic to ℓ2 (which has separable dual). Theorem 2.8. Let κ be a cardinal and X = c0(κ) or X = ℓp(κ) with 1 < p < ∞. (1) If κ < cov(M) then X has the WLP. (2) If κ ≥ non(SN ) then X does not have the WLP. Proof. It is enough to prove the result when X = c0(κ) due to Theorem 2.7. Since c0(κ)∗ = ℓ1(κ) has density character κ, it follows from Theorem 2.1 that c0(κ) has the WLP if κ < cov(M). Suppose non(SN ) ≤ κ ≤ c. Due to Theorem 1.2, there exist a closed Lebesgue null set F and a set E = {xα}α<κ in R such that E + F does not have Lebesgue measure zero. Without loss of generality, we may assume that E, F ⊂ [0, 1] and (E + F )∩[0, 1] does not have Lebesgue measure zero. Set F0 := (x0 + F )∩[0, 1] and Fα := ((xα + F ) ∩ [0, 1])\(cid:16)Sβ<α Fβ(cid:17) for every 0 < α < κ. Let χFα : [0, 1] → {0, 1} be the characteristic function of Fα for every α < κ and f : [0, 1] → X the function defined by the formula f (t) = Pα<κ χFα (t)eα for every t ∈ [0, 1], where {eα}α<κ is the canonical basis of c0(κ). Since the sets Fα are pairwise disjoint, the function f is well-defined. Each χFα is not continuous at Fα, since Fα cannot contain an interval of [0, 1]. Therefore, f is not weakly continuous a.e. because Sα<κ Fα = (E + F ) ∩ [0, 1] does not have Lebesgue measure zero. RIEMANN INTEGRABILITY VERSUS WEAK CONTINUITY 7 We claim that f is Riemann integrable. For every α < κ and every tagged i=1 we have partition P = {(si, Ii)}N Xi=1 χFα(P) = N µ(Ii)χFα (si) ≤ µ(Ii − xα)χF (si − xα) = χF (P ′) N Xi=1 for a suitable tagged partition P ′ with kPk = kP ′k. Since F ⊂ [0, 1] is a closed Lebesgue measure zero set, the characteristic function χF is Riemann integrable due to Lebesgue's Theorem. Then, for every ε > 0 there exists δ > 0 such that χF (P) < ε for every tagged partition P with kPk < δ. Therefore, for every ε > 0 there exists δ > 0 such that χFα(P) < ε for all tagged partitions P with kPk < δ and for every α < κ. Thus, f is Riemann integrable since kf (P)k = supα<κ χFα(P) < ε for every tagged partition P of [0, 1] with kPk < δ. (cid:3) The facts that the countable ℓ1-sum of spaces with the WLP has the WLP (Theorem 2.11) and that L1(λ) has the WLP if dens(L1(λ)) < cov(M) (Theorem 2.12) will be a consequence of the following lemma. Lemma 2.9. Let (Ω, Σ, λ) be a probability space and P = {PA : A ∈ Σ} a family of operators on a Banach space X such that A ∩ B = ∅; (1) PA + PΩ\A = PΩ = idX for every A ∈ Σ; (2) kPA(x)k ≤ kxk for every x ∈ X and every A ∈ Σ; (3) kPA(x)k+kPB(x′)k ≤ max{kx+x′k,kx−x′k} for every x, x′ ∈ X whenever (4) limλ(A)→0 kPA(x)k = 0 for every x ∈ X. Let f : [0, 1] → X be a Riemann integrable function. Then there is a measurable set E ⊆ [0, 1] with µ(E) = 1 such that, for every sequence (tn)∞ n=1 in [0, 1] converg- ing to some t ∈ E, the set {f (tn) : n ∈ N} is P-uniformly integrable, in the sense that Proof. The proof is similar to that of [5, Lemma 2.3] and [28, Lemma 3]. Fix β > 0 and denote by Eβ the set of points t ∈ [0, 1] such that for every δ > 0 there exist t′ ∈ [0, 1] with t′ − t < δ and a set A ∈ Σ with λ(A) < δ such that lim λ(A)→0 sup n∈N(cid:13)(cid:13)PA(f (tn))(cid:13)(cid:13) = 0. kPA(f (t) − f (t′))k > β. Let µ∗ be the Lebesgue outer measure in [0, 1]. We show that µ∗(Eβ) = 0 with a proof by contradiction. Suppose µ∗(Eβ) > 0. Since f is Riemann integrable, we can choose a partition P = {J1, . . . , Jm} of [0, 1] such that < βµ∗(Eβ) (2) m µ(Jj)(f (ξj ) − f (ξ′ (cid:13)(cid:13)(cid:13)(cid:13) Xj=1 j))(cid:13)(cid:13)(cid:13)(cid:13) for all choices ξj , ξ′ where Ij = Int(Jj) for each j = 1, . . . , m. Thus, j ∈ Jj, 1 ≤ j ≤ m. Let S = {j ∈ {1, . . . , m} : Ij ∩ Eβ 6= ∅}, (3) µ∗(Ij ∩ Eβ) = µ∗(Eβ). Xj∈S It is not restrictive to suppose S = {1, . . . , n} for some 1 ≤ n ≤ m. 8 GONZALO MART´INEZ-CERVANTES k Because of the definition of Eβ and I1, there exist points t1 ∈ I1 ∩ Eβ and 1 ∈ I1 such that kf (t1) − f (t′ t′ 1))k > β for some A ∈ Σ, hence kµ(I1)(f (t1) − f (t′ 1))k > βµ(I1). Fix 1 ≤ k < n and assume that we have already chosen points tj, t′ j ∈ Ij for all 1 ≤ j ≤ k with the property that 1)k ≥ kPA(f (t1) − f (t′ > β(cid:18) k Xj=1 Define x := µ(Ij)(cid:19). µ(Ij )(f (tj) − f (t′ (cid:13)(cid:13)(cid:13)(cid:13) Xj=1 µ(Ij )(f (tj) − f (t′ j))(cid:13)(cid:13)(cid:13)(cid:13) α := kxk − β(cid:18) k µ(Ij )(cid:19) > 0. Xj=1 Due to (4), we can choose δ > 0 such that kPA(x)k < α whenever A ∈ Σ satisfies λ(A) < δ. Take tk+1, t′ k+1 ∈ Ik+1 and a set A ∈ Σ with λ(A) < δ such that kPA(f (tk+1) − f (t′ k+1))k > β, so y := µ(Ik+1)(f (tk+1) − f (t′ j)) ∈ X and k+1)) satisfies Pj=1 k kPA(y)k > βµ(Ik+1). By the choice of A, (1) and (3), we also have (interchanging the role of tk+1 and t′ k+1 if necessary) k+1 (cid:13)(cid:13)(cid:13)(cid:13) Xj=1 µ(Ij)(f (tj ) − f (t′ j))(cid:13)(cid:13)(cid:13)(cid:13) ≥ kPA(y)k + kPAc(x)k ≥ kPA(y)k + kxk − kPA(x)k > k k+1 n µ(Ij ). Xj=1 > βµ(Ik+1) + α + β Thus, there exist tj, t′ µ(Ij )(f (tj) − f (t′ Xj=1 µ(Ij ) − kPA(x)k > β j ∈ Ij for all 1 ≤ j ≤ n such that µ(Ij)(cid:19) (3) Xj=1 j))(cid:13)(cid:13)(cid:13)(cid:13) Therefore, E := [0, 1] \Sn∈N E 1 m ∈ N. Since t /∈ E 1 t′ − t < δm and every set A ∈ Σ with λ(A) < δm, kPA(f (t) − f (t′))k ≤ > β(cid:18) n Xj=1 is measurable with µ(E) = 1. Fix t ∈ E and , there exists δm > 0 such that for every t′ ∈ [0, 1] with which contradicts the inequality (2). So we can conclude that µ∗(Eβ) = 0. ≥ βµ∗(Eβ), 1 m . (cid:13)(cid:13)(cid:13)(cid:13) m n Thus, for every m ∈ N, every sequence (tn)∞ with λ(A) < δm, n=1 converging to t and every A ∈ Σ 1 m Now the conclusion follows from (4). kPA(f (tn))k ≤ kPA(f (t))k + for n big enough depending only on m. (cid:3) Let {Xi}i∈Γ be a family of Banach spaces. We denote by πj : (Li∈Γ Xi) → Xj the canonical projection onto Xj for each j ∈ Γ. 2.11 and 2.12: We will need the following property of ℓ1-sums and the space L1(λ) for Theorems RIEMANN INTEGRABILITY VERSUS WEAK CONTINUITY 9 Proof. The second part is essentially Lemma 2 of [28]. The proof of the first part is x, y ∈ (Li∈Γ Xi)ℓ1 and any disjoint sets A, B ⊂ Γ. disjoint sets A, B ∈ Σ. Lemma 2.10. Let (Ω, Σ, λ) be a probability space and {Xi}i∈Γ a family of Banach spaces. Then: (1) max{kx + yk,kx − yk} ≥ Pi∈A kπi(x)k +Pi∈B kπi(y)k for every vectors (2) max{kf + gk,kf − gk} ≥RA fdλ +RB gdλ for any f, g ∈ L1(λ) and any analogous and we include it for the sake of completeness. Let x, y ∈ (Li∈Γ Xi)ℓ1, A, B ⊂ Γ be disjoint sets and ε > 0. Without loss of generality, we may assume that A = {an : n ∈ N} and B = {bn : n ∈ N} are countable subsets. Consider the func- i )ℓ∞ defined by x∗(u) = Pi∈A x∗ tionals x∗, y∗ ∈ (Li∈Γ Xi)∗ i (πi(u)) i (πi(u)) for every u ∈ (Li∈Γ Xi)ℓ1, where each x∗ and y∗(u) = Pi∈B y∗ i ∈ X ∗ i satisfies kx∗ i k = 1, x∗ i k = ky∗ i (πi(y)) = kπi(y)k if i = bn. Then, since A, B are disjoint, kx∗ + y∗k = kx∗ − y∗k = 1. Therefore, kx + yk + kx − yk ≥ hx + y, x∗ + y∗i + hx − y, x∗ − y∗i = 2hx, x∗i + 2hy, y∗i = ℓ1 = (Li∈Γ X ∗ i (πi(x)) = kπi(x)k if i = an and y∗ i , y∗ = 2(cid:18)Xi∈A x∗ i (πi(x)) +Xi∈B y∗ i (πi(y))(cid:19) = 2(cid:18)Xi∈A kπi(x)k +Xi∈B kπi(y)k(cid:19), so max{kx + yk,kx − yk} ≥Pi∈A kπi(x)k +Pi∈B kπi(y)k. Theorem 2.11. Let {Xi}i∈N be Banach spaces with the WLP. Then the space X := (Li∈N Xi)ℓ1 has the WLP. Proof. We are going to apply Lemma 2.9. Take Ω := N, Σ := P(N) the power set of N, λ(A) :=Pn∈A 2−n and P = {PA : A ∈ Σ} with if i ∈ A if i /∈ A πi(PA(x)) =(πi(x) 0 (cid:3) for every A ∈ Σ and every x ∈ X. Property (3) of Lemma 2.9 is Lemma 2.10(1) and property (4) holds because if λ(A) < 1 2n , then A ⊂ {n, n + 1, . . .}, so kPA(x)k =Xi∈A kπi(x)k ≤Xi≥n kπi(x)k for every x ∈ X. Therefore, we can apply Lemma 2.9, so there exists a measurable set E ⊂ [0, 1] with µ(E) = 1 such that for every sequence (tn)∞ n=1 in [0, 1] converging to some t ∈ E the set {f (tn) : n ∈ N} is P-uniformly integrable. We can assume that, for each i ∈ N, the map t 7→ πi(f (t)) is weakly continuous at each point of E because each Xi has the WLP. n=1 in X converges weakly to x ∈ X It is a well known fact that a sequence (xn)∞ if and only if it satisfies the following two conditions: (i) πi(xn) → πi(x) weakly in Xi for every i ∈ N; (ii) for every ε > 0 there is a finite set J ⊆ N such that supn∈N kPN\J (xn)k ≤ ε. Since P-uniform integrability is equivalent to (ii), it follows that f is weakly (cid:3) continuous at each point of E. A similar idea to that of Theorem 2.11 let us prove the following theorem, which improves [28, Theorem 5] and [5, Proposition 2.10]. Theorem 2.12. Let (Ω, Σ, λ) be a probability space. 10 GONZALO MART´INEZ-CERVANTES (1) If dens(L1(λ)) < cov(M) then L1(λ) has the WLP. (2) If dens(L1(λ)) ≥ non(SN ) then L1(λ) does not have the WLP. Proof. Fix a Riemann integrable function f : [0, 1] → L1(λ). Take PA(x) := xχA for every A ∈ Σ and every x ∈ L1(λ). The family of operators {PA : A ∈ Σ} fulfills the requirements of Lemma 2.9 (bear in mind Lemma 2.10). Then P-uniform integrability is the usual uniform integrability and therefore a set is bounded and P-uniformly integrable if and only if it is relatively weakly compact due to Dunford's Theorem (see [1, Theorem 5.2.9]). Lemma 2.9 ensures that there exist a measurable n=1 in [0, 1] converging set E ⊂ [0, 1] with µ(E) = 1 such that for every sequence (tn)∞ to some t ∈ E, the set {f (tn) : n ∈ N} is relatively weakly compact. Let C ⊂ Σ be a dense family of λ-measurable sets, i.e. such that inf C∈C λ(A △ C) = 0 for every A ∈ Σ. Let (hn)∞ n=1 be a relatively weakly compact sequence in L1(λ) and h ∈ L1(λ). Since C is a dense family of λ-measurable sets, if RC hn dµ → RC h dµ for every C ∈ C, then h = ω-lim hn. Suppose dens(L1(λ)) < cov(M). Then C can be taken such that C < cov(M). Therefore, we can assume that, for each C ∈ C, the Riemann integrable map t 7→RC f (t) dλ is continuous at each point of E. Then, for every sequence (tn)∞ in [0, 1] converging to a point t ∈ E, we have f (t) = ω-lim f (tn). Now suppose ν = dens(L1(λ)) ≥ non(SN ). Due to Maharam's Theorem (see [16, p. 127, Theorem 9]), L1(λ) contains an isometric copy of L1(µν), where µν is the usual product probability measure on {0, 1}ν. Since L1(µν ) contains an isomorphic copy of ℓ2(ν) (see [16, p. 128, Theorem 12]) and ℓ2(ν) does not have the WLP (Theorem 2.8), we conclude that L1(λ) does not have the WLP. (cid:3) n=1 Theorem 2.11 can be extended to arbitrary ℓ1-sums: Theorem 2.13. The arbitrary ℓ1-sum of a family of Banach spaces with the WLP has the WLP. Proof. The proof uses some ideas of [18]. Let f : [0, 1] → X := (Li∈Γ Xi)ℓ1 be a Riemann integrable function, where {Xi}i∈Γ is a family of Banach spaces with the WLP. For each J ⊂ Γ, we denote by PJ : X → X the function defined by πi (PJ (x)) = πi(x) if i ∈ J and πi (PJ (x)) = 0 in any other case. Let (rn)∞ n=1 be an enumeration of the rational numbers in [0, 1] and fix a countable set L ⊂ Γ such that PL(f (rn)) = f (rn) for every n ∈ N. Then, f = (f − PLf ) + PLf . Since PLf is Riemann integrable and takes values in the space XL := {x ∈ X : πi(x) = 0 for each i /∈ L}, which is isomorphic to a countable ℓ1-sum of spaces with the WLP, by Theorem 2.11 PLf is weakly continuous almost everywhere. 0 f (t)dt = 0 and that f is null over a dense set. Let Therefore, we can assume that R 1 for each n ∈ N and each subset J ⊂ Γ. If J1 ⊂ J2 ⊂ Γ, then AJ2 Claim: For every n ∈ N there exists a countable set J ⊂ Γ with µ(cid:16)AJ Suppose this is not the case. Then, there exist n ∈ N and δ > 0 with µ(cid:16)AJ AJ n := {t ∈ [0, 1] : kPJ c (f (t))k ≥ n ⊂ AJ1 n . 1 n} n(cid:17) = 0. n(cid:17) > δ RIEMANN INTEGRABILITY VERSUS WEAK CONTINUITY 11 N < (4) δ n (cid:13)(cid:13)(cid:13)(cid:13) Xj=1 for all choices ξj, ξ′ for every countable subset J ⊂ Γ (if for every m ∈ N we can take a countable n(cid:17) = 0). Let set Jm ⊂ Γ with µ(cid:16)AJm P = {I1, I2, . . . , IN} be a partition of [0, 1] such that m , then J = Sm∈N Jm verifies µ(cid:16)AJ n (cid:17) < 1 j))(cid:13)(cid:13)(cid:13)(cid:13) µ(Ij)(f (ξj ) − f (ξ′ j ∈ Ij, 1 ≤ j ≤ N. n(cid:17) = µ(cid:16)AJ Let J ⊂ Γ be a countable subset. Since null over a dense set, we can suppose that there exist ξ1 ∈ Int(I1)∩ AJ such that kµ(I1)(f (ξ1) − f (ξ′ (4) we have µ(I1) < δ < PN Int(I2) ∩ AJ1 kµ(I1)(f (ξ1) − f (ξ′ n(cid:17) > δ and f is Pj=1 1 ∈ I1 1))k ≥ 1 1). By n (cid:17) and so it is not restrictive to suppose j=1 µ(cid:16)Ij ∩ AJ1 2 ∈ I2 such that 1)) + µ(I2)(f (ξ2) − f (ξ′ n and ξ′ n µ(I1). Let J1 = supp f (ξ1) ∪ supp f (ξ′ n 6= ∅. Thus, due to Lemma 2.10, we can choose ξ2, ξ′ µ(cid:16)Ij ∩ AJ (µ(I1) + µ(I2)). 2))k ≥ 1 n N Fix 1 ≤ k < N and assume that we have already chosen points ξj , ξ′ 1 ≤ j ≤ k with the property that j ∈ Ij for all N (cid:13)(cid:13)(cid:13)(cid:13) Xj=1 µ(Ij)(f (ξj ) − f (ξ′ δ n . ≥ j))(cid:13)(cid:13)(cid:13)(cid:13) But this is a contradiction with (4). Therefore, the Claim is proved. n (cid:17) = 0. Fix Thus, for every n ∈ N there exists a countable set Jn such that µ(cid:16)AJn J :=Sn∈N Jn. Theorem 2.11 guarantees the existence of a set F ⊂ [0, 1] of measure one such that PJ (f ) is weakly continuous at every point of F . Let E = F \Sn∈N AJ n. Then, µ(E) = 1, f = PJ (f ) + PJ c(f ), PJ (f ) is weakly continuous at each point of E and PJ c (f ) is norm continuous at each point of E (if tn → t ∈ E, then, for every m ∈ N, tn /∈ AJ Corollary 2.14 ([24, 20]). ℓ1(κ) has the LP for any cardinal κ. m for n big enough so kPJ c(f )(tn)k < 1 m ). (cid:3) Proof. Since ℓ1(κ) has the Schur property, ℓ1(κ) has the LP if and only it has the WLP. Therefore, the conclusion follows from Theorem 2.13. (cid:3) µ(Ij )(f (ξj ) − f (ξ′ ≥ k 1 n Xj=1  µ(Ij )  . j ))(cid:13)(cid:13)(cid:13)(cid:13) j=1 supp f (ξj) ∪ supp f (ξ′ j), which is countable. By (4) we have j=1 µ(cid:16)Ij ∩ AJk n (cid:17), hence it is not restrictive to suppose that k+1 ∈ Ik+1 such k+1 n 6= ∅ and therefore that there exist points ξk+1, ξ′ µ(Ij )  . µ(Ij )(f (ξj ) − f (ξ′ n  ≥ k+1 1 j ))(cid:13)(cid:13)(cid:13)(cid:13) µ(Ij ) = 1 > δ, it follows that there exist ξj, ξ′ Xj=1 j ∈ Ij for every 1 ≤ j ≤ N k k (cid:13)(cid:13)(cid:13)(cid:13) Xj=1 Set Jk := Sk µ(Ij ) < δ < PN Pj=1 Int(Ik+1) ∩ AJk that (cid:13)(cid:13)(cid:13)(cid:13) Xj=1 N Since Pj=1 such that 12 GONZALO MART´INEZ-CERVANTES As an application of 2.13 we also obtain the following result: Corollary 2.15. Let K be a compact Hausdorff space. Then, C(K)∗ has the WLP if dens(L1(λ)) < cov(M) for every regular Borel probability λ on K. Proof. For every compact Hausdorff space K, the Banach space C(K)∗ is isometric to a ℓ1-sum of L1(λ) spaces, where each λ is a regular Borel probability measure on K (see the proof of [1, Proposition 4.3.8]). Thus, C(K)∗ has the WLP if each space L1(λ) has the WLP, due to Theorem 2.13. Hence, the result follows from Theorem 2.12. (cid:3) Corollary 2.16. If K is a compact Hausdorff space in the class M S (i.e. L1(λ) is separable for every regular Borel probability on K), then C(K)∗ has the WLP. Some classes of compact spaces in the class M S are metric compacta, Eber- lein compacta, Radon-Nikod´ym compacta, Rosenthal compacta and scattered com- pacta. For more details on this class, we refer the reader to [8], [17] and [25]. The LP is a three-space property, i.e. if X is a Banach space and Y is a subspace of X such that Y and X/Y have the LP, then X has the LP [24, Proposition 1.19]. This result follows from Michael's Selection Theorem. However, as far as we are concerned, it is not known whether the WLP is a three-space property. We have a positive result in the following case: Theorem 2.17. Let X be a Banach space and Y a subspace of X. If Y is reflexive, dens(Y ) < cov(M) and X/Y has the WLP, then X has the WLP. Proof. Let Q : X → X/Y be the quotient operator and φ : X/Y → X be a norm- norm continuous map such that Qφ = 1X/Y given by Michael's Selection Theorem (see [10, Section 7.6]). Let f : [0, 1] → X be a Riemann integrable function. Then, since Qf is Riemann integrable and X/Y has the WLP, there exists a set E ⊂ [0, 1] with µ(E) = 1 such that Qf is weakly continuous at every point of E. Set (5) C = {x ∈ X : ∃ (tn)∞ n=1 converging to some t ∈ E with x = ω- lim f (tn)}. Let {x∗ First we are going to see that dens(C) < cov(M). Let x ∈ C and (tn)∞ n=1 as in (5). Then Qx = ω-lim Qf (tn) = Qf (t). Therefore, x = φ(Qx) + (x − φ(Qx)) with φ(Qx) ∈ φ(Qf (E)) and x− φ(Qx) ∈ Y . Notice that φ(Qf (E)) is separable because of the ω-separability of Qf (E) and Mazur's Lemma. Thus, C ⊂ φ(Qf (E)) + Y satisfies dens(C) < cov(M). α}α∈Γ ⊂ X ∗ be a set separating points of C with Γ < cov(M). Set E0 ⊂ E with µ(E0) = 1 such that x∗ α ◦ f is continuous at every point of E0 for every α ∈ Γ. Notice that this can be done because the set of discontinuity points of each x∗ α ◦ f is an Fσ Lebesgue null set and Γ < cov(M). We claim that f is weakly continuous at each point of E0. Let t ∈ E0 and (tn)∞ n=1 be a sequence converging to t. Since Qf (t) = ω-lim Qf (tn), the set {Qf (tn) : n ∈ N} is relatively weakly compact in X/Y . From the reflexivity of Y , it follows that Q is a Tauberian operator, so {f (tn) : n ∈ N} is relatively weakly compact in X (see [11, Theorem 2.1.5 and Corollary 2.2.5]). Therefore, it is enough to prove the uniqueness of the limit of the subsequences of (f (tn))∞ f (tnk ). Then, x, f (t) ∈ C and x∗ n=1. Let x = ω-lim k α(x) = lim k x∗ α(f (tnk )) = x∗ α(f (t)) for every α ∈ Γ, so x = f (t). (cid:3) RIEMANN INTEGRABILITY VERSUS WEAK CONTINUITY 13 3. Weak continuity does not imply integrability It is not true that every weakly continuous function is Riemann integrable [2]. In fact, V. Kadets proved the following theorem: Theorem 3.1 ([15]). If X is a Banach space without the Schur property, then there is a weakly continuous function f : [0, 1] → X which is not Riemann integrable. The proof of the previous theorem together with Josefson-Nissenzweig Theorem (see [7, Chapter XII]) gives the following corollary: Corollary 3.2. Given an infinite-dimensional Banach space X, there always exists a weak* continuous function f : [0, 1] → X ∗ which is not Riemann integrable. In [29], Wang and Yang extend the previous result to a general locally convex topology weaker than the norm topology. In this section, we generalize these results in Theorem 3.4. Following the terminology used in [9], we say that a subset M of a Banach space is spaceable if M ∪ {0} contains a closed infinite-dimensional subspace. We start with the definitions of τ -Dunford-Pettis operator and the τ -Schur prop- erty, that coincide with the classical definitions of Dunford-Pettis or completely continuous operator and the Schur property when τ is the weak topology. Definition 3.3. Let X and Y be Banach spaces and τ a locally convex topology on X weaker than the norm topology. An operator T : X → Y is said to be τ - Dunford-Pettis (τ -DP for short) if it carries bounded τ -null sequences to norm null sequences. A Banach space X is said to have the τ -Schur property if the identity operator I : X → X is τ -DP. Theorem 3.4. Let X and Y be Banach spaces and τ be a locally convex topology on X weaker than the norm topology. If T : X → Y is an operator which is not τ -DP, then the family of all bounded τ -continuous functions f : [0, 1] → X such that T f is not Riemann integrable is spaceable in ℓ∞([0, 1], X), the space of all bounded functions from [0, 1] to X with the supremum norm. Proof. The proof uses ideas from [15]. Since T is not τ -DP, we can take a bounded sequence (xn)∞ n=1 that is τ -convergent to zero such that kT xnk = 1 for all n ∈ N. Let K ⊂ [0, 1] be a copy of the Cantor set constructed by removing from [0, 1] an 2 , . . . I n 1 in the middle of [0, 1] and removing open intervals I n open interval I 1 2n from the middles of the remaining intervals in each step. Suppose that the removed intervals are so small that µ(K) > 2 3 . Let Ca([0, 1]) be the closed subspace of C([0, 1]) consisting of all continuous functions g : [0, 1] → R antisymmetric with respect to the axe x = 1 2 and with g(0) = g(1) = 0. For every g ∈ Ca([0, 1]) and every open interval I = (a, b) in [0, 1], we define the functions gI : [0, 1] → R and fg : [0, 1] → X as follows 1 , I n The function φ : Ca([0, 1]) → ℓ∞([0, 1], X) given by the formula φ(g) := fg for every g ∈ Ca([0, 1]) is a linear map and satisfies kφ(g)k = (supn kxnk)kgk for every g( t−a b−a ) gI (t) =( 0 fg(t) =( 0 if t /∈ (a, b), if t ∈ [a, b]. if t ∈ K, if t ∈ I n k . gI n k (t)xn 14 GONZALO MART´INEZ-CERVANTES g ∈ Ca([0, 1]). Therefore, φ is a multiple of an isometry. Thus, V := φ(Ca([0, 1]) is an infinite-dimensional closed subspace of ℓ∞([0, 1], X). We are going to check that each function fg 6= 0 is τ -continuous but T fg is not τ−→ 0, fg is Riemann integrable. Since g is continuous, g(0) = g(1) = 0 and xn τ -continuous. Suppose T fg is Riemann integrable. Then, y∗(cid:18)Z 1 0 T fg(t)dt(cid:19) =Z 1 0 y∗T fg(t)dt =Xk,n y∗(T xn)ZI n k gI n k (t)dt = 0 N for each y∗ ∈ Y ∗. The only possible value for the Riemann integral of T fg is 0 due to the above equality. Choose a partition P = {J1, J2, . . . , JN} of [0, 1]. Let Int Jj ∩ K 6= ∅}. We can take m ∈ N such that if j ∈ A A = {j : 1 ≤ j ≤ N, then Jj contains some interval I m k . Hence, if j ∈ A, there is tj ∈ Jj such that fg(tj) = kgkxm. If j /∈ A, then we pick any tj ∈ Int Jj. From the inequality Pj∈A µ(Jj) ≥ µ(K) > 2 µ(Jj)T fg(tj)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13) Xj=1 kgkµ(Jj)T xm(cid:13)(cid:13)(cid:13)(cid:13) −(cid:13)(cid:13)(cid:13)(cid:13)Xj /∈A 1 3kgk. Then, T fg is Riemann integrable if and only if g = 0 if and only if fg = 0. (cid:3) µ(Jj)T fg(tj)(cid:13)(cid:13)(cid:13)(cid:13) ≥(cid:13)(cid:13)(cid:13)(cid:13)Xj∈A t∈[0,1]kT fg(t)k = sup µ(Jj)T fg(tj ) +Xj /∈A 2 3kgk − =(cid:13)(cid:13)(cid:13)(cid:13)Xj∈A µ(Jj)T fg(tj)(cid:13)(cid:13)(cid:13)(cid:13) > 1 3 3 , we deduce ≥ The next corollary gives an affirmative answer to a question posed by Sofi in [26]. Corollary 3.5. Given an infinite-dimensional Banach space X, the set of all weak* continuous functions f : [0, 1] → X ∗ which are not Riemann integrable is spaceable in ℓ∞([0, 1], X ∗). Proof. X ∗ is not ω∗-Schur for any infinite-dimensional Banach space X due to the Josefson-Nissenzweig Theorem. Thus, the conclusion follows from Theorem 3.4. (cid:3) Given a Banach space X, a function f : [0, 1] → X is said to be scalarly Riemann integrable if every composition x∗f with x∗ ∈ X ∗ is Riemann integrable. We can also characterize Dunford-Pettis operators thanks to Theorem 3.4. The equivalence (1) ⇔ (3) in the following corollary was mentioned without proof in [23]. Corollary 3.6. Let X and Y be Banach spaces and T : X → Y be an operator. The following statements are equivalent: (1) T is Dunford-Pettis. (2) T f is Riemann integrable for every ω-continuous function f : [0, 1] → X. (3) T f is Riemann integrable for every scalarly Riemann integrable function f : [0, 1] → X. Proof. (2) ⇒ (1) is a consequence of Theorem 3.4. Since every ω-continuous func- tion f : [0, 1] → X is scalarly Riemann integrable, (3) implies (2). Therefore, it remains to prove (1) ⇒ (3). Suppose T is Dunford-Pettis and fix (Pn)∞ n=1 a sequence of tagged partitions of [0, 1] with kPnk n−→ 0. Let f : [0, 1] → X be a scalarly Riemann integrable function. Then, x∗f (Pn) n−→ R 1 0 x∗f (t)dt for every RIEMANN INTEGRABILITY VERSUS WEAK CONTINUITY 15 x∗ ∈ X ∗. Thus, f (Pn) is a ω-Cauchy sequence in X, so T f (Pn) is norm convergent to some y ∈ Y . The limit y does not depend on the sequence of tagged partitions, nk n−→ 0, then n)∞ since if (P ′ n=1 is any other sequence of tagged partitions with kP ′ n)k n−→ 0. f (Pn)−f (P ′ n) is weakly null and this in turn implies that kT f (Pn)−T f (P ′ Thus, T f is Riemann integrable. (cid:3) Acknowledgements I would like to thank Antonio Avil´es and Jos´e Rodr´ıguez the useful suggestions and the help provided in some proofs of this paper. References [1] F. Albiac and N. J. Kalton. Topics in Banach Space Theory. Graduate texts in Mathematics. Springer, New York, London, 2006. [2] A. Alexiewicz and W. Orlicz. Remarks on Riemann-integration of vector-valued functions. Studia Math., 12:125 -- 132, 1951. [3] T. Bartoszynski and H. Judah. Set Theory: On the Structure of the Real Line. Ak Peters Series. Taylor & Francis, 1995. [4] T. Bartoszynski and S. Shelah. Closed measure zero sets. Ann. Pure Appl. Logic, 58(2):93 -- 110, 1992. [5] J. M. Calabuig, J. Rodr´ıguez, and E. A. S´anchez-P´erez. Weak continuity of Riemann inte- grable functions in Lebesgue-Bochner spaces. Acta Math. Sin. (Engl. Ser.), 26(2):241 -- 248, 2010. [6] G. C. da Rocha Filho. Integral de Riemann Vetoral e Geometria de Espa¸cos de Banach. PhD thesis, Universidade de Sao Paulo, 1979. [7] J. Diestel. Sequences and series in Banach spaces. Graduate texts in mathematics. Springer- Verlag, 1984. [8] M. Dzamonja and K. Kunen. Properties of the class of measure separable compact spaces. Fundam. Math., 147(3):261 -- 277, 1995. [9] P. Enflo, V. I. Gurariy, and J. B. Seoane-Sep´ulveda. Some results and open questions on spaceability in function spaces. Trans. Am. Math. Soc., 366(2):611 -- 625, 2014. [10] M. Fabian, P. Habala, P. H´ajek, V. Montesinos, and V. Zizler. Banach Space Theory: The Basis for Linear and Nonlinear Analysis. CMS Books in Mathematics. Springer New York, 2011. [11] M. Gonz´alez and A. Mart´ınez-Abej´on. Tauberian Operators. Operator Theory: Advances and Applications. Birkhauser Basel, 2010. [12] R. Gordon. Riemann Integration in Banach Spaces. Rocky Mountain J. Math., 21(3):923 -- 949, 1991. [13] L. M. Graves. Riemann integration and Taylor's theorem in general analysis. Trans. Amer. Math. Soc., 29(1):163 -- 177, 1927. [14] R. Haydon. Darboux integrability and separability of types in stable Banach spaces. S´emin. analyse fonctionnelle, Paris 1983-84, Publ. Math. Univ. Paris VII 20, 95 -- 115, 1984. [15] V. M. Kadets. On the Riemann integrability of weakly continuous functions. Quaest. Math., 17(1):33 -- 35, 1994. [16] H. E. Lacey. The Isometric Theory of Classical Banach Spaces. Grundlehren der mathema- tischen Wissenschaften. Springer Berlin Heidelberg, 2012. [17] W. Marciszewski and G. Plebanek. On measures on Rosenthal compacta. J. Math. Anal. Appl., 385(1):185 -- 193, 2012. [18] F. Miraglia and G. C. da Rocha Filho. The measurability of Riemann integrable functions with values in Banach spaces and applications. Sao Paulo Ime/usp, 1984. [19] K. M. Naralenkov. Asymptotic Structure of Banach Spaces and Riemann Integration. Real Anal. Exchange, 33(1):113 -- 126, 2007. [20] A. S. Nemirovski, M. Ju. Ochan, and R. Redjouani. Conditions for the Riemann integrability of functions with values in a Banach space. Mosc. Univ. Math. Bull., 27(3-4):124 -- 126, 1973. [21] E. Odell. A nonseparable Banach space not containing a subsymmetric basic sequence. Israel J. Math., 52(1 -- 2):97 -- 109, 1985. 16 GONZALO MART´INEZ-CERVANTES [22] J. Pawlikowski. A characterization of strong measure zero sets. Israel J. Math., 93(1):171 -- 183, 1996. [23] A. Pelczynski and G. C. da Rocha Filho. Operadores de Darboux. Seminario Brasileiro de Analise, 12, Sao Jose dos Campos, 1980. Trabalhos Apresentados. Rio de Janeiro, SBM, 1980. [24] M. Pizzotti. Darboux-integrabilidade e mensurabilidade de fun¸coes Riemann-integr´aveis definidas em compactos. PhD thesis, Universidade de Sao Paulo, 1989. [25] G. Plebanek and D. Sobota. Countable tightness in the spaces of regular probability measures. Fundam. Math., 229(2):159 -- 169, 2015. [26] M. A. Sofi. Weaker forms of continuity and vector-valued Riemann integration. Colloq. Math., 129(1):1 -- 6, 2012. [27] C. Wang. On the weak property of Lebesgue of Banach spaces. Journal of Nanjing University Mathematical Biquarterly (English Series), 13(2):150 -- 155, 1996. [28] C. Wang and K. Wan. On the Weak Property of Lebesgue of L1(Ω, Σ, µ). Rocky Mountain J. Math., 31(2):697 -- 703, 2001. [29] C. Wang and Z. Yang. Some Topological Properties of Banach Spaces and Riemann Integra- tion. Rocky Mountain J. Math., 30(1):393 -- 400, 2000. Departamento de Matem´aticas, Facultad de Matem´aticas, Universidad de Murcia, 30100 Espinardo (Murcia), Spain E-mail address: [email protected]
1512.08741	2	1512	2016-09-30T14:33:08	Ideal structures in vector-valued polynomial spaces	[ "math.FA" ]	This paper is concerned with the study of geometric structures in spaces of polynomials. More precisely, we discuss for $E$ and $F$ Banach spaces, whether the class of weakly continuous on bounded sets $n$-homogeneous polynomials, $\mathcal P_w(^n E, F)$, is an HB-subspace or an $M(1,C)$-ideal in the space of continuous $n$-homogeneous polynomials, $\mathcal P(^n E, F)$. We establish sufficient conditions under which the problem can be positively solved. Some examples are given. We also study when some ideal structures pass from $\mathcal P_w(^n E, F)$ as an ideal in $\mathcal P(^n E, F)$ to the range space $F$ as an ideal in its bidual $F^{**}$.	math.FA	math	IDEAL STRUCTURES IN VECTOR-VALUED POLYNOMIAL SPACES VER ´ONICA DIMANT1, SILVIA LASSALLE1, ´ANGELES PRIETO2 Abstract. This paper is concerned with the study of geometric structures in spaces of polynomials. More precisely, we discuss for E and F Banach spaces, whether the class of n-homogeneous polynomials, Pw(nE, F ), which are weakly continuous on bounded sets, is an HB -subspace or an M (1, C)-ideal in the space of continuous n-homogeneous polynomials, P(nE, F ). We establish sufficient conditions under which the problem can be positively solved. Some examples are given. We also study when some ideal structures pass from Pw(nE, F ) as an ideal in P(nE, F ) to the range space F as an ideal in its bidual F ∗∗. 6 1 0 2 p e S 0 3 ] . A F h t a m [ 2 v 1 4 7 8 0 . 2 1 5 1 : v i X r a 1. Introduction Let X be a (real or complex) Banach space and let J be a closed subspace of X. According to the Hahn -- Banach theorem, every continuous linear functional g ∈ J ∗ has an extension f ∈ X ∗, with the same norm. A long standing problem is to determine when every functional on J has a unique norm-preserving extension to X. This question is closely related to geometric properties of both spaces which, in many cases, imply the existence of a norm-one projection on X ∗ whose kernel is J ⊥ := {x∗ ∈ X ∗ : x∗(y) = 0, for all y ∈ J}, the annihilator of J. When there exists such a projection J is said to be an ideal in X. A canonical example of this fact is that X is always an ideal in its bidual X ∗∗. The notion of M-ideal, introduced by Alfsen and Effros and widely studied in the book by Harmand, Werner and Werner [18] is one of these geometric properties ensuring unique Hahn -- Banach extensions. Recall that J is an M - ideal in X if it is an ideal in X with associated projection q such that for each f ∈ X ∗ one has kf k = kqf k + kf − qf k. The fact that J is an M-ideal in X has a strong impact on both J and X, and sometimes seems to be too restrictive. So, we will be interested in studying some weaker properties among those implying unique norm-preserving extensions. Recall that a closed subspace J is HB -smooth in X if every element in J ∗ has a unique norm-preserving extension to an element in X ∗. A closed subspace J is strongly HB -smooth in X if there exists a linear projection q on X ∗ whose kernel is J ⊥ such that for each f ∈ X ∗ with f 6= qf one has kqf k < kf k. 2010 Mathematics Subject Classification. Primary 46G25; Secondary 47H60, 46B04, 47L22. Key words and phrases. HB -subspaces, homogeneous polynomials, weakly continuous on bounded sets polynomials. 1 2 V. DIMANT, S. LASSALLE, A. PRIETO The interplay between uniqueness of the extension and strong HB -smoothness was clarified by Oja [22]. Namely, the uniqueness of the extensions and being an ideal are independent notions for a subspace J, strong HB -smoothness implies both, and if J is an HB -smooth ideal in X then J is strongly HB -smooth in X. A particular case of HB -smoothness is the notion of HB -subspace, introduced by Hennefeld [19]. A closed subspace J is an HB-subspace of X if there exists a projection q on X ∗ whose kernel is J ⊥ such that for each f ∈ X ∗ with f 6= qf one has kqf k < kf k and kf − qf k ≤ kf k. Finally, given C ∈ (0, 1], a closed subspace J is an M(1, C)-ideal in X if J is an ideal of X with associated projection q on X ∗ such that for each f ∈ X ∗ one has kqf k + Ckf − qf k ≤ kf k. The last inequality is called the M(1, C)-inequality. Note that when C = 1, the notion of M-ideal is covered and to be M(1, C)-ideal immediately implies strong HB -smoothness. However, the notions of M(1, C)-ideal and HB -subspace are independent. On the one hand, Cabello and Nieto [6, Example 3.7] showed that if X is a nonreflexive separable M-ideal in its bidual, then ℓp(X) as a subspace of its bidual, 1 < p < ∞, is an HB -subspace that cannot be renormed to be an M(1, C)-ideal for any 0 < C < 1. On the other hand, Cabello, Nieto and Oja [7, Example 4.3] showed that for any 0 < C < 1, there is a renorming of c0, c0 due to Johnson and Wolfe, such that the space of compact operators on c0, K(c0) is an M(1, C)-ideal in the space of all continuous operators L(c0) without being an HB -subspace. Several authors have been interested in this kind of properties for arbitrary subspaces of Banach spaces, and also for distinguished particular cases. The space of compact operators K(E, F ) between Banach spaces E and F as a subspace of the space of all continuous linear operators L(E, F ) received special interest (see, for example, [6, 7, 19, 21, 22, 23, 24, 26]). The strongest of the abovementioned properties is the one of being an M-ideal. All other properties which are more flexible still allow us to deal with uniqueness of Hahn -- Banach extensions. Here, we will be concerned with P(nE, F ), the space of continuous n-homogen- eous polynomials between Banach spaces E and F . In the polynomial context, the space of compact mappings is usually replaced by Pw(nE, F ) the subspace of homogeneous polynomials which are weakly continuous on bounded sets. Recall that a polynomial P ∈ P(nE, F ) is in Pw(nE, F ) if it maps bounded weakly convergent nets into convergent nets. Note that we could have considered poly- nomials in P(nE, F ) mapping bounded sets into relatively compact sets, which are called compact polynomials. For linear operators to be compact and to be weakly continuous on bounded sets are equivalent notions. For n-homogeneous polynomials with n > 1, every polynomial in Pw(nE, F ) is compact (as can be derived from results in [3] and [4]), but the converse might not be true. Every scalar-valued continuous polynomial is compact but it is not necessarily weakly continuous on bounded sets. The prototypical example of this situation is given IDEAL STRUCTURES IN VECTOR-VALUED POLYNOMIAL SPACES 3 by P (x) = Pk x2 k, for all x = (xk)k ∈ ℓ2. Therefore, we will focus our atten- tion on determining the presence of ideal structures for Pw(nE, F ) as a subspace of P(nE, F ). To be more precise, our main concern is to study the notion of HB -subspace in the polynomial setting. Some previous results in this direction can be found in [12], where the problem of determining when Pw(nE) is an M-ideal in P(nE) was considered. A vector- valued approach of the same question was treated in [14]. Note that the searching of ideal structures for Pw(nE, F ) as a subspace of P(nE, F ) makes sense when the spaces Pw(nE, F ) and P(nE, F ) do not coincide. The equality Pw(nE, F ) = P(nE, F ) is a long standing nontrivial problem, considered for instance in [1, 5, 16, 17]. The plan of the paper is as follows. First, we review the notation and the basic facts that will be used in Sections 3 and 4. Then, in Section 3, we investigate sufficient conditions under which the subspace Pw(nE, F ) enjoys an additional geometric structure inside P(nE, F ) and we exhibit some particular examples. In the last section, we study some ideal structures for the range space F as a subspace of F ∗∗, when they are fulfilled by Pw(nE, F ) as a subspace of P(nE, F ). 2. Notation and basic facts Before proceeding, we fix some notation. Every time we write E or F we will be considering Banach spaces over the real or complex field, K. The closed unit ball of E will be denoted by BE and the unit sphere by SE. As usual, E∗ and E∗∗ stand for the dual and bidual of E, respectively. The space of linear bounded operators from E to F will be denoted by L(E, F ) (and L(E) when E = F ); its subspace of compact mappings will be denoted by K(E, F ) (K(E) in the case E = F ). A function P : E → F is an n-homogeneous polynomial if there exists a (unique) symmetric n-linear form A : E × · · · × E → F such that n {z } P (x) = A(x, . . . , x), for all x ∈ E. The space of all continuous n-homogeneous polynomials from E to F , P(nE, F ), endowed with the supremum norm kP k = sup{kP (x)k : x ∈ BE}, is a Banach space. L(b⊗ Every polynomial P in P(nE; F ) can be associated with a linear operator in n,s πs E; F ), where πs is the symmetric projective tensor norm. We will identify P with its linearization without further mention. Even though this identification preserves the norm, there is no Hahn -- Banach Theorem for homogeneous poly- nomials of degree 2 or greater. However, Aron and Berner [2] and Davie and Gamelin [9] showed that for every P ∈ P(nE, F ) there is a norm-preserving ex- tension of P to P ∈ P(nE∗∗, F ∗∗) such that P (x) = P (x) for all x ∈ E. The construction of this canonical extension is based on the Arens extension of the 4 V. DIMANT, S. LASSALLE, A. PRIETO symmetric mapping A associated to the polynomial P . To obtain the Arens ex- tension, we simply extend by weak-star continuity, one variable at a time, the n variables of A. This process depends on the order that the variables are extended and the final result might not be a symmetric mapping. However, the n! possible extensions coincide on the diagonal and P is well defined. For the particular case in which P belongs to Pw(nE, F ), the range of P is also in F (as can be derived from [3] and [8, Proposition 2.5]). This fact will be used repeatedly in Section 4. In this paper, we will present several results in which at least one of the spaces involved enjoys the metric compact approximation property. Recall that a Ba- nach space E has the metric compact approximation property if there is a net of compact operators (Kα) on E such that Kα → IdE pointwise and supα kKαk ≤ 1. Usually, the net (Kα) is called a metric compact approximation of the identity. If in addition K ∗ α → IdE ∗ pointwise, the net (Kα) is called a shrinking met- ric compact approximation of the identity. As usual, K α denotes the operator IdE − Kα. For dual spaces we have the following intermediate property. We say that E∗ has a metric compact approximation of the identity with adjoint operators if there exists a net (Kα) ⊂ K(E) such that K ∗ α converges to IdE ∗ pointwise and supα kKαk ≤ 1. These notions are closely related with ideal structures on Banach spaces. For instance, [21, Theorem 1.1] asserts that the following conditions are equivalent: (i) F has the metric compact approximation property (ii) K(E, F ) is an ideal in L(E, F ) for every Banach space E. So, it is natural to expect that the metric compact approximation property shows up when describing Pw(nE, F ) as an ideal in P(nE, F ). One further ingredient will appear in our discussion. In [10], Delpech obtained an appropriate connection between the moduli of asymptotic uniform smoothness and convexity and weak sequential continuity of polynomials. In [13], Dimant, Gonzalo and Jaramillo followed his approach to obtain results on compactness or weak-sequential continuity of multilinear mappings. Here, we will impose restrictions on the growth of the moduli of the underlying spaces E or F to ensure that Pw(nE, F ) enjoys an appropriate property in P(nE, F ) (see Theorem 3.4 and Proposition 3.13). Some definitions are in order. For an infinite dimensional Banach space E, the modulus of asymptotic point- wise smoothness is defined for kxk = 1 and t > 0 by ρE(t; x) = inf dim(E/H)<∞ sup kx + hk − 1, h∈H,khk≤t and the modulus of asymptotic uniform smoothness is defined for t > 0 by ρE(t) = sup kxk=1 ρE(t; x). The space E is asymptotically uniformly smooth if limt→0 = 0. ρE(t) t For an infinite dimensional Banach space, the modulus of asymptotic pointwise convexity is defined for kxk = 1 and t > 0 by δE(t; x) = sup dim(E/H)<∞ inf h∈H,khk≥t kx + hk − 1, IDEAL STRUCTURES IN VECTOR-VALUED POLYNOMIAL SPACES 5 and the modulus of asymptotic uniform convexity is defined for t > 0 by δE(t) = inf kxk=1 δE(t; x). The space E is asymptotically uniformly convex if δE(t) > 0, for every 0 < t ≤ 1. Finally, E has modulus of asymptotic uniform convexity of power p if there exists C > 0 such that δE(t) ≥ Ctp, for all 0 < t ≤ 1. We refer to [15] for the necessary background on polynomials on Banach spaces. 3. Sufficient conditions When working with polynomials, the lack of linearity provides, in many cases, difficulties that can be overcome not without certain detours. The value of n for which Pw(nE, F ) has the chance to be a nontrivial M-ideal in P(nE, F ) cannot be chosen arbitrarily. In fact, in the scalar-valued case [12] it was proved that, whenever P(mE) \ Pw(mE) 6= ∅ for some m, there exists a unique value n, called the critical degree, for which Pw(nE) can be a non-trivial M-ideal in P(nE). The critical degree of E is defined as cd(E) := min{k ∈ N : Pw(kE) 6= P(kE)}. In the vector-valued case, the critical degree is defined by analogy [14] as cd(E, F ) := min{k ∈ N : Pw(kE, F ) 6= P(kE, F )}, and the problem of whether Pw(nE, F ) is an M-ideal in P(nE, F ) is worth being studied only for polynomials of degree n with cd(E, F ) ≤ n ≤ cd(E). Although we are interested in studying ideal structures which are more flexible than to be an M-ideal, in order to show positive results we could not get rid of some restrictions on the degree of homogeneity. We start with a lemma which, under certain conditions on n, gives a version of Johnson's projection [20, Lemma 1.1] for the polynomial case. Lemma 3.1. Let E, F be Banach spaces and let n < cd(E). Suppose that F has the metric compact approximation property. Then Pw(nE, F ) is an ideal in P(nE, F ). Proof. Let (Lβ) ⊂ K(F ) be a metric compact approximation of the identity. As (Lβ) is bounded, there exists a subnet, still denoted by (Lβ), that converges w∗ to some L0 ∈ K(F )∗∗. Define Λ : P(nE, F )∗ → P(nE, F )∗ by Λ(f )(P ) = lim β f (Lβ ◦ P ). (3.1) Note that Λ is well defined. In fact, if P ∈ P(nE, F ) and τP : K(F ) → P(nE, F ) is the composition operator, τP (K) = K ◦ P , its transpose τ ∗ P satisfies that τ ∗ P (f ) ∈ K(F )∗, for any f ∈ P(nE, F )∗. So lim β f (Lβ ◦ P ) = lim β τ ∗ P (f )(Lβ) = L0(τ ∗ P (f )). It is clear that Λ is linear and kΛk ≤ 1. It is also a projection: since any P ∈ Pw(nE, F ) is compact and (Lβ) converges to the identity on compact sets 6 V. DIMANT, S. LASSALLE, A. PRIETO we see that limβ Lβ ◦ P = P . Thus, Λ(f )(P ) = lim β f (Lβ ◦ P ) = f (P ), for P ∈ Pw(nE, F ). Now, by [14, Lemma 1.8], as n < cd(E), the net of polyno- mials (Lβ ◦ Q) belongs to Pw(nE, F ), for every Q ∈ P(nE, F ). Hence, f (Lβ ◦ Q) = Λ(f )(Q), Λ(Λ(f ))(Q) = lim β Λ(f )(Lβ ◦ Q) = lim β and Λ2 = Λ. Finally, it is easy to check that ker Λ = Pw(nE, F )⊥. Then, Λ is a norm one projection on P(nE, F )∗ with ker Λ = Pw(nE, F )⊥. (cid:3) Remark 3.2. Every time Pw(nE, F ) is an ideal in P(nE, F ) with associated pro- jection Λ, we have the decomposition P(nE, F )∗ = Pw(nE, F )∗ ⊕ Pw(nE, F )⊥, and any f ∈ P(nE, F )∗ has a unique representation such that f = g + h, with g = Λ(f ) ∈ Pw(nE, F )∗, Now, if F has a metric compact approximation of the identity (Lβ) ⊂ K(F ), we may (and will) suppose that (Lβ) is w∗-convergent in K(F )∗∗ and that the projection Λ is defined as in (3.1), Λ(f )(P ) = limβ f (Lβ ◦ P ). Then, with Lβ = IdF − Lβ, we have the following facts that were already used: h = f − g ∈ Pw(nE, F )⊥. (3.2) • limβ Λ(f )(Lβ ◦ P ) = 0, for all f ∈ P(nE, F )∗ and all P ∈ P(nE, F ). • limβ Lβ ◦ Q = Q, for all Q ∈ Pw(nE, F ). Indeed, the first assertion follows by [14, Lemma 1.8]. For the second one, note that Lβ converges uniformly to the identity on compact sets. Finally, note that since kΛk ≤ 1, we automatically have kgk ≤ kf k. If in addition kLβk ≤ 1, we also obtain khk ≤ kf k since h(P ) = (f − Λf )(P ) = limβ f (Lβ ◦ P ) for all P ∈ P(nE, F ). As mentioned before, Delpech used the moduli of asymptotic uniform convex- ity and smoothness of a Banach space to obtain properties of weak sequential continuity of polynomials. Dimant, Gonzalo and Jaramillo [13] showed the con- nection between these moduli and compactness or weak sequential continuity of multilinear mappings. The moduli play their role when dealing with Pw(nE, F ) as an HB -subspace of P(nE, F ). We present a refinement of [10, Lemma 10.3], that will be used in the sequel. Lemma 3.3. Let E be an infinite dimensional Banach space and let (wα) ⊂ E be a weakly null bounded net. (a) If x ∈ BE, then limαkx + wαk ≤ 1 + ρE(limαkwαk). (b) If (xα) ⊂ BE is a net contained in a compact set, then limαkxα + wαk ≤ 1 + ρE(limαkwαk). Proof. To prove (a) first note that [10, Lemma 10.3] remains valid if we consider weakly null nets instead of weakly null sequences. That is, limαkx + wαk ≤ 1 + ρE(limαkwαk), for any x ∈ SE. Now, fix a nonzero x ∈ BE and consider each IDEAL STRUCTURES IN VECTOR-VALUED POLYNOMIAL SPACES 7 x + wα as a convex combination of x kxk + wα and −x kxk + wα. Applying the above inequality to ±x kxk we get limα(cid:13)(cid:13)(cid:13)(cid:13) ±x kxk + wα(cid:13)(cid:13)(cid:13)(cid:13) ≤ 1 + ρE(limαkwαk), and the statement follows. Now, suppose that (b) does not hold. Then, we may find subnets (xβ), (wβ), and x0 ∈ BE, so that limβ xβ = x0 and limβkxβ + wβk > 1 + ρE(limαkwαk). As ρE is increasing, limβkxβ + wβk > 1 + ρE(limβkwβk). Note that for any subnet (βi) such that limi kxβi + wβik exists, so too does the limit limi kx0 + wβik, and both coincide. This implies that limβkx0 + wβk = limβkxβ + wβk. It follows that limβkx0 + wβk > 1 + ρE(limβkwβk), which contradicts (a). (cid:3) We are ready to describe Pw(nE, F ) as an HB -subspace of P(nE, F ), under certain conditions on F and n. Theorem 3.4. Let E be a Banach space, and let n < cd(E). Let F be an infinite dimensional Banach space with a shrinking metric compact approximation of the identity (Lβ) ⊂ K(F ) such that supβ kLβk ≤ 1 and suppose that F is asymptoti- cally uniformly smooth. Then, Pw(nE, F ) is an HB-subspace of P(nE, F ). Proof. Consider the projection Λ, given in (3.1), under which Pw(nE, F ) is an ideal in P(nE, F ). For any f ∈ P(nE, F )∗, write f = g + h as in (3.2). Then, as commented in Remark 3.2, kgk ≤ kf k and khk ≤ kf k. In order to finish, we have to prove that kgk < kf k, for h 6= 0. Fix ε > 0 and take P ∈ P(nE, F ) and Q ∈ Pw(nE, F ) such that kP k = kQk = 1, h(P ) > khk − ε and g(Q) > kgk − ε. Since limβ Lβ ◦ Q = Q, we can choose β0 satisfying g(Lβ0 ◦ Q) > kgk − 2ε. Change, if necessary, Q to λQ (with λ = 1) to obtain g(Lβ0 ◦ Q) > kgk − 2ε. For t > 0, consider Lβ0 ◦ Q + tLβ ◦ P and take a net (xβ) ⊂ BE with limβkLβ0 ◦ Q + tLβ ◦ P kP(nE,F ) = limβk(Lβ0 ◦ Q)(xβ) + t(Lβ ◦ P )(xβ)kF . Now, note that ((Lβ ◦ P )(xβ)) is weakly null. Indeed, (P (xβ)) is bounded and for any y∗ ∈ F ∗, limβ(Lβ)∗y∗ = 0. Then, lim β h(Lβ ◦ P )(xβ), y∗i = lim β hP (xβ), (Lβ)∗y∗i = 0 for any y∗ ∈ F ∗. On the other hand, the compact set Lβ0(BF ) ⊂ BF contains the net ((Lβ0 ◦ Q)(xβ)). Therefore, Lemma 3.3 can be applied to get limβkLβ0 ◦ Q + tLβ ◦ P kP(nE,F ) ≤ 1 + ρF (limβkt(Lβ ◦ P )(xβ)kF ) ≤ 1 + ρF (t). As observed in Remark 3.2, limβ g(Lβ ◦ P ) = 0. Then, g(Lβ ◦ P ) < ε, for β ≥ β1. Also, h(Lβ ◦ P ) > khk − ε, since Lβ ◦ P belongs to Pw(nE, F ) and h(P ) = h(Lβ ◦ P ). 8 V. DIMANT, S. LASSALLE, A. PRIETO Combining the previous estimates, we conclude for t > 0 and β ≥ β1 that (kgk − 2ε) − tε + t(khk − ε) < g(Lβ0 ◦ Q) − tg(Lβ ◦ P ) + th(Lβ ◦ P ) ≤ f (Lβ0 ◦ Q + tLβ ◦ P ). Then, for t > 0 and ε > 0, kgk + tkhk − 2ε(1 + t) ≤ limβf (Lβ0 ◦ Q + tLβ ◦ P ) ≤ kf k limβkLβ0 ◦ Q + tLβ ◦ P k ≤ kf k(1 + ρF (t)). Thus, kgk + tkhk ≤ kf k(1 + ρF (t)). Now, suppose that kgk = kf k, then tkhk ≤ kf kρF (t), for t > 0. Since F is asymptotically smooth, limt→0 h = 0, which completes the proof. = 0 and (cid:3) ρF (t) t Our next result gives another set of sufficient conditions, also related with the notion of smoothness, under which Pw(nE, F ) is an HB -subspace in P(nE, F ). It is reminiscent of [23, Theorem 1]. The proof is similar to that of the above theorem and we omit it. Theorem 3.5. Let E, F be Banach spaces, and let n < cd(E). Suppose that there exists a metric compact approximation of the identity (Lβ) ⊂ K(F ) such that supβ kLβk ≤ 1 and for any ε > 0 there exist µ > 0 and β0 so that sup kLβy + µLβzk ≤ 1 + εµ , for all β ≥ β0. (3.3) kyk,kzk≤1 Then, Pw(nE, F ) is an HB-subspace of P(nE, F ). Following [19, Definition 3.6] we say that a Schauder basis of a Banach space is uniformly smooth if for every ε > 0 there exists δ > 0 such that kx+yk+kx−yk < 2 + εkyk, whenever x and y have disjoint supports with respect to the basis, kxk = 1 and kyk < δ. Note that by using convex combinations, the definition can be restated for x with kxk ≤ 1. The next result should be compared with [19, Theorem 3.7]. Corollary 3.6. Let E, F be Banach spaces, and let n < cd(E). Suppose that F has a uniformly smooth 1-unconditional basis. Then Pw(nE, F ) is an HB-subspace of P(nE, F ). Proof. Let ΠN be the natural projection on F onto the subspace generated by the first N elements of the 1-unconditional basis. Then, (ΠN ) is a metric com- pact approximation of the identity and satisfies supN kΠN k ≤ 1. Also, (3.3) in Theorem 3.5 holds. Indeed, take ε > 0 and consider µ = δ as in the definition of uniform smoothness of the basis. For any y, z ∈ BF and N ∈ N, being the basis 1-unconditional, we have kΠN y + µΠN zk ≤ 1 + µε/2. Thus, an immediate application of Theorem 3.5 gives the result. (cid:3) The above corollary can be applied to show some examples of HB -subspaces of polynomials where the spaces ℓp and the Lorentz sequence spaces d(w, p) appear. Recall that, for 1 < p < ∞, both ℓp and d(w, p) have uniformly smooth 1- unconditional bases. Also, the critical degree of ℓp is the integer number satisfying IDEAL STRUCTURES IN VECTOR-VALUED POLYNOMIAL SPACES 9 p ≤ cd(ℓp) < p + 1 and cd(ℓp, ℓq) is the integer satisfying p q + 1. For the case of cd(ℓp, d(w, q)), a restatement of (I) and (II) in [14, p. 705] reads as q ≤ cd(ℓp, ℓq) < p cd(ℓp, d(w, q)) = max{k ∈ N : k < + 1 and w 6∈ ℓ( p (k−1)q )∗}. p q Example 3.7. Let 1 < p, q < +∞. (a) Let E be a Banach space, and let n < cd(E). Then, • Pw(nE, ℓq) is an HB -subspace of P(nE, ℓq). • Pw(nE, d(w, q)) is an HB -subspace of P(nE, d(w, q)). (b) Let cd(ℓp, ℓq) < n < cd(ℓp). Then, Pw(nℓp, ℓq) is an HB -subspace but not an M-ideal in P(nℓp, ℓq). (c) Let cd(ℓp, d(w, q)) < n < cd(ℓp). Then, Pw(nℓp, d(w, q)) is an HB -subspace but not an M-ideal in P(nℓp, d(w, q)). The same result holds for n = cd(ℓp, d(w, q)) for the case cd(ℓp, d(w, q)) < p q . The statements about not being M-ideals in the previous examples are proved in [14, Theorems 3.2 and 3.9]. Now, we consider conditions satisfied by the domain space E so that we also have geometric structures in P(nE; F ). Similarly to what happens in Lemma 3.1, we will describe Pw(nE, F ) as an ideal of P(nE, F ) whenever E∗ has a metric compact approximation of the identity with adjoint operators. Here, no restric- tions on the degree of the polynomials are imposed. Lemma 3.8. Let E, F be Banach spaces such that E∗ has a metric compact approximation of the identity with adjoint operators. Then, Pw(nE, F ) is an ideal of P(nE, F ) for all n ∈ N. Proof. Let (Kα) ⊂ K(E) be a net satisfying limα K ∗ αx∗ = x∗, for all x∗ ∈ E∗ and supα kKαk ≤ 1. Without loss of generality, we may assume that (Kα) is weak∗-convergent in K(E)∗∗. Therefore, as in Lemma 3.1, the mapping Λ : P(nE, F )∗ → P(nE, F )∗ given by Λ(f )(P ) = lim α f (P ◦ Kα), (3.4) is well defined. It is clear that Λ is linear and kΛk ≤ 1. By [14, Lemma 2.1], limα kP − P ◦ Kαk = 0 for every P ∈ Pw(nE, F ). Then, Λ(f )(P ) = f (P ), for all P ∈ Pw(nE, F ). Furthermore, Λ is a projection: for all Q ∈ P(nE, F ), (Q ◦ Kα) belongs to Pw(nE, F ). Thus, for all Q ∈ P(nE, F ), Λ(Λ(f ))(Q) = lim α Λ(f )(Q ◦ Kα) = lim α f (Q ◦ Kα) = Λ(f )(Q), and Λ2 = Λ. It is easy to check that ker Λ = Pw(nE, F )⊥ and the result follows. (cid:3) The next result gives a sufficient condition to obtain the dual space Pw(nE, F )∗ as a quotient. We denote by π the projective tensor norm. Recall that P denotes the canonical extension of P in P(nE, F ) to P(nE∗∗, F ∗∗). 10 V. DIMANT, S. LASSALLE, A. PRIETO Proposition 3.9. Let E, F be Banach spaces such that Pw(nE, F ) does not con- n,s elementary tensor u ⊗ y∗ by j(u ⊗ y∗)(P ) = y∗(P (u)), is a quotient mapping. tain ℓ1. Then, the application j : b⊗ Proof. Take v ∈ b⊗ with ui ∈ b⊗ j(Xi n,s πs E∗∗ and y∗ ui ⊗ y∗ i )(P ) = Xi πs E∗∗b⊗πF ∗ → Pw(nE, F )∗, given on any πs E∗∗b⊗πF ∗. For each representation of v of the formPi ui⊗y∗ i ∈ F ∗ for all i, we have kuikky∗ i k. y∗ i (P (ui)) ≤ kP kXi n,s i , So, j is continuous and kjk = 1. Using Haydon's characterization of spaces not containing ℓ1, we may write the unit ball of Pw(nE, F )∗ as the closed convex hull of its extreme points. Now, by [14, Proposition 1.2], with ez(P ) = P (z) for z ∈ E∗∗, we obtain BPw(nE,F )∗ = Γ(ExtPw(nE,F )∗) ⊂ Γ(ez ⊗ y∗ : z ∈ SE ∗∗, y∗ ∈ SF ∗) ⊂ j(B b⊗ n,s πs E ∗∗ b⊗πF ∗) ⊂ BPw(nE,F )∗. Then, all the inclusions are (actually) equalities and j is a quotient mapping. (cid:3) In the next result we show that the natural hypothesis on E and F guarantee that Pw(nE, F ) does not contain ℓ1, and the above proposition can be applied. We will appeal to the result by Stegall which asserts that if a Banach space E has a separable subspace whose dual is nonseparable, then E∗ lacks the Radon -- Nikod´ym property, see for instance [11, Theorem VII.2.6]. Proposition 3.10. Let E, F be Banach spaces such that E∗∗ and F ∗ have the Radon -- Nikod´ym property. Then, Pw(nE, F )∗ has the Radon -- Nikod´ym property and hence Pw(nE, F ) does not contain ℓ1 for all n ∈ N. Proof. For any P ∈ Pw(nE, F ) we consider its associated symmetric multilinear map A and define TP ∈ L(E, Pw(n−1E, F )) as the operator given by TP (x)(x) = A(x, x, . . . , x). By [3, Theorem 2.9], TP is a well-defined compact operator and kP k ≤ kTP k ≤ ekP k. Then, the mapping Φ : Pw(nE, F ) → K(E, Pw(n−1E, F )) given by Φ(P ) = TP is an isomorphism with its image. As the Radon -- Nikod´ym property is preserved by isomorphisms, induction on n and [25, Theorem 1.9] yield the result. (cid:3) Lemma 3.11. Let E, F be Banach spaces such that Pw(nE, F ) does not contain ℓ1 and such that n < cd(E). Suppose that E∗ has a metric compact approximation of the identity with adjoint operators given by (Kα) ⊂ K(E). Then, limα P ◦ K α = 0 in the topology σ(P(nE, F ), Pw(nE, F )∗), for any P ∈ P(nE, F ). n,s fined by j(u ⊗ y∗)(P ) = y∗(P (u)) is a quotient mapping. We will show for i that limαhu, P ◦ K αi = 0. So, as for any n,s i=1 wi ⊗ wi ⊗ · · · ⊗ wi ⊗ y∗ Proof. By Proposition 3.9, the application j : b⊗ u = PM h ∈ Pw(nE, F )∗ there exists u ∈ b⊗ K αi = PM i ) = PM πs E∗∗b⊗πF ∗ → Pw(nE, F )∗, de- πs E∗∗b⊗πF ∗ so that j(u) = h and u can be approximated by such u's, the result follows. Take u as above. Then hu, P ◦ i ). As n < cd(E), P is w∗ − w∗ continuous and, since limα(K α)∗∗(wi) = 0 in the w∗-topology, we obtain limαhu, P ◦ K αi = 0, which completes the proof. (cid:3) i=1 P ((K α)∗∗(wi))(y∗ i=1 P ◦ K α(wi)(y∗ IDEAL STRUCTURES IN VECTOR-VALUED POLYNOMIAL SPACES 11 Proposition 1.4 in [13] provides an appropriate equivalence of asymptotic uni- form convexity of power p. With a slight modification of its proof we drop the hypothesis of separability and obtain a refinement, analogous to condition (c) in Lemma 3.3, as follows. Lemma 3.12. Let E be an infinite dimensional Banach space and let 1 < p < ∞. The following statements are equivalent. (a) E has modulus of asymptotic uniform convexity of power p. (b) There exists a constant C > 0 such that for every x ∈ SE and every bounded weakly null net (wα) in E, we have limαkx + wαkp ≥ 1 + C limαkwαkp. (c) There exists a constant C > 0 such that for every net (xα) in a compact set of BE and every bounded weakly null net (wα) in E, we have limαkxα + wαkp ≥ limαkxαkp + C kwαkp. Proposition 3.13. Let E, F be Banach spaces such that E∗∗ and F ∗ enjoy the Radon -- Nikod´ym property. Suppose that E has modulus of asymptotic uniform convexity of power n = cd(E, F ), with n < cd(E), and suppose that E has a shrinking metric compact approximation of the identity (Kα) ⊂ K(E) satisfying supα kK αk ≤ 1 and limαkIdE − 2Kαk ≤ 1. Then, there exists C > 0 such that Pw(nE, F ) is an M(1, C)-ideal of P(nE, F ). Furthermore, Pw(nE, F ) is an HB-subspace of P(nE, F ). Proof. We proceed as in Theorem 3.4. Consider Λ as in (3.4), under which Pw(nE, F ) is an ideal in P(nE, F ) and write f ∈ P(nE, F )∗, f = g + h as in (3.2), where kgk = kΛ(f )k ≤ kf k. Consider ε > 0 and take P ∈ P(nE, F ), Q ∈ Pw(nE, F ) with kP k = kQk = 1, h(P ) ≥ khk − ε and g(Q) ≥ kgk − ε. For any α, the polynomial P − P ◦ K α is weakly continuous on bounded sets at 0 (see, for instance, the proof of [12, Proposition 2.2]). Since n = cd(E, F ), the net (P − P ◦ K α)α is in Pw(nE, F ). Then, with h ∈ Pw(nE, F )⊥, h(P ) = h(P ◦ K α) ≥ khk − ε, for all α. Also, as limα Q ◦ Kα = Q, there exists α0 so that g(Q◦Kα) > kgk−2ε and kIdE −2Kα0k < 1+ε, for all α ≥ α0. Changing Q◦Kα0 to λQ◦ Kα0 with λ = 1, if necessary, we may assume that g(Q◦ Kα0) > kgk −2ε. Now, with C > 0 (to be fixed later) we have kf k kQ ◦ Kα0 + CP ◦ K αk ≥ f (Q ◦ Kα0 + CP ◦ K α) ≥ g(Q ◦ Kα0) + Ch(P ◦ K α) − Cg(P ◦ K α) ≥ kgk + Ckhk − (2 + C)ε − Cg(P ◦ K α). By Lemma 3.11, we may find α1 ≥ α0 so that g(P ◦ K α) < ε for α ≥ α1. Thus, we obtain kf k kQ ◦ Kα0 + CP ◦ K αk ≥ kgk + Ckhk − 2ε(1 + C), for all α ≥ α1. Now, take (xα) ⊂ BE such that limαkQ ◦ Kα0 + CP ◦ K αk = limαkQ(Kα0xα) + CP (K αxα)k and note that (Kα0xα) is contained in a compact subset of BE and 12 V. DIMANT, S. LASSALLE, A. PRIETO that (K αxα) is weakly null. Since E has modulus of asymptotic convexity of power n we apply Lemma 3.12, with C > 0 as in item (c), and get limαkQ ◦ Kα0 + CP ◦ K αk ≤ limαkQkkKα0xαkn + CkP kkK αxαkn = limαkKα0xαkn + CkK αxαkn ≤ limαkKα0xα + K αxαkn ≤ limαkKα0 + K αkn ≤ kIdE − 2Kα0kn, where the last inequality, being standard, can be found for instance in [18, p. 300]. Then, we may find α2 > α1 so that kQKα0 + CP K α2k < 1 + ε, and therefore, kgk + Ckhk − 2ε(1 + C) ≤ kf k(1 + ε). Since ε > 0 is arbitrary, kgk + Ckhk ≤ kf k and Pw(nE, F ) is an M(1, C)-ideal in P(nE, F ). To prove that Pw(nE, F ) is also an HB -subspace of P(nE, F ) note that for h 6= 0, kgk < kgk + Ckhk ≤ kf k. On the other hand, for α > α1, kf k ≥ f (P ◦ K α) ≥ h(P ◦ K α) − g(P ◦ K α) ≥ khk − 2ε, implying that kf k ≥ khk, which completes the proof. (cid:3) 4. Ideal structures inherited by the range space Our purpose in this section is to give sufficient conditions on the spaces E and F under which those geometric properties enjoyed by Pw(nE, F ) as a subspace of P(nE, F ) are inherited by the range space F as a subspace of F ∗∗. We start with HB -smoothness presenting an extension to the polynomial setting of [24, Theorem 7]. Our proof also follows their ideas. Proposition 4.1. Let E, F be Banach spaces such that there exists a surjection ρ : E → F . If Pw(nE, F ) is HB-smooth in P(nE, F ) for some n ∈ N, then F is HB-smooth in F ∗∗. Proof. Denote by NA(E∗∗) the subset of norm-attaining elements in E∗∗ and consider A := {ρ∗∗(x∗∗) : x∗∗ 6= 0, x∗∗ ∈ NA(E∗∗)}. As ρ is surjective, ρ∗∗ is also surjective and, by the Bishop-Phelps theorem, A = F ∗∗. By [24, Theorem 1 (c)] it is enough to show that for every ρ∗∗(x∗∗) ∈ A and any sequence (yk) ⊂ F with ky1k < 1, kyk+1 − ykk < 1 there are y ∈ F and k0 ∈ N so that kρ∗∗x∗∗ − y ± yk0k < k0. 0) = kx∗∗ 0 (x∗ 0 k and define, 0(x)nyk ∈ Pw(nE, F ). It is Fix x∗∗ 0 6= 0 in NA(E∗∗), take x∗ 0 ∈ SE ∗ such that x∗∗ for k ∈ N, the n-homogeneous polynomial Pk(x) = x∗ clear that kP1k < 1 and kPk+1 − Pkk ≤ kyk+1 − ykk < 1, for all k ∈ N. Now, define Q(x) = ρ(x) x∗ Q given by Q(x∗∗) = ρ∗∗(x∗∗) x∗∗(x∗ 0(x)n−1kx∗∗ 0)n−1kx∗∗ 0 k. 0 k and consider its Aron -- Berner extension IDEAL STRUCTURES IN VECTOR-VALUED POLYNOMIAL SPACES 13 By [24, Theorem 1 (a)] there exist R ∈ Pw(nE, F ) and k0 ∈ N with kQ − R ± Pk0k < k0. As the Aron -- Berner extension preserves the norm, we obtain 0 kn. The result follows by taking y = kQ(x∗∗ 0 ) − R(x∗∗ R(x∗∗ 0 ) 0 kn . kx∗∗ As an immediate consequence we have the following result. 0 )k < k0kx∗∗ 0 ) ± P k0(x∗∗ (cid:3) Corollary 4.2. Let E be a Banach space. If Pw(nE, E) is HB-smooth in P(nE, E) for some n ∈ N, then E is HB-smooth in E∗∗. Note that the above corollary says that Pw(nℓ1, ℓ1) is not HB -smooth in P(nℓ1, ℓ1) for any n ∈ N. Now we address the notion of HB -subspace, when the range space F is a quo- tient of the space E. The following technical result, inspired by [26, Proposition 2.3], will be useful. Lemma 4.3. Let J be an ideal in the Banach space E under the projection q . Suppose that J is HB-smooth and λ ∈ (0, 2]. The following statements are equivalent. (i) kIdE ∗ − λqk ≤ 1. (ii) For each x ∈ BE there exists a net (yα) ⊂ J such that limα yα = x in the σ(E, J ∗)-topology and limαkx − λyαk ≤ 1. (iii) For each x ∈ BE and ε > 0 there exists a net (yα) ⊂ J such that limα yα = x in the σ(E, J ∗)-topology and limαkx − λyαk ≤ 1 + ε. Proof. To prove that (i) implies (ii), consider the set of indices A := {α = (N, M, ε) : N ∈ FIN(E∗∗), M ∈ FIN(E∗), ε > 0}, where FIN denotes the set of all finite dimensional subspaces, with the usual order. By the Principle of Local Reflexivity, for any α ∈ A there exists Tα : N → E so that • hTαx∗∗, x∗i = hx∗∗, x∗i, for x∗∗ ∈ N, x∗ ∈ M, • kTαk ≤ 1 + ε, • TαN ∩E = IdE. Fix x ∈ BE and consider yα = Tα(q y∗ ∈ J ∗ and ε > 0, then if α ≥ ({q ∗x) ∈ E, defined for α large enough. Fix ∗x}, {y∗}, ε), as J ∗ is the range of q, we have hy∗, yαi = hy∗, Tα(q ∗x)i = hy∗, q ∗xi = hqy∗, xi = hy∗, xi, and limα yα = x in the σ(E, J ∗)-topology. On the other hand, also for α large enough, kx − λyαk = kTα(x − λq ∗x)k ≤ (1 + ε)kx − λq ∗xk ≤ ≤ (1 + ε)kIdE ∗∗ − λq ∗k = (1 + ε)kIdE ∗ − λqk ≤ 1 + ε. Then, limαkx − λyαk ≤ 1 and (ii) follows. Clearly, (ii) implies (iii). To prove that (iii) implies (i), fix ε > 0, x ∈ BE and choose (yα) ⊂ J satisfying (iii). For each x∗ ∈ BE ∗, as limα x∗(yα) = limα qx∗(yα) = qx∗(x), we have that x∗(x) − λqx∗(x) = lim α x∗(x) − λx∗(yα) ≤ kx∗k limαkx − λyαk ≤ 1 + ε. 14 V. DIMANT, S. LASSALLE, A. PRIETO Since ε is arbitrary, the implication follows. (cid:3) Proposition 4.4. Let E, F be Banach spaces such that there exists a quotient If Pw(nE, F ) is an HB-subspace of P(nE, F ), for some mapping ρ : E → F . n ∈ N, then F is an HB-subspace of F ∗∗. Proof. First, note that F is an ideal in its bidual and call q the associated pro- jection. By Proposition 4.1, F is HB -smooth in F ∗∗. Then, by [22, Theorem], F is strongly HB -smooth in F ∗∗ and we only have to show that kf − qf k ≤ kf k for all f ∈ F ∗∗∗. In order to do so, we will see that condition (iii) in Lemma 4.3 is satisfied for λ = 1. Take y∗∗ ∈ BF ∗∗ and ε > 0. Choose w∗ ∈ SF ∗ so that y∗∗(w∗) > 0 and (cid:18) ky∗∗k y∗∗(w∗)(cid:19)n < 1 + ε. Now, with µ = y∗∗(w∗) define P ∈ P(nE, F ) by P (x) = µρ(x)(ρ∗w∗)n−1(x), for each x ∈ E. By Lemma 4.3, there exists a net (Pα) ⊂ Pw(nE, F ) converging to P in the σ(P(nE, F ), Pw(nE, F )∗)-topology and such that limαkP − Pαk ≤ 1. As ρ is a quotient mapping, ρ∗ is an isometry and we may find x∗∗ ∈ E∗∗ with ρ∗∗(x∗∗) = y∗∗ and kx∗∗k = ky∗∗k. Since the Aron -- Berner extension of each Pα has range in F , we define yα = P α(x∗∗/µ) ∈ F , for all α. Note that each x∗∗ ⊗ y∗ ∈ E∗∗ ⊗ F ∗ acts in a natural way as an element of P(nE, F )∗ and, therefore, as an element of Pw(nE, F )∗. Then, as P (z) = µρ∗∗(z)zn−1(ρ∗w∗), for z ∈ E∗∗, we have P (x∗∗/µ) = y∗∗ and y∗∗(y∗) = (P (x∗∗/µ))(y∗) = lim α Thus, yα → y∗∗ in the w∗-topology. Also, y∗(Pα(x∗∗/µ)) = lim α y∗(yα). limαky∗∗ − yαk = limαkP (x∗∗/µ) − P α(x∗∗/µ)k ≤ limαkP − Pαkkx∗∗/µkn ≤ (ky∗∗k/µ)n < 1 + ε. Another application of Lemma 4.3 gives that kIdF ∗∗∗ − qk ≤ 1 and, therefore, F is an HB -subspace of F ∗∗. (cid:3) Finally, we focus on M(1, C)-ideal structures. Recall that if J is an ideal in a Banach space E satisfying the M(1, C)-inequality the following condition holds: For any m ∈ N, y1, y2, . . . , ym ∈ BJ , x ∈ BE, and ε > 0, there is z ∈ J such that kyi + Cx − zk ≤ 1 + ε, for 1 ≤ i ≤ m, [7, Lemma 2.2]. In fact, when dealing with E = J ∗∗, it is true that being an M(1, C)-ideal is equivalent to an appropriate 2-ball property of type (1, C). Namely, we have the following equivalence, which can be proved with the arguments appearing in the proof of [7, Lemma 2.3]. Lemma 4.5. Let E be a Banach space, and letC ∈ (0, 1]. The following state- ments are equivalent. (i) E is an M(1, C)-ideal of E∗∗. (ii) For all x ∈ SE, x∗∗ ∈ SE ∗∗ and ε > 0, there exists x0 ∈ E with k ± x + Cx∗∗ − x0k < 1 + ε. IDEAL STRUCTURES IN VECTOR-VALUED POLYNOMIAL SPACES 15 We use the above characterization to give an analogous statement to Proposi- tion 4.4 in the case of M(1, C)-ideals. Proposition 4.6. Let E, F be Banach spaces such that there exists a quotient mapping ρ : E → F . If Pw(nE, F ) is an M(1, C)-ideal of P(nE, F ), for some n ∈ N and C > 0, then F is an M(1, C)-ideal of F ∗∗. Proof. Let us prove that condition (ii) in Lemma 4.5 is satisfied. Fix y ∈ SF , (1 − δ)n−1 < 1+ε and y∗ ∈ SF ∗ y∗∗ ∈ SF ∗∗ and ε > 0. Choose δ > 0 such that δ + with y∗∗(y∗) ≥ 1 − δ. Define, for x ∈ E, P ∈ P(nE, F ) and Q ∈ Pw(nE, F ) by 1 + δ P (x) = (ρ∗y∗)n−1(x)ρ(x) and Q(x) = (ρ∗y∗)n(x)y, with kP k, kQk ≤ kρkn = 1. Due to [7, Lemma 2.2], there exists R ∈ Pw(nE, F ) so that k ± Q + CP − Rk ≤ 1 + δ. As ρ∗ is an isometry, there is x∗∗ ∈ SX ∗∗ with ρ∗∗(x∗∗) = y∗∗. Extending by Aron -- Berner, for z ∈ E∗∗, P (z) = z(ρ∗y∗)n−1ρ∗∗(z), and Q(z) = z(ρ∗y∗)n y. Then, with µ = y∗∗(y∗), P (x∗∗) = µn−1y∗∗ and Q(x∗∗) = µny, k ± µn y + Cµn−1 y∗∗ − R(x∗∗)k = k ± Q(x∗∗) + CP (x∗∗) − R(x∗∗)k ≤ 1 + δ. As R ∈ Pw(nE, F ), the range of R is also in F and we may take y0 = R(x∗∗)/µn−1. Thus, k ± µy + Cy∗∗ − y0k ≤ 1 + δ (1 − δ)n−1 . Finally, k ± y + Cy∗∗ − y0k ≤ ky − µyk + k ± µy + Cy∗∗ − y0k < 1 + ε, and the result follows. (cid:3) In the above proposition, the case C = 1 corresponds to the structure of an M-ideal. The corollary follows directly and seems to be new in this context. Corollary 4.7. Let E, F be Banach spaces such that there exists a quotient map- ping ρ : E → F . If Pw(nE, F ) is an M -ideal of P(nE, F ), for some n ∈ N, then F is an M -ideal of F ∗∗. As we have already noted in Corollary 4.2, it is now immediate to derive ver- sions of Proposition 4.4, Proposition 4.6 and Corollary 4.7 for the case E = F . Acknowledgement. The work of this paper was initiated while the third author was visiting the Departamento de Matem´atica, Universidad de San Andr´es. She is greatly indebted to Profs. Dimant and Lassalle for the invitation during her sabbatical, and for the warm hospitality extended by them and their families. Work partially supported by PAI-UdeSA 2013, CONICET-PIP 0624, ANPCyT PICT 1456, UBACyT X038 and MINECO grant MTM 2012-3431. 16 V. DIMANT, S. LASSALLE, A. PRIETO References 1. R. Alencar, K. Floret. Weak-strong continuity of multilinear mappings and the Pelczynski- Pitt theorem. J. Math. Anal. Appl. 206 (1997), no. 2, 532 -- 546. 2. R. Aron, P. Berner. A Hahn -- Banach extension theorem for analytic mappings. Bull. Math. Soc. France 106 (1978), 3 -- 24. 3. R. Aron, C. Herv´es, M. Valdivia. Weakly continuous mappings on Banach spaces. J. Funct. Anal. 52 (1983), 189 -- 204. 4. R. Aron, J. B. Prolla. Polynomial approximation of differentiable functions on Banach spaces. J. Reine Angew. Math. 313 (1980), 195 -- 216. 5. C. Boyd, R. Ryan. Bounded weak continuity of homogeneous polynomials at the origin. Arch. Math. (Basel) 71 (1998), no. 3, 211 -- 218. 6. J. C. Cabello, E. Nieto. On properties of M -ideals. Rocky Mountain J. Math. 28 (1998), no. 1, 61-93. 7. J. C. Cabello, E. Nieto, E. Oja. On ideals of compact operators satisfying the M (r, s)- inequality. J. Math. Anal. Appl. 220 (1998), 334 -- 348. 8. D. Carando, S. Lassalle. E ′ and its relation with vector-valued functions on E. Ark. Mat. 42 (2004), no. 2, 283 -- 300. 9. A. Davie, T. Gamelin. A theorem on polynomial-star approximation. Proc. Amer. Math. Soc. 106 (1989), no. 2, 351 -- 356. 10. S. Delpech. Approximations hold´eriennes de fonctions entre espaces d'Orlicz. Modules asymptotiques uniformes. Doctoral thesis. Universit´e de Bordeaux, 2005. 11. J. Diestel, J. J. Uhl. Vector measures. Mathematical Surveys 15. American Mathematical Society, Providence, R. I., 1977. 12. V. Dimant. M -ideals of homogeneous polynomials. Studia Math. 202 (2011), 81 -- 104. 13. V. Dimant, R. Gonzalo, J. A. Jaramillo. Asymptotic structure, ℓp-estimates of sequences, and compactness of multilinear mappings. J. Math. Anal. Appl. 359 (2009), 680 -- 693. 14. V. Dimant, S. Lassalle. M -structures in vector-valued polynomial spaces. J. Convex Anal. 19 (2012), no. 3, 685 -- 711. 15. S. Dineen. Complex analysis on infinite-dimensional spaces. Springer Monographs in Math- ematics. Springer-Verlag London, Ltd., London, 1999. 16. M. Gonz´alez, J. Guti´errez. The polynomial property (V). Arch. Math. (Basel) 75 (2000), no. 4, 299 -- 306. 17. R. Gonzalo, J. A. Jaramillo. Compact polynomials between Banach spaces. Proc. Roy. Irish Acad. Sect. A 95 (1995), no. 2, 213 -- 226. 18. P. Harmand, D. Werner, W. Werner. M -ideals in Banach spaces and Banach algebras. Lecture Notes in Mathematics, 1547. Springer-Verlag, Berlin, 1993. 19. J. Hennefeld. M -ideals, HB-subspaces, and compact operators. Indiana Univ. Math. J. 28 (1979), no. 6, 927 -- 934. 20. J. Johnson. Remarks on Banach spaces of compact operators. J. Funct. Anal. 32 (1979), 304 -- 311. 21. V. Lima, A. Lima. Ideals of operators and the metric approximation property. J. Funct. Anal. 210 (2004), no. 1, 148 -- 170. 22. E. Oja. Strong uniqueness of the extension of linear continuous functionals according to the Hahn -- Banach theorem. (Russian) Mat. Zametki 43 (1988), no. 2, 237 -- 246, 302; translation in Math. Notes 43 (1988), no. 1 -- 2, 134 -- 139. 23. E. Oja. Isometric properties of the subspace of compact operators in the space of continuous linear operators. (Russian) Mat. Zametki 45 (1989), no. 6, 61 -- 65, 111; translation in Math. Notes 45 (1989), no. 5 -- 6, 472 -- 475. 24. E. Oja, M. Poldvere. On subspaces of Banach spaces where every functional has a unique norm-preserving extension. Studia Math. 117 (1996), no. 3, 289 -- 306. 25. W. M. Ruess, C. P. Stegall. Extreme points in duals of operator spaces. Math. Ann. 261 (1982), no 4, 535 -- 546. IDEAL STRUCTURES IN VECTOR-VALUED POLYNOMIAL SPACES 17 26. D. Werner. M -ideals and the "basic inequality". J. Approx. Theory 76 (1994), no. 1, 21 -- 30. 1Departamento de Matem´atica, Universidad de San Andr´es, Vito Dumas 284, (B1644BID) Victoria, Buenos Aires, Argentina and CONICET. E-mail address: [email protected]; [email protected] 2Departamento de An´alisis Matem´atico, Facultad de CC. Matem´aticas, Uni- versidad Complutense de Madrid, Plaza de Ciencias, 3, 28040 Madrid, Spain. E-mail address: [email protected]
1807.10790	2	1807	2018-08-26T06:06:33	Interpolation of weighted Sobolev spaces	[ "math.FA" ]	In this work we present a newly developed study of the interpolation of weighted Sobolev spaces by the complex method. We show that in some cases, one can obtain an analogue of the famous Stein-Weiss theorem for weighted $L^{p}$ spaces. We consider an example which gives some indication that this may not be possible in all cases. Our results apply in cases which cannot be treated by methods in earlier papers about interpolation of weighted Sobolev spaces. They include, for example, a proof that $\left[W^{1,p}(\mathbb{R}^{d},\omega_{0}),W^{1,p}(\mathbb{R}^{d},\omega_{1})\right]_{\theta}=W^{1,p}(\mathbb{R}^{d},\omega_{0}^{1-\theta}\omega_{1}^{\theta})$ whenever $\omega_{0}$ and $\omega_{1}$ are continuous and their quotient is the exponential of a Lipschitz function. We also mention some possible applications of such interpolation in the study of convergence in evolution equations.	math.FA	math	INTERPOLATION OF WEIGHTED SOBOLEV SPACES MICHAEL CWIKEL AND AMIT EINAV ABSTRACT. In this work we present a newly developed study of the interpo- lation of weighted Sobolev spaces by the complex method. We show that in some cases, one can obtain an analogue of the famous Stein-Weiss the- orem for weighted Lp spaces. We consider an example which gives some indication that this may not be possible in all cases. Our results apply in cases which cannot be treated by methods in earlier papers about interpo- lation of weighted Sobolev spaces. They include, for example, a proof that hW 1,p³Rd ,ω0´ ,W 1,p³Rd ,ω1´iθ = W 1,p³Rd ,ω1−θ 0 ωθ are continuous and their quotient is the exponential of a Lipschitz function. We also mention some possible applications of such interpolation in the study of convergence in evolution equations. 1´ whenever ω0 and ω1 CONTENTS Introduction 1. 1.1. Goals and aims 1.2. The setting of the problem 1.3. Main results 1.4. Structure of the paper Acknowledgement 2. Preliminaries: Interpolation spaces and the complex interpolation method 3. The simpler inclusion 4. Towards the difficult inclusion 5. Approximation theorems 6. Proofs of the main theorem, and of its main consequences 7. More about the space W p (θ, rω) 7.1. Can W p (θ, rω) be smaller than W 1,p (ωθ)? 7.2. A simpler subspace of£W 1,p (U , ω0),W 1,p (U , ω1)¤θ 8. Final Remarks 8.1. Equivalence of weights 8.2. Possible applications to evolution equations 8.3. Future research and open problems Appendix A. Additional Proofs 2 2 5 8 11 11 11 14 18 30 39 40 41 43 45 45 45 47 47 The first named author's work was supported by the Technion V.P.R. Fund, and by the Fund for Promotion of Research at the Technion. The second named author was supported by the Austrian Science Fund (FWF) grant M 2104-N32. 1 2 MICHAEL CWIKEL AND AMIT EINAV A.1. An alternative characterization of the space W A.2. From locally Lipschitz to globally Lipschitz - A proof of Fact (a) A.3. A Stein-Weiss like theorem in the case where rω is not Lipschitz A.4. A comment about homogeneous weighted Sobolev spaces of 1,p l oc (U ) univariate functions References 48 49 50 51 52 1. INTRODUCTION 1.1. Goals and aims. Consider two weighted Sobolev spaces, W s0,p0 (U , ω0) and W s1,p1(U , ω1) on an open subset U of Rd . (Precise definitions of these spaces and of other notions, which are mentioned only in general terms in this subsection, will be given later). A natural question one can ask oneself is the following: Question 1. If T is a linear operator which is bounded on both W s0,p0 (U , ω0) and W s1,p1(U , ω1), does this imply that T is also bounded on various other weighted Sobolev spaces W s,p (U , ω)? For certain choices of U , weight functions ω0, ω1, and s0, s1, p0 and p1 the answer is known to be affirmative. In most cases, the affirmation of Question 1 is obtained with the explicit, or implicit, help of Alberto Calderón's complex interpolation spaces. As in many (but not all) papers dealing with these inter- polation spaces, we will use Calderón's notation [A0, A1]θ for them. Thus, so far, the affirmation of Question 1 comes hand in hand with the affirmation of the following question: Question 2. Does the complex interpolation space [W s0,p0 (U , ω0),W s1,p1(U , ω1)]θ coincide with W s,p (U , ω) for some s, p and ω? Our goal in this work is to investigate Question 2 further, and provide some new affirmative answers to it - thereby also finding more cases where Question 1 has an affirmative answer. We consider that our results can be potentially useful in the study of asymptotic behaviour of solutions of certain evolution equations. Before continuing, let us recall two previously treated cases of Questions 1 and 2: The Stein-Weiss Theorem: Suppose that s0 = s0 = 0, i.e., suppose that our two weighted Sobolev spaces are simply weighted Lp spaces. Then a celebrated re- sult, often referred to as the Stein-Weiss Theorem, which goes back to the paper [20] of Eli Stein and Stein's joint paper [21] with Guido Weiss, gives a positive answer to Question 1 for all p0 and p1 in [1,∞] and all choices of positive mea- surable functions ω0 and ω1 on U (and even in a more general setting, when (U , d x) is replaced by an arbitrary measure space). More explicitly, in this case, the operator T is bounded on W 0,p (U , ω) whenever p and ω satisfy (1.1) 1 p = 1− θ p0 + θ p1 , and ω 1 1−θ p0 0 ω θ p1 1 . p = ω INTERPOLATION OF WEIGHTED SOBOLEV SPACES 3 for some θ ∈ (0,1). This result was obtained before Calderón developed his the- ory of complex interpolation spaces, but in the light of that theory the Stein- Weiss Theorem can also be expressed by the formula £W 0,p0 (U , ω0),W 0,p1 (U , ω1)¤θ = W 0,p (U , ω) (1.2) which holds isometrically for each θ ∈ (0,1) and for p and ω as in (1.1). (When p0 = p1 = ∞ the weight ω of course cannot be determined from (1.1) but in that case (1.2) holds for all choices of ω for a trivial reason related to the discussion below in Remarks 1.1 and 2.8.) A result of Jörgen Löfström: In his paper [14], among various other results, Löf- ström gives an affirmative answer to Question 2 in the setting where s0 and s1 1/p0 can be non-zero. However, the set U must be all of Rd and the powers ω and 0 1/p1 of the given weight functions are required to satisfy a special condition ω 1 which he calls "polynomial regularity". With these conditions fulfilled, Löfström is able to use the Fourier transform and the Mihlin multiplier theorem and to define the weighted Sobolev spaces W s0,p0¡Rd , ω0¢ and W s1,p1¡Rd , ω1¢ via prop- erties of the Fourier transforms of their elements. In this setting Löfström proves that (1.3) hW s0,p0³Rd , ω0´,W s1,p1³Rd , ω1´iθ = W sθ,p³Rd , ω´ to within equivalence of norms, for each p0 and p1 ∈ (1,∞) and each θ ∈ (0,1) where sθ = (1−θ)s0 +θs1 and p and ω are given by (1.1). Taking into account the particular different notation and formulation of definitions used by Löfström, one can see that this result is expressed by the formula (5.3) in Theorem 4 on page 208 of [14]. An interesting advantage of Löfström's method is that he can extend his definition of weighted Sobolev spaces to the case where their orders of smoothness are not necessarily integers, and in fact he proves his formula (1.3) in this more general setting. Stein-Weiss formula (1.2). Of course, in the special case where s0 = s1 = 0, the formula (1.3) becomes the In this paper, in contrast to Löfström's work, we shall only consider the case s0 = s1 = 1, p0 = p1 = p ∈ [1,∞). However, making these restrictions gives us much more flexibility in our choices of the set U and the weight functions ω0 and ω1 than is available in [14]. We can also include the case where p = 1. For a relatively large class of weight functions, and for U = Rd , we obtain that (1.4) to within equivalence of norms, for all p ∈ [1,∞) and θ ∈ (0,1), where here, and also in fact throughout this paper, ωθ is the weight function obtained, for each θ, from whichever weight functions ω0 and ω1 are currently under consideration, by the formula £W 1,p (U , ω0),W 1,p (U , ω1)¤θ = W 1,p (U , ωθ) (1.5) ωθ = ω1−θ 0 ωθ 1. 4 MICHAEL CWIKEL AND AMIT EINAV Remark 1.1. When p = ∞, a case which we have not formally included in our discussion, the formula (1.4) is an uninteresting triviality (see Remark 2.8 below) which holds isometrically for all choices of the weight functions ω0 and ω1 (as can be justified later by part (iv) of Theorem 2.7). Here is one easily formulated special case of our results, which gives some indication of the considerable flexibility we have in our choices of the weight functions ω0 and ω1: When the set U is chosen to be all of Rd , we can prove that (1.4) holds whenever ω0/ω1 is the exponential of a Lipschitz function and ω0 and ω1 are continuous. In fact we can make do here with a much weaker condition than continuity which is satisfied by various functions which are not even equivalent to continuous functions. In fact (1.4) is valid not only for U = Rd , but also for a certain collection of other open sets which satisfy properties which we shall specify below. Moreover, (U , ω) of functions in W 1,p (U , ω) that if we consider the natural subspaces W can be approximated by a natural class of compactly supported functions, an analogue of (1.4) for these spaces is valid for an even larger collection of open sets U . 1,p 0 Our strategy for obtaining (1.4) will be to show that, under somewhat milder conditions on ω0 and ω1 than those imposed in Löfström's theorem, and for ap- propriate choices of U , the interpolation space£W 1,p (U , ω0),W 1,p (U , ω1)¤θ sat- isfies the continuous inclusions W p (U , θ, rω) ⊂£W 1,p (U , ω0),W 1,p (U , ω1)¤θ ⊂ W 1,p(U , ωθ) for a certain space W p (U , θ, rω) which is the intersection of W 1,p (U , ωθ) with an appropriate weighted Lp space. These inclusions may also be of independent interest. We can then deduce that (1.4) holds under some explicitly formulated conditions (see Theorem 1.22), which ensure that W 1,p (U , ωθ) and W p (U , θ, rω) coincide. Interestingly enough, we will also present a somewhat intricate example where these spaces do not coincide, but we cannot exclude the possibility that (1.4) might still hold for that example. We are now ready, in the following subsections, to give more explicit and more detailed formulations of our results and of the definitions of the notions which play roles in them. Remark 1.2. There are a number of other papers and books which contain other interpolation results involving weighted Sobolev spaces (and also sometimes other related spaces such as weighted Besov spaces and/or weighted Triebel- Lizorkin spaces) but the results in them do not quite correspond to the kind of results that we are seeking here. Some of them deal with the real method, rather than the complex method of interpolation. Some deal with interpolation be- tween weighted Sobolev spaces and weighted Lp spaces. Also, in some of them, the definitions of weighted Sobolev spaces differ in some non-trivial way from the definition that we have chosen to use here. We can refer, for example, to [5, 7, 8, 9, 10, 15, 16, 17, 18, 19, 22]. Among the many other papers and some INTERPOLATION OF WEIGHTED SOBOLEV SPACES 5 books which deal with other aspects of weighted Sobolev spaces and some of their applications we can mention, for example, the works [11, 12, 13] of Kufner, Kufner-Opic, and Kufner-Sändig. 1.2. The setting of the problem. In what is to follow, d will always be a positive integer, p will always lie in the interval [1,∞], and U will always denote a non- empty open subset of Rd . For most of our results p will be restricted to the range [1,∞). The letter ω, with or without a subscript, will always denote a weight function on U , i.e. a Lebesgue measurable function ω : U → (0,∞). Often the weight functions which are of interest in various applications are con- tinuous. However, for our results here which establish the formula (1.4) in cer- tain cases, we will not need the weight functions ω0 and ω1 to be continuous, nor even equivalent to continuous functions. Instead we will require them to satisfy a rather weaker condition which we are now about to define. (We will however need ω0/ω1 to have some smoothness properties.) We will be considering weight functions ω : U → (0,∞) which, like continuous functions, satisfy (1.6) 0 < inf x∈K ω(x) ≤ sup x∈K ω(x) < ∞, for every compact set K ⊂ U . Consequently, it will sometimes be convenient to use the notation (1.7) m(K , ω) := inf x∈K ω(x), and M(K , ω) := sup x∈K ω(x). Definition 1.3. A weight function ω : U → (0,∞) on an open subset U of Rd which satisfies (1.6) will be said to satisfy the compact boundedness condition on U . Remark 1.4. A weight function ω satisfying this condition can be very far from being continuous. As we already intimated above, it can, for example, have the property that, for every choice of a pair of positive constants C1 and C2 and of a continuous function f , the inequality C1 f (x) ≤ ω(x) ≤ C2 f (x) cannot hold for all x ∈ U , not even for almost all x ∈ U . As an example of this, let U = (0,∞) and let ω be the weight function on U defined by ω(x) =P∞n=0 nχ(2n,2n+1]+ P∞n=0 n2χ(2n,+1,2n+2]. Remark 1.5. We will have occasions to use the obvious fact that, if ω0 and ω1 are weight functions which satisfy the compact boundedness condition on some set U , then the same is true for every power, product and quotient of these weights, and in particular for ω1−θ 0 ωθ 1 and ω0/ω1. In the literature there are two standard ways of defining weighted Lp spaces with respect to a given weight ω. In one of them the weighted space consists of all functions f for which ω f is in the (unweighted) Lp space. But here we shall choose the second way, which proceeds by replacing the ambient underlying measure, in our case Lebesgue measure d x, with the weighted measure ω(x)d x. In fact we shall need to use weighted Lp spaces of vector valued functions. 6 MICHAEL CWIKEL AND AMIT EINAV Definition 1.6. Given a weight function ω on U and p ∈ [1,∞), and m ∈ N, we define the Banach space Lp¡U , ω, Cm¢ =nφ : U → Cm φ is measurable and °°φ°°Lp (U,ω,Cm ) < ∞o where °°φ°°Lp (U,ω,Cm ) =Ã m Xi=1U¯¯φi (x)¯¯ p ω(x)d x! 1 p , for each φ =¡φ1,..., φm¢. We will sometimes use the abbreviated notation Lp (U , ω) for Lp(U , ω, C), i.e. when m = 1. We will also use the notation Lp (U , Cm ) for the "unweighted" case, where ω is identically 1. As is customary in such definitions, there is a natural "limiting" extension to the case where p = ∞, which in our context takes the following form: Definition 1.7. For U , ω and m as above we define the Banach space L∞¡U , ω, Cm¢ =nφ : U → Cm φ is measurable and °°φ°°L∞(U,ω,Cm ) < ∞o where °°φ°°L∞(U,ω,Cm) = esssup x∈Uµ max i∈{1,2,...,m}¯¯φi (x)¯¯¶ for each φ =¡φ1,..., φm¢. Remark 1.8. The consequence of the choice that we made in formulating Defi- nition 1.6 is that here the weighted L∞ space in fact does not depend on ω and coincides isometrically with the unweighted space L∞(U , Cm ) which we shall occasionally use below. Had we chosen the first way of defining weighted Lp spaces we would have obtained a more "interesting" space here. We can now introduce the main spaces that we wish to study. Definition 1.9. Given an open set U of Rd , a weight function ω on U and p ∈ [1,∞], we define W 1,p (U , ω) =nφ : U →C φ has a weak gradient ∇φ on U and ¡φ,∇φ¢ ∈ Lp³U , ω, Cd+1´o, where the weak gradient ∇φ (i.e. the vector of its weak first order partial deriva- tives) is defined in the usual way via the theory of distributions, i.e. for every ψ ∈ C∞c (U ) it satisfies (1.8) U ∇ψ(x)φ(x)d x = −U ψ(x)∇φ(x)d x. INTERPOLATION OF WEIGHTED SOBOLEV SPACES 7 The norm1 which we will define on W 1,p (U , ω) is given by °°φ°°W 1,p (U,ω) =°°¡φ,∇φ¢°°Lp(U,ω,Cd+1) . Remark 1.10. The requirement that φ and ∇φ satisfy the equation (1.8) in which the integrals are evidently understood to be Lebesgue integrals, forces both of these functions to be locally integrable on U (with respect to (unweighted) Lebesgue measure). Remark 1.11. In the case where the weight function ω is identically equal to 1 the above weighted Sobolev spaces coincide with their original and more frequently studied "unweighted" analogues. We shall occasionally need to use these un- weighted spaces here and (analogously to what we did for unweighted Lp spaces) we shall use the same notation as above for them, but with ω simply omitted. Note (cf. Remark 1.8) that, for every weight function ω, W 1,∞(U , ω) coincides isometrically with the unweighted Sobolev space W 1,∞(U ). The study of weighted Sobolev spaces is certainly not new. The definition which we have chosen to use for them here is only one of a number of various different definitions chosen by authors of the many papers and several books (already mentioned above in §1.1) which deal with weighted Sobolev spaces and in some cases also give or indicate various applications of them to other topics in analysis. Consider the normed space W k,p(Ω, σ) introduced on page 11 of [11] and also, with slightly different notation, on page 537 of [12]. In the case where Ω = U and k = 1 and all the elements of the vector σ = {σα : α ≤ 1} are chosen to be ω this space coincides isometrically with our space W 1,p (U , ω). (In [11] it is assumed that p ∈ [1,∞) and that the open set Ω is also connected.) For most of our purposes in this paper we need the weighted Sobolev spaces with which we are working to be complete. Accordingly, most of our results here will be formulated subject to the "abstract" requirement that the weighted Sobolev spaces being considered in them are complete. When seeking to apply these results we can keep in mind that there are various "concrete" conditions on a weight function ω which are sufficient to ensure that W 1,p (U , ω) is complete. One of these is a condition which appears in Theorem 1.11 of [12, pp. 540 -- 541] in the case where p > 1. More relevantly for us, and as we shall show below in Lemma 5.2, W 1,p(U , ω) is complete whenever ω satisfies the compact bounded- ness condition. This is a condition which will be imposed anyway, also for other purposes, on the weight functions appearing in our main theorem 1.16, and in Theorems 1.20, 1.21 and 1.22 which follow from it. Remark 1.12. We have not attempted to find examples of p, U and ω for which W 1,p(U , ω) is not complete. However Example 1.12 on pp. 541 -- 543 of [12] shows explicitly that a certain variant of W 1,p (U , ω), where different weight functions 1Of course in this definition, and in Definitions 1.6 and 1.7, we permit ourselves the usual abuse of terminology where the word "function" really means an equivalence class of functions which are equal to each other almost everywhere. 8 MICHAEL CWIKEL AND AMIT EINAV are used for the weighted Lp norms of the function and of its weak first order partial derivatives, can fail to be complete. In some of our results we will also be requiring the weighted Sobolev spaces W 1,p(U , ω) appearing in them to have the property that each of their elements can be approximated by a sequence of compactly supported C∞ or Lipschitz functions. Here too there are concrete conditions on U and ω which suffice to imply such properties. The interested reader can find more information about such issues in [11] and [12]. 1.3. Main results. As we've stated in §1.1, our main goal in this work is to find settings in which an analogue of the Stein-Weiss theorem holds for weighted Sobolev spaces, or for suitable substitutes for these spaces. In what follows, we will denote by Li pc (U ), or Li pc for short, the space of all Lip- schitz functions ϕ : U → C with compact support. To avoid any ambiguity here, we should point out that ϕ ∈ Li pc(U ) means that the closure of©x ∈ U : ϕ(x) 6= 0ª in Rd is a compact subset, not only of Rd , but also of U . Definition 1.13. Let ω be a weight function such that W 1,p (U , ω) is a Banach (U , ω) to be the sub- space, containing Li pc (U ). Define the Banach space W space of W 1,p (U , ω) equipped with the same norm as that of W 1,p (U , ω) which is the closure of Li pc (U ) with respect to that norm. 1,p 0 1,p 0 (U , ω) applicable. In particular we should mention that if ω satisfies the compact boundedness condition on U then, as we shall show below in Lemma 5.2, W 1,p (U , ω) has the two above mentioned properties required to make the above definition of W As certain steps in the proof of our main theorem will show, Li pc is a very natu- ral space to consider for our purposes. In addition, for many choices of the open set U , the subspace C∞c (U ) of Li pc(U ) can be shown to have the same closure in W 1,p (U , ω) norm as that of Li pc (U ). We will also sometimes need to consider functions which satisfy the following variant of the Lipschitz condition. Definition 1.14. A function φ : U → C is said to be locally Lipschitz if, for each compact subset K of U , the restriction of φ to K is Lipschitz. Remark 1.15. It is easy to verify that φ : U → C is locally Lipschitz if and only if its restriction to the set V is Lipschitz whenever V is an open set whose closure is a compact subset of U . Let us now state our main theorem: Theorem 1.16. Let ω0 and ω1 be weight functions on U that satisfy the com- pact boundedness condition (1.6). Assume furthermore that the function rω : U → (0,∞) defined by (1.9) rω(x) := ω0(x) ω1(x) INTERPOLATION OF WEIGHTED SOBOLEV SPACES 9 is locally Lipschitz on U . For each p ∈ [1,∞) and θ ∈ (0,1), let W p (U , θ, rω) be the space with the norm W p (U , θ, rω) =©φ : U → C φ ∈ W 1,p (U , ωθ), φ¯¯∇log (rω)¯¯ ∈ Lp (U , ωθ)ª ωθ(x)d x¶ °°φ°°W p (U,θ,rω) =µ°°φ°° W 1,p (U,ωθ) +U¯¯φ(x)¯¯ p¯¯∇log (rω(x))¯¯ 0 (U , θ, rω) to be the closure of Li pc (U ) in W p (U , θ, rω) with respect 1 p , p p p (1.10) and define W to this norm. Then, for every p ∈ [1,∞) and θ ∈ (0,1), we have that £W 1,p (U , ω0) ,W 1,p (U , ω1)¤θ ⊂ W 1,p (U , ωθ), (1.11) 0 (U , θ, rω) ⊂hW (U , ω1)iθ (U , ω0),W (1.12) 1,p 0 1,p 0 and W p . p 0 (U , θ, rω), Moreover, there exists a positive constant Cp which depends only on p, such that, for every φ ∈ W (1.13) °°φ°°W 1,p (U,ωθ) ≤°°φ°°[W 1,p (U,ω0),W 1,p (U,ω1)]θ ≤ Cp°°φ°°W p (U,θ,rω) Remark 1.17. The exact value of the constant in (1.13) interests us rather less than the fact that it is independent of our choices of weight functions and θ, and depends only on p. But we also note that the arguments that will eventually combine to give us a proof of Theorem 1.16 will also show that the said constant can be chosen to be (1.14) Cp = min β>0 2e β max(1, e β pp2βe). Remark 1.18. From the definition of [A0, A1]θ it is obvious (cf. part (iii) of Theo- rem 2.7) that 1,p 0 (U , ω0),W 1,p 0 hW (U , ω1)iθ ⊂£W 1,p (U , ω0) ,W 1,p (U , ω1)¤θ . We can thus combine this inclusion together with (1.11) and (1.12) to obtain one chain of inclusions. p can assume the value 0, it will be helpful for us to permit ourselves the following slight abuse of notation: Although the non-negative function ωθ¯¯∇log (rω)¯¯ Definition 1.19. Let Lp¡U , ωθ¯¯∇log (rω)¯¯ tions φ : U → C for which the semi-norm, which we shall denote by p¢ denote the space of measurable func- (1.15) is finite. °°φ°°Lp(U,ωθ∇ log(rω)p) =µRd¯¯φ(x)¯¯ p p¯¯∇log (rω(x))¯¯ 1 p , ωθ(x)d x¶ 10 MICHAEL CWIKEL AND AMIT EINAV While Theorem 1.16 is quite general, it might also be desirable to have a "cleaner" 1,p 0 version of it which deals only with spaces of the kind "W 1,p " or only with spaces of the kind "W " rather than a mixture of these two kinds. Such versions can indeed be obtained when suitable additional conditions are imposed on U and/or on the weight functions ω0 and ω1. The following two theorems give instances of this. Theorem 1.20. Let U , ω0 and ω1 be as in Theorem 1.16. Assume furthermore (U , ω1) that, for some p ∈ [1,∞), whenever φ is an element of W there exists a sequence©φnªn∈N of functions in Li pc(U ) which converges to φ in W 1,p (U , ω0) or in W 1,p (U , ω1) and is a bounded sequence in the other space. Then, as well as the inclusions (1.11) and (1.12), we also have that (U , ω0)∩ W 1,p 0 1,p 0 (1.16) W p 0 (U , θ, rω) ⊂hW 1,p 0 (U , ω0) ,W 1,p 0 (U , ω1)iθ ⊂ W 1,p 0 (U , ωθ) for that same value of p and for all θ ∈ (0,1). Theorem 1.21. Let U , ω0 and ω1 be as in Theorem 1.16. Assume furthermore that, for some p ∈ [1,∞), (1.17) W and (1.18) W 1,p 0 ¡U , ω j¢ = W 1,p¡U , ω j¢ , 0 (U , θ, rω,) = W p (U , θ, rω), p j = 0,1, θ ∈ (0,1). Then, as well as the inclusions (1.11) and (1.12), we also have that (1.19) W p (U , θ, rω) ⊂£W 1,p (U , ω0) ,W 1,p (U , ω1)¤θ ⊂ W 1,p (U , ωθ) for that same value of p and for all θ ∈ (0,1). In particular if U = Rd and if ω0 and ω1 satisfy the conditions imposed in The- orem 1.16, then (1.17) and (1.18) both hold for every p ∈ [1,∞) and therefore so does (1.19). With the help of Theorems 1.20 and 1.21 we are able to obtain the follow- ing theorem which identifies new conditions which suffice for a Stein-Weiss like theorem to hold for pairs of W 1,p spaces: Theorem 1.22. Let U be a non-empty open subset of Rd and let g : U → R be a Lipschitz function. Let ω0 and ω1 be weight functions on U such that at least one of them satisfies the compact boundedness condition and ω0(x) = ω1(x)e g (x). (i) Then the conditions of Theorem 1.16 are satisfied. (ii) If U = Rd then (1.20) hW 1,p³Rd , ω0´,W 1,p³Rd , ω1´iθ = W 1,p³Rd , ωθ´ to within equivalence of norms, for every p ∈ [1,∞) and θ ∈ (0,1). INTERPOLATION OF WEIGHTED SOBOLEV SPACES 11 (iii) If the open set U and the above two weight functions ω0 and ω1 satisfy the conditions of Theorem 1.20 then 1,p 0 (U , ω0) ,W 1,p 0 hW (U , ω1)iθ = W 1,p 0 (U , ωθ) (iv) If the number p, the open set U and the above two weight functions ω0 and to within equivalence of norms, for every p ∈ [1,∞) and θ ∈ (0,1). ω1 satisfy the conditions of Theorem 1.21 then £W 1,p (U , ω0) ,W 1,p (U , ω1)¤θ = W 1,p (U , ωθ) to within equivalence of norms, for that value of p and for every θ ∈ (0,1). 1.4. Structure of the paper. In §2 we will recall some basic definitions and facts about interpolation spaces, and in particular Calderón's complex interpolation spaces. In §3 we will show how a generalisation of the Stein-Weiss theorem p 1,p 0 easily implies the inclusion of£W 1,p (U , ω0) ,W 1,p (U , ω1)¤θ in W 1,p (U , ωθ) and 0 (U , θ, rω) in£W 1,p (U , ω0),W 1,p (U , ω1)¤θ, which will eventually yield our main we shall also consider the analogous inclusion for the W spaces. §4 is the main section of this work. In it we will discuss the more difficult inclusion of W result. With the help of some approximation theorems which we will present in §5, we will prove our main theorem and several interesting consequences of it in §6. In §7 we will discuss further properties of W p (U , θ, rω), such as the fact that it can sometimes happen that W p (U , θ, rω) is strictly smaller than W 1,p (U , ωθ). In the final section §8 we will share some final thoughts about this, and future, work. These will include a discussion of potential applications to rates of con- vergence in evolution equations. There is also an appendix at the end of the paper. It contains some auxiliary proofs which we felt should not hinder the flow of our presentation in the pre- ceding sections. Acknowledgement. A. Einav would like to thank Anton Arnold and Tobias Wöhrer for stimulating discussions about using interpolation theory to estimate rates of convergence to equilibrium in Fokker-Planck equations, which led him to start investigating the main problem presented in this work. 2. PRELIMINARIES: INTERPOLATION SPACES AND THE COMPLEX INTERPOLATION METHOD This short section has been included for the convenience of those readers who may not be familiar with the theory of interpolation spaces, in particular of those which are generated by the complex interpolation method (sometimes re- ferred to simply as the complex method). We briefly recall some basic definitions and some basic properties of these interpolation spaces. For more information about this subject we refer the reader to Calderón's fundamental paper [3] or to [2] or many other excellent textbooks. The notations and definitions which we shall use are fairly standard.2 2But note that there are some differences between the notation used in [3] and in [2]. 12 MICHAEL CWIKEL AND AMIT EINAV The basic setting of interpolation theory is a pair of normed vector spaces A0 and A1 which have the property that there exists a Hausdorff topological vector space A satisfying A0 ⊂ A , A1 ⊂ A , where the inclusions of A0 and A1 in A are linear and continuous. If A0 and A1 are also Banach spaces, then it is customary to say that (A0, A1) is a Banach couple, and also to sometimes use the abbreviated notation A instead of (A0, A1). Definition 2.1. Let A = (A0, A1) be a Banach couple. A normed vector space A is called an intermediate space with respect to (A0, A1) if A0 ∩ A1 ⊂ A ⊂ A0 + A1 with continuous inclusions, where the norms on A0 ∩ A1 and A0 + A1 are given by and kakA0∩A1 = max¡kakA0 ,kakA1¢ kakA0+A1 = inf©ka0kA0 +ka1kA1 a0 + a1 = aª. The notation T : (A0, A1) → (A0, A1) (or T : A → A for short) is taken to mean that T is a linear operator which maps A0+ A1 into itself and also satisfies T (A0) ⊂ A0 and T (A1) ⊂ A1 and is bounded from A0 into A0 and from A1 into A1. An intermediate space A is called an interpolation space with respect to (A0, A1) if it has the additional property that every linear operator T : (A0, A1) → (A0, A1) is also a bounded map of A into itself. We often need to consider a more elaborate version of the previous definition, for operators which map between the spaces of two possibly different Banach couples: Definition 2.2. Let A = (A0, A1) and B = (B0, B1) be Banach couples. The nota- tion T : (A0, A1) → (B0, B1) (or T : A → B for short) is taken to mean that T is a linear operator which maps A0 + A1 into B0 + B1 and also satisfies T (A0) ⊂ B0 and T (A1) ⊂ B1 and is bounded from A0 into B0 and from A1 into B1. Let A be an intermediate space with respect to (A0, A1) and let B be an intermedi- ate space with respect to (B0, B1). Then we say that A and B are relative in- terpolation spaces with respect to (A0, A1) and (B0, B1) if every linear operator T : (A0, A1) → (B0, B1) is a bounded map of A into B . There are many methods for creating interpolation spaces with respect to a Banach couple and relative interpolation spaces with respect to a pair of Ba- nach couples. In our work here we will only concern ourselves with the so-called complex method, which was introduced and developed by Calderón in [3], and which we will now describe. In what follows we will let S denote the infinite strip {z ∈ C 0 ≤ Rez ≤ 1} in the complex plane. INTERPOLATION OF WEIGHTED SOBOLEV SPACES 13 Definition 2.3. Let A = (A0, A1) be a Banach couple of complex Banach spaces. We define F (A0, A1), also denoted by F³A´, to be the space F³A´ =½ f : S →A0 + A1 f is continuous and bounded on S and analytic on int (S) t → f ( j + i t ) is a continuous bounded map of R into A j , for j = 0,1¾, t∈R°° f (i t )°°A0 ,sup t∈R°° f (1+ i t )°°A1¶. and define a norm on F³A´ by °° f°°F³A´ = maxµsup It can very readily be shown (using e.g., obvious very minor modifications of the argument in paragraph 22 of [3, p. 129]) that F³A´ with the above norm is a Banach space, and we will make use of this fact. For each given θ ∈ [0,1] we define [A0, A1]θ to be the linear space of all elements of a ∈ A0 + A1 for which there exists f ∈ F³A´ such that a = f (θ). This space, when equipped with the norm kak[A0,A1]θ = inf½°° f°°F³A´ f ∈ F³A´, f (θ) = a¾, is a Banach space. The basic and very useful interpolation property of the spaces defined by the previous definition is expressed by the following theorem of Calderón. (See §4 of [3, p. 115], also e.g., Theorem 4.1.2 of [2, p. 88].) Theorem 2.4. Let A = (A0, A1) and B = (B0, B1) be Banach couples of complex Ba- nach spaces. Then [A0, A1]θ and [B0, B1]θ are relative interpolation spaces with respect to A and B for each θ ∈ (0,1). Furthermore, for each linear operator T : A → B , the norms of T as a operator from A0 to B0, from A1 to B1 and from [A0, A1]θ to [B0, B1]θ satisfy the inequality kTkB([A0,A1]θ,[B0,B1]θ) ≤ kTk1−θ B(A0,B0)kTkθ B(A1,B1) . Remark 2.5. The above definition, describing the complex method for creating interpolation spaces, indicates why one may think of such spaces as 'in between' spaces. The original spaces A0 and A1 are somehow identified with elements of F³A´ when restricted to Rez = 0 or Rez = 1. The θ−interpolation space [A0, A1]θ is identified with elements of F³A´ restricted to the intermediate line Rez = θ. Remark 2.6. Our definition here of the space F³A´, which is often, but not al- that we have not required that limt→∞°° f ( j + i t )°°A j = 0 for j ∈ {0,1}. However it ways, the definition chosen to appear in papers about the complex interpolation method is in fact slightly different from the original definition appearing in [3] in is well known (implicit already in [3] and mentioned e.g. in [4, p. 1007]) and very 14 MICHAEL CWIKEL AND AMIT EINAV easy to check (by multiplying functions in F (A) by e β(z−θ)2 for arbitrarily small positive β) that this difference does not effect the definition of [A0, A1]θ nor its norm. Here are some more properties of complex interpolation spaces which will be relevant for our purposes: and Theorem 2.7. Let (A0, A1) be a Banach couple of complex Banach spaces and let θ be a number in (0,1). Then (i) A0 ∩ A1 is a dense subspace of [A0, A1]θ, (ii) each element a of A0 ∩ A1 satisfies the inequality A0 kakθ . (2.1) (iii) Suppose, furthermore, that, for j ∈ {0,1}, B j is a closed subspace of A j equipped kak[A0,A1]θ ≤ kak1−θ with the same norm as A j . Then (B0, B1) is a Banach couple, and [B0, B1]θ ⊂ [A0, A1]θ, and each b ∈ [B0, B1]θ satisfies A1 kbk[A0,A1]θ ≤ kbk[B0,B1]θ . (iv) If A0 = A1 with equality of norms, then [A0, A1]θ = A0 = A1 with equality of norms. For the statement and proof of (i) we refer to page 116 and then pages 132 -- 133 of [3], or to Theorem 4.2.2 and Lemma 4.2.3 of [2, pp. 91 -- 92]. The rather obvi- ous, immediate and well known proof of (ii) proceeds via the function f (z) = a/kak1−z which is a norm one element of F (A0, A1). Part (iii) is an im- mediate consequence of Definition 2.3. Part (iv) is an obvious consequence of Definition 2.3 and part (i). A0 kakz A1 Remark 2.8. We note that part (iv) of this theorem combined with Remark 1.11, justifies the claim made in Remark 1.1. With the basics of interpolation theory and the complex method in hand, we can start proving our main theorem. 3. THE SIMPLER INCLUSION Before beginning our systematic study of the interpolation spaces we of course need to be sure that £W 1,p (U , ω0),W 1,p (U , ω1)¤θ ¡W 1,p (U , ω0),W 1,p (U , ω1)¢ is a Banach couple, (3.1) when we either assume that W 1,p (U , ω0) and W 1,p (U , ω1) are both Banach spaces or impose explicit conditions on U , ω0 and ω1 which guarantee that. In fact (3.1) holds under such assumptions, because both of the spaces W 1,p (U , ω0) and W 1,p (U , ω1) are continuously embedded in the Banach space Lp (U , ω) when ω is chosen to be the weight function ω = min {ω0, ω1}. Our main theorem, Theorem 1.16, establishes two inclusions. The reader has INTERPOLATION OF WEIGHTED SOBOLEV SPACES 15 probably already noticed that one of them, (1.11), is actually an obvious and almost immediate consequence of the Stein-Weiss theorem, or of a slightly gen- eralised version of it due to Calderón. Nevertheless, we will provide an explicit proof of that inclusion in this section. We will also consider the case where this inclusion can be generalised to the W Here is the version that we need of Calderón's generalization of the Stein-Weiss theorem. spaces. 1,p 0 Theorem 3.1. The formula £Lp¡U , ω0, Cm¢ , Lp¡U , ω1, Cm¢¤θ = Lp¡U , ωθ, Cm¢ , (3.2) holds with equality of norms for every m ∈ N, every p ∈ [1,∞), every θ ∈ (0,1), every open subset U of Rd , and all pairs of weight functions ω0 and ω1 on U , when ωθ is defined by (1.5). Remark 3.2. In fact the formula (3.2) is a special case of the isometric formula £Lp0 (U , ω0, Cm), Lp1¡U , ω1, Cm¢¤θ = Lpθ¡U , ωθ,pθ , Cm¢ (3.3) which holds for every p0 and p1 in [1,∞) and each θ ∈ (0,1) when we take pθ and ωθ,pθ to be the number p and the function ω specified by (1.1). This can be proved by a straightforward variant of the proof of Theorem 2 in [20]. It can also be seen to be a very special case of a result3 in [3]. We shall now use this theorem to easily show the aforementioned inclusion, as well as the associated norm inequality. Theorem 3.3. Let U be an open subset of Rd and suppose that ω0 and ω1 are weight functions which both satisfy the compact boundedness condition on U . Then, for each p ∈ [1,∞), we obtain that¡W 1,p (U , ω0),W 1,p (U , ω1)¢ is a Banach couple and that, for each θ ∈ (0,1), its complex interpolation spaces satisfy and (3.4) £W 1,p (U , ω0),W 1,p (U , ω1)¤θ ⊂ W 1,p (U , ωθ) , °°φ°°W 1,p (U,ωθ) ≤°°φ°°[W 1,p (U,ω0),W 1,p (U,ω1)]θ , Proof. In view of Lemma 5.2, W 1,p (U , ω0) and W 1,p (U , ω1) are both Banach for every φ ∈£W 1,p (U , ω0),W 1,p (U , ω1)¤θ. spaces, and therefore also, as explained at the beginning of this section,¡W 1,p (U , ω0) ,W 1,p (U , ω1)¢ is a Banach couple. Define a linear operator T : W 1,p (U , ω0)+ W 1,p (U , ω1) → Lp³U , ω0, Cd+1´+ Lp³U , ω1Cd+1´ 3It is a special case of the isometric formula [X0(B0), X1(B1)]θ = X 1−θ 0 X θ stated as the second part of the result labelled "i)" on p. 125 of [3], holds for all choices of the Banach couple (B0, B1) whenever X0 and X1 are Banach lattices of measurable functions on the same underlying σ-finite measure space and X 1−θ X θ 1 has a certain property. That property holds in our case in view of the Lebesgue Dominated Convergence Theorem. 0 1 ¡[B0, B1]θ¢ which, as 16 by MICHAEL CWIKEL AND AMIT EINAV T φ :=¡φ,∇φ¢ . By the definition of W 1,p (U , ω) we have that T φ is a norm 1 map from W 1,p¡U , ω j¢ into Lp¡U , ω j , Cd+1¢ for j = 0,1. Using the standard interpolation theorem, The- orem 2.4, and Theorem 3.1, we conclude that T :£W 1,p (U , ω0),W 1,p (U , ω1)¤θ → Lp³U , ωθ, Cd+1´ is bounded with norm less than, or equal to, 1. Thus, if φ ∈£W 1,p (U , ω0),W 1,p (U , ω1)¤θ then¡φ,∇φ¢ ∈ Lp¡U , ωθ, Cd+1¢, i.e. φ ∈ W 1,p (U , ωθ). Moreover, °°φ°°W 1,p (U,ωθ) =°°¡φ,∇φ¢°°Lp(U,ωθ,Cd+1) =°°T φ°°Lp(U,ωθ,Cd+1) =°°T φ°°[Lp(U,ω0,Cd+1),Lp(U,ω1,Cd+1)]θ ≤°°φ°°[W 1,p (U,ω0),W 1,p (U.ω1)]θ This concludes the proof of this theorem. Thus it also establishes (1.11) and the left side of the norm inequality (1.13) in the statement of Theorem 1.16. (cid:3) . Remark 3.4. It should perhaps also be pointed out that the proof of Theorem 3.3 does not require the weight functions ω0 and ω1 to satisfy the compact bound- edness condition. In fact it suffices if they have any other property which en- sures that W 1,p (U , ω0) and W 1,p (U , ω1) are both Banach spaces. (For example, some such properties are considered in the latter part of Lemma 5.2.) Essentially the same proof as for Theorem 3.3 also shows that, for any p0 and p1 in [1,∞), if W 1,p0 (U , ω0) and W 1,p1 (U , ω1) are both Banach spaces then they satisfy the norm one inclusion £W 1,p0 (U , ω0),W 1,p1 (U , ω1)¤θ ⊂ W 1,p (U , ω) for each θ ∈ (0,1), where p and ω are as in (1.1). This is an immediate conse- quence of the isometry (3.3). We have not yet considered under what conditions we can use some version of our approach to obtain the reverse of this inclusion when p0 6= p1. Such a result would complement the particular case considered in Formula (5.3) of Theorem 4 of [14, p. 208] in which U = Rd and ω0, ω1 and ωθ are all required to be so-called polynomially regular weight functions. We still have one more task in this section, which is to consider under what conditions we can replace the spaces W 1,p (U , ω) by their subspaces W (U , ω) in the inclusion (1.11). We do this in the following theorem which will lead us eventually to Theorem 1.20. 1,p 0 1,p 0 (U , ω0)∩ W Theorem 3.5. Let U be an open subset of Rd and suppose that ω0 and ω1 are weight functions which both satisfy the compact boundedness condition on U . Suppose further that, for some value of p ∈ [1,∞), whenever φ is an element of (U , ω1), then there exists a sequence©φnªn∈N of functions in W Li pc(U ) which converges to φ in W 1,p (U , ω0) or W 1,p (U , ω1) and is a bounded sequence in the other space. Then, for that value of p and for each θ ∈ (0,1), hW (U , ω1)iθ ⊂ W (U , ω0) ,W (U , ωθ) 1,p 0 1,p 0 1,p 0 1,p 0 INTERPOLATION OF WEIGHTED SOBOLEV SPACES 17 and, furthermore, (3.5) for each φ ∈hW 1,p 0 1,p 0 °°φ°°W (U , ω0),W (U,ωθ) ≤°°φ°°hW (U , ω1)iθ 1,p 0 . 1,p 0 (U,ω0),W 1,p 0 (U,ω1)iθ Remark 3.6. We suspect that, given a pair of weight functions ω0 and ω1 which are known to both satisfy the compact boundedness condition, quite mild ad- ditional conditions might suffice to ensure that they also satisfy the second hy- pothesis required in Theorem 3.5. Perhaps it is even not necessary to impose any additional conditions for this to happen. Pending further investigation of this, we simply point out that this second hypothesis is obviously satisfied whenever one of the weight functions ω0 and ω1 is dominated by some constant multiple of the other. In that case there even exists a sequence©φnªn∈N in Li pc (U ) which converges to φ in both W 1,p (U , ω0) and W 1,p (U , ω1). Proof of Theorem 3.5. As pointed out in Remark 1.5, the fact that ω0 and ω1 both satisfy the compact boundedness condition immediately implies that ωθ also satisfies this condition for each θ ∈ (0,1). Consequently, by Lemma 5.2, W 1,p (U , ω) is a Banach space which contains Li pc (U ) whenever ω is any one of the weight functions ω0, ω1 or ωθ. This enables us to define W (U , ω) in accordance with Definition 1.13 for each of these choices of ω. Then the relevant definitions im- 1,p 0 mediately imply (cf. part (iii) of Theorem 2.7) that³W a Banach couple, and that 1,p 0 (U , ω0),W 1,p 0 (U , ω1)´ is and that (3.6) 1,p 0 1,p 0 (U , ω0),W hW °°φ°°[W 1,p (U,ω0),W 1,p (U,ω1)]θ ≤°°φ°°hW (U , ω1)iθ ⊂£W 1,p (U , ω0),W 1,p (U , ω1)¤θ (U,ω1)iθ (U,ω0),W 1,p 0 1,p 0 . for each φ ∈hW 1,p 0 (U , ω0),W 1,p 0 By definition, Li pc(U ) is contained in W fore also in W that 1,p 0 1,p 0 (U , ω1)iθ . 1,p 0 (U , ω0)∩ W Li pc(U ) ⊂hW 1,p 0 (U , ω0),W 1,p 0 (U , ω1)iθ . (U , ω1) and there- (U , ω1). This implies (cf. part (i) of Theorem 2.7) (U , ω0) and in W 1,p 0 1,p 0 In view of this inclusion and since we also know that Li pc(U ) is also contained in W (U , ωθ), we are now ready to proceed to our next step of the proof, which will be to show that the inequality (3.5) holds for every φ ∈ Li pc(U ). By defini- (U,ω) =°°φ°°W 1,p (U,ω) for such functions φ and this, combined with the inequality (3.4) of Theorem 3.3 and (3.6), indeed accomplishes this step. In view of the standard fact recalled in part (i) of Theorem 2.7, we can assert that . Therefore, W tion °°φ°°W (U , ω0),W 1,p 0 1,p 0 (U , ω0)∩ W 1,p 0 (U , ω1) is dense inhW 1,p 0 since W we show that each element φ in W (U , ωθ) is a Banach space, the proof of this theorem will be complete if (U , ωθ) (U , ω1) is also in W 1,p 0 1,p 0 (U , ω0)∩W 1,p 0 1,p 0 (U , ω1)iθ 1,p 0 18 MICHAEL CWIKEL AND AMIT EINAV and satisfies the inequality (3.5). (U , ω1). We keep in mind Let φ be an arbitrary element of W that Theorem 3.3 combined with parts (i) and (iii) of Theorem 2.7 guarantee (U , ω0)∩ W 1,p 0 1,p 0 with the properties guaranteed by the hypotheses of the present theorem, so and that φ must also be in W 1,p(U , ωθ). Now let©φnªn∈N be a sequence in Li pc(U ) that at least one of the two numerical sequences n°°φn − φ°°W 1,p (U,ω0)on∈N n°°φn − φ°°W 1,p (U,ω1)on∈N converges to 0 and both are bounded. We now use the result about the complex interpolation method which is recalled in part (ii) and inequality (2.1) of Theorem 2.7. In our current context the inequality (2.1) be- comes °°φn − φ°°hW 1,p 0 (U,ω0),W 1,p 0 which gives us that (3.7) (U,ω1)iθ ≤°°φn − φ°° =°°φn − φ°° lim n→0°°φ− φn°°hW 1,p 0 (U,ω0),W 1,p 0 θ W 1,p 0 (U,ω1) θ W 1,p (U,ω1) 1−θ W 1,p 0 1−θ (U,ω0) ·°°φn − φ°° W 1,p (U,ω0) ·°°φn − φ°° (U,ω1)iθ = 0, and therefore also (cf. part (iii) of Theorem 2.7) that limn→0°°φ− φn°°[W 1,p (U,ω0),W 1,p (U,ω1)]θ = 0. This in turn implies, by Theorem 3.3 that (3.8) lim n→0°°φ− φn°°W 1,p (U,ωθ) = 0 and therefore φ is not only an element of W 1,p (U , ωθ) but also of W So we can deduce from (3.8) and Definition 1.13 that 1,p 0 (U , ωθ). 1,p 0 °°φ°°W (U,ωθ) =°°φ°°W 1,p (U,ωθ) = lim Since an earlier step of this proof gives us that each φn satisfies (3.5), the above (U,ω1)iθ n→0°°φn°°W 1,p (U,ωθ) = lim limit must be less than or equal to the limit limn→∞°°φn°°hW which exists and equals°°φ°°hW (U , ω1), as well as be- (U , ωθ), also satisfies (3.5). As already explained above, this suffices (cid:3) 1,p 0 1,p shown that our arbitrary element φ of W 0 ing in W to complete the proof of our theorem. (U,ω1)iθ (U , ω0)∩ W n→0°°φn°°W in view of (3.7). Thus we have (U,ωθ) . 1,p 0 (U,ω0),W (U,ω0),W 1,p 0 1,p 0 1,p 0 1,p 0 1,p 0 In the next section we will make the significant step towards showing the other inclusion and inequality of our main theorem. This, however, will not be as simple. 4. TOWARDS THE DIFFICULT INCLUSION W 1,p 0 The goal of this section is to consider the relationship between the two spaces p . We do not find it surprising that in general we cannot obtain an analogous relationship between the "full" spaces 0 (U , θ, rω) andhW W p (U , θ, rω) and£W 1,p (U , ω0),W 1,p (U , ω1)¤θ, and can only deal with their above mentioned smaller but "significant" respective subspaces. However, as we have (U , ω1)iθ (U , ω0) ,W 1,p 0 INTERPOLATION OF WEIGHTED SOBOLEV SPACES 19 mentioned in §1, in many cases these smaller spaces are either everything, or "big" enough to warrant a theorem on their own. We begin with a definition which specifies some conditions which may or may not be satisfied by a given pair of weight functions ω0 and ω1. As we shall see, for most of the results of this section, we will require our weight functions to satisfy these conditions. Definition 4.1. We say that a pair of weight functions (ω0, ω1) is a (θ, p)−admissible pair of weight functions on U (or, in short, a (θ, p)−admissible pair) for some θ ∈ (0,1) and p ∈ [1,∞] if the following holds: (i) W 1,p (U , ω0), W 1,p (U , ω1) and W 1,p (U , ωθ) are Banach spaces that contain Li pc (U ). (ii) The function log (rω(x)) = log³ ω0(x) ω1(x)´ is locally Lipschitz on U . Remark 4.2. In view of the inequality ω1−θ 1 ≤ max {ω0, ω1} it is easy to see that if W 1,p(U , ω0) and W 1,p (U , ω1) both contain Li pc(U ), then in fact it follows au- tomatically that W 1,p (U , ωθ) also contains Li pc(U ) for every θ ∈ (0,1). 0 ωθ From this point onward, till the end of this section, we will assume that we are dealing with a given fixed open set U and often we will not mention U in the notation that we use. Remark 4.3. It is simple to observe that if (ω0, ω1) forms a (θ, p)−admissible pair for some θ ∈ (0,1) and p ∈ [1,∞), then W p (θ, rω) also contains Li pc. (Relevant definitions, i.e. of W p (θ, rω) and of rω, appear in the statement of Theorem 1.16.) Indeed, given any φ ∈ Li pc, we first note that it is contained in W 1,p (ωθ), in view of condition (i) of Definition 4.1. Then the other requirement for the member- ship of φ in W p (θ, rω) follows immediately from condition (ii) of the same def- inition, which ensures that the auxiliary function φ(x)¯¯∇log (rω(x))¯¯ is continu- ous. Since that function also has compact support, it is obviously an element of Lp (ωθ) as required. Remark 4.4. In most of our main results in this paper we require ω0 and ω1 to both satisfy the compact boundedness condition. In view of Remark 1.5, Lemma 5.2 and Lemma 5.6, this implies that (i) holds for all θ ∈ (0,1) and p ∈ [1,∞). From this we see that Definition 4.1, and the results in this section which use it, are somewhat more general than they need to be for obtaining those main results. But this greater generality may be useful in future investigations where more general weight functions are considered. Our main goal in this section is to prove the following result: Theorem 4.5. Let (ω0, ω1) form a (θ, p)−admissible pair for some θ ∈ (0,1) and p ∈ [1,∞). Then, we have that (4.1) W p (ω0),W 0 (θ, rω) ⊂hW 1,p 0 Moreover, there exists a positive constant Cp which depends only on p, such that (4.2) °°φ°°[W 1,p (ω0),W 1,p (ω1)]θ ≤ Cp°°φ°°W p (θ,rω) 1,p 0 (ω1)iθ for every φ ∈ W p 0 (θ, rω). 20 MICHAEL CWIKEL AND AMIT EINAV The proof of Theorem 4.5 has two main parts and will occupy the rest of this section. In the first part, for each given φ ∈ Li pc, we will construct an element f φ of F³W (ω1)´ which satisfies f φ(θ) = φ and also the norm esti- mates necessary for showing that φ satisfies (4.2). Then, in the second part, we will extend this construction to elements φ which are in the closure of Li pc in W p (θ, rω). (ω0) ,W 1,p 0 1,p 0 Definition 4.6. Given θ ∈ (0,1), an arbitrary function φ ∈ Li pc and a number β > 0, we define the function (z−θ) (4.3) log(rω(x))+β(z−θ)2 φ(x), φ p f β,θ(x, z) = e x ∈ U , z ∈ S. φ (Recall, cf. (1.9) that rω denotes the ratio ω0/ω1.) Thus f β,θ denotes the map from S into the space of measurable functions on U , which is such that the im- age f β,θ(z) of each z ∈ S is the function of x defined by (4.3). φ φ The reader might be surprised by the inclusion of another parameter β > 0 in the formula (4.3). This parameter appears in the exponential factor e β(z−θ)2 in φ β,θ, and we need that factor in order to control the size of the the formula for f β,θ(x, z) as z tends to ∞ on the strip S. (The same function e β(z−θ)2 has a quite f similar role elsewhere, as mentioned in Remark 2.6.) As will be evident in the course of the proofs of the results presented in this section, β could be chosen arbitrarily, but we can also note that there is one choice of β which will minimize the value that we can obtain for the above mentioned constant. For convenience, we will often drop the subscripts of β and θ and accordingly write f φ to denote the function f We begin our preparations for achieving the first of the above mentioned main parts of our proof by recalling a rather standard definition and mentioning an important theorem which we will use in several places. φ β,θ. 1,p Definition 4.7. For each p ∈ [1,∞] and each open subset U of Rd , the space l oc (U ) consists of all measurable functions f : U → C which have a weak gra- W dient ∇ f in U and for which, for every compact subset K of U , the functions χK f and χK ∇ f are elements, respectively, of the (unweighted) spaces Lp (U ) and of Lp (U , Cd ). We remark that it is quite straightforward, but perhaps a little tedious, to 1,p l oc (U ) if and only if, for every check that a function f : U → C is an element of W open subset V of U whose closure is a compact subset of U , the restriction of f to V is an element of W 1,p (V ). We refer the interested reader to the appendix for the proof of this fact. Theorem 4.8 (Rademacher's Theorem). Let U be an open set. Then, (1) a function φ : U → C is locally Lipschitz if and only if φ ∈ W 1,∞l oc (U ). INTERPOLATION OF WEIGHTED SOBOLEV SPACES 21 (2) Moreover, if φ is locally Lipschitz on U , then it is differentiable almost everywhere4 and its gradient equals its weak gradient almost everywhere5. The proof of this theorem can be obtained from material in Chapter 5 of [6] (subchapter 5.8.2 b., pages 279-281). Part (1), as pointed out in the remark on p. 280 of [6], can be obtained by an easy adaptation of the arguments of the proof of Theorem 4 on pp. 279 -- 280. Part (2) is the case p = ∞ of Theorem 5 ([6, pp. 280 -- 281]) combined with Theorem 6 ([6, p. 281]). The alternative characterisation of W 1,p l oc (U ) mentioned just after Definition 4.7 is relevant here. Remark 4.9. When we use Theorem 4.8, it is sometimes helpful to also keep in mind the very obvious fact that, at every point x ∈ U where the pointwise gradi- is the Lipschitz constant of the restriction of φ to a bounded open set containing x. ent ∇φ(x) of a locally Lipschitz function φ exists, it satisfies¯¯∇φ(x)¯¯ ≤ L where L Lemma 4.10. Let (ω0, ω1) form a (θ, p)−admissible pair for some θ ∈ (0,1) and p ∈ [1,∞), and let φ ∈ Li pc. Then, for each fixed z ∈ S, we have that f φ(x, z) ∈ Li pc and consequently f φ(·, z) ∈ W (ω1) ⊂ W (ω0)∩ W (ω0)+ W (ω1). (4.4) 1,p 0 1,p 0 1,p 0 1,p 0 Before we begin proving this lemma, it is convenient to present a few simple facts that we will use in the proof of this lemma, as well as in subsequent proofs: (a) If the function ψ : U → C is locally Lipschitz and has compact support, then it is Lipschitz on all of U . This rather standard result is perhaps slightly less obvious than may seem at first glance. For the sake of completeness, we include its quite straightforward and short proof in the appendix. (b) Every Lipschitz function is bounded on every compact set. Therefore every locally Lipschitz function also has this property. (c) The pointwise product of two bounded Lipschitz functions is itself also a bounded Lipschitz function. So, by (b), the pointwise product of two locally Lipschitz functions is also locally Lipschitz. (d) The composition of two Lipschitz functions is a Lipschitz function. There- fore the composition of two locally Lipschitz functions is a locally Lipschitz function. (e) A function mapping R to C whose real and imaginary parts have bounded derivatives on every bounded interval is a locally Lipschitz function. p (z−θ) t+β(z−θ)2 Proof of Lemma 4.10. Let us define g1 : U → R by g1(x) = log (rω(x)) and define . Then f φ(x, z) = φ(x)g2¡g1(x)¢ for all x ∈ U . g2 : R → C by g2(t ) = e Since φ has compact support so does φ(x)g2¡g1(x)¢. So, in view of fact (a) it suffices to show that φ(x)g2¡g1(x)¢ is locally Lipschitz, which we will now do 4When we say "almost everywhere" we always mean with respect to d−dimensional Lebesgue 5This equality is to be understood in the following way: each component of the pointwise gradient belongs to the equivalence class of the corresponding component of the weak gradient. measure. 22 MICHAEL CWIKEL AND AMIT EINAV (cid:3) 1,p 0 1,p 0 lemma. (ω0),W (ω1)´. with the help of the other above mentioned obvious facts. We first use (e) to show that g2 is locally Lipschitz. The (θ, p)−admissibility of (ω0, ω1) implies that g1 is also locally Lipschitz. So (d) gives us that g2 ◦ g1 is locally Lipschitz. Finally, since φ is obviously locally Lipschitz, (c) completes the proof that φ(x)g2¡g1(x)¢ is locally Lipschitz and therefore also the proof of this Lemma 4.11. Let (ω0, ω1) form a (θ, p)−admissible pair for some θ ∈ (0,1) and p ∈ [1,∞), and let φ ∈ Li pc. Then the function f φ is an element of the space F³W Proof. For reasons which the reader can quite possibly guess and which anyway will become apparent later, we begin this proof by establishing some properties of the functions¡log (rω(x))¢k and¡log (rω(x))¢k φ(x) for each non-negative inte- ger k. We first note that the (θ, p)−admissibility of (ω0, ω1), together with facts (d) and (e) that were stated after Lemma 4.10, ensure that, for each such k, the function¡log (rω(x))¢k is locally Lipschitz. Consequently, using fact (c) and then fact a, we obtain that¡log (rω(·))¢k φ(·) is in Li pc. This in turn implies, again be- cause (ω0, ω1) is (θ, p)−admissible, that this same function is also an element of W Our next task will be to estimate the norm of this function in each of these spaces. The fact that φ has compact support ensures the existence of an open set Vφ whose closure is a compact subset of U such that at every point x ∈ U \ Vφ we have that φ(x) = 0 and also that the pointwise gradient ∇φ(x) exists and also equals 0. We observe that, for j ∈ {0,1}, (ω0) and of W (ω1). 1,p 0 1,p 0 φ(·)°°° °°°¡log (rω(·))¢k x∈Vφ¯¯log (rω(x))¯¯!pk ≤Ãsup p p p¯¯log (rω(x))¯¯ Lp(ω j ) = Λpk°°φ°° Lp(ω j ) =U¯¯φ(x)¯¯ °°φ°° x∈Vφ¯¯log(rω(x))¯¯ Λ := sup pk ω j (x)d x p Lp(ω j ) , where Λ denotes the quantity which is finite, as the supremum of a Lipschitz function on a bounded set. Next we need to make some more preparations for performing a somewhat more elaborate calculation to estimate the size of°°°∇³¡log (rω(·))¢k We note the following: • For any a, b ∈ R we have that φ(·)´°°° (4.5) (a+b)p ≤ 2p¡ap +bp¢, p ≥ 0. p . Lp(ω j ) INTERPOLATION OF WEIGHTED SOBOLEV SPACES 23 • By part (2) of Theorem 4.8 we know that the pointwise gradients∇log (rω(x)) and ∇φ(x) of the locally Lipschitz functions log (rω(x)) and φ(x) both ex- ist for all x in a certain subset U# of U which is such that U \U# has mea- sure 0. We shall let Ξ denote the supremum (4.6) Ξ := sup x∈Vφ∩U#¯¯∇log (rω(x))¯¯ , which is obviously finite (cf. Remark 4.9). p 1 p p Lp(ω j )¶ φ(x). another application of Rademacher's theorem tells us that the pointwise statements hold for the pointwise and weak gradients of the functions φ • Furthermore it is also obvious that the pointwise gradient of¡log (rω(x))¢k exists and equals k¡log (rω(x))¢k−1 ∇log (rω(x)) for every x ∈ U#. As al- ready observed above, the function¡log (rω(x))¢k is locally Lipschitz. So gradient of¡log(rω(x))¢k on U# is also its weak gradient on U . Analogous and¡log (rω(x))¢k Having made these preparations, we can now see that, for j ∈ {0,1} and for each k ∈ N, φ(·)´°°° °°°∇³¡log (rω(·))¢k +2p k pVφ¯¯φ(x)¯¯ Lp(ω j ) ≤ 2pVφ¯¯∇φ(x)¯¯ p¯¯∇log (rω(x))¯¯ ≤ 2p Λp(k−1)³Λp°°∇φ°° p¯¯log (rω(x))¯¯ Lp(ω j ) + k p Ξp°°φ°° p¯¯log (rω(x))¯¯ Lp(ω j )´ ω j (x)d x ω j (x)d x p(k−1) pk p W 1,p(ω j ) p W 1,p(ω j ) φ(·) ∈ W and for k = 0 we obviously obtain that ≤ 2p Λp(k−1)¡Λp + k p Ξp¢°°φ°° °°°∇³¡log (rω(·))¢k φ(·)´°°° Lp(ω j ) ≤°°φ°° We already noted above that ¡log (rω(·))¢k can conclude furthermore, from the preceding estimates, that, for each k ∈ N, 0 (ω j ) =µ°°°¡log (rω(·))¢k φ(·)°°°W °°°¡log(rω(·))¢k ≤³Λpk°°φ°° Lp(ω j ) + 2p Λp(k−1)¡Λp + k p Ξp¢°°φ°° ≤³2p Λp(k−1)¡2−p Λp + Λp + k p Ξp¢°°φ°° W 1,p(ω j )´ ≤ 2Λk−1¡2Λp + k p Ξp¢1/p°°φ°°W 1,p (ω j ) . φ(·)°°°W We estimate this last expression, using (4.5) again, but this time with 1 of p, and thus obtain that Lp(ω j ) +°°°∇³¡log (rω(·))¢k p Λ+ k Ξ´°°φ°°W 1,p (ω j ) . °°°¡log (rω(·))¢k 0 (ω j ) ≤ 21+ 1 p Λk−1³2 W 1,p(ω j )´ φ(·)°°° (ω j ) for j = 0,1. Now we φ(·)´°°° p in place p 1 p p 1 p 1,p 0 p p p p , . 1,p p (4.7) 1,p 1 24 MICHAEL CWIKEL AND AMIT EINAV The analogue of this inequality for k = 0 is just the trivial fact that For each N ∈ N we define the function f N : S → Li pc by setting 1,p 0 °°φ°°W N (x, z) = e β(z−θ)2 N φ φ (ω j ) =°°φ°°W 1,p (ω j ) . Xk=0¡log (rω(x))¢k (z − θ)k p k k! φ(x). (4.8) f In view of (4.7) we have that Here, analogously to the convention adopted in Definition 4.6, it must be under- φ stood that f N denotes the map from S into the space of measurable functions on U , which is such that the image f N (z) of each z ∈ S is the function of x de- φ(·) is known to be (ω0). Also, a simple fined by (4.8). For each k ∈ N∪ {0}, the function¡log (rω(·))¢k 1,p 0 1,p 0 φ (ω0) ∩ W an element of Li pc and therefore also of W calculation shows that the entire function z 7→ e β(z−θ)2 (z − θ)k 1,p 0 is bounded on S. (Later we will do this calculation in more detail, to find an explicit bound for the modulus of this function on S.) From these two facts it φ is obvious that f N indeed maps S into Li pc and, furthermore, that it is an ele- ment of F (A0, A1), where, throughout this proof, we will take (A0, A1) to be the (ω0) and letting Banach couple obtained by setting A0 = A1 = W (ω0)∩ W A0 and A1 be normed by°°ψ°°A0 =°°ψ°°A1 = maxn°°ψ°°W next step will be to use the inequalities (4.7) to show thatn f Xk=M¡log (rω(x))¢k (z − θ)k sequence in F (A0, A1). For that it will clearly suffice to consider the quantity 1,p 0 is a Cauchy (ω1)o. Our for all integers M and N which satisfy 0 ≤ M ≤ N and to show that E (M, N ) tends to 0 when M and N tend to ∞. We see that p k k! 1,p 0 1,p 0 N N p k k! e β(j+i t−θ)2 E (M, N )≤ e β(z−θ)2 N e β(z−θ)2¡log (rω(x))¢k (z − θ)k E (M, N ) :=°°°°° Xk=M°°°°° t∈R ¯¯¯ ¡ j + i t − θ¢k¯¯¯ φ(x)°°°A0 =°°°¡log (rω(x))¢k °°°¡log (rω(x))¢k ≤ 21+ 1 p Λ+ k Ξ´ max p Λk−1³2 °°φ°°A0 =°°φ°°Ai = max °°°¡log (rω(x))¢k φ(x)°°°A1 j∈{0,1}°°φ°°W 1,p(ω j ) j∈{0,1}°°φ°°W 1,p(ω j ) Xk=M max j=0,1 p k k! sup 1 = φ (ω0) ,°°ψ°°W NoN∈N φ(x)°°°°°F (A0,A1) φ(x)°°°°°F (A0,A1) φ(x)°°°A j . for all k ∈ N. The analogue of this for k = 0 is simply and obviously INTERPOLATION OF WEIGHTED SOBOLEV SPACES 25 We also have, for each k ∈ N, that sup (4.9) ≤ sup t≥0 max j=0,1 t∈R¯¯¯ e β(1−t 2)¡1+ t 2¢ e β(j+i t−θ)2 ¡ j + i t − θ¢k¯¯¯ e−β(1+t 2)¡1+ t 2¢ k 2 k 2 = e 2β sup t≥0 e−βy 2 y k, = e 2β sup y≥1 and a simple computation shows that (4.10) y k e−βy 2 max y≥0 k 2 =µ k 2βe¶ for each k > 0. These preceding inequalities show that E (M, N )≤ max where, for each k ∈ N, 21+ 1 γk = p Λk−1³2 and for k = 0, since max j=0,1 supt∈R¯¯¯ 1 Xk=M j∈{0,1}°°φ°°W 1,p(ω j ) p Λ+ k Ξ´ µ k 2βe¶ e β(j+i t−θ)2¯¯¯ ≤ e β, γ0 = e β. p k k! N γk k 2 e 2β, For each k ∈ N we have that 1 1 k 2 k , · 1 1 Λ Λ 1 1 1 = γk+1 γk = 2 pk + 1 p2βe ·µ k + 1 k ¶ ·µ1+ k¶ pk + 1 p Λ+ (k + 1)Ξ´ p(k + 1) ·³2 p Λ+ k Ξ´ ³2 p Λ+ (k + 1)Ξ´ pp2βe ·³2 p Λ+ k Ξ´ ³2 γk+1 γk = 0. This shows that the positive term from which it is clear that limk→∞ series P∞k=0 γk is convergent and this implies that E (M, N ) indeed has the be- haviour required to show thatn f is a Cauchy sequence in F (A0, A1). There- fore n f converges in F (A0, A1) norm to an element g ∈ F (A0, A1). In our setting, where A0 and A1 are the same space with the same norm, each ele- ment h ∈ F (A0, A1) is a continuous bounded A0 valued function on S which is analytic in the interior of S. Therefore an obvious extension of the Phragmen- Lindelöf theorem to Banach space valued analytic functions shows that NoN∈N NoN∈N φ φ kh(z)kA0 ≤ sup ζ∈∂Skh(ζ)kA0 = khkF (A0,A1) 26 MICHAEL CWIKEL AND AMIT EINAV for each z ∈ S. In particular, this means that, for each fixed z ∈ S, g (z) ∈ A0 and (4.11) lim f N→∞°°° φ N (z)− g (z)°°° A0 = 0. φ lows from (4.11) that, for each fixed z ∈ S the sequencen f Since g (z) is an element of A0 and therefore a measurable function, or rather an equivalence class of measurable functions on U , it will be convenient to let g (x, z) denote the value of g (z) at each (or almost every) point x of U . It fol- converges to g (z) also in Lp (ω) when ω is either one of the weight functions ω0 and ω1. By standard results, a norm convergent sequence in any (weighted or unweighted) Lp space must have a subsequence which converges pointwise at almost every point of the underlying measure space. So there exists a subset Uz of U pos- sibly depending on z, and an unbounded subsequence {Nk}k∈N of positive in- tegers, also possibly depending on z, such that U \ Uz has measure zero and g (x, z) = limk→∞ f But, from the formula (4.8) we see that the "whole" sequencenf verges pointwise for every x ∈ U and that its pointwise limit is N (x, z)oN∈N (x, z) for all x ∈ Uz . N (z)oN∈N con- φ Nk φ (z−θ) p e log(rω(x))+β(z−θ)2 φ(x) = f φ β,θ(x, z). This shows that f φ(x, z) = g (x, z) for every x ∈ Uz which means that f φ(z) and (ω1) for each z ∈ S. Consequently g (z) are the same element of W f φ coincides with g and is therefore an element of F (A0, A1). Since A0 is con- tinuously embedded in W (ω1) (ω0)∩ W 1,p 0 1,p 0 it follows that F (A0, A1) is continuously embedded in F³W and this completes our proof of Lemma 4.11. (ω0) and A1 is continuously embedded in W (ω0) ,W 1,p 0 1,p 0 1,p 0 1,p 0 Remark 4.12. Note that the fact f φ(z) and g (z) are the same element of W W (ω1) for each z ∈ S provides an alternative proof of the inclusion (4.4). 1,p 0 1,p 0 1,p 0 (ω0),W 4.11, via its proof, also provides the means to obtain a expression which is an As well as showing that f φ is an element of F³W upper bound for°° f φ°°F³W (ω1)´, Lemma (ω1)´. But that very complicated expression6 is not useful for our needs here. Instead we need the upper bound which will be provided by the next result: Proposition 4.13. Let (ω0, ω1) form a (θ, p)−admissible pair for some θ ∈ (0,1) and p ∈ [1,∞), and let φ be an element of Li pc. Then the element f (ω0),W 1,p 0 1,p 0 (ω0),W 1,p 0 1,p 0 (cid:3) (ω1)´ (ω0)∩ 1,p 0 (4.12) φ φ β,θ of F³W β,θ°°°F³W (ω0),W 1,p 0 f °°° (ω1)´ ≤ 2e β max(1, 1,p 0 (ω1)´ defined by (4.3) satisfies pp2βe)°°φ°°W p (θ,rω) . e β 6It depends in quite complicated ways on φ, β, p, ω0 and ω1 including dependence on ex- pressions involving the set Vφ and the numbers Λ and Ξ which themselves have complicated dependence on φ, ω0 and ω1. INTERPOLATION OF WEIGHTED SOBOLEV SPACES 27 Consequently, we also obtain that (4.13) 1,p 0 (ω0),W 1,p 0 °°φ°°hW (ω1)iθ ≤ Cp°°φ°°W p (θ,rω) for every φ ∈ Li pc, for a positive constant Cp depending only on p, which can be chosen to be de- fined as in (1.14). φ Remark 4.14. In the statement of Proposition 4.13 we have used the notation β,θ instead of its abbreviated variant f φ, and we will sometimes also do this in f the following proof of this proposition, in some places where it is necessary to emphasise the dependence of this element on the parameter β. Proof of Proposition 4.13. We first notice that ¯¯ f φ(x, z)¯¯ = rω(x) Rez−θ p e β((Rez−θ)2−(Imz)2)¯¯φ(x)¯¯ . Then we recall that, as shown just before (4.6) in the proof of Lemma 4.11, there exists a set U# (which of course depends on our given function φ ∈ Li pc) which is almost all of U and consists of all points x for which the pointwise gradients of log (rω(x)) and of φ(x) both exist. It obviously follows that, at every point x in this same set U# and for each fixed z ∈ S, the pointwise gradient with respect to x of f φ(x, z) exists and equals (z−θ) p ∇x f φ(x, z) = e and therefore also log(rω(x))+β(z−θ)2µ∇x φ(x)+ p e β((Rez−θ)2−(Imz)2)¯¯¯¯ Rez−θ z − θ p φ(x)∇x log (rω(x))¶, φ(x)∇x log (rω(x))¯¯¯¯ z − θ p . ∇xφ(x)+ ¯¯∇x f φ(x, z)¯¯ = rω(x) Since φ ∈ Li pc, Lemma 4.10 ensures that we also have f φ (·, z) ∈ Li pc. Therefore, by (2) of Theorem 4.8, the weak gradients of each of the functions φ and f φ(·, z) exist and coincide almost everywhere with their respective pointwise gradients. We will need to use this fact because the norms appearing in (4.12) are defined via weak gradients, rather than pointwise gradients. We will also need to use the fact that, for all x ∈ U and j ∈ {0,1}, ω j (x) = ωθ(x). rω(x) j−θω j (x) =µ ω0(x) ω1(x)¶ j−θ We compute p ω j (x)d x e rω(x) j−θe p Lp(ω j ) =U pβ³( j−θ)2 °° f φ(·, j + i t )°° =U −t 2´¯¯φ(x)¯¯ p Lp (ωθ) . Then we also see that the weak gradients of f φ(·, z) and of φ satisfy z − θ °°∇x f φ(·, j + i t )°° p Lp(ω j ) =U ∇x φ(x)+ pβ³( j−θ)2 −t 2´¯¯φ(x)¯¯ ωθ(x)d x ≤ e pβ°°φ°° pβ³( j−θ)2 rω(x) j−θe p p −t 2´¯¯¯¯ p ω j (x)d x φ(x)∇x log (rω(x))¯¯¯¯ 28 MICHAEL CWIKEL AND AMIT EINAV =U e ∇x φ(x)+ z − θ p which, by (4.5), does not exceed (4.14) pβ³( j−θ)2 −t 2´¯¯¯¯ −t 2´ ¯¯∇x φ(x)¯¯ p ωθ(x)d x φ(x)∇x log (rω(x))¯¯¯¯ + t 2´ p 2 +³¡ j − θ¢2 p e p p pβ³( j−θ)2 ¯¯φ(x)¯¯ 2pU In turn, in order to estimate this expression we use the fact that¯¯j − θ¯¯ ≤ 1, for one of its terms. For the other term we recall and use the calculations of (4.9) and (4.10) which together, when k = 1, give us that ≤µ 1 2βe¶ p¯¯∇x log (rω(x))¯¯ β³( j−θ)2 e 2β. e 1 2 1 2 −t 2´³¡ j − θ¢2 These considerations enable us to deduce that the expression (4.14) is domi- nated by + t 2´ p  ωθ(x)d x. 2p e pβ°°∇x φ°° p Lp (ωθ) + 2p We conclude that p 1,p 0 1,p 0 (ω0),W °° f φ°°F³W (ω1)´ = sup ≤ Lp (ωθ) + 2p e pβ°°∇x φ°° e pβ°°φ°° ≤ 2p e pβ°°φ°° W 1,p (ωθ) + 2p p e 2pβ p p . 1,p 0 ¡2βe¢ Lp(ωθ∇ log(rω)p) 2 p p °°φ°° t∈R³°° f φ(·, i t )°°W (ω0) ,°° f φ(·,1+ i t )°°W Lp (ωθ) + 2p 2 p p °°φ°° ¡2βe¢ Lp(ωθ∇ log(rω)p) 2 p p °°φ°°  (ω1)´ Lp(ωθ∇ log(rω)p)  e 2pβ e 2pβ 1,p 0 1 p p p p . p p ¡2βe¢ 1 p Recalling the notation introduced in (1.10), we see that the preceeding ex- pression is dominated by 2p e pβ,2p max    with°°φ°°hW 1,p 0 (ω0),W This establishes (4.12). Since f 1 p p e 2pβ 2 p p  °°φ°°W p (θ,rω) = 2e β max(1,  ¡2βe¢  β,θ(x, θ) = φ(x), we conclude that φ ∈hW β,θ°°°F³W (ω1)iθ ≤°°° (ω1)´ (ω0),W 1,p 0 1,p 0 1,p 0 φ φ f e β pp2βe)°°φ°°W p (θ,rω) . . Combining this inequality (ω0) ,W 1,p 0 with (4.12) and taking the infimum over all values of β in the interval (0,∞), (an infimum which is obviously attained) establishes (4.13) when Cp is the constant defined in (1.14). This completes the proof of the proposition. (cid:3) We are finally ready to present the proof of Theorem 4.5. In fact nearly all of the tools needed for the proof have already been developed by the preceding results of this section. It remains only to use an approximation argument to 1,p 0 (ω1)iθ INTERPOLATION OF WEIGHTED SOBOLEV SPACES 29 extend the result which we already have obtained for functions in Li pc to all functions in W p 0 (θ, rω). Proof of Theorem 4.5. Let φ be an arbitrary function in the space W view of the definition of this space, there exists a sequence of functions©φnªn∈N ∈ Li pc such that p 0 (θ, rω). In 0. and since, in view of (4.13), °°φ− φn°°W p (θ,rω) −→n→∞ (ω1)iθ (ω1)iθ ≤ Cp°°φn − φm°°W p (θ,rω) −→n,m→∞ (ω1)iθ (ω0),W 1,p 0 (ω0),W 1,p 0 1,p 0 1,p 0 1,p 0 0, . AshW (ω1)iθ 1,p 0 (ω0),W 1,p 0 such (ω1)iθ Since Li pc ⊂hW 1,p 0 (ω0) ,W 1,p 0 °°φn − φm°°hW 1,p 0 (ω0),W we find that©φnªn∈N is a Cauchy sequence inhW is a Banach space we conclude that there exists g ∈hW (ω1)iθ −→n→∞ (ω0),W that 1,p 0 1,p 0 °°φn − g°°hW 0. The relevant definitions immediately imply7 (cf. part (iii) of Theorem 2.7) that (ω0),W 1,p 0 1,p 0 hW (ω1)iθ ⊂£W 1,p (ω0),W 1,p (ω1)¤θ , and k·k[W 1,p (ω0),W 1,p (ω1)]θ ≤ k·khW According to Theorem 3.3, g ∈ W 1,p (ωθ), and (ω1)iθ °°φn − g°°W 1,p (ωθ) ≤°°φn − g°°[W 1,p (ω0),W 1,p (ω1)]θ (ω0),W (ω0),W 1,p 0 1,p 0 . ≤°°φn − g°°hW 1,p 0 1,p 0 (ω1)iθ −→n→∞ 0. p Therefore ©φnªn∈N converges to g in W 1,p (ωθ). However, since ©φnªn∈N con- 0 (θ, rω) it must also converge to φ in W 1,p (ωθ). As W 1,p (ωθ) is verges to φ in W a Banach space due to the (θ, p)−admissibility of the pair (ω0, ω1), we conclude that g = φ. Thus φ is an element of hW and©φnªn∈N con- verges to it in the norm of this space. To complete this proof it remains only to observe that inequality (4.2) follows immediately from the facts that (ω1)iθ (ω0),W 1,p 0 1,p 0 lim n→∞°°φn°°hW 1,p 0 (ω0),W 1,p 0 (ω1)iθ =°°φ°°hW 1,p 0 (ω0),W 1,p 0 (ω1)iθ and together with (4.13). lim n→∞°°φn°°W p (θ,rω) =°°φ°°W p (θ,rω) , (cid:3) 7In fact we already took note of and used this simple fact while proving Theorem 3.5. 30 MICHAEL CWIKEL AND AMIT EINAV 5. APPROXIMATION THEOREMS Theorems 3.3 and 4.5 are the principal ingredients that we require for proving our main theorem, Theorem 1.16, as well as for most of the proof of Theorem 1.21. As further preparation for these tasks, we require some approximation re- sults that will allow us to show that the conditions imposed in Theorems 1.16 and 1.21 in fact also suffice to ensure that Li pc¡Rd¢ and even also C∞c ¡Rd¢ are dense in W 1,p¡Rd , ω¢ and also in W p¡Rd , θ, rω¢. These approximation results are the focus of this section, with the following as our main goal: Theorem 5.1. Let ω0 and ω1 be weight functions that satisfy the compact bound- edness condition on Rd . Assume in addition that the function rω(x) defined in (1.9) is locally Lipschitz on Rd . Then, for each p ∈ [1,∞), (5.1) and (5.2) 1,p 0 W 1,p³Rd , ω j´ = W W p³Rd , θ, rω´ = W ³Rd , ω j´ for j ∈ {0,1} 0 ³Rd , θ, rω´ for each θ ∈ (0,1). p The proof of Theorem 5.1 relies on several approximation results, which we will establish via a series of lemmas. We start by showing that the space W 1,p (U , ω) has two important properties whenever ω satisfies the compact boundedness condition. Lemma 5.2. Let U be an open subset of Rd , let p ∈ [1.∞) and let ω be a weight function which satisfies the compact boundedness condition on U . Then (i) W 1,p (U , ω) is a Banach space, and (ii) W 1,p (U , ω) is contained in W 1,p l oc (U ). In fact these two properties also hold if our hypothesis on the weight function ω is replaced by the weaker requirement that (5.3) inf x∈K ω(x) > 0 for every compact subset K of U . Furthermore, in order to obtain that (i) holds, it suffices to assume that ω only satisfies an even weaker condition, namely that (5.4) 1 ω1/p χK ∈ Lp′ (U ) for every compact subset K of U . where, as usual, 1/p + 1/p′ = 1. Remark 5.3. For p = ∞ the space W 1,∞(U , ω) has both of the properties (i) and (ii) in the statement of Lemma 5.2 for all choices of the weight function ω. This is simply because, as already pointed out in Remark 1.11, W 1,∞(U , ω) coincides isometrically with the unweighted space W 1,∞(U ), for which (ii) is a triviality and (i) is a standard result. (Cf. e.g., Theorem 2 of [6, p. 249].) INTERPOLATION OF WEIGHTED SOBOLEV SPACES 31 Remark 5.4. In fact, one can also deduce that W 1,p (ω) is a Banach space for p ∈ (1,∞) from Theorem 1.11 of [12]. For p > 1 the condition (5.4) is just a refor- mulation of the condition ω ∈ Bp(Ω) introduced in Definition 1.4 on page 538 of [12] with Ω in the role of our set U . So part (i) of this lemma and our proof of it are very closely related to the statement and proof of Theorem 1.11 of [12, pp. 540 -- 541]. Here, unlike in that theorem, we also deal with the case where p = 1 and we can observe that the condition (5.4), when p = 1, would be the natural substitute for the condition ω ∈ Bp(Ω) of [12] and in fact is exactly the same as (5.3). Proof of Lemma 5.2. One should keep in mind that we are adopting the frequently used convention that, for each point a of Cd , a denotes the ℓ2 norm of a. Since the ℓ2 and ℓp norms are equivalent on Cd we have, for every f ∈ Lp (U , ω, Cd ), that (5.5) r°° f°°Lp (U,ω,Cd ) ≤°°¯¯ f¯¯°°Lp (U,ω) =µU¯¯ f (x)¯¯ ≤ R°°f°°Lp (U,ω,Cd ) Using Hölder's inequality and then (5.5), for each f ∈ Lp¡U , ω, Cd¢, for each compact subset K of U and for each g ∈ L∞(U ) which vanishes on U \ K we obtain that for some constants r and R depending only on d and p. ω(x)d x¶1/p p χK χK χK f (x)g (x)d x¯¯¯¯ Then, by quite similar reasoning, for K as above, but when the function f is in U ¯¯¯¯ (5.6) ω1/p°°°°Lp′ (U) ω1/p°°°°Lp′ (U) ω1/p°°°°Lp′ (U) ≤U¯¯ f (x)¯¯¯¯g (x)¯¯ d x ≤°°g°°L∞(U)U¯¯f (x)¯¯ χK (x)d x ≤°°g°°L∞(U)°°¯¯ f¯¯ ω1/p°°Lp (U)°°°° =°°g°°L∞(U)°°¯¯ f¯¯°°Lp (U,ω)°°°° ≤ R°°g°°L∞(U)°° f°°Lp (U,ω,Cd )°°°° ≤U¯¯ f (x)¯¯¯¯g (x)¯¯ d x ≤ pd°° f°°L∞(U,ω)°°g°°Lp (U,ω)°°°° After these preparations we turn to showing that (i) holds: Since ω is measur- L∞¡U , Cd¢ and vanishes on U \K and when g is an arbitrary element of Lp (U , ω), pd°° f°°L∞(U,ω)U¯¯g (x)¯¯ χK (x)d x ¯¯¯¯ able and strictly positive and since °°φ°°Lp (U,ω) ≤°°φ°°W 1,p (U,ω), it is obvious that k·kW 1,p (ω) is a norm. Now let©φnªn∈N be an arbitrary Cauchy sequence in W 1,p (U , ω). This obviously implies that©φnªn∈N is also a Cauchy sequence in Lp (U , ω) and that©∇φnªn∈N is a Cauchy sequence in Lp (U , ω, Cd ). Since Lp (U , ω) and Lp (U , ω, Cd ) are both Banach spaces, these sequences ©φnªn∈N and©∇φnªn∈N converge re- spectively, in the norms of these spaces, to an element φ ∈ Lp (U , ω) and an ele- ment ξ ∈ Lp (U , ω, Cd ). Therefore, in order to show that W 1,p (U , ω) is complete, f (x)g (x)d x¯¯¯¯ . χK ω1/p°°°°Lp′ (U) ≤ we have that U (5.7) 32 MICHAEL CWIKEL AND AMIT EINAV it remains only to show that ξ is the weak gradient of φ on U . Let ψ be an arbitrary function in C∞c (U ) and let Kψ denote a compact subset of U on which ψ is supported. We claim that (5.8) and also that (5.9) lim n→∞U ψ(x)¡∇φn(x)− ξ(x)¢ d x = 0. n→∞U ∇ψ(x)¡φ(x)− φn(x)¢ d x = 0 lim χKψ ω1/p°°°Lp′ (U) We shall obtain these two formulae with the help of the inequalities (5.6) and (5.7). We shall choose K in both of these inequalities to be Kψ. If we require ω to satisfy the compact boundedness condition or if we merely make do with requiring one of the weaker variants (5.3) and (5.4) of this condition, we will ob- is finite. We use this fact and substitute f (x) = this gives us (5.8). Then we substitute f (x) = ∇ψ(x) and g (x) = φ(x)− φn(x) in Using (5.8) and then (5.9), we see that tain that the quantity°°° ∇φn(x)− ξ(x) and g (x) = ψ(x) in (5.6). Since limn→∞°°∇φn − ξ°°Lp(U,ω,Cd ) = 0, (5.7). Since limn→∞°°φ− φn°°Lp (U,ω) = 0, this gives us (5.9). n→∞U ∇ψ(x)φn(x)d x U ∇ψ(x)φ(x)d x = lim n→∞U ψ(x)∇φn(x)d x = − lim = −U ψ(x)ξ(x)d x for each ψ ∈ C∞c (U ). This shows that ξ is indeed the weak gradient of φ on U and completes the proof of part (i) of the lemma, also when ω only satisfies (5.3) or (5.4). We now turn our attention to part (ii). Here our hypothesis on ω is either that it satisfies the compact boundedness condition, or merely one "half" of that con- dition, namely (5.3). For each compact subset K of U and each φ ∈ W 1,p (U , ω) we have, recalling the notation introduced in (1.7), that Lp(U,ω,Cd )´ °°χK φ°° Lp (U) +°°χK ∇φ°° Lp(U,Cd ) ≤ 1 1 p p p m(K , ω)p ³°°χK φ°° m(K , ω)p ³°°φ°° m(K , ω)p °°φ°° 1 ≤ = p p Lp (U,ω) +°°χK∇φ°° Lp (U,ω) +°°∇φ°° Lp(U,ω,Cd )´ p p W 1,p (U,ω) . This establishes the inclusion claimed in part (ii), and therefore completes the proof of the lemma. (cid:3) Remark 5.5. In the formulations of several of the following results of this section we will find it convenient to require the relevant weight function to satisfy the INTERPOLATION OF WEIGHTED SOBOLEV SPACES 33 compact boundedness condition, even though, just as for the preceding lemma, some of these results hold under weaker hypotheses. Lemma 5.6. Let ω be a weight function that satisfies the compact boundedness condition on an open set U . Then Li pc (U ) ⊂ W 1,p (U , ω) for each p ∈ [1,∞]. Proof. Recalling Rademacher's theorem, Theorem 4.8, we know that if φ ∈ Li pc (U ) then it is in W 1,∞l oc (U ), and therefore has a weak gradient ∇φ on U . Obviously (cf. Remark 4.9) ∇φ ∈ L∞(U , Cd ) and°°∇φ°°L∞(U,Cd ) is bounded by the Lipschitz con- stant of φ. Also φ, like every other compactly supported Lipschitz function, is bounded. This completes the proof in the case where p = ∞. To deal with the case p ∈ [1,∞), we note that φ and ∇φ both vanish almost ev- erywhere on U \suppφ, and that the Lebesgue measure of suppφ is finite. Then, using the notation of (1.7), and once more the boundedness of φ we conclude that p p p p and U¯¯φ(x)¯¯ U¯¯∇φ(x)¯¯ ω(x)d x ≤°°φ°° ω(x)d x ≤°°∇φ°° L∞(U)¯¯suppφ¯¯ M¡suppφ, ω¢ < ∞8 L∞(U,Cd )¯¯suppφ¯¯ M¡suppφ, ω¢ < ∞. This shows that φ ∈ W 1,p (U , ω) also for p ∈ [1,∞), completing the proof. Lemma 5.7. Let ω be a weight function that satisfies the compact boundedness condition on an open set U . Then, given p ∈ [1,∞), and functions φ ∈ W 1,p (U , ω) and ξ : U → C such that ξ is Lipschitz on U , we have that their pointwise product φξ has a weak gradient which can be expressed in terms of the weak gradients of φ and ξ by the formula (cid:3) ∇¡φξ¢(x) = ξ(x)∇φ(x)+ φ(x)∇ξ(x) for almost every x ∈ U . 1,p Proof. Much as in the proof of the previous lemma, we know, due to Rademacher's Theorem and Remark 4.9, that ξ ∈ W 1,∞l oc (U ) (and in fact ξ ∈ W 1,∞ (U )), and con- sequently ξ ∈ W 1,r l oc (U ) for every r ≥ 1. We also note that, due to Lemma 5.2, l oc (U ). Let ψ be an arbitrary function in C∞c (U ) and let Kψ = suppψ. Stan- φ ∈ W dard approximation theorems which are used in the theory of Sobolev spaces (see for example Theorem 1 of [6, pp. 250-251] and its proof, which is based on Theorem 6 of [6, pp. 630-631]) enable us to find an open set V containing Kψ, whose closure is compact and contained in U , and functions ξn : U → C whose restrictions to V are in C∞ (V ) and which satisfy (5.10) and (5.11) lim n→∞V ξn(x)− ξ(x)q d x = 0 lim n→∞V ∇ξn(x)−∇ξ(x)q d x = 0. 8Here (and also elsewhere) we use the standard notation E for the d−dimensional Lebesgue measure of any measurable subset E of Rd . 34 MICHAEL CWIKEL AND AMIT EINAV where q is the Hölder conjugate of p. Since ψ(x)ξn(x) ∈ C∞c (U ) for each n ∈ N we see that U φ(x)ξn(x)∇ψ(x)d x =U φ(x)∇¡ψ(x)ξn(x)¢ d x −U φ(x)ψ(x)∇ξn(x)d x = −U ψ(x)¡ξn(x)∇φ(x)+ φ(x)∇ξn(x)¢ d x. 1,p We recall that ψ and ∇ψ both vanish on U \ V . Therefore, in all the integrals in the preceding calculation, we can replace the domain of integration U by V . l oc (U ) (due to (ii) of Lemma 5.2), we also observe that φ∇ψ and ψ∇φ Since φ ∈ W are both elements of Lp (U , Cd ) and ψφ ∈ Lp (U ). So we can use (5.10), (5.11) and Hölder's inequality to pass to the limit in the preceding calculation and thus to obtain that U φ(x)ξ(x)∇ψ(x)d x = −U ψ(x)¡ξ(x)∇φ(x)+ φ(x)∇ξ(x)¢ d x. Since this holds for every ψ ∈ C∞c (U ), the proof is complete. (cid:3) From this point onward, till the end of this section, we will only deal with the case where U = Rd , and, accordingly, we shall usually remove any mention of the underlying set (namely Rd ) from our notation. Note that in the following lemma we will be permitting a function ρ which seems to be playing the role of a weight function to possibly also take the value 0. So, analogously to the conventions adopted in Definition 1.19 and in (1.15), it may happen that Lp (ρ) and k·kLp (ρ) have to be understood to be, respectively, a semi- normed space and the semi-norm on that space. Lemma 5.8. Let ρ : Rd → [0,∞) and φ : Rd → C be arbitrary measurable functions which satisfy for some p ∈ [1,∞). For each n ∈ N define the function ξn : Rd → [0,∞) by p Rd¯¯φ(x)¯¯ ξn(x) = 1 2− xn  0 ρ(x)d x < ∞, x < n n < x < 2n x ≥ 2n (5.12) Then, the d -dimensional ℓp norm of its weak gradient satisfies (i) ξn ∈ Li pc¡Rd¢ with Lipschitz constant 1 k∇ξn(x)kℓp ≤ (5.13) 1 p d n . n , and at almost every point x ∈ Rd (ii) φξn ∈ Lp¡ρ¢ and (5.14) Rd¯¯φ(x)ξn(x)− φ(x)¯¯ p ρ(x)d x −→n→∞ 0. INTERPOLATION OF WEIGHTED SOBOLEV SPACES 35 Proof. For part (i) the fact that ξn is in Li pc¡Rd¢ with Lipschitz constant 1 n is immediate. It is also obvious that, at each point x ∈ Rd such that x 6= n and x 6= 2n, the pointwise gradient ∇ξn(x) exists and equals either 0 or − x nx This gives us (5.13) since every x ∈ Rd satisfies . 1 kxkℓp ≤ d p kxkℓ∞ . We also have For part (ii) we note that ¯¯φ(x)ξn(x)¯¯ ≤¯¯φ(x)¯¯ which ensures that φξn ∈ Lp(ρ). ρ ∈ L1 and since φ(x)ξn(x) for all n ∈ N and for almost every x ∈ Rd . Since¯¯φ¯¯ converges pointwise to φ(x) we obtain (5.14) as a consequence of the Domi- nated Convergence Theorem. (cid:3) 1− ξn(x)p ≤¯¯φ(x)¯¯ ¯¯φ(x)ξn(x)− φ(x)¯¯ =¯¯φ(x)¯¯ p p p p Lemma 5.9. Let ν be a weight function that satisfies the compact boundedness condition and let ξn be as in (5.12). Then, for each p ∈ [1,∞) and for every φ ∈ W 1,p (ν), we have that φξn ∈ W 1,p (ν) and (5.15) 0. Proof. Our hypothesis on ν enables us to use Lemma 5.7 and so we see that φξn has a weak gradient on Rd , which is given by °°φ− φξn°°W 1,p (ν) −→n→∞ ξn(x)∇φ(x)+ φ(x)∇ξn(x). Since ξn and ∇ξn are bounded, for each fixed n, and since φ ∈ W 1,p (ν) we see that φξn ∈ Lp (ν) and ∇¡φξn¢ ∈ Lp¡ν, Cd¢. This implies that φξn ∈ W 1,p (ν). More- over, with the help of (5.13), we have that °°∇φ−∇¡φξn¢°°Lp(ν,Cd) ≤°°∇φ− ξn∇φ°°Lp(ν,Cd) +°°φ∇ξn°°Lp(ν,Cd ) d 1 p 1 p ≤µRd 1− ξn(x)p¯¯∇φ(x)¯¯ p ℓp ν(x)d x¶ + n °°φ°°Lp (ν) . So, as in the proof of part (ii) of Lemma 5.8, the Dominated Convergence Theo- rem obviously provides what is needed to show that And of course Lemma 5.8 also gives us that lim n→∞°°∇φ−∇¡φξn¢°°Lp(ν,Cd) = 0. These two conditions show that (5.15) holds, and thus complete the proof of the present lemma. (cid:3) lim n→∞°°φ− φξn°°Lp (ν) = 0. A very useful feature of Lemmas 5.8 and 5.9 is the fact that the approximating sequence in the relevant semi-normed weighted Lp space or weighted Sobolev space is always constructed in the same way, regardless of which p or weight function we use! This fact immediately proves the following obvious extension of these two lemmas which will be essential for us when we seek to approxi- mate functions in the normed space W p (θ, rω), precisely because its norm is the 36 MICHAEL CWIKEL AND AMIT EINAV sum of the (semi-)norms of two differently weighted spaces. As in Lemma 5.8 and Definition 1.19, here again we will encounter non-negative functions acting essentially as weights. This time there will be a finite collection of possibly dif- ferent non-negative functions ρβ and semi-normed spaces Lqβ(ρβ) associated with them. Lemma 5.10. Let ν1,..., ν j be weight functions that satisfy the compact bound- edness condition, and let ρ1,... , ρk be arbitrary non-negative measurable func- tions. Given j + k exponents p1,..., p j , q1,...., qk in [1,∞), we let X be the linear space X =³T j β=1 Lqβ¡ρβ¢´ equipped with the norm α=1 W 1,pα (να)´∩³Tk k j °°φ°°X := Xα=1°°φ°°W 1,pα (να) + Xβ=1°°φ°°L qβ(ρβ) . For each n ∈ N, let ξn be the function defined by (5.12). Then, for every φ ∈ X , the function φξn is also in X and lim n→∞°°φ− φξn°°X = 0. Now that we know how to approximate functions of interest to us by com- pactly supported functions, our next step will be to approximate those approxi- mating functions in turn by smoother functions, not merely Lipschitz functions lowing standard result: but even elements of C∞c ¡Rd¢. An essential ingredient for doing this is the fol- Lemma 5.11. Let η ∈ C∞c ¡Rd¢ be a given non-negative function which is sup- ported in B1(0) and satisfies ´Rd η(x)d x = 1. For p ∈ [1,∞) let φ be a given func- tion in Lp¡Rd¢ that is compactly supported. Define: η¡n(x − y)¢ φ(y)d y = ηn ∗ φ(x), φn(x) := ndRd where ηn(x) = nd η(nx). Then (i) φn ∈ C∞c ¡Rd¢ and is supported in Dn =©x ∈ Rd dist¡x, suppφ¢ ≤ 1 nª. (ii) φn converges to φ in Lp¡Rd¢. (iii) If in addition φ ∈ W 1,p¡Rd¢ we have that φn ∈ W 1,p¡Rd¢ and lim n→∞°°φ− φn°°W 1,p(Rd ) = 0. We refer to standard textbooks for the standard proof 9 of this result. We will now obtain a variant of Lemma 5.10 as our next step towards proving our principal result in this section, Theorem 5.1. This variant will be stated in much greater generality than we need for our immediate purposes here - but that comes at no extra cost in the proof. That greater generality may well turn out to be useful in future extensions of this research. 9For example, it is a straightforward variant of the proof of Theorem 6 of [6, p. 630]. See rele- vant material on pp. 629-631 of [6] and also, for example, on pp. 148-149 of [23]. INTERPOLATION OF WEIGHTED SOBOLEV SPACES 37 Theorem 5.12. Let X be the normed space introduced in Lemma 5.10 and sup- pose that each of the weight functions ν1,..., ν j appearing in its definition sat- isfies the compact boundedness condition. Suppose also that each of the non- negative measurable functions ρ1,..., ρk appearing in its definition is bounded from above on every compact set and that (5.16) max©q1,...., qkª ≤ max©p1,..., p jª . Then C∞c (Rd ) is a dense subspace of X . Proof. The inclusion C∞c (Rd ) ⊂ X follows immediately from the fact that all the functions ν1,..., ν j and ρ1,..., ρk are bounded above on every compact set. Let φ be an arbitrary compactly supported element of X and let φn = ηn ∗ φ for each n ∈ N, where the functions ηn are as specified in Lemma 5.11. Then, by Lemma 5.11, all of the functions φn are in C∞c (Rd ) and they and their gradients vanish on the complement of the compact set D1 =©x ∈ Rd : dist(x,suppφ) ≤ 2ª. Furthermore, our hypotheses ensure that the supremum x∈D1½ M := sup max α∈{1,..., j },β∈{1,...,k}©να(x), ρβ(x)ª¾ is finite. This enables us to use the inequalities (5.17) and (5.18) 1 qβ (ρβ) ≤ M °°φ− φn°°L °°φ− φn°°W 1,pα (να) ≤ M qβ (Rd ) qβ °°φ− φn°°L pα °°φ− φn°°W 1,pα (Rd ) 1 which hold for all n ∈ N, α ∈ {1,..., j } and β ∈ {1,..., k}. The fact that infx∈D1 να(x) > 0 ensures that φ ∈ W 1,pα¡Rd¢ and therefore also φ ∈ Lpα(Rd ) for each α ∈©1,..., jª. Consequently, in view of (5.16), Hölder's inequality and the fact that the support of φ is compact, we also have that φ ∈ Lqβ(Rd ) for each β ∈ {1,.., k}. The membership of φ in all of these unweighted Lp spaces permits us to use Lemma 5.11 to obtain that the right sides of (5.17) and (5.18) tend to 0 as n tends to ∞ for all α ∈ {1,..., j } and β ∈ {1,..., k}. Consequently we also have that lim n→∞°°φ− φn°°X = 0. This shows that every compactly supported element in X can be approximated in X norm by a sequence of functions in C∞c ¡Rd¢. Since we already know from Lemma 5.10 that every element in X can be approximated in X norm by a se- quence of compactly supported functions in X , this completes the proof of the theorem. (cid:3) We are now ready to prove Theorem 5.1: Proof of Theorem 5.1. We shall in fact prove a stronger result, namely that C∞c (Rd ) is dense in each of the spaces W 1,p¡Rd , ω0¢, W 1,p¡Rd , ω1¢ and W p¡Rd , θ, rω¢. 38 MICHAEL CWIKEL AND AMIT EINAV This of course will imply (5.1) and (5.2), since C∞c ¡Rd¢ ⊂ Li pc¡Rd¢. We begin by applying Theorem 5.12 in the special case where j = k = 1, p1 = q1 = p, ν1 = ωm, ρ1 = 0 and where m = 0 or m = 1. In this case the theorem shows that C∞c ¡Rd¢ is dense in X = W 1,p¡Rd , ωm¢. This establishes (5.1). It remains to prove (5.2). We start by recalling (cf. Remark 1.5) that if ω0 and ω1 satisfy the compact boundedness condition, then obviously so do ωθ = ω1−θ 0 ωθ 1 and rω. As rω is locally Lipschitz, we use Rademacher's theorem to conclude that it is differentiable almost everywhere with essentially bounded gradient on each compact set. Consequently, due to the strict positivity of rω, the function log rω is also differentiable almost everywhere and its gradient, given by ∇log (rω(x)) = ∇rω(x) rω(x) is also essentially bounded on each compact set. We summarise the above: , • ωθ satisfies the compact boundedness condition. • ωθ¯¯∇log (rω)¯¯ ρ1 = ωθ¯¯∇log rω¯¯ p is essentially bounded from above on every compact set. These two properties are exactly what we need to enable us to apply Theorem 5.12 once more, where this time we choose j = k = 1, p1 = q1 = p, ν1 = ωθ and p . These choices make X coincide with W p (θ, ωθ) to within equivalence of norms, and so the density of C∞c (Rd ) in X , which is again guar- anteed by Theorem 5.12, establishes (5.2) and therefore completes the proof of the theorem. (cid:3) In all of our results in this section from Lemma 5.8 onwards we have been dealing only with the case where the underlying open set U is all of Rd . But let us conclude this section with the following remark about the possibility of extending our approximation theorems to the case where U is a proper open subset of Rd : Remark 5.13. As the reader may have noticed from the various steps leading to the proof of Theorem 5.12, the two main ingredients which we used to show the theorem were the fact that we can approximate a function in W 1,p¡Rd , ω¢ by a compactly supported function and Lemma 5.11. As a more general version of Lemma 5.11 can be shown to be valid for any open set U ⊂ Rd instead of Rd (see e.g. [6] pp. 629-631), we see that if the open set U and the weight function ω are such that any φ ∈ W 1,p (U , ω) can be approximated by compactly supported functions, then we will have that W 1,p 0 (U , ω) = W 1,p (U , ω) , and also that C∞c (U ) is dense in W 1,p (U , ω) and in W p (U , θ, rω). This raises the question of what properties of U and ω might suffice to ensure that compactly supported functions are dense in W 1,p (U , ω). It would seem that this density property will hold if ω(x) is required to tend to 0 sufficiently rapidly p W 0 (U , θ, rω) = W p (U , θ, rω), INTERPOLATION OF WEIGHTED SOBOLEV SPACES 39 when x approaches the boundary of U . Thus it might be natural to consider this question for weight functions which depend solely on the distance to the boundary, such as those which appear in the weighted Sobolev spaces discussed on pages 17-20 of [11]. 6. PROOFS OF THE MAIN THEOREM, AND OF ITS MAIN CONSEQUENCES In this short section we pool all the resources that we have developed in Sec- tions §3, §4 and §5 to prove our main theorem, Theorem 1.16 and its main con- sequences: Theorems 1.20, 1.21 and 1.22. Proof of Theorem 1.16. Using Theorem 3.3 we obtain the inclusion (1.11) and one side of the inequality (1.13). To complete the proof it remains to establish the inclusion (1.12) and the other side of (1.13). We start doing this by observ- ing that, since ω0, ω1, and therefore obviously also ωθ, each satisfy the compact boundedness condition (cf. Remark 1.5), we can apply Lemma 5.2 and Lemma 5.6 to obtain that (1) For every p ∈ [1,∞), each of the spaces W 1,p (ω0), W 1,p (ω1) and also W 1,p (ωθ), for every θ ∈ (0,1), is a Banach space that contains Li pc. One of the hypotheses of Theorem 1.16 is that rω is locally Lipschitz on U . The strict positivity of rω therefore implies that (2) The function log rω is locally Lipschitz on U . The above two conditions (1) and (2) tell us (cf. Definition 4.1) that the pair (ω0, ω1) is¡θ, p¢−admissible for every θ ∈ (0,1) and p ∈ [1,∞). This permits us to invoke Theorem 4.5, which applies to any open subset U of Rd and whose con- clusions (4.1) and (4.2) are exactly the above mentioned inclusion and inequality which are required to complete the proof of Theorem 1.16. (cid:3) Proof of Theorem 1.20. The first of the two inclusions in (1.16) has already been established as part of the proof of Theorem 1.16. The additional condition im- posed in Theorem 1.20 for some p ∈ [1,∞) is exactly what is required to enable Theorem 3.5 to be applied in order to establish the second inclusion in (1.16) for that value of p and for all θ ∈ (0,1). (cid:3) Proof of Theorem 1.21. The proof of Theorem 1.16 shows that the inclusions (1.11) and (1.12) hold also here. So, to prove the first part of Theorem 1.21 we simply substitute (1.17) and (1.18) in (1.11) and (1.12) to obtain (1.19). To prove the sec- ond part of the theorem, in which it is assumed that U = Rd , we simply apply Theorem 5.1 to give us that (1.17) and (1.18) both hold for all p ∈ [1,∞). Proof of Theorem 1.22. Since g (x) is continuous and the exponential function is continuous and strictly positive, we conclude that if ω1 (or ω0) satisfies the compact boundedness condition, then so does ω0 (or ω1). Next, we note that (cid:3) which implies, as g (x) is Lipschitz and the exponential function is C 1 on R, that rω(x) is locally Lipschitz. Thus we have shown that the conditions of Theorem rω(x) = e g (x), 40 MICHAEL CWIKEL AND AMIT EINAV 1.16 are satisfied, as claimed in part (i) of Theorem 1.22. We next show part (ii), namely that (1.20) holds to within equivalence of norms when U = Rd . Theorem 1.16 (or even simply Theorem 3.3) gives us "half" of (1.20), i.e. the inclusion "⊂" and the corresponding norm inequality. So we only need to show that W 1,p³Rd , ωθ´ ⊂hW 1,p³Rd , ω0´ ,W 1,p³Rd , ω1´iθ and to obtain the reverse norm inequality corresponding to this reverse inclu- sion. In view of (1.12) and of Theorem 1.16 and Theorem 5.1 we can in turn reduce this task to showing that the inclusion W 1,p³Rd , ωθ´ ⊂ W p³Rd , θ, rω´ holds and is continuous. For any given element φ ∈ W 1,p (ωθ) we see that (6.1) (6.2) This implies that φ is automatically in W p (θ, rω) and p p °°φ°° Lp(ωθ∇ log(rω)p) =U¯¯φ(x)¯¯ L∞°°φ°° W 1,p (ωθ) +°°φ°° L∞¢ ≤°°∇g°° °°φ°°W p (θ,rω) =³°°φ°° ≤¡1+°°∇g°° p 1 p p ωθ(x)d x p¯¯∇g (x)¯¯ p Lp (ωθ) . 1 p p Lp(ωθ∇ log(rω)p)´ p °°φ°°W 1,p (ωθ) . This establishes (6.1) and the required associated norm inequality and thus takes care of part (ii). (In fact (6.2) also shows that the three norms appearing in (1.13) are equivalent to each other.) Finally we establish the remaining parts (iii) and (iv) of Theorem 1.22 by using exactly the same considerations, together with Theorems 1.20 and 1.21. (cid:3) In the last two sections of the paper we will discuss additional topics related to our study, including an example of when W p (θ, rω) is not W 1,p (ωθ), and some potential applications and open problems. 7. MORE ABOUT THE SPACE W p (θ, rω) In this section we will discuss two more properties of the space W p (θ, rω). Namely, we will answer the following two questions: Question 3. Is it possible for W p (θ, rω) to not be W 1,p (ωθ)? Question 4. Can we find a smaller, yet "big enough", subspace of W p (θ, rω) that will allow us to only consider the weight ωθ? Our answers will give us more information about the space£W 1,p (U , ω0),W 1,p (U , ω1)¤θ. INTERPOLATION OF WEIGHTED SOBOLEV SPACES 41 7.1. Can W p (θ, rω) be smaller than W 1,p (ωθ)? We consider it very unlikely that the Stein-Weiss like formula (1.4) could hold for all choices of the set U and weight functions ω0 and ω1. Here we shall begin some attempts towards finding a counterexample. One might expect that when U has a very complicated structure or when some or all of the hypotheses of Theorem 1.16 do not hold, this could prevent (1.4) from holding. But the examples which we are about to present hint that maybe (1.4) can fail even when U is simply all of Rd and the above mentioned hypothe- ses do hold. From this point onwards we will take U to be Rd and remove almost all mention of the underlying set from our notation. As can be seen from (1.19) in Theorem 1.21, a sufficient condition for obtaining a Stein-Weiss like theorem when U = Rd is that W p (θ, rω) = W 1,p (ωθ) and indeed (since obviously W p (θ, rω) ⊂ W 1,p (ωθ)) this condition is effectively what we have used in the proof of Theorem 1.22. (Cf. (6.1).) So a natural first step in searching for the above mentioned counterexample is to look for pairs of weight functions for which (7.1) W p (θ, rω) ⊂6= W 1,p (ωθ) . We shall describe such pairs. In fact, for each choice of θ ∈ (0,1), we will find such a pair which satisfies the hypotheses of Theorem 1.16. But we should em- phasize that at this stage we do not know whether these weight functions, which satisfy (7.1), will also turn out to provide a counterexample to (1.4). To simplify the discussion we shall make do with presenting explicit formulas for our weight functions only in the case p = 1. However it is not difficult to modify our exam- ples so that they apply when p ∈ (1,∞). Let us start by stating the relevant properties that a function, say φ, which is in W 1,p (ωθ) but not in W p (θ, rω) must have: Since φ ∈ W 1,p (ωθ) we have that φ has a weak gradient on Rd and (7.2) On the other hand, since φ 6∈ W p (θ, rω) we must have that (7.3) p Rd¯¯φ(x)¯¯ p ωθ(x)d x < ∞, Rd¯¯∇φ(x)¯¯ Rd¯¯φ(x)¯¯ p¯¯∇log (rω(x))¯¯ ωθ(x)d x = ∞. p ωθ(x)d x < ∞. We shall simply choose φ ≡ 1. Then the gradient condition in (7.2) is automat- ically satisfied, and conditions (7.2) and (7.3) become the following two condi- tions on ω0 and ω1: (7.4) ωθ(x)d x = ∞. Let us simplify things further by also choosing ω0 ≡ 1. In this case ∇logµ ω0(x) ω1(x)¶¯¯¯¯ ωθ ∈ L1³Rd´, Rd¯¯¯¯ p ωθ(x) = ω1(x)θ, 42 MICHAEL CWIKEL AND AMIT EINAV and we can see that showing that there is a weight function ωθ which satisfies (7.4) in this case is equivalent to finding a weight function ω (which will play the role of ωθ) such that: (7.5) Rd ω(x)d x < ∞, Rd ∇ω(x)p ω(x)1−p d x = ∞. It is clear that finding such a function ω will in fact show that (7.4) can be sat- isfied for every choice of θ ∈ (0,1) and for the particular value of p appearing in both (7.5) and (7.4). At this stage we shall specialise to the case p = 1. In this case (7.5) is simply the statement that ω is a function in L1¡Rd¢ whose gradient is not in L1¡Rd , Cd¢. For various obvious examples of integrable functions ω which have some monotone decay at infinity (for example, like the reciprocal of a polynomial, or of an expo- nential) it turns out that (7.5) does not hold. However, as we shall now see, if we modify such a function by adding a term which oscillates rapidly near in- finity and is suitably bounded, then the resulting function ω does satisfy (7.5): Consider the strictly positive C 1 function (7.6) = ω(x) = 1 1 ¡1+x2¢αµ1+ sin³xβ´+ ¡1+x2¢αµ1+ 1+x2¶+ 1+x2¶ sin¡xβ¢ ¡1+x2¢α . 1 1 We shall now show that one can easily choose positive numbers α and β for which ω will satisfy (7.5) for p = 1. In fact, a similar computation shows that for any p ∈ (1,∞), one can find an α and a β for which (7.6) will satisfy (7.5). Clearly, if α > d 2 then ω(x) ∈ L1¡Rd¢ since ω(x) ≤ 3 Next, as and and ∇ω(x) = 2x ¡1+x2¢α+2 x 1 − −2α ¡1+x2¢α . 1+x2¶ x − βxβ−2 cos¡xβ¢ ¡1+x2¢α 1+x2¶ x¯¯¯¯¯ ≤ ¡1+x2¢α+1 ¡1+x2¢α , ≤ ¡1+x2¢α+1µ1+ 2αsin¡xβ¢ ¡1+x2¢α+1 x + ¡1+x2¢α+1µ1+ ¯¯¯¯¯ ¡1+x2¢α+2¯¯¯¯¯ ¯¯¯¯¯ ¡1+x2¢α+1 x¯¯¯¯¯ sin¡xβ¢ ≤ 2x 1 1 1 1 ¯¯¯¯¯ 1 ¡1+x2¢α , INTERPOLATION OF WEIGHTED SOBOLEV SPACES 43 We now compute we see that if α > d 2 then ∇ω ∈ L1¡Rd , Cd¢ if and only if x¯¯¯¯¯ Iα,β :=Rd¯¯¯¯¯ xβ−2 cos¡xβ¢ ¡1+x2¢α xβ−1¯¯cos¡xβ¢¯¯ Iα,β =Rd ¡1+x2¢α r β+d−2¯¯cos¡r β¢¯¯ Sd−1¯¯ d r ≥¯¯ 2α ∞ ¡1+ r 2¢α Sd−1¯¯ =y=r β ¯¯ 2αβ ∞ If, for example, we choose α = d and β = d + 1, then this last integral is infinite. Since we then also have α > d 2 as required in the previous part of this discussion, this completes our proof that (7.1) can hold for p = 1 and for every θ ∈ (0,1) for suitable choices of weight functions. r β+d−2−2α¯¯¯ ¯¯cos¡y¢¯¯ d r cos³r β´¯¯¯ =¯¯¯ Sd−1¯¯¯ d−1−2α β y 1 ∞ 0 d x < ∞. d x 1 d r Remark 7.1. As we already mentioned, there exist other examples of pairs of weight functions which satisfy (7.1) for p ∈ (1,∞). Some of them nevertheless also satisfy (1.4). We refer the reader to the appendix where we give an example of a pair of weight functions (even in C∞¡Rd¢) which in fact satisfies both (7.1) and (1.4) for every p ∈ [1,∞) and θ ∈ (0,1). 7.2. A simpler subspace of£W 1,p (U , ω0),W 1,p (U , ω1)¤θ. The main result of this paper shows that, under relatively mild conditions, the interpolation space £W 1,p (U , ω0),W 1,p (U , ω1)¤θ lies between W 1,p (U , ωθ) and a smaller subspace W p (U , θ, rω), which can be expressed as W 1,p (U , ωθ)∩ X (with an appropriate norm) where X is the semi- normed weighted space Lp¡U , ωθ¯¯∇log rω¯¯ For some applications, such as those which we will shortly discuss in 8.2, it will be convenient to find a more simply defined space X , which is embedded con- tinuously in X . Such a space will therefore yield the chain of inclusions p¢ (see Definition 1.19). W 1,p (U , ωθ)∩ X ⊂£W 1,p (U , ω0),W 1,p (U , ω1)¤θ ⊂ W 1,p (U , ωθ). A simple X to choose, if possible, would be a standard weighted Lebesgue space with respect to the weight function ωθ alone. We shall now discuss one possible way which can enable us to choose an X of this kind, however with an expo- nent larger than p. It requires us to impose one more condition on our weight functions. Definition 7.2. Let ω0 and ω1 be weight functions on an open subset U of Rd and let ωθ and rω be as defined throughout this paper (by (1.5) and (1.9)). Suppose 44 MICHAEL CWIKEL AND AMIT EINAV (7.7) that log rω has a weak gradient on U . For each q ∈ [1,∞) and each θ ∈ [0,1] let M¡U , θ, rω, q¢ be the quantity 1 q M¡U , θ, rω, q¢ :=µU¯¯∇log (rω(x))¯¯ q ωθ(x)d x¶ . (Note that here θ may also equal 0 or 1.) We will often use the abbreviated nota- tion M¡θ, q¢ for M¡U , θ, rω, q¢. Lemma 7.3. Let U , ω0, ω1, ωθ and rω have the properties stated in Definition 7.2. Suppose, furthermore, that there exist q ∈ (p,∞) and θ ∈ (0,1) for which M¡θ, q¢ < ∞. Then, in terms of the notation of Definition 1.19, we have that qp and (7.8) L q−p (U , ωθ) ⊂ Lp¡U , ωθ¯¯∇log (rω)¯¯ k·kLp(U,ωθ∇ log(rω)p) ≤ M¡θ, q¢k·kL p¢ qp q−p (U,ωθ) for these values of q and θ. Proof. Using Hölder's inequality with the exponent q p we find that U¯¯φ(x)¯¯ ≤µU¯¯∇log (rω(x))¯¯ q p¯¯∇log (rω(x))¯¯ ωθ(x)d x¶ p q µU¯¯φ(x)¯¯ p ωθ(x)d x This proves both the inclusion and (7.8). pq q−p ωθ(x)d x¶ q−p q . (cid:3) With this in hand we can deduce the following: Theorem 7.4. Let ω0 and ω1 be a pair of weight functions on U that satisfy the compact boundedness condition. Assume furthermore that is locally Lipschitz on U , and that U , ω0 and ω1 are such that the conditions of rω(x) = ω0(x) ω1(x) Theorem 1.21 hold. Then, if M¡θ, q¢, defined by (7.7), is finite for some q > p, we have that W 1,p (U , ωθ)∩ L qp q−p (U , ωθ) ⊂£W 1,p (U , ω0),W 1,p (U , ω1)¤θ . Moreover, there exists a constant Cp > 0 which depends only on p, such that for any φ ∈ W 1,p (U , ωθ)∩ L q−p (U , ωθ), qp (7.9) °°φ°°[W 1,p (U,ω0),W 1,p (U,ω1)]θ p W 1,p (U,ωθ) + M¡θ, q¢p°°φ°° p L ≤ Cpµ°°φ°° 1 p q−p (U,ωθ)¶ qp Proof. This theorem is an immediate consequence of Theorem 1.21 and Lemma 7.3. The inequality (7.9) follows immediately from (1.13) of Theorem 1.16 and (7.8). (cid:3) INTERPOLATION OF WEIGHTED SOBOLEV SPACES 45 Obviously, a modification of the above theorem can be made to accommo- date the more general case of Theorem 1.16 when we start considering appro- priate closures. We conclude this subsection, and section, with the following remark: Remark 7.5. A simple application of Hölder's inequality shows that if M¡0, q¢ and M¡1, q¢ are finite then < ∞ for every θ ∈ (0,1). M¡θ, q¢ ≤ M¡0, q¢1−θ M¡1, q¢θ 8. FINAL REMARKS In this last section of our work, we reflect on possible generalisations of our main results via notions of equivalence of weight functions, and possible appli- cations to evolution equations. Then, finally, we discuss some further possibili- ties for research related to the issue of interpolation of weighted Sobolev spaces. 8.1. Equivalence of weights. As our main results of this work depend strongly on the underlying weights of our respective spaces, one natural question can arise: Question 5. Can one change the weight under consideration to a 'better one' (i.e. suitable for our theorems) without changing the underlying Sobolev space? The answer to that is in the affirmative - in the obvious case when one deals with equivalent weights. Definition 8.1. Adopting standard terminology, we say that two strictly positive measurable functions ω and η are equivalent, if there exists a constant A > 0 such that A−1η(x) ≤ ω(x) ≤ Aη(x). It is immediate to check that if ω0 and ω1 satisfy the compact boundedness condition, and are equivalent, respectively, to η0 and η1, then the latter func- tions also satisfy the same condition. Moreover, W 1,p (ω) and W 1,p¡η¢ are iso- morphic to each other for equivalent weights. (Remark 1.4 has some relevance in this context.) Accordingly, one can immediately state and almost immediately prove more general versions of our main theorem and its main consequences, in which the weight functions ω0 and ω1 are required merely to be equivalent to other weight functions which satisfy the requirements in the current statements of those the- orems. 8.2. Possible applications to evolution equations. Many linear evolution equa- tions of the form (8.1) give rise to an evolution semigroup, {T (t )}t≥0, which acts smoothly on several spaces simultaneously. Moreover, in fields such as parabolic PDEs and kinetic theory, one frequently encounters evolution semigroups whose action on the ∂t f (t , x) = L ( f )(t , x) 46 MICHAEL CWIKEL AND AMIT EINAV initial datum of the problem produces functions which have increased smooth- ness, decay and integrability (hypercontractivity). One good example of semi- groups with such behaviour are the so-called Markov semigroups. We refer the reader to [1] for more information on the subject. In many cases, one is interested in the decay, or convergence to equilibrium as t tends to ∞, of the solution at time t of (8.1) with respect to an appropriate norm, such as k·kW 1,p(Rd ,ω). Under suitable hypotheses, Theorem 1.21, together with observations that we made in §7.2, allows us to interpolate rates of decay/convergence between two different weighted Sobolev spaces to an intermediate space of the same kind. More explicitly, if we know that for a given semigroup {T (t )}t≥0 the solution to (8.1) decays or converges to an equilibrium at rates that are given by the norms of T (t ) on appropriate subspaces of both W 1,p (U , ω0) and W 1,p (U , ω1), then it may be possible to deduce analogous information about the norm of T (t ) on a suitable subspace of W 1,p (U , ωθ), for θ ∈ (0,1). We shall now give one example of a setting where this can be done. Here we as- sume that the norms of T (t ) decay to zero. The underlying open set for all of the spaces is to be understood to be Rd . Theorem 8.2. Let (ω0, ω1) be a pair of weight functions that satisfy the compact boundedness condition on Rd , and such that rω, defined by (1.9), is locally Lips- chitz on Rd . Consider a semigroup of operators {T (t )}t>0 such that, for some fixed p ∈ [1,∞), (8.2) T (t ) :¡W 1,p (ω0),W 1,p (ω1)¢ →¡W 1,p (ω0),W 1,p (ω1)¢ for each t > 0. Assume furthermore that there exist q ∈ (p,∞), θ ∈ (0,1), t0 > 0 and a class of initial data G such that for every g ∈ G T (t0)g ∈ W 1,p (ωθ)∩ L qp q−p (ωθ). Suppose furthermore that M¡θ, q¢, defined in (7.7), is finite for these values of θ and q. Then we have, for every g ∈ G and t > t0, that T (t )g ∈ W 1,p (ωθ) and also that °°T (t )g°°W 1,p (ωθ) ≤ C kT (t − t0)k1−θ µ°°T (t0)g°°W 1,p (ωθ) + M¡θ, q¢°°T (t0)g°°L where the constant C depends only on p. B(W 1,p (ω0))kT (t − t0)kθ q−p (ωθ)¶ qp B(W 1,p (ω1)) Proof. We start by noting that, due to Theorem 7.4, for each g ∈ G we have that T (t0)g ∈£W 1,p (ω0),W 1,p (ω1)¤θ . Recalling Theorem 3.3 and the semigroup property T (t ) = T (t − t0)T (t0) INTERPOLATION OF WEIGHTED SOBOLEV SPACES 47 which holds for every t ≥ t0, we see that for each g ∈ G and t > t0 we have that T (t )g ∈£W 1,p (ω0),W 1,p (ω1)¤θ ⊂ W 1,p (ωθ). Furthermore, using this and (3.4) together with (8.2) and Theorem 2.4, we obtain that °°T (t )g°°W 1,p (ωθ) ≤°°T (t )g°°[W 1,p (ω0),W 1,p (ω1)]θ =°°T (t − t0)¡T (t0)g¢°°[W 1,p (ω0),W 1,p (ω1)]θ ≤ kT (t − t0)k1−θ B(W 1,p (ω0))kT (t − t0)kθ . B(W 1,p (ω1))°°T (t0)g°°[W 1,p (ω0),W 1,p (ω1)]θ Using (7.9) yields the inequality required to complete the proof of this theorem. (cid:3) 8.3. Future research and open problems. We feel excited about future possi- bilities in continuing to try and understand whether or not one can obtain a Stein-Weiss like theorem for weighted Sobolev spaces under more general hy- potheses on the underlying set and the weight functions. In particular, some of the problems which we think are worth investigating are the following: Question. Can one find an open set U and weights ω0 and ω1 on it for which W p (U , θ, rω) ⊂6=£W 1,p (U , ω0),W 1,p (U , ω1)¤θ? (U , ω) = W 1,p (U , ω) Question. Can one find necessary and sufficient conditions under which 0 (U , θ, rω) = W p (U , θ, rω)? and W 1,p 0 W p Let us mention that pages 550-554 of [12] contain a discussion of the closure of C∞c (U ) (there denoted by C∞0 (U )) in various weighted Sobolev spaces. A dif- ferent but somewhat related topic is considered on pages 43-49 of [11]. Question. Can one obtain similar results to Theorems 1.16, 1.20, 1.21 and 1.22 for weighted homogeneous Sobolev spaces W 1,p (U , ω) (i.e. where we only con- sider the norm of the weak gradient)? (The answer is evidently yes for each p ∈ [1,∞) in the very special case where d = 1 and U = R and both of the weight functions satisfy the compact bounded- ness condition. In that case the maps φ 7→ φ′ and its inverse ψ 7→´ x 0 ψ(t )d t give us an isometric identification between the two couples ¡ W 1,p (R, ω0), W 1,p (R, ω1)¢ and ¡Lp (R, ω0), Lp (R, ω1)¢ and so we can apply the "classical" Stein-Weiss theorem. A more detailed expla- nation can be found in the appendix.) We are confident that more can be done (and asked), and will be done in the next few years. APPENDIX A. ADDITIONAL PROOFS In this appendix we include some additional proofs that we felt would hinder the flow of the main body of the paper. 48 MICHAEL CWIKEL AND AMIT EINAV 1,p l oc (U ). In this part of the A.1. An alternative characterization of the space W appendix we will provide some rather straightforward arguments which can be used to justify the following claim, made immediately after Definition 4.7: 1,p l oc (U ) if and only if, for every open subset V of U whose clo- Claim A.1. f ∈ W sure is a compact subset of U , the restriction of f to V is an element of W 1,p (V ). In order to prove this claim, we require the following notation: For a given open set U , and each n ∈ N we define the open set Un = {x ∈ U dist(x, ∂U ) > 1/n}∩ Bn(0) (where Bn(0) is the open ball of radius n in Rd centred at the origin). The closure of Un is contained in U and in Bn(0) and is therefore a compact subset of U . Proof of Claim A.1. Suppose that the function f : U → C has the property that, for each n ∈ N, its restriction fn := f¯¯Un to Un has a weak gradient on Un. Let us denote this weak gradient by gn. We do not yet need to suppose also that fn ∈ W 1,p (Un) but later, when we do, we will also be able to assert that gn ∈ Lp (Un, Cd ). Due to the uniqueness of the weak gradient, and the fact that Un ⊂ Un1 ⊂ U for all n1 > n, we see that if n1 > n we have that gn1Un = gn. For each n ∈ N let vn : U → Cd be the function which coincides with gn on Un and equals ~0 on U \Un. Thus we have that (A.1) vn1(x) = gn1(x) = gn1(x)¯¯Un = gn(x) = vn(x) for all x ∈ Un and all n1 > n and all n ∈ N. This shows that, for each fixed n ∈ N, the pointwise limit limk→∞ vk(x) exists for all x ∈ Un. Since U =Sn∈NUn this in turn implies that this same pointwise limit limk→∞ vk(x) exists for all x ∈ U and so we can define a new function g : U → Cd by setting g (x) := lim k→∞ vk(x) for all x ∈ U . In view of (A.1) we have that, for each n ∈ N, (A.2) g (x) = gn(x) for all x ∈ Un We will now show that f has a weak gradient on U and that g is that gradient. Given an arbitrary φ ∈ C∞c (U ) let K be a compact set contained in U such that φ and therefore also ∇φ both vanish on U \K . There exists n ∈ N such that K ⊂ Un, and therefore the restriction of φ to Un is in C∞c (Un). We also have that (A.3) φ(x) = 0 and ∇φ(x) =~0 for all x ∈ U \Un. Consequently, using (A.3), then the definitions of fn and gn and then (A.2) and then (A.3) again, we obtain that U f ∇φd x =Un = −Un f ∇φd x =Un φg d x = −U φg d x. fn∇φd x = −Un φgnd x INTERPOLATION OF WEIGHTED SOBOLEV SPACES 49 Since the preceding argument holds for every φ ∈ C∞c (U ), this shows that g in- deed is the (necessarily unique) weak gradient of f on U . Suppose now that f has the property that, for every open subset V of U such that V is a compact subset of U , the restriction of f to V is an element of W 1,p (V ). Then the preceding argument shows that f has a weak gradient ∇ f on U which equals the function g obtained above. Furthermore, the functions fn and gn de- fined as above now satisfy fn ∈ W 1,p (Un) and gn ∈ Lp (Un, Cd ) and they coincide with the restrictions of f and of g respectively to the set Un . Thus we have χUn f ∈ Lp (U ) and χUn∇ f ∈ Lp (U , Cd ). Let K be an arbitrary compact subset of U . For some n ∈ N we have that K ⊂ Un and so χK f ∈ Lp (U ) and χK∇ f ∈ Lp (U , Cd ). This shows that f ∈ W We now prove the reverse implication. Suppose that f ∈ W an arbitrary open subset of U whose closure K := V is a compact subset of U . l oc (U ) and that ∇ f denotes its weak gradient on U . Let V be 1,p l oc (U ). 1,p 1,p Let g = f¯¯V and let ∇g denote the Cd -valued function defined only on V which 1,p l oc (U ) implies that χK f and is the weak gradient of g on V . The fact that f ∈ W χK∇ f are elements, respectively of Lp (U ) and Lp (U , Cd ). The restriction of χK f to V must of course equal g . Furthermore, by the definition of weak gradients and by their uniqueness, the restriction to V of χK ∇ f must equal ∇g . The fact l oc (U ) implies that χK f and χK ∇ f are elements, respectively of Lp (U ) that f ∈ W and Lp (U , Cd ). Therefore g and ∇g are in Lp (V ) and Lp (V, Cd ) respectively. In other words g ∈ W 1,p(V ), and this completes our proof of Claim A.1. A.2. From locally Lipschitz to globally Lipschitz - A proof of Fact (a). In this part of the appendix we will prove the following claim, which was formulated as the fact (a) just after the statement of Lemma 4.10: (cid:3) Claim A.2. If the function ψ : U → C is locally Lipschitz and has compact sup- port, then it is Lipschitz on all of U . Proof. Suppose that ψ : U → C is locally Lipschitz and has compact support. Then there exists a compact subset K of U such that ψ vanishes on U \ K . The number δ := dist(K , Rd \U ) is strictly positive and we let K1 be the set K1 =½x ∈ Rd : dist(x, K ) ≤ δ 2¾. Obviously K1 ⊂ U . The compactness of K implies that K1 is closed and bounded and therefore also compact. Our hypotheses on ψ and the compactness of K1 ensure that the quantities and C1 := sup½¯¯ψ(x)− ψ(y)¯¯ ¯¯x − y¯¯ : x, y ∈ K1, x 6= y¾ C2 := sup©¯¯ψ(x)¯¯ : x ∈ K1ª 50 MICHAEL CWIKEL AND AMIT EINAV are both finite. To complete the proof we will show that (A.4) ¯¯ψ(x)− ψ(y)¯¯ ≤¯¯x − y¯¯ max½C1, 2C2 δ ¾ for every x and y in U . The left side of (A.4) equals 0 whenever x and y are both in U \ K1 and it is bounded by C1¯¯x − y¯¯ whenever x and y are both in K1. Thus we only need to consider what happens when x ∈ K1 and y ∈ U \ K1. • If x ∉ K then we again have¯¯ψ(x)− ψ(y)¯¯ = 0. • If x ∈ K then¯¯x − y¯¯ > δ/2 and so ¯¯ψ(x)− ψ(y)¯¯ =¯¯ψ(x)¯¯ ≤ C2 = 2C2 δ · δ 2 ≤ 2C2 δ ¯¯x − y¯¯. We conclude that (A.4) indeed holds for all x and y in U and our proof is com- plete. (cid:3) A.3. A Stein-Weiss like theorem in the case where rω is not Lipschitz. In this subsection of the appendix we will focus on showing that there exist weight functions, ω0 and ω1, for which one can obtain a Stein-Weiss like theorem, (i.e. for which (1.4) is valid) even though rω, defined in (1.9), is not Lipschitz. This will show that Theorem 1.22 gives a sufficient yet not necessary condition for obtaining such a result. Our starting point for finding such an example is the following simple lemma: Lemma A.3. Let ρ0 and ρ1 be weight functions on Rd which satisfy the hypotheses of Theorem 1.22. (In particular this implies that the function log³ ρ0 on Rd .) Let ω1 be a weight function on Rd which satisfies cρ1(x) ≤ ω1(x) ≤ C ρ1(x) for all x ∈ Rd (A.5) for two positive constants c and C . If we set ω0 = ρ0, then we have that ρ1´ is Lipschitz hW 1,p³Rd , ω0´ ,W 1,p³Rd , ω1´iθ = W 1,p³Rd , ωθ´ to within equivalence of norms. Proof. This is an immediate consequence of Theorem 1.22 combined with the observations made in Subsection 8.1. (cid:3) Lemma A.3 gives us a possible approach for constructing our desired exam- ple: We start with a pair of weights that satisfy our Stein-Weiss like theorem, Theorem 1.22, and modify one of the weights in a way that keeps the new weight equivalent to the original weight, but violates the Lipschitz condition on rω. We shall do that now. Consider the weight functions ρ0 ≡ 1 and ρ1(x) = e−p1+x2 and ρ1 both satisfy the compact boundedness condition and that . It is obvious that ρ0 logµ ρ0(x) ρ1(x)¶ =p1+x2 INTERPOLATION OF WEIGHTED SOBOLEV SPACES 51 is Lipschitz on Rd . Thus, we conclude that ρ0 and ρ1 satisfy the conditions of Theorem 1.22. Now, let ω1 be the function ω1(x) = e−p1+x2−sin³ex2´. Clearly ω1 satisfies (A.5) with c = 1/e and C = e. Since rω = ω0/ω1 = 1/ω1 we see that ∇log (rω(x)) = ∇³p1+x2 + sin³ex2´´ = x p1+x2 + 2xex2 cos³ex2´. The above is unbounded on Rd , which gives us the example we needed. In fact, one can do more with this example. One can even use it to show that this pair of weight functions has the property that: W p¡Rd , θ, rω¢ is strictly smaller than W 1,p¡Rd , ωθ¢ for every p ∈ [1,∞) and θ ∈ (0,1). We leave the proof of this claim to the reader. This example also leads us naturally to ask: Question. Can one find weight functions ω0 and ω1 on Rd , that are not equiv- alent to weights that satisfy the conditions of Theorem 1.22 but for which we nevertheless have (1.4)? We'd like to conclude this subsection by noting that a theme common to this example and the one presented in Subsection 7.1 is that they both use weight functions which are the product of a "well behaved" function with a highly os- cillatory but bounded function. A.4. A comment about homogeneous weighted Sobolev spaces of univariate functions. In this last subsection we will show that, when the open set U is R, one can easily obtain a Stein-Weiss like theorem for the homogeneous weighted Sobolev spaces, W 1,p(R, ω), as a direct consequence of the regular Stein-Weiss theorem (as we claimed at the end of Subsection 8.3). We start with a definition Definition A.4. For each weight function ω : R → (0,∞) and each p ∈ [1,∞] let W 1,p(R, ω) be the space of equivalence classes modulo constants of measurable functions f : R → C which have a weak derivative f ′ in Lp (R, ω). This space is normed by°° f°° W 1,p (R,ω) =°° f ′°°Lp (R,ω). Suppose that ω satisfies the compact boundedness condition, or at least that it satisfies the weaker condition (5.4) which we used in Lemma 5.2. Then every function g in Lp (R, ω) is locally integrable and therefore the function G(x) := x 0 g (t )d t is defined for all x ∈ R and is absolutely continuous on every bounded inter- val. Moreover, its pointwise derivative G′ exists almost everywhere and coin- cides with g almost everywhere. The fact that g is also the weak derivative of G 52 MICHAEL CWIKEL AND AMIT EINAV is immediate due to the fact that we are on R and G is absolutely continuous. From all the above we see that the map T defined by T g (x) = x 0 g (t )d t is a continuous linear map from Lp (R, ω) onto W 1,p (R, ω). The inverse mapping of W 1,p(R, ω) onto Lp (R, ω) is of course simply the derivative map D defined by D f = f ′. It is also clear that T and D are both isometries. Since Lp (R, ω) is complete, we conclude that W 1,p (R, ω) is complete for all weight functions ω which satisfy (5.4). Now, given p0 and p1 in [1,∞) and weight functions ω0 and ω1 on R which both satisfy the compact boundedness condition (as we saw, we can also make do with somewhat weaker assumptions), we see that¡ W 1,p0 (R, ω1), W 1,p1 (R, ω1)¢ is a Banach couple. We use Theorem 2.4 to obtain that the operators T and D satisfy and T :£Lp0 (R, ω0), Lp1 (R, ω1)¤θ →£ W 1,p0 (R, ω1), W 1,p1 (R, ω1)¤θ D :£ W 1,p0 (R, ω1), W 1,p1 (R, ω1)¤θ →£Lp0 (R, ω0), Lp1 (R, ω1)¤θ , with norms that do not exceed 1. Since the weight function ωθ,pθ defined by (1.1) also has the compact bounded- ness property, we see that T and D are also isometries, respectively, of Lpθ¡R, ωθ,pθ¢ onto W 1,pθ¡R, ωθ,pθ¢ and of W 1,pθ¡R, ωθ,pθ¢ onto Lpθ¡R, ωθ,pθ¢. Combining these facts with the formula (3.3) shows that with equality of norms. £ W 1,p0 (R, ω1), W 1,p1 (R, ω1)¤θ = W 1,pθ¡R, ωθ,pθ¢, REFERENCES [1] D. Bakry, I. Gentil, and M. Ledoux. Analysis and Geometry of Markov Diffusion Operators, vol- ume 348 of Grundlehren der mathematischen Wissenschaften. Springer International Pub- lishing, Cham, 2014. [2] J. Bergh and J. Löfström, Interpolation Spaces, an introduction. Grundlehren der Mathema- tischen Wissenschaften, No. 223. Springer-Verlag, Berlin-New York, 1976. [3] A. P. Calderón, Intermediate spaces and interpolation, the complex method. Studia Math., 24 (1964), 113 -- 190. http://matwbn.icm.edu.pl/ksiazki/sm/sm24/sm24110.pdf [4] M. Cwikel, Complex interpolation, a discrete definition and reiteration, Indiana Univ. Math. J. 27 (1978), 1005 -- 1009. [5] D. E. Edmunds and H. Triebel, Function spaces, entropy numbers, differential operators. Cam- bridge Univ. Press, Cambridge, 1996. [6] C. L. Evans, Partial differential equations. Second edition. Graduate Studies in Mathematics, 19. American Mathematical Society, Providence, RI, 2010. xxii+749 pp. [7] P. Grisvard, Espaces intermédiaires entre espaces de Sobolev avec poids. (French) Ann. Scuola Norm. Sup. Pisa. 17 (1963), 255 -- 296. [8] D. D. Haroske and H. Triebel, Entropy numbers in weighted function spaces and eigenvalue distributions of some degenerate pseudodifferential operators. I, Math. Nachr. 167 (1994), 131 -- 156. INTERPOLATION OF WEIGHTED SOBOLEV SPACES 53 [9] D. D. Haroske and H. Triebel, Entropy numbers in weighted function spaces and eigenvalue distributions of some degenerate pseudodifferential operators. II, Math. Nachr. 168 (1994), 109 -- 137. [10] D. D. Haroske and H. Triebel, Wavelet bases and entropy numbers in weighted function spaces, Math. Nachr. 278 (2005), 108 -- 132. [11] A. Kufner, Weighted Sobolev spaces. Translated from the Czech. A Wiley-Interscience Publi- cation. John Wiley & Sons, Inc., New York, 1985. 116 pp. ISBN: 0-471-90367-1. [12] A. Kufner and B. Opic, How to define reasonably weighted Sobolev spaces. Comment. Math. Univ. Carolin. 25 (1984), 537 -- 554. [13] A. Kufner and A-M Sändig, Some applications of weighted Sobolev spaces, Teubner Texts in Mathematics, 100. B. G. Teubner Verlagsgesellschaft, Leipzig, 1987. 268 pp. [14] J. Löfström, Interpolation of weighted spaces of differentiable functions on Rd . Ann. Mat. Pura. Appl. 132 (1982), 189 -- 214. [15] S. G. Pyatkov, Interpolation of some function spaces and indefinite Sturm-Liouville problems, (English summary) Differential and integral operators (Regensburg, 1995), 179 -- 200, Oper. Theory Adv. Appl., 102, Birkhäuser, Basel, 1998. [16] S. G. Pyatkov. Interpolation of weighted Sobolev spaces. Siberian Adv. Math. 10 (2000), no. 3, 83 -- 132. [17] S. G. Pyatkov, Interpolation of weighted Sobolev spaces. (Russian. Russian summary) Mat. Tr. 4 (2001), no. 1 122 -- 173. [18] S. G. Pyatkov, Interpolation of Sobolev spaces and indefinite elliptic spectral problems, (Eng- lish summary) Recent advances in operator theory in Hilbert and Krein spaces, 265 -- 290, Oper. Theory Adv. Appl., 198, Birkhäuser Verlag, Basel, 2010. [19] H.-J. Schmeisser and H. Triebel, Topics in Fourier analysis and function spaces, Leipzig: Geest & Portig, 1987; Chichester; Wiley, 1987. [20] E. Stein, Interpolation of linear operators. Trans. Amer. Math. Soc. 83 (1956), 482 -- 492. [21] E. M. Stein and G. Weiss, Interpolation of operators with change of measures, Trans. Amer. Math. Soc. 87 (1958), 159 -- 172. [22] H. Triebel, Interpolation theory, function spaces, differential operators, North Holland Pub- lishing Company 1978. [23] R. Wheeden and A. Zygmund, Measure and integral. An introduction to real analysis. Pure and Applied Mathematics, Vol. 43. Marcel Dekker, Inc., New York-Basel, 1977. DEPARTMENT OF MATHEMATICS, TECHNION - ISRAEL INSTITUTE OF TECHNOLOGY, HAIFA 32000, ISRAEL E-mail address: [email protected] INSTITUT FÜR ANALYSIS UND SCIENTIFIC COMPUTING, TECHNISCHE UNIVERSITÄT WIEN, WIED- NER HAUPTSTRASSE 8-10 A-1040 VIENNA, AUSTRIA E-mail address: [email protected]
1106.3351	1	1106	2011-06-16T21:15:59	The spine of a Fourier-Stieltjes algebra: corrigenda	[ "math.FA", "math.OA" ]	Some unfortunate errors from our paper math/0505591 are corrected.	math.FA	math	THE SPINE OF A FOURIER-STIELTJES ALGEBRA: CORRIGENDA MONICA ILIE AND NICO SPRONK It has come to the authors' attention that there are several errors in our paper [3]. Fortunately, these errors are correctable. Some of these errors carry on to modest, though mainly cosmetic, errors in follow-up papers [5] and [4]. We are grateful to El¸cim Elgun for pointing out the gap in the proof of [3, Theorem 2.2] and the error in the proof of [3, Theorem 5.1]. We are also grateful to Pekka Salmi for pointing out the flaw in the statement of [3, Theorem 4.2] and suggesting its correct form. We appeal to [3] for pertinent notation and terminology. 1. On non-quotient locally precompact topologies The following should replace [3, Theorem 2.2]. Theorem 1.1. Let τ ∈ T (G). Then the following hold. (i) There exists a unique τnq in Tnq(G) such that τ is a quotient of τnq. (ii) If G is abelian, then τnq = τ ∨ τap. In [4, Theorem 2.2] it is claimed that τnq = τ ∨ τap for general locally compact groups G and τ in T (G). This is false. For example let G = SL2(R). We have that, on G, τap is the trivial topology ε = {∅, G}. Now if q : G → G/{−I, I} is the quotient map and τ = q−1(τG/{−I,I}) is the coarsest topology making q continuous, then we have that τnq = τG whereas τ ∨ τap = τ ( τG. may be used to show that s 7→ (cid:0)ητ1 The proof of part (ii) proceeds exactly as does the proof of [3, Theorem 2.2]. In particular, in the second paragraph of that proof, [3, Lemma 2.3] τ1 is a bicontin- uous isomorphism, since Gτ1, being abelian, is maximally almost periodic. The proof of part (i), however, demands more care. We fix τ0 in T (G) ap) : Gτ1 → Gτ × Gap τ (s), ητ1 and let (1.1) Qτ0 = {τ ∈ T (G) : τ0 is a quotient of τ }. Date: October 6, 2018. 2000 Mathematics Subject Classification. Primary 43A30; Secondary 43A60, 43A07, 46L07, 22B05. Key words and phrases. Locally precompact topology, spine. We like to thank NSERC for partial funding of our work. 1 2 MONICA ILIE AND NICO SPRONK Lemma 1.2. (i) If τ1, τ2 ∈ Qτ0 then τ1 ∨ τ2 ∈ Qτ0 as well. (ii) If τ1, τ2 ∈ T (G) satisfy τ0 ⊆ τ1 ⊆ τ2 and τ2 ∈ Qτ0 , then τ1 is a quotient of τ2. Proof. (i) For j = 1, 2 we let K j τ0 . Our assumptions provide that for j = 1, 2, K j ∼= Gτ0 . We identify Gτ1∨τ2 as a subgroup of Gτ1 × Gτ2 and define K = Gτ1∨τ2 ∩ (K 1 0 ). We let q : Gτ1∨τ2 → Gτ1∨τ2/K denote the quotient map. We obtain the following commutative diagram 0 = ker ητj 0 is compact and Gτj /K j 0 × K 2 0 Gτ1∨τ2 q Gτ1∨τ2/K / Gτ1 × Gτ2 ητ1 τ0 ×ητ2 τ0 +WWWWWWWWWWWWWWWWWWWWWW / Gτ1 × Gτ2 /(K 1 0 × K 2 0 ) ∼= Gτ1 /K 1 0 × Gτ2/K 2 0 ∼ Gτ0 × Gτ0. Identifying Gτ0 with the diagonal subgroup of Gτ0 × Gτ0, this diagram shows that the map ητ1∨τ2 τ0 Gτ1 ∨τ2 is a proper map. τ0 × ητ2 = ητ1 τ0 (ii) We recall that our assumptions give the following commuting diagram (1.2) Gτ2 ητ2 τ1 H H H H ητ2 τ0 H H H H #H Gτ2 /K 2 0 / Gτ1 ητ1 τ0 ∼ Gτ0 τ0 is compact. Then K 2 0 = ker ητ2 where K 2 0 , and is thus compact. We let q2 1 be the quotient map and τ = (q2 τ1 is injective. The commuting diagram (1.2) and the first isomorphism theorem give the commuting diagram ), so that Gτ ∼= Gτ2/K 1 1 : Gτ2 → Gτ2 /K 2 1 , τ ⊇ τ1 and ητ 1 ◦ητ2)−1(τGτ2 /K 2 1 = ker ητ2 τ1 ⊂ K 2 1 (1.3) Gτ2 ητ2 τ / Gτ ητ τ1 / Gτ1 R R R R R R F F F q F F R R R R F F (R R R R R ητ2 τ0 "F ητ1 τ0 Gτ /K ∼ Gτ0 where K = K 2 0 /K 2 1 and q : Gτ → Gτ /K is the quotient map. Since ητ τ1 is injective, it suffices to prove that it is open for ητ2 τ1 to be a quotient map, i.e. we obtain that τ1 = τ where ητ2 is a quotient map. To this τ end, let U ⊂ Gτ be relatively compact open set. Then U K is also relatively compact and open. Hence ητ τ1 (U K), and furthermore τ1(U K) is closed and equal to ητ (1.4) ητ τ1 U K : U K → ητ τ1(U K) is a homeomorphism. The commuting diagram (1.3) tells us that ητ τ1(U K) = (ητ1 τ0 )−1(cid:0)q(U )(cid:1) / + / # / ( / " / THE SPINE which is open in Gτ1. Hence by (1.4), ητ τ1U K : U K → ητ τ1 (U K) is a homeomorphism onto an open subset so ητ τ1(U ) is open. 3 (cid:3) Proof of Theorem 1.1 (i). We first note that Qτ , as defined in (1.1), is a directed system: if τ1, τ2 ∈ Qτ then τ1 ∨ τ2 ∈ Qτ by Lemma 1.2 (i). Moreover, it follows Lemma 1.2 (ii) that if τ1, τ2 ∈ Qτ with τ1 ⊆ τ2, then ητ2 τ1 is a proper map. Hence the inverse mapping system τ1 : τ ′ ∈ Qτ , τ1 ⊂ τ2 in Qτ } {Gτ ′, ητ2 gives rise to the projective limit Gτ ′ : ητ2 τ1 (sτ2) = sτ1 if τ1 ⊆ τ2 in Qτ Gτ ′ : ητ ′ τ (sτ ′) = sτ for τ ′ in Qτ GQτ = lim ←− τ ′∈Qτ Gτ ′ =   =   (sτ ′) ∈ Y τ ′∈Qτ (sτ ′) ∈ Y τ ′∈Qτ     which is locally compact by [3, Proposition 2.1]. We let τnq denote the coars- est topology which makes the map s 7→ (cid:0)ητ ′(s)(cid:1) : G → GQτ , continuous. We now show that τnq ∈ Qτ . First observe that ητnq ∼= GQτ → Gτ We have that τnq ⊇ τ ′ for every τ ′ in Qτ . : Gτnq τ is given by the map (sτ ′) 7→ sτ . With this identification we have that ker ητnq τ = lim ←− τ ′∈Qτ ker ητ ′ τ and is thus compact. Moreover ητnq τ is open, since for any basic open set V = Y τ ′∈Qτ (cid:31){τ ′ 1,...,τ ′ n} Gτ ′ × nY j=1 Uτ ′ j ⊂ Y τ ′∈Qτ Gτ ′ where each Uτ ′ j is open in Gτ ′ j , we have ητnq τ (V ∩ GQτ ) = n\ j=1 τ ′ j η τ (Uτ ′ j ) ) is open by assumption. where each η τ ′ j τ (Uτ ′ j Finally, if there were τ1 in T (G) of which τnq is a quotient, then by the first isomorphism theorem we would have that τ1 ∈ Qτ . Hence τ1 ⊆ τnq. Thus τnq is a non-quotient topology, and the unique such one of which τ is a quotient. (cid:3) For any locally compact group G for which τnq = τ ∨τap for any τ in T (G), we obtained in [3, Section 2.4] that Tnq(G) = T (G) ∨ τap and is thus an ideal in, and hence a subsemilattice of, the semilattice (T (G), ∨). We note that for any τ for which Gτ is maximally almost periodic, we have τnq = τ ∨ τap. 4 MONICA ILIE AND NICO SPRONK Unfortunately, it is not clear whether Tnq(G) is a subsemilattice of T (G), in general. However the following is immediate. Corollary 1.3. If τ1, τ2 ∈ Tnq(G) define τ1 ∨τ2 = (τ1 ∨ τ2)nq. Then (Tnq(G), ∨) is a quotient semilattice of (T (G), ∨). The only inobvious aspect of this corollary is the associativity of ∨. We observe that τ1 ∨ (τ2 ∨τ3) admits τ1 ∨ τ2 ∨ τ3 as a quotient, and hence, by Lemma 1.2 (ii), is itself a quotient of (τ1 ∨ τ2 ∨ τ3)nq. Symmetrically, the same is true of (τ1 ∨τ2) ∨ τ3. Hence τ1 ∨(τ2 ∨τ3) = (τ1 ∨ τ2 ∨ τ3)nq = (τ1 ∨τ2)∨τ3. Unless it can be shown that τ1 ∨τ2 = τ1 ∨ τ2 for all τ1, τ2 ∈ Tnq(G), some changes have to be made to the exposition in [3, Section 4], where ∨ must always be replaced by, or understood to be, ∨. Fortunately, this change does not appear to affect any of the results or proofs in this section, or any later part of the paper, in more than a cosmetic manner. 2. Topology of the spine compactification As shown in [3, Theorem 4.1], with appropriate notational changes as suggested by Corollary 1.3, the spectrum of the algebra A∗(G) is given by G∗ = G S∈HD(G) GS where each GS is a projective limit over a hereditary directed subset S of (Tnq(G), ∨). The statement of [3, Theorem 4.2] is flawed, and should be replaced with the following. Theorem 2.1. The topology on G∗ is given as follows: for any s0 in G∗, say s0 ∈ GS0 for some S0 in HD(G), a neighbourhood base at s0 is formed by the sets U (Vτ ; Wτ1, . . . , Wτn ) =(cid:8) s ∈ G∗ :s ∈ GS for some S ⊇ Sτ in HD(G) with sτ ∈ Vτ , and sτj ∈ Wτj if S ⊇ Sτj , for j = 1, . . . , n (cid:9) where τ ∈ S0, τ1, . . . , τn ∈ Tnq(G) \ S0, Vτ is an open neighbourhood of s0,τ in Gτ , and each Wτj is a cocompact subset of Gτj . Pekka Salmi has pointed out to us that the error in the description of [3, Theorem 4.2], implies that all groups GS , for S ∈ HD(G) are locally compact. This is false, as is implicit in [3, Section 6.3], or is shown in [1, Theorem 2]. Unfortunately, the proof of [3, Theorem 4.2] requires slight modification. THE SPINE 5 Proof of Theorem 2.1. We should first observe that the family of sets described above indeed is a base for a topology. It is straightforward to check that for τ, τ ′ ∈ S0 and τ1, . . . , τn, τ ′ m in Tnq(G) \ S0 that 1, . . . , τ ′ m) , . . . , Wτ ′ U (Vτ ; Wτ1 , . . . , Wτn ) ∩ U (Vτ ′; Wτ ′ 1 ⊃ U (cid:16)(ητ ∨τ ′ τ )−1(Vτ ) ∩ (ητ ∨τ ′ τ ′ )−1(Vτ ′); Wτ1 , . . . , Wτn , Wτ ′ 1 , . . . , Wτ ′ n(cid:17) m for neighbourhoods Vτ of s0,τ , Vτ ′ of s0,τ ′, and cocompact subsets Wτ1, . . . , Wτ ′ of Gτ1 , . . . , Gτ ′ m , respectively. We note that for a net (si)i∈I in G∗, the following are equivalent: (i) si → s0 in G∗ with the topology described above; (ii) for each τ0 ∈ S0 there is i0 such that i ≥ i0 implies that si ∈ GSi for some Si ⊃ Sτ0 and limi≥i0 si,τ = s0,τ ; and for each τ 6∈ S0 for which Iτ = {i : si ∈ Si for some Si ⊃ Sτ } is admits no maximal element in I, and for any co-compact Wτ ⊂ Gτ , there is iτ in I for which si,τ ∈ Wτ if i ∈ Iτ and i ≥ iτ ; (iii) χsi → χs0 weak* in A∗(G)∗. The equivalence of (i) and (ii) is clear. If we write u in A∗(G) as u = Pτ ∈Tnq(G) uτ as in [3, (4.2)], and suppose (ii), above, then χsi(u) = X uτ (si,τ ) + X uτ (si,τ ) → X uτ (s0,τ ) = χs0(u) τ ∈S0\Si τ ∈Si\S0 τ ∈S0 where si ∈ Si for each i; which shows (iii). Likewise, selecting u = uτ , a repeat of the computation above shows that (iii) implies (ii). (cid:3) We remark that [3, Corollary 4.3] remains unchanged. The neighbourhood in the proof of part (ii), therein, should be changed to U (Vτ0) = {s ∈ G∗ : s ∈ GS ′ for some S ′ ⊃ Sτ0 and Sτ0 ∈ Vτ0}. Furthermore, a modest change must be made in the proof of [3, Proposition 4.6 (ii)]. If a net (ei) of idempotents converges to s in GS , then for all τ ∈ S, there is an iτ for which Si ⊃ Sτ , where ei ∈ GSi, for i ≥ iτ , and limi≥iτ ei,τ = sτ ; and for τ in Tnq(G) \ S, there is iτ for which Si 6⊃ Sτ for i ≥ iτ , which may be seen by observing that otherwise such ei,τ must be within the cocompact set Wτ = Gτ \ {ητ (e)}. 3. On abelian groups The fourth paragraph of the proof of [3, Theorem 5.1] contains an error bicontinuous is false, for obviously only its inverse is continuous. Fortunately the claim which that paragraph is attempting to establish, namely that if in its claim that the map from cGτ to the diagonal subgroup of cGτ × bGd is τ in T (G) is such that cGτ is the group bG but with a finer locally compact : dGτ ′ → bG is continuous and injective. The in T (G), the dual map cητ ′ assumptions on τ above imply that bητ is also surjective. Since the map group topology τ , then τ ∈ Tnq(G), remains true. Recall that for any τ ′ 6 MONICA ILIE AND NICO SPRONK ητnq τ example. Since ητ = ητnq dητnq : Gτnq → Gτ is proper, its dual map dητnq is continuous, bijective and open, hence τ = cτnq. Thus by Pontryagin ◦ητnq we have bητ = dητnq τ duality τ = τnq. ◦dητnq τ τ τ is open; see [2, (23.24)(d)], for . Thus it follows that References [1] C.F. Dunkl, D.E. Ramirez. Locally compact subgroups of the spectrum of the measure algebra II Semigroup Forum, 3:267 -- 269, 1971. [2] E. Hewitt and K.A. Ross. Abstract Harmonic Analysis I Second Ed., Grundlehern der mathematics Wissenshaften 115, Springer, New York, 1979. [3] M. Ilie and N. Spronk. The spine of a Fourier-Stieltjes algebra Proc. Lond. Math. Soc. (3), 94:273 -- 301, 2007. [4] M. Ilie and N. Spronk. The algebra generated by idempotents in a Fourier-Stieltjes algebra Houston J. Math., 33:1131 -- 1145, 2007. [5] V. Runde and N. Spronk. Operator amenability of Fourier-Stieltjes algebras. II Bull. Lond. Math. Soc., 39:194 -- 202, 2007. Monica Ilie, Department of Mathematical Sciences, Lakehead University, 955 Oliver Road, Thunder Bay, ON, P7B 5E1, Canada [email protected] Department of Pure Mathematics, University of Waterloo, [email protected] Waterloo, ON, N2L 3G1, Canada
1903.00370	1	1903	2019-03-01T15:28:50	Compact operators between lattice normed spaces	[ "math.FA" ]	In this paper we continue the study of compact-like operators in lattice normed spaces started recently by Aydin, Emelyanov, Erkur\c{s}un \"Ozcand and Marabeh. We show among others, that every p-compact operator between lattice normed spaces is p-bounded. The paper contains answers of almost all questions asked by these authors.	math.FA	math	On compact operators between lattice normed spaces Youssef Azouzi and Mohamed Amine Ben Amor Research Laboratory of Algebra, Topology, Arithmetic, and Order Department of Mathematics Faculty of Mathematical, Physical and Natural Sciences of Tunis Tunis-El Manar University, 2092-El Manar, Tunisia Abstract In this paper we continue the study of compact-like operators in lattice normed spaces started recently by Aydin, Emelyanov, Erkur¸sun Ozcand and Marabeh. We show among others, that every p-compact operator between lattice normed spaces is p-bounded. The paper con- tains answers of almost all questions asked by these authors. 1 Introduction In [3], the authors introduced a new notion of compact operators in Lattice- normed spaces and studied some of their properties. These operators act on spaces equipped with vector valued norms taking their values in some vector lattices. Recall that an operator from a normed space X to a normed space Y is said to be compact if the image of every norm bounded sequence (xn) in X has a norm convergent subsequence. This notion has been generalized in the setting of lattice normed spaces giving rise to two new notions: se- quentially p-compactness and p-compactness (p referred to the vector valued norm). Notice that these notions coincide in the classical case of Banach spaces. In general setting with vector lattice valued norms boundedness and convergence are considered with respect to these 'norms'. Also as notions of relatively uniform convergence and almost order boundedness have been 1 generalized, new properties for the operator are considered like semicompact- ness. Recall that if (E, p, V ) and (F, q, W ) are Lattice-normed spaces and T is a linear operator from E to F , then T is said to be p-compact (respec- tively, rp-compact) if for every p-bounded net (xα) in E, there is a subnet (cid:0)T xϕ(β)(cid:1) that p-converges (respectively, rp-converges) to some y ∈ F . The operator is said to be sequentially p-compact if nets and subnets are replaced by sequences and subequences above. In this paper we prove some new re- sults in this direction. Namely we show that every p-compact operator is p-bounded. As a consequence we get that every rp-compact is p-bounded. Also we give an example of sequentially p-compact operator which fails to be p-bounded. As a consequence we deduce that a sequentially p-compact need not be p-compact. In fact these two notions are totally independent. Example of p-compact operators that fail to be sequentially p-compact is given. As mentioned above the study of p-compact operators between lattice normed spaces was started in [3]. That paper contains several new results but also some open questions. Almost all these questions will be answered in our paper. 2 Preliminaries The goal of this section is to introduce some basic definitions and facts. For general informations on vector lattices, Banach spaces and lattice-normed spaces, the reader is referred to the classical monographs [1] and [6]. Consider a vector space E and a real Archimedean vector lattice V . A map p : E → V is called a vector norm if it satisfies the following axioms: 1) p(x) ≥ 0; p(x) = 0 ⇔ x = 0; (x ∈ E). 2) p(x1 + x2) ≤ p(x1) + p(x2); (x1, x2 ∈ E). 3) p(λx) = λp(x); (λ ∈ R, x ∈ E). A triple (E, p, V ) is a lattice-normed space if p(.) is a V -valued vector norm in the vector space E. When the space E is itself a vector lattice the triple (E, p, V ) is called a lattice-normed vector lattice. A set M ⊂ E is called p-bounded if p (M) ⊂ [−e, e] for some e ∈ V+. A subset M of a lattice-normed vector lattice (E, p, V ) is called p-almost order bounded if, for any w ∈ V+, there is xw ∈ E+ such that p((x − xw)+) = p(x − xw ∧ x) ≤ w for any x ∈ M. 2 p Let (xα)α∈∆ be a net in a lattice-normed space (E, p, V ). We say that (xα)α∈∆ is p-convergent to an element x ∈ E and write xα −→ x, if there exists a decreasing net (eγ)γ∈Γ in V such that inf γ∈Γ(eγ) = 0 and for every γ ∈ Γ there is an index α(γ) ∈ ∆ such that p(v − vα) ≤ eγ for all α ≥ α(γ). Notice that if V is Dedekind complete, the dominating net (eγ) may be chosen over the same index set as the original net. We say that (xα) is p-unbounded convergent to x (or for short, up-convergent to x) if xα − x ∧ u −→ 0 for all u ∈ V+. It is said to be relatively uniformly p-convergent to x ∈ X (written rp as, xα −→ x) if there is e ∈ E+ such that for any ε > 0, there is αε satisfying p(xα − x) ≤ εe for all α ≥ αε. p When E = V and p is the absolute value in E, the p-convergence is the order convergence, the up-convergence is the unbounded order convergence, and the rp-convergence is the relatively uniformly convergence. We refer to [5] and [4] for the basic facts about nets in topological spaces and vector lattices respectively. We will use [6, 8] as unique source for unexplained terminology in Lattice-Normed Spaces. Since the most part of this paper is devoted to answer several open questions in [3], the reader must have that paper handy, from which we recall some definitions. Definition 1 Let X, Y be two lattice-normed spaces and T ∈ L(X, Y ). Then 1. T is called p-compact if, for any p-bounded net (xα) in X, there is a p subnet xαβ such that T xαβ −→ y in Y for some y ∈ Y . 2. T is called sequentially p-compact if, for any p-bounded sequence xn in −→ y in Y for some X, there is a subsequence (xnk ) such that T xnk y ∈ Y . p 3. T is called p-semicompact if, for any p-bounded set A in X, the set T (A) is p-almost order bounded in Y . 3 p-compact operators are p-bounded It is well known that compact operators between Banach spaces are bounded. This result remains valid for general situation of p-compact operators as it will be shown in our first result, which answers positively Question 2 in [3]. 3 Theorem 2 Every p-compact operator between two Lattice-normed spaces is p-bounded. Proof. Assume, by contradiction, that there exists a p-compact operator T : (E, p, V ) −→ (F, q, W ) which is not p-bounded. Then there exists a p-bounded subset A of E such that T (A) is not q-bounded. So, for every u ∈ W + there exists some xu ∈ A satisfying q(T (xu)) 6≤ u. Since the net (xu)u∈W + is p-bounded there is a subnet (cid:0)yv = xϕ(v)(cid:1)v∈Γ and an element f ∈ F such that (T yv) tail, which means that for some v0 in Γ and some w ∈ W + we have, q −→ f . It follows that the net (T yv) has a q-bounded q(cid:0)T xϕ(v)(cid:1) ≤ w, for v ≥ v0. (1) Pick v1 in Γ such that ϕ(v) ≥ w for all v ≥ v1. It follows that for v ≥ v0 ∨ v1, we have q(T xϕ(v)) 6≤ ϕ(v) and so q(T xϕ(v)) 6≤ w, which is a contradiction with 1. and the proof comes to its end. The following lemma, which connects unbounded order convergence with pointwise convergence, is a known fact, although a quick proof is included for the sake of completeness. Lemma 3 Let E = RX be the Riesz space of all real-valued functions defined on a nonempty set X. The following statements are equivalent: (i) The net (fα)α∈A is uo-convergent in E. (ii) for every x ∈ X, the net (fα (x))α∈A is convergent in R. uo−→ f in the Dedekind complete Riesz Proof. (i) =⇒ (ii) Assume that fα space E. Then there is a net (gα)α∈A which decreases to 0 and for some α0 we have fα − f ∧ 1 ≤ gα for all α ≥ α0. (2) Since (gα (x)) decreases to 0 for every x ∈ X, it follows easily from 2 that fα (x) − f (x) converges to 0, as desired. (ii) =⇒ (i) Assume now that fα converges simply to some f ∈ E and let h ∈ E+. Define a net (gα) by putting gα (x) = sup β≥α (fβ − f ∧ h) (x) , x ∈ X. 4 it is clear that gα decreases to 0 and fα − f ∧ h ≤ gα. This shows that fα uo−→ f and we are done. Consider the Riesz space F of all bounded real valued functions defined on the real line with countable support and denote by E the direct sum R1 ⊕ F, where 1 denotes the constant function taking the value 1. This example will be of great interest for us. The following lemma establishes some of its properties. Recall that a vector sublattice Y of a vector lattice X is said to be regular if every subset in Y having a supremum in Y has also a supremum in X and these suprema coincide. For more information about this notion and nice characterizations of it via unbounded order convergence the reader is referred to [4]. Lemma 4 The space E introduced above has the following properties. (i) E is a regular vector sublattice of RR. (ii) E is Dedekind σ-complete but not Dedekind complete. Proof. (i) It is clear that E is a vector sublattice of RR. To show that it is regular assume that (gα)α∈A is a net in E satisfying gα ↓ 0 in E. Let g = inf gα α in RR and x ∈ R. Then h = g (x) 1{x} ∈ E and 0 ≤ h ≤ gα for all α. This implies that h = 0 and then g (x) = 0. Hence g = 0 and the regularity is proved. (ii) Let (gn) be an order bounded sequence in E and write gn = λn + fn, with λn ∈ R and fn ∈ F. Let Ω be the union of the supports of fn, then Ω is countable. Let g be the supremum of (gn) in RR, that is, g (a) = sup gn (a) , for all a ∈ R. It will be sufficient to show that g ∈ E. To this end observe that g (x) = α := sup αn for all x ∈ R\Ω. Now put f = (g − α) 1Ω. Then f ∈ F and g = α + f ∈ E as required. Next we show that E is not Dedekind complete. Consider the net (gx)x∈[0,1] in E defined by gx = x1{x}. It is a bounded net in E and its supremum in RR does not belong to E. As E is regular in RR this net can not have a supremum in E. Remark 5 Consider the following operator: T : L1 [0, 1] −→ c0; f 7−→ T f = (cid:18)Z 1 0 f (t) sin ntdt(cid:19)n≥1 . 5 It is mentioned in [1], that T is not order bounded; it is perhaps more conve- nient to consider the same operator defined on L1 [0, 2π] . In this case if we define un by u (t) = sin nt for t ∈ [0, 2π] , then un ≤ 1, however (T un) = (en) is not bounded in c0, where (en) denotes the standard basis of c0. This state- ment implies also that T is not sequentially order compact. Because (en) has no order bounded subsequence, it follows that (T un) can not admit an order convergent subsequence. So the statement made in [3] that T is p-bounded is not correct. The above example is presented in [3] to show that sequentially p-compact operators need not be p-bounded. Although the operator given in that ex- ample fails to be sequentially p-compact, the assertion that sequentially p- compact operators need not be p-bounded is true. This will be shown in our next example. Example 6 Consider the Riesz spaces E and F defined just before Lemma 4 and let T be the projection defined on E with range F and kernel R1. We claim that T is sequentially order compact, but not order bounded. Let (fn) be an order bounded sequence in E. Then fn ≤ λ for some real λ > 0 and for all n. Write fn = gn + λn with λn real and gn ∈ F and observe that gn ≤ 2λ for all n. We have also gn ≤ 2λ1A ∈ F where A is the union of the supports of gn, n = 1, 2, ... A standard diagonal process yields a subsequence (gkn) of (gn) which converges pointwise on A and then on R since all functions gnkvanish on R\A. Hence (gkn) is uo-convergent in RR. As (gkn) is order bounded this implies that (gkn) is order convergent in RR. gkn belongs to F, which shows that (gkn) is order Observe moreover that sup p≥n convergent in F. The fact that T is not order bounded is more obvious: it is clear that the image of the net (cid:0)1{x}(cid:1)x∈[0,1] by T is not order bounded in F. As an immediate consequence of Theorem 2 and Example 6 we deduce that sequentially p-compactness does not imply p-compactness. At this stage one might expect that the converse is true. Does p-compactness imply se- quentially p-compactness? This is an open question left in [3]. Unfortunately the answer is again negative. Example 7 Let X be the set of all strictly increasing maps from N to N and E = RX be the space of all real-valued functions defined on X, equipped with the product topology. 6 1. First we will prove that the identity map, I, is a p-compact operator on the lattice-normed space (E, , E). To this aim, pick a p-bounded net (fα)α∈A in E, that is, fα ≤ f for some f ∈ E+ and for every α ∈ A. It follows that fα ∈ Yx∈X [−f (x), f (x)]. topology, is compact by Tychonoff 's Theorem. Thus (fα) has a con- Notice that the space Qx∈X[−f (x), f (x)], equipped with the product vergent subnet (gβ)β∈B in Qx∈X[−f (x), f (x)] to some g. This means that gβ(x) −→ g(x) for all x ∈ X. According to Lemma 3, gβ is uo-convergent to g in E. Since bounded o−→ g. This uo-convergent nets are order convergent, we have that gβ proves that I is a p-compact operator. 2. We prove now that I is not sequentially p-compact. Let (ϕn) be a sequence in {−1, 1}X which has no convergent subsequence (see Ex- ample 3.3.22 in [7]). This sequence is order bounded in E and every subsequence (ψn) of (ϕn) does not converges in {−1, 1}X, that is, for some x ∈ X, ψn (x) diverges. According to Lemma 3, (ψn) is not uo- convergent in E. Since (ψn) is order bounded it does not converge in order. This finishes the proof. In classical theory of Banach spaces the identity map is compact if and only if the space is finite-dimensional. In contrast of this the situation is not clear in general case. We already have seen an example of infinite-dimensional space on which the identity map is p-compact. This question has been in- vestigated in [3] where the authors showed that IL1[0,1] fails to be compact however, Iℓ1 is p-compact. In the next example we show that IL∞[0,1] is not p-compact, answering a question asked in [3]. Example 8 The identity operator I on the lattice normed space (L∞[0, 1], . , L∞[0, 1]) is neither p-compact nor sequentially p-compact. To this end, consider the sequence of Rademacher function given by : rn : [0, 1] −→ R t 7−→ sgn (sin(2nπt)) 7 for all n ∈ N, which is order bounded since rn = 1. Suppose now that (rn) has o→ r for some r ∈ L∞ [0, 1] . an order convergent subnet (rnα)α∈Γ . Then rnα 0 rnαrnβ dµ = 0. On the other hand (cid:0)rnαrnβ(cid:1)β Let α ∈ A. For every β > α, R 1 converges in order to rnαr in L∞ [0, 1] and then in L1 [0, 1] . Since the integral is order continuous, we deduce that Z 1 0 rnαrdµ = 0. This equality holds for every α ∈ A, and a similar argument leads to Z 1 0 r2dµ = 0, which is a contradiction since r = 1, and the claim is now proved. 4 Semicompact operators The notion of semicompact operators has been introduced by Zaanen in [9] and extended in the framework of lattice normed spaces in [3]. Let (X, p, E) be a lattice normed space and (Y, q, F ) be an lattice normed vector lattice. A linear operator T : X → Y is called p-semicompact if it maps p-bounded sets in X to p-almost order bounded sets in Y . We recall that a subset B of Y is said to be p-almost order bounded if for any w ∈ F+, there is yw ∈ Y such that q((y − yw)+) = q(y − yw ∧ y) ≤ w for all y ∈ B. Semicompact operators from Banach spaces to Banach lattices fail, in gen- eral, to be compact (see [1]). This yields trivially that p-semicompactness does not imply p-compactness. However, the converse is true in the classical case as has been shown in Theorem 5.71 in [1]. And one can expect to extend this result in general situation. This is already the subject of Question 4 in [3]. Unfortunately the answer is again negative. Before stating our coun- terexample let us recall that every order bounded operator from a vector lattice E to a Dedekind complete vector lattice F has a modulus ([1]). Example 9 Let E be a Dedekind complete Banach lattice with order contin- uous norm and T be a norm-compact operator in L(E) such that T has no 8 modulus, and therefore T can not be order bounded. For the existence of such operator we refer the reader to the Krengel's example in [1, p 277.]. Consider now the following lattice-normed vector spaces (E, k.k, R) and (E, p, R2) , where p(x) = (cid:18)kxk 0 (cid:19) for all x ∈ E. It is straightforward to prove that T is again p-compact operator and we claim that T is not p-semicompact. To this end we we will argue by contradiction and we assume that T is p- semicompact. Fix an element u ∈ E+ and let w = (cid:18)0 1(cid:19), then there exists zw such that p ((T (x) − zw)+) ≤ w for all x ∈ [−u, u], which means that (T (x) − zw)+ = 0. Noting that this occurs for x and −x we see that T (x) ≤ zw for all x ∈ [−u, u] . This shows that T is order bounded, a contradiction. and our proof comes to an end. A slight modification of the proof of Example 9 leads to a more general result. The proof of it will be left for the reader. Proposition 10 Let (E, p, V ) be a lattice normed space and (F, q, W ) a lat- tice normed vector lattice. We assume that q (F )d is not trivial. Then every semicompact operator T : (E, p, V ) −→ (F, q, W ) is p-bounded as an operator from (E, p, V ) to (F, . , F ) . 5 rp-compact operators As every rp-compact operator between lattice-normed spaces is p-compact, the following result is an immediate consequence of Theorem 2. Theorem 11 Let (E, p, V ) and (F, q, W ) be lattice-normed spaces and T be in L(E, F ). If T is rp-compact then T is p-bounded. In the following example we will prove that sequentially p-compact oper- ators need not be rp-compact. Example 12 Let E be the Riesz space defined above. We claim that the identity operator I : E → E is sequentially p-compact but fails to be rp- compact. Let (xn) be a bounded sequence in E, that is, xn ≤ x for some 9 x ∈ E+. Write x = α + f, and xn = αn + fn where α ∈ R+ and f ∈ F and αn ∈ R, fn ∈ F for n = 1, 2, ... It is easily seen that αn ≤ α, fn ≤ x+α. By a standard diagonal argument there exists a subsequence such that fϕ(n) (a) converges for every a ∈ R and αϕ(n) converges in R.T his shows that xn uo−→ x converges pointwise on R and its limit is clearly in E. By Lemma 3, xn o−→ x in RR as it is an order bounded sequence. Now using in RR and then xn Lemma 27 in [2] and Lemma 4 we deduce that x ∈ E. On the other hand, let F be the collection of finite subsets of R+ ordered by inclusion and consider the net (gA)A∈F where gα = 1α. Then (gα) is order bounded in E but has no convergent subnet. Since gα ↑ 1R+ in RR and E is regular in RR, it follows that (gα) is not order convergent in E and so are all its subnets. References [1] C.D. Aliprantis and O. Burkinshaw, Positive Operators, Springer, 2006. [2] Y. Azouzi, Completeness for vector lattices, J. Math. Anal. Appl. 472 (2019) 216 -- 230 [3] A. Aydın, E.Yu. Emelyanov, N. Erkur¸sun Ozcand, M.A.A. Marabeh, Indag. Math. (N.S.) 29 (2018) 633 -- 656. [4] N. Gao, V.G. Troitsky, F. Xanthos, Uo-convergence and its applications to Ces´aro means in Banach lattices, Isr. J. Math. 220 (2) (2017) 649 -- 689. [5] J. L. Kelley, General topology. Courier Dover Publications, 2017. [6] A.G. Kusraev, Dominated Operators, in: Mathematics and its Applica- tions, vol. 519, Kluwer Academic Publishers, Dordrecht, 2000. [7] V. Runde, A Taste of Topology. Springer, Berlin (2005) [8] B.Z. Vulikh, Introduction To the Theory of Partially Ordered Spaces, Wolters-Noordhoff Scientifc Publications, Ltd., Groningen, 1967. [9] A. C. Zaanen, Riesz spaces II, North-Holland, Amsterdam, 1983. 10
1712.08450	1	1712	2017-12-22T14:04:24	On the weighted fractional Poincare-type inequalities	[ "math.FA" ]	Weighted fractional Poincar\'e-type inequalities are proved on John domains whenever the weights defined on the domain are depending on the distance to the boundary and to an arbitrary compact set in the boundary of the domain.	math.FA	math	ON THE WEIGHTED FRACTIONAL POINCAR´E-TYPE INEQUALITIES RITVA HURRI-SYRJ ANEN AND FERNANDO L ´OPEZ-GARC´IA Abstract. Weighted fractional Poincar´e-type inequalities are proved on John do- mains whenever the weights defined on the domain depend on the distance to the boundary and to an arbitrary compact set in the boundary of the domain. 1. Introduction In this article we study a version of the classical fractional Poincar´e-type inequality where the domain in the double integral in the Gagliardo seminorm is replaced by a smaller one:(cid:18)(cid:90) (cid:19)1/p ≤ C (cid:18)(cid:90) (cid:90) u(x) − uΩpdx Ω Ω B(x,τ d(x)) (cid:19)1/p u(x) − u(y)p x − yn+sp dydx . (1.1) The parameter τ in the double integral belongs to (0, 1) and d(x) denotes the distance from x to ∂Ω. The inequality (1.1) was introduced in [4]. It is well-known that the fractional classical Poincar´e inequality is valid for any bounded domain, while this new version (1.1) depends on the geometry of the domain. In [4] it was proved that the inequality (1.1) is valid on John domains and, hence, in particular on Lipschitz domains. An example of a domain where the inequality (1.1) is not valid was also given. We refer the reader to [5] and [2] where the fractional Sobolev-Poincar´e versions of (1.1) are considered. For a weighted version of (1.1) where weights are power functions to the boundary we refer to [3]. The main result of our paper is the following theorem where the distance to an arbitrary set of the boundary has been added as a weight. Theorem 1.1. Let Ω in Rn be a bounded John domain and 1 < p < ∞. Given a compact set F in ∂Ω, and the parameters β ≥ 0 and s, τ ∈ (0, 1), there exists a constant C such that (cid:18)(cid:90) Ω u(x) − uΩ,ωpdpβ F (x)dx (cid:19)1/p (cid:18)(cid:90) (cid:90) (cid:19)1/p (1.2) for all functions u ∈ Lp(Ω, d(x)pβ), where d(x) and dF (x) denote the distance from x to ∂Ω and F respectively, and uΩ,ω is the weighted average u(x) − u(y)p x − yn+sp dps(x)dpβ (cid:82) F (x)dydx ≤ C Ω u(z)dpβ F (z)dz. Ω B(x,τ d(x)) 1 dpβ F (Ω) Date: May 24, 2021. 2010 Mathematics Subject Classification. Primary: 46E35 ; Secondary: 26D10. Key words and phrases. Fractional Poincar´e inequalities, Hardy-type operator, Tree covering, Weights. 1 2 In addition, the constant C in (1.2) can be written as C = Cn,p,β τ s−nK n+β, where K is the geometric constant introduced in (5.1). We would like to emphasize two points in this result: The first one is that no extra conditions are required for the compact set F in ∂Ω. The second point is that the estimate shows how the constant depends on the given τ and a certain geometric condition of the domain. Some of the essential auxiliary parts for the proofs for weighted inequalities are from [7] and [8] where a useful decomposition technique was introduced by the second author. Our work was stimulated by the papers of Augusto C. Ponce, [10], [11], [12], where more general fractional Poincar´e inequalities for functions defined on Lipschitz domains were investigated. The paper is organized as follows: In Section 2, we introduce some definitions and preliminary results. In Section 3, we show how to use decompositions of functions to extend the validity of certain inequalities on "simple domains", such as cubes, to more complex ones. We are interested in extending the results from cubes to John domains. In Section 4, we apply the results obtained in the previous section to estimate the constant in the unweighted version of (1.2) on cubes. Especially we are interested in how the constant depends on τ . This result is auxiliary of our main theorem but it might be of independent interest. In Section 5, we show the validity of the weighted fractional Poincare inequality studied in this paper with the estimate of the constant and a generalization to the type of inequalities considered by Ponce. 2. Notation and preliminary results Throughout the paper Ω in Rn is a bounded domain with n ≥ 2, 1 < p < ∞, and 1 < q < ∞ with 1 q = 1, unless otherwise stated. Moreover, given η : Ω → R a weight (i.e., a positive measurable function) and 1 ≤ r ≤ ∞, we denote by Lr(Ω, η) the space of Lebesgue measurable functions u : Ω → R equipped with the norm p + 1 (cid:18)(cid:90) (cid:19)1/r (cid:107)u(cid:107)Lr(Ω,η) := u(x)rη(x) dx if 1 ≤ r < ∞, and Ω (cid:107)u(cid:107)L∞(Ω,η) := ess sup x∈Ω u(x)η(x). Finally, given a set A we denote by χA(x) its characteristic function. Definition 2.1. Let C be the space of constant functions from Rn to R and {Ut}t∈Γ a collection of open subsets of Ω that covers Ω except for a set of Lebesgue measure zero; Γ is an index set. It also satisfies the additional requirement that for each t ∈ Γ the set Ut intersects a finite number of Us with s ∈ Γ. This collection {Ut}t∈Γ is called we say that a collection of functions {gt}t∈Γ in L1(Ω) is a C-orthogonal decomposition of g subordinate to {Ut}t∈Γ if the following three properties are satisfied: an open covering of Ω. Given g ∈ L1(Ω) orthogonal to C (i.e.,(cid:82) g ϕ = 0 for all ϕ ∈ C), (1) g =(cid:80) (3) (cid:82) gt ϕ = 0, for all ϕ ∈ C and t ∈ Γ. t∈Γ gt. (2) supp(gt) ⊂ Ut. Ut We also refer to this collection of functions by a C-decomposition. We say that {gt}t∈Γ is a finite C-decomposition if gt (cid:54)≡ 0 only for a finite number of t ∈ Γ. distance to the space of constant functions C by replacing its left hand side by Inequality (1.2), and similar Poincar´e type inequalities, can be written in terms of a 3 (cid:18)(cid:90) inf α∈C Ω (cid:19)1/p . u(x) − αpdpβ F (x)dx The technique used in this paper may also be considered when the distance to other vector spaces V are involved, in which case, a V-orthogonal decomposition of func- tions is required. We direct the reader to [9] where a generalized version of the Korn inequality is studied by using decomposition of functions. Let us denote by G = (V, E) a graph with vertices V and edges E. Graphs in this paper have neither multiple edges nor loops and the number of vertices in V is at most countable. A rooted tree (or simply a tree) is a connected graph G in which any two vertices are connected by exactly one simple path, and a root is simply a distinguished vertex a ∈ V . Moreover, if G = (V, E) is a rooted tree with a root a, it is possible to define a partial order "(cid:22)" in V as follows: s (cid:22) t if and only if the unique path connecting t with the root a passes through s. The height or level of any t ∈ V is the number of vertices in {s ∈ V : s (cid:22) t with s (cid:54)= t}. The parent of a vertex t ∈ V is the vertex s satisfying that s (cid:22) t and its height is one unit smaller than the height of t. We denote the parent of t by tp. It can be seen that each t ∈ V different from the root has a unique parent, but several elements in V could have the same parent. Note that two vertices are connected by an edge (adjacent vertices) if one is the parent of the other. Definition 2.2. Let Ω be in Rn be a bounded domain. We say that an open covering {Ut}t∈Γ is a tree covering of Ω if it also satisfies the properties: t∈Γ χUt(x) ≤ N χΩ(x), for almost every x ∈ Ω, where N ≥ 1. (1) χΩ(x) ≤(cid:80) (2) Γ is the set of vertices of a rooted tree (Γ, E) with a root a. (3) There is a collection {Bt}t(cid:54)=a of pairwise disjoint open cubes with Bt ⊆ Ut∩ Utp. Definition 2.3. Given a tree covering {Ut}t∈Γ of Ω we define the following Hardy-type operator T on L1-functions: g, (2.1) (2.2) T g(x) := (cid:88) a(cid:54)=t∈Γ (cid:90) Wt χt(x) Wt (cid:91) s(cid:23)t Us , where Wt := and χt is the characteristic function of Bt for all t (cid:54)= a. We may refer to Wt by the shadow of Ut. Note that the definition of T is based on the a-priori choice of a tree covering {Ut}t∈Γ of Ω. Thus, whenever T is mentioned in this paper there is a tree covering {Ut}t∈Γ of Ω explicitly or implicitly associated to it. The following fundamental result was proved in [8, Theorem 4.4], which shows the existence of a C−decomposition of functions subordinate to a tree covering of the domain. g ∈ L1(Ω) such that(cid:82) Theorem 2.4. Let Ω in Rn be a bounded domain with a tree covering {Ut}t∈Γ. Given Ω gϕ = 0, for all ϕ ∈ C, and supp(g)∩ Us (cid:54)= ∅ for a finite number of s ∈ Γ, there exists a C-decompositions {gt}t∈Γ of g subordinate to {Ut}t∈Γ (refer to Definition 2.1). Moreover, let t ∈ Γ. If x ∈ Bs where s = t or sp = t then 4 gt(x) ≤ g(x) + Ws Bs T g(x), (2.3) where Wt denotes the shadow of Ut defined in (2.2). Otherwise (2.4) Remark 2.5. The C-decomposition stated in Theorem 2.4 is finite. This fact is not in the statement of [8, Theorem 4.4] but it is easily deduced from its proof. gt(x) ≤ g(x). In the next lemma, the continuity of the operator T is shown. We refer the reader to [7, Lemma 3.1] for its proof. Lemma 2.6. The operator T : Lq(Ω) → Lq(Ω) defined in (2.1) is continuous for any 1 < q ≤ ∞. Moreover, its norm is bounded by (cid:19)1/q (cid:18) qN q − 1 . (cid:107)T(cid:107)Lq→Lq ≤ 2 Here N is the overlapping constant from Definition 2.2. If q = ∞, the previous inequality means (cid:107)T(cid:107)L∞→L∞ ≤ 2. Actually, for being T an averaging operator, it can be easily observed that (cid:107)T(cid:107)L∞→L∞ = 1, but it does not affect our work. Notice that Lq(Ω, ω−q) ⊂ L1(Ω) if the weight ω : Ω → R>0 satisfies that ωp ∈ L1(Ω). Then, the operator T introduced in Definition 2.3 for functions in L1(Ω) is well-defined in Lq(Ω, ω−q). Lemma 2.7. Let Ω in Rn be a bounded domain, {Ut}t∈Γ a tree covering of Ω and ω : Ω → R a weight which satisfies ωp ∈ L1(Ω). If ω satisfies that (2.5) for all a (cid:54)= t ∈ Γ, then the Hardy-type operator T defined in (2.1) and subordinate to {Ut}t∈Γ is continuous from Lq(Ω, ω−q) to itself. Moreover, its norm for 1 < q < ∞ is bounded by ess sup y∈Wt ω(x), ω(y) ≤ C2 ess inf x∈Bt (cid:19)1/q (cid:18) qN q − 1 C2, (cid:107)T(cid:107)L→L ≤ 2 where L denotes Lq(Ω, ω−q), and N is the overlapping constant from Definition 2.2. Proof. Given g ∈ Lq(Ω, ω−q) we have (cid:90) (cid:90) (cid:90) Ω Ω Ω = = (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:88) T g(x)qω−q(x) dx ω−q(x) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:88) a(cid:54)=t∈Γ a(cid:54)=t∈Γ χt(x) Wt Wt ω−1(x) χt(x) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)q dx g(y) dy g(y)ω−1(y) ω(y) dy (cid:90) (cid:90) Wt Wt (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)q dx. Now, condition (2.5) implies that ω(y) ≤ C2ω(x) for almost every x ∈ Bt and y ∈ Wt. 5 Thus, ≤ (cid:90) (cid:90) Ω Ω = C q 2 = C q 2 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)q dx g(y)ω−1(y) dy (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)q g(y)ω−1(y) dy dx χt(x) a(cid:54)=t∈Γ T g(x)qω−q(x) dx (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:88) Wt ω−1(x) C2 ω(x) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:88) (cid:90) (cid:90) (cid:90) (cid:12)(cid:12)T (gω−1)(cid:12)(cid:12)q dx. (cid:18) χt(x) Wt a(cid:54)=t∈Γ Wt Ω Ω (cid:90) Wt (cid:19) Finally, gω−1 belongs to Lq(Ω) and T is continuous from Lq(Ω) to itself; we refer to Lemma 2.6, hence(cid:90) T g(x)qω−q(x) dx ≤ 2q qN q − 1 Ω 2 (cid:107)g(cid:107)q C q Lq(Ω,ω−q). (cid:3) (cid:19)1/p 3. A decomposition and Fractional Poincar´e inequalities (cid:82) Let Ω in Rn be an arbitrary bounded domain and {Ut}t∈Γ an open covering of Ω. The weight ω : Ω → R>0 satisfies that ωp ∈ L1(Ω). In addition, uΩ denotes the average 1Ω Ω u(z)dz. For weighted spaces of functions, uΩ,ω represents the weighted average Now, given a bounded domain U in Rn and a nonnegative measurable function µ : U × U → R we introduce the following Poincar´e type inequality 1 ω(Ω) (cid:82) Ω u(z)ω(z)dz, where ω(Ω) :=(cid:82) (cid:18)(cid:90) c∈R(cid:107)u − c(cid:107)Lp(U,ωp) ≤ C inf (cid:90) Ω ω(z)dz. u(x) − u(y)pµ(x, y) dydx (3.6) where u ∈ Lp(U, ωp). Notice that the right hand side in this inequality might be infinite. The validity of (3.6) depends on U , p, µ and ω. The function µ(x, y) might be zero, however, ω(x) is strictly positive almost everywhere in Ω. U U , Let us mention three examples. Examples 3.1. (1) The weighted fractional Poincar´e inequality with µ(x, y) = x−yn+sp , where s ∈ (0, 1) , is the classical fractional Poincar´e inequality which 1 is clearly valid for any arbitrary bounded domain. (2) If µ(x, y) = χBx (y) x−yn+sp , where Bx is the ball centered at x with radius τ d(x) for s, τ ∈ (0, 1), then the inequality represents a more recently studied fractional Poincar´e inequality whose validity depends on the geometry of the domain (re- fer to [4] for details). (3) Finally, µ(x, y) = ρ(x−y) x−yp , where ρ is a certain nonnegative radial function, is another inequality which has also been studied recently (refer to [10] for details). 6 Inequality (3.6) deals with an estimation of the distance to C of an arbitrary function u in Lp(Ω, ωp). The local-to-global argument used in this paper to study this Poincar´e type inequalities is based on the fact that Lp(Ω, ωp) is the dual space of Lq(Ω, ω−q) and the existence of decompositions of functions in Lq(Ω, ω−q) orthogonal to C. Let us properly define this set and a subspace: W := {g ∈ Lq(Ω, ω−q) : W0 := {g ∈ W : supp(g) intersects a finite number of Ut}. (3.8) The integrability of ωp implies that Lq(Ω, ω−q) ⊂ L1(Ω), then W and W0 are well- defined. Following Remark 2.5, the C-decomposition of functions in W0 stated in Theorem 2.4 is finite, which is not valid in general for functions in W. This property verified by the functions in W0 simplifies the proof of Lemma 3.3, which motivates the definition of this space. gϕ = 0 for all ϕ ∈ C} (cid:90) (3.7) Now, we introduce the spaces W ⊕ ωpC = {g + αωp / g ∈ W and α ∈ C} S := W0 ⊕ ωpC = {g + αωp / g ∈ W0 and α ∈ C}. (3.9) It is not difficult to observe that Lq(Ω, ω−q) = W ⊕ ωpC and S is a subspace of Lq(Ω, ω−q). The following lemma, which was proved in [8, Lemma 3.1], states that S is also dense in Lq(Ω, ω−q) and uses in its proof the requirement that says that for each t ∈ Γ the set Ut intersects a finite number of Us with s ∈ Γ. Lemma 3.2. The space S is dense in Lq(Ω, ω−q). Moreover, if g + αωp is an element in S then (cid:107)g(cid:107)Lq(Ω,ω−q) ≤ 2(cid:107)g + αωp(cid:107)Lq(Ω,ω−q). Lemma 3.3. If there exists an open covering {Ut}t∈Γ of Ω such that (3.6) is valid on Ut for all t ∈ Γ, with a uniform constant C1, and there exists a finite C-orthogonal decomposition of any function g in W0 subordinate to {Ut}t∈Γ, with the estimate 0(cid:107)g(cid:107)q Lq(Ω,ω−q), (cid:88) t∈Γ (cid:107)gt(cid:107)q (cid:32)(cid:88) Lq(Ut,ω−q) ≤ C q (cid:90) (cid:90) t∈Γ Ut Ut then, there exists a constant C such that (cid:107)u − uΩ,ω(cid:107)Lp(Ω,ωp) ≤ C u(x) − u(y)pµ(x, y) dydx (cid:33)1/p (3.10) is valid for any u ∈ Lp(Ω, ωp). Moreover, the constant C = 2C0C1 holds in (3.10). Proof. Without loss of generality we can assume that uΩ,ω = 0. We estimate the norm on the left hand side of the inequality by duality. Thus, let g + ωpψ be an arbitrary function in S, we refer to Lemma 3.2. Then, by using the finite C-orthogonal decomposition of g we conclude that (cid:90) Ω u(g + αωp) = = ug = u Ω ugt = gt (cid:90) (u − ct)gt. (3.11) (cid:90) (cid:90) (cid:88) Ω t∈Γ (cid:90) (cid:88) (cid:88) t∈Γ t∈Γ Notice that the identity in the second line is valid for any t ∈ Γ and ct ∈ R. Ut Ut Next, by using the Holder inequality in (3.11), the fact that (3.6) is valid on Ut with a uniform constant C1 and, finally, the Holder inequality over the sum, we obtain 7 (cid:90) Ω ≤ C1 ≤ C1 Ut Ut t∈Γ u(g + αωp) ≤(cid:88) (cid:18)(cid:90) (cid:90) (cid:88) (cid:32)(cid:88) (cid:90) (cid:90) (cid:32)(cid:88) (cid:90) (cid:32)(cid:88) (cid:90) (cid:90) (cid:90) t∈Γ t∈Γ Ut Ut Ut ≤ C0C1 ≤ 2C0C1 t∈Γ Ut Ut inf c∈R(cid:107)u − c(cid:107)Lp(Ut,ωp)(cid:107)gt(cid:107)Lq(Ut,ω−q) (cid:19)1/p (cid:107)gt(cid:107)Lq(Ut,ω−q) t∈Γ u(x) − u(y)pµ(x, y) dydx (cid:33)1/p(cid:32)(cid:88) (cid:33)1/p (cid:33)1/p u(x) − u(y)pµ(x, y) dydx u(x) − u(y)pµ(x, y) dydx t∈Γ (cid:107)g(cid:107)Lq(U,ω−q) (cid:107)gt(cid:107)q Ut (cid:33)1/q Lq(Ut,ω−q) u(x) − u(y)pµ(x, y) dydx (cid:107)g + αωp(cid:107)Lq(U,ω−q). Finally, as S is dense in Lq(Ω, ω−q), by taking the supremum over all the functions g + αωp in S with (cid:107)g + αωp(cid:107)Lq(Ω,ω−q) ≤ 1 we prove the result. (cid:3) 4. On fractional Poincar´e inequalities on cubes In this section, we use the results stated in the previous two sections to show a certain fractional Poincar´e inequality on an arbitrary cube Q. Thus, in order to show the existence of the C-decomposition, which is used later to apply Lemma 3.3, we define a tree covering {Ut}t∈Γ of Q. This covering is only used in this section and for cubes. In the following section, we work with a different bounded domain, an arbitrary bounded John domain, which requires a different covering. However, let us warn the reader that we will keep the notation {Ut}t∈Γ used in Section 3. The validity of the local inequality stated in the following proposition is well-known. We refer the reader to [3] for its proof. Proposition 4.1. The fractional Poincar´e inequality c∈R(cid:107)u(x) − c(cid:107)Lp(U ) ≤ inf (cid:18)diam (U )n+sp (cid:90) (cid:90) U u(y) − u(x)p y − xn+sp dydx U U (cid:19)1/p holds for any bounded domain U in Rn and 1 ≤ p < ∞. The following proposition is a special case of [4, Lemma 2.2]. In the present paper, we give a different proof which let us estimate the dependance of the constant with respect to τ . Proposition 4.2. Let Q in Rn be a cube with side length l(Q) = L, 1 < p < ∞ and τ ∈ (0, 1). Then, the following inequality holds (cid:18)(cid:90) (cid:90) u(y) − u(x)p y − xn+sp dydx Q∩B(x,τ L) Q (cid:19)1/p , c∈R(cid:107)u(x) − c(cid:107)Lp(Q) ≤ Cn,p τ s−nLs inf where Cn,p depends only on n and p. Proof. This result follows from Lemma 3.3 on the cube Q, where µ(x, y) = x−yn+sp and ω ≡ 1. So, let us start by defining an appropriate tree covering of Q to obtain, via Theorem 2.4 and Remark 2.5, a finite C-decomposition of any functions in W0. Let and {At}t∈Γ the regular partition of Q with m ∈ N be such that mn open cubes. The side length of each cube is l(At) = L m. In the example shown in Figure 1, m = 4 and the index set Γ has 16 elements. τ < m ≤ 1 + n+3 τ √ √ n+3 1 8 Figure 1. A tree covering of Q The tree covering of Q that we are looking for will be defined by enlarging the sets in the covering {At}t∈Γ in an appropriate way but keeping the tree structure of Γ, which is introduced in the following lines. Indeed, we pick a cube Aa, whose index will be the root, and inductively define a tree structure in Γ such that the unique chain connecting t with a is associated to a chain of cubes connecting Qt with Qa, with minimal number of cubes, such that two consecutive cubes share a n − 1 dimensional face. In Figure 1, the cube Aa is in the lower left corner and the tree structure is represented using black arrows that"descend" to the root. Now that Γ has a tree structure, we define the tree covering {Ut}t∈Γ of Q with the rectangles Ut := (At ∪ Atp)◦ if t (cid:54)= a and Ua := Aa. In order to have a better understanding of the construction, notice that Ut ∩ Utp = Atp for all t (cid:54)= a. Moreover, the index set Γ in the example with its tree structure has 7 levels, from level 0 to level 6 (refer to page 3 for definitions), with only one index of level 6, whose rectangle Ut appears in Figure 1 in a different color. Now, let us define the collection {Bt}t(cid:54)=a of pairwise disjoint open cubes Bt ⊆ Ut∩Utp or equivalently Bt ⊆ Atp. Given t (cid:54)= a, we split Atp into 3n cubes with the same size. The open set Bt is the cube in the regular partition of Atp whose closure intersects the n − 1 dimensional face Atp in the intersection (At ∩ Atp). There are 3n−1 cubes with that property but we pick Bt to be the one which does not share any part of any other n − 1 dimensional face of Atp. The cubes in {Bt}t(cid:54)=a have side length equal to L 3m and are represented in Figure 1 by the 15 grey gradient small cubes. By its construction, it is easy to check that {Bt}t(cid:54)=a is a collection of pairwise disjoint open cubes Bt ⊆ Ut ∩ Utp, hence, {Ut}t∈Γ is a tree covering of Q with N = 2n (it could also be less). By Theorem 2.4, there is a finite C-decomposition of functions {gt}t∈Γ subordinate to {Ut}t∈Γ which satisfies (2.3) and (2.4). Moreover, it can be seen that Bs ≤ Q Ws Bs = (3m)n, for all s ∈ Γ, thus, for all t ∈ Γ and x ∈ Ut. Next, using the continuity of T stated in Lemma 2.6 and some straightforward calculations we conclude gt(x) ≤ g(x) + (3m)nT g(x), 9 (cid:18) (cid:19) (cid:107)gt(cid:107)q Lq(Ut) ≤ 2q−1N 1 + (3m)nq2q qN q − 1 (cid:107)g(cid:107)q Lq(Q) (cid:88) t∈Γ Hence, we have a finite C-decomposition of any function in W0 subordinate to {Ut}t∈Γ with the constant in the estimate equal to ≤ 22q+2n2q q − 1 ≤ 22q+23nqn2q q − 1 (cid:18)22q+23nqn2q q − 1 (3m)nq (cid:107)g(cid:107)q √ 1 + (cid:16) (cid:19)1/q (cid:16) 1 + √ C0 = (cid:17)nq τ−nq (cid:107)g(cid:107)q Lq(Q). Lq(Q) n + 3 (cid:17)n √ n + 3 τ−n. n+3 Now, from Proposition 4.1 and using that m > τ (3.6) is valid on each Ut with an uniform constant , we can conclude that inequality Thus, using Lemma 3.3 we can claim that (cid:107)u − uQ(cid:107)Lp(Q) ≤ 2C0C1 (cid:33)1/p . u(x) − u(y)p x − yn+sp dydx Finally, diam(Ut) ≤ √ the control on the overlapping of the tree covering given by N = 2n, it follows that n + 3 L m ≤ τ L, thus Ut ⊂ B(x, τ L) for any x ∈ Ut, thus, using C1 = (n + 3)n/2p(τ L)s. (cid:90) (cid:90) Ut Ut (cid:90) t∈Γ (cid:32)(cid:88) (cid:18)(cid:90) (cid:19)1/q (cid:16) Q∩B(x,τ L) Q √ n + 3 1 + (cid:17)n (cid:19)1/p u(x) − u(y)p x − yn+sp dydx (n + 3)n/2p(2n)1/p. , (4.1) (cid:3) (cid:107)u − uQ(cid:107)Lp(Q) ≤ Cn,p τ−n(τ L)s where Cn,p = 2 (cid:18)22q+23nqn2q q − 1 5. On fractional Poincar´e inequalities on John domains In this section, we apply the results obtained in the previous sections on an arbitrary bounded John domain Ω. Its definition is recalled below. The weight ω(x) is defined as dF (x)β, where dF (x) denotes the distance from x to an arbitrary compact set F in ∂Ω and β ≥ 0. In the particular case where F = ∂Ω, d∂Ω(x) is simply denoted as d(x). Notice that ωp belongs to L1(Ω) for being Ω bounded and β nonnegative. A Whitney decomposition of Ω is a collection {Qt}t∈Γ of closed pairwise disjoint dyadic cubes, which verifies (1) Ω =(cid:83) t∈Γ Qt. (2) diam(Qt) ≤ dist(Qt, ∂Ω) ≤ 4diam(Qt). (3) 1 4diam(Qs) ≤ diam(Qt) ≤ 4diam(Qs), if Qs ∩ Qt (cid:54)= ∅. 10 Here, dist(Qt, ∂Ω) is the Euclidean distance between Qt and the boundary of Ω, denoted by ∂Ω. The diameter of the cube Qt is denoted by diam(Qt) and the side length is written as (cid:96)(Qt). Two different cubes Qs and Qt with Qs ∩ Qt (cid:54)= ∅ are called neighbors. This kind of covering exists for any proper open set in Rn (refer to [13, VI 1] for details). Moreover, each cube Qt has less than or equal to 12n neighbors. And, if we fix 0 < < 1 4 and define (1 + )Qt as the cube with the same center as Qt and side length (1 + ) times the side length of Qt, then (1 + )Qt touches (1 + )Qs if and only if Qt and Qs are neighbors. Given a Whitney decomposition {Qt}t∈Γ of Ω we refer by an expanded Whitney decomposition of Ω to the collection of open cubes {Q∗ t}t∈Γ defined by Observe that this collection of cubes satisfies that Q∗ t := Q◦ t . 9 8 χΩ(x) ≤ 12n(cid:88) t∈Γ (x) ≤ (12n)2χΩ(x) χQ∗ t for all x ∈ Rn. We recall the definition of a bounded John domain. A bounded domain Ω in Rn is a John domain with constants a and b, 0 < a ≤ b < ∞, if there is a point x0 in Ω such that for each point x in Ω there exists a rectifiable curve γx in Ω, parametrized by its arc length written as length(γx), such that dist(γx(t), ∂Ω) ≥ a length(γx) for all t ∈ [0, length(γx)] t and length(γx) ≤ b. Examples of John domains are convex domains, uniform domains, and also domains with slits, for example B2(0, 1)\[0, 1). The John property fails in domains with zero angle outward spikes. John domains were introduced by Fritz John in [6] and they were renamed by O. Martio and J. Sarvas as John domains later. There are other equivalent definitions of John domains. In these notes, we are interested in a definition of the style of Boman chain condition (see [1]) in terms of Whitney decompositions and trees. This equivalent definition is introduced in [8]. Definition 5.1. A bounded domain Ω in Rn is a John domain if for any Whitney decomposition {Qt}t∈Γ, there exists a constant K > 1 and a tree structure of Γ, with a root a, that satisfies Qs ⊆ KQt, (5.1) for any s, t ∈ Γ with s (cid:23) t. In other words, the shadow of Qt written as Wt is contained in KQt; refer to (2.2). Moreover, the intersection of the cubes associated to adjacent indices, Qt and Qtp, is an n − 1 dimensional face of one of these cubes. Now, given a Whitney decomposition {Qt}t∈Γ of a bounded John domain Ω in Rn, with constant K in the sense of (5.1), we define the tree covering {Ut}t∈Γ of expanded Whitney cubes such that Ut := Q∗ t . (5.2) 11 The overlapping is bounded by N = 12n. Now, each open cube Bt in the collection {Bt}t(cid:54)=a shares the center with the n− 1 dimensional face Qt ∩ Qtp and has side length lt 64, where lt is the side length of Qt. It follows from the third condition in the Whitney decomposition, and some calculations, that this collection is pairwise disjoint and Moreover, it can be seen that Bt ⊂ Q∗ t ∩ Q∗ tp = Ut ∩ Utp. Wt Bt ≤ (K 9 8lt)n ( lt 64)n = 72n K n, (5.3) for all t ∈ Γ, with t (cid:54)= a. Lemma 5.2. Let Ω in Rn be a John domain with the constant K in the sense of (5.1), F in ∂Ω a compact set and dF (x) the distance from x to F . Then, dF (y) ≤ 3K sup y∈Wt for all t ∈ Γ. √ n inf x∈Bt dF (x), A similar inequality is also valid if we consider the weight dβ F (x) with a nonnegative power of the distance to F . Thus, this lemma implies, via Lemma 2.7, the continuity −qβ of the operator T from Lq(Ω, d ) to itself with an estimation of its constant. Then, there exists a C-decomposition with a weighted estimate for a certain weight. F Proof. Given t ∈ Γ, with t (cid:54)= a, x ∈ Bt and y ∈ Wt := ∪s(cid:23)tUs, we have to prove that dF (y) ≤ 3KdF (x). Notice that d(x) ≤ dF (x) for all x ∈ Ω. Moreover, Qs ⊆ KQt for all s (cid:23) t, then Wt ⊆ KUt. In addition, dF (y) ≤ y − x + dF (x) ≤ diam(Wt) + dF (x) ≤ Kdiam(Ut) + dF (x) = K 9 8 diam(Qt) + dF (x). Finally, using the second property stated in the Whitney decomposition it follows that 3Qt ⊂ Ω. Then, as dist(Q∗ t , ∂Ω) ≥ dist(Q∗ t , (3Qt)c) ≥ 15 16 lt, doing some calculations we can assert that n dist(Q∗ diam(Qt) ≤ 16 √ 15 t , ∂Ω) ≤ 16 9 √ n dist(Q∗ t , ∂Ω). Thus, dF (y) ≤ 2K ≤ 2K √ √ n dist(Q∗ t , ∂Ω) + dF (x) √ n d(x) + dF (x) ≤ 2K n dF (x) + dF (x). (cid:3) Now we are able to prove Theorem 1.1 and also to give the dependence of the constant C on the given value of τ and the constant K from (5.1). Proof of Theorem 1.1. This result follows from Lemma 3.3 with the tree covering {Ut}t∈Γ of Ω defined in (5.2), ω(x) := dβ F (x) and 12 µ(x, y) := dps(x) dpβ F (x) χB(x,τ d(x))(y) x − yn+sp . (5.4) Notice that ωp belongs to L1(Ω), the condition assumed at the beginning of Section 3. The validity of (3.6) on a cube Ut, with a uniform constant C1, follows from Proposition 4.2. Indeed, by using the fact that Ut is an expanded Whitney cube by a factor 9/8 and F ⊆ ∂Ω, it follows that F (x) ≤ 2β inf dβ x∈Ut sup x∈Ut dβ F (x). Thus, we have c∈R(cid:107)u(x) − c(cid:107)Lp(Ut,dpβ inf F ) ≤Cn,p τ s−nLs t 2β (cid:18)(cid:90) (cid:90) Ut Ut u(x) − u(y)p x − yn+sp dpβ F (x)χB(x,τ Lt)(y) dydx (cid:19)1/p , where Lt is the side length of Ut and Cn,p is the constant in (4.1). Now, observe that Lt ≤ d(x) for all x ∈ Ut. Indeed, if x ∈ Qt then √ Lt = 9 8 lt < n lt = diam(Qt) ≤ dist(Qt, ∂Ω) ≤ d(x), where lt is the side length of Qt. Now, if x ∈ Ut \ Qt then n lt ≤ dist(Qt, ∂Ω) ≤ dist(Ut, ∂Ω) + 1 √ √ n lt, 16 √ hence, 15 16 n lt ≤ dist(Ut, ∂Ω) and 8 lt < 15 Lt = 9 16 √ n lt ≤ dist(Ut, ∂Ω) ≤ d(x). Then, the validity of Lt ≤ d(x) for all x ∈ Ut implies (3.6) for all Ut, where µ(x, y) is the function defined in (5.4), and the uniform constant (5.5) Next, by Theorem 2.4, there is a finite C-decomposition of functions {gt}t∈Γ subor- dinate to {Ut}t∈Γ of any function g in W0 which satisfies (2.3) and (2.4). Moreover, using (5.3), it can be seen that C1 = Cn,pτ s−n2β. gt(x) ≤ g(x) + (72K)nT g(x), for all t ∈ Γ and x ∈ Ut. √ Now, ω(x) := dβ F (x) fulfills the hypothesis of Lemma 2.7 where the constant in n)β. The last assertion uses Lemma 5.2. Thus, the operator T is (2.5) is C2 = (3K −qβ continuous from L := Lq(Ω, d F (cid:18) qN ) to itself with the norm √ n)β. (3K (cid:19)1/q q − 1 (cid:107)T(cid:107)L→L ≤ 2 13 (cid:33)(cid:41) (x) Hence,(cid:88) −qβ F ) (cid:90) (cid:33) (x) g(x)q d −qβ F (cid:107)gt(cid:107)q t∈Γ ≤ 2q−1 ≤ 2q−1N ≤ 2q−1N Lq(Ut,d (cid:40)(cid:32)(cid:88) (cid:26)(cid:90) (cid:26) t∈Γ Ω Ut g(x)q d (cid:18) q 1 + (72K)qn2q −qβ F (cid:18) qN (cid:19) (cid:18) q (cid:19) q − 1 √ (3K q − 1 ≤ 4qN 2(72K)qn q − 1 √ = 4q 122n 72qn (3 n)qβ + (72K)qn (cid:90) (cid:19) (x) dx + (72K)qn √ n)qβ (3K Ω n)qβ(cid:107)g(cid:107)q Lq(Ω,d −qβ F ) K q(n+β)(cid:107)g(cid:107)q (cid:90) Ut (cid:32)(cid:88) (cid:27) t∈Γ (cid:107)g(cid:107)q T g(x)q d −qβ F (cid:27) T g(x)q d −qβ F (x) dx Lq(Ω,d −qβ F ) . −qβ F ) (cid:19)1/q Lq(Ω,d (cid:18) q q − 1 Therefore, we have a C-decomposition subordinate to {Ut}t∈Γ with constant √ C0 = 4(12)2n/q(72)n(3 n)β K n+β. (5.6) Finally, inequality (3.10) and the control on the overlapping of the tree covering by (cid:3) N = 12n implies (1.2). Remark 5.3. Notice that the proof of Theorem 1.1 provides an explicit constant C = 2C0C1 for inequality (1.2), where C0 and C1 are described respectively in (5.6) and (5.5). The next result, similar to Proposition 4.1, follows from the Holder inequality (equiv- alently, from Minkowski's integral inequality). Proposition 5.4. Let ρ : Rn \ {0} → R be a positive radial Lebesgue measurable function which is increasing with respect to the radius. Then, the fractional Poincar´e type inequality (cid:107)u(x) − uU(cid:107)Lp(U ) ≤ diam (U )n/p ρ(diam(U )) U1/p u(y) − u(x)p y − xn(ρy − x)p dydx (cid:18)(cid:90) (cid:90) U U holds for any bounded domain U in Rn and 1 < p < ∞, where uU := 1U Proof.(cid:90) (cid:90) (cid:90) (cid:90) (cid:12)(cid:12)(cid:12)(cid:12) 1 (cid:90) u(x) − u(y)dy ≤ diam(U )n{ρ(diam(U ))}p U U U (cid:12)(cid:12)(cid:12)(cid:12)p (cid:90) U (cid:90) dx ≤ 1 U u(x) − u(y)p U U u(x) − u(y)pdydx x − yn(ρx − y)p dydx. U U u(x) − uUpdx = U (cid:3) Remark 5.5. If ρ(x) = xs, with s ∈ (0, 1), we recover the classical fractional Poincar´e inequality. (cid:19)1/p (cid:82) (5.7) U u(y)dy. 14 We generalize the fractional Poincar´e inequality stated in Theorem 1.1 by replacing the fractional derivatives given by the power functions xs, with 0 < s < 1, by general increasing and positive radial functions ρx. Theorem 5.6. Let Ω in Rn be a bounded John domain and 1 < p < ∞. Given an arbitrary compact set F in ∂Ω, a parameter β ≥ 0 and a positive radial Lebesgue measurable function ρ : Rn \{0} → R increasing with respect to the radius, there exists a constant C such that u(x) − uΩ,ωpdpβ (cid:19)1/p (cid:18)(cid:90) F (x)dx Ω (cid:19)1/p (cid:18)(cid:90) (cid:90) ≤ C u(x) − u(y)p (5.8) for all function u ∈ Lp(Ω, d(x)pβ). We denote by d(x) and dF (x) the distance from x to ∂Ω and F respectively, and by uΩ,ω the weighted average x − yn(ρx − y)p [ρ(2d(x))]pdpβ (cid:82) F (x) dydx Ω∩B(x,d(x)) Ω Ω u(z)dpβ F (z)dz. 1 dpβ F (Ω) In addition, the constant C in (5.8) can be written as C = Cn,p,β K n+β, where K is the geometric constant introduced in (5.1). Proof. This proof mimics the one of Theorem 1.1 with Proposition 5.4 instead of Propo- sition 4.2. Indeed, we will use again the tree covering {Ut}t∈Γ of Ω defined in (5.2) and the weight ω(x) = dβ F (x), however, in this case µ(x, y) is defined as µ(x, y) := [ρ(2d(x))]pdpβ F (x) x − yn(ρx − y)p . We only have to show that (3.6) is verified on Ut, for all t, with uniform constant. This fact follows from (5.7) by using the inequality diam(Ut) ≤ 2d(x) for all x ∈ Ut. (cid:3) Acknowledgements This research was initiated when the second author visited the University of Helsinki during the Summer of 2016. This author gratefully acknowledges Professor Michel Lapidus from University of California Riverside for his financial support for the ex- penses of the trip. References [1] S. Buckley, P. Koskela, and G. Lu, Boman equals John, Proceedings of the 16th Rolf Nevanlinna Colloquium (1995), 91 -- 99. [2] B. Dyda, L. Ihnatsyeva, and A. Vahakangas, On improved fractional Sobolev-Poincar´e inequalities, Ark. Mat., 54 (2) (2016), 437 -- 454. [3] I. Drelichman, and R. G. Dur´an, Improved Poincar´e inequalities in fractional Sobolev spaces, arXiv: 1705.04227v1 (2017). [4] R. Hurri-Syrjanen, and A. V. Vahakangas, On fractional Poincar´e inequalities, J. Anal. Math. 120 (2013), 85 -- 104. [5] R. Hurri-Syrjanen, and A. V. Vahakangas, Fractional Sobolev-Poincar´e and fractional Hardy inequalities, Mathematika 61 (2015), 385 -- 401. [6] F. John, Rotation and strain Comm. Pure Appl. Math. 14 (1961), 391 -- 413. [7] F. L´opez-Garc´ıa, A decomposition technique for integrable functions with applications to the divergence problem, J. Math. Anal. Appl. 418 (2014), 79 -- 99. 15 [8] F. L´opez-Garc´ıa, Weighted Korn inequality on John domains, Studia Math. 241 (2018), 17 -- 39. [9] F. L´opez-Garc´ıa, Weighted generalized Korn inequality on John domains, submitted for publication (2017). [10] Augusto C. Ponce, A variant of of Poincar´e' s inequality, C. R. Acad. Sci. Paris, Ser. I 337 (2003), 253 -- 257. [11] Augusto C. Ponce, A new approach to Sobolev spaces and connections to Γ-convergence, Calc. Var. 19 (2004), 229 -- 255. [12] Augusto C. Ponce, An estimate in the spirit of Poincar´e's inequality, J. Eur. Math. Soc. 6 (2004), 1 -- 15. [13] E. M. Stein, Singular Integrals and Differentiability Properties of Functions, Princeton Univ. Press, Princeton, NJ, 1970. Department of Mathematics and Statistics, FI-00014 University of Helsinki, Finland E-mail address: [email protected] Department of Mathematics and Statistics, California State Polytechnic Univer- sity Pomona, 3801 West Temple Avenue, Pomona, CA (91768), US E-mail address: [email protected]
1702.08671	1	1702	2017-02-28T07:09:30	On The Absolute Value of The Product and the Sum of Linear Operators	[ "math.FA" ]	Let $A,B\in B(H)$. In the present paper, we establish simple and interesting facts on when we have $\|A\|\|B\|=\|B\|\|A\|$, $\|AB\|=\|A\|\|B\|$, $\|A\pm B\|\leq \|A\|+\|B\|$, $\|\|A\|-\|B\|\|\leq \|A\pm B\|$ and $\\|\|A\|-\|B\|\\|\leq \\|A\pm B\\|$, where $\|\cdot\|$ denotes the absolute value (or modulus) of an operator. The results give some other interesting consequences.	math.FA	math	ON THE ABSOLUTE VALUE OF THE PRODUCT AND THE SUM OF LINEAR OPERATORS MOHAMMED HICHEM MORTAD Abstract. Let A, B ∈ B(H). In the present paper, we establish simple and interesting facts on when we have AB = BA, AB = AB, A ± B ≤ A + B, A − B ≤ A ± B and kA − Bk ≤ kA ± Bk, where · denotes the absolute value (or modulus) of an operator. The results give some other interesting consequences. 1. Introduction Let H be a complex Hilbert space and let A, B ∈ B(H). We say that A is positive, and we write A ≥ 0, if < Ax, x >≥ 0 for all x ∈ H. Since H is a complex Hilbert space, a positive operator is clearly self-adjoint. We say that A ≥ B if they are both self-adjoint and A − B ≥ 0. Recall also that if A ≥ 0, then there is a unique positive operator B such that B 2 = A. We call it the (positive) square root of A and we denote it by √A (or A 2 ). Next, we gather basic results on square roots of sums and products. 1 Lemma 1.1. Let A, B ∈ B(H) be such that AB = BA and A, B ≥ 0. Then • AB ≥ 0. • √AB = √A√B. • √A + B ≤ √A + √B. The unique positive square root of the positive operator A∗A is commonly known as the absolute value (or modulus) of A. We denote it by A, that is, A = √A∗A. Notice that kAk = k A k always holds. We usually warn students to be careful with this notation as it may mislead them to think that e.g. A = A∗, A + B ≤ A + B or AB = AB would hold. Counterexamples are easily found in the setting of 2 by 2 ma- trices. Notice that A is normal if and only if A = A∗. Also, a priori if A, B are arbitrary, then there is no reason why we should expect AB = BA to hold (for instance, just think of positive operators). 2010 Mathematics Subject Classification. Primary 47A63, Secondary 47A62, 47B15, 47B20. Key words and phrases. Absolute Value. Triangle Inequality. Normal, Hyponormal, Self- adjoint and Positive Operators. Commutativity. Fuglede Theorem. 1 2 M. H. MORTAD Even when AB = BA, the equality AB = BA need not hold. For example, let A = (cid:18) 1 1 0 1 (cid:19) and B = (cid:18) 0 1 0 0 (cid:19) . Then as we can easily verify: AB = (cid:18) 0 1 0 0 (cid:19) = BA whereas AB 6= BA. terexample (cf. Proposition 2.1). Observe that we have purposely avoided normal operators in our coun- The main aim of this paper is to investigate when relations of the types • AB = BA; • AB = AB; • A + B ≤ A + B; • A − B ≤ A + B; • k A − B k ≤ kA + Bk; hold. It turns out that normality and sometimes hyponormality plus commu- tativity are sufficient for these relations to hold. This comes to corroborate the resemblance to complex numbers which is already known to many. Notice also that commutativity is not unnatural as we already have it in (C,×). The idea here is to start from scratch, and use as basic results as possible to make the paper accessible to a wide audience. We note that for example, we have wittingly avoided the use of the spectral theorem of normal operators. Therefore, most of the results here can be taught at elementary courses in Operator Theory. It is worth noticing that there is a big amount of papers which have dealt with inequalities involving absolute values and/or norms of operators. The literature is so rich that we rather refer readers to books which have gathered most of these results. For example, see [1], [4] and [10]. Finally, we assume the reader is familiar with other basic results on Op- erator Theory. A well established reference is [2]. We do recall two crucial results though. Theorem 1.2. (Löwner-Heinz Inequality, see [8] for a simple proof ) If A ≥ B ≥ 0, then Aα ≥ Bα for any α ∈ [0, 1]. Remark. It is known to readers that A ≥ B ≥ 0 implies that A2 ≥ B 2 when AB = BA. Since we will be dealing with sums and products of commuting normal operators, the use of the celebrated Fuglede-Putnam theorem is inevitable. The following lemma will be used below without further notice. Lemma 1.3. Let A, B ∈ B(H) where A is normal. Then, we have AB = BA ⇐⇒ A∗B = BA∗ ⇐⇒ AB∗ = B∗A ⇐⇒ A∗B∗ = B∗A∗. ABSOLUTE VALUE OF THE PRODUCT AND THE SUM OF OPERATORS 3 2. Main Results: Absolute Value and Products We start with the following: Proposition 2.1. Let A, B ∈ B(H) be such that AB = BA. If A is normal, then AB = BA. Remark. The preceding result was proved in [9] by assuming that both A and B are normal. Proof. Since AB = BA and A is normal, we have A∗B = BA∗. We then clearly have from the previous two relations: AB = BA =⇒ A∗AB = A∗BA = BA∗A. Hence AB = BA. Since A is self-adjoint, the previous equality gives (by taking adjoints) AB∗ = B∗A. Hence B∗AB = AB∗B =⇒ B∗BA = AB∗B =⇒ BA = AB, as required. (cid:3) We have already observed above that in general AB 6= AB. The following result is somewhat inspired by a one in [5]. Theorem 2.2. Let A, B ∈ B(H) be self-adjoint such that AB is normal. Then AB = AB. Remark. It was noted in [6] that if S, T are two non-commuting self-adjoint operators, then the inequality ST ≤ ST never holds. So, in our result the normality of the product transforms the non valid inequality into a true full equality. Remark. Notice that AB being a normal product of two self-adjoint operators does not necessarily imply that AB is self-adjoint, i.e. we do not necessarily have AB = BA. If, however, we impose further that A ≥ 0 (or B ≥ 0), then AB becomes self-adjoint. See e.g. [7]. Proof. Since A and B are self-adjoint, we may write B(AB) = BAB = (AB)∗B. Since AB and (AB)∗ are normal, the Fuglede-Putnam theorem gives B(AB)∗ = (AB)∗∗B or merely B 2A = AB 2. Consequently, B 2A2 = AB 2A = A2B 2. On the other hand, we easily see that AB2 = (AB)∗AB = AB(AB)∗ = AB 2A = A2B 2 4 M. H. MORTAD and so as required. AB = √A2B 2 = √A2√B 2 = AB, (cid:3) Since AB is self-adjoint, we have: Corollary 2.3. Let A, B ∈ B(H) be self-adjoint such that AB is normal. Then AB is self-adjoint, i.e. AB = BA. Moreover, if we also assume that A, B ≥ 0, then AB ≥ 0. The assumptions of the previous theorem cannot just be dropped. We give a counterexample for each hypothesis. • Let A = (cid:18) 2 0 1 0 (cid:19) . 0 −1 (cid:19) and B = (cid:18) 0 1 AB = (cid:18) 0 −1 0 (cid:19) 2 Then each of A and B is self-adjoint but AB is not normal for We can easily check that AB = (cid:18) 1 0 0 2 (cid:19) , A = (cid:18) 2 0 0 1 (cid:19) and B = (cid:18) 1 0 0 1 (cid:19) , i.e. • Let AB 6= AB. A = (cid:18) 0 1 2 0 (cid:19) and B = (cid:18) 0 2 1 0 (cid:19) . Then, neither A nor B is normal. Their product AB is, however, self-adjoint (hence normal!) because AB = (cid:18) 1 0 0 4 (cid:19) Next, we have A = (cid:18) 2 0 0 1 (cid:19) , B = (cid:18) 1 0 0 2 (cid:19) and AB = (cid:18) 1 0 0 2 (cid:19) . Accordingly, AB = (cid:18) 1 0 0 2 (cid:19) 6= (cid:18) 2 0 0 2 (cid:19) = AB. An akin result to Theorem 2.2 is: Theorem 2.4. Let A, B ∈ B(H) be such that AB = BA. If A is normal, then AB = AB. ABSOLUTE VALUE OF THE PRODUCT AND THE SUM OF OPERATORS 5 Remark. It was also noted in [6] that if S, T are two commuting normal operators, then the inequality ST ≤ ST holds. So, our result here is stronger. Proof. Since AB = BA and A is normal, we get A∗B = BA∗ or AB∗ = B∗A. Hence AB2 = (AB)∗AB = B∗A∗AB = A∗B∗AB = A∗AB∗B. By Proposition 2.1, A∗AB∗B = B∗BA∗A. Consequently, AB = √A∗AB∗B = √A∗A√B∗B = AB, as required. (cid:3) Before generalizing the previous result, we give some direct consequences. The first one is a funny application. Corollary 2.5. Let A, B ∈ B(H) be such that AB = BA. If A and B are normal, then AB = A∗B = AB∗ = A∗B∗ = B∗A∗ = B∗A = BA∗ = BA. Proof. Since A and B are normal, A = A∗ and B = B∗. As AB = BA, then AB = BA for A (or B!) is normal. Now, apply Theorem 2.4 to each of the eight products. (cid:3) Corollary 2.6. Let A, B ∈ B(H) be such that AB = BA. If A is normal and B is invertible, then AB−1 = AB−1. Proof. Since AB = BA and B is invertible, we have AB−1 = B−1A. Theo- rem 2.4 does the remaining job. (cid:3) Corollary 2.7. Let A ∈ B(H) be normal and invertible. Then A−1 = A−1. Proof. It is clear that So, the self-adjoint A is right invertible and so it is invertible (cf. [3]) and: I = AA−1 = AA−1. A−1 = A−1. Theorem 2.4 may be generalized as follows: Proposition 2.8. Let (Ai)i=1,··· ,n be a family of pairwise commuting ele- ments of B(H). If all (Ai)i=1,··· ,n but one are normal, then A1A2 ··· An−1An = A1A2··· An−1An. (cid:3) 6 M. H. MORTAD Proof. If A1, A2,··· , An−1 are normal, then just apply the preceding theorem by using a proof by induction. Otherwise, just use commutativity to push the non normal factor to the right as many times as possible until it will be i=1 Ai". Then proceed as just (cid:3) the last factor on the right of the product "Qn It is simple to see that A2 = A2 does not hold in general. For instance, indicated three lines above. let A = (cid:18) 0 2 1 0 (cid:19) . Then A2 = (cid:18) √2 0 0 √2 (cid:19) 6= A2 = (cid:18) 2 0 0 1 (cid:19) . But for normal A, things are better. Corollary 2.9. Let A ∈ B(H) be normal and invertible. Let n ∈ Z. Then An = An. Proof. The case n ≥ 0 follows from Proposition 2.8. The case n < 0 follows from Proposition 2.8 and Corollary 2.7. (cid:3) 3. Main Results: Absolute Value and Sums versions. We now turn to the triangle inequality w.r.t. · . We have two different Before all else, we state a result (perhaps known to many) which will be called on later. Its proof relies on the following yet simpler result. Lemma 3.1. If A ∈ B(H) is anti-symmetric, i.e. A∗ = −A, then A2 ≤ 0. Lemma 3.2. Let T ∈ B(H) be hyponormal (i.e. T T ∗ ≤ T ∗T , that is, kT ∗xk ≤ kT xk for all x ∈ H). Then T + T ∗ ReT = 2 √T ∗T = T. ≤ Proof. It is clear that T − T ∗ is anti-symmetric and so by Lemma 3.1: (T − T ∗)2 ≤ 0. Hence (T − T ∗)2 ≤ 0 ⇐⇒ T 2 + T ∗2 − T T ∗ − T ∗T ≤ 0. But, T is hyponormal and so −T T ∗ − T ∗T ≥ −2T ∗T . So, T 2 + T ∗2 − 2T ∗T ≤ 0 + T T ∗ + T ∗T ≤ T 2 + T ∗2 (T + T ∗)2 ≤ 4T ∗T or T + T ∗ ≤ 2T + 2T ∗T ≤ 4T ∗T. or Therefore, T 2 + T ∗2 by Theorem 1.2. ABSOLUTE VALUE OF THE PRODUCT AND THE SUM OF OPERATORS 7 Remembering that S− = 1 2 (S − S) ≥ 0 whenever S is self-adjoint, we conclude that as required. T + T ∗ ≤ T + T ∗ ≤ 2T = 2√T ∗T , (cid:3) The following fairly simple result is also useful to us. Lemma 3.3. Let A, B ∈ B(H) such that A is normal and B is hyponormal. If AB = BA, then A∗B is hyponormal. Proof. Let x ∈ H. As AB = BA, by the normality of A and the hyponor- mality of B we have k(A∗B)∗xk = kB∗Axk ≤ kBAxk = kABxk = kA∗Bxk, establishing the hyponormality of A∗B. (cid:3) Here is the first version of the triangle inequality. Theorem 3.4. Let A, B ∈ B(H) be such that AB = BA. If A is normal and B is hyponormal, then the following triangle inequality holds: A + B ≤ A + B. Proof. Since A is normal and AB = BA, we know from Proposition 2.1 that AB = BA. Hence A + B2 ≤ (A + B)2 ⇐⇒ (A + B)∗(A + B) ≤ A∗A + B∗B + 2√A∗A√B∗B We already know from above that √A∗A√B∗B = √A∗AB∗B. So, to ⇐⇒ A∗B + B∗A ≤ 2√A∗A√B∗B. prove the desired triangle inequality, we are only required to prove A∗B + B∗A ≤ 2√A∗AB∗B. But the following holds: A∗AB∗B = AA∗B∗B = AB∗A∗B = B∗AA∗B. If we set T = A∗B, then are done with the proof if we come to show that T + T ∗ ≤ 2√T ∗T . But this is just Lemma 3.2 once we show that A∗B is hyponormal. This is in effect the case as A∗B is hyponormal by Lemma 3.3. Therefore, under the assumptions of our theorem we have shown that A + B2 ≤ (A + B)2. Hence, by Theorem 1.2, we have ended up with A + B ≤ A + B, and this is precisely what we wanted to prove. (cid:3) 8 M. H. MORTAD Remark. The foregoing result need not hold if commutativity is dropped even if A and B are self-adjoint. The reader may check this easily via the following example: A = (cid:18) −1 1 1 −1 (cid:19) and B = (cid:18) 2 0 0 0 (cid:19) . Corollary 3.5. Let T ∈ B(H) be normal. Then T ≤ ReT + ImT. Proof. Write T = ReT + iImT where ReT and ImT are commuting self- adjoint operators. Then apply Theorem 3.4. (cid:3) We have another simple consequence of Theorem 3.4. Corollary 3.6. Let A, B ∈ B(H) be such that AB = BA. If A is normal and B is hyponormal, then the following triangle inequality holds: A − B ≤ A + B. Proof. Since AB = BA, we know that A(−B) = (−B)A. Also −B is hy- ponormal. Then, apply Theorem 3.4. (cid:3) Theorem 3.4 may be generalized to a finite sum of operators. Before, recall that the sum of two commuting normal operators remains normal. This too may be generalized (the proof by induction is omitted). Proposition 3.7. Let (Ai)i=1,··· ,n be a family of normal pairwise commuting elements of B(H). Then A1 + A2 + ··· + An is normal. We are ready for the promised generalization of Theorem 3.4 whose proof is again a proof by induction. Corollary 3.8. Let (Ai)i=1,··· ,n be a family of pairwise commuting elements of B(H). If all (Ai)i=1,··· ,n are normal except one which is assumed to be hyponormal, then A1 + A2 + ··· + An ≤ A1 + A2 + ··· + An. It is known that an inequality of the type kA − Bk ≤ kA ± Bk is not true in general even if A and B are self-adjoint. Proposition 3.9. Let A, B ∈ B(H) be such that AB = BA. If A and B are normal, then the following inequality holds: kA − Bk ≤ kA + Bk. Probably the following lemma has been noted elsewhere but we state it here anyway with a proof. Lemma 3.10. Let S, T ∈ B(H) be self-adjoint where S ≥ 0. If −S ≤ T ≤ S, then kTk ≤ kSk. ABSOLUTE VALUE OF THE PRODUCT AND THE SUM OF OPERATORS 9 Proof. By assumption, for all x ∈ H − < Sx, x >≤< T x, x >≤< Sx, x > or merely < T x, x > ≤< Sx, x > . Therefore, kTk = sup kxk=1 < T x, x > ≤ sup kxk=1 as desired. Let us prove Proposition 3.9. < Sx, x >= kSk, (cid:3) Proof. Since A and B are commuting normal operators, we know that A + B too is normal. Since A + B commutes with B, by Corollary 3.6 we have A = A + B − B ≤ A + B + B =⇒ A − B ≤ A + B. Similarly, as A + B commutes with A, we get Whence B − A ≤ A + B. −A + B ≤ A − B ≤ A + B. By Lemma 3.10 (and remembering that kTk = k T k for T ∈ B(H)), we obtain as required. kA − Bk ≤ k A + B k = kA + Bk, (cid:3) Remark. In the previous proposition, if B is only hyponormal, then at the moment we are only sure that: B − A ≤ A + B because we can only prove that A + B is hyponormal. We will remedy this little problem shortly. Corollary 3.11. Let A, B ∈ B(H) be such that AB = BA. If A and B are normal, then the following inequality holds: kA − Bk ≤ kA − Bk. Proposition 3.9 can be improved as it is a particular case of the following remarkable result: Proposition 3.12. Let A, B ∈ B(H) be such that AB = BA. If A is normal and B is hyponormal, then the following inequality holds: A − B ≤ A − B. Proof. We easily see as AB = BA that A − B2 ≤ A − B2 ⇐⇒ A2 + B2 − 2AB ≤ A2 + B2 − A∗B − B∗A ⇐⇒ A∗B + B∗A ≤ 2√A∗A√B∗B ⇐⇒ A∗B + B∗A ≤ 2√B∗AA∗B. 10 M. H. MORTAD But, this is always true in virtue of Lemma 3.2 as A∗B is hyponormal. Therefore, we have shown A glance at Theorem 1.2 finally gives A − B2 ≤ A − B2. A − B ≤ A − B. (cid:3) Corollary 3.13. Let A, B ∈ B(H) be such that AB = BA. If A is normal and B is hyponormal, then the following inequality holds: Proof. Since B is hyponormal, so is −B. The rest is obvious. A − B ≤ A + B. Here is the improvement of Proposition 3.9: (cid:3) Corollary 3.14. Let A, B ∈ B(H) be such that AB = BA. If A is normal and B is hyponormal, then the following inequality holds: Proof. By Proposition 3.9 and Corollary 3.13, we know that kA − Bk ≤ kA ± Bk. Then, calling on Lemma 3.10 yields A − B ≤ A ± B. kA − Bk = k A − B k ≤ k A ± B k = kA ± Bk. (cid:3) If we want to drop commutativity in Theorem 3.4, then this is at the cost of adding an extra condition. Also, we only have to assume that one of the two operators is normal. Theorem 3.15. Let A, B ∈ B(H) be such that AB = BA. If A is normal and A∗B + B∗A ≤ 0, then A + B ≤ A + B. A + B = √A∗A + A∗B + B∗A + B∗B ≤ √A∗A + B∗B. Since AB = BA and A is normal, Proposition 2.1 implies that AB = BA or A2B2 = B2A2. Finally, Lemma 1.1 does the remaining job, i.e. it gives us and this completes the proof. A + B ≤ A + B (cid:3) Proof. Clearly, As A∗B + B∗A ≤ 0, then By Theorem 1.2, we have (A + B)∗(A + B) = A∗A + A∗B + B∗A + B∗B. A∗A + A∗B + B∗A + B∗B ≤ A∗A + B∗B. ABSOLUTE VALUE OF THE PRODUCT AND THE SUM OF OPERATORS 11 acknowledgement The author wishes to thank Mr. S. Dehimi for a discussion which led to a slight improvement of the result of Theorem 3.4. References [1] R. Bhatia, Matrix analysis, Graduate Texts in Mathematics, 169. Springer-Verlag, New York, 1997. [2] J. B. Conway, A course in functional analysis, Graduate Texts in Mathematics, 96. Springer, 1990 (2nd edition). [3] S. Dehimi, M. H. Mortad, Right (or left) invertibility of bounded and un- (submitted). bounded operators and applications to the spectrum of products, https://arxiv.org/pdf/1505.02719v1.pdf. [4] T. Furuta, Invitation to linear operators: From matrices to bounded linear operators on a Hilbert Space, Taylor & Francis, Ltd., London, 2001. [5] K. Gustafson, M. H. Mortad, Conditions implying commutativity of unbounded self- adjoint operators and related topics, J. Operator Theory, 76/1 (2016) 159-169. [6] R. Harte, The triangle inequality in C ∗ algebras, Filomat No. 20, part 2 (2006), 51-53. [7] M. H. Mortad, An application of the Putnam-Fuglede theorem to normal products of selfadjoint operators, Proc. Amer. Math. Soc. 131 (2003) 3135-3141. [8] G. K. Pedersen, Some operator monotone functions, Proc. Amer. Math. Soc., 36 (1972) 309-310. [9] R. Zeng, Young's inequality in compact operators-the case of equality, JIPAM. J. In- equal. Pure Appl. Math., 6/4 (2005), Article 110, 10 pp. [10] X. Zhan, Matrix inequalities, Lecture Notes in Mathematics, 1790. Springer-Verlag, Berlin, 2002. Department of Mathematics, University of Oran 1, Ahmed Ben Bella, B.P. 1524, El Menouar, Oran 31000, Algeria. Mailing address: Pr Mohammed Hichem Mortad BP 7085 Seddikia Oran 31013 Algeria E-mail address: [email protected], [email protected].
1711.04122	2	1711	2018-01-29T21:22:52	Bohr's equivalence relation in the space of Besicovitch almost periodic functions	[ "math.FA" ]	Based on Bohr's equivalence relation which was established for general Dirichlet series, in this paper we introduce a new equivalence relation on the space of almost periodic functions in the sense of Besicovitch, $B(\mathbb{R},\mathbb{C})$, defined in terms of polynomial approximations. From this, we show that in an important subspace $B^2(\mathbb{R},\mathbb{C})\subset B(\mathbb{R},\mathbb{C})$, where Parseval's equality and Riesz-Fischer theorem holds, its equivalence classes are sequentially compact and the family of translates of a function belonging to this subspace is dense in its own class.	math.FA	math	Noname manuscript No. (will be inserted by the editor) Bohr's equivalence relation in the space of Besicovitch almost periodic functions J.M. Sepulcre · T. Vidal Received: date / Accepted: date Abstract Based on Bohr's equivalence relation which was established for gen- eral Dirichlet series, in this paper we introduce a new equivalence relation on the space of almost periodic functions in the sense of Besicovitch, B(R, C), defined in terms of polynomial approximations. From this, we show that in an important subspace B2(R, C) ⊂ B(R, C), where Parseval's equality and Riesz-Fischer theorem holds, its equivalence classes are sequentially compact and the family of translates of a function belonging to this subspace is dense in its own class. Keywords Almost periodic functions · Besicovitch almost periodic functions · Bochner's theorem · Exponential sums · Fourier series Mathematics Subject Classification (2010) 42A75 · 42A16 · 42B05 · 46xx · 42Axx · 30B50 · 30Bxx 1 Introduction The class of almost periodic functions, whose theory was created and developed in its main features by H. Bohr during the 1920's, is the class of continuous functions possessing a certain structural property, which is a generalization of pure periodicity. This theory opened a way to study a wide class of trigono- metric series of the general type and even exponential series. In this context, we can cite, among others, the papers [3,4,5,6,8,10]. Let f (t) be a real or complex function of an unrestricted real variable t. The notion of almost periodicity given by Bohr involves the fact that f (t) must be continuous, and for every ε > 0 there corresponds a number l = l(ε) > 0 such J.M. Sepulcre · T. Vidal Department of Mathematics University of Alicante 03080-Alicante, Spain E-mail: [email protected], [email protected] 8 1 0 2 n a J 9 2 ] . A F h t a m [ 2 v 2 2 1 4 0 . 1 1 7 1 : v i X r a 2 J.M. Sepulcre, T. Vidal that each interval of length l contains a number τ satisfying f (t+τ )−f (t) ≤ ε for all t. As in [8], we will denote as AP (R, C) the space of almost periodic functions in the sense of this definition (Bohr's condition). A very important result of this theory is the approximation theorem according to which the class of almost periodic functions AP (R, C) coincides with the class of limit functions of uniformly convergent sequences of trigonometric polynomials of the type a1eiλ1t + . . . + aneiλnt (1) with arbitrary real exponents λj and arbitrary complex coefficients aj. More- over, S. Bochner observed that Bohr's notion of almost periodicity of a function f is equivalent to the relative compactness, in the sense of uniform convergence, of the family of its translates {f (t + h)}, h ∈ R. In the course of time, some variants and extensions of Bohr's concept have been introduced, most notably by A. S. Besicovitch, W. Stepanov and H. Weyl. We refer the reader to the papers by Besicovitch [3, Chapter II], Bohr and Foelner [7], Corduneanu [8], and by Andres, Bersani and Grande [1] and the references therein for a comprehensive treatment of this subject. In particular, A.S. Besicovitch enlarged the class of almost periodic func- tions by considering the convergence of sequences of functions in a more general sense than uniform convergence. In this way, the Besicovitch spaces Bp(R, C), 1 ≤ p < ∞, are obtained by the completion of the trigonometric polynomials of the form (1) with respect to the seminorms lim sup l→∞ (2l)−1Z l −l f (t)pdt!1/p . This topology is certainly weaker than that of the uniform convergence. In particular, the space B1(R, C) is denoted by B(R, C) and contains AP (R, C), B2(R, C) and all variants of almost periodic functions which were mentioned above. Moreover, for any function f ∈ B(R, C) there exists the mean value M (f ) = lim l→∞ (2l)−1Z l −l f (t)dt (2) and, at most, a countable set of values of λk ∈ R such that ak = a(f, λk) = M (f (t)e−λkt) 6= 0. Thus the series Pk≥1 akeiλkt is called the Fourier series of f [8, Section 4.2]. Also, λk and ak are called the Fourier exponents and coefficients of the function f , respectively. In the case that f ∈ B2(R, C), with Pk≥1 akeiλkt the Fourier series of f , it is accomplished the Parseval's equality [3, p. 109] ak2 = M (f (t)2) < ∞. Xk≥1 In this respect, if f1, f2 ∈ B2(R, C), then f1 and f2 have the same Fourier series if and only if M (f1(t) − f2(t)2) = 0. That is, two functions satisfying Bohr's equivalence relation in the space of Besicovitch almost periodic functions 3 this condition belong to the same class of equivalence defined in terms of the Fourier series. This equivalence relation is inherent to the classes Bp(R, C) and it is different from the generalization of Bohr's equivalence of Definition 2 which is the main tool of this paper. Besicovitch's generalization is interesting because, for this extension, the analogue of the Riesz-Fischer theorem is also valid, that is to say, any trigono- B2(R, C) almost periodic function [3, p. 110] (in this sense, B2(R, C) is also called AP2(R, C) in [8]). This is not the case for some Stepanov or Weyl func- metric series Pn≥1 aneiλnt with Pn≥1 an2 finite is the Fourier series of a tions [10]. As a consequence of the above, a Fourier series Pn≥1 aneiλnt so thatPn≥1 an2 < ∞ represents an equivalence class of functions in B2(R, C). In this paper we extend Bohr's equivalence relation to the Fourier series associated with the Besicovitch almost periodic functions, and hence to the Besicovitch almost periodic functions too. In this way, in view of the analogue of the Riesz-Fischer theorem and with respect to the topology of B2(R, C), the main result of our paper shows that, fixed an almost periodic function in B2(R, C), the limit points of the set of its translates are precisely the functions which are equivalent to it (see Theorem 2 in this paper). This means that the Bochner-type property, which is satisfied for the Besicovitch classes of almost periodic functions defined as above in terms of polynomials approximations (see [1, Definition 5.5, Definition 5.17 and Theorem 5.34] or [8, Section 3.4, p. 65]), is now refined for B2(R, C) in the sense that we show that the condi- tion of almost periodicity in the Besicovitch sense implies that every sequence of translates has a subsequence that converges in B2(R, C) to an equivalent function. 2 Preliminary definitions and results on exponential sums We shall refer to the expressions of the type P1(p)eλ1p + . . . + Pj(p)eλj p + . . . as exponential sums, where the frequencies λj are complex numbers and the Pj(p) are polynomials in p. In this paper we are going to consider some func- tions which are associated with a concrete subclass of these exponential sums, where the parameter p will be changed by t in the real case. In this way, as in [11], we take the following definition. Definition 1 Let Λ = {λ1, λ2, . . . , λj, . . .} be an arbitrary countable set of distinct real numbers, which we will call a set of exponents or frequencies. We will say that an exponential sum is in the class SΛ if it is a formal series of type ajeλj p, aj ∈ C, λj ∈ Λ. Xj≥1 (3) We next introduce an equivalence relation on the classes SΛ. 4 J.M. Sepulcre, T. Vidal Definition 2 Given an arbitrary countable set Λ = {λ1, λ2, . . . , λj , . . .} of distinct real numbers, consider A1(p) and A2(p) two exponential sums in the class SΛ, say A1(p) =Pj≥1 ajeλj p and A2(p) =Pj≥1 bjeλj p. We will say that ∼A2) if for each integer A1 is equivalent to A2 (in that case, we will write A1 value n ≥ 1, with n ≤ ♯Λ, there exists a Q-linear map ψn : Vn → R, where Vn is the Q-vector space generated by {λ1, λ2, . . . , λn}, such that ∗ bj = ajeiψn(λj ), j = 1, . . . , n. ∗ equivalence relation. It is plain that the relation ∼ considered in the foregoing definition is an Let GΛ = {g1, g2, . . . , gk, . . .} be a basis of the vector space over the ra- tionals generated by a set Λ of exponents, which implies that GΛ is linearly independent over the rationals and each λj is expressible as a finite linear combination of terms of GΛ, say λj = qj Xk=1 rj,kgk, for some rj,k ∈ Q. (4) By abuse of notation, we will say that GΛ is a basis for Λ. Moreover, we will say that GΛ is an integral basis for Λ when rj,k ∈ Z for any j, k. By taking this into account, the equivalence relation introduced in Definition 2 can be characterized in terms of a basis for Λ (see [11, Proposition 1']). ∗ ∗ ∗ Proposition 1 Given Λ = {λ1, λ2, . . . , λj, . . .} a set of exponents, consider A1(p) and A2(p) two exponential sums in the class SΛ, say A1(p) =Pj≥1 ajeλj p and A2(p) = Pj≥1 bjeλj p. Fixed a basis GΛ for Λ, for each j ≥ 1 let rj be ∼A2 if and only the vector of rational components satisfying (4). Then A1 if for each integer value n ≥ 1, with n ≤ ♯Λ, there exists a vector xn = (xn,1, xn,2, . . . , xn,k, . . .) ∈ R♯GΛ such that bj = aje<rj ,xn>i for j = 1, 2, . . . , n. Furthermore, if GΛ is an integral basis for Λ then A1 ∼A2 if and only if there exists x0 = (x0,1, x0,2, . . . , x0,k, . . .) ∈ R♯GΛ such that bj = aje<rj ,x0>i for every j ≥ 1. Proof For each integer value n ≥ 1, let Vn be the Q-vector space generated by {λ1, . . . , λn}, V the Q-vector space generated by Λ, and GΛ = {g1, g2, . . . , gk, . . .} ∼A2, by Definition 2 for each integer value n ≥ 1, with a basis of V . If A1 n ≤ ♯Λ, there exists a Q-linear map ψn : Vn → R such that bj = ajeiψn(λj ), j = ij 1, 2 . . . , n. Hence bj = ajei P k=1 rj,kψn(gk), j = 1, 2 . . . , n or, equivalently, bj = ajei<rj ,xn>, j = 1, 2 . . . , n, with xn := (ψn(g1), ψn(g2), . . .). Conversely, suppose the existence, for each integer value n ≥ 1, of a vector of the form xn = (xn,1, xn,2, . . . , xn,k, . . .) ∈ R♯GΛ such that bj = aje<rj ,xn>i, j = 1, 2 . . . , n. Thus a Q-linear map ψn : Vn → R can be defined from ψn(gk) := xn,k, k ≥ 1. Therefore ψn(λj) = Pij k=1 rj,kψ(gk) = < rj, xn >, j = 1, 2 . . . , n, and the result follows. Bohr's equivalence relation in the space of Besicovitch almost periodic functions 5 ∗ Now, suppose that GΛ is an integral basis for Λ and A1 ∼A2. Thus, by above, for each fixed integer value n ≥ 1, let xn = (xn,1, xn,2, . . .) ∈ R♯GΛ be a vector such that bj = ajei<rj ,xn>, j = 1, 2 . . . , n. Since each component of rj is an integer number, without loss of generality, we can take xn ∈ [0, 2π)♯GΛ as the unique vector in [0, 2π)♯GΛ satisfying the above equalities, where we assume xn,k = 0 for any k such that rj,k = 0 for j = 1, . . . , n. Therefore, under this assumption, if m > n then xm,k = xn,k for any k so that xn,k 6= 0. In this way, we can construct a vector x0 = (x0,1, x0,2, . . . , x0,k, . . .) ∈ [0, 2π)♯GΛ such that bj = aje<rj,x0>i for every j ≥ 1. Indeed, if r1,k 6= 0 then the component x0,k is chosen as x1,k, and if r1,k = 0 then each component x0,k is defined as xn+1,k where rj,k = 0 for j = 1, . . . , n and rn+1,k 6= 0. Conversely, if there exists x0 = (x0,1, x0,2, . . . , x0,k, . . .) ∈ R♯GΛ such that bj = aje<rj ,x0>i for every j ≥ 1, then it is clear that A1 ∼A2 under Definition 2. ∗ Λ then λj =Pij Λ, and so on. In this way, if λj ∈ G∗ In terms of the natural basis, we next prove another characterization which On the other hand, we will say that GΛ is the natural basis for Λ, and we will denote it as G∗ Λ, when it is constituted by elements in Λ. That is, firstly if λ1 6= 0 then g1 := λ1 ∈ G∗ Λ. Secondly, if {λ1, λ2} are Q-rationally independent, then g2 := λ2 ∈ G∗ Λ. Otherwise, if {λ1, λ3} are Q-rationally independent, then g2 := λ3 ∈ G∗ Λ then rj,mj = 1 and rj,k = 0 for k 6= mj, where mj is such that gmj = λj. In fact, each element in G∗ Λ is of the form gmj for j such that λj is Q-linear independent of the previous elements in the basis. Furthermore, if λj /∈ G∗ k=1 rj,kgk, where {g1, g2, . . . , gij} ⊂ {λ1, λ2, . . . , λj−1}. will be used later. Proposition 2 Given Λ = {λ1, λ2, . . . , λj, . . .} a set of exponents, consider A1(p) and A2(p) two exponential sums in the class SΛ, say A1(p) =Pj≥1 ajeλj p and A2(p) =Pj≥1 bjeλj p. Fixed the natural basis G∗ Λ = {g1, g2, . . . , gk, . . .} for Λ, for each j ≥ 1 let rj ∈ R♯G∗ Λ be the vector of rational components verifying ∼A2 if and only if there exists x0 = (x0,1, x0,2, . . . , x0,k, . . .) ∈ (4). Then A1 [0, 2π)♯G∗ Λ such that for each j = 1, 2, . . . it is satisfied bj = aje<rj ,x0+pj >i for some pj = (2πnj,1, 2πnj,2, . . .) ∈ R♯G∗ Λ} and Proof Suppose that A1 In = {1, 2, . . . , k, . . . , n : λk ∈ G∗ Λ}. Let j ∈ I, then rj,mj = 1 and rj,k = 0 for k 6= mj, where mj is such that gmj = λj. Thus, by Proposition 1, let xj = (xj,1, xj,2, . . .) ∈ R♯G∗ ∼A2. Consider I = {1, 2, . . . , k, . . . : λk ∈ G∗ Λ, with nj,k ∈ Z. Λ be a vector such that ∗ ∗ ij k=1 rj,kxj,k = ajeirj,mj xj,mj = ajeixj,mj . bj = ajei<rj ,xj> = ajei P Define x0 = (x0,1, x0,2, . . .) ∈ R♯G∗ Λ = R♯I as x0,mj := xj,mj for j ∈ I. Thus, by taking pj = (0, 0, . . .), the result trivially holds for those j's such that λj ∈ G∗ Λ, i.e. for j ∈ I. Now, let j be such that λj /∈ G∗ Λ, i.e. j /∈ I. By Proposition 1, let xj = (xj,1, xj,2, . . .) ∈ R♯G∗ Λ be a vector such that bp = apei<rp,xj> = apei P ij k=1 rp,kxj,k , p = 1, 2, . . . , j. 6 J.M. Sepulcre, T. Vidal Note that if p = 1, 2, . . . , j is such that λp ∈ G∗ bp = apeirp,mp xj,mp , Λ, then which necessarily implies that rp,mp xj,mp = rp,mp xp,mp + 2πnp, i.e. xj,mp = xp,mp + 2πnj,p for some nj,p ∈ Z. Hence bj = ajei<rj ,xj > = ajei P ij k=1 rj,kxj,k = aje i Pp∈Ij−1 rj,mp xj,mp = i Pp∈Ij−1 rj,mp (xp,mp +2πnj,p) aje = ajei<rj ,x0+pj >, Λ without loss of generality. where pj = (2πnj,1, 2πnj,2, . . . , 0, 0, . . .). Moreover, by changing conveniently the vectors pj, we can take x0 ∈ [0, 2π)♯G∗ Conversely, suppose the existence of x0 = (x0,1, x0,2, . . . , x0,k, . . .) ∈ R♯G∗ satisfying bj = aje<rj ,x0+pj>i for some pj = (2πnj,1, 2πnj,2, . . .) ∈ R♯G∗ Λ, with nj,k ∈ Z. Let rj,k = pj,k with pj,k and qj,k coprime integer numbers, and define qn,k := lcm(q1,k, q2,k, . . . , qn,k) for each k = 1, 2, . . .. Thus, for any integer number n ≥ 1, take xn = x0+mn, where mn,k = 2πp1,kp2,k ··· pn,kqn,k, k = 1, 2, . . .. Therefore, it is satisfied bj = aje<rj,xn>i for each j = 1, 2, . . . , n, which implies that A1 qj,k Λ ∗ ∼A2. As corollary, we can formulate the following result. Corollary 1 Given Λ = {λ1, λ2, . . . , λj, . . .} a set of exponents, consider A1(p) and A2(p) two exponential sums in the class SΛ, say A1(p) = Pj≥1 ajeλj p and A2(p) = Pj≥1 bjeλj p. Fixed a basis GΛ = {g1, g2, . . . , gk, . . .} for Λ, for each j ≥ 1 let rj ∈ R♯GΛ be the vector of rational components verifying ∼A2 if and only if there exists x0 = (x0,1, x0,2, . . . , x0,k, . . .) ∈ (4). Then A1 [0, 2π)♯GΛ such that for each j = 1, 2, . . . it is satisfied bj = aje<rj,x0+qj >i for some qj ∈ R♯GΛ which are of the form qj = T · pt j, where T is the change of basis matrix, with respect to the natural basis, and pj is of the form (2πnj,1, 2πnj,2, . . . , 2πnj,k, . . .), nj,k ∈ Z. ∗ In particular, note that the coefficients of equivalent exponential sums have the same modulus. 3 The finite exponential sums of the classes PR,Λ This section is focused on the following classes of finite exponential sums. Definition 3 Let Λ = {λ1, . . . , λn} be a set of n ≥ 1 distinct real numbers. We will say that a function f : R 7→ C is in the class PR,Λ if it is of the form (5) f (t) = a1eiλ1t + . . . + aneiλnt, aj ∈ C, λj ∈ Λ, j = 1, . . . , n. The functions f (t) of type (5) are also called trigonometric polynomials. Note that Definition 2 can be particularized to the classes PR,Λ. Further- more, if Λ is finite it is clear that it is always possible to find an integral basis for Λ. In this context, we next prove the following important result. Bohr's equivalence relation in the space of Besicovitch almost periodic functions 7 Theorem 1 Given Λ = {λ1, λ2, . . . , λn} a set of n ≥ 1 exponents, let a1eiλ1t+ . . . + aneiλnt and b1eiλ1t + . . . + bneiλnt be two equivalent functions in the class PR,Λ. Fixed d > 0 and ε > 0, there exists τ > d such that n Xj=1 ajeiλj τ − bj < ε. Proof Let GΛ = {g1, . . . , gm}, for a certain m ≥ 1, be linearly independent over the rationals so that each λj ∈ Λ is expressible as a linear combination of its terms, say λj = m Xk=1 rj,kgk, for some rj,k = pj,k qj,k ∈ Q, j = 1, 2, . . . , n. (6) Consider ε > 0, q := lcm(qj,k : j = 1, . . . , n, k = 1, . . . , m), r := max{rj,k : j = 1, . . . , n, k = 1, . . . , m} > 0 and a := max{aj : j = 1, 2, . . . , n} > 0. Since a1eiλ1t + . . . + aneiλnt and b1eiλ1t + . . . + bneiλnt are equivalent, Proposition 1 assures the existence of a vector of real numbers x0 = (x0,1, x0,2, . . . , x0,m) such that bj = aje<rj ,x0>i = ajei Pm k=1 rj,kx0,k , j = 1, 2, . . . , n. (7) Now, as the numbers ck = gk 2πq , k = 1, 2, . . . , m, are rationally independent, we next apply Kronecker's theorem [9, p.382] with the following choice: ck, ε1 = 2πq , k = 1, 2, . . . , m. In this manner we assure the existence of a real number τ > d > 0 and integer numbers e1, e2, . . . , em such that a·m·n·r·E > 0 and dk = x0,k ε that is τ ck − ek − dk =(cid:12)(cid:12)(cid:12)(cid:12) τ gk 2πq − ek − < ε1, x0,k 2πq(cid:12)(cid:12)(cid:12)(cid:12) τ gk = 2πqek + x0,k + ηk, with ηk < ε1. Therefore, from (6) and (7), with t ∈ R, we have (8) = ajeiλj τ − ajei Pm n Xj=1(cid:12)(cid:12)(cid:12) n ajeiλj τ − bj = Xj=1 aj(cid:12)(cid:12)(cid:12) eiτ λj − ei Pm n n Xj=1 k=1 rj,kx0,k(cid:12)(cid:12)(cid:12) ≤ a eiτ Pm k=1 rj,kgk − ei Pm n eiτ λj − ei Pm Xj=1(cid:12)(cid:12)(cid:12) k=1 rj,kx0,k(cid:12)(cid:12)(cid:12) , a Xj=1(cid:12)(cid:12)(cid:12) k=1 rj,kx0,k(cid:12)(cid:12)(cid:12) ≤ k=1 rj,kx0,k(cid:12)(cid:12)(cid:12) which, from (8), is equal to a n Xj=1(cid:12)(cid:12)(cid:12) ei Pm k=1(rj,k2πqek+rj,kx0,k+rj,kηk) − ei Pm = k=1 rj,kx0,k(cid:12)(cid:12)(cid:12) 8 J.M. Sepulcre, T. Vidal a n Xj=1(cid:12)(cid:12)(cid:12) anr ei Pm k=1 rj,kηk − 1(cid:12)(cid:12)(cid:12) ≤ a Xk=1 Xk=1 ηk < anr m m n Xj=1(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ε ≤ m Xk=1 rj,kηk(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) = ε. a · m · n · r As an immediate consequence of Theorem 1, we obtain the following corol- lary (compare with [11, Corollary 3]). Corollary 2 Given Λ = {λ1, λ2, . . . , λn} a finite set of exponents, let f1(t) = Pn j=1 ajeiλj t and f2(t) =Pn j=1 bjeiλj t be two equivalent functions in the class PR,Λ. Fixed ε > 0, there exists a relatively dense set of real numbers τ such that f1(t + τ ) − f2(t) < ε ∀t ∈ R. Proof Fixed τ > 0, note that for any t ∈ R it is accomplished that f1(t + τ ) − f2(t) ≤ n Xj=1 ajeiλj (t+τ ) − bjeiλj t = n Xj=1 ajeiλj τ − bj. Thus, by Theorem 1 and given d > 0, there exists τ1 > d such that f1(t + τ1) − f2(t) < ε/2 ∀t ∈ R. (9) Moreover, since f1(t) is almost periodic, there exists a real number l = l(ε) such that every interval of length l contains at least one translation number τ , associated with ε, satisfying f1(t + τ ) − f1(t) ≤ ε/2 for all t ∈ R. (10) Consequently, from (9) and (10) we deduce the existence of a relatively dense set of real numbers τ such that any t ∈ R satisfies f1(t + τ + τ1) − f2(t) ≤ f1(t + τ1 + τ ) − f1(t + τ1) + f1(t + τ1) − f2(t) < ε. This proves the result. It was proved in [11, Proposition 2] that, with respect to the topology of uniform convergence, the equivalence classes in PR,Λ/ ∼ are sequentially compact. We can analogously prove that this property is also true with respect to the topology of B2(R, C) (see the proof in [11, Proposition 2]). ∗ Proposition 3 Let Λ be a finite set of exponents and G an equivalence class in PR,Λ/ ∼. Thus G is sequentially compact. ∗ Bohr's equivalence relation in the space of Besicovitch almost periodic functions 9 4 Besicovitch almost periodic functions in terms of an equivalence relation For our purposes, we next focus our attention on the Besicovitch space B(R, C), whose functions are obtained by the completion of the trigonometric polynomi- −l f (t)dt(cid:17) (see for example als with respect to the seminorm lim supl→∞(cid:16) 1 2lR l [8, Section 3.4]). In particular, the space of functions B(R, C) contains those of the space of the almost periodic functions AP (R, C) and those functions of B2(R, C). We recall that every function in B(R, C) is associated with a real its Fourier series. exponential sum with real frequencies of the formPj≥1 ajeiλj t, which is called Definition 4 Let Λ = {λ1, λ2, . . . , λj, . . .} be an arbitrary countable set of distinct real numbers. We will say that a function f : R → C is in the class FB2,Λ if it is an almost periodic function in B2(R, C) whose associated Fourier series is of the form (11) ajeiλj t, aj ∈ C, λj ∈ Λ. Xj≥1 It is worth noting that, in general, when we write that a function f is in B(R, C) we do not have in mind the function f itself, it does represent a whole class of equivalent functions according to the relation f1 ≃ f2 if and only if lim sup l→∞ 1 2l Z l −l f (t) − g(t)2dt! = 0. In terms of Definition 2, we can define an equivalence relation on the func- tions in B(R, C), in particular on the classes FB2,Λ. More specifically, we es- tablish the following definition. Definition 5 Given Λ = {λ1, λ2, . . . , λj, . . .} a set of exponents, let f1 and f2 denote two equivalence classes of B(R, C)/ ≃ whose associated Fourier series are given by Xj≥1 ajeiλj t and Xj≥1 bjeiλj t, aj, bj ∈ C, λj ∈ Λ. We will say that f1 is equivalent to f2 if for each integer value n ≥ 1 there exists a Q-linear map ψn : Vn → R, where Vn is the Q-vector space generated by {λ1, λ2, . . . , λn}, such that bj = ajeiψn(λj ), j = 1, . . . , n. In that case, we will write f1 ∗ ∼f2. The next important lemma allows us to prove that if a function f2 is equivalent (in the sense of Definition 5) to a function f1 belonging to the space B2(R, C), then f2 also belongs to B2(R, C). This is clearly a consequence of Riesz-Fischer theorem [3, p. 110]. 10 J.M. Sepulcre, T. Vidal Lemma 1 Let f1(t) ∈ B2(R, C) be an almost periodic function whose Fourier series is given by Pj≥1 ajeiλj t, aj ∈ C, where {λ1, . . . , λj, . . .} is a set of dis- tinct exponents. Consider bj ∈ C such that Pj≥1 bjeiλj t and Pj≥1 ajeiλj t are equivalent. Then Pj≥1 bjeiλj t is the Fourier series associated with an almost periodic function f2(t) ∈ B2(R, C) so that f1 Proof Take Λ = {λ1, . . . , λj , . . .}. By the hypothesis, f1 ∈ FB2,Λ ⊂ B2(R, C) is determined by the series Pj≥1 ajeiλj t, aj ∈ C, λj ∈ Λ. Moreover, since Pj≥1 ajeiλj t ∗ ∼Pj≥1 bjeiλj t, we deduce from Corollary 1 that bj = aj for j ≥ 1 and hence ∼f2. ∗ aj2 < ∞. Xj≥1 bj2 =Xj≥1 By Riesz-Fischer theorem [3, p. 110], there exists a function f2 ∈ B2(R, C) such that the values bn are the Fourier coefficients of f2. As it was said before, it is worth noting that a Fourier seriesPn≥1 aneiλnt, such that Pn≥1 an2 < ∞, represents an equivalence class (according to the relation ≃) of functions in B2(R, C) (not a single function). In fact, as we pointed out in introduction, since two almost periodic functions in the Besi- covitch sense are connected in B2(R, C) when they have the same Fourier series ([3, p. 148] or [8, Section 4.2]), we immediately deduce from the results above the following corollary. Corollary 3 Let f1(t) and f2(t) be two equivalent functions in B(R, C). If f1(t) ∈ B2(R, C), then f2(t) ∈ B2(R, C). The following result is concerned with the concept of convergence in B2(R, C) which is certainly weaker than the uniform convergence. Under this topology, ∼ are closed. In fact, more we next show that the equivalence classes of FB2,Λ/ specifically, they are sequentially compact. ∗ ∗ ∼. Thus G is sequentially compact. Proposition 4 Let Λ be a set of exponents and G an equivalence class in FB2,Λ/ ∼. For Proof Let {fl}l≥1 be a sequence in an equivalence class G in FB2,Λ/ each l = 1, 2, . . ., suppose that the Fourier series which is associated with fl(t) is given by ∗ al,jeiλj t with al,j ∈ C, λj ∈ Λ. Xj≥1 Fixed a basis GΛ = {g1, g2, . . . , gk, . . .} for Λ, let rj = (rj,1, rj,2, . . .) be the vector satisfying < rj, g >= λj for each j ≥ 1, where g = (g1, g2, . . . , gk, . . .). Since f1 ∼fl for each l = 1, 2, . . ., we deduce from Proposition 1 that for each integer value n ≥ 1 there exists xl,n = (xl,n,1, xl,n,2, . . .) ∈ R♯GΛ such that ∗ al,j = a1,jei<rj ,xl,n>, j = 1, 2 . . . , n. (12) Bohr's equivalence relation in the space of Besicovitch almost periodic functions 11 Given l ≥ 1, let Pl,k(t) =Pj≥1 pj,kal,jeiλj t, k = 1, 2, . . ., be the Bochner-Fej´er polynomials which converge to fl with respect to the topology of B2(R, C) (and converge formally to its Fourier series on R) [3, p. 105, Theorem II]. It is worth noting that for each k only a finite number of the factors pj,k dif- fer from zero, and these factors pj,k do not depend on l [3, p. 48]. Thus, by taking into account (12), it is clear that {Pl,1(t)}l≥1 is a sequence of equivalent trigonometric polynomials and, by Proposition 3, there exists a subsequence {Plm,1,1(t)}m≥1 ⊂ {Pl,1(t)}l≥1 convergent to a certain P1(t) = Pj≥1 pj,1ajeiλj t ∈ PR,Λ1 , where Λ1 = {λj ∈ Λ : pj,1 6= 0}, which is in the same equivalence class as P1,1(t). Furthermore, by Proposition 1, this means that there exists x(1) 0,2, . . .) ∈ Rm1 such that 0 = (x(1) 0,1, x(1) pj,1aj = pj,1a1,jei<rj ,x (1) 0 >, j = 1, 2 . . . , with λj ∈ Λ1, where m1 is the number of elements of any basis for Λ1. Equivalently aj = a1,jei<rj ,x (1) 0 >, j = 1, 2 . . . , with λj ∈ Λ1. Analogously, from the sequence {Plm,1,2(t)}m≥1, we can draw a subsequence {Plm,2,2(t)}m≥1 ⊂ {Plm,1,2(t)}m≥1 convergent to a certain P2(t) =Xj≥1 pj,2ajeiλj t ∈ PR,Λ2 , where Λ2 = {λj ∈ Λ : pj,2 6= 0}∪ Λ1, which is in the same equivalence class as P1,2(t). This implies that there exists x(2) 0,2, . . .) ∈ Rm2 such that 0 = (x(2) 0,1, x(2) aj = a1,jei<rj ,x (2) 0 >, j = 1, 2 . . . , with λj ∈ Λ2, where m2 is the number of elements of any basis for Λ2. In general, for each k = 2, 3, . . ., we can extract a subsequence {Plm,k,k(t)}m≥1 ⊂ {Plm,k−1,k(t)}m≥1 convergent to a certain Pk(t) =Xj≥1 pj,kajeiλj t ∈ PR,Λk , where Λk = {λj ∈ Λ : pj,k 6= 0} ∪ Λk−1, which is in the same equivalence 0 = (x(k) class as P1,k(t) and hence there exists x(k) 0,2, . . .) ∈ Rmk (mk is the number of elements of any basis for Λk) such that 0,1, x(k) aj = a1,jei<rj ,x (k) 0 >, j = 1, 2 . . . , with λj ∈ Λk. (13) So we get by induction a sequence {Pk(t)}k≥1 of trigonometric polynomials which converges formally to the series ajeiλj t, λj ∈ Λ, Xj≥1 (14) 12 J.M. Sepulcre, T. Vidal and, since (13) is satisfied for any k = 1, 2, . . ., we can construct, for each integer value n ≥ 1, a vector x0,n ∈ R♯GΛ such that aj = a1,jei<rj ,x0,n>, j = 1, 2 . . . , n with λj ∈ Λ. Hence the series (14) is equivalent toPj≥1 a1,jeiλj t and, by Lemma 1, it is the Fourier series associated with an almost periodic function h(t) ∈ B2(R, C) such that h ∼f1. Consequently, {Pk(t)}k≥1 converges with respect to the topology of B2(R, C) to h(t) ∈ G and we can extract a subsequence of {fl(t)}l≥1 which also converges in B2(R, C) to h(t). ∗ As a consequence of Proposition 4, in the topology of B2(R, C), we next show that the family of translates of a function f ∈ FB2,Λ is closed on its equivalence class of FB2,Λ/ Corollary 4 Let Λ be a set of exponents and f ∈ FB2,Λ. Thus the limit points of the set of functions Tf = {fτ (t) := f (t + τ ) : τ ∈ R} are functions which are equivalent to f . ∼. ∗ Proof Since it is plain that the functions included in Tf = {fτ (p) := f (t + τ ) : τ ∈ R} are in the same equivalence class of f (see in [8, Section 4.2] the Fourier series of the translates of a function in the Besicovitch spaces), the result follows easily from Proposition 4. Now Corollary 4 can be improved with the following result. Indeed, we next prove that, fixed a function f ∈ FB2,Λ, the limit points of the set of the translates Tf = {f (t + τ ) : τ ∈ R} of f are precisely the almost periodic functions which are equivalent to f . ∗ ∼ Theorem 2 Let Λ be a set of exponents, G an equivalence class in FB2,Λ/ and f ∈ G. Thus the set of functions Tf = {fτ (t) := f (t + τ ) : τ ∈ R} is dense in G. Proof Let f (t) be a function in the class FB2,Λ. We know by Corollary 4 that the limit points of the set of functions Tf = {fτ (t) := f (t + τ ) : τ ∈ R} are functions in B2(R, C) which are equivalent to f . We next demonstrate that any function h(t) which is equivalent to f (t) is also a limit point of Tf . If ♯Λ < ∞, given εn = 1 n , n ∈ N, Corollary 2 assures the existence of an increasing sequence {τn}n≥1 of positive real numbers such that any n ∈ N verifies f (t + τn) − h(t)2 < εn ∀t ∈ R. Hence M (fτn(t)− h(t)2) → 0 as n goes to ∞ (see (2) for the definition of the mean value M (f )), and the result holds for the case ♯Λ < ∞. Consider ♯Λ = ∞ and let Pj≥1 ajeiλj t and Pj≥1 bjeiλj t be the Fourier series of f ∈ FB2,Λ and ∼f , respectively. Take ε1 = Pj>1 aj2 > 0, then Theorem 1 assures the h existence of τ1 > 0 such that ∗ (cid:12)(cid:12)(cid:12) a1eiλ1(t+τ1) − b1eiλ1t(cid:12)(cid:12)(cid:12) < √ε1 ∀t ∈ R, Bohr's equivalence relation in the space of Besicovitch almost periodic functions 13 which implies < ε1. (15) 2 (cid:12)(cid:12)a1eiλ1τ1 − b1(cid:12)(cid:12) Thus, from (15) and aj = bj for any j ≥ 1 (Corollary 1), we have that Xj≥1 (aj + bj)2 = ajeiλ1τ1 − bj2 < ε1 +Xj>1 ajeiλj τ1 − bj2 ≤ ε1 +Xj>1 Consequently, aj2 = 5ε1. ε1 + 4Xj>1 M (fτ1(t) − h(t)2) < 5ε1. Similarly, take ε2 = Pj>2 aj2 > 0, then Theorem 1 assures the existence of τ2 > τ1 such that 2 which implies < √ε2, Xj=1(cid:12)(cid:12)(cid:12) ajeiλj (t+τ2) − bjeiλj t(cid:12)(cid:12)(cid:12)   Xj=1(cid:12)(cid:12)ajeiλj τ2 − bj(cid:12)(cid:12)   2 2 < ε2. (16) Therefore, from (16) and aj = bj for any j ≥ 1, we have Xj≥1 ajeiλ1τ2 − bj2 = a1eiλ1τ2 − b12 + a2eiλ1τ2 − b22 +Xj>2 ajeiλj τ2 − bj2 ≤ (a1eiλ1τ2 − b1 + a2eiλ1τ2 − b2)2 +Xj>2 (aj + bj)2 = ε2 + 4Xj>2 M (fτ2(t) − h(t)2) < 5ε2. ≤ ε2 +Xj>2 ajeiλj τ2 − bj2 ≤ aj2 = 5ε2. Consequently, In general, by repeating this process, we can construct an increasing sequence {τn}n≥1 such that each τn satisfies that which implies n Xj=1(cid:12)(cid:12)(cid:12) ajeiλj (t+τn) − bjeiλj t(cid:12)(cid:12)(cid:12)   Xj=1(cid:12)(cid:12)ajeiλj τn − bj(cid:12)(cid:12)   n 2 < √εn, < εn. (17) 14 J.M. Sepulcre, T. Vidal with εn =Pj>n aj2. Thus, from (17) we have ajeiλ1τn − bj2 = ajeiλj τn − bj2 ≤ n M (fτn(t) − h(t)2) =Xj≥1 ajeiλj τn − bj2 +Xj>n Xj=1 ajeiλj τn − bj  2 +Xj>n n  Xj=1  ≤ εn +Xj>n ajeiλj τn − bj2 ≤ (aj + bj)2 = εn + 4Xj>n aj2 = 5εn. Note that Pj≥1 aj2 < ∞, then Pj>n aj2 tends to 0 when n goes to ∞. Consequently, the sequence of functions {f (t + τn)}n≥1 converges in B2(R, C) to h(t), and the result holds. Corollary 5 Let f ∈ B2(R, C) and f1 ∼f . There exists an increasing un- bounded sequence {τn}n≥1 of positive numbers such that the sequence of func- tions {f (t + τn)}n≥1 converges in B2(R, C) to f1(t). In fact, given ε > 0 there exists a satisfactorily uniform set of positive numbers τ such that ∗ M (f (t + τ ) − f1(t)2) < ε. Proof Let f ∈ B2(R, C), then f ∈ FB2,Λ for some set Λ of exponents. Let G be the equivalence class in FB2,Λ/ ∼f . Thus, by Theorem 2 (see also its proof), there exists an increasing unbounded sequence {τn}n≥1 of positive numbers such that the sequence of functions {f (t+τn)}n≥1 converges in B2(R, C) to f1(t). Equivalently, given ε > 0 there exists n0 ∈ N such that ∼ so that f ∈ G and let f1 ∗ ∗ M (f (t + δn) − f1(t)2) < ε/2 ∀n ≥ n0. Moreover, since f (t) is almost periodic in the sense of Besicovitch, there exist a set S = {τk} ⊂ R and l = l(ε) > 0 such that the ratio of the maximum number of elements of S included in an interval (a, a + l) to the minimum number is less than 2 and satisfy M (f (t + τk) − f (t)2) < ε/2. Hence any τk satisfies M (f (t + δn + τk) − f1(t)2) ≤ M (f (t + δn + τk) − f (t + δn)2)+ + M (f (t + δn) − f1(t)2) < ε ∀n ≥ n0, which proves the result. Bohr's equivalence relation in the space of Besicovitch almost periodic functions 15 It is known that the almost periodic functions in the Besicovitch spaces Bp(R, C), 1 ≤ p < ∞, satisfy the Bochner-type property consisting of the relative compactness of the set {f (t + τ )}, τ ∈ R, associated with an arbitrary function f ∈ Bp(R, C) (see [1, Theorem 5.34] or [8, Section 3.4]). As an im- portant consequence of Theorem 2, we next refine this property for the case of B2(R, C) in the sense that we show that the condition of almost periodic- ity, in the sense of Besicovitch, of a function f (t) implies that every sequence {f (t + τn)}, τn ∈ R, of translates of f has a subsequence that converges with the topology of B2(R, C) to a function which is equivalent to f . Corollary 6 If f ∈ B2(R, C), then the compact closure of its set of translates coincides with its equivalence class. ∗ Proof First of all, we recall that any function f ∈ B(R, C) has an associated Fourier series. Let f ∈ B2(R, C), then f ∈ FB2,Λ for some set Λ of exponents. Now, let G be the equivalence class in FB2,Λ/ ∼ so that f ∈ G. By Theorem 2, all the limit points of the translates of f are exponential sums which are included in G and, in fact, the compact closure of the set of the translates of f coincides with G. Remark 1 Given Λ = {λ1, λ2, . . . , λj , . . .} a set of exponents, consider A1(p) and A2(p) two exponential sums in the class SΛ, say A1(p) =Pj≥1 ajeλj p and A2(p) = Pj≥1 bjeλj p. Let V be the Q-vector space generated by Λ. We will say that A1 is B-equivalent to A2 if there exists a Q-linear map ψn : V → R such that bj = ajeiψn(λj ), j = 1, 2, . . . . It is easy to prove that, fixed a basis GΛ for Λ, A1 is B-equivalent to A2 if and only if there exists x0 = (x0,1, x0,2, . . . , x0,k, . . .) ∈ R♯GΛ such that bj = aje<rj,x0>i for every j ≥ 1, where the rj's are the vectors of rational components verifying (4). From this and Proposition 1, it is worth noting that Definition 2 and defi- nition of B-equivalence are equivalent in the case that it is possible to obtain an integral basis for the set of exponents Λ. Consequently, all the results of this paper which can be formulated in terms of an integral basis are also valid under the B-equivalence. (in particular, those related to the finite exponential sums in Section 3). References 1. J. Andres, A.M. Bersani, R.F. Grande, Rendiconti di Matematica, Serie VII, 26 (2006), 121-188. 2. T.M. Apostol, Modular functions and Dirichlet series in number theory, Springer-Verlag, New York, 1990. 3. A.S. Besicovitch, Almost periodic functions, Dover, New York, 1954. 4. S. Bochner, A new approach to almost periodicity, Proc. Nat. Acad. Sci., 48 (1962), 2039-2043. 5. H. Bohr, Almost periodic functions, Chelsea, New York, 1951. 16 J.M. Sepulcre, T. Vidal 6. H. Bohr, Contribution to the theory of almost periodic functions, Det Kgl. danske Viden- skabernes Selskab. Matematisk-fisiske meddelelser. Bd. XX. Nr. 18, Copenhague, 1943. 7. H. Bohr, E. Foelner, On some types of Functional Spaces, Acta Math., 76 (1945), 31-155. 8. C. Corduneanu, Almost Periodic Oscillations and Waves, Springer, New York, 2009. 9. G.H. Hardy, E.M. Wright, An Introduction to the Theory of Numbers, Oxford Science, Oxford, 1979. 10. B. Jessen, Some aspects of the theory of almost periodic functions, in Proc. Internat. Congress Mathematicians Amsterdam, 1954, Vol. 1, North-Holland, 1954, pp. 304–351. 11. J.M. Sepulcre, T. Vidal, Almost Periodic Functions in terms of Bohr's Equivalence Relation, Ramanujan J., DOI: 10.1007/s11139-017-9950-1, 2017. Corrigendum sent to the journal. See also arXiv: 1801.08035 [math.CV].
1711.04470	1	1711	2017-11-13T08:47:17	New variations of power increasing sequences	[ "math.FA" ]	The aim of this paper is to generalize a main theorem concerning weighted mean summability to absolute matrix summability which plays a vital role in summability theory and applications to the other sciences by using quasi-$f$-power sequences.	math.FA	math	New variations of power increasing sequences S¸ebnem Yıldız Abstract. The aim of this paper is to generalize a main theorem concerning weighted mean summability to absolute matrix summability which plays a vital role in summability theory and applications to the other sciences by using quasi-f -power sequences. 1. Introduction Definition 1.1. [1] A positive sequence (bn) is said to be an almost increas- ing sequence if there exists a positive increasing sequence (cn) and two positive constants M and N such that M cn 6 bn 6 N cn. Definition 1.2. [19] A positive sequence X = (Xn) is said to be quasi-f -power increasing sequence if there exists a constant K = K(X, f ) > 1 such that KfnXn > fmXm for all n > m > 1, where f = {fn(σ, β)} =(cid:8)nσ(logn)β, β > 0, 0 < σ < 1(cid:9). Definition 1.3. The sequence (λn) is said to be of bounded variation, denoted ∞ by (λn) ∈ BV, if ∆λn < ∞. If we take β = 0, then we have a quasi-σ-power Pn=1 increasing sequence. Every almost increasing sequence is a quasi-σ-power increasing sequence for any non-negative σ, but the converse is not true for σ > 0 (see [13]). For any sequence (λn) we write that ∆2λn = ∆λn − ∆λn+1 and ∆λn = λn − λn+1. n and tα n we denote the nth Ces`aro means of order α, with α > −1, of the sequence (sn) and (nan), respectively, that is (see [8]) Let P an be a given infinite series with the partial sums (sn). By uα (1.1) where (1.2) uα n = 1 Aα n n Xv=0 Aα−1 n−vsv and tα n = 1 Aα n n Xv=0 Aα−1 n−vvav, Aα n = (α + 1)(α + 2)...(α + n) n! = O(nα), Aα −n = 0 for n > 0. 2010 Mathematics Subject Classification. 26D15; 42A24; 40F05; 40G99. Key words and phrases. Summability factors, absolute matrix summability, Fourier series, infinite series, Holder inequality, Minkowski inequality. 1 2 YILDIZ k > 1, if Definition 1.4. [10],[12] The series P an is said to be summable C, αk, (1.3) nk−1uα n − uα n−1k = tα nk < ∞. 1 n ∞ Xn=1 ∞ Xn=1 Xv=0 If we take α = 1, then C, αk summability reduces to C, 1k summability. Let (pn) be a sequence of positive real numbers such that n (1.4) Pn = pv → ∞ as n → ∞, (P−i = p−i = 0, i > 1). The sequence-to-sequence transformation (1.5) tn = 1 Pn pvsv n Xv=0 defines the sequence (tn) of the Riesz mean or simply the ( ¯N , pn) mean of the sequence (sn) generated by the sequence of coefficients (pn) (see [11]). Definition 1.5. [2] The series P an is said to be summable (cid:12)(cid:12) ∞ ¯N , pn(cid:12)(cid:12)k, k > 1, if (1.6) Xn=1(cid:18) Pn pn(cid:19)k−1 tn − tn−1 k< ∞. summability is the same as C, 1k (resp. ¯N , pn ) summability. In the special case when pn = 1 for all values of n (resp. k = 1), (cid:12)(cid:12) The following theorems are known dealing with the (cid:12)(cid:12) ¯N , pn(cid:12)(cid:12)k ¯N , pn(cid:12)(cid:12)k summability fac- Theorem 2.1. [14] Let (Xn) be a almost increasing sequence. If the sequences 2. The Known Results tors of infinite series. (Xn), (λn), and (pn) satisfy the conditions m Xn=1 (2.1) (2.2) (2.3) (2.4) (2.5) λmXm = O(1) as m → ∞, nXn∆2λn = O(1) as m → ∞, Pn n = O(Pm) m Xn=1 m pn Pn m Xn=1 Xn=1 tnk = O(Xm) as m → ∞, tnk n = O(Xm) as m → ∞, then the series P anλn is summable (cid:12)(cid:12) ¯N , pn(cid:12)(cid:12)k, k > 1. NEW VARIATIONS OF POWER INCREASING SEQUENCES 3 Theorem 2.2. [6] Let (Xn) be a quasi-σ-power increasing sequence. If the sequences (Xn), (λn) and (pn) satisfy the conditions (2.1)-(2.3), and (2.6) (2.7) pn Pn tnk X k−1 n = O(Xm) as m → ∞, tnk nX k−1 n = O(Xm) as m → ∞, m m Xn=1 Xn=1 then the series P anλn is summable (cid:12)(cid:12) ¯N , pn(cid:12)(cid:12)k, k > 1. creasing sequence instead of a quasi-σ-power increasing sequence. Later on, Bor has proved the following theorem by taking quasi-f-power in- Theorem 2.3. [7] Let (Xn) be a quasi-f -power increasing sequence. If the sequences (Xn), (λn) and (pn) satisfy all the conditions of Theorem 2.2, then the series P anλn is summable (cid:12)(cid:12) ¯N , pn(cid:12)(cid:12)k, k > 1. Let A = (anv) be a normal matrix, i.e., a lower triangular matrix of nonzero diagonal entries. Then A defines the sequence-to-sequence transformation, mapping the sequence s = (sn) to As = (An(s)), where (2.8) An(s) = anvsv, n = 0, 1, ... n Xv=0 if Definition 2.1. [18] The series P an is said to be summable A, pnk, k > 1, (2.9) where (2.10) ∞ Xn=1(cid:18) Pn pn(cid:19)k−1 k < ∞, (cid:12)(cid:12) ¯∆An(s)(cid:12)(cid:12) ¯∆An(s) = An(s) − An−1(s). If we take pn = 1 for all values of n, then we have Ak summability (see [20]). And also if we take anv = pv Pn we take anv = pv Pn to C, 1k summability (see [10]). , then we have (cid:12)(cid:12) ¯N , pn(cid:12)(cid:12)k summability. Furthermore, if and pn = 1 for all values of n, then A, pnk summability reduces 3. The Main Results The Fourier series play an important role in many areas of applied mathematics and mechanics. Recently some papers have been done concerning absolute matrix summability of infinite series and Fourier series (see [3]-[5], [15]-[17], [21]-[23]). The aim of this paper is to generalize Theorem 2.3 for A, pnk summability method for these series by taking quasi-f-power increasing sequence instead of a quasi-σ- power increasing sequence. 4 YILDIZ Given a normal matrix A = (anv), we associate two lower semimatrices ¯A = (¯anv) and A = (anv) as follows: (3.1) and (3.2) ¯anv = n Xi=v ani, n, v = 0, 1, ... a00 = ¯a00 = a00, anv = ¯anv − ¯an−1,v, n = 1, 2, ... It may be noted that ¯A and A are the well-known matrices of series-to-sequence and series-to-series transformations, respectively. Then, we have (3.3) and (3.4) An(s) = anvsv = n Xv=0 ¯anvav n Xv=0 ¯∆An(s) = anvav. n Xv=0 Using this notation we have the following theorem. Theorem 3.1. Let (Xn) be a quasi-f -power increasing sequence. Let k > 1 and A = (anv) be a positive normal matrix such that (3.5) (3.6) (3.7) (3.8) ano = 1, n = 0, 1, ..., an−1,v > anv, for n > v + 1, Pn(cid:19) ann = O(cid:18) pn an,v+1 = O(ann). 1 v n−1 Xv=1 If the sequences (Xn), (λn) and (pn) satisfy all the conditions of Theorem 2.3, then the series P anλn is summable A, pnk, k > 1. It may be remarked that if we take A = ( ¯N , pn), the conditions (3.5)-(3.7) are satisfied automatically and the condition (3.8) is satisfied by the condition (2.3). We need the following lemmas for the proof of our theorem. Lemma 3.1. [3] Under the conditions of Theorem 2.1 we have that (3.9) (3.10) nXn∆λn = O(1) as n → ∞, Xn∆λn < ∞. ∞ Xn=1 NEW VARIATIONS OF POWER INCREASING SEQUENCES 5 Proof of Theorem 3.1 Let Xn be a be a quasi-f -power increasing sequence and (In) denotes the A- transform of the series P∞ n=1 anλn. Then, we have ¯∆In = n Xv=1 anvavλv. Applying Abel's transformation to this sum, we have that anvavλv v v = n−1 Xv=1 ∆( anvλv v ) v Xr=1 rar + annλn n rar n Xr=1 ∆( anvλv v )(v + 1)tv + annλn n + 1 n tn ¯∆In = = = n n−1 Xv=1 Xv=1 Xv=1 n−1 ¯∆anvλvtv v + 1 v + n−1 Xv=1 an,v+1∆λvtv v + 1 v + n−1 Xv=1 an,v+1λv+1 tv v + annλntn n + 1 n = In,1 + In,2 + In,3 + In,4. To complete the proof of Theorem 3.1, by Minkowski's inequality, it is sufficient to show that (3.11) ∞ Xn=1(cid:18) Pn pn(cid:19)k−1 In,r k< ∞, for r = 1, 2, 3, 4. First, by applying Holder's inequality with indices k and k′, where k > 1 and 1 k + 1 k′ = 1, we have that m+1 Xn=2(cid:18) Pn = O(1) m+1 v + 1 pn(cid:19)k−1 Xn=2(cid:18) Pn In,1 k6 ¯∆anv(cid:12)(cid:12) λvtv)k pn(cid:19)k−1(n−1 Xn=2(cid:18) Pn Xv=1 (cid:12)(cid:12) ¯∆anv(cid:12)(cid:12))k−1 ¯∆anv(cid:12)(cid:12) λvktvk ×(n−1 pn(cid:19)k−1 n−1 Xv=1(cid:12)(cid:12) Xv=1(cid:12)(cid:12) m+1 v , using ∆anv = anv − an,v+1 = ¯anv − ¯an−1,v − ¯an,v+1 + ¯an−1,v+1 = anv − an−1,v, and from (3.5) and (3.6) we have ¯∆anv = n−1 Xv=1 = n−1 Xv=1 Xv=0 n Xv=1 Xv=0 n−1 n−1 anv − an−1,v = (an−1,v − anv) an−1,v − an−1,0 − anv + an0 + ann = 1 − an−1,0 − 1 + an0 + ann 6 ann, 6 YILDIZ n=v+1 ¯∆anv 6 avv we have, pn(cid:19)k−1 ak−1 nn (n−1 Xv=1 ¯∆anvλvktvk) In,1 k= O(1) λvk−1λvtvk m+1 Xn=2(cid:18) Pn Xn=v+1 m+1 ¯∆anv m−1 λvtvkavv = O(1) ∆λv Xv=1 arr trk X k−1 r v Xr=1 + O(1)λm avv tvk X k−1 v m Xv=1 ∆λvXv + O(1)λmXm = O(1) as m → ∞, by virtue of the hypotheses of Theorem 3.1 and Lemma 3.1. Also, we have that m m+1 = O(1) and using Pm+1 pn(cid:19)k−1 Xn=2(cid:18) Pn Xv=1 Xv=1 Xv=1 = O(1) = O(1) m−1 1 m v X k−1 m+1 Xn=2(cid:18) Pn = O(1) = O(1) = O(1) m+1 m+1 m+1 pn(cid:19)k−1 Xn=2(cid:18) Pn Xn=2(cid:18) Pn Xn=2(cid:18) Pn Xv=1 Xv=1 Xv=1 Xv=1 m−1 m−1 m−1 m an,v+1∆λvtv)k an,v+1∆λvtv v Xv v + 1 pn(cid:19)k−1(n−1 Xv=1 Xv)k tvk) ×(n−1 Xv=1 1 X k v an,v+1∆λvXv an,v+1∆λvXv m+1 In,2 k6 Xn=2(cid:18) Pn pn(cid:19)k−1(n−1 Xv=1 pn(cid:19)k−1(n−1 Xv=1 nn (n−1 pn(cid:19)k−1 Xv=1 Xn=v+1 X k−1 ak−1 tvk 1 v m+1 1 v 1 X k v an,v+1∆λvXv)k−1 ∆λvXv)k−1 tvk) ×(m−1 Xv=1 Xv=1 Xr=1 trk rX k−1 vX k−1 v∆λv tvk 1 m m v r = O(1) v∆λv an,v+1 = O(1) = O(1) ∆(v∆λv) trk rX k−1 r + O(1)m∆λm v Xr=1 = O(1) ∆(v∆λv)Xv + O(1)m∆λmXm = O(1) vXv∆2λv + O(1) = O(1) as m → ∞, Xv∆λv + O(1)m∆λmXm m−1 Xv=1 Xn=1(cid:18) Pn pn(cid:19)k−1 In,4k = O(1) = O(1) pn(cid:19)k−1 1 X k−1 n Xn=1(cid:18) Pn Xn=1 ann m ak nnλnktnk = O(1) ak−1 nn annλnk−1λntnk Xn=1(cid:18) Pn pn(cid:19)k−1 λntnk = O(1) as m → ∞, NEW VARIATIONS OF POWER INCREASING SEQUENCES 7 by virtue of the hypotheses of Theorem 3.1 and Lemma 3.1. Furthermore, as in In,1, we have m+1 Xn=2(cid:18) Pn = O(1) = O(1) = O(1) an,v+1λv+1 m+1 m+1 In,3 k6 pn(cid:19)k−1 Xn=2(cid:18) Pn Xn=2(cid:18) Pn Xv=1 tvk pn(cid:19)k−1(n−1 Xn=2(cid:18) Pn Xv=1 v ) ×(n−1 pn(cid:19)k−1(n−1 an,v+1λv+1k tvk Xv=1 Xv=1 pn(cid:19)k−1 Xv=1 Xn=v+1 λv+1λv+1k−1 tvk v an,v+1 = O(1) Xv=1 X k−1 λv+1 ak−1 nn m+1 m+1 n−1 1 1 v m m v X k−1 v 1 v tv v )k an,v+1)k−1 Xv=1 m λv+1 tvk v an,v+1 = O(1) = O(1) as m → ∞, tvk 1 v X k−1 v λv+1 m+1 Xn=v+1 an,v+1 by virtue of the hypotheses of Theorem 3.1 and Lemma 3.1. Again, as in In,1, we have that m m m by virtue of hypotheses of the Theorem 3.1 and Lemma 3.1. This completes the proof of Theorem 3.1. 4. An application of absolute matrix summability to Fourier series Let f be a periodic function with period 2π and integrable (L) over (−π, π). Without any loss of generality the constant term in the Fourier series of f can be taken to be zero, so that (4.1) where a0 = We write 1 π Z π −π ∞ f (t) ∼ (ancosnt + bnsinnt) = Xn=1 Cn(t). ∞ Xn=1 f (t)dt, an = 1 π Z π −π f (t)cos(nt)dt, bn = 1 π Z π −π f (t)sin(nt)dt. (4.2) (4.3) φα(t) = φ(t) = 1 2 {f (x + t) + f (x − t)} , (t − u)α−1φ(u) du, (α > 0). α tα Z t 0 It is well known that if φ(t) ∈ BV(0, π), then tn(x) = O(1), where tn(x) is the (C, 1) mean of the sequence (nCn(x)) (see [9]). 8 YILDIZ Using this fact, Bor has obtained the following main result dealing with the trigono- metric Fourier series. Theorem 4.1. [6] Let (Xn) be a quasi-σ-power increasing sequence. If φ1(t) ∈ BV(0, π), and the sequences (pn), (λn), and (Xn) satisfy the conditions of Theorem Theorem 4.2. [7] Let (Xn) be a quasi-f -power increasing sequence. If φ1(t) ∈ BV(0, π), and the sequences (pn), (λn), and (Xn) satisfy the conditions of Theorem 2.3, then the series P Cn(x)λn is summable ¯N , pnk, k > 1. 2.3, then the series P Cn(x)λn is summable ¯N , pnk, k > 1. We now apply the above theorems to the weighted mean in which A = (anv) when 0 6 v 6 n, where Pn = p0 + p1 + ... + pn. Therefore, is defined as anv = pv Pn it is well known that ¯anv = Pn − Pv−1 Pn and an,v+1 = pnPv PnPn−1 . So, one can easily verify that the conditions of Theorem 3.1 reduce to those of Theorem 2.3 and also we can obtain new results dealing with absolute matrix summability of Fourier series in the following manner. Theorem 4.3. Let A be a positive normal matrix satisfying the conditions of Theorem 3.1. Let (Xn) be a quasi-σ-power increasing. If φ1(t) ∈ BV(0, π), and the sequences (pn), (λn), and (Xn) satisfy the conditions of Theorem 3.1, then the series P Cn(x)λn is summable A, pnk, k > 1. Theorem 4.4. Let A be a positive normal matrix satisfying the conditions of Theorem 3.1. Let (Xn) be a quasi-f -power increasing sequence. If φ1(t) ∈ BV(0, π), and the sequences (pn), (λn), and (Xn) satisfy the conditions of Theorem 3.1, then the series P Cn(x)λn is summable A, pnk, k > 1. 5. APPLICATIONS n−v/Aα We may now ask whether there are some examples other than weighted mean methods of matrices A that satisfy the hypotheses of Theorem 3.1. For example, apply Theorem 3.1 to the Ces`aro method of order α with 0 < α 6 1 in which A is given by anv = Aα−1 n, and by applying Theorem 3.1, Theorem 4.3 and Theorem 4.4 to weighted mean so, the following results can be easily verified. 1. If we take anv = pv Pn have Theorem 2.3, Theorem 4.1 and Theorem 4.2. 2. If we take β = 0 and anv = pv Pn Theorem 2.2 and Theorem 4.1. 3. If we take pn = 1 for all values of n in Theorem 3.1, Theorem 4.3 and Theorem 4.4, then we have a new result dealing with Ak summability. 4. If we take anv = pv Pn and Theorem 4.4, then we have a new result concerning C, 1k summability. and pn = 1 for all values of n in Theorem 3.1, Theorem 4.3 in Theorem 3.1, Theorem 4.3 and Theorem 4.4, then we in Theorem 3.1 and Theorem 4.4, then we have NEW VARIATIONS OF POWER INCREASING SEQUENCES 9 References [1] Bari, N. K., Steckin, S.B, Best approximation and differential properties of two conjugate functions. Trudy. Moskov. Mat. Obsc. (in Russian) 5, 483-522 (1956) [2] Bor, H., On two summability methods. Math. Proc. Cambridge Philos Soc. 97, 147-149 (1985) [3] Bor, H., Quasi-monotone and almost increasing sequences and their new appli- cations. Abstr. Appl. Anal. Art. ID 793548, 6 PP.(2012) [4] Bor, H., On absolute weighted mean summability of infinite series and Fourier series. Filomat 30, 2803-2807 (2016) [5] Bor, H., Some new results on absolute Riesz summablity of infinite series and Fourier series. Positivity 20, 3 599-605 (2016) [6] Bor, H., An Application of power increasing sequences to infinite series and Fourier series. Filomat 31,6 1543-1547 (2017) [7] Bor, H., Absolute weighted arithmetic mean summability factors of infinite series and trigonometric Fourier series. Filomat 31, 15 4963-4968 (2017) [8] Ces`aro, E., Sur la multiplication des s`eries. Bull. Sci. Math. 14, 114-120 (1890) [9] Chen, K. K., Functions of bounded variation and the cesaro means of Fourier series. Acad. Sin. Sci. Record 1 283-289 (1945) [10] Flett, T. M., On an extension of absolute summability and some theorems of Littlewood and Paley. Proc. Lond. Math. Soc. 7, 113-141 (1957) [11] Hardy, G. H., Divergent Series. Clarendon Press, Oxford (1949) [12] Kogbetliantz, E., Sur l`es series absolument sommables par la methode des moyennes arithmetiques. Bull. Sci. Math. 49, 234-256 (1925) [13] Leindler, L., A new application of quasi power increasing sequences. Publ. Math. Debrecen 58, 791-796 (2001) [14] Mazhar, S. M., Absolute summability factors of infinite series. Kyungpook Math. J. 39, 67-73 (1999) [15] Ozarslan, H. S., Yıldız, S¸., A new study on the absolute summability factors of Fourier series. J. Math. Anal. 7 31-36 (2016) 10 YILDIZ [16] Ozarslan, H. S., Yıldız, S¸., On the local property of summability of factored Fourier series, Int. J. Pure Math. 3, 1-5, (2016). [17] Sarıgol, M. A., On the local properties of factored Fourier series. Appl. Math. Comp. 216 3386-3390 (2010) [18] Sulaiman, W. T., Inclusion theorems for absolute matrix summability methods of an infinite series. IV. Indian J. Pure Appl. Math. 34, 11 1547-1557 (2003) [19] Sulaiman, W. T., Extension on absolute summability factors of infinite series. J. Math. Anal. Appl. 322, 1224-1230 (2006) [20] Tanovic-Miller, N., On strong summability. Glas. Mat. Ser III 14 , (34) 87-97 (1979) [21] Yıldız, S¸., On absolute matrix summability factors of infinite series and Fourier series. GU J. Sci. 30, 1 363-370 (2017) [22] Yıldız, S¸., On Riesz summability factors Fourier series. Trans. A. Razmadze Math. Inst. http://dx.doi.org/10.1016/j.trmi.2017.06.003. (2017) [23] Yıldız, S¸., On Application of Matrix Summability to Fourier Series,Math. Methods Appl. Sci., DOI: 10.1002/mma.4635, (2017) Department of Mathematics, Ahi Evran University, Kırs¸ehir, Turkey E-mail address: [email protected]; [email protected]
1103.1005	4	1103	2011-09-24T17:23:16	On the reproducing kernel of a Pontryagin space of vector valued polynomials	[ "math.FA", "math.AC" ]	We give necessary and sufficient conditions under which the reproducing kernel of a Pontryagin space of $d \times 1$ vector polynomials is determined by a generalized Nevanlinna pair of $d \times d$ matrix polynomials.	math.FA	math	ON THE REPRODUCING KERNEL OF A PONTRYAGIN SPACE OF VECTOR VALUED POLYNOMIALS B. ´CURGUS AND A. DIJKSMA Abstract. We give necessary and sufficient conditions under which the repro- ducing kernel of a Pontryagin space of d × 1 vector polynomials is determined by a generalized Nevanlinna pair of d × d matrix polynomials. 1. Introduction 1.1. By Baire's category theorem, the Pontryagin space B in the title is neces- sarily finite dimensional (see Remark 2.1 below) and hence is a reproducing kernel space. Indeed, if (cid:0)B, [· ,· ]B(cid:1) is an n-dimensional Pontryagin space of d × 1 vec- tor polynomials and if B(z) is a d × n matrix polynomial whose columns Bk(z), k ∈ {1, . . . , n}, form a basis of B, then the reproducing kernel of B is the d × d matrix polynomial in z and w∗ given by K(z, w) = B(z)G−1B(w)∗, z, w ∈ C, where G is the n × n Gram matrix associated with B(z), that is, G = [gjk]n j,k=1, gjk = [Bk, Bj]B, j, k ∈ {1, . . . , n} (see [3, Example 2.1.8] and the remark following it). The reproducing kernel of a reproducing kernel space is unique but can often be written in various ways. In this paper we give necessary and sufficient conditions under which K(z, w) above is a polynomial Nevanlinna kernel. This means that it can be written in the form K(z, w) = KM,N (z, w) := M (z)N (w)∗ − N (z)M (w)∗ where M (z) and N (z) are d × d matrix polynomials such that z − w∗ , z, w ∈ C, z 6= w∗, for all z ∈ C and M (z)N (z∗)∗ − N (z)M (z∗)∗ = 0 rank(cid:2)M (z) N (z)(cid:3) = d (1.1) If, in addition, the equality in (1.1) holds for all z ∈ C, then the Nevanlinna kernel KM,N (z, w) is called a full Nevanlinna kernel. The following theorem is the main result in this paper. It is proved in Section 4. for at least one z ∈ C. Date: June 7, 2018. 2000 Mathematics Subject Classification. 46C20, 46E22, 47A06. Key words and phrases. Polynomials, generalized Nevanlinna pair, Pontryagin space, repro- ducing kernel, Smith normal form, Forney indices, symmetric operator, defect number, self-adjoint extension, Q-function. 1 2 B. ´CURGUS AND A. DIJKSMA Theorem 1.1. Let B be a (finite dimensional) Pontryagin space of d × 1 vec- tor polynomials. Denote by SB the operator of multiplication by the independent variable in B and by Eα the operator of evaluation at a point α ∈ C. Then the re- producing kernel of B is a polynomial Nevanlinna kernel if and only if the following two conditions hold: (A) The operator SB is symmetric in B. (B) For some α ∈ C we have ran(cid:0)SB − α(cid:1) = B ∩ ker Eα. In this case the reproducing Nevanlinna kernel is full if and only if the equality in (B) holds for all α ∈ C. We think Theorem 1.1 is new, possibly even in the positive definite case, that is, the case where the space B is a reproducing kernel Hilbert space of vector polynomials. In that case B in the theorem is a special case of L. de Branges' Hilbert spaces of entire functions. For scalar functions, see [12]; for vector functions, see [13] and [14]. In particular, [14, Theorems 1-3] are closely related to Theorem 1.1. For results on the indefinite scalar case we refer to the series of papers on Pontryagin spaces of entire functions by M. Kaltenback and H. Woracek. More specifically, [26, Theorem 5.3] is closely related to Theorem 1.1 with d = 1, [27, Proposition 2.8] can be used to obtain a scalar version of Theorem 1.2 below, and [26, Lemma 6.4] is linked with Theorem 5.1 in Section 5. The emphasis in this paper is on vector polynomials and an indefinite setting. In the proof of Theorem 1.1 we use the following result which shows that the 1.2. condition (B) in Theorem 1.1 completely determines the structure of B as a linear space. We believe Theorem 1.2 is also new, but closely related to results around [20, Proposition 2.3]. For the proof of Theorem 1.2 we refer to Section 3. Theorem 1.2. Let B be a finite dimensional linear space of d×1 vector polynomials and let α ∈ C. The equality (1.2) holds if and only if there exist nonnegative integers µ1, . . . , µd and a d × d matrix polynomial W (z) with det W (α) 6= 0 such that the space B consists of all vector where pj(z) runs through all scalar polynomials of degree strictly less than µj, j ∈ {1, . . . , d}. The matrix W (z) can be chosen such that polynomials of the form W (z)(cid:2)p1(z)··· pd(z)(cid:3)⊤ ran(cid:0)SB − α(cid:1) = B ∩ ker Eα (cid:8)α ∈ C : det W (α) 6= 0(cid:9) =(cid:8)α ∈ C : ran(cid:0)SB − α(cid:1) = B ∩ ker Eα(cid:9). (1.3) It follows that the dimension of B in Theorem 1.2 is µ1 +··· + µd. If the conditions (A) and (B) of Theorem 1.1 hold, then the numbers µ1, . . . , µd are the Forney indices Nevanlinna kernel of B. Moreover, the defect numbers of SB coincide with the of the block matrix polynomial (cid:2)M (z) N (z)(cid:3) corresponding to the reproducing cardinality of the set (cid:8)j ∈ {1, . . . , d} : µj > 0(cid:9), see Remarks 4.1 and 4.3. This from the block matrix polynomial(cid:2)M (z) N (z)(cid:3). offers a direct way of determining the dimension of the reproducing kernel space B with reproducing Nevanlinna kernel KM,N (z, w) and the defect numbers of SB In the scalar case (d = 1) the space B in the above theorems is analogous to the so-called Szego space, in the Hilbert space setting defined and studied in [34, 35] and in the Pontryagin space setting in [1]. In the literature there are many papers characterizing special forms of the reproducing kernel of a reproducing kernel space. ON THE REPRODUCING KERNEL 3 Of those related to a reproducing kernel Pontryagin space we mention [7, Section 6] and [2]. We refer to the references in these papers for papers dealing with the Hilbert space case. The characterizations in these works are often in terms of a special identity to be satisfied by the difference-quotient operator on the space. In some cases, such as in [2, Theorem 4.1] and [12, Problems 51, Theorem 23] the invertibility of K(z, z) for some values of z plays a role in proving the asserted representation of the kernel K(z, w). We give in Section 6 some examples where detK(z, w) = 0 for all z, w ∈ C, see Example 6.6 and Example 6.7. 1.3. A pair {M (z), N (z)} of d× d matrix functions M (z) and N (z) is called a gen- eralized Nevanlinna pair if the functions are meromorphic on C\ R, the intersection of the domains of holomorphy hol(M ) of M (z) and hol(N ) of N (z) is symmetric with respect to the real axis, (1.4) (1.5) M (z)N (z∗)∗ − N (z)M (z∗)∗ = 0 rank(cid:2)M (z) N (z)(cid:3) = d for at least one for all and the Nevanlinna kernel z ∈ hol(M ) ∩ hol(N ), z ∈ hol(M ) ∩ hol(N ), M (z)N (w)∗−N (z)M (w)∗ (1.6) KM,N (z, w) := , z, w ∈ hol(M ) ∩ hol(N ), z 6= w∗, has a finite number of negative squares. Here, by a finite number of negative squares we mean that the set of numbers of negative eigenvalues counted according to multiplicity of the self-adjoint matrices of the form z − w∗ with j KM,N (zj, zi)xi(cid:3)n (cid:2) x∗ i,j=1 n ∈ N, xi ∈ Cd, zi ∈ hol(M ) ∩ hol(N ), zi 6= z∗ j , i, j ∈ {1, . . . , n} has a maximum. If this maximum is κ, then we say that the pair and the kernel have κ negative squares. If κ = 0 the adjective "generalized" is omitted; in that case the matrix functions are holomorphic at least on C\ R. The number of positive squares is defined in the same way. The pair and kernel are called full if the equality in (1.5) holds for all z ∈ hol(M )∩ hol(N ). If a (generalized) Nevanlinna pair {M (z), N (z)} is such that N (z) = Id, the d × d identity matrix, then it is identified with its first entry M (z) and M (z) is a (generalized) Nevanlinna function. Nevanlinna pairs and generalized Nevanlinna pairs have been used in interpola- tion and moment problems (see [30], [4, 5] and [8]), the description of generalized resolvents (see [28]) and in the theory of boundary value problems with eigenvalue dependent boundary conditions (see [17, 18], [15] and [9]). Theorem 1.1 arose in our study [10] of an eigenvalue problem for an ordinary differential operator in a Hilbert space with boundary conditions which depend polynomially on the eigenvalue pa- rameter. In that paper we linearize the original problem by extending the Hilbert space with a finite dimensional Pontryagin space of d × 1 vector polynomials. This paper concerns the structure of such spaces. 1.4. The Nevanlinna pair in a Nevanlinna kernel is not unique (see the paragraph before Example 6.7) and if {M (z), N (z)} is a pair that determines the kernel, then the polynomial matrix N (z) may be such that detN (z) = 0 for all z ∈ C. In Section 5 we prove that one can always choose the pair so that detN (z) 6≡ 0 and 4 B. ´CURGUS AND A. DIJKSMA the rational generalized Nevanlinna matrix function N (z)−1M (z) is essentially a Q- function of the symmetric operator SB. We show that every self-adjoint extension of SB with nonempty resolvent set gives rise to a reproducing Nevanlinna kernel for the space B. The proof of Theorem 1.1 given in Section 4 is geometric, the proof of the first if statement in Theorem 1.1 given in Section 5 is analytic. The last two examples in Section 6, Example 6.6 and Example 6.7, also serve to show that this analytic proof is constructive. In Section 6 we present three corollaries of Theorem 1.1 and four examples. In Section 2 we fix the notation related to vector and matrix polynomials and we recall the Smith normal form and the Forney indices of a matrix polynomial. Moreover, we prove some lemmas on the structure of a degenerate subspace of a finite dimensional Pontryagin space, the defect numbers of a simple symmetric relation in such a space and on polynomial Hermitian kernels. Although most proofs in this paper are based on methods from linear algebra, in the sequel we assume that the reader is familiar with (i) Pontryagin spaces and (multi-valued) operators on such spaces such as symmetric and self-adjoint relations (as in [24], [19] and [11]), (ii) generalized Nevanlinna matrix functions (as in [30, 31]) and (iii) reproducing kernel Pontryagin spaces (as in [6, Chapter 1] and [3, Chapter 7]). The notion of a Q-function of a simple symmetric operator in a Pontryagin space is recalled in Section 5. Acknowledgements. The authors thank Prof. Marius van der Put for many discussions about Theorem 3.4 and another proof of it and Prof. Daniel Alpay for acquainting them with his unpublished note [1] which among other things lead to Remark 2.1 and pointing out the references [34, 35]. We also thank the referee for useful comments. 2. Notation and basic objects 2.1. The symbols N, R, and C denote the sets of positive integers, real numbers and complex numbers. For d ∈ N the vector space of all d × 1 vectors is written as Cd and Id stands for the d × d identity matrix. The k-th row of Id will be denoted by ed,k. For k ∈ {1, . . . , n} the subspace of Cd spanned by ed,1, . . . , ed,k will be called a top coordinate subspace of Cd; it will be denoted by Cd k. The corresponding d× d projection matrix is denoted by Pd,k. We consider Cd 0 = {0} a top coordinate subspace spanned by the empty set. By Cd[z] we denote the vector space over C of all polynomials with coefficients in Cd. The space Cd is identified with the subspace of all constant polynomials in Cd[z]. If d = 1 we simply write C[z] and C. For f ∈ Cd[z] \ {0} with Matrix polynomials are written as B(z), M (z), N (z), . . ., that is, with their argu- ment z; we use the bold face P(z), S(z), . . ., for d × 2d matrix polynomials. Vector polynomials are sometimes written with and sometimes without their argument. The Fraktur alphabet A, B, C, H, . . . is used to denote vector subspaces of Cd[z]. One exception to this is that L will be used for a subspace of C2d[z]. An inner and for the zero polynomial 0 we define f (z) = a0 + a1z + ··· + anzn deg f = max(cid:8)k ∈ {0, . . . , n} : ak 6= 0(cid:9) and deg 0 = −∞. ON THE REPRODUCING KERNEL 5 product on B is denoted by [· ,· ]B. In a vector space, the symbol ⊕ denotes the direct sum of subspaces. Remark 2.1. A Banach space with a countable Hamel basis is separable and hence, by [32], it is finite dimensional. Since (cid:8)zn : n ∈ {0} ∪ N(cid:9) is a countable Hamel basis of C[z], the space Cd[z] and all its subspaces also have countable Hamel bases. Therefore any Pontryagin subspace of Cd[z] is finite dimensional. In spite of this fact, to emphasize the finite dimensionality, we continue to speak of finite dimensional Pontryagin subspaces of Cd[z]. We introduce some special subspaces of Cd[z]. Let n ∈ {0} ∪ N. The symbol Cd[z]<n stands for the set of all f ∈ Cd[z] such that deg f < n. In particular, Cd[z]<1 = Cd and Cd[z]<0 = {0}. A subspace C of Cd[z] is called canonical if there exist nonnegative integers µk, k ∈ {1, . . . , d}, such that C = dMk=1(cid:0)C[z]<µk(cid:1)ed,k =(cid:8)[p1(z)··· pd(z)]⊤ : pk(z) ∈ C[z], degpk < µk, k ∈ {1, . . . , d}(cid:9). The numbers µ1, . . . , µd will be called the degrees of C. Without loss of generality we can assume that they are ordered: µ1 ≥ ··· ≥ µd ≥ 0. Then a canonical subspace is uniquely determined by its degrees. Clearly, the dimension of C is the sum of its degrees. Next we introduce some useful operators on Cd[z]. By Pd,k, k ∈ {1, . . . , d}, we denote the natural extension of Pd,k to Cd[z], by S : Cd[z] → Cd[z] the operator of multiplication by the independent variable, that is, (Sf )(z) = zf (z), f ∈ Cd[z], and by Eα : Cd[z] → Cd the evaluation operator at the point α ∈ C: Eα(f ) = f (α), f ∈ Cd[z]. It follows from the fundamental theorem of algebra that ran(cid:0)S − α(cid:1) = ker Eα. (2.1) A wide class of operators on Cd[z] is induced by d× d matrix polynomials. If M (z) is such a polynomial we define the operator M : Cd[z] → Cd[z] by (cid:0)M f(cid:1)(z) = M (z)f (z), f ∈ Cd[z]. Clearly, M S = SM . A square matrix polynomial is unimodular if its determinant is identically equal to a nonzero constant. If M (z) is a unimodular matrix polynomial we will call M a unimodular operator. In this case M is a bijection and its inverse is also a unimodular operator. 2.2. a Smith normal form representation (see for example [22, Satz 6.3] or [25]): In the sequel we use that any nonzero d × n matrix polynomial B(z) admits (2.2) B(z) = U (z)"D(z) 0 0# V (z), 0 where U (z) is a d × d unimodular matrix polynomial, V (z) is an n × n unimod- ular matrix polynomial and the matrix in the middle is a d × n matrix in which, for some l ∈ {1, . . . , min{d, n}}, D(z) is a diagonal l × l matrix polynomial with 6 B. ´CURGUS AND A. DIJKSMA monic diagonal entries: D(z) = diag(cid:0)b1(z), . . . , bl(z)(cid:1) such that bi(z) is divisible by bi+1(z), i ∈ {1, . . . , l − 1}. Notice that rank B(α) = l if and only if b1(α) 6= 0. If for some z ∈ C the rank of B(z) is d (n, respectively), then l = d (l = n) and the zero block row (column) in the matrix in the middle of the right hand side in (2.2) is not present. Remark 2.2. The matrix in the middle of the right hand side in (2.2) is uniquely determined by B(z). In this paper B(z) often is a matrix polynomial whose columns form a basis of a subspace B of Cd[z]. Then for any d × n matrix polynomial B1(z) whose columns also form a basis of B, the middle term of its Smith normal form is identical to that of B(z). Thus, the number l and the monic polynomials bj(z), j ∈ {1, . . . , l}, above are uniquely determined by the subspace B of Cd[z]. 2.3. Let S(z) be a d × 2d polynomial matrix. For j ∈ {1, . . . , d} let σj be the degree of the j-th row of S(z). By definition, a degree of a row is the degree of its transpose. Define S∞, the internal degree and the external degree of S(z) by: S∞ = lim z→∞ z   −σ1 . . . 0 · · · . . . · · · 0 . . . −σd z   S(z), extdeg S(z) = σ1 + ··· + σd, intdeg S(z) = max{ deg m(z) : m(z) is a d × d minor of S(z)}. and For a proof of the following theorem we refer to [33]. Theorem 2.3. Let P(z) be a d × 2d matrix polynomial with rankP(z) = d for all z ∈ C. Let S(z) be a matrix polynomial in the family (2.3) (cid:8)U (z)P(z) : U (z) unimodular(cid:9). The following statements are equivalent: rankS∞ = d. (b) (c) extdeg S(z) = intdeg S(z). (d) S(z∗)∗ has the "predictable degree property": (a) extdeg S(z) = min(cid:8)extdeg U (z)P(z) : U (z) unimodular(cid:9). For every u(z) =(cid:2)u1(z) ··· ud(z)(cid:3)⊤ ∈ Cd[z] we have deg(cid:0)S(z∗)∗u(z)(cid:1) = max(cid:8)σj + deg uj(z), j ∈ {1, . . . , d}(cid:9). A matrix polynomial S(z) in the family (2.3) satisfying the conditions (a) -- (d) is called row reduced. The multiset {σ1, . . . , σd} of row degrees for each row reduced matrix in the family (2.3) is the same. Its elements are called the Forney indices of any of the matrices in the family (2.3), in particular of P(z). We extend this definition to the case where the d × 2d matrix polynomial P(z) has full rank for some z ∈ C. For that we use the following lemma which is a standard tool in system theory, see for example [21]. Lemma 2.4. Let P(z) be a d × 2d matrix polynomial with rankP(z) = d for some z ∈ C. Then P(z) admits the factorization: (2.4) where G(z) is a d × d matrix polynomial with det G(z) 6≡ 0 and T(z) is a d × 2d matrix polynomial with rankT(z) = d for all z ∈ C. This factorization is essentially P(z) = G(z)T(z) z ∈ C, for all ON THE REPRODUCING KERNEL 7 unique, meaning that if also P(z) = G1(z)T1(z) for all z ∈ C, where G1(z) and T1(z) have the same properties as G(z) and T(z), then for some unimodular d × d matrix polynomial E(z): G1(z) = G(z)E(z)−1 and T1(z) = E(z)T(z), z ∈ C. The Forney indices of P(z) in the lemma are by definition the Forney indices of the matrix polynomial T(z) in the factorization (2.4). By the second part of the lemma, this definition is independent of the choice of the matrix G(z) in this factorization. For convenience of the reader we give a proof of Lemma 2.4 based on the Smith normal form of a matrix polynomial. Proof of Lemma 2.4. Let P(z) have the Smith normal form (2.2). The assumptions imply that l = d and that the matrix in the middle of (2.2) is equal to(cid:2)D(z) 0(cid:3). 0(cid:3) V (z). Then the factorization (2.4) Set G(z) = U (z)D(z) and T(z) = (cid:2)Id holds and G(z) and T(z) have the properties mentioned in the lemma. To prove uniqueness we use the fact that, since T(z) and T1(z) have full rank for all z ∈ C, they have right inverses, see [25]. These are 2d × d matrix polynomials S(z) and S1(z) such that T(z)S(z) = Id and T1(z)S1(z) = Id for all z ∈ C. Define the matrix polynomials E(z) = T1(z)S(z) and F (z) = T(z)S1(z). Then the equality G(z)T(z) = G1(z)T1(z) implies E(z) = G1(z)−1G(z) and F (z) = G(z)−1G1(z), hence E(z)F (z) = Id for all but finitely many z ∈ C. By continuity the last equality holds for all z ∈ C, hence E(z) is unimodular and has the stated properties. (cid:3) 2.4. The next two lemmas concern finite dimensional Pontryagin spaces. By the positive (negative) index of a Pontryagin space K we mean the dimension of a maximal positive (negative) subspace of K; evidently, the dimension of K is equal to the sum of the indices. Lemma 2.5. Let K be a Pontryagin space with positive and negative index equal to n. Let L be a subspace of K with dimL = 2n − τ . If L contains a maximal neutral subspace of K, then L⊥ is the isotropic part of L and L/L⊥ is a Pontryagin space with positive and negative index equal to n − τ . Proof. Let N be a maximal neutral subspace contained in L. Since N ⊥ = N , the inclusion N ⊆ L, yields L⊥ ⊂ N ⊂ L. Therefore, L⊥ is the isotropic part of L and dimL⊥ = τ . Let L = L⊥ +L− +L+ be a pseudo-fundamental decomposition of L. Since N is a neutral subspace of L, we have n = dimN ≤ τ + dim L±. Therefore 2n − τ = dimL = τ + dimL− + dimL+ ≥ τ + n − τ + n − τ = 2n − τ. (cid:3) This proves that dimL− = dimL+ = n − τ . Recall that a symmetric relation S in a Pontryagin space K is simple if S has no non-real eigenvalues and K = span{ker(S∗ − z) : z ∈ C \ R}. Below mul S∗ stands for the multi-valued part of the adjoint S∗ of S: mul S∗ = {g ∈ K : {0, g} ∈ S∗}. Lemma 2.6. Let S be a simple symmetric relation in a finite dimensional Pontrya- gin space of dimension n. Then the spaces mul S∗, ker S∗, and S∗ ∩ zI, z ∈ C, have the same dimension d′, say. In particular, the defect numbers of S are both equal to d′. Furthermore, dim ranS = dim S = dim domS = n − d′ and dim S∗ = n + d′. Proof. First notice that by [9, Proposition 2.4] S is an operator and S has no eigenvalues. The following statements are equivalent: 8 B. ´CURGUS AND A. DIJKSMA (a) dim(cid:0)mul S∗(cid:1) = d′. (b) codim(cid:0)dom S(cid:1) = d′. (c) codim(cid:0)ran(S − z∗)(cid:1) = d′ for all z ∈ C. (d) dim(cid:0)S∗ ∩ zI(cid:1) = d′ for all z ∈ C. that d′ = dim(cid:0)ker S∗(cid:1). The equalities n − d′ = dim domS = dim S = dim ranS The relation (dom S)⊥ = mul S∗ implies the equivalence (a)⇔(b). The equivalence (b)⇔(c) follows from the fact that S − z∗ is one-to-one. By taking the orthogonal complements we obtain the equivalence (c)⇔(d). Notice that (d) with z = 0 implies follow from (b) and the fact that S is an injective operator. Since dim S∗ = 2n − dim S the last equality follows. (cid:3) 2.5. A d× d matrix function K(z, w) will be called a polynomial Hermitian kernel if it is a polynomial of two variables z and w∗ and K(z, w)∗ = K(w, z), z, w ∈ C. This implies that the degree of K(z, w) as a polynomial in z equals the degree of K(z, w) as a polynomial in w∗. If we denote this common degree by p − 1, then K(z, w) can be expanded as (2.5) K(z, w) = p−1Xk=0 Ajkzjw∗k, z, w ∈ C, where Ajk, j, k ∈ {0, . . . , p − 1}, are d × d matrices. Since K(z, w) is a Hermitian kernel, the dp × dp block matrix p−1Xj=0 A = A00 ... Ap−1,0 A0,p−1 ··· . . . ··· Ap−1,p−1 ...  (2.6) is self-adjoint. It also follows that the number of negative squares of K(z, w) equals the number of negative eigenvalues of A and the number of positive squares of K(z, w) equals the number of positive eigenvalues of A. The dimension of the reproducing kernel space corresponding to K(z, w) is the rank of A. These obser- vations are used in the proof of the following lemma. Lemma 2.7. Let K(z, w) be a d× d matrix polynomial Hermitian kernel of degree p − 1. For q ∈ N set Lq(z, w) = i (zq − w∗q)K(z, w), z, w ∈ C. If q ≥ p, then the positive and the negative index of the reproducing kernel Pon- tryagin space with kernel Lq(z, w) are equal and coincide with the dimension of the reproducing kernel Pontryagin space with kernel K(z, w). Proof. Write K(z, w) in the form (2.5) and denote by A the matrix (2.6). We calculate the coefficients of the matrix polynomial Lq(z, w) for q ≥ p: Lq(z, w) = izq Ajkzjw∗k − iw∗q p−1Xk=0 p−1Xj=0 p−1Xk=0 p−1Xj=0 = iAjkzq+jw∗k + Ajkzjw∗k p−1Xk=0 (−i)Ajkzjw∗(q+k) p−1Xj=0 p−1Xk=0 p−1Xj=0 ON THE REPRODUCING KERNEL 9 = = q+p−1Xj=0 q+p−1Xj=0 p−1Xj=0 iA(j−q)kzjw∗k + q+p−1Xk=0 p−1Xk=0 q+p−1Xk=0 (cid:0)iA(j−q)k − iAj(k−q)(cid:1)zjw∗k, (−i)Aj(k−q)zjw∗k where we set Ajk = 0 whenever j < 0 or k < 0 or j > p − 1 or k > p − 1. In other words, the 2d(p + q) × 2d(p + q) self-adjoint matrix formed by the coefficients of Lq(z, w) is given by where the 0 in the center is a d(q − p) × d(q − p) matrix. With 0 −iA 0 0 0 0 iA 0 0  Idp 0 iIdp 0 Id(q−p) 0 iIdp 0 Idp B = √2 E∗BE = E∗ E = 1 0 −iA 0 0 0 0 iA 0 0  E = A 0 0 0 0  0 −A . 0 0 we have EE∗ = Id(q+p) and Therefore the rank of B is twice the rank of A. Moreover, B has equal numbers of positive and negative eigenvalues. Since the positive and negative index of the reproducing kernel Pontryagin space with kernel Lq(z, w) coincide with the number of positive and negative eigenvalues of B the lemma is proved. (cid:3) A polynomial reproducing Nevanlinna kernel introduced in the Introduction is a polynomial Hermitian kernel. Since in the proof of Theorem 1.1 the polynomials in a Nevanlinna pair never appear separate we adopt the following equivalent definition of a polynomial Nevanlinna kernel: A d × d matrix function K(z, w) is called a polynomial Nevanlinna kernel if it can be represented as (2.7) P(z)Q−1P(w)∗ = i (z − w∗)K(z, w) for all z, w ∈ C, where Q is a 2d × 2d self-adjoint matrix with d positive and d negative eigenvalues and P(z) is a d × 2d matrix polynomial such that P(z) has rank d for some z ∈ C. With (2.8) Q = Q1 :=" 0 −iId 0 # iId and P(z) =(cid:2)M (z) N (z)(cid:3) the definition in the Introduction is obtained from the new one. The assumptions on Q imply that there exists a constant invertible matrix T such that Q = T Q1T ∗. Now, if we write P(z)T = (cid:2)M (z) N (z)(cid:3), we have K(z, w) = KM,N (z, w). Since P(z) is a polynomial, the condition that rank P(z) = d for some z ∈ C implies that rank P(z) = d for all but finitely many z ∈ C. A polynomial Nevanlinna kernel will be called a full Nevanlinna kernel if P(z) can be chosen such that rank P(z) = d for all z ∈ C. 10 B. ´CURGUS AND A. DIJKSMA 3. Proof of Theorem 1.2 3.1. Let B be a vector subspace of Cd[z]. By SB we denote the range restriction of S to B, that is, In graph notation this means: (cid:0)SBf(cid:1)(z) = zf (z), f ∈ dom SB. dom SB = B ∩ S−1B, SB =(cid:8){f, g} : f, g ∈ B, g(z) = zf (z) for all z ∈ C(cid:9). By (2.1), for α ∈ C we have (3.1) ran(cid:0)SB − α(cid:1) =(cid:0)S − α(cid:1)(cid:0)B ∩ S−1B(cid:1) ⊆ ran(cid:0)S − α(cid:1) ∩ B = B ∩ ker Eα. The reverse inclusion is equivalent to the implication f ∈ B, α ∈ C, f (α) = 0 ⇒ f (z) = (z − α)g(z) for some g ∈ dom SB. In some cases this implication does not hold. For example, it does not hold for any α ∈ C in the space B ⊂ C2[z] given by B =(cid:26)(cid:20) a0 + a2z2 b0 + b1z (cid:21) : a0, a2, b0, b1 ∈ C(cid:27) . z − α(cid:3)⊤ which is 0 at z = α, but B does not Indeed, B contains (cid:2)z2 − α2 contain(cid:2)z + α 1(cid:3)⊤ . That the implication, or equivalently, equality in (3.1) holds, is characterized in terms of canonical subspaces of Cd[z] in Theorem 1.2 in the Introduction. This section is devoted to the proof of this theorem. Let B(z) be a d × n matrix polynomial whose columns form a basis for B, n = dim B. Then, as will be shown in the proof of Theorem 1.2, the sets in (1.3) are equal to (cid:8)α ∈ C : b1(α) 6= 0(cid:9), where b1(z) is the scalar polynomial in the Smith normal form (2.2) of B(z). We will first prove Theorem 1.2 for the case where the sets in (1.3) are equal to C, see Theorem 3.4 below. In this case W (z) is unimodular. The proof of Theorem 3.4 is based on the following three lemmas. Lemma 3.1. Let B be a finite dimensional subspace of Cd[z] such that ran(cid:0)SB − α(cid:1) = B ∩ ker Eα for all α ∈ C. If dom SB ⊆ B′ ⊆ B, then ran(cid:0)SB′ − α(cid:1) = B′ ∩ ker Eα for all α ∈ C. Proof. Let f ∈ B′ ∩ ker Eα. Then f ∈ B ∩ ker Eα = ran(cid:0)SB − α(cid:1), that is, g ∈ dom SB′ and f = (SB′ − α)g. This proves B′ ∩ ker Eα ⊆ ran(cid:0)SB′ − α(cid:1). Since f = Sg− αg for some g ∈ domSB ⊆ B′. From f, g ∈ B′ we infer g, Sg ∈ B′. Hence the reverse inclusion is obvious, the lemma is proved. (cid:3) Lemma 3.2. Let B be an n-dimensional subspace of Cd[z]. Then (3.2) if and only if there exists a unimodular operator W such that B = W Cd is a top coordinate subspace of Cd. B ∩ ker Eα = {0} for all α ∈ C n, where Cd n Proof. If n = 0, the statements are trivial with W (z) = Id. From now on we assume n ≥ 1. If B(z) is any d × n matrix polynomial whose columns form a basis of B, then, clearly, (3.3) (cid:8)α ∈ C : rank B(α) = n(cid:9) =(cid:8)α ∈ C : B ∩ ker Eα = {0}(cid:9). ON THE REPRODUCING KERNEL 11 Assume (3.2). Let B(z) be a d × n matrix polynomial whose columns form a basis of B. By (3.3), for all α ∈ C the rank of B(α) is n and n ≤ d. Hence B(z) V (z), where U (z) and V (z) are unimodular. Define admits the Smith normal form (see (2.2)): B(z) = U (z)(cid:2)In Id−n# . W (z) = U (z)"V (z) (3.4) 0 0 0(cid:3)⊤ it follows that B = W Cd n. This proves the only if statement. polynomial W (z) such that B = W Cd Then W (z) is a unimodular d×d matrix polynomial and from B(z) = W (z)(cid:2)In 0(cid:3)⊤ Cd. Then the columns of B(z) = W (z)(cid:2)In 0(cid:3)⊤ To prove the if statement, assume that there exists a d × d unimodular matrix n is a top coordinate subspace of form a basis of B and the rank of (cid:3) B(α) is n for all α ∈ C. The equality (3.2) follows from (3.3). Lemma 3.3. Let B be an n-dimensional subspace of Cd[z] and let C be a canonical subspace of Cd[z] with degrees µ1 ≥ ··· ≥ µd ≥ 0 of which k are positive. Assume C + SC ⊆ B and (3.5) n, where Cd B ∩ ker Eα ⊆ C + SC for all α ∈ C. Then there exists a unimodular operator W which acts as the identity on C + SC and is such that (3.6) where m = n − (µ1 + ··· + µd)(≥ 0). B = W(cid:0)Cd m + C + SC(cid:1), Notice that Cd m + C + SC is a canonical subspace. If m ≤ k, then Cd m + C + SC coincides with C + SC and W = Id. Proof. If C = {0} the statement follows from Lemma 3.2. From now on we assume C 6= {0}. Then µ1 > 0, consequently k ∈ {1, . . . , d} and Cd k ⊆ C. We consider two cases: k = d and k < d. (i) Assume k = d. Then Cd ⊆ C ⊆ B. Let f ∈ B. It can be written as f (z) = f (0)+zh(z) = f (0)+(Sh)(z). Then Sh = f −f (0) ∈ B. Since(cid:0)Sh(cid:1)(0) = 0, by (3.5) we get Sh ∈ B∩ ker E0 ⊆ C + SC, which implies f = f (0) + Sh ∈ C + SC. That is, B = C + SC. In this case m = d and with W = Id the lemma is proved. (ii) Assume k < d. If C + SC = B, then (3.6) holds with W = Id and m = k, implying that Cd m ⊆ C. From now on we assume that C + SC is a proper subspace of B. Recall that Pd,k is the coordinate projection. A trivial, but important observation is (3.7) Let α ∈ C be arbitrary and let f ∈ B be such that (Id − Pd,k)f (α) = 0. By (3.7), there exists a p ∈ C + SC such that p(α) = Pd,kf (α), hence for all α ∈ C. k = ran Pd,k Eα(cid:0)C + SC(cid:1) = Cd (f − p)(α) = (Id − Pd,k)f (α) + Pd,kf (α) − p(α) = 0, that is, f − p ∈ ker Eα. Since also f − p ∈ B, (3.5) implies f − p ∈ C + SC. Thus both p and f − p belong to C + SC, implying that f ∈ C + SC. We have proved the implication: (3.8) f ∈ B, α ∈ C and (Id − Pd,k)f (α) = 0 ⇒ f ∈ C + SC. 12 B. ´CURGUS AND A. DIJKSMA Let L0 be a subspace of B be such that (cid:0)C + SC(cid:1) ∩ L0 = {0} The dimension of L0 is and B =(cid:0)C + SC(cid:1) ⊕ L0. Let B0(z) be a d× j matrix polynomial whose columns form a basis of L0. Decom- pose B0(z) as j = n −(cid:0)µ1 + ··· + µd + k(cid:1) ≥ 1. B0,b(z)# , B0(z) ="B0,t(z) where B0,t(z) is a k × j matrix polynomial and B0,b(z) a (d − k) × j matrix poly- nomial. We will prove that for all α ∈ C. rank B0,b(α) = rank(Id − Pd,k)B0(α) = j (Id − Pd,k)f (α) = 0. By (3.8), f ∈ (cid:0)C + SC(cid:1) ∩ L0, consequently f = 0, that is, (3.9) The first equality is trivial. To prove the second let α ∈ C be arbitrary and x ∈ Cj be such that (Id − Pd,k)B0(α)x = 0. Set f (z) = B0(z)x. Then f ∈ L0 and B0(z)x = 0 for all z ∈ C. Since the columns of B0(z) form a basis of L0, this implies x = 0. This proves (3.9). Hence j ≤ d− k. If j = d− k, the (d− k)× (d− k) matrix polynomial Wb(z) := B0,b(z) is unimodular. If j < d − k we can extend B0,b(z) to a unimodular (d − k) × (d − k) matrix polynomial (also denoted by) Wb(z) with det Wb(α) 6= 0 in the same way as the matrix B(z) was extended to W (z) in (3.4) by the means of the Smith normal form. In both cases the first j columns of Wb(z) are the columns of B0,b(z). Let Wt(z) be the k × (d − k) matrix obtained from the k × j matrix B0,t(z) by adding d − k − j zero columns on the right. Define the d × d matrix polynomial W (z) by W (z) =" Ik 0(d−k)×k Wb(z)# . Wt(z) Then W (z) is unimodular and W (z)ed,k+l, l = 1, . . . , j, are the columns of the matrix B0(z). The operator W acts as the identity on C+ SC and W Cd k + L0, where m = Cd m = k + j = k + n −(cid:0)µ1 + ··· + µd + k(cid:1) = n −(cid:0)µ1 + ··· + µd(cid:1). m + C + SC(cid:1) = Cd k + L0 + C + SC = B. Theorem 3.4. Let B be a finite dimensional subspace of Cd[z]. The equality Hence W(cid:0)Cd (3.10) holds if and only if there exist a d × d unimodular matrix polynomial W (z) and a canonical subspace C of Cd[z] such that B = W C. ran(cid:0)SB − α(cid:1) = B ∩ ker Eα for all α ∈ C (cid:3) Proof. We first prove the if statement. To prove (3.10) it suffices to show that B ∩ ker Eα ⊆ ran(SB − α). Let f ∈ B ∩ ker Eα. Then f (α) = 0 and f = W g for some g ∈ C. Since W is unimodular, g(α) = 0. Since C is canonical, the polynomial g(z)/(z − α) belongs We prove the only if statement by induction on the dimension of B. Assume (3.10). The theorem is obviously true if dim B = 0. Lemma 3.2 implies that it to C. Therefore f (z)/(z − α) = W (z)(cid:0)g(z)/(z − α)(cid:1) ∈ B, hence f ∈ ran(SB − α). ON THE REPRODUCING KERNEL 13 is true if dim B = 1 for then B ∩ Eα = {0}. Let n ∈ N and state the inductive hypothesis: If A is a subspace of Cd[z] with dim A < n and such that for all α ∈ C, ran(cid:0)SA − α(cid:1) = A ∩ ker Eα (3.11) then there exists a unimodular d × d matrix polynomial operator F (z) such that F A is a canonical subspace of Cd[z]. Let B be a finite dimensional subspace of Cd[z] such that (3.10) holds and dim B = n. Then A = dom SB is a proper subspace of B. Therefore dim A < n. If A = {0}, then B ∩ ker Eα = ran(cid:0)SB − α(cid:1) = {0} and the theorem follows from Lemma 3.2. Now we assume A 6= {0}. By Lemma 3.1 the subspace A satisfies (3.11). By the inductive hypothesis there exists a unimodular matrix polynomial F (z) such that D := F A is a canonical subspace of Cd[z]. Since F and S commute we have D = F A = F dom SB = dom SF B, hence D + SD ⊆ U B. To apply Lemma 3.3 to F B we need to verify (3.5). Let f ∈ B be such that (F f )(α) = 0. Then f (α) = 0 and, by (3.10), there exists a g ∈ dom SB = A such that f = SBg − αg ∈ A + SA. Therefore, F f ∈ D+SD, which verifies (3.5). Lemma 3.3 applied to F B yields that there exists a unimodular operator U such that U −1F B is a canonical subspace of Cd[z]. This proves the theorem with W = F −1U (cid:3) 3.2. The following lemma will be used to deduce Theorem 1.2 from Theorem 3.4. Lemma 3.5. Let B be an n-dimensional subspace of Cd[z] and let B(z) be a d× n matrix polynomial whose columns form a basis of B. Let l be the size of the square diagonal matrix in the Smith normal form (2.2) of B(z). Then (cid:8)α ∈ C : ran(SB − α) = B ∩ ker Eα(cid:9) =(cid:8)α ∈ C : rank B(α) = l(cid:9) if and only if the set on the left hand side is nonempty. In this case dim ran SB = dim B − l. Proof. The only if statement follows from the fact that the set on the right hand side in (3.12) is nonempty. Before proving the if statement we show (3.12) (3.13) For all α ∈ C we have ran(SB − α) ⊆ B ∩ ker Eα, and hence dim ran SB ≤ dim B − l. dim ran SB = dim ran(SB − α) ≤ dim(B ∩ ker Eα) = dim B − rank B(α). Consequently, l = maxα∈C rank B(α) ≤ dim B − dim ran SB. This proves (3.13). of (3.12). Then equality holds in (3.13). Indeed, this follows from To prove the if statement assume that α0 ∈ C is in the set on the left hand side dim B − l ≥ dim ran SB = dim ran(SB − α0) = dim(cid:0)B ∩ ker Eα0(cid:1) = dim B − rank B(α0) ≥ dim B − l. This proves the last statement in the lemma. Now the equality (3.12) follows from the following sequence of equivalences which hold for all α ∈ C: rank B(α) = l ⇔ dim(B ∩ ker Eα) = dim B − l ⇔ dim(B ∩ ker Eα) = dim ran SB 14 B. ´CURGUS AND A. DIJKSMA ⇔ dim(B ∩ ker Eα) = dim ran(SB − α) ⇔ ran(SB − α) = B ∩ ker Eα. (cid:3) Proof of Theorem 1.2. We first prove the if statement. It suffices to prove the inclusion B ∩ ker Eα ⊆ ran(SB − α), as the reverse inclusion always holds. Let f ∈ B ∩ ker Eα. Then f (α) = 0 and f = W g with g ∈ C. Since W (α) is invertible, g(α) = 0. As C is canonical, the polynomial h(z) = g(z)/(z − α) belongs to C. Therefore W (z)h(z) ∈ B and (x − α)W (z)h(z) = f (z), which implies f ∈ ran(SB − α). To prove the only if statement, assume that (1.2) holds for α = α0. Let B(z) be a d × n matrix polynomial whose columns form a basis of B. Let B(z) = U (z)"D(z) 0 0# V (z) 0 be the Smith normal form (2.2) of B(z) where D(z) is an l× l diagonal matrix with nonzero diagonal entries. Now define the space B1 ⊂ Cd[z] as the span over C of the columns of Set B1(z) = U (z)"Il F (z) = U (z)"D(z) 0 0 0 0# V (z). Id−l# U (z)−1. 0 Then B = F B1 and detF (α0) 6= 0. Moreover, since {α ∈ C : detF (α) 6= 0} = {α ∈ C : rankB(α) = l} and by Lemma 3.5, (1.3) holds for F (z). From detF (α0) 6= 0 it follows that ran(SB − α0) = B ∩ ker Eα0 ⇒ ran(SB1 − α0) = B1 ∩ ker Eα0 . Since rank B1(α) = l for all α ∈ C, Lemma 3.5 implies that ran(SB1 − α) = B1 ∩ ker Eα for all α ∈ C. By Theorem 3.4, there exists a unimodular matrix U (z) such that C = U −1B1 is a canonical subspace of Cd[z], hence B = W C with W = F U . Finally, (1.3) holds, because U is unimodular and F (z) satisfies (1.3). (cid:3) 3.3. Theorem 1.2 can also be formulated in terms of matrix polynomials: Theorem 3.6. Let B be an n-dimensional subspace of Cd[z], n ≥ 1. Let B(z) be a d × n matrix polynomial whose columns form a basis of B. Let b1(z) and l be as in the Smith normal form (2.2) of B(z). Then l + dim dom SB = dim B if and only if there exist (a) a d×d matrix polynomial W (z) whose determinant has the same zeros as b1(z), (b) nonnegative integers m and δ0 ≥ δ1 ≥ ··· ≥ δm with δ0 + ··· + δm = n and (c) an invertible n × n constant matrix T such that (3.14) B(z) = W (z)hPδ0 Pδ1 z ··· Pδm zmi T for all z ∈ C, where Pδ stands for the d × δ matrix: Pδ =(cid:2)Iδ . 0(cid:3)⊤ ON THE REPRODUCING KERNEL 15 Proof. For all α ∈ C we have ker(SB − α) ⊆ B ∩ kerEα. For all α ∈ C with b1(α) 6= 0 we have l + dim dom SB = dim ran(cid:0)EαB(cid:1) + dim ran(SB − α) ≤ dim ran(cid:0)EαB(cid:1) + dim(B ∩ kerEα) = dim B and equality holds if and only if ker(SB − α) = B ∩ kerEα. To prove the only if statement, assume l + dim dom SB = dim B. Then we can apply Theorem 1.2: There exist a matrix polynomial W (z) satisfying (a) and a canonical subspace C of Cd[z] such that B = W C. Let µ1 ≥ ··· ≥ µd be the degrees of C. Since n ≥ 1, we have µ1 ≥ 1. set m = µ1 − 1 and δj = #{i ∈ {1, . . . , d} : µi > j}, j ∈ {0, . . . , m}. Then the equality in (b) holds. Since the columns of the matrix(cid:2)Pδ0 form a basis for C, there exists a matrix T satisfying (c) such that (3.14) holds. ··· Pδmzm(cid:3) To prove the if statement, we note that (a)-(c) and (3.14) imply that B = W C with C as above, and hence Theorem 1.2 can be applied and together with the if and only if statement at the beginning of the proof yield that l + dim dom SB = dim B. (cid:3) Remark 3.7. Denote by (Sj)B the range restriction of Sj to B. Then in item (b) of Theorem 3.6: m is the nonnegative integer with and {0} = dom(Sm+1)B ( dom(Sm)B Moreover, if we set δ−1 = d, then the numbers δj = dim dom(Sj)B − dim dom(Sj+1)B, µk = 1 + max(cid:8)j ∈ {−1, 0, . . . , m} : δj ≥ k(cid:9), (3.15) j ∈ {0, . . . , m}. k ∈ {1, . . . , d}, are the degrees of the canonical space W −1B. In the next section we will see that if B ⊂ Cd[z] is a Pontryagin space which satisfies the conditions (A) and (B) of Theorem 1.1, then the numbers (3.15) are the Forney indices of a matrix polynomial P(z) in a representation (2.7) of the Nevanlinna reproducing kernel K(z, w) of B; see Remark 4.3. 4. Proof of Theorem 1.1 4.1. We divide the proof of Theorem 1.1 in two parts. In the first part we prove the if statements and in the second part we prove the only if statements. In the first part we will need characterizations of the defect numbers of the operator SB of multiplication by the independent variable in the Pontryagin space B which are collected in the following remark. Remark 4.1. Clearly, SB has no eigenvalues and for any subset Ω of C containing more than d × max{degf : f ∈ B} elements we have ∩w∈Ω ran(SB − w∗) = {0} or, equivalently, B = span(cid:8)ker(S∗ B − w) : w ∈ Ω(cid:9). Now assume (A) of Theorem 1.1. Then, by the above observations, SB is a simple symmetric operator and hence its defect numbers coincide and are equal to the codimension of ranSB, see Lemma 2.6. It follows from Lemma 3.5 that the defect 16 B. ´CURGUS AND A. DIJKSMA numbers of SB are also equal to the integer l introduced in Remark 2.2. Hence l ∈ {1, . . . , min{d, n}}, where n = dim B. Now also assume (B) of Theorem 1.1. Then l can be characterized in a different way. Indeed, by Theorem 1.2, there exist a canonical subspace C ⊆ Cd[z] with degrees µ1 ≥ ··· ≥ µd ≥ 0 and a d × d matrix polynomial W (z) with det W (α) 6= 0 such that B = W C. Since, by Lemma 2.6, we have n − l = dimdomSB = dimranSB and since multiplication by z and by W (z) commute, l is uniquely determined by the inequalities: (4.1) µ1 ≥ ··· ≥ µl ≥ 1 and µl+1 = ··· = µd = 0. Proof of the if statements in Theorem 1.1. Assume (A) and (B). We show that B has a reproducing Nevanlinna kernel in steps (i) -- (iv). In step (v) we prove the last if statement in the theorem. (i) By Theorem 1.2 there exist a canonical subspace C ⊆ Cd[z] with degrees µ1 ≥ ··· ≥ µd ≥ 0 and a d × d matrix polynomial W (z) with det W (α) 6= 0 such that B = W C. Then, by Remark 4.1, the defect numbers of the symmetric operator SB are both equal to l, where l is determined by the inequalities (4.1). It follows that the elements of B are of the form: (4.2) f (z) ∈ B ⇒ f (z) = W (z)(cid:20)x(z) 0 (cid:21) , where x(z) is an l × 1 vector polynomial and 0 denotes the zero vector of size (d − l) × 1. Let n = dim B and let B(z) be a d × n matrix polynomial whose columns form a basis of B. Let G be the corresponding Gram matrix and write the reproducing kernel K(z, w) of B as K(z, w) = B(z)G−1B(w)∗, z, w ∈ C. By (B), this representation implies that for each w ∈ C which belongs to the set in (3.12) the columns of K(z, w) span an l-dimensional subspace of B, in formula: (4.3) dim(cid:8)K(· , w)x : x ∈ Cd(cid:9) = l whenever w ∈(cid:8)α ∈ C : rank B(α) = l(cid:9). (ii) In the following we use graph notation in the space B⊕ B. The operator SB is identified with its graph in B⊕ B and its adjoint S∗ B is the orthogonal complement of SB in B ⊕ B equipped with the Lagrange inner product (cid:2)(cid:2){f, g},{p, q}(cid:3)(cid:3) = −i(cid:0)[g, p]B − [f, q]B(cid:1), Let w ∈ C, x ∈ Cd and {f, Sf} ∈ SB be arbitrary. Then {f, g},{p, q} ∈ B ⊕ B. = 0 = −i(cid:0)x∗(Sf )(w) − wx∗f (w)(cid:1) B ∩ (w∗I) for all w ∈ C. Ac- cording to the definition of defect number (see [9, p.369]) and by (4.3), it follows (cid:2)(cid:2){f, Sf},{K(·, w)x, w∗K(·, w)x}(cid:3)(cid:3) = −i(cid:0)(cid:2)Sf, K(·, w)x(cid:3)B −(cid:2)f, w∗K(·, w)x(cid:3)B(cid:1) and hence (cid:8){K(·, w)x, w∗K(·, w)x} : x ∈ Cd(cid:9) ⊆ S∗ that for all w ∈(cid:8)α ∈ C \ R : rank B(α) = l(cid:9) (cid:8){K(·, w)x, w∗K(·, w)x} : x ∈ Cd(cid:9) = S∗ B ∩ (w∗I), L0 := span(cid:8){K(·, w)x, w∗K(·, w)x} : w ∈ C, x ∈ Cd(cid:9) because for such w's both sets have dimension l. Consider the subspace (4.4) (4.5) ON THE REPRODUCING KERNEL 17 of S∗ B. Since SB has no eigenvalues, the generalized von Neumann formula given in [9, Theorem 3.7] implies that for d + 1 distinct points w0, . . . , wd from the set (cid:8)α ∈ C : Imα > 0, rank B(α) = l(cid:9) we have B ∩ (w∗ B = SB + S∗ S∗ 0I) + dXj=1 S∗ B ∩ (wj I). Combined with (4.4) and (4.5) this yields SB + L0 ⊆ S∗ (4.6) B = SB + L0. S∗ B ⊆ SB + L0, and hence z, w ∈ C. (iii) Let(cid:0)B1, [· ,· ]B1(cid:1) be the reproducing kernel Pontryagin space whose kernel is L1(z, w) = i (z − w∗)K(z, w), B, [[· ,· ]](cid:1) → B1 defined by T(cid:0){f, g}(cid:1) = Sf − g, {f, g} ∈ S∗ the operator T :(cid:0)S∗ We claim that its positive and negative index are l. To prove the claim we consider B, and show that it is a partial isometry onto B1 with null space kerT = SB. The last equality is easy to verify. That ran T = B1 follows from (4.6) as it implies (with w ∈ C and x ∈ Cd): That T is isometric follows from (4.6), the symmetry of SB and the equalities (with w, v ∈ C and x, y ∈ Cd): (cid:0)T(cid:0){K(·, w)x, w∗K(·, w)x}(cid:1)(cid:1)(z) = (z − w∗)K(z, w)x = −iL1(z, w)x. (cid:2)(cid:2)(cid:8)K(·, w)x,w∗K(·, w)x(cid:9),(cid:8)K(·, v)y, v∗K(·, v)y(cid:9)(cid:3)(cid:3) = y∗L1(v, w)x = −i(cid:16)(cid:2)w∗K(·, w)x, K(·, v)y(cid:3)B −(cid:2)K(·, w)x, v∗K(·, v)y(cid:3)B(cid:17) = i(cid:0)v − w∗(cid:1)y∗K(v, w)x =(cid:2)−iL1(·, w)x,−iL1(·, v)y(cid:3)B1 =hT(cid:0){K(·, w)x, w∗K(·, w)x}(cid:1), T(cid:0){K(·, v)y, v∗K(·, v)y}(cid:1)iB1 . The claim now follows because(cid:0)S∗ B/SB, [[· ,· ]](cid:1) is a Pontryagin space with positive and negative index l (see [9, Theorem 2.3(c)]) and T establishes a unitary mapping between this space and B1. (iv) Let B1(z) be a d × 2l matrix polynomial whose columns form a basis of B1, and let Q1 be the corresponding 2l × 2l Gram matrix. Then Q1 is self-adjoint and, by the claim proved in (iii), has l positive and l negative eigenvalues. Let B2(z) be the d × 2(d − l) matrix polynomial defined by B2(z) = W (z)(cid:20) 0 Id−l 0 Id−l(cid:21) , where the zero matrices are of size l × (d − l). Define the d× 2d matrix polynomial P(z) by P(z) =(cid:2)B1(z) B2(z)(cid:3) and the 2d × 2d block diagonal matrix Q by 0 iId−l Q ="Q1 0 Q2# , where Q2 =" 0 0 # . −iId−l Then Q is self-adjoint and has d positive and d negative eigenvalues. We claim that (I) rankP(z) = d for some z ∈ C, and 18 B. ´CURGUS AND A. DIJKSMA (II) P(z)Q−1P(w)∗ = i(z − w∗)K(z, w) for all z, w ∈ C. We prove (I): The inclusion B1 = T (S∗ exists an l × 2l matrix polynomial X(z) such that B) ⊆ B + SB and (4.2) imply that there B1(z) = W (z)(cid:20)X(z) 0 (cid:21) , where now 0 stands for the (d−l)×2l zero matrix. The complex number α satisfying (B) belongs to the sets in (3.12) and (1.3) and hence rankX(α) = rankB1(α) = dimEαB1 = dimspan(cid:8)L1(α, w)x : w ∈ C, x ∈ Cd(cid:9) = dimspan(cid:8)i(α − w∗)K(α, w)x : w ∈ C, x ∈ Cd(cid:9) = dimspan(cid:8)K(α, w)y : w ∈ C, y ∈ Cd(cid:9) = dimEαB = rankB(α) = l. The equality P(z) = W (z)(cid:20)X(z) 0 0 Id−l 0 Id−l(cid:21) implies that rankP(α) = d. This proves (I). We prove (II): P(z)Q−1P(w)∗ = B1(z)Q−1 1 B1(w)∗ + B2(z)Q−1 2 B2(w)∗ = L1(z, w) + W (z)(cid:20) 0 = i(z − w∗)K(z, w). Id−l 0 Id−l(cid:21)(cid:20) 0 −iId−l iId−l 0 (cid:21)(cid:20)0 0 Id−l Id−l(cid:21) W (w)∗ Items (I) and (II) show that K(z, w) is a polynomial Nevanlinna kernel for B. This completes the proof of the if statement. (v) If (B) holds for all α ∈ C, then, by Theorem 3.4, W (z) is unimodular and the proof of (I) shows that then rankP(z) = d for all z ∈ C. 4.2. lemma. In the proof of the only if statements in Theorem 1.1 we use the following (cid:3) Lemma 4.2. Let Q be a self-adjoint 2d × 2d matrix with d positive and d negative eigenvalues. Let P(z) be a d × 2d matrix polynomial such that (a) P(z)Q−1P(z∗)∗ = 0 for all z ∈ C, (b) rankP(z) = d for all z ∈ C, and (c) P(z) is row reduced and has row degrees σ1, . . . , σd, assumed ordered so that σ1 ≥ ··· ≥ σd and σ1 = degP(z) =: p. Equip C2d[z]<p with the inner product [f, g]Q = p−1Xj=0 b∗ p−1−j Q−1aj, f (z) = ajzj, g(z) = p−1Xj=0 p−1Xj=0 bjzj, aj, bj ∈ C2d, ON THE REPRODUCING KERNEL 19 and consider the following subspace of C2d[z]<p : zp−1−kw∗kP(w)∗x : w ∈ C, x ∈ Cd). Lp = span(p−1Xk=0 p =nf (z) ∈ C2d[z]<p : f (z) = P(z∗)∗u(z) with u(z) ∈ Cd[z]o. Then the orthogonal complement of Lp in (C2d[z]<p, [· ,· ]Q) is (4.7) L⊥ p is a Pontryagin space with positive and It is the isotropic part of Lp and Lp/L⊥ negative index σ1 + ··· + σd. Proof. For an element f (z) = Pp−1 hold: j=0 ajzj ∈ C2d[z]<p the following equivalences f (z) ∈ L⊥ p ⇔ p−1Xk=0 w∗ka∗ k!Q−1P(w)∗ = 0 for all w ∈ C, ⇔ P(z)Q−1f (z) = 0 ⇔ f (z) = P(z∗)∗uz for all z ∈ C, for some uz ∈ Cd and all z ∈ C. The last equivalence follows from (a) and (b). To prove that the vector uz depends polynomially on z we use that the Smith normal form (2.2) of P(z) is given by: P(z) = U (z)(cid:2)Id 0(cid:3)V (z), where U (z) and V (z) are unimodular matrices. Then f (z) = P(z∗)∗uz = V (z∗)∗(cid:20)Id 0(cid:3) U (z∗)−∗f (z) 0(cid:21) U (z∗)∗uz ⇒ uz = V (z∗)−∗(cid:2)Id and the right hand side belongs to Cd[z]. This proves (4.7). Since P(z∗)∗ has full rank for every z ∈ C, it acts as an injection on Cd[z], dimL⊥ p = dim(cid:8)u(z) ∈ Cd[z] : deg(cid:0)P(z∗)∗u(z)(cid:1) < p(cid:9). The number on the right hand side can be expressed in terms of the Forney indices of P(z). Indeed, since P(z) is row reduced, it has the "predictable degree property" (see Theorem 2.3): therefore (4.8) Consequently, the space on the right hand side of in (4.8) equals deg(cid:0)P(z∗)∗u(z)(cid:1) = max(cid:8)σj + deg uj(z) : j ∈ {1, . . . , d}(cid:9). (cid:8)u(z) ∈ Cd[z] : deg uj(z) < p − σj, j ∈ {1, . . . , d}(cid:9), p = dp +(cid:0)σ1 + ··· + σd(cid:1). dim Lp = dim C2d[z]<p − dim L⊥ whose dimension is dp −(cid:0)σ1 + ··· + σd(cid:1). Hence dim L⊥ p = dp −(cid:0)σ1 + ··· + σd(cid:1) and To prove the last two statements of the lemma we apply Lemma 2.5 with n = dp and τ = dp− (σ1 +··· + σd). The assumptions about Q in the lemma readily imply that C2d[z]<p is a 2dp-dimensional Pontryagin space with negative index dp. It remains to construct a maximal neutral subspace of C2d[z]<p which is contained in Lp. We begin with the subspace H = ranH, where the operator H : Cd[z] → C2d[z] u(z). For example, if P(z) is written as: maps u(z) ∈ Cd[z] into the polynomial part of P(cid:0)1/z∗(cid:1)∗ P(z) = P0 + zP1 + ··· + zpPp, 20 B. ´CURGUS AND A. DIJKSMA then for k ∈ {0} ∪ N and x ∈ Cd H(cid:0)zkx(cid:1) = zjP ∗ k−j x if k < p, kXj=0 zk−p zjP ∗ p−jx if p ≤ k. pXj=0  {0} ∪ N and x, y ∈ Cd we have These formulas imply that H is neutral in (cid:0)C2d[z]<p, [· ,· ]Q(cid:1). Indeed, for k, m ∈ hH(cid:0)zkx(cid:1), H(cid:0)zmy(cid:1)iQ Xj+k=n and the last expression equals 0 because the assumption (a) is equivalent to k−jx = Xi+j=p−1−m+k k = 0 for all n ∈ {0, . . . , 2p}. min{k,m}Xj=0 y∗Pp−1−m+j Q−1P ∗ y∗PiQ−1P ∗ j x Pj Q−1P ∗ i,j∈{0,...,p} = j,k∈{0,...,p} (4.9) From Since, by (b), P0 = P(0) has full rank, H is degree preserving and hence injective. Therefore, dim H = dp = (1/2) dim(cid:0)C2d[z]<p(cid:1) and H is maximal neutral. subspace of(cid:0)C2d[z]<p, [· ,· ]Q(cid:1). The proof of the lemma is complete if we show that Define the mapping R : C2d[z]<p → C2d[z]<p by (Rf )(z) = zp−1f (1/z). Then R is unitary with respect to [· ,· ]Q and hence N := RH is also a maximal neutral N ⊆ Lp. For that we consider the polynomials of the form 2p−1Xk=0 zkw∗kx, w ∈ C, x ∈ Cd. z∗(cid:19)∗ 2p−1Xk=0 zkw∗kx! = P(cid:18) 1 j−1Xk=0 2p−1Xk=0 H 2p−1Xk=0 zkw∗kx! = zkw∗k! P ∗ 2p−1Xk=0 2p−1−jXk=p 2p−1Xj=0 1 p−1Xk=0 zkw∗kP(w)∗x + higher order terms. w∗k zj−k + w∗j Since the space Cd[z]<2p is spanned by polynomials in (4.9), each element of Cd[z]<p is also a sum of polynomials in (4.9). As H is degree preserving, the polynomials in H = H(cid:0)Cd[z]<p(cid:1) have degrees < p and therefore they have the form (4.10) with zero higher order terms. Thus p−1Xk=0 zkw∗(k+j) zkw∗k = zkw∗k + we obtain (4.10) 1 zj and zj j x H ⊂ span(p−1Xk=0 zkw∗kP(w)∗x : w ∈ C, x ∈ Cd) and N = RH ⊆ Lp. This proves Lemma 4.2. (cid:3) Lp(z, w) = i (zp − w∗p)K(z, w) = p−1Xk=0 Then Bp = span(P(z)Q−1 and (cid:2)P(z)Q−1f, P(z)Q−1g(cid:3)Bp where z, w ∈ C. zp−1−kw∗k!P(z)Q−1P(w)∗, zp−1−kw∗kP(w)∗x : w ∈ C, x ∈ Cd) p−1Xk=0 p−1Xk=0(cid:16)v∗(p−1−k)P(v)∗y(cid:17)∗ p−1Xk=0 zp−1−kv∗kP(v)∗y. g(z) = = Q−1(cid:16)w∗kP(w)∗x(cid:17), f (z) = zp−1−kw∗kP(w)∗x, p−1Xk=0 ON THE REPRODUCING KERNEL 21 Proof of the only if statements in Theorem 1.1. Assume that the reproducing ker- nel of B is a polynomial Nevanlinna kernel K(z, w): i (z − w∗)K(z, w) = P(z)Q−1P(w)∗ (4.11) where Q is a self-adjoint 2d × 2d matrix with d positive and d negative eigenvalues and P(z) is a d × 2d matrix polynomial with rankP(z) = d for some z ∈ C. Note that (4.11) implies (a) of Lemma 4.2: z, w ∈ C, for all (4.12) P(z)Q−1P(z∗)∗ = 0 for all z ∈ C. We prove (A) and (B) in the steps (i) -- (iv), in step (v) we prove the last only if statement in the theorem. (i) In this step we prove (B) under the assumption that (b) and (c) of Lemma 4.2 hold. Denote by Bp the reproducing kernel Pontryagin space with kernel Comparing this inner product with the one defined in Lemma 4.2, we find that P(z)Q−1 considered as a multiplication operator maps Lp ⊂ C2d[z]<p isometri- cally onto Bp and its null space is L⊥ p (see the second of the three equivalences in the beginning of the proof of Lemma 4.2). Hence, dim Bp = 2(σ1 + ··· + σd) and the positive and the negative index of Bp equal σ1 + ··· + σd. According to Lemma 2.7, we have dimB = σ1 +··· + σd. The space B is spanned by the columns of K(z, w), w ∈ C and for j ∈ {1, . . . , d} the degree of the j-th row of K(z, w) as Since both spaces have dimension σ1 + ··· + σd, equality prevails: a polynomial in z is equal to max(cid:8)0, σj − 1(cid:9). Therefore B ⊆Ld j=1(cid:0)C[z]<σj(cid:1)ed,k. (4.13) B0 = This implies (B). dMj=1(cid:0)C[z]<σj(cid:1)ed,k. (ii) In this step we prove (A) under the assumption that (b) and (c) of Lemma 4.2 hold. Set M (z) = z−σ1 ... 0 ··· . . . ··· 0 ... z−σd  . 22 B. ´CURGUS AND A. DIJKSMA Then P∞ = limz→∞ M (z)P(z) and by (4.12) we have (4.14) P∞Q−1P∗ ∞ = lim z→∞ M (z)P(z)Q−1P(z∗)∗D(z∗)∗ = 0. ∞ is Since P∞ has full rank, (4.14) implies that the linear span of the columns of P∗ a maximal neutral subspace of(cid:0)C2d, [· ,· ]Q(cid:1) and this span coincides with the null P(z)a ∈ B ⇔ P∞a = 0. space of P∞Q−1. We claim that for a ∈ C2d (4.15) To prove the claim assume first that P(z)a ∈ B. From (4.13) we see that the degree of the j-th entry of the vector polynomial P(z)a is strictly less than σj, j ∈ notice that by the definition of P∞ the row degrees of the matrix polynomial P0(z) = P(z) − M (z)−1P∞ are strictly less than σj, j ∈ {1. . . . , d}. By (4.13) we have that P0(z)a ∈ B for all a ∈ C2d. Now assume P∞a = 0. Then {1, . . . , d}. Hence P∞a = limz→∞ M (z)(cid:0)P(z)a(cid:1) = 0. As to the converse, first P(z)a = P0(z)a + M (z)−1P∞a = P0(z)a ∈ B. This completes the proof of (4.15). Consider f ∈ B. Since B is finite dimensional it can be written as (4.16) f (z) = K(z, wi)xi, m ∈ N, wi ∈ C, xi ∈ Cd, i ∈ {1, . . . , m}. The next sequence of equivalences follows from (4.15) and the observation after (4.14): mXi=1 f ∈ dom SB ⇔ Sf ∈ B (cid:2)Sf, g(cid:3)B −(cid:2)f, Sg(cid:3)B = nXj=1 mXi=1 vjy∗ mXi=1 j K(vj, wi)xi − mXi=1 nXj=1 wix∗ ∗ i K(wi, vj)yj i )K(z, wi)xi ∈ B ⇔ mXi=1 (z − w∗ ⇔ P(z)Q−1 mXi=1 ⇔ P∞Q−1 mXi=1 ⇔ mXi=1 P(wi)∗xi = P∗ P(wi)∗xi! ∈ B P(wi)∗xi! = 0 ∞x for some x ∈ Cd. Let f ∈ B be given by (4.16) and let g ∈ B be of the form g(z) = K(z, vj)yj, n ∈ N, vj ∈ C, yj ∈ Cd, j ∈ {1, . . . , n}. Assume that f, g ∈ dom SB. Then there exist x, y ∈ Cd such that P(vi)∗yi = P∗ ∞y P(wi)∗xi = P∗ ∞x and and using the reproducing kernel property of K(z, w) we have nXj=1 mXi=1 ON THE REPRODUCING KERNEL 23 mXi=1 w∗ i y∗ j K(vj, wi)xi nXj=1 j K(vj, wi)xi = = vjy∗ j K(vj, wi)xi − nXj=1 mXi=1 mXi=1(cid:0)vj − w∗ nXj=1 i(cid:1)y∗ mXi=1 nXj=1 j P(vj )!Q−1 mXi=1 = −i nXj=1 = −i y∗ = −i y∗P∞Q−1P∗ = 0. ∞x y∗ j P(vj)Q−1P(wi)∗xi P(wi)∗xi! This proves that SB is symmetric. (iii) In this step we only assume (b) of Lemma 4.2: rankP(z) = d for all z ∈ C. Then there is a unimodular d×d matrix polynomial U (z) such that S(z) = U (z)P(z) is row reduced with ordered row degrees σ1 ≥ ··· ≥ σd. Then U is an isometry from B onto the reproducing kernel Pontryagin space C with kernel (4.17) − i According to what has already been proved in (i) S(z)Q−1S(w)∗ . z − w∗ U B = C = dMj=1(cid:0)C[z]<σj(cid:1)ed,k. Thus B = U −1C and, by Theorem 3.4, (B) holds for all α ∈ C. According to part (ii) of this proof, SUB is symmetric, hence SB = U −1SUBU is also symmetric, that is, (A) holds. (iv) Finally we prove that (A) and (B) hold if rankP(z) = d for some z ∈ C as in the beginning of this proof. In that case there exist a d × d matrix polynomial G(z) with detG(z) 6≡ 0 and a d × 2d matrix polynomial S(z) with rankS(z) = d for all z ∈ C such that P(z) = G(z)S(z) for all z ∈ C, see Lemma 2.4. If by A we denote the reproducing kernel space with Nevanlinna kernel (4.17), then, by what has been proved in (iii), the operator SA is symmetric and for almost all α ∈ C we have ran(SA − α) = A ∩ ker Eα. Now (A) and (B) follow since the multiplication operator G corresponding to G(z) is an isomorphism from A onto B. (v) The last only if statement in the theorem follows from step (iii) above and Theorem 3.4. (cid:3) Remark 4.3. Assume that B ⊂ Cd[z] is a Pontryagin space which satisfies the conditions (A) and (B) of Theorem 1.1. Then, by Theorem 1.1, there is a gen- eralized Nevanlinna pair {M (z), N (z)} such that the d × 2d matrix polynomial P(z) = [M (z)N (z)] provides a representation (2.7), with Q given by (2.8), for the Nevanlinna reproducing kernel K(z, w) of B. In addition, by Theorem 1.2 there is a canonical subspace C such that B = W C for some d × d matrix polynomial W (z) with detW (z) 6≡ 0. The proof of Theorem 1.1 and Lemma 2.4 show that the multiset of the Forney indices of P(z) coincides with the multiset of the degrees of 24 B. ´CURGUS AND A. DIJKSMA C. This implies that the Forney indices are independent of the Nevanlinna repre- sentation (2.7) of the kernel K(z, w). In the special case when the defect numbers of SB are equal to d this fact can also be proved directly by using [23, Theorem 1.3]. In view of Remark 4.1, this remark substantiates the observations about the Forney indices and the defect numbers after Theorem 1.2 in the Introduction. 5. Q-functions 5.1. Let M (z) be a generalized Nevanlinna d × d matrix function and denote by L(M ) the reproducing kernel Pontryagin space with reproducing kernel KM (z, w) = KM,Id(z, w). By [16, Theorem 2.1], the operator S in L(M ) of multiplication by the independent variable is a simple symmetric operator with equal defect numbers and its adjoint is given by S∗ = span(cid:8){KM (· , w∗)x, wKM (· , w∗)x} : x ∈ Cd, w ∈ hol(M )(cid:9) =(cid:8){f, g} ∈ L(M )2 : ∃ x, y ∈ Cd such that g(z) − zf (z) ≡ x − M (z)y(cid:9). It follows that for all w ∈ hol(M ) ker(S∗ − w) =(cid:8)KM (· , w∗)x : x ∈ Cd(cid:9) = ran E∗ w, where Ew is considered as a mapping Ew : L(M ) → Cd. Taking orthogonal com- plements we see that ran(S − α) = L(M ) ∩ ker Eα, α ∈ hol(M ). Thus (A) and (B) of Theorem 1.1 hold. Moreover, [16, Theorem 2.1] and its proof imply that there is a constant invertible d × d matrix T such that 0(cid:21) , T M (z)T ∗ = M0 +(cid:20)cM (z) 0 0 Clearly, Q(z) depends on the choice of the pair {A, Γz} and if this choice has to be mentioned explicitly we shall say that Q(z) is a Q-function for S associated with the where M0 is a constant self-adjoint d × d matrix and, if the defect numbers of S a Q-function for S. The theorem below concerns a converse implication. But first we recall the notion of a Q-function. are denoted by l, cM (z) is a generalized Nevanlinna l × l matrix function which is Let S be a simple symmetric operator in a Pontryagin space K with defect numbers equal to l. Let A be a self-adjoint extension of S in K with a nonempty resolvent set ρ(A). Let µ ∈ ρ(A) \ R and define a function Γµ : Cl → K such that it is a linear bijection from Cl onto ker(S∗ − µ). Finally, for z ∈ ρ(A) define the defect mappings Γz : Cl → K by Γz =(cid:0)I + (z − µ)(A − z)−1(cid:1) Γµ, K = span{Γzc : z ∈ ρ(A) ∩ (C \ R), c ∈ Cl} Then Γz is a bijection from Cl onto ker(S∗ − z), (5.1) and, by the resolvent identity, Γ∗ by definition an l × l matrix function that satisfies the equation (5.2) Q(z) − Q(w)∗ z ∈ ρ(A). wΓz = Γ∗ = Γ∗ wΓz, z, w ∈ ρ(A). z − w∗ z∗ Γw∗, w, z ∈ ρ(A). A Q-function for S is ON THE REPRODUCING KERNEL 25 pair {A, Γz}. Q(z) is uniquely determined up to an additive constant self-adjoint d × d matrix Q0: Q(z) = Q0 − iImµ Γ∗ µΓµ + (z − µ∗)Γ∗ µΓz, Q0 = Q∗ 0. From (5.1) and the defining relation (5.2) it follows that Q(z) is a generalized Nevanlinna l × l matrix function with κ negative squares where κ is the negative index of the Pontryagin space K; in particular Q(z)∗ = Q(z∗). Q-functions in an indefinite setting were introduced and studied by M.G. Krein and H. Langer in [28, 29]. 5.2. The following theorem shows that the Nevanlinna pair {M (z), N (z)} of ma- trix polynomials M (z) and N (z) in Theorem 1.1 can be chosen such that detN (z) 6≡ 0 and such that N (z)−1M (z) is essentially the Q-function for SB. As before, by L(Q) we denote the reproducing kernel space with reproducing kernel given by (5.2). Theorem 5.1. Let B be a finite dimensional Pontryagin subspace of Cd[z] for which the conditions (A) and (B) of Theorem 1.1 hold. Denote by l ∈ {1, . . . d} the equal defect numbers of the symmetric operator SB. Let Q(z) be an l × l matrix Q-function for SB. Then there is a d × d matrix polynomial N (z) with detN (z) 6≡ 0 such that M (z) = N (z)diag(Q(z), 0) is a d× d matrix polynomial and B = N (L(Q) ⊕ {0}). In particular, {M (z), N (z)} is a Nevanlinna pair of matrix polynomials and KM,N (z, w) is the reproducing kernel of B. Proof. Assume (A) and (B) of Theorem 1.1. By Theorem 1.2 there is a d × d matrix function W (z) with detW (z) 6≡ 0 such that B = W C, where C is a canonical subspace of Cd[z]. By Remark 4.1 the defect numbers of the symmetric operator SB are both equal to l with l ≤ d. We consider two cases: l = d and l < d. (i) l = d. Let Q(z) be the Q-function for SB associated with the pair {A, Γz}, where A is a self-adjoint extension of SB and the defect mappings Γz are defined above with l = d. Since SB is simple, the mapping f 7→ g with f (z) = Γ∗ z∗ Γw∗ x, g(z) = (Γw∗x) (z), x ∈ Cd, w ∈ ρ(A), can be extended by linearity to a unitary mapping U from L(Q) onto B. That U is isometric follows from (cid:2)Γ∗ wΓz x, Γ∗ vΓz y(cid:3)L(Q) = y∗Γ∗ wΓv x = y∗Γ∗ v∗ Γw∗ x = [Γw∗x, Γv∗ y]B, x, y ∈ Cd. We claim that U is the operator of multiplication by a d× d matrix function. To prove the claim we use the equality B = W C. Since the defect numbers of SB are equal to d, the degrees of C are all ≥ 1 (see Remark 4.1) and hence the d columns of W (z) belong to B and are linearly independent over C. We denote by Γ∗ z∗ W the d × d matrix function defined by z∗ W )x = Γ∗ z∗ (W x), (Γ∗ x ∈ Cd, z ∈ ρ(A). We show that its inverse exists for z ∈ Ω := ρ(A) ∩ {z ∈ C : detW (z) 6= 0}. Suppose there is an x ∈ Cd such that (Γ∗ z∗ W ) x = 0. Then for all y ∈ Cd 0 =(cid:2)(Γ∗ z∗ W ) x, y(cid:3)Cd =(cid:2)Γ∗ z∗(W x), y(cid:3)Cd =(cid:2)W x, Γz∗ y(cid:3)B, hence W x ∈ ker(S∗ follows that x = 0. This proves that (Γ∗ B − z∗)⊥ = ran(SB − z) = B ∩ Ez. That is, W (z)x = 0 and it z∗ W )−1 is well defined for all z ∈ Ω. We set 26 B. ´CURGUS AND A. DIJKSMA N (z) = W (z)(Γ∗ with multiplication by N (z) if we have proved that z∗ W )−1. Clearly, detN (z) 6≡ 0. We have shown that U coincides N (z)Γ∗ z∗Γw∗ x = (Γw∗x) (z), x ∈ Cd, w ∈ ρ(A), z ∈ Ω, or, equivalently, that with y(z, w, x) := W (z)−1 (Γw∗x) (z) ∈ Cd Γ∗ z∗ Γw∗ x = Γ∗ z∗ (W y(z, w, x)) , x ∈ Cd, w ∈ ρ(A), z ∈ Ω. But this equality holds, since Γ∗ z∗ (ran(SB − z)) = {0} and Γw∗ x − W y(z, w, x) ∈ B ∩ Ez = ran(SB − z). This completes the proof of the claim that U is multiplication by N (z). It follows from [6, Theorem 1.5.7] and its proof that the formula for the kernel K(z, w) of B is given by K(z, w) = N (z) Q(z) − Q(w)∗ N (w)∗ z − w∗ and hence B = NL(Q). It remains to show that M (z) = N (z)Q(z) and N (z) are matrix polynomials. Since the elements of the space B are polynomials, the matrix function z 7→ K(z, w) is a matrix polynomial, hence the matrix function M (z) − N (z)Q(w)∗ = N (z)Q(z) − N (z)Q(w)∗ = (z − w∗)K(z, w)N (w)−∗ is a matrix polynomial in z. Thus if N (z) is a matrix polynomial, then so is M (z). It remains to show that N (z) is a polynomial. For this we note that the above formula implies that for x ∈ Cd N (z) (z − µ∗)K(z, µ)N (µ)−∗ − (z − w∗)K(z, w)N (w)−∗ Q(µ∗) − Q(w)∗ x = x. µ∗ − w∗ µ∗ − w∗ The right hand side is a matrix polynomial in z and hence it follows from the equality that N (z) is a matrix polynomial if we can show that To prove this equality we argue by contradiction and suppose it is not true. Then there is a nonzero vector x ∈ Cd orthogonal to the set on the right hand side, that is, µΓw∗ x : w ∈ ρ(A) ∩ (C \ R), x ∈ Cd(cid:9). Cd = span(cid:8)Γ∗ (cid:2)Γw∗ y, Γµx(cid:3)B = 0, w ∈ ρ(A) ∩ (C \ R), y ∈ Cd. Since SB is simple and Γµ is injective, we find that Γµx = 0 and that x = 0, which contradicts the choice of the nonzero vector x. (ii) l < d. Then C = C1 ⊕ {0}, where C1 is a canonical subspace of Cl[z] of which the degrees are all ≥ 1. Using the relation B = W (C1 ⊕ {0}) we equip C1 with an indefinite inner product that makes W an isomorphism. Then SB and SC1 are isomorphic: SC1 = W SBW −1, hence SC1 is symmetric and has defect numbers equal to l. Thus (A) holds and it is not difficult to verify that also (B) holds on C1. Finally, since Q(z) is the Q-function for SB associated with the pair {A, Γz}, Q(z) is the Q-function for SC1 associated with the pair {W −1AW, W −1Γz}. This all shows that we may apply part (i) of this proof (with W (z) = Il): There exists an l× l matrix polynomial N1(z) with detN1(z) 6≡ 0 such that N1(z)Q(z) is an l× l matrix polynomial and C1 = N1L(Q). It follows that if N (z) = W (z) diag(N1(z), Id−l), ON THE REPRODUCING KERNEL 27 then detN (z) 6≡ 0, N (z) diag(Q(z), 0) is a d × d matrix polynomial and B = N (L(Q) ⊕ {0}). (cid:3) 6. Corollaries and examples In the next corollary we extend Theorem 1.1 to finite dimensional Pontryagin spaces of rational vector functions. A rational Nevanlinna kernel is a kernel of the form KM,N (z, w) as in (1.6), in which M (z) and N (z) are rational d × d matrix functions satisfying (1.4) and (1.5). Corollary 6.1. Let B be a finite dimensional Pontryagin space of rational d × 1 vector functions and let Ω ⊂ C be the finite set of all the poles of the functions in B. Denote by SB the operator of multiplication by the independent variable in B and by Eα the operator of evaluation at a point α ∈ C. Then the reproducing kernel of B is a rational Nevanlinna kernel if and only if the following two conditions hold: (a) The operator SB is symmetric in B. (b) For some α ∈ C \ Ω we have ran(cid:0)SB − α(cid:1) = B ∩ ker Eα. Proof. Assume (a) and (b). Let q(z) be the monic scalar polynomial of minimal degree such that B′ := {q(z)f (z) : f ∈ B} consists of polynomials. Equip B′ with the Pontryagin space inner product that makes the mapping q : B → B′ of multiplication by q(z) a unitary mapping. Then items (A) and (B) of Theorem 1.1 hold for B′. Hence B′ has a polynomial reproducing Nevanlinna kernel KM,N (z, w). It follows that B has reproducing kernel KM/q,N/q(z, w). Now assume KM,N (z, w) is a rational reproducing Nevanlinna kernel of B. Let r(z) be a polynomial such that r(z)M (z) and r(z)N (z) are polynomials and hence form a polynomial Nevanlinna pair {r(z)M (z), r(z)N (z)}. Then KrM,rN (z, w) is a polynomial reproducing Nevanlinna kernel of the space B′′ := {r(z)f (z) : f (z) ∈ B} equipped with the inner product that makes multiplication by r(z) an isomor- phism from B onto B′′. Since the elements of B′′ are polynomials, we can apply Theorem 1.1 to conclude that items (A) and (B) hold for the space B′′. Since multiplication by r(z) and by z commute, (A) implies (a). By Theorem 1.2 and (1.3), the equality (6.1) holds for all but finitely many α ∈ C. Choose α ∈ C \ Ω such that (6.1) is valid. Then for this α item (b) holds. ran(cid:0)SB′′ − α(cid:1) = B′′ ∩ ker Eα (cid:3) Corollary 6.2. Let (B, [· , · ]B) be a finite dimensional Pontryagin subspace of Cd[z] whose reproducing kernel is a Nevanlinna kernel determined by a generalized Nevanlinna pair. Let J be a fundamental symmetry on B. Then the Hilbert space (B, [J · , · ]B) has a reproducing Nevanlinna kernel determined by a Nevanlinna pair if and only if SB is symmetric in this space. The corollary follows from Theorem 1.1, because condition (B) is independent of the topology on B. Example 6.3. Consider the subspace B of C2[z] spanned by the columns of the matrix B(z) =(cid:20)1 0 0 1(cid:21) z 0 z2 0 28 B. ´CURGUS AND A. DIJKSMA and equipped with the inner product [· , · ]B so that G = 1 0 0 0 0 1 0 0 0 0 0 0 0 1 1 0   and J =   . is the Gram matrix associated with B(z): G = [B, B]B. The spectral decomposition of G is G = U JU ∗ with unitary matrix 0 √2 0 0 −1 1 0 0 0 0 √2 0 0 1 1 0 −1 0 0 0 1 0 0 1 0 0 0 0 1 √2 0 0 0 1 U = It follows that (B, [· , · ]B) is a Pontryagin space with positive index 3 and negative index 1. The equality [BU, BU ]B = J defines a fundamental decomposition of B with corresponding fundamental symmetry J determined by J BU = BU J. In the Hilbert space inner product [· , · ]J := [J · , · ]B we have [BU, BU ]J = J 2 = In and hence [B, B]J = In. The operator SB is symmetric in the Pontryagin space (B, [· , · ]B), but not in the Hilbert space (B, [· , · ]J ). Since B is a canonical subspace of C2[z], Theorem 1.2 implies that Theorem 1.1 (B) holds in B. Hence, according to Theorem 1.1, the Pontryagin space (B, [· , · ]B) has a reproducing Nevanlinna kernel, whereas the reproducing Hilbert space (B, [· , · ]J ) does not have a reproducing Nevanlinna kernel. (cid:3) Corollary 6.4. Let B be a finite dimensional Pontryagin subspace of Cd[z] whose reproducing kernel is a Nevanlinna kernel. Let B0 be a Pontryagin subspace of B. Then the reproducing kernel of B0 is a Nevanlinna kernel if and only if for some α ∈ C we have ran(cid:0)SB0 − α(cid:1) = B0 ∩ ker Eα. The corollary follows from Theorem 1.1, because the hypothesis implies that SB0, being a subset of SB, is symmetric in B0, that is, that (A) holds for SB0. Example 6.5. Consider the Hilbert subspace B0 = span(cid:26)(cid:20)1 0(cid:21) ,(cid:20)z2 0(cid:21)(cid:27) of the space B in Example 6.3. Then for arbitrary α ∈ C we have ran(cid:0)SB0 − α(cid:1) = {0} and B0 ∩ ker Eα = span(cid:26)(cid:20)z2 − α2 Thus the condition ran(cid:0)SB0 −α(cid:1) = B0∩ker Eα does not hold for any α ∈ C. Hence Corollary 6.4 implies that the reproducing kernel of B0, which is calculated to be (cid:21)(cid:27) . 0 K(z, w) ="1 + z2w∗2 0 0 0# , z, w ∈ C, is not a Nevanlinna kernel. This fact can be verified using [23, Theorem 1.3]. First observe that for all z, w ∈ C we have (z − w∗)K(z, w) = M (z)N (w)∗ − N (z)M (w)∗ =(cid:2)M (z) N (z)(cid:3)" N (w)∗ −M (w)∗# , ON THE REPRODUCING KERNEL 29 where N (z) =(cid:20)1 0 z2 0(cid:21) and M (z) = zN (z), z ∈ C. Now [23, Theorem 1.3 and Section 4] imply that for any 2 × 2 matrix polynomials M1(z) and N1(z) such that (6.2) there exists a 4 × 4 invertible matrix S such that (z − w∗)K(z, w) =(cid:2)M1(z) N1(z)(cid:3)" N1(w)∗ −M1(w)∗# , (cid:2)M1(z) N1(z)(cid:3) =(cid:2)M (z) N (z)(cid:3)S, z ∈ C. z, w ∈ C, Hence (6.2) yields that rank(cid:2)M1(z) N1(z)(cid:3) = 1 for all z ∈ C. Consequently K(z, w) is not a Nevanlinna kernel. Using the same results from [23] one can also show that the scalar reproducing kernel K(z, w) = 1 + z2w∗2 of the Hilbert space with orthonormal basis {1, z2} is not a Nevanlinna kernel. ✷ We end the paper with two examples in which det K(z, w) ≡ 0. These examples also show that the proof of Theorem 5.1 is constructive. Example 6.6. Consider the space B with reproducing kernel K(z, w) = −1 0 0 −w∗ 0 0 −1 −z 0  . We show that, even though detK(z, w) ≡ 0, the kernel is a Nevanlinna kernel. We follow the proof of the first part of Theorem 5.1 and construct two Nevanlinna pairs that determine K(z, w). The space B is spanned by the columns of the 3× 4 matrix polynomial It follows that B is a canonical subspace of C3[z] with degrees 1, 1 and 2. The Gram matrix associated with B(z) is given by hence B is a Pontryagin space with positive and negative index 2. The operator of multiplication by z on B is given by It is easy to see that (A) and (B) of Theorem 1.1 are satisfied. The defect numbers of SB are both equal to 3, see Remark 4.1. The Q-function of SB associated with the self-adjoint extension 0 0 0 0 1 −I 1 0 0 1 0 0 B(z) = z . G = [B, B]B =(cid:20) 0 −I 0 (cid:21) ,  : a ∈ C SB =   , B  B : a, b, c, d ∈ C   A =   , B B a 0 b 0 0 0 a 0 0 0 0 a 0 c 0 d . 30 B. ´CURGUS AND A. DIJKSMA is of SB (which is multi-valued and has a nonempty resolvent set) and the defect mappings Γz : C3 → ker(S∗ − z) defined by Γz =(cid:0)I + (z − i)(A − z)−1(cid:1) Γi = B Q(z) = Q0 + 0 1/z −iz 1/z 0 0 iz 0 i/z i 0 0 0 0 i/z 0 0 0 0 1  0. 0 , Q0 = Q∗ z2 and Y (z) = 0 0 0 −iz 0 0 −1 0 −iz 0 0  X(z) = Y (z)Q0 + −i 0 iz 0 0 −i We find that X(z) = Y (z)Q(z) and Y (z) = (Γ∗ linna pair of matrix polynomials: z∗ I)−1 form a full generalized Nevan- such that K(z, w) = KX,Y (z, w) and (cid:2)X(z) Y (z)(cid:3) is row reduced with Forney indices 1, 1 and 2, which is in accordance with Remark 4.3. The Q-function associated with the self-adjoint operator extension A of SB and defect mappings Γz defined by AB = B 0 0 0 0 1 0 0 0 0 0 0 1 0 0 0 0  , Γz = B −1/z2 −1/z 0 0 0 0 i/z −(z − i)/z2 0 0 0 i/z  , Q1 = Q∗ 1. is given by 0 0 0 z2 −i z z2 i z iz2 + z + i Q(z) = Q1 + −iz2 + z − i Again we find that M (z) = N (z)Q(z) and N (z) = (Γ∗   M (z) = N (z)Q1 + 0 −iz2 + z − i −iz , N (z) = (cid:2)M (z) N (z)(cid:3) is row reduced with Forney indices 1, 1 and 2. 0 0 0 0 z 1 0 0 which form a full generalized Nevanlinna pair such that K(z, w) = KM,N (z, w) and (cid:3) Two generalized Nevanlinna pairs {X(z), Y (z)} and {M (z), N (z)} of d×d matrix polynomials define the same Nevanlinna kernel if one is a J-unitary transformation of the other, that is, if they are connected via the formulas z∗ I)−1 are matrix polynomials: 0 −iz 0 0 z2 0 z + i −iz 0  M (z) = X(z)A + Y (z)C, N (z) = X(z)B + Y (z)D, where A, B, C and D are constant d × d matrices such that if we set U =(cid:20)A B C D(cid:21) and J =(cid:20) 0 −iId iId 0(cid:21) , ON THE REPRODUCING KERNEL 31 then U is J-unitary: U JU ∗ = J. The pairs {X(z), Y (z)} and {M (z), N (z)} in Ex- ample 6.6 (with arbitrary constant self-adjoint matrices Q0 and Q1) are connected via a J-unitary transformation. The following example shows that the converse of the foregoing statement does not hold. Example 6.7. Consider the Nevanlinnna pair {X(z), Y (z)} given by X(z) = diag(z, 0, z2), Y (z) = diag(0, z, z). it is spanned by B(z) = (cid:2)0 Then the space B with reproducing kernel KX,Y (z, w) = diag(0, 0, zw∗) is a 1- dimensional Hilbert space: and the corre- sponding Gram matrix is G = [B, B]B = 1. Note that detX(z) ≡ 0, detY (z) ≡ 0 that the pair {X(z), Y (z)} can be replaced by a Nevanlinna pair {M (z), N (z)} such that detN (z) 6≡ 0. Since P(z) does not have full rank for all z ∈ C, to calculate the Forney indices we must first apply Lemma 2.4. We write P(z) as P(z) = G(z)T(z) with and P(z) :=(cid:2)X(z) Y (z)(cid:3) does not have full rank at z = 0: P(0) = 0. We show 0 z(cid:3)⊤ G(z) = 0 0 z 0 z z 0 0 0 , T(z) = 0 0 1 0 z 0 0 0 0 0 0 0 0 1 0 1 0 0 . Since T(z) has full rank for all z ∈ C and is row reduced, the Forney indices of P(z) are those of T(z) and they are µ1 = 1, µ2 = µ3 = 0. This fits in well with the observations after Theorem 1.2 indicating that dim B = 1 and the defect numbers of the symmetric operator SB = {{0, 0}} are both equal to 1. We follow part (ii) of the proof of Theorem 5.1 and write B = W (C1 ⊕ {0}) with W (z) = 0 0 z 0 1 0 1 0 0 and C1 = C. We make the multiplication operator W an isometry when C is equipped with the Euclidean inner product. Then SC1 = {{0, 0}} is symmetric. The defect subspaces ker(S∗ C1 − z) all coincide with C and A is a self-adjoint extension of SC1 if and only if A = Am, the operator of multiplication by m, m ∈ R, or and relations have a non-empty resolvent set. Choose µ ∈ C \ R, γ ∈ C \ {0} and define Γµ : C → ker(S∗ C1 − µ) = C by Γµx = γx, x ∈ C. Then the Q-function q(z) of SC1 associated with {A, Γz} is given by A = Arel =(cid:8){0, c} : c ∈ C(cid:9). Since C1 is a Hilbert space, all self-adjoint operators where q0 is an arbitrary real number, and q(z) = q0 + z∗ 1)−1 = (Γ∗ m − µ2γ2 m − z γ2z if A = Am, if A = Arel, m − z γ∗(m − µ∗) 1 γ∗ if A = Am, if A = Arel. 32 B. ´CURGUS AND A. DIJKSMA We find that KX,Y (z, w) = KM,N (z, w) with matrix polynomials M (z) = N (z) q(z) 0 0 0 0 0 0 0 0 and N (z) = W (z) (Γ∗ z∗ 1)−1 0 0 0 0 1 0 0 1  . It is easy to see that the Nevalinna pairs {X(z), Y (z)} and {M (z), N (z)} are not related via a J-unitary transformation. (cid:3) References [1] D. Alpay, handwritten manuscript (French), unpublished. [2] D. Alpay, A structure theorem for reproducing kernel Pontryagin spaces, J. Comput. Appl. Math. 99 (1998) 413 -- 422. [3] D. Alpay, The Schur algorithm, reproducing kernel spaces and system theory, SMF/AMS Texts and Monographs 5, Panorama et Synth`eses 6, 1998. [4] D. Alpay, P. Bruinsma, A. Dijksma, H.S.V. de Snoo, Interpolation problems, extensions of symmetric operators and reproducing kernel spaces I, Oper. Theory Adv. Appl. 50, Birkhauser, Basel, 1991, pp. 35 -- 82. [5] D. Alpay, P. Bruinsma, A. Dijksma, H.S.V. de Snoo, Interpolation problems, extensions of symmetric operators and reproducing kernel spaces II, Integral Equations Operator Theory 14 (1991) 465 -- 500; missing Section 3, Integral Equations Operator Theory 15 (1991) 378 -- 388. [6] D. Alpay, A. Dijksma, J. Rovnyak, H.S.V. de Snoo, Schur functions, operator colligations, and reproducing kernel Pontryagin spaces, Oper. Theory Adv. Appl. 96, Birkhauser, Basel, 1997. [7] D. Alpay, H. Dym, On applications of reproducing kernel spaces to the Schur algorithm and rational J unitary factorization, Oper. Theory Adv. Appl. 18, Birkhauser, Basel, 1986, pp. 89 -- 159. [8] A. Amirshadyan, V. Derkach, Interpolation in generalized Nevanlinna and Stieltjes classes, J. Operator Theory 42 (1999) 145 -- 188. [9] T.Ya. Azizov, B. ´Curgus, A. Dijksma, Standard symmetric operators in Pontryagin spaces: A generalized von Neuman formula and minimality of boundary coefficients, J. Funct. Anal. 198 (2003) 361 -- 412. [10] T.Ya. Azizov, B. ´Curgus, A. Dijksma, Eigenvalue problems with boundary conditions de- pending polynomially on the eigenparameter (tentative title), in preparation. [11] J. Behrndt, V.A. Derkach, S. Hassi, H.S.V. de Snoo, A realization theorem for generalized Nevanlinna families, to appear in Oper. Matrices. [12] L. de Branges, Hilbert spaces of entire functions, Prentice Hall, Englewood Cliffs, 1968. [13] L. de Branges, The expansion theorem for Hilbert spaces of entire functions, in: Entire Functions and Related Parts of Analysis (Proc. Sympos. Pure Math., La Jolla, Calif., 1966), Amer. Math. Soc., Providence, 1968, pp. 79 -- 148. [14] L. de Branges, J. Rovnyak, Canonical models in quantum scattering theory, in: Perturbation Theory and its Applications in Quantum Mechanics (Proc. Adv. Sem. Math. Res. Center, U.S. Army, Theoret. Chem. Inst., Univ. of Wisconsin, Madison, Wis., 1965) Wiley, New York, 1966, pp. 295 -- 392. [15] B. ´Curgus, A. Dijksma, T.T. Read, The linearization of boundary eigenvalue problems and reproducing kernel Hilbert spaces, Linear Algebra Appl. 329 (2001) 97 -- 136. [16] A. Dijksma, H. Langer, A. Luger, Y. Shondin, Minimal realizations of scalar generalized Nevanlinna functions related to their basic factorization, Oper. Theory Adv. Appl. 154, Birkhauser, Basel, 2004, pp. 69 -- 90. [17] A. Dijksma, H. Langer, H.S.V. de Snoo, Selfadjoint Πκ-extensions of symmetric subspaces: an abstract approach to boundary problems with spectral parameter in the boundary condi- tions, Integral Equations Operator Theory 7 (1984) 459 -- 515; Addendum, Integral Equations Operator Theory 7 (1984) 905. [18] A. Dijksma, H. Langer, H.S.V. de Snoo, Hamiltonian systems with eigenvalue depending boundary conditions, Oper. Theory Adv. Appl. 35, Birkhauser, Basel, 1988, pp. 37 -- 83. [19] A. Dijksma, H.S.V. de Snoo, Symmetric and selfadjoint relations in Krein spaces I, Oper. Theory Adv. Appl. 24, Birkhauser, Basel, 1987, pp. 145 -- 166. ON THE REPRODUCING KERNEL 33 [20] H. Dym, D. Volok, Zero distribution of matrix ploynomials, Linear Algebra Appl. 425 (2007) 714 -- 738. [21] P. Fuhrmann, On duality in some problems of geometric control. Acta Appl. Math. 91 (2006) 207 -- 251. [22] F.R. Gantmacher, Matrizentheorie, VEB Deutscher Verlag der Wissenschaften, Berlin, 1986. [23] I. Gohberg, T. Shalom, On Bezoutians of nonsquare matrix polynomials and inversion of matrices with nonsquare blocks, Linear Algebra Appl. 137-138 (1990) 249 -- 323. [24] I.S. Iohvidov, M.G. Krein, H. Langer, Introduction to the spectral theory of operators in spaces with an indefinite metric, Akademie-Verlag, Berlin, 1982. [25] T. Kailath, Linear systems, Prentice-Hall, Englewood Cliffs, 1980. [26] M. Kaltenback, H. Woracek, Pontryagin spaces of entire functions I, Integral Equations Op- erator Theory 33 (1999) 34 -- 64. [27] M. Kaltenback, H. Woracek, Pontryagin spaces of entire functions V, Acta Sci. Math. (Szeged) 77 (2011) 179 -- 292. [28] M.G. Krein, H. Langer, Defect subspaces and generalized resolvents of an Hermitian operator in the space Πκ, Funct. Anal. Appl. 5 (1971) 136 -- 146; ibid 5 (1971) 217 -- 228. [29] M.G. Krein, H. Langer, Uber die Q-Funktion eines π-hermiteschen Operators im Raume Πκ, Acta Sci. Math. (Szeged) 34 (1973) 191 -- 230. [30] M.G. Krein, H. Langer, Uber einige Fortsetzungsprobleme, die eng mit der Theorie her- mitescher Operatoren im Raume Πκ zusammenhangen. I. Einige Funktionenklassen und ihre Darstellungen, Math. Nachr. 77 (1977) 187 -- 236. [31] M.G. Krein, H. Langer, Some propositions on analytic matrix functions to the theory of operators in the space Πκ, Acta. Sci. Math. (Szeged) 43 (1981) 181 -- 205. [32] H. Elton Lacey, The Hamel dimension of any infinite dimensional separable Banach space is c, Amer. Math. Monthly 80 (3) (1973) 298. [33] R. J. McEliece, The algebraic theory of convolutional codes. Handbook of coding theory, Vol. I, II, North-Holland, Amsterdam, 1998, pp. 1065 -- 1138. [34] Xian-Jin Li, The Riemann hypothesis for polynomials orthogonal on the unit circle, Math. Nachr. 166 (1994) 229 -- 258. [35] Xian-Jin Li, On Reproducing Kernel Hilbert Spaces of Polynomials, Math. Nachr. 186 (1997) 115 -- 148. Department of Mathematics, Western Washington University, Bellingham, WA 98225, USA E-mail address: [email protected] Johann Bernoulli Institute of Mathematics and Computer Science, University of Groningen, P.O. Box 407, 9700 AK Groningen, The Netherlands E-mail address: [email protected]
1609.01083	1	1609	2016-09-05T10:26:42	Approaching bilinear multipliers via a functional calculus	[ "math.FA" ]	We propose a framework for bilinear multiplier operators defined via the (bivariate) spectral theorem. Under this framework we prove Coifman-Meyer type multiplier theorems and fractional Leibniz rules. Our theory applies to bilinear multipliers associated with the discrete Laplacian on $\mathbb{Z}^d,$ general bi-radial bilinear Dunkl multipliers, and to bilinear multipliers associated with the Jacobi expansions.	math.FA	math	APPROACHING BILINEAR MULTIPLIERS VIA A FUNCTIONAL CALCULUS BŁAŻEJ WRÓBEL Abstract. We propose a framework for bilinear multiplier operators defined via the (bi- variate) spectral theorem. Under this framework we prove Coifman-Meyer type multiplier theorems and fractional Leibniz rules. Our theory applies to bilinear multipliers associated with the discrete Laplacian on Zd, general bi-radial bilinear Dunkl multipliers, and to bilinear multipliers associated with the Jacobi expansions. 6 1 0 2 p e S 5 ] . A F h t a m [ 1 v 3 8 0 1 0 . 9 0 6 1 : v i X r a The theory of spectral multipliers is now a well established and vast branch of linear harmonic analysis. Its origins lie in trying to extend the Fourier multiplier operators on R given by 1. Introduction f 7→ 1 2π ZR m(ξ) f (ξ)eixξ dξ, x ∈ R, to other settings. Here m is a bounded function on R while f (ξ) = RR f (x)e−ixξ dx, ξ ∈ R. For a self-adjoint operator L its spectral multipliers are the operators m(L) defined by the spectral theorem. In the Fourier case L is merely i d dx . As in the Fourier case the boundedness of m(L) on L2 is equivalent with the boundedness of m. The main task in the theory of spectral multipliers is to extend the boundedness of m(L) to Lp, for some 1 < p < ∞, p 6= 2. The bilinear multipliers for the Fourier transform are the operators (1.1) Fm(f1, f2)(x) = m(ξ1, ξ2) f1(ξ1) f2(ξ2)eix(ξ1+ξ2) dξ, x ∈ R, 1 4π2 ZZR2 with m : R2 → C being a bounded function. As far as we know, in the bilinear case, there has been no systematic approach to extend the operators Fm outside of the Fourier transform setting. The main idea behind the creation of this paper is to provide a theory for bilinear mul- tipliers defined by the (bivariate) spectral theorem that parallels the correspondence between the linear Fourier multipliers and spectral multipliers. Our starting point is the observation that (1.1) may be rephrased as Fm(f1, f2)(x) = m(i∂1, i∂2)(f1 ⊗ f2)(x, x), x ∈ R; here ∂1, ∂2 denote the partial derivatives, while m(i∂1, i∂2) is defined by the bi-variate spectral theorem. Note that ∂1 = ∂ ⊗ I and ∂2 = I ⊗ ∂, where ∂ denotes the derivative on R, while I is the identity operator. We investigate the possibility of replacing i∂1 and i∂2 by some other operators L1 = L ⊗ I and L2 = I ⊗ L. The bilinear multipliers we consider are of the form (1.2) Bm(f1, f2)(x) = m(L1, L2)(f1 ⊗ f2)(x, x), x ∈ X. 2010 Mathematics Subject Classification. 42B15, 47B38, 26D10. Key words and phrases. bilinear multiplier, joint spectral theorem, fractional Leibniz rule. 1 2 BŁAŻEJ WRÓBEL Here L is a self-adjoint non-negative operator on L2(X, ν), and m(L1, L2) is defined by the bi-variate spectral theorem. We also assume that L is injective on its domain, and that the contractivity condition (CT) (see p. 4) and the well definiteness condition (WD) (see p. 5) are satisfied. These assumptions should be regarded as technical ones. The main assumptions on L that are in force in this paper are the existence of a Mikhlin-Hörmander functional calculus (MH), see p. 4, together with a product formula for the spectral multipliers of L, see (PF) on p. 6. There are two main goals of our paper. Firstly, we would like to prove Coifman-Meyer type multiplier theorems outside of the Fourier transform setting. Secondly, we would like to apply these results to obtain fractional Leibniz rules. The classical Coifman-Meyer multiplier theorem [7] says that the Mikhlin-Hörmander con- dition supξ∈R2 ξα1+α2∂αm(ξ) ≤ Cα, α ∈ N2, implies the boundedness of Fm from Lp1 × Lp2 to Lp, 1/p = 1/p1 + 1/p2, p1 > 1, p2 > 1, p > 1/2. This was proved by Coifman and Meyer for p > 1, while for p > 1/2 it is due to Grafakos and Torres [12] and Kenig and Stein [15]. There are also Coifman-Meyer type multiplier theorems which are known in settings other than the Fourier transform. For bilinear multipliers on the torus a theorem of Coifman-Meyer type may be deduced from Fan and Sato [9, Theorems 1-3]. Similarly, for bilinear multipliers on the integers such a theorem follows from Blasco [5, Theorem 3.4]. Next, in the product Dunkl setting a Coifman-Meyer type multiplier theorem was proved by Amri, Gasmi, and Sifi [3]. The main result of this paper is the following generalized Coifman-Meyer type theorem. Theorem (Theorem 2.3). Let m : (0,∞)2 → C satisfy the Hörmander's condition λα1+α2 ∂αm(λ) ≤ Cα, λ ∈ (0,∞)2, for sufficiently many multi-indices α ∈ N2. Then Bm given by (1.2) is bounded from Lp1(X) × Lp2(X) to Lp(X), where 1/p1 + 1/p2 = 1/p, with p1, p2, p > 1. Theorem 2.3 is formally stated and proved in Section 2. The main difficulty in obtaining the theorem lies in finding an appropriate proof of the classical Coifman-Meyer multiplier theorem, which is prone to modifications towards our setting. The proof we present in Section 2 follows the scheme by Muscalu and Schlag [18, pp. 67-71]. An important ingredient in our proof is a spectrally defined Littlewood-Paley theory. For this method to work the assumption (PF) is very useful. It might be interesting to try to replace (PF) with a less rigid condition. An application of Theorem 2.3 provides Coifman-Meyer type multiplier results for bilinear multipliers given by (1.2) in three cases different than the Fourier transform setting. In Theo- rem 3.1 we treat bilinear multipliers for L being the discrete Laplacian on Zd. This is close to [5, Theorem 3.4], however our results here are of a different kind. In Theorem 4.1 we consider bi-radial bilinear Dunkl multipliers, here L is the general Dunkl Laplacian. In Corollary 4.2 we also reprove [3, Theorem 4.1]. Finally, in Theorem 5.1 we give a Coifman-Meyer type multiplier result for Jacobi trigonometric polynomials, here L is the Jacobi operator. The second main goal of this paper is to obtain fractional Leibniz rules for operators different from the Laplacian. The fractional Leibniz rule states that, if ∆Rd is the Laplacian on Rd, then for each s ≥ 0 and 1/p = 1/p1 + 1/p2, p1, p2 > 1, p > 1/2, we have, k(−∆Rd)s(f g)kp . k(−∆Rd)s(f )kp1kgkp2 + k(−∆Rd)s(g)kp2kfkp1. The proof of this inequality can be found in Grafakos and Ou [11], see also Bourgain and Li [6] for the endpoint case. The fractional Leibniz rule is also known as the Kato-Ponce inequality, as Kato and Ponce studied a similar estimate [13] (see also [14]). Generalizations of Kato-Ponce or similar inequalities were considered by many authors. For example Muscalu, Pipher, Tao, and BILINEAR MULTIPLIERS VIA A FUNCTIONAL CALCULUS 3 Thiele [17] extended this inequality by admitting partial fractional derivatives in R2, Bernicot, Maldonado, Moen, and Naibo [4] proved the Kato-Ponce inequality in weighted Lebesgue spaces, while Frey [10] obtained a fractional Leibniz rule for general operators satisfying Davies- Gaffney estimates and p1 = p = 2, p2 = ∞ In the present paper we obtain fractional Leibniz rules of the form kLs(f g)kp .s kLs(f )kp1kgkp2 + kLs(g)kp2kfkp1, where s > 0 and 1/p1 +1/p2 = 1/p, with p1, p2, p > 1, in two other settings. In Corollary 3.2 we prove a fractional Leibniz rule for L being the discrete Laplacian on Zd, while in Corollary 4.3 we justify a fractional Leibniz rule when L is the Dunkl Laplacian in the product setting. The proofs of these fractional Leibniz rules rely on two properties of L. Firstly, we need appropriate Coifman-Meyer type multiplier results; these are Theorems 3.1 and 4.1 and are deduced from Theorem 2.3. Secondly, we require the existence of certain operators related to L that satisfy (or almost satisfy) an integer order Leibniz rule. As we do not know such an operator in the Jacobi setting we do not provide a fractional Leibniz rule there. The article is organized as follows. In Section 2 we provide a general Coifman-Meyer type multiplier result, see Theorem 2.3. This is then a basis to establish Coifman-Meyer type multiplier results in various cases. In Section 3 we apply Theorem 2.3 for the discrete Laplacian on Zd, see Theorem 3.1. As a consequence, in Corollary 3.2 we also obtain a fractional Leibniz rule. Next, in Section 4 we deduce from Theorem 2.3 a Coifman-Meyer multiplier theorem for general bi-radial Dunkl multipliers, see Theorem 4.1. From this result we obtain a fractional Leibniz rule for the Dunkl Laplacian in the product case, see Corollary 4.3. Finally, in Section 5, using Theorem 2.3 we prove a bilinear multiplier theorem for Jacobi trigonometric polynomial expansions. It is straightforward to extend the result of this paper to the multilinear setting. However, to keep the presentation simple, we decided to limit ourselves to the bilinear case. Troughout the paper we use the variable constant convention, where C, Cp, Cs, etc. may denote different constants that may change even in the same chain of inequalities. We write X . Y, whenever X ≤ CY, with C being independent of significant quantities. Similarly, by X ≈ Y we mean that C −1Y ≤ X ≤ CY. By S(Rd) we denote the space of Schwartz functions. The symbols Z and N denote the sets of integers and non-negative integers, respectively. For a multi-index α ∈ N2 by α we denote its length α1 + α2. Throughout the paper, for a function ψ : [0,∞) → C we set ψk(λ) = ψ(2−kλ), λ ∈ [0,∞). 2. General bilinear multipliers We say that a function µ : (0,∞) → C satisfies the (one-dimensional) Mikhlin-Hörmander condition of order ρ ∈ N if it is differentiable up to order ρ and (2.1) sup kµkM H(ρ) := sup j≤ρ λ∈(0,∞)λj dj dλj µ(λ) < ∞. Similarly, we say that m : (0,∞)2 → C satisfies the (two-dimensional) Mikhlin-Hörmander condition of order s ∈ N, if the partial derivatives ∂αm exist for multi-indices α ≤ s and (2.2) kmkM H(s) := sup α≤s sup λ∈(0,∞)2 λα∂αm(λ1, λ2) < ∞. 4 BŁAŻEJ WRÓBEL Consider a non-negative self-adjoint operator L on L2(X, ν) with domain Dom(L). Here (X, ν) is a σ-finite measure space with ν being a Borel measure. Throughout the paper we assume that L generates a symmetric contraction semigroup, namely ke−tLfkLp(X,ν) ≤ kfkLp(X,ν), f ∈ Lp(X, ν) ∩ L2(X, ν), (CT) and that L is injective on Dom(L). Then, for µ : (0,∞) → C, the spectral theorem allows us to define the multiplier operator µ(L) =R(0,∞) µ(λ)dE(λ) on the domain Dom(µ(L)) =(cid:26)f ∈ L2(X, ν) : Z(0,∞) µ(λ)2 dEf,f (λ) < ∞(cid:27). Here E is the spectral measure of L, while Ef,f is the complex measure defined by Ef,f (·) = hE(·)f, fiL2(X,ν). We shall need the following assumption on L; (MH) L has a Mikhlin-Hörmander functional calculus of a finite order ρ > 0. More pre- cisely, every function µ that satisfies (2.1) gives rise to an operator µ(L) which is bounded on all Lp(X, ν), 1 < p < ∞, and kµ(L)kLp(X,ν)→Lp(X,ν) ≤ CpkµkM H(ρ). Note that if L = (−∆R)1/2 then (MH) follows from the Mikhlin-Hörmander multiplier theorem. There are two consequence of (MH) which will be needed later. The first of them is well known and follows from Khintchine's inequality. Proposition 2.1. Let ψ : [0,∞) → C be a function supported in [ε, ε−1], for some ε > 0, and assume that ψ ∈ C ρ([0,∞)). Then the square function is bounded on Lp(X, ν), p > 1, and f 7→ Sψ(f ) =(cid:0)Xk∈Z ψk(L)f2(cid:1)1/2 (2.3) kSψ(f )kLp(X,ν) ≤ Cε kψkC ρ([0,∞))kfkLp(X,ν). The second of the required consequences is proved in [23, Corollary 3.2]. Proposition 2.2. Let ϕ : [0,∞) → C be compactly supported, and assume that ϕ ∈ C α([0,∞)) for some α > ρ + 2. Then the maximal operator is bounded on Lp(X, ν), p > 1, and f 7→ Mϕ(f ) = sup k∈Z ϕk(L)f (2.4) kMϕ(f )kLp(X,ν) ≤ kϕkC ρ+2([0,∞))kfkLp(X,ν). To simplify the proof of our main Theorem 2.3 we will need an auxiliary subspace of L2(X, ν). Namely, consider the spaces (2.5) A2 = {g ∈ L2(X, ν) : g = E(ε,ε−1)g, for some ε > 0} and A = A2 ∩ \1<p<∞ Then, (MH) implies that A is dense in Lp(X, ν) for 1 < p < ∞. For the convenience of the reader we shall justify this statement. Let ψ : [0,∞) → C be a smooth function which is supported in [1/2, 2] and such that Pk∈Z ψk(λ) = 1, λ > 0. Then, for each N ∈ N and f ∈ L1(X, ν) ∩ L∞(X, ν) the partial sum SN f =PN k=−N ψk(L)f belongs Lp(X, ν). BILINEAR MULTIPLIERS VIA A FUNCTIONAL CALCULUS 5 to A by (MH). We claim that SN f → f in Lp(X, ν). To see this we take 1 < r < ∞ if p ≤ 2 or r > p if p > 2. Then we observe that kSN fkLr(X,ν) is uniformly bounded in N (this follows from (MH)) and that SN f → f in L2(X, ν) (this follows from the spectral theorem, since E{0} = 0 by the injectivity of L). Therefore, the log-convexity of Lp norms proves the claim. Finally, a density argument together with the fact that kSN fkLp(X,ν) is uniformly bounded in N shows that A is dense in Lp(X, ν) and finishes our task. Besides being dense in Lp(X, ν) the space A has the nice property that each f ∈ A satisfies f = PN (f ) k=−N (f ) ψk(L)f, where N (f ) is a fixed integer depending on f and ψ is the function from the previous paragraph. This allows us to deal easily with some rather delicate questions on convergence in the proof of Theorem 2.3. We proceed to define formally the bilinear multipliers studied in this paper. To do this we will need the operators L1 = L ⊗ I and L2 = I ⊗ L. These may be regarded as non-negative self-adjoint operators on L2(X × X, ν ⊗ ν), see [21, Theorem 7.23] and [25, Proposition A.2.2]. Moreover, the spectral measure of L1 is EL ⊗ I, while the spectral measure of L2 is I ⊗ EL. Thus, the operators L1 and L2 commute strongly and the bivariate spectral theorem, see e.g. [21, Theorem 5.21], allows us to consider multiplier operators m(L1, L2) =Z(0,∞)2 m(λ) dE⊗(λ) on the domain Dom(m(L1, L2)) =(cid:26)F ∈ L2(X × X, ν ⊗ ν) : Z(0,∞) m(λ)2 dE⊗ F,F (λ) < ∞(cid:27). Here m : [0,∞)2 → C is a Borel measurable function, E⊗ = EL ⊗ EL is the joint spectral mea- sure of (L1, L2), while E⊗ In the most general form the bilinear multiplier operators studied in the paper are given by F,F is the complex measure defined by (E⊗)F,F (·) = hE⊗(·)F, FiL2(X×X,ν⊗ν). (2.6) Bm(f1, f2)(x) = m(L1, L2)(f1 ⊗ f2)(x, x) = ZZ(0,∞)2 m(λ) d(Ef1 ⊗ Ef2)! (x, x), where L1 = L ⊗ I and L2 = I ⊗ L. Since the diagonal {(x, x) : x ∈ X} may be of measure 0 in (X × X, ν ⊗ ν) the equation (2.6) is not formal. In order to make it rigorous we assume that: (WD) if f1, f2 ∈ A and m : (0,∞)2 → C is bounded, then • m(L1, L2)(f1 ⊗ f2) has a continuous representative on X × X • km(L1, L2)(f1 ⊗ f2)kL∞((X×X),ν⊗ν) ≤ Cf1,f2kmkL∞((0,∞)2). Thus, restricting m(L1, L2)(f1 ⊗ f2)(x1, x2) to the diagonal, we have a formal definition of Bm(f1, f2), for f1, f2 ∈ A. For instance, if L = (−∆R)1/2, then the operator Bm is closely related to the bilinear multiplier for the Fourier transform, namely, Bm(f1, f2)(x) =ZZR2 m(ξ1,ξ2) f (ξ1) f (ξ2)eix(ξ1+ξ2) dξ. If m is bounded and f1, f2 ∈ A then m(ξ1, ξ2) f (ξ1) f (ξ2) ∈ L1(Rd), and thus Bm(f1, f2)(x) is well defined (in fact continuous) by the Lebesgue dominated convergence theorem. 6 BŁAŻEJ WRÓBEL We need one more assumption to prove the main theorem. Namely, we require that: (PF) there is b > 0 with the following property: if ϕ and ψ are bounded smooth functions such that supp ϕk ⊆ [0, 2k−b] and supp ψk ⊆ [2k−2, 2k+2], k ∈ Z, then ϕk(L)(f1) · ψk(L)(f2) = ψk(L)[ϕk(L)(f1) · ψk(L)(f2)], for f1, f2 ∈ A, where ψk is a smooth function which is bounded by 1, equals 1 on [2k−3−b, 2k+3+b] and vanishes outside [2k−5−b, 2k+5+b]. We remark that, since f1, f2 ∈ A, the function g = ϕk(L)(f1)·ψk(L)(f2) belongs to L2(X, ν), so that an application of ψk(L) to g is legitimate. Note that when L = (−∆R)1/2 the formula (PF) can be easily deduced by using the convolution structure on the frequency space associated with Fourier multipliers. In what follows we often abbreviate Lp := Lp(X, ν) and k · kp := k · kLp. Let p, p1, p2 > 1. We say that a bilinear operator B is bounded from Lp1 × Lp2 to Lp if f1, f2 ∈ A. kB(f1, f2)kp ≤ Ckf1kp1kf2kp2, Note that in this case B has a unique bounded extension from Lp1 × Lp2 to Lp. The main result of this paper is a Coifman-Meyer type general bilinear multiplier theorem. Theorem 2.3. Let L be a non-negative self-adjoint operator on L2(X, ν), which is injective on its domain and satisfies (CT), (MH), (WD), and (PF). Assume that m : (0,∞)2 → C satisfies the Mikhlin-Hörmander condition (2.2) of an order s > 2ρ + 4. Then the bilinear multiplier operator Bm, given by (2.6), is bounded from Lp1 × Lp2 to Lp, where 1/p1 + 1/p2 = 1/p, and p1, p2, p > 1. Moreover, for such p, p1, p2, there is C = C(p1, p2, p, s) such that (2.7) kBm(f1, f2)kp ≤ C kmkM H(s) kf1kp1kf2kp2. Proof. Let ψ be a smooth function supported in [1/2, 2] and such that Pk ψk ≡ 1. We set F = f1 ⊗ f2 : X × X → C and split Bm(f1, f2)(x) = Xk1,k2∈Z = Xk1−k2≤b+2 . . . + Xk1>k2+b+2 . . . + Xk2>k1+b+2 [ψk1(L1)ψk2(L2)m(L1, L2)](F )(x, x) . . . := T1 + T2 + T3. There is no issue of convergence here as for f1, f2 ∈ A each of the sums defining T1, T2, and T3 is finite. We estimate separately each of the operators Ti, i = 1, 2, 3, starting with T1. This is the easiest part, in fact here the assumption (PF) is redundant. For k ∈ Z set mk(λ1, λ2) = ψk(λ1) Xk2 : k−k2≤b+2 ψk2(λ2)m(λ) = ψk(λ1)φk(λ2)m(λ), with φ(λ2) =Pj≤b+2 ψj(λ2), so that supp φ ⊆ [2−b−3, 2b+3], and supp ψ ⊗ φ ⊆ [2−1, 21] × [2−b−3, 2b+3]. Let ψ be another smooth function, which vanishes outside [2−b−4, 2b+4] and equals 1 on [2−b−3, 2b+3]. Then mk(λ1, λ2) = [ ψk(λ1) ψk(λ2)]ψk(λ1)φk(λ2)m(λ), BILINEAR MULTIPLIERS VIA A FUNCTIONAL CALCULUS 7 Moreover, supp mk ⊆ [2k−b−4, 2k+b+4]2, and, consequently, Mk(λ) := mk(2kλ) is supported in [−2b+4, 2b+4]2 := [−a, a]2. Thus, Mk can be expanded into a double Fourier series inside [−a, a]2, i.e., cn,keπin1λ1/aeπin2λ2/a, λ ∈ [−a, a]2, with the Fourier coefficients 1 Mk(λ) = Xn1,n2∈Z cn,k = 4a2 ZZ[−a,a]2 [ψ ⊗ φ]m(2kξ) eπin1ξ1/aeπin2ξ2/a dλ. Now, using integration by parts, together with the assumption (2.2), and the fact that ψ ⊗ φ is compactly supported away from 0, we we obtain the uniform in k ∈ Z bound (2.8) cn,k ≤ C kmkM H(s) (1 + n)−s, n ∈ Z2. We remark that here, in order to conclude (2.8), it is perfectly enough to assume the Marcinkiewicz 'product' condition instead of (2.2). Dγm(λ) ≤ Cλ1γ1λ2γ2, Coming back to mk we now write, for λ ∈ [2k−b−4, 2k+b+4]2, ψk(λ1)φk(λ2)m(λ) = Xn∈Z2 cn,k e2πin12−kλ1/ae2πin22−kλ2/a. Thus, mk can be expressed as cn,k[ ψk(λ1)e(2π/a)in12−kλ1][ ψk(λ2)e(2π/a)in22−kλ2] mk(λ1, λ2) = Xn∈Z2 := Xn∈Z2 cn,kψn1 k (λ1)ψn2 k (λ2). By (2.8) and the bivariate spectral theorem we have that cn,k [ ψk(L1)e(2π/a)in12−kL1(f1)](x1)[ ψk(L2)e(2π/a)in22−kL2](f2)(x2), mk(L1, L2)(F )(x1, x2) = Xn∈Z2 for a.e. x1, x2 ∈ X; here we have convergence in L2(X × X, ν ⊗ ν). Moreover, (2.8) and the assumption (WD) imply that the above sum converges also pointwise (and gives a continuous function on X × X). Consequently, for x ∈ X we have T1(f1, f2)(x) =Xk∈Z mk(L1, L2)(F )(x, x) = Xn∈Z2 Xk∈Z k (L)(f1)(x) · ψn2 k (L)(f2)(x), cn,k ψn1 where we have used the fact that the sum in k is finite when f1, f2 ∈ A. Now Schwarz's in- equality (first inequality below), and Hölder's inequality together with (2.8) (second inequality below), lead to the estimate (2.9) kT1(f1, f2)kp ≤ Xn∈Z2 . kmkM H(s) Xn∈Z2 sup k∈Z cn,k(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:0)Xk∈Z (1 + n)−s(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:0)Xk∈Z ψn2 k (L)(f2)2(cid:1)1/2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)p k (L)(f1)2(cid:1)1/2(cid:0)Xk∈Z ψn1 k (L)(f1)2(cid:1)1/2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)p1 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:0)Xk∈Z k (L)(f2)2(cid:1)1/2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)p2 ψn1 ψn2 . 8 BŁAŻEJ WRÓBEL Thus, taking into account the presence of the modulations e2πinj2−kλj/a in the definition of ψnj k , j = 1, 2, and using Proposition 2.1 we obtain (cid:18)Xk∈Z (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ψnj k (L)(fj)2(cid:19)1/2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)pj . (1 + nj)ρ kfjkpj . However, since we have the rapidly decaying factor in (2.9), if s > 2ρ + 4, we arrive at the desired bound Now we pass to estimating T2 and T3. Since the proofs are mutatis mutandis the same, we kT1(f1, f2)kp . kmkM H(s) kf1kp1kf2kp2. treat only the former operator. Setting ϕ =Pj<−b−2 ψj we rewrite T1 as [ψk1(L1)ψk2(L2)m(L1, L2)](F )(x, x) T2(f1, f2)(x) = Xk1>k2+b+2 [ψk1(L1)(cid:18) Xk2<k1−b−2 [ψk(L1)ϕk(L2)m(L1, L2)](f1 ⊗ f2)(x, x), ψ(2−k2 L2)(cid:19)m(L1, L2)](f1 ⊗ f2)(x, x) =Xk1 =Xk where ϕ(λ2) = Pk2<−b−2 ψk2(λ2). Then clearly supp ϕ ⊆ [0, 2−b−1]. Recall that in the above decomposition of T2 all the appearing sums in k, k1, and k2, are in fact finite since f1, f2 ∈ A. Set mk := ψkϕkm and note that mk is supported in [2k−1, 2k+1]× [0, 2k−b−1], this is because supp ψ ⊗ ϕ ⊆ [2−1, 21] × [0, 2−b−1]. Similarly to the case of T1 we expand the function Mk = mk(2kλ) in a Fourier series. Namely, let ψ be a smooth function vanishing outside [2−2, 22] and equal to 1 on [2−1, 21], and let ϕ be a smooth function vanishing outside [0, 2−b] and equal to 1 on [0, 2−b−1]. Then mk(λ1, λ2) = [ ψk(λ1) ϕk(λ2)]ψk(λ1)ϕk(λ2)m(λ), Moreover, supp mk ⊆ [2k−1, 2k+1] × [0, 2k−b−1], and, consequently, Mk(λ) = mk(2kλ) is sup- ported in [−2, 2]2. Hence, Mk can be expanded into a double Fourier series inside [−2, 2]2, i.e., for λ ∈ [−2, 2]2, π π cn,ke 2 in1λ1e 2 in2λ2, with the Fourier coefficients cn,k = [ψ ⊗ ϕ]m(2kξ) e π 2 in1ξ1e π 2 in2ξ2 dξ. Mk(λ) = Xn1,n2∈Z 16ZZ[−2,2]2 1 As with T1, we now use integration by parts, together with the assumption (2.2). Here it is important that we assume the stronger Mikhlin-Hörmander condition instead of merely the Mikhlin-Marcinkiewicz condition. Indeed, from integration by parts we obtain, for arbitrary β cn,k = O((1 + n)−β)ZZ[−2,2]2 dβ dξβ ([ψ ⊗ ϕ]m(2kξ)) e π 2 in1ξ1e π 2 in2ξ2 dξ. However, as ψ ⊗ ϕ does not vanish for λ2 close to zero, in order to conclude that the above integral is uniformly bounded we do need (2.2). In summary we proved that (2.8) holds also in this case. BILINEAR MULTIPLIERS VIA A FUNCTIONAL CALCULUS 9 Coming back to mk we now write, for λ ∈ [2k−2, 2k+2] × [0, 2k−b] ψk(λ1)ϕk(λ2)m(λ) = Xn∈Z2 cn,k e π 2 in12−kλ1e π 2 in22−kλ2. Thus, mk, k ∈ Z, can be expressed as mk(λ1, λ2) =Xn := Xn∈Z cn,kψn1 k (λ1)ϕn2 k (λ2). cn,k[ ψk(λ1)e π 2 in12−kλ1][ ϕk(λ2)e π 2 in22−kλ2] With the aid of (WD) and (2.8), arguing as on p. 7 we see that mk(L1, L2)(F )(x, x) = Xn∈Z2 cn,k ψn1 k (L)(f1)(x) · ϕn2 k (L)(f2)(x), where the series on the right converges pointwise to a continuous function on X. Summarizing the above, we have just decomposed T2(f1, f2)(x) =Xk∈Z mk(L1, L2)(F )(x, x) = Xn∈Z2 Xk∈Z cn,k ψn1 k (L)(f1)(x) · ϕn2 k (L)(f2)(x). Now, let ψ be a real-valued smooth function equal to 1 on [2−3−b, 23+b] and vanishing outside [2−5−b, 25+b]. Since, for each n = (n1, n2) ∈ Z2, the function ϕn2 is supported in [0, 2k−b], and the function ψn1 k is supported in [2k−2, 2k+2], using the assumption (PF) we have k cn,k ψk(L)[ψn1 k (L)(f1) · ϕn2 k (L)(f2)](x). Hence, if h is a function in Lq, 1/p + 1/q = 1, then we obtain T2(f1, f2)(x) = Xn∈Z2 Xk∈Z T2(f1, f2)(x)h(x) dν(x) =ZX Xn∈Z2 Xk∈Z ZX and, consequently, cn,k ψn1 k (L)(f1)(x)· ϕn2 k (L)(f2)(x) ψk(L)(h)(x) dν(x), (2.10) sup k∈Z ZX T2(f1, f2)(x)h(x) dν ≤ Xn∈Z2 k (L)(f1)2(cid:19)1/2 ×ZX(cid:20)(cid:18)Xk∈Z ψn1 (1 + n)−s(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:0)Xk∈Z . kmkM H(s) Xn∈Z2 k (L)(f1)2(cid:19)1/2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)p1 k (L)(f2)(cid:13)(cid:13)(cid:13)(cid:13)p2 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:18)Xk (cid:13)(cid:13)(cid:13)(cid:13) k ϕn2 ψn1 sup where we used Proposition 2.1 with ψ in the second inequality above. Similarly to the estimate for T1, applying Propositions 2.1 and 2.2 leads to ϕn2 k∈Z cn,k× sup k (L)(f2)(cid:21)(cid:18)Xk∈Z k (L)(f1)2(cid:1)1/2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)p1(cid:13)(cid:13)(cid:13)(cid:13) ψn1 dν ψk(L)(h)2(cid:19)1/2 k∈Z ϕn2 sup , k (L)(f2)(cid:13)(cid:13)(cid:13)(cid:13)p2 . (1 + n1)ρkf1kp1 (cf. (2.3)) . (1 + n2)ρ+2kf2kp2 (cf. (2.4)). 10 BŁAŻEJ WRÓBEL Finally, the rapidly decaying factor in (2.10) gives, for s > 2ρ + 4, the desired bound The proof of Theorem 2.3 is thus completed. kT2(f1, f2)kp . kmkM H(s) kf1kp1kf2kp2. (cid:3) 3. Bilinear multipliers on Zd In the present section we formalize Theorem 2.3 for bilinear multiplier operators on Zd. We also prove a fractional Leibniz rule for the discrete Laplacian. Let ej = (0, . . . , 1, . . . , 0) ∈ Zd be the j-th coordinate vector. Consider the discrete Laplacian on Zd, given by ∆Zd(f )(n) = 2d f (n) − d d Xj=1 (f (n + ej) + f (n − ej)) = 2dIf (n) − Xj=1 (f ∗ δej + f ∗ δ−ej ). The multilinear operators (2.6) for the discrete Laplacian are defined via Fourier analysis on Zd. Namely, let Td ≡ (−1/2, 1/2]d be the d-dimensional torus, let ξ ∈ Td f (k)e2πin·ξ, FZd(f )(ξ) = Xn∈Zd be the Fourier transform on Zd, and define Sin2(ξ) = 4 d Xj=1 sin2(πξj), ξ ∈ Td. Then, since the formula (2.6) takes the form FZd(∆Zd(f ))(ξ) = Sin2(ξ)FZd (f )(ξ), ξ ∈ Td, Bm(f1, f2)(n) := m((−∆Zd)1/2 ⊗ I, I ⊗ (−∆Zd)1/2)(f1 ⊗ f2)(n, n) (3.1) =ZTdZTd m( Sin(ξ1), Sin(ξ2))FZd (f1)(ξ1)FZd(f2)(ξ2)e−2πin(ξ1+ξ2) dξ, where n ∈ Zd. Note that the space A2 from (2.5) in this case is given by A2 = {g ∈ L2(Td) : FZd(g)(ξ) = 0 for some ε > 0 and all ξ < ε.} Throughout this section we denote by Lp the space lp(Zd) equipped with the counting measure. Using Theorem 2.3 we prove the following Coifman-Meyer multiplier theorem for the discrete Laplacian. Theorem 3.1. Assume that m satisfies Hörmander's condition (2.2) of order s > d + 4. Then the bilinear multiplier operator given by (3.1) is bounded from Lp1 × Lp2 to Lp, where 1/p1 + 1/p2 = 1/p, and p1, p2, p > 1. Moreover, the bound (2.7) holds. Proof. It is well known that L = (−∆Zd)1/2 is injective on L2 and satisfies (CT). Moreover, it also satisfies (WD) since for f1, f2 ∈ A we have FZd(f1)(ξ1)FZd(f2)(ξ2) ∈ L1(Td × Td). From [1, Theorem 1.1] it follows that −∆Zd has a Mikhlin-Hörmander functional calculus (of order [d/2] + 1). Then, clearly, the same is true for (−∆Zd)1/2. Hence, (MH) has been justified. To apply Theorem 2.3 it remains to show that L = (−∆Zd)1/2 satisfies (PF). We prove 2 log2 d. Since the spectrum of (−∆Zd)1/2 is contained in [0, 2√d], we have it with b = 7 + 1 BILINEAR MULTIPLIERS VIA A FUNCTIONAL CALCULUS 11 2 log2 d. Hence, it suffices to show (PF) for k ≤ 2 + 1 ψk((−∆Zd)1/2) ≡ 0, if k > 2 + 1 2 log2 d. Using elementary Fourier analysis on Zd we see that to prove (PF) it is enough to show that ψk ◦ Sin = 1 ((ψk ◦ Sin)FZd(f1)) ∗Td ((ϕk ◦ Sin))FZd (f2)), where ψk, ψk, and ϕk are the functions from (PF). In other words that we are left with proving that if Sin(ξ) < 2k−3−b or Sin(ξ) > 2k+3+b, then (3.2) ψk( Sin(ξ − η))FZd (f1)(ξ − η) · ϕk( Sin(η))FZd (f2)(η) dη = 0. on the support of ZTd The formula (3.3) leads to sin π(ξj) ≤ sin π(ξj − ηj) + sin πηj, j = 1, . . . , d, and, consequently, sin π(t − s) = sin πt cos πs − sin πs cos πt, √d( Sin(ξ − η) + Sin(η)), Sin(ξ) ≤ η ∈ Td. s, t ∈ T, From the above it follows that if Sin(ξ) > 2k+3+b, then for every η ∈ Td the integrand in (3.2) vanishes. It remains to show that also Sin(ξ) < 2k−3−b forces (3.2). We argue by contradiction assuming that Sin(ξ) < 2k−3−b yet the integral in (3.2) is non-zero. Then, for some η ∈ Td, we must have ψk( Sin(ξ − η)) ϕk( Sin(η)) 6= 0, which implies that (3.4) Note that since k ≤ 2 + 1 2 log2 d, the integral in (3.2) runs over Sin(η) ≤ 2k−b ≤ 2−1, and, consequently, we consider only those η satisfying cos πηj > √3/2 > 1/2, for every j = 1, . . . , d. Now, using (3.3) (with t − s = ξj, s = −ηj) we obtain 2k−1 ≤ Sin(ξ − η) ≤ 2k+1 Sin(η) ≤ 2k−b. and sin πξj ≥ sin π(ξj − ηj) cos πηj − cos π(ξj − ηj) sin πηj ≥ 1 2 sin π(ξj − ηj) − sin πηj. Summing the above estimate in j and using Schwarz inequality we arrive at d d 1 2 Xj=1 sin πξj ≥ √d Sin(ξ) ≥ Xj=1 Now, since Sin(ξ) < 2k−3−b, using (3.4) we arrive at 1 √d sin π(ξj − ηj) − Xj=1 2k−b−3 > Sin(ξ) > which is a contradiction. √d2k−b(cid:1) = 1 √d(cid:0)2k−1 − d sin πηj ≥ 1 2 Sin(ξ − η) − √d Sin(η). (2k−1 − 2k−7) > 1 √d 2k−2 = 2k−b+5, (cid:3) As a corollary of Theorem 3.1 we prove a fractional Leibniz rule for the discrete Laplacian on Zd. For Re(z) ≥ 0 and h ∈ L2 the complex derivative (−∆Zd)zh is given by FZd((−∆Zd)zh)(ξ) = Sin ξ2zFZd(h)(ξ), ξ ∈ Td. This coincides with taking the n-th composition of (−∆Zd) when z = n is a non-negative integer. Clearly, (−∆Zd)z is bounded on L2. Moreover, when z = s ∈ R, s ≥ 0, then (−∆Zd)s is also bounded on all Lp, 1 ≤ p ≤ ∞. To see this we just use the Taylor series expansion of the function xs = (1 − (1 − x))s, with x replaced by (−∆Zd)/(4d). This is legitimate since (−∆Zd)/(4d) is a contraction on all Lp spaces. Our fractional Leibniz rule is the following. 12 BŁAŻEJ WRÓBEL k(−∆Zd)s(f g)kp . k(−∆Zd)sfkp1kgkp2 + k(−∆Zd)sgkp2kfkp1, Corollary 3.2. Let 1/p = 1/p1 + 1/p2, with p, p1, p2 > 1. Then, for every s > 0, (3.5) where f, g ∈ A. Remark 1. Note that if f, g ∈ A then f g ∈ L2, hence (−∆Zd)s(f g) makes sense. Remark 2. Since (−∆Zd)s is bounded on all Lp spaces, 1 ≤ p ≤ ∞, a version of (3.5) without the Laplacians on the right hand side is obvious. This is in contrast with the fractional Leibniz rule on Rd. In the proof of the corollary we shall need two lemmata. The first of them follows from the lp(Z) boundedness of the discrete Hilbert transform. Lemma 3.3. The one-dimensional linear multiplier operator H(f )(n) =Z 1/2 is bounded on all lp(Z) spaces, 1 < p < ∞. 0 FZ(f )(x)e2πiξn dξ, The second of the lemmata is the following. n ∈ Z Lemma 3.4. Let d = 1. Assume that ϕ : (0,∞)2 → C is a bounded function that satisfies the Mikhlin-Hörmander condition (MH) of order 6. Then, for Re(z) ≥ 0 we have (3.6) (−∆Z)z(Bϕ(f, g))(n) =ZZT2 ϕ(2 sin πξ1, 2 sin πξ2)2 sin π(ξ1 + ξ2)2ze2πi(ξ1+ξ2)n FZ(f )(ξ1)FZ(g)(ξ2) dξ, where f, g ∈ A, and n ∈ Z. Proof. From Theorem 3.1 and the assumptions on ϕ it follows that Bϕ(f, g) ∈ ℓ2(Z). Thus, the left hand side of (3.6) makes sense as a function on ℓ2(Z). Moreover, a continuity argument shows that it suffices to demonstrate (3.6) for Re(z) > 0. Thus, for P being a polynomial we obtain ϕ(ξ1, ξ2) (4 sin2 π(ξ1 + ξ2))ke2πi(ξ1+ξ2)n FZ(f )(ξ1)FZ(g)(ξ2) dξ. Set ϕ(ξ1, ξ2) = ϕ(2 sin πξ1, 2 sin πξ2). Since −∆Z(e2πit·)(n) = 4(sin2 πt)e2πitn, for t ∈ T and n ∈ Z, we deduce that (−∆Z)k(e2πit·)(n) = 22k sin πt2ke2πitn, k ∈ N. Hence, for k, n ∈ N, we have (−∆Z)k(Bϕ(f, g))(n) =ZZT2 P (−∆Z)(Bϕ(f, g))(n) =ZZT2 where n ∈ Z. tions in place of polynomials. In particular, taking λ 7→ λz, Re(z) > 0, we obtain (3.6). Finally, a density argument shows that the above formula remains true for continuous func- (cid:3) ϕ(ξ1, ξ2) P (4 sin2 π(ξ1 + ξ2))e2πi(ξ1+ξ2)n FZ(f )(ξ1)FZ(g)(ξ2) dξ, We proceed to the proof of the corollary. Proof of Corollary 3.2. We claim that it is enough to prove the corollary in dimension d = 1. Indeed, fix s > 0 and assume that (3.5) is true in this case. Let ∆Z be the one dimensional discrete Laplacian on Z. Define Lj := −∆Z ⊗ I(j), j = 1, . . . , d, to be the one-dimensional discrete Laplacian acting on the j-th variable, so that, clearly, −∆Zd = Pd j=1 Lj. Since each BILINEAR MULTIPLIERS VIA A FUNCTIONAL CALCULUS 13 Lj generates a symmetric contraction semigroup, using e.g. the multivariate multiplier theorem [24, Corollary 3.2] we see that the operator is bounded on Lp, p > 1. In other words, we have the bound (X Lj)s(X Ls j)−1 k(−∆Zd)s(f g)kp . k d Xj=1 Ls j(f g)kp ≤ d Xj=1 kLs j(f g)kp. Since the multiplier Ls Corollary 3.2]) in order to conclude the proof of our claim it is thus enough to show that j(−∆Zd)−s is bounded on all Lp, p > 1, (this again follows from [24, (3.7) for every j = 1, . . . , d. kLs j(f g)kp . kLs jfkp1kgkp2 + kLs jgkp2kfkp1, For notational simplicity we justify (3.7) only for j = 1, the proofs for other j are analogous. For a sequence h : Zd → C denote hn(k) := h(k, n), k ∈ Z, n ∈ Zd−1. Clearly, we have (f g)n(·) = fn(·)gn(·). Then, using (3.5) in the dimension d = 1 (first inequality below), together with the simple fact that (a + b)p ≈ ap + bp (second and last inequalities below), and Hölder's inequality with exponents p1/p, p2/p > 1 (third inequality below) we obtain 1(f g)kp = Xn∈Zd−1 kLs kLs . Xn∈Zd−1(cid:0)kLs . Xn∈Zd−1 kLs .(cid:0) Xn∈Zd−1 kLs +(cid:0) Xn∈Zd−1 kLs .(cid:0)kLs 1(fn(·)gn(·))kp 1(gn)klp1 (Z)kfnklp2 (Z)(cid:1)p 1(gn)kp lp1 (Z)kgnkp lp1 (Z)(cid:1)p/p1(cid:0) Xn∈Zd−1 kgnkp2 lp2 (Z)(cid:1)p/p2(cid:0) Xn∈Zd−1 kfnkp1 lp1 (Z)kfnkp lp2 (Z)(cid:1)p/p2 lp1 (Z)(cid:1)p/p1 = kLs 1((f g)n(·))kp lp(Z) = Xn∈Zd−1 kLs 1(fn)klp1 (Z)kgnklp2 (Z) + kLs 1(fn)kp 1(fn)kp1 1(gn)kp2 lp2 (Z) + kLs 1(g)kp2kfkp1(cid:1)p. 1(f )kp1kgkp2 + kLs 1(f )kp p1kgkp p2 + kLs Hence, (3.7) is proved. lp2 (Z) lp(Z) 1(g)kp p2kfkp p1 Having justified the claim we now focus on proving (3.5) for d = 1. Till the end of the proof of the corollary we work on Z and write lp and k · kp for lp(Z) and k · klp(Z), respectively. Let η0 and η1 be smooth functions satisfying supp η0 ⊆ [0, 1/4], supp η1 ⊆ [1/8, 10] and η0 + η1 = 1 on [0, 4]. For a function h ∈ A we set h0 = η0((−∆Z)1/2)(h) and h1 = η1((−∆Z)1/2)(h), so that h = h0 + h1. From [1, Theorem 1.1] it follows that, for each fixed s > 0 the mul- tiplier (−∆Z)−sη1(−∆Z) is bounded on all lp, 1 < p < ∞. Moreover, h0, h1 ∈ A. Since h1 = (−∆Z)−sη1(−∆Z)[(−∆Z)s(h)], we thus have the estimate kh1kp . k(−∆Z)shkp. Hence, using the boundedness of (−∆Z)s and Hölder's inequality we obtain k(−∆Z)s(fi1gi2)kp . kfi1kp1kgi2kp2 . k(−∆Z)sfkp1kgkp2 + k(−∆Z)sgkp1kfkp2, 14 BŁAŻEJ WRÓBEL for i1, i2 ∈ {0, 1} not both equal to 0. In summary, to finish the proof it is enough to demon- strate that k(−∆Z)s(f0g0)kp . k(−∆Z)sfkp1kgkp2 + k(−∆Z)sgkp1kfkp2. Clearly, FZ(f0)(x) = η0( sin πx)FZ(f )(x) and FZ(g)(y) = η0( sin πy)FZ(g0)(y). Hence, denoting I = [0, 1/2) and using Lemma 3.4 together with (3.3) we now write (−∆Z)s(f0g0)(n) = 22sZTZT = 22s Xǫ∈{−1,1}2Zǫ1IZǫ2I sin πξ1q1 − sin2 πξ2 + sin πξ2q1 − sin2 πξ12sη0( sin πξ1)η0( sin πξ2) [ sin πξ1 cos πξ2 + sin πξ1 cos πξ22sη0( sin πξ1)η0( sin πξ2)] × e2πi(ξ1+ξ2)n FZ(f )(ξ1)FZ(g)(ξ2) dξ × e2πi(ξ1+ξ2)n FZ(f )(ξ1)FZ(g)(ξ2) dξ := Xǫ∈{−1,1}2 Tǫ(f, g)(n), n ∈ Z. Thus, in order to finish the proof it is enough to show that, for ǫ ∈ {−1, 1}2 it holds (3.8) kTǫ(f, g)kp . k(−∆Z)sfkp1kgkp2 + k(−∆Z)sgkp2kfkp1. It is enough to justify (3.8) only for T1,1 and T1,−1 as the proofs for T−1,1 and T−1,−1 are symmetric. In what follows we let φ be a function in C ∞([0,∞)) supported in [0, 1/4] and such that φ(t) + φ(t−1) = 1. Note that then φ(λ2/λ1) satisfies Hörmander's condition (2.2) of arbitrary order. Let (η⊗ 0 )(λ) = η0(λ1)η0(λ2), λ ∈ [0,∞)2. To justify (3.8) for T1,1 we set φ(λ2/λ1)(η⊗ 2/4)1/2 + λ2(1 − λ2 1/4)1/22s 0 )(λ), 1,1(λ) = λ1(1 − λ2 ms 1,1(λ) = λ1(1 − λ2 ms 2/4)1/2 + λ2(1 − λ2 1/4)1/22s φ(λ1/λ2)(η⊗ 0 )(λ). λ2s 1 λ2s 2 λ2z 1 λ2z 2 Then, using (3.1) (in the case d = 1) we rewrite T1,1 as T1,1(f, g) = Bms 1,1(H(−∆Z)sf, Hg) + B ms 1,1(Hf, H(−∆Z)sg). In view of Lemma 3.3, to demonstrate (3.8) it suffices to show that kBms 1,1 (f, g)kp + kB ms 1,1 (f, g)kp ≤ Ckfkp1kgkp2 . This, however, follows directly from Theorem 3.1, since, for each s > 0, the multipliers ms and ms 1,1, satisfy Hörmander's condition (2.2) of arbitrary order. 1,1, Finally, we prove (3.8) for T1,−1. For Re(z) ≥ 0 we set 1,−1(λ) = λ1(1 − λ2 mz 1,−1(λ) = λ1(1 − λ2 mz 2/4)1/2 − λ2(1 − λ2 1/4)1/22z 2/4)1/2 − λ2(1 − λ2 1/4)1/22z φ(λ2/λ1)(η⊗ 0 )(λ), φ(λ1/λ2)(η⊗ 0 )(λ). Then using (3.1) (in the case d = 1) we rewrite T1,−1 as T1,−1(f, g) = Bms 1,−1(H(−∆Z)sf, (I − H)g) + B ms 1,−1 (Hf, (I − H)(−∆Z)sg). BILINEAR MULTIPLIERS VIA A FUNCTIONAL CALCULUS 15 Note that A is preserved by (−∆Z)s. Thus, by Lemma 3.3, to demonstrate (3.8) it is enough to prove, for f, g ∈ A, the bounds (3.9) kBms kB ms 1,−1 (Hf, (I − H)g)kp ≤ CkHfkp1k(I − H)gkp2 , 1,−1 (Hf, (I − H)g)kp ≤ CkHfkp1k(I − H)gkp2 . We focus only on the first estimate, the reasoning for the second being analogous. We are going to apply Stein's complex interpolation theorem [22] for each fixed f ∈ A. The argument used here takes ideas from the proof of [16, Theorem 1.4]. For further reference we note that the formula (3.10) Bmz 1,−1 (Hf, (I − H)g)(n) =Z 1/2 φ(cid:18) sin πξ2 sin πξ1(cid:19)η0( sin πξ1)η0( sin πξ2) × sin πξ1p1 − sin2 πξ2 − sin πξ2p1 − sin2 πξ12z sin πξ12z e2πi(ξ1+ξ2)n FZ(f )(ξ1)FZ(g)(ξ2) dξ; Z 0 −1/2 0 makes sense not only for f, g ∈ A but more generally, for f, g ∈ ℓ2. Mikhlin-Hörmander condition (2.2) of order 8. Thus, Theorem 3.1 (with d = 1) gives Let n be an even integer larger than 8. Then the multipliers mn+iv 1,−1 , v ∈ R, satisfy the kBmn+iv 1,−1 Now, Lemma 3.4 applied to ϕ(λ) = φ(λ2/λ1)η⊗ (Hf, (I − H)g)kp ≤ C(1 + v)8kHfkp1k(I − H)gkp2 , 0 (λ), λ ∈ (0,∞)2, implies v ∈ R. Bmiv 1,−1 (Hf, (I − H)g) = (−∆Z)iv(cid:2)Bφ(λ2/λ1)η⊗ 0 (H(−∆Z)−ivf, (I − H)g)(cid:3). By [1, Theorem 1.1] we have k(−∆Z)ivkℓq→ℓq ≤ Cq(1 + v)4, 1 < q < ∞. Hence, Theorem 3.1 applied to the multiplier φ(λ1/λ2)η⊗ 0 produces kBmiv 1,−1 (Hf, (I − H)g)kp ≤ C(1 + v)8kHfkp1k(I − H)gkp2, v ∈ R. By (3.10), for fixed f ∈ A, the family {Bmz 1,−1 (Hf, (I − H)g)}Re(z)>0 consists of analytic operators. This family has admissible growth, more precisely, for each finitely supported g, h we have Consequently, an application of Stein's complex interpolation theorem is permitted and leads to the first inequality in (3.9). The proof of the corollary is thus finished. (cid:3) (cid:12)(cid:12)hBmz 1,−1 (Hf, (I − H)g), hil2(Z)(cid:12)(cid:12) ≤ Cf,g,h, Re(z) ≤ s. 4. Bilinear radial multipliers for the generic Dunkl transform Here we apply Theorem 2.3 for bilinear multiplier operators associated with the generic Dunkl transform. In the case when the underlying group of reflections is isomorphic to Z2 we also prove a fractional Leibniz rule. Let R be a root system in Rd and G the associated reflection group (see [19, Chapter 2]). Let σα(x) denote the reflection of x in the hyper-plane orthogonal to α ∈ Rd and let κ be a nonnegative, G invariant function on R. The differential-difference (rational) Dunkl operators, are defined as δjf (x) = ∂jf (x) + Xα∈R+ αjκ(α) f (x) − f (σα(x)) hα, xi , j = 1, . . . , d. 16 BŁAŻEJ WRÓBEL Here f is a Schwartz function, R+ is a fixed positive subsystem of R and hx, yi = Pd j=1 xjyj is the standard inner product. The fundamental property of the operators δj is that, similarly to the usual partial derivatives (which appear when we take κ ≡ 0), they commute, i.e. δlδj = δjδl, l, j = 1, . . . , d. The operators δj are also symmetric on L2 = L2(Rd, w(x)dx), with i=1 hα, xi2κ(α) . Moreover they leave S(Rd) invariant. Additionally the w(x) = wκ(x) := Qd Leibniz rule (4.1) δj(f1f2)(x) = δj(f1)(x)f2(x) + δj(f1)(x)f2(x), x ∈ Rd, holds under the extra assumption that one of the functions f1, f2 is invariant under G. The easiest case of Dunkl operators arrises when G ∼ Zd reflections through the coordinate axes. In this case 2. In other words G consists of δjf (x) = ∂jf (x) + κj f (x) − f (σj(x)) xj , j = 1, . . . , d, where κj ≥ 0, while σj(x) denotes the reflection of x in the hyperplane orthogonal to the j-th coordinate vector. In this case the weight wκ(x) takes the product form wκ(x) =Qd j=1 wκj (xj), x ∈ Rd. transform. It is defined by In the (general) Dunkl setting there is an analogue of the Fourier transform, called the Dunkl Df (ξ) = cκZRd E(−iξ, x)f (x)wκ(x) dx where E(z, w) = Eκ(z, w) = Eκ(w, z) is the so called Dunkl kernel. A defining property of this kernel is the equation (4.2) The operator D has properties similar to the Fourier transform. Namely, we have the Plancherel formula δj,x(Eκ(iξ, x)) = iξjEκ(iξ, x), x ∈ Rd. (4.3) ZRd f (x)g(x) w(x) dx = cκ ZRd D(f )(ξ)D(h)(ξ) w(ξ) dξ, and the inversion formula, (4.4) f (x) = D2f (−x) = cZRd D(f )(ξ)E(iξ, x) w(ξ) dξ, f ∈ S(Rd). Additionally, the Dunkl transform diagonalizes simultaneously the Dunkl operators δi, i.e. (4.5) Dδjf = iξjD. δjDf = −D(ixjf ), The Dunkl Laplacian is given by ∆κ =Pd the operator −∆κ may be formally defined as a non-negative self-adjoint operator on L2(Rd, w). The same is true for L := (−∆κ)1/2. Then, for a bounded function µ the spectral multiplier µ(L) is uniquely determined on S(Rd) by (4.6) D(∆κf )(ξ) = −ξ2 D(f )(ξ), i . Using the identity ξ ∈ Rd, i=1 δ2 D(µ(L)f )(ξ) = µ(ξ)D(f )(ξ) ξ ∈ Rd. BILINEAR MULTIPLIERS VIA A FUNCTIONAL CALCULUS 17 Consider now L1 := L ⊗ I and L2 = I ⊗ L. Analogously to the case of bilinear Fourier multipliers the formula (2.6) can given by the Dunkl transform. Namely, for a bounded function m : [0,∞)2 → C we have Bm(f1, f2)(x) (4.7) =ZRdZRd m(ξ1,ξ2)D(f )(ξ1)D(g)(ξ2) E(iξ1, x)E(iξ2, x) w(ξ1)w(ξ2)dξ1dξ2. The above formula is valid pointwise e.g. for Schwartz functions f1 and f2 on Rd. We observe that in this section the space A2 from (2.5) is (4.8) Thus, by (4.5) the Dunkl derivatives δj, j = 1, . . . , d, preserve A2. lution. For x, y ∈ Rd The Dunkl translation is defined by A2 = {g ∈ L2(Rd, wκ) : there is ε > 0 such that D(g)(ξ) = 0 for ξ 6∈ [ε, ε−1]}. In this section we will heavily rely on the concepts of Dunkl translation and Dunkl convo- τ yf (x) = cκZRd D(f )(ξ)E(iξ, x)E(iξ, y) w(ξ) dξ. The inversion formula (4.4) and the properties of the Dunkl kernel imply For f, g ∈ A the Dunkl convolution is D(τ yf )(ξ) = E(−iξ, y)D(f )(ξ). f ∗κ g(x) =ZRd f (y) τxg(y) w(y) dy, The first result of this section is the following Coifman-Meyer type theorem. In what follows where g(x) = g(−x). It is known that the Dunkl transform turns this convolution into multi- plication, i.e. (4.9) D(f ∗κ g)(x) = D(f )(x)D(g)(x), f, g ∈ A. we set λκ = (d− 1)/2 +Pα∈R+ κ(α) and for brevity write Lp := Lp(Rd, wκ) and k·kp = k·kLp. Theorem 4.1. Assume that m satisfies the Mikhlin-Hörmander condition (2.2) of an order s > 2λκ + 6. Then the bilinear multiplier operator given by (4.7) is bounded from Lp1 × Lp2 to Lp, where 1/p1 + 1/p2 = 1/p, and p1, p2, p > 1. Moreover, the bound (2.7) holds. [D(f ) ∗κ D(g)](x) = D(f g)(x), Proof. We are going to apply Theorem 2.3. In order to do so we need to check that its as- sumptions are satisfied for the operator L = (−∆κ)1/2. To see that L is injective on its domain we merely note that wκ(ξ) dξ is absolutely continuous with respect to Lebesgue measure. The contractivity condition (CT) follows from [19, Theorem 4.8] and the subordination method. The assumption (WD) is straightforward from (4.7) and the Lebesgue dominated convergence theorem, while (MH) was proved by Dai and Wang [8, Theorem 4.1] (with arbitrary ρ > λκ+1). Thus we are left with verifying the property (PF), which we prove with b = 2. This will be deduced by using the convolution structure associated with Dunkl operators. Let ϕk and ψk, be smooth functions such that supp ϕk ⊆ [0, 2k−2] and supp ψk ⊆ [2k−1, 2k+1]. Let ψk be a smooth function equal 1 on [2k−5, 2k+5] and vanishing outside of [2k−7, 2k+7]. Taking the Dunkl transform of the both sides of (PF) and using (4.6) we see that our task is equivalent to proving the formula D(ϕk(L)(f1)ψk(L)(f2)) = ψk(ξ)D(ϕk(L)(f1)ψk(L)(f2)), ξ ∈ Rd. 18 BŁAŻEJ WRÓBEL Denote gj = D(fj), j = 1, 2. By (4.9) and (4.6) the equation above is exactly [(ϕk( · )g1) ∗κ (ψk( · )g2)](ξ) = ψk(ξ)[(ϕk( · )g1) ∗κ (ψk( · )g2)](ξ), By definition of ψ to prove the last formula it is enough to show that ξ ∈ Rd. supp[h1 ∗κ h2] ⊆ [2k−5, 2k+5], (4.10) for any functions h1 supported in B(0, 2k−2) and h2 supported in B(0, 2k+1)\ B(0, 2k−1). Take ξ 6∈ [2k−5, 2k+5] and y ∈ B(0, 2k−2). We claim that τ ξ h2(y) = 0. This implies (4.10). Till the end of the proof we thus focus on proving the claim. Let γξ,y be the distribution given by γξ,y(f ) = (τ ξf )(y), f ∈ S(Rd). In [2, Theorem 5.1] Amri, Anker, and Sifi proved that γξ,y is supported in the spherical shell Sξ,y :=(cid:8)z ∈ Rd : ξ − y ≤ z ≤ ξ + y(cid:9). Therefore, if we prove that supp h2 ∩ Sξ,y = ∅, then τ ξh2(y) = 0. Recall that we have ξ 6∈ [2k−5, 2k+5] and y ∈ B(0, 2k−2). Take z ∈ Sξ,y and consider two possibilities, either ξ < 2k−5 or ξ > 2k+5. In the first case we obtain z ≤ 2k−5 + 2k−2 < 2k−1, while in the second z ≥ ξ − y ≥ 2k+5 − 2k−2 > 2k+1. Thus, in both the cases z 6∈ supp h2, and the proof of (PF) is completed. (cid:3) Theorem 4.1 is quite far from a general bilinear Dunkl multiplier theorem, i.e. when the multiplier function m is not necessarily radial in each of its variables. However, in the case d = 1 (and G ∼ Z2), Theorem 4.1 implies [3, Theorem 4.1] by Amri, Gasmi, and Sifi. We slightly abuse the notation and, for ϕ : R2 → C, f1, f2 ∈ A, and x ∈ R, define (4.11) ϕ(ξ)D(f1)(ξ1)D(f2)(ξ2) E(iξ1, x)E(iξ2, x) w(ξ1)w(ξ2) dξ. Bϕ(f1, f2)(x) =ZRZR This will cause no confusion with (4.7), as till the end of the present section we only use Bϕ given by (4.11). Corollary 4.2 (Theorem 4.1 of [3]). Let G ∼ Z2. Assume that ϕ : R2 → C satisfies the Mikhlin-Hörmander condition on R2 of an order s > 2λκ + 6, namely ξ∈R2 ξα∂αϕ(ξ1, ξ2) < ∞. kϕkM H(R2,s) := sup (4.12) sup α≤s Then the bilinear multiplier operator given by (4.11) is bounded from Lp1 × Lp2 to Lp, where 1/p1 + 1/p2 = 1/p, and p1, p2, p > 1. Remark. When κ = 0 we recover the Coifman-Meyer multiplier theorem in the Fourier trans- form setting. Proof of Corollary 4.2 (sketch). Let Π(f )(x) = D−1(χξ>0)D(f )(ξ))(x) be the projection onto the positive Dunkl frequencies. The corollary can be deduced from the boundedness of Π on all Lp spaces 1 < p < ∞. (cid:3) For Re z ≥ 0, let (−∆κ)z be the complex Dunkl derivative D[(−∆κ)z(h)](ξ) = ξ2zD(h)(ξ), ξ ∈ Rd. The natural L2 domain of this operator is DomL2((−∆κ)z) = {h ∈ L2 : ξ2 Re zD(h)(ξ) ∈ L2}. BILINEAR MULTIPLIERS VIA A FUNCTIONAL CALCULUS 19 By Plancherel's formula for the Dunkl transform (−∆κ)z(h) ∈ L2 for h ∈ A. The second main result of this section is the following fractional Leibniz rule for (−∆κ)s, in the case G ∼ Zd 2. Corollary 4.3. Let G ∼ Zd s > 0, we have 2 and take 1/p = 1/p1 + 1/p2, with p, p1, p2 > 1. Then, for any k(−∆κ)s(f g)kp . k(−∆κ)s(f )kp1kgkp2 + kfkp1k(−∆κ)s(g)kp2 , where f, g ∈ A and at least one of the functions f or g is invariant by G. Before proving the fractional Leibniz rule we need a lemma which is an analogue of Lemma 3.4. Its proof is similar, however a bit more technical. Therefore we give more details. Lemma 4.4. Take d = 1 and let G ∼ Z2. Assume that at least one of the functions f, g ∈ A is G-invariant. Take Re(z) ≥ 0 and let ϕ : R2 → C be a bounded function that satisfies the Mikhlin-Hörmander condition (4.11) of order s > 2λκ + 6. Then (−∆κ)z(Bϕ(f, g))(x) =ZZR2 for almost all x ∈ Rd. Remark. It is not obvious why Bϕ(f, g) ∈ DomL2((−∆κ)z). This is explained in the proof of the lemma. ϕ(ξ)ξ1+ξ22z D(f )(ξ1)D(g)(ξ2) E(iξ1, x)E(iξ2, x) w(ξ1)w(ξ2)dξ, Proof. Since the argument is symmetric in f and g we assume that f is G-invariant. Denote EG(iξ1, x) = G−1Pg∈G E(iξ1, gx), and observe that EG is G-invariant in x. Then, since both f and D(f ) are G-invariant our task reduces to proving that (4.13) (−∆κ)z(Bϕ(f, g))(x) =ZRZR ϕ(ξ)ξ1+ξ22z D(f )(ξ1)D(g)(ξ2) EG(iξ1, x)E(iξ2, x) w(ξ1)w(ξ2)dξ, for almost all x ∈ Rd. For z = n ∈ N this formula is a direct computation, and follows from the Leibniz rule. Indeed, by (4.1) and (4.2) we have δ(Bϕ(f, g))(x) =ZZR2 =ZZR2 ϕ(ξ)D(f )(ξ1)D(g)(ξ2) i(ξ1 + ξ2) EG(iξ1, x)E(iξ2, x) w(ξ1)w(ξ2)dξ, ϕ(ξ)D(f )(ξ1)D(g)(ξ2) δx[EG(iξ1, x)E(iξ2, x)] w(ξ1)w(ξ2)dξ the interchange of differentiation and integration being allowed since f, g ∈ A. Iterating the above equality 2n times we obtain (4.13) for z = n. We remark that (4.13) for z ∈ N also explains why does (−∆κ)z(Bϕ(f, g)) make sense for general Re(z) ≥ 0. Indeed, let n be an integer larger than Re(z). Then, to prove that Bϕ(f, g) ∈ DomL2((−∆κ)z) it is enough to show that Bϕ(f, g) ∈ DomL2((−∆κ)n). Now, using (4.13) for z = n, together with the binomial formula and (4.5), we arrive at 2n (−∆κ)n(Bϕ(f, g))(x) Xj=0(cid:18)2n j (cid:19)ZRZR = ϕ(ξ)D(δj f )(ξ1)D(δ2n−j g)(ξ2) EG(iξ1, x)E(iξ2, x) w(ξ1)w(ξ2)dξ, 20 BŁAŻEJ WRÓBEL with δ being the Dunkl operator on R. Since f, g belong to A2 the same is true for δj f and δ2n−j g. Thus, an application of Corollary 4.2 proves that Bϕ(f, g) ∈ DomL2((−∆κ)n), as desired. We come back to demonstrating (4.13) for general Re(z) ≥ 0. Note first that by a continuity Tz(f, g)(x) =ZZR2 ϕ(ξ)ξ1 + ξ22z D(f )(ξ1)D(g)(ξ1) E(iξ1, x)E(iξ2, x) w(ξ1)w(ξ2)dξ. argument it suffices to consider Re(z) > 0. Denoting our task is reduced to proving that (4.14) for h ∈ A2∩S(R) (recall that A2 is given by (4.8)). This is enough because A2∩S(R) is dense in L2. From (4.13) for z ∈ N we deduce that for any polynomial P it holds h(−∆κ)z(Bϕ(f, g)), hiL2 = hTz(f, g), hiL2 , (4.15) P (−∆κ)(Bϕ(f, g))(x) =ZZR2 ϕ(ξ)P (ξ1 + ξ22)D(f )(ξ1)D(g)(ξ2) EG(iξ1, x)E(iξ2, x) w(ξ1)w(ξ2)dξ. For brevity we denote by T P (f, g)(x) the right hand side of (4.15). Note that D(f ), D(g), and D(h) are supported in [−N, N ] for some large N. Let {Pr(t)}r∈N, be a sequence of polynomials that converges uniformly to tz on [0, 4N 2]. Then, (4.3), (4.5), and (4.15) imply (4.16) Pr(ζ2)D(Bϕ(f, g))(ζ)D(¯h)(ζ) w(ζ) dζ = hPr(−∆κ)(Bϕ(f, g)), hiL2 ZR = hT Pr (f, g), hiL2 . Now, since suppD(¯h) ⊆ [−N, N ] and D(Bϕ(f, g))D(¯h) ∈ L1, the dominated convergence theorem shows that the left hand side of (4.16) converges to h(−∆κ)z(Bϕ(f, g)), hiL2 as r → ∞. Similarly, since D(f ) and D(g) are supported in [−N, N ] the expression T Pr (f, g)(x) is uniformly bounded in r ∈ N and x ∈ R and converges to Tz(f, g)(x) as r → ∞. As h ∈ S(R) the dominated convergence theorem implies limr→∞hT Pr (f, g), hiL2 = hTz(f, g), hiL2 . Therefore, (4.14) is justified and hence, also (4.13). This completes the proof of Lemma 4.4. (cid:3) We now pass to the proof of Corollary 4.3. Proof. By repeating the argument from the beginning of the proof of Corollary 3.2 (with sums replaced by integrals) our task is reduced to d = 1. We devote the present paragraph to a brief justification of this statement Here we need the fact that for s ≥ 0 and Lj = −δ2 j , the operators (Lj)s(−∆κ)−s as well as (−∆κ)s(Pd j=1(Lj)s)−1, are bounded on all Lp, 1 < p < ∞. This is true by e.g. [24, Corollary 3.2], since in the product setting each Lj, j = 1, . . . , d, generates a symmetric contraction semigroup. Then we are left with showing that (4.17) kLs j(f g)kp . kLs jfkp1kgkp2 + kLs jgkp1kfkp2 cf. (3.7). The proof of (4.17) is similar to that of (3.7), thus we give a sketch when j = 1. For t ∈ R and x ∈ Rd−1, consider the auxiliary functions fx(t) = f ((t, x)) and gx(t) = g((t, x)). Then, setting w(1) i=2 wκi(x), we write κ (x) =Qd 1(f g)kp =ZRd−1 kLs kLs 1(fx(·)gx(·))kp Lp(R,wκ1 ) w(1) κ (x) dx. BILINEAR MULTIPLIERS VIA A FUNCTIONAL CALCULUS 21 From this point on we repeat the steps in the proof of (3.7). Namely, we apply the fractional Leibniz rule for d = 1 and Hölder's inequality (for integrals). We omit the details here. From now on we focus on proving Corollary 4.3 for d = 1. Setting Let φ be a function in C ∞([0,∞)) supported in [0, 1/4] and such that φ(t) + φ(t−1) = 1. T1(f, g)(x) =ZZR2 T2(f, g)(x) =ZZR2 and using Lemma 4.4 with ϕ ≡ 1 we rewrite φ(ξ2/ξ1)ξ1 + ξ22s D(f )(ξ1)D(g)(ξ2) E(iξ1, x)E(iξ2, x) w(ξ1)w(ξ2)dξ, φ(ξ1/ξ2)ξ1 + ξ22s D(f )(ξ1)D(g)(ξ2) E(iξ1, x)E(iξ2, x) w(ξ1)w(ξ2)dξ. (−∆κ)s(f g) = T1(f, g) + T2(f, g). From now on the proof resembles that of Corollary 3.2 (in fact it is even easier). We need to prove, for f, g ∈ A, the estimate kT1(f, g)kp ≤ Ck(−∆κ)sfkp1kgkp2 , kT2(f, g)kp ≤ Ckfkp1k(−∆κ)sgkp2. We focus only on the first inequality, as the proof of the second is analogous. For Re(z) ≥ 0 we set mz(ξ1, ξ2) = ξ1 + ξ22z ξ12z φ(ξ2/ξ1), ξ ∈ R2, so that T1(f, g) = Bms((−∆κ)sf, g). Since A is preserved under (−∆κ)s our task is reduced to showing that, for s > 0 it holds (4.18) kBms (f, g)kp ≤ Ckfkp1kgkp2 , f, g ∈ A As in Section 3 we are going to apply Stein's complex interpolation theorem. To do this we need to extend Bmz (f, g) outside of A × A, by allowing g to be a simple function. This may be achieved by a limiting process. Namely, instead of mz we consider mz . Then, ε = mze−εξ2 (4.19) Bms ε(f, g)(x) e−εξ2 φ(ξ2/ξ1) ξ1 + ξ22z :=ZRdZRd ξ12z D(f )(ξ1)D(g)(ξ2) E(iξ1, x)E(iξ2, x) w(ξ1)w(ξ2)dξ converges pointwise to Bms(f, g) as ε → 0+, whenever f, g ∈ A. Therefore, by Fatou's Lemma, to prove (4.18) for Bms it is enough to prove it for each Bms ε (f, g)kp ≤ Ckfkp1kgkp2 , ε, ε > 0, as long as kBms where C is independent of ε. The gain is that now (4.19) is well defined for g ∈ L2, in particular it is valid for simple functions. , j = 1, 2, v ∈ R, satisfy Hörmander's condition Let n > 2λκ + 6. Then the multipliers mn+iv (4.12) of order 2λκ + 6. Thus, using Corollary 4.2 (with d = 1) we obtain ε kBmn+iv ε (f, g)kp ≤ Cn(1 + v)2λκ+2kf1kp1kf2kp2, v ∈ R. Now, Lemma 4.4 applied to ϕ(ξ) = φ(ξ2/ξ1)e−ε(ξ2) implies Bmiv ε (f, g) = (−∆κ)iv(cid:2)Bϕ((−∆κ)−ivf, g)(cid:3). 22 BŁAŻEJ WRÓBEL Thus, using [8, Theorem 4.1] followed by Corollary 4.2 (for the multiplier φ(ξ2/ξ1)e−ε(ξ2)) we obtain kBmiv (f, g)kp ≤ C(1 + v)2λκ+2kf1kp1kf2kp2, v ∈ R. By definition Bmz ε (f, g) =ZZR2 ξ1 + ξ2z ξ1z φ(ξ2/ξ1)e−εξ2 D(f )(ξ1)D(g)(ξ2) E(iξ1, x)E(iξ2, x) w(ξ1)w(ξ2)dξ1dξ2. Hence, for fixed f1 ∈ A the family {Bmz (f, g)}Re(z)>0 consists of analytic operators. This family has admissible growth, more precisely, for each simple function h we have Consequently, using Stein's complex interpolation theorem is permitted and leads to (4.18). The proof of the corollary is thus finished. (cid:3) (cid:12)(cid:12)hBmz ((−∆κ)zf, g), hiL2(cid:12)(cid:12) ≤ Cf,g,h,s, Re(z) ≤ s. 5. Bilinear multipliers for Jacobi trigonometric polynomials In this section we give a bilinear multiplier theorem for expansions in terms of Jacobi trigono- metric polynomials. Contrary to the previous sections we do not prove a fractional Leibniz rule here. The reason for this is that there is no natural first order operator in the Jacobi setting that satisfies a Leibniz-type rule of integer order. Let α, β > −1/2 be fixed, and let P α,β n α, β. For n ∈ N and −1 < x < 1 these are given by the Rodrigues formula be the one-dimensional Jacobi polynomials of type P α,β n (x) = (−1)k 2nn! (1 − x)−α(1 + x)−β dn dxnh(1 − x)α+k(1 + x)β+ki. We now substitute x = cos θ, θ ∈ [0, π], and consider the trigonometric Jacobi polynomials P α,β n (cos θ). This is an orthogonal and complete system in L2(dµα,β), where θ θ dµα,β(θ) =(cid:16) sin (cos θ), where kP α,β 2(cid:17)2β+1 2(cid:17)2α+1(cid:16) cos n (cos ·)k2 = (cα,β dθ. Throughout this chapter we abbreviate Lp := Lp([0, π], µα,β ) and k·kp := k·kLp . Now, setting Pn(θ) = P α,β n )−1 we obtain a complete orthonormal system in L2. Each P α,β is an eigenfunction of the differential operator n (θ) = cα,β n P α,β n k J = J α,β = − d2 dθ2 − α − β + (α + β + 1) cos θ sin θ with the corresponding eigenvalue being (n+ α+β+1 observe that γ > 0. 2 In this setting the spectral multipliers of J 1/2 are given by µ(J 1/2)f = Xn∈N µ(cid:0)n + γ(cid:1)hf,PkiL2 Pk. d dθ +(cid:16) α + β + 1 2 (cid:17)2 ; )2. In what follows we set γ = (α+β +1)/2; BILINEAR MULTIPLIERS VIA A FUNCTIONAL CALCULUS 23 If µ : R+ → C is bounded, then µ(J 1/2) is a bounded operator on L2. In this section the formula (2.6) defining bilinear multipliers becomes (5.1) Bm(f1, f2)(θ) = m(J 1/2 ⊗ I, I ⊗ J 1/2)(x, x) = Xn1∈N,n2∈N m(cid:16)n1 + γ, n2 + γ(cid:17)hf1,Pn1ihf2,Pn2i Pn1(θ)Pn2(θ). The space A from (2.5) coincides with the linear span of {Pn}n∈N. We prove the following Coifman-Meyer type multiplier theorem. Theorem 5.1. Assume that m satisfies Hörmander's condition (2.2) of order s > 4(α+β)+15. Then the bilinear multiplier operator given by (5.1) is bounded from Lp1 × Lp2 to Lp, where 1/p1 + 1/p2 = 1/p, and p1, p2, p > 1. Moreover, the bound (2.7) is valid. Remark. The theorem implies a Coifman-Meyer type multiplier result for bilinear multipliers associated with the modified Hankel transform. This follows from a transference results of Sato [20]. Proof. Once again the proof hinges on Theorem 2.3. We need to verify that L = J 1/2 satisfies its assumptions. The injectivity condition is clear since 0 is not an eigenvalue of J 1/2. The contractivity assumption (CT) can be inferred from the formula e−tJ (f ◦ cos)(θ) = e−t(α+β+1)2/4T α,β t f (cos θ) relating the semigroup e−tJ with the semigroup T α,β from the Jacobi polynomial setting, as T α,β is well known to be Markovian. The condition (WD) is straightforward, since A is the t linear span of Jacobi trigonometric polynomials. The Mikhlin-Hörmander functional calculus (MH) for J 1/2 (with ρ = 2α + 2β + 13/2) was obtained in [26, Corollary 4.3]. It remains to show (PF). Here we need the following identity t (5.2) Pn1(θ)Pn2(θ) = j=n1+n2 Xj=n1−n2 cn1,n2(j)Pj (θ). The above is well known to hold for general orthogonal polynomials on an interval contained in R, hence also for Pj as they are merely a reparametrisation of the Jacobi polynomials. We prove that (PF) holds with b = 3. Take f, g ∈ A. Then f2 = Xn2∈N c2 n2Pn2, f1 = Xn1∈N c1 n1Pn1, where all but a finite number of c1 n, c2 n vanish. Denote Since ϕk and ψk are supported in [0, 2k−3] and [2k−1, 2k+1], respectively, we have Ra,b = {n ∈ N : 2a − γ ≤ n ≤ 2b − γ}. whereas ϕk(L)(f1) = Xn1∈N : n1+γ≤2k−3 ψk(L)(f2) = Xn2∈Rk−1,k+1 c1 n1 ϕk(n1 + γ)Pn1 c2 n2 ψk(n2 + γ)Pn2. 24 BŁAŻEJ WRÓBEL Now, if n1 + γ ≤ 2k−3 and 2k−1 ≤ n2 + γ ≤ 2k+1, then we must also have n1 − n2 ≥ 2k−1 − 2k−3 ≥ 2k−2 and n1 + n2 ≤ 2k−3 − γ + 2k+1 − γ ≤ 2k+2 − 2γ. Since γ > 0, we see that if n1− n2 ≤ n ≤ n1 + n2, then 2k−2 ≤ n + γ ≤ 2k+2. Consequently, in view of (5.2), the operator ψk(L) leaves invariant each product Pn1·Pn2 , hence, also ϕk(L)(f1)· ψk(L)(f2). The proof of (PF) is thus completed. (cid:3) Acknowledgments. I thank prof. Krzysztof Stempak for suggesting the idea to combine joint spectral multipliers with multilinear multipliers, prof. Christoph Thiele for a discussion on the Coifman-Meyer multiplier theorem and useful remarks during the preparation of the paper, prof. Camil Muscalu for a discussion on multilinear multipliers, prof. Herbert Koch, prof. Fulvio Ricci, and dr Gian Maria Dall'Ara for discussions on the fractional Leibniz rule, dr hab. Wojciech Młotkowski for bringing to my attention the formula (5.2), and dr Jotsaroop Kaur for discussions on Hermite polynomials. Part of the research presented in this paper was carried over while the author was Assegnista di ricerca at the Università di Milano-Bicocca. The research was supported by Italian PRIN 2010 "Real and complex manifolds: geometry, topology and harmonic analysis"; Polish funds for sciences, National Science Centre (NCN), Poland, Research Project 2014/15/D/ST1/00405; and by the Foundation for Polish Science START Scholarship. References [1] G. Alexopoulos, Spectral multipliers on discrete groups, Bull. Lond. Math. Soc. (4) 33 (2001), 417 -- 424. [2] B. Amri, J. P. Anker, and M. Sifi, Three results in Dunkl analysis, Colloquium Math. (1) 118 (2010), pp. 299-312. [3] B. Amri, A. Gasmi, M. Sifi, Linear and Bilinear Multiplier Operators for the Dunkl Transform, Mediterr. J. Math. (4) 7 (2010), 503 -- 521. [4] F. Bernicot, D. Maldonado, K. Moen, and V. Naibo, Bilinear Sobolev-Poincaré inequalities and Leibniz-type rules, J. Geom. Anal. (2) 24 (2014), pp. 1144-1180. [5] O. Blasco, Bilinear multipliers and transference, Int. J. Math. Math. Sci. (4) 2005, pp. 545-554. [6] J. Bourgain, D. Li, On an endpoint Kato-Ponce inequality, Diff. Int. Eq. (11/12) 27 (2014), pp. 1037-1072. [7] R. Coifman, Y. Meyer, Nonlinear harmonic analysis, operator theory, and PDE in Beijing Lectures in Harmonic Analysis (Beijing, 1984), Annals of Mathematics Studies 112, Princeton Univ. Press, Princeton, NJ, 1986, pp. 3 -- 45. [8] F. Dai, H. Wang, A transference theorem for the Dunkl transform and its applications, J. Funct. Anal. (12) 258 (2010), pp. 4052-4074. [9] D. Fan, S. Sato, Transference on certain multilinear multiplier operators, J. Austral. Math. Soc., 70 (2001), pp.37-55. [10] D. Frey, Paraproducts via H ∞-functional calculus, Rev. Math. Iberoam. (2) 29 (2013), 635-663. [11] L. Grafakos, S. Oh, The Kato-Ponce inequality, Commun. Part. Diff. Eq. (6) 39 (2014), pp. 1128-1157. [12] L. Grafakos, R. Torres, Multilinear Calderón-Zygmund theory, Adv. Math. 165 (2002), pp. 124-164. [13] T. Kato, G. Ponce, Commutator estimates and the Euler and Navier-Stokes equations, Commun. Pure Appl. Math., 41 (1988), pp. 891-907. [14] C. Kenig, G. Ponce, and L. Vega, Well-posedness and scattering results for the generalized Korteweg-de Vries equation via the contraction principle, Comm. Pure Appl. Math., 46 (1993), pp. 527-620. [15] C. Kenig, E. M. Stein, Multilinear estimates and fractional integrals, Math. Res. Lett. 6 (1999) pp. 1-15. [16] A. Gulisashvili, M. A. Kon, Exact Smoothing Properties of Schrödinger Semigroups, Amer. J. Math. (6) 118 (1996), pp. 1215-1248. [17] C. Muscalu, J. Pipher, T. Tao, and C. Thiele, Bi-parameter paraproducts, Acta Math. 193 (2004), pp. 269-296. [18] C. Muscalu, W. Schlag, Classical and multilinear harmonic analysis, Vol II, Cambridge Studies in advanced mathematics 138, 2013. BILINEAR MULTIPLIERS VIA A FUNCTIONAL CALCULUS 25 [19] M. Rösler, Dunkl operators: theory and applications, in Orthogonal Polynomials and Special Functions (Leuven, 2002), Lecture Notes in Math., Vol. 1817, Springer, Berlin, 2003, 93Ű135. [20] E. Sato, Transference of bilinear operators between Jacobi series and Hankel transforms, Taiwanese J. Math. (4) 15 (2011), pp. 1561-1573. [21] K. Schmüdgen, Unbounded Self-adjoint Operators on Hilbert Space, Grad. Texts in Math. 265 (2012). [22] E. Stein, Interpolation of Linear Operators, Trans. Amer. Math. Soc. (2) 83 (1956), pp. 482-492. [23] B. Wróbel, On the consequences of a Mihlin-Hörmander functional calculus: maximal and square function estimates, preprint 2015, arXiv:1507.08114. [24] B. Wróbel, Joint spectral multipliers for mixed systems of operators, J. Fourier Anal. Appl. (2016) online first, DOI:10.1007/s00041-016-9469-7, pp. 1-43. [25] B. Wróbel, Multivariate spectral multipliers, PhD thesis, Scuola Normale Superiore, Pisa and Uniwersytet Wrocławski (2014), http://arxiv.org/abs/1407.2393. [26] B. Wróbel, Multivariate spectral multipliers for tensor product orthogonal expansions, Monatsh. Math. 168 (2012), 124 -- 149. Błażej Wróbel: Mathematical Institute, Universität Bonn, Endenicher Allee 60, D -- 53115 Bonn, Germany & Instytut Matematyczny, Uniwersytet Wrocławski, pl. Grunwaldzki 2/4, 50-384 Wrocław, Poland [email protected] E-mail address: [email protected]
1005.0391	1	1005	2010-05-03T20:03:30	On the higher rank numerical range of the shift	[ "math.FA", "math.CV", "math.OA" ]	For any n-by-n complex matrix T and any $1\leqslant k\leqslant n$, let $\Lambda_{k}(T)$ the set of all $\lambda\in \C$ such that $PTP=\lambda P$ for some rank-k orthogonal projection $P$ be its higher rank-k numerical range. It is shown that if $\bbS$ is the n-dimensional shift on ${\C}^{n}$ then its rank-k numerical range is the circular disc centred in zero and with radius $\cos\dfrac{k\pi}{n+1}$ if $1<k\leqslant[\frac{n+1}{2}]$ and the empty set if $[\frac{n+1}{2}]<k\leqslant n$, where $[x] $ denote the integer part of $x$. This extends and rafines previous results of U. Haagerup, P. de la Harpe \cite{Haagerup} on the classical numerical range of the n-dimensional shift on${\C}^{n}$. An interesting result for higher rank-$k$ numerical range of nilpotent operator is also established.	math.FA	math	ON THE HIGHER RANK NUMERICAL RANGE OF THE SHIFT OPERATOR HAYKEL GAAYA centred in zero and with radius cos Abstract. For any n-by-n complex matrix T and any 1 6 k 6 n, let Λk(T ) the set of all λ ∈ CI such that P T P = λP for some rank-k orthogonal pro- jection P be its higher rank-k numerical range. It is shown that if Sn is the n-dimensional shift on CI n then its rank-k numerical range is the circular disc 2 (cid:3) and the empty set if (cid:2) n+1 2 (cid:3) < k 6 n, where [x] denote the integer part of x. This extends and rafines previous results of U. Haagerup, P. de la Harpe [8] on the classical numerical range of the n-dimensional shift on CI n. An interesting result for higher rank-k numerical range of nilpotent operator is also established. if 1 < k 6 (cid:2) n+1 kπ n + 1 1. Introduction Let H be a complex separable Hilbert space and B(H) the collection of all bounded linear operator on H. The numerical range of an operators T in B(H) is the subset W (T ) = {< T x, x >∈ CI ; x ∈ H, kxk6 1} of the plane, where < ., . > denotes the inner product in H and the numerical range of T is defined by We denote by S the unilateral shift acting on the Hardy space HI 2 of the square ω2(T ) = sup {z; z ∈ W (T )} . summable analytic functions. S : HI 2 → HI 2 7→ zf (z) f Beurling's theorem implies that the non zero invariant subspaces of S are of the forme φ HI 2, where φ is some inner function . Let S(φ) denote the compression of S to the space H(φ) = HI 2 ⊖ φ HI 2 : S(φ)f (z) = P (zf (z)), where P denotes the ortogonal projection from HI 2 onto H(φ). The space H(φ) is a finite-dimensional exactly when φ is a finite Blaschke product. The numerical radius and numerical range of the model operator S(φ) seems to be important and have many applications. In [1], Badea and Cassier showed that there is relation- ship between numerical radius of S(φ) and Taylor coefficients of positive rational functions on the torus and more recently in [6], the author gave an extension of this result. However the evaluation of the numerical radius of S(φ) under an explicit form is always an open problem. The reader may consult [6] for an estimate of S(φ) 2000 Mathematics Subject Classification. 47A12, 47B35. Key words and phrases. Operator theory, Numerical radius, Numerical range, higher rank numerical range, Eigenvalues, Toeplitz forms. 1 2 HAYKEL GAAYA where φ is a finite Blashke product with unique zero. In the particular case where φ(z) = zn, S(φ) is unitarily equivalent to Sn where Sn =  0 1 . . . . . . . . . 1 0 .   n+1o In [8]; it is proved that W (Sn) is the closed disc Dn = nz ∈ CI ; z6 cos π and ω2(Sn) = cos π n+1 and more general Theorem 1.1 ([8]). Let T be an operator on H such that T n = 0 for some n ≥ 2. One has: ω2(T ) 6 kT kcos π n + 1 and ω2(T ) = kT kcos π n+1 when T is unitarily equivalent to kT kSn. In this mathematical note, we extend this result to the higher rank-k numerical range of the shift. The notion of the the higher rank-k numerical range of T ∈ B(H) is introduced in [4] and it's denoted by: Λk(T ) = {λ ∈ CI : P T P = λP for some rank-k orthogonal projection P } , The introduction of this notion was motivated by a problem in quantum error correction; see [5]. If P is a rank-1 orthogonal projection then P = x ⊗ x for some x ∈ CI n and P T P =< T x, x > P . Then when k = 1, this concept is reduces to the classical numerical range W (T ), which is well known to be convex by the Toeplitz- Hausdorff theorem; for exemple see [10] for a simple proof. In [2], it's conjectured that Λk(T ) is convex, and reduced the convexity problem to the problem of showing that 0 ∈ Λk(T ′) where T ′ =(cid:18) Ik X Y −Ik (cid:19) for arbitrary X, Y ∈ Mk (the algebra of k × k complex matrix). They further reduced this problem to the existance of a Hermitian matrix H satisfying the matrix equation (1.1) Ik + M H + HM ∗ − HRH = H for arbitrary M ∈ Mk and a positive definite R ∈ Mk. In [16], H. Woerdeman proved that equation (1.1) is equivalent to Ricatti equation: (1.2) HRH − H(M ∗ − Ik/2) − (M − Ik/2)H − Ik = 0k, and using the theory of Ricatti equations (see [9], Theorem 4), the equation (1.2) is solvable which prove the convexity of Λk(T ). In [4], the authors showed that if dimH < ∞ and T ∈ B(H) is a Hermitian matrix with eigenvalues λ1 6 λ2 · · · 6 λn then the rank-k nuemrical range Λk(T ) coincides with [λk, λn+1−k] which is a non- degenerate closed interval if λk < λn+1−k, a singleton set if λk = λn+1−k and an empty set if λk > λn+1−k. In [13], the authors proved that if dimH = n Λk(T ) = \θ∈[0,2π[(cid:8)µ ∈ CI : eiθµ + e−iθµ 6 λk(cid:0)eiθT + e−iθT ∗(cid:1)(cid:9) , ON THE HIGHER RANK NUMERICAL RANGE OF THE SHIFT OPERATOR 3 for 1 6 k 6 n, where λk(H) denote the kth largest eigenvalue of the hermitian matrix H ∈ Mn. This result establishes that if dimH = n and T ∈ B(H) is a normal matrix with eigenvalues λ1, . . . , λn then Λk(T ) = \ 16j1<···<jn−k+16n conv(cid:8)λj1 , . . . , λjn+1−k(cid:9) . We close this section by the following properties wich are easly checked. The reader may consult [2],[3],[4],[5],[7] and [11]. P1. For any a and b ∈ CI , Λk(aT + bI) = aΛk(T ) + b. P2. Λk(T ∗) = Λk(T ). P3. Λk(T ⊕ S) ⊇ Λk(T ) ∪ Λk(S). P4. For any unitary U ∈ B(H), Λk(U ∗T U ) = Λk(T ). P5. If T0 is a compression of T on a subspace H0 of H such that dimH0 ≥ k, then Λk(T0) ⊆ Λk(T ). P6. W (T ) ⊇ Λ2(T ) ⊇ Λ3(T ) ⊇ . . . . Some results from [1] will be also developed in this context in a forthcoming paper. 2. main theorem In the following theorem we give the higher rank-k numerical range of the n- dimensional shift on CI n. Theorem 2.1. For any n ≥ 2 and 1 6 k 6 n, Λk(Sn) coincides with the circular disc {z ∈ CI : z6 cos kπ n + 1 } if 1 6 k 6(cid:2) n+1 2 (cid:3) and the empty set if(cid:2) n+1 2 (cid:3) < k 6 n. Proof. First observe that Λk(Sn) = \θ∈[0,2π[(cid:8)µ ∈ CI : eiθµ + e−iθµ 6 λk(cid:0)eiθSn + e−iθS ∗ n(cid:1)(cid:9) n(cid:1)(cid:27) λk(cid:0)eiθSn + e−iθS ∗ n(cid:1)(cid:27) λk(cid:0)eiθSn + e−iθS ∗ = \θ∈[0,2π[(cid:26)µ ∈ CI : Re(eiθµ) 6 eiθ(cid:26)z ∈ CI : Re(z) 6 = \θ∈[0,2π[ 1 2 1 2 (2.1) On the other hand, we have eiθSn + e−iθS ∗ n =   0 eiθ 0 ... 0 0 e−iθ 0 eiθ ... 0 0 0 e−iθ 0 ... 0 0 . . . . . . . . . . . . . . . . . . 0 0 0 ... 0 eiθ 0 . . . 0 . . . 0 ... e−iθ 0 .   Note that eiθSn + e−iθS ∗ n is a Toeplitz matrix associated to the Toeplitz form fθ(t) = 2 cos(θ + t). 4 HAYKEL GAAYA The eigenvalues satisfy the caracteristic equation ∆n(λ) = Det(cid:0)eiθSn + e−iθS ∗ n(cid:1) 0 = (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) −λ e−iθ eiθ −λ eiθ 0 ... ... 0 0 0 0 0 0 0 ... 0 . . . 0 . . . . . . . . . . . . 0 ... . . . . . . −λ e−iθ eiθ −λ . . . e−iθ −λ ... 0 0 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Expanding this determinant, we obtain the recurrence relation ∆n(λ) = −λ∆n−1 − ∆n−2, n = 2, 3, 4, . . . , This recurrence relation holds also for n = 1 provided we put ∆0 = 1 and ∆−1 = 0. In order to find an explicit representation of ∆n(λ), we write convenently λ = 2 cos(θ + t) = fθ(t) and form the caracteristic equation ρ2 = −λρ − 1 = −2ρ cos(θ + t) − 1 with the roots −ei(θ+t) and −e−i(θ+t) so that ∆n(2 cos(θ + t)) = (−1)n(Aein(θ+t) + Be−in(θ+t)) where the constants A and B can be determined from the cases n = −1 and n = 0. Thus ∆n(2 cos(θ + t)) = (−1)n sin((n + 1)(θ + t)) sin(θ + t). This yields the eigenvalues λν = 2 cos( νπ n + 1 ), ν = 1, 2, . . . n. This implies of course that Λk(Sn) = \θ∈[0,2π[ eiθ(cid:26)z ∈ CI : Re(z) 6 cos( kπ n + 1 )(cid:27) Thus Λk(Sn) is the intersection of closed half planes. We note that cos( positive if and only if k 6(cid:2) n+1 2 (cid:3). Case 1.If k 6(cid:2) n+1 Case 2. If k >(cid:2) n+1 Λk(Sn) ⊆ (cid:26)z ∈ CI : Re(z) 6 cos( = ∅. 2 (cid:3) In this case Λk(Sn) is circular disc {z ∈ CI : z6 cos 2 (cid:3), then kπ n + 1 )(cid:27)\ eiπ(cid:26)z ∈ CI : Re(z) 6 cos( kπ n + 1 ) is kπ n + 1 }. kπ n + 1 )(cid:27) (cid:3) This completes the proof. Theorem 2.2. For any integer k ≥ 1 Λk(S) = D(0, 1) ON THE HIGHER RANK NUMERICAL RANGE OF THE SHIFT OPERATOR 5 Proof. Let a fixed k ≥ 1, we have D(0, 1) ⊆ Λk(S) which is du to (P.5) and theorem (2.1). Now let λ in Λk(S) then there exists a rank-k orthogonal projection P such that P SP = λP . Let denote by Uθ the unitary operator on HI 2 defined by Uθ(f )(z) = f (ze−iθ), then if we denote by Q the rank-k orthogonal projection θ we can easly check that QSQ = λeiθ which implies that Λk(S) is a circular UθP U ∗ disc centred in 0. On the other hand if 1 ∈ Λk(S) then 1 ∈ W (S) and there exists a unitary f ∈ HI 2 such that < Sf, f >= 1 wich implies that 1 is un eigenvalue for S which is absurd. (cid:3) On the sequel of this paper, let denote by ρ(k, r) =(cid:26) k/r [k/r] + 1 unless if k/r is is integer where k and r are arbitrary numbers. Lemma 2.3. For a fixed n ≥ 1 and r ≥ 1, let denote by λ1 > · · · > λn; n real numbers and (λ′ p)16p6nr a finite sequence defined by: 1 = · · · = λ′ r = λ1, . . . , λ′ (n−1)r+1 = · · · = λ′ λ′ nr = λn. Then for each 1 6 k 6 nr, the kth largest term of (λ′ t)16t6nr is λρ(k,r). Proof. The claim is obvious in the case where r = 1. We may assume r ≥ 2. We prove the result by induction on k. If k = 1, then the largest term is λ1 = λρ(1,r). So the result hold for k = 1. Assume that k > 1, and the reslut is valid for the mth largest term of (λ′ t)16t6nr whenever m < k. Case 1. Suppose that ρ(k − 1, r) = k−1 k − 1 = pr. By induction assumption, we have λρ(k−1,r) = λ′ that the kth largest term of (λ′ r , then there exists 1 6 p 6 n − 1 such that pr = λp, which implies t)16t6nr is λ′ pr+1 = λp+1 = λ k−1 r +1 = λ[ k r ]+1 = λρ(k,r). Case 2. Suppose that ρ(k − 1, r) = [ k−1 r ] + 1, then there exist 1 6 q 6 n − 1 and 1 6 s 6 r − 1 such that k − 1 = qr + s. First, note that ρ(k − 1, r) = ρ(k, r). On the other hand, by induction assumption, we have λρ(k−1,r) = λ′ qr+s = λq+1. Consequently the kth largest term of (λ′ t)16t6nr is λ′ qr+s+1 = λq+1 = λρ(k−1,r) = λρ(k,r). The proof is now complete. (cid:3) Let DT = (IN − T ∗T )1/2 be the defect operator of T and DT the closed range of DT . Let denote by r = dimDT . Theorem 2.4. Consider T ∈ B(H) such that kT k6 1 and T n = 0. Then Λk(T ) is contained in the circular disc {z ∈ CI : z6 cos( ρ(k,r)π n+1 )} if 1 6 ρ(k, r) 6 [ n+1 2 ] and empty if ρ(k, r) > [ n+1 2 ]. Proof. If T is a contaction with T n = 0, then T can be viewed as a compression of n acting on the Hilbert space DT ⊗ CI n. Consider the isometry H → DT ⊗ CI n, Ir ⊗S ∗ V (x) =Xn t=1 DT T t−1x ⊗ et 6 HAYKEL GAAYA where {el}n l=1 is the canonical basis of CI n. Note that V T x =Xn t=1 It follows that DT T tx ⊗ et =Xn−1 t=1 DT T tx ⊗ et = (Ir ⊗ S ∗ n)V x. T = V ∗(Ir ⊗ S ∗ n)V and from (P.5) (2.2) Λk(T ) = Λk(V ∗(Ir ⊗ S ∗ n)V ) ⊆ Λk(Ir ⊗ S ∗ n), for any 1 6 k 6 nr. Now, Λk(Ir ⊗ S ∗ n) n) + e−iθ(Ir ⊗ S ∗ n)∗(cid:1)(cid:9) n) + e−iθ(Ir ⊗ Sn)(cid:1)(cid:9) = \θ∈[0,2π[(cid:8)µ ∈ CI : eiθµ + e−iθµ 6 λk(cid:0)eiθ(Ir ⊗ S ∗ = \θ∈[0,2π[(cid:8)µ ∈ CI : eiθµ + e−iθµ 6 λk(cid:0)eiθ(Ir ⊗ S ∗ = \θ∈[0,2π[(cid:8)µ ∈ CI : eiθµ + e−iθµ 6 λk(cid:0)Ir ⊗ (eiθSn + e−iθS ∗ n)(cid:1)(cid:9) = \θ∈[0,2π[(cid:8)µ ∈ CI : eiθµ + e−iθµ 6 λk(cid:0)⊕r n)(cid:1)(cid:9) = \θ∈[0,2π[ eiθ(cid:26)z ∈ CI : Re(z) 6 cos( i (eiθSn + e−iθS ∗ ρ(k, r)π n + 1 ) (cid:27) where the last equality is due to the lemma (2.2) and theorem (2.1). Thus Λk(Ir ⊗ S ∗ n) =( D(0, cos( ρ(k,r)π n+1 )) ∅ if 1 6 ρ(k, r) 6 [ n+1 2 ] if [ n+1 2 ] < ρ(k, r) 6 n Therefore, if 1 6 k 6 nr, (2.2) implies that Λk(T ) ⊆ D(0, cos( ρ(k,r)π and empty if [ n+1 2 ] < ρ(k, r) 6 n. Finally, if k > nr, Λk(T ) = ∅ from (P6). n+1 )) if 1 6 ρ(k, r) 6 [ n+1 2 ] (cid:3) Corollary 2.5 (U. Haagerup, P. de la Harpe,[8]). Consider T ∈ B(H) such that kT k6 1 and T n = 0. Then we have ω2(T ) 6 cos( π n+1 ). Proof. T = V ∗(Ir ⊗ S ∗ n)V where V : H → DT ⊗ CI n, Now V (x) =Xn t=1 DT T t−1x ⊗ et. W (T ) = Λ1(T ) = Λ1(V ∗(Ir ⊗ Sn)V ) ⊆ Λ1(Ir ⊗ Sn) = D(0, cos π n + 1 ). (cid:3) Acknowledgements: The author would like to express his gratitude to Gilles Cassier for his help and his good advices. ON THE HIGHER RANK NUMERICAL RANGE OF THE SHIFT OPERATOR 7 References [1] C. Badea and G. Cassier, Constrained von neumann inequalities, Adv. Math. 166 (2002), no. 2, 260 -- 297. [2] M.-D. Choi, M. Giesinger, J. A. Holbrook, and D. W. Kribs, Geometry of higher-rank nu- merical ranges, Linear and Multilinear Algebra 56 (2008), 53-64. [3] M.-D. Choi, J. A. Holbrook, D.W. Kribs, and K. Zyczkowski, Higher-rank numerical ranges of unitary and normal matrices, Operators and Matrices 1 (2007), 409-426. [4] M.-D. Choi, D. W. Kribs, and K. Zyczkowski, Higher-rank numerical ranges and compression problems, Linear Algebra Appl. 418 (2006), 828-839. [5] M.-D. Choi, D. W. Kribs, and K. Zyczkowski, Quantum error correcting codes from the compression formalism, Rep. Math. Phys. 58 (2006), 77-91. [6] H. Gaaya, On the numerical radius of the truncated adjoint Shift. to appear. [7] H.L. Gau, C.K. Li, P.Y. Wu, Higher-rank numerical ranges and dilations, J. Operator Theory, in press. [8] U.Haagerup and P. dela Harpe, The numerical radius of a nilpotent operator on a Hilbert space, Proc. Amer. Math. Soc. 115(1992), 371 -- 379. [9] P. Lancaster and L. Rodman, Algebraic Riccati equations, Oxford Science Publications, The Clarendon Press, Oxford University Press, New York, 1995. [10] C.-K. Li, A simple proof of the elliptical range theorem, Proc. Amer. Math. Soc. 124 (1996), 1985-1986. [11] C.-K. Li, Y. T. Poon and N.-S. Sze, Condition for the higher rank numerical range to be non-empty, Linear and Multilinear Algebra, to appear. [12] C.-K. Li, Y. T. Poon and N.-S. Sze, Higher rank numerical ranges and low rank perturbations of quantum channels, preprint. http://arxiv.org/abs/0710.2898. [13] C.-K. Li, N.-S. Sze, Canonical forms, higher rank numerical ranges, totally isotropic sub- spaces, and matrix equations, Proc. Amer. Math. Soc. Volume 136, Number 9, September 2008, Pages 3013 -- 3023 . [14] C. Pop, On a result of Haagerup and de la Harpe. Rev. Roumaine Math. Pures Appl. 43 (1998), no. 9-10, 869 -- 871. [15] G. W. Stewart and J.-G. Sun, Matrix Perturbation Theory, Academic Press, New York, 1990. [16] H. Woerdeman, The higher rank numerical range is convex, Linear and Multilinear Algebra 56 (2008), 65-67. ‡.Institute Camille Jordan, Office 107 University of Lyon1, 43 Bd November 11, 1918, 69622-Villeurbanne, France. E-mail address: ‡[email protected]
1306.2423	1	1306	2013-06-11T05:31:18	Numerical Radii for Tensor Products of Matrices	[ "math.FA" ]	For $n$-by-$n$ and $m$-by-$m$ complex matrices $A$ and $B$, it is known that the inequality $w(A\otimes B)\le\\|A\\|w(B)$ holds, where $w(\cdot)$ and $\\|\cdot\\|$ denote, respectively, the numerical radius and the operator norm of a matrix. In this paper, we consider when this becomes an equality. We show that (1) if $\\|A\\|=1$ and $w(A\otimes B)=w(B)$, then either $A$ has a unitary part or $A$ is completely nonunitary and the numerical range $W(B)$ of $B$ is a circular disc centered at the origin, (2) if $\\|A\\|=\\|A^k\\|=1$ for some $k$, $1\le k<\infty$, then $w(A)\ge\cos(\pi/(k+2))$, and, moreover, the equality holds if and only if $A$ is unitarily similar to the direct sum of the $(k+1)$-by-$(k+1)$ Jordan block $J_{k+1}$ and a matrix $B$ with $w(B)\le\cos(\pi/(k+2))$, and (3) if $B$ is a nonnegative matrix with its real part (permutationally) irreducible, then $w(A\otimes B)=\\|A\\|w(B)$ if and only if either $p_A=\infty$ or $n_B\le p_A<\infty$ and $B$ is permutationally similar to a block-shift matrix \[[ {array}{cccc} 0 & B_1 & & & 0 & \ddots & & & \ddots & B_k & & & 0 {array} ]\] with $k=n_B$, where $p_A=\sup\{\ell\ge 1: \\|A^{\ell}\\|=\\|A\\|^{\ell}\}$ and $n_B=\sup\{\ell\ge 1 : B^{\ell}\neq 0\}$.	math.FA	math	Numerical Radii for Tensor Products of Matrices Hwa-Long Gaua∗, Kuo-Zhong Wangb, Pei Yuan Wub aDepartment of Mathematics, National Central University, Chung-Li 32001, Taiwan bDepartment of Applied Mathematics, National Chiao Tung University, Hsinchu 30010, Taiwan Abstract. For n-by-n and m-by-m complex matrices A and B, it is known that the inequality w(A ⊗ B) ≤ kAkw(B) holds, where w(·) and k · k denote, respectively, the numerical radius and the operator norm of a matrix. In this paper, we consider when this becomes an equality. We show that (1) if kAk = 1 and w(A ⊗ B) = w(B), then either A has a unitary part or A is completely nonunitary and the numerical range W (B) of B is a circular disc centered at the origin, (2) if kAk = kAkk = 1 for some k, 1 ≤ k < ∞, then w(A) ≥ cos(π/(k + 2)), and, moreover, the equality holds if and only if A is unitarily similar to the direct sum of the (k + 1)-by-(k + 1) Jordan block Jk+1 and a matrix B with w(B) ≤ cos(π/(k + 2)), and (3) if B is a nonnegative matrix with its real part (permutationally) irreducible, then w(A ⊗ B) = kAkw(B) if and only if either pA = ∞ or nB ≤ pA < ∞ and B is permutationally similar to a block-shift matrix 0 B1 0 . . . . . . Bk 0     with k = nB, where pA = sup{ℓ ≥ 1 : kAℓk = kAkℓ} and nB = sup{ℓ ≥ 1 : Bℓ 6= 0}. Keywords: numerical range; numerical radius; tensor product; Sn-matrix; nonneg- ative matrix AMS Subject Classifications: 15A60; 15A69; 15B48 ∗Corresponding author. Email: [email protected] 1 Introduction and Preliminaries For any n-by-n complex matrix A, its numerical range W (A) is, by definition, the subset {hAx, xi : x ∈ Cn,kxk = 1} of the complex plane C, where h·,·i and k·k denote the standard inner product and its associated norm in Cn, respectively. The numerical radius w(A) of A is max{z : z ∈ W (A)}. It is known that W (A) is a nonempty compact convex subset of C, and w(A) satisfies kAk/2 ≤ w(A) ≤ kAk, where kAk denotes the usual operator norm of A. For other properties of the numerical range and numerical radius, the reader may consult [7], [9, Chapter 22] or [12, Chapter 1]. The tensor product (or Kronecker product) A⊗B of an n-by-n matrix A = [aij]n i,j=1 and an m-by-m matrix B is the (mn)-by-(mn) matrix   .   a11B · · · a1nB ... ... an1B · · · annB It is known that A ⊗ B and B ⊗ A are unitarily similar and kA ⊗ Bk = kAk · kBk. Other properties of the tensor product can be found in [12, Chapter 4]. The main concern of this paper is the relations between the numerical radius of A⊗ B and those of A and B. For one direction, we have w(A⊗B) ≤ min{kAkw(B),kBkw(A)}. This can be proven by using the unitary dilation of contractions, as to be done below. On the other hand, we also have w(A ⊗ B) ≥ w(A)w(B). We are interested in when these become equalities. In the present paper, we obtain various conditions, neces- sary or sufficient, for w(A ⊗ B) = kAkw(B) to hold. The discussions on the equality w(A ⊗ B) = w(A)w(B) will be the subject of a subsequent paper of ours. For the ease of exposition, we introduce two indices for an n-by-n matrix A: the power norm index pA and nilpotency index nA of A. They are defined, respectively, by pA = sup{k ≥ 1 : kAkk = kAkk} 2 and sup{k ≥ 1 : Ak 6= 0n} 0 if A 6= 0n, if A = 0n, nA =  where 0n denotes the n-by-n zero matrix. We start in Section 2 by proving that if kAk = 1 and w(A ⊗ B) = w(B), then either A has a unitary part or A is completely nonunitary and W (B) is a circular disc centered at the origin (Theorem 2.2). The proof depends on the dilation of A to a direct sum of Sℓ-matrices with ℓ ≤ n, the Poncelet property of the numerical ranges of matrices of the latter class, and Anderson's theorem on the circular disc numerical range. As a by-product, we obtain a lower bound for w(A) when A satisfies kAk = kAkk = 1 for some k, 1 ≤ k < n: w(A) ≥ cos(π/(k+2)), and determine exactly when this bound is attained: this is the case if and only if A is unitarily similar to Jk+1 ⊕ B, where Jk+1 is the (k + 1)-by-(k + 1) Jordan block 0 1 0 . . . . . . 1 0     and B is a finite matrix with w(B) ≤ cos(π/(k + 2)) (Theorem 2.10). This generalizes the classical result of Willams and Crimmins [17] for k = 1. We conclude Section 2 with a result on nilpotent contractions, namely, we prove that if A is an n-by-n matrix with kAk = 1, then a necessary and sufficient condition for pA = nA < ∞ to hold is that A be unitarily similar to a direct sum Jk+1 ⊕ B, where k = pA and Bk+1 = 0 (Theorem 2.13). Finally, in Section 3, we consider B to be a nonnegative matrix with Re B (= (B + B∗)/2) (permutationally) irreducible. We obtain in Theorem 3.1 a complete characterization for w(A ⊗ B) = kAkw(B), namely, this is the case if and only if either pA = ∞ or nB ≤ pA < ∞ and B is permutationally similar to a block-shift 3 matrix of the form with k = nB. 0 B1 0   . . . . . . Bk 0   As was mentioned before, the inequality w(A ⊗ B) ≤ kAkw(B) for n-by-n and m-by-m matrices A and B is known. It is a consequence of [10, Theorem 3.4] because A ⊗ B is the product of A ⊗ Im and In ⊗ B, and the latter two matrices doubly commute, that is, A ⊗ Im commutes with both In ⊗ B and its adjoint In ⊗ B∗. Here we give a simple proof based on the unitary dilation of contractions. Proposition 1.1. If A and B are n-by-n and m-by-m matrices, respectively, then w(A ⊗ B) ≤ min{kAkw(B),kBkw(A)}. Proof. We need only prove that w(A⊗B) ≤ kAkw(B), and may assume that kAk = 1. Then the (2n)-by-(2n) matrix U =  A (In − A∗A)1/2 (In − AA∗)1/2 −A∗   is unitary. Let U be unitarily similar to the diagonal matrix diag (u1, . . . , u2n), where uj = 1 for all j. Then w(A ⊗ B) ≤ w(U ⊗ B) = w( 2n Xj=1 ⊕ujB) = max j w(ujB) = w(B) = kAkw(B). We conclude this section with some basic properties of the indices pA and nA of a matrix A. Proposition 1.2. Let A be an n-by-n matrix. Then (a) 1 ≤ pA ≤ n − 1 or pA = ∞, 4 (b) pA = n − 1 if and only if A is a nonzero multiple of a Sn-matrix, and (c) the following conditions are equivalent: (1) pA = ∞, (2) kAk = ρ(A), (3) kAk = w(A), and if kAk = 1, then the above are also equivalent to (4) A has a unitary part. Here ρ(A) denotes the spectral radius of A, that is, ρ(A) is the maximum modulus of eigenvalues of A. Recall that an n-by-n matrix A is of class Sn if it is a contraction (kAk ≤ 1), its eigenvalues are all in D ≡ {z ∈ C : z < 1}, and rank (In − A∗A) = 1. Any contraction A is unitarily similar to the direct sum of a unitary matrix U, called the unitary part of A, and a completely nonunitary contraction A′, called the c.n.u. part of A. The latter means that A′ is not unitarily similar to any direct sum with a unitary summand. Proof of Proposition 1.2. (a) was obtained by Pt´ak in 1960 (cf. [15, Theorem 2.1]) and (b) was proven in [4, Theorem 3.1]. As for (c), the implication (1) ⇒ (2) is by [9, Problem 88], (2) ⇒ (3) by the known inequalities ρ(A) ≤ w(A) ≤ kAk, (3) ⇒ (2) by [9, Problem 218 (b)], and (2) ⇒ (1) by the inequalities ρ(A) ≤ kAkk1/k ≤ kAk for all k ≥ 1. If kAk = ρ(A) = 1, then, letting λ be an eigenvalue of A with λ = 1, we have the unitary similarity of A and a matrix of the form  . Since kAk = λ = 1  implies that B = 0, A is unitarily similar to [λ]⊕ C and thus has a unitary part. This proves (2) ⇒ (4). That (4) ⇒ (2) is trivial. 0 C   λ B Proposition 1.3. Let A be an n-by-n matrix. Then 5 (a) 0 ≤ nA ≤ n − 1 or nA = ∞, (b) nA = n − 1 if and only if A is similar to the n-by-n Jordan block Jn, (c) nA = ∞ if and only if A is not nilpotent, and (d) pA ≤ nA for A 6= 0n. We omit its easy proofs. In the following, we use σ(A) to denote the spectrum of A, that is, σ(A) is the set of eigenvalues of A. An n-by-n matrix A is a dilation of an m-by-m matrix B (or B is a compression of A) if there is an n-by-m matrix V such that B = V ∗AV and V ∗V = Im. This is equivalent to A being unitarily similar to a matrix of the form   B ∗ ∗ ∗ .   2 Contractions We start with a simple condition which yields the equality w(A ⊗ B) = kAkw(B). Lemma 2.1. If A is an n-by-n matrix with pA = ∞, then w(A ⊗ B) = kAkw(B) for any m-by-m matrix B. In particular, this is the case for A a contraction with a unitary part. Proof. Since pA = ∞ implies, by Proposition 1.2 (c), that kAk = w(A). If λ is a number in W (A) with λ = w(A), then λ = kAk. Since A is unitarily similar to a matrix of the form   .   It follows that kAkw(B) = w(λB) ≤ w(A ⊗ B). On the other hand, we also have w(A ⊗ B) ≤ kAkw(B) by Proposition 1.1. Thus w(A ⊗ B) = kAkw(B) holds. , we have the unitary similarity of A ⊗ B and   λB ∗ ∗ ∗ λ ∗ ∗ ∗   The next theorem is one of the main results of this section. It gives a necessary condition for the equality w(A ⊗ B) = kAkw(B). 6 Theorem 2.2. Let A and B be n-by-n and m-by-m matrices, respectively. If kAk = 1 and w(A ⊗ B) = w(B), then either A has a unitary part or A is c.n.u. and W (B) is a circular disc centered at the origin. We first prove this for the case when A is an Sn-matrix. The numerical ranges of such matrices are known to have the Poncelet property, namely, if A is of class Sn, then, for any point λ on the unit circle ∂D, there is a unique (up to unitary similarity) (n + 1)-by-(n + 1) unitary dilation U of A such that λ is an eigenvalue of U and each edge of the (n + 1)-gon ∂W (U) intersects W (A) at exactly one point (cf. [2, Theorem 2.1 and Lemma 2.2]). Lemma 2.3. Let A be an Sn-matrix and B an m-by-m matrix. If w(A⊗ B) = w(B), then W (B) is a circular disc centered at the origin. Proof. Let U1, . . . , Um+1 be (n + 1)-by-(n + 1) unitary dilations of A with σ(Ui) ∩ σ(Uj) = ∅ for all i and j, 1 ≤ i 6= j ≤ m + 1. We may assume that Uj = diag (λ1j, . . . , λn+1,j) for each j, where λij = 1 for all i and j. Let Vj be an (n + 1)- by-n matrix such that A = V ∗ j Vj = In for each j. Since kAk = 1 and j UjVj and V ∗ w(A ⊗ λB) = w(A ⊗ B) = w(B) = w(λB) for any λ, λ = 1, we may further assume that w(B) is in W (A ⊗ B). Let x be a unit vector in Cn ⊗ Cm such that h(A ⊗ B)x, xi = w(B). We decompose (Vj ⊗ Im)x as y1j ⊕ · · · ⊕ yn+1,j with yij, 1 ≤ i ≤ n + 1, in Cm for each j. Then w(B) = h(A ⊗ B)x, xi n+1 = = h(Uj ⊗ B)(Vj ⊗ Im)x, (Vj ⊗ Im)xi = h(λ1jB ⊕ · · · ⊕ λn+1,jB)(y1j ⊕ · · · ⊕ yn+1,j), y1j ⊕ · · · ⊕ yn+1,ji Xi=1 hλijByij, yiji Xi=1 hByij, yiji. ≤ n+1 7 Letting ηij = hB(yij/kyijk), yij/kyijki for each yij 6= 0, we obtain w(B) = Xyij 6=0 λijkyijk2ηij ≤ Xyij 6=0 kyijk2ηij ≤ Xyij 6=0 kyijk2w(B) = w(B) since kyijk2 = k(Vj ⊗ Im)xk2 = kxk2 = 1. Xi Thus we have equalities throughout the above sequence, which yields that w(B) = λijηij for yij 6= 0. SincePi kyijk2 = 1, this must hold for at least one i, say, ij. Hence λijjw(B) = ηijj is in ∂W (B) for each j. Note that such λij jw(B)'s, 1 ≤ j ≤ m + 1, are distinct from each other by our assumption on the disjointness of the spectra of the Uj's. This shows that the boundary of W (B) and the circle z = w(B) intersect at at least m + 1 points. Since W (B) is contained in {z ∈ C : z ≤ w(B)}, we apply [3, Theorem] or [20]) to infer that W (B) = {z ∈ C : z ≤ Anderson's theorem (cf. w(B)}. Proof of Theorem 2.2. We assume that A is c.n.u. Then A can be dilated to the direct sum A′⊕· · ·⊕ A′ of rank (In− A∗A) many copies of some Sℓ-matrix A′ with ℓ ≤ n (cf. [18, Theorem 1.4] or [21, Lemma 3 (a)]). Hence A⊗ B dilates to (A′⊕· · ·⊕ A′)⊗ B = (A′ ⊗ B) ⊕ · · · ⊕ (A′ ⊗ B). We have w(B) = w(A⊗ B) ≤ w((A′ ⊗ B)⊕· · ·⊕ (A′ ⊗ B)) = w(A′ ⊗ B) ≤ kA′kw(B) = w(B). Thus w(A′ ⊗ B) = w(B). It follows from Lemma 2.3 that W (B) is a circular disc centered at the origin. An easy consequence of Theorem 2.2 is that the converse of Lemma 2.1 is also true. Corollary 2.4. For an n-by-n matrix A, the equality w(A ⊗ B) = kAkw(B) holds for all matrices B if and only if pA = ∞. 8 Proof. For the necessity, assume that kAk = 1 and let B be any matrix with its numerical range not a circular disc centered at the origin. Theorem 2.2 yields that A has a unitary part. Then pA = ∞ follows immediately. In Theorem 2.2, if B is the Jordan block Jm, then we have the following charac- terizations for w(A ⊗ B) = kAkw(B). Theorem 2.5. Let A be an n-by-n matrix with kAk = 1. Then the following condi- tions are equivalent: (a) W (A ⊗ Jm) = W (Jm), (b) w(A ⊗ Jm) = w(Jm), (c) A ⊗ Jm is unitarily similar to Jm ⊕ B for some matrix B with w(B) ≤ w(Jm), and (d) kAm−1k = 1. If, in addition, n = m, then the above conditions are also equivalent to (e) either A has a unitary part or A is of class Sn, and (f) pA = ∞ or n − 1. Note that W (Jm) = {z ∈ C : z ≤ cos(π/(m + 1))} (cf. [8, Proposition 1]). Proof of Theorem 2.5. The implication (a) ⇒ (b) is trivial. To prove (b) ⇒ (c), note that (A⊗Jm)m = Am⊗J m m = 0nm and kA⊗Jmk = kAkkJmk = 1. If x is a unit vector in Cn ⊗ Cm such that h(A ⊗ Jm)x, xi = w(A ⊗ Jm), then w(A ⊗ Jm) = w(Jm) = cos(π/(m + 1)) implies that the subspace K of Cn ⊗ Cm generated by the vectors x, (A⊗ Jm)x, . . . , (A⊗ Jm)m−1x is reducing for A⊗ Jm, and the restriction of A⊗ Jm [8, Theorem 1 (2)]). Hence A ⊗ Jm is unitarily to K is unitarily similar to Jm (cf. 9 similar to Jm ⊕ B, where B is the restriction of A ⊗ Jm to K ⊥. We obviously have w(B) ≤ w(A ⊗ Jm) = w(Jm). For (c) ⇒ (d), note that Am−1 ⊗ J m−1 m is unitarily similar to J m−1 m ⊕ Bm−1 under (c). Hence kAm−1k = kAm−1 ⊗ J m−1 m k = kJ m−1 m ⊕ Bm−1k = max{kJ m−1 m k,kBm−1k} = 1. To prove (d) ⇒ (c), let x be a unit vector in Cn such that kAm−1xk = 1. Then kAm−jxk = 1 for all j, 1 ≤ j ≤ m. Let {e1, . . . , em} be the standard basis for Cm, let xj = Am−jx ⊗ ej, 1 ≤ j ≤ m, and let K be the subspace of Cn ⊗ Cm generated by x1, . . . , xm. Then (A ⊗ Jm)x1 = 0 and (A ⊗ Jm)xj = xj−1 for 2 ≤ j ≤ m. Since {x1, . . . , xm} is an orthonormal basis of K, this shows that (A ⊗ Jm)K ⊆ K and the restriction of A ⊗ Jm to K is unitarily similar to Jm. On the other hand, it follows from kA ⊗ Jmk = kAkkJmk = 1 and (A ⊗ Jm)∗xm = (A∗ ⊗ J ∗ m)(x ⊗ em) = (A∗x) ⊗ (J ∗ mem) = (A∗x) ⊗ 0 = 0 that K is reducing for A⊗ Jm, and hence A⊗ Jm is unitarily similar to Jm⊕ B, where B is the restriction of A ⊗ Jm to K ⊥. Obviously, we have w(B) ≤ w(A ⊗ Jm) ≤ kAkw(Jm) = w(Jm). To prove (c) ⇒ (a), note that the unitary similarity of Jm and eiθJm for all real θ implies the same for A ⊗ Jm and eiθ(A ⊗ Jm). Thus W (A ⊗ Jm) is a circular disc centered at the origin. (c) implies that w(A ⊗ Jm) = w(Jm), which means that the radii of the two circular discs W (A ⊗ Jm) and W (Jm) are equal. Therefore, W (A ⊗ Jm) = W (Jm) holds. Now assume that n = m and that kAn−1k = 1. If kAnk = 1, then pA = ∞ and hence A has a unitary part by Proposition 1.2 (a) and (c). On the other hand, if kAnk < 1, then A is of class Sn by [4, Theorem 3.1]. This shows that (d) ⇒ (e). Next, if (e) is true, then pA = ∞ or n− 1 depending on whether A has a unitary part 10 or A is of class Sn (cf. [4, Theorem 3.1] for the latter). This proves (f). Finally, if pA = ∞, then kAkk = 1 for all k ≥ 1, and, in particular, kAn−1k = 1. On the other hand, if pA = n − 1, then kAn−1k = kAkn−1 = 1. This proves (f) ⇒ (d). The next proposition gives a characterization of w(A⊗ B) = kAkw(B) when B is of class Sm. Proposition 2.6. Let A be an n-by-n matrix with kAk = 1, and B be an Sm-matrix. Then w(A ⊗ B) = w(B) if and only if either A has a unitary part or A is c.n.u., kAm−1k = 1 and B is unitarily similar to Jm. Its proof depends on a special property of Sn-matrices. The following lemma is from [19, Lemma 5]. Here we give a shorter geometric proof. Lemma 2.7. Let A be an Sn-matrix. Then W (A) is a circular disc centered at the origin if and only if A is unitarily similar to Jn. Proof. If W (A) is as asserted, then the Poncelet property of W (A) says that it is circumscribed by (n + 1)-gons with vertices on the unit circle. As the circular disc {z ∈ C : z ≤ cos(π/(n + 1))}(= W (Jn)) is circumscribed by any regular (n + 1)-gon on the unit circle, if the radius of W (A) is not equal to cos(π/(n + 1)), then we infer from a geometrical consideration that W (A) cannot have the Poncelet property. Thus W (A) must equal W (Jn). The unitary similarity of A and Jn then follows from [2, Theorem 3.2]. The converse is trivial. Proof of Proposition 2.6. If w(A ⊗ B) = w(B), then, by Theorem 2.2, either A has a unitary part or A is c.n.u. and W (B) is a circular disc centered at the origin. In the latter case, Lemma 2.7 yields the unitary similarity of B and Jm, and then Theorem 2.5 gives kAm−1k = 1. The converse also follows from Theorem 2.5. 11 Note that, under the conditions of Proposition 2.6, if A is c.n.u., then we auto- matically have m ≤ n. This is because if, otherwise, m > n, then kAm−1k = 1 yields, by Proposition 1.2 (a) and (c), that A has a unitary part. A specific example of the results obtained so far is in the next proposition. Proposition 2.8. Let n and m be positive integers. Then W (Jn ⊗ Jm) = W (Jℓ), where ℓ = min{n, m}, and thus w(Jn ⊗ Jm) = min{w(Jn), w(Jm)}. Proof. Assume that m ≤ n. Since the principal submatrix of Jn ⊗ Jm formed by its rows and columns numbered 1, m + 2, 2m + 3, . . ., and (m − 1)m + m is Jm, we have that Jn ⊗ Jm is a dilation of Jm. Thus w(Jm) ≤ w(Jn ⊗ Jm). The reversed inequality w(Jn ⊗ Jm) ≤ kJnkw(Jm) = w(Jm) is by Proposition 1.1. Therefore, w(Jn ⊗ Jm) = w(Jm) holds. As was seen in the proof of (c) ⇒ (a) in Theorem 2.5, W (Jn ⊗ Jm) is a circular disc centered at the origin. Thus the equality of w(Jn ⊗ Jm) and w(Jm) implies that of W (Jn ⊗ Jm) and W (Jm). Besides Sn-matrices, another generalization of the Jordan blocks is the companion matrices. Recall that a companion matrix is one of the form   0 1 0 1 · −an −an−1 · · · · · · · 0 1 · −a2 −a1 ,   12 whose characteristic and minimal polynomials are both equal to zn +Pn The numerical ranges of such matrices have been studied in [5, 6, 1]. j=1 ajzn−j. Proposition 2.9. Let A be an n-by-n (n ≥ 2) companion matrix. Then the following conditions are equivalent: (a) w(A ⊗ A) = kAkw(A), (b) A is unitary, A = Jn, or A is unitarily similar to a direct sum [aωj n]⊕ B, where a > 1, ωn = ei(2π/n), 0 ≤ j ≤ n − 1, and B is an Sn−1-matrix with eigenvalues (1/a)ωk n, 0 ≤ k ≤ n − 1 and k 6= j, and (c) pA = nA = ∞ or n − 1. Proof. To prove (a) ⇒ (b), let A′ = A/kAk. Then (a) gives w(A′ ⊗ A′) = w(A′). By Theorem 2.2, either A′ has a unitary part or it is c.n.u. with numerical range a circular disc centered at the origin. In the former case, either A is normal or is unitarily similar to a matrix of the form [aωj n] ⊕ B, where a = kAk ≥ 1 and B is of size n − 1 with eigenvalues (1/a)ωk [5, Theorem 1.1 and Corollary 1.3]). If A is normal or a = 1, then A is unitary by [5, Corollary 1.2]. Hence we may assume that a > 1. Thus the eigenvalues of B are all contained in D. Moreover, by [1, Theorem 2.1], we have rank (In−1 − B∗B) = 1. These two together imply, by way of the singular value decomposition of B, that kBk = 1. Hence B is of class Sn−1. On the other hand, if it is the latter case, then W (A) is also a circular n, 0 ≤ k ≤ n − 1 and k 6= j (cf. disc centered at the origin. Therefore, A = Jn by [5, Theorem 2.9]. This proves (b). For (b) ⇒ (c), if A is unitary (resp., A = Jn), then, obviously, pA = nA = ∞ (resp., pA = nA = n − 1). On the other hand, if A is unitarily similar to the asserted [aωj n] ⊕ B, then kAk = max{a,kBk} = a = ρ(A). Thus pA = nA = ∞ by Proposition 1.2 (c) and 1.3. Finally, for (c) ⇒ (a), if pA = nA = ∞, then (a) is a consequence of Lemma 2.1. On the other hand, if pA = nA = n − 1, then An = 0n. This implies that A = Jn and thus (a) holds by Proposition 2.8. 13 The next theorem is a consequence of Theorem 2.5. It gives a lower bound, in terms of pA, for w(A) when A is an n-by-n matrix with kAk = 1. Theorem 2.10. If A is an n-by-n matrix with kAk = kAkk = 1 for some k ≥ 1, then w(A) ≥ cos(π/(k + 2)). Moreover, in this case, the following conditions are equivalent: (a) w(A) = cos(π/(k + 2)), (b) A is unitarily similar to Jk+1 ⊕ B, where B is a finite matrix with w(B) ≤ cos(π/(k + 2)), and (c) W (A) = {z ∈ C : z ≤ cos(π/(k + 2))}. For the proof of (a) ⇒ (b), we need the following lemma. Lemma 2.11. Let A =   0 a1 0 . . . . . . an−2 0 an−1 a   and B = 0 a1 0 . . . . . .     an−2 0 be n-by-n and (n− 1)-by-(n− 1) matrices, respectively, where n ≥ 2 and aj is nonzero for all j. Then w(A) > w(B). Proof. We prove this by induction on n. If n = 2, then A =  in which case we obviously have w(A) > 0 = w(B). Assume now that the assertion is true for the matrix A of size at most n − 1 (n ≥ 3), and let A and B be of the above form. By considering eiθA for a suitable real θ instead of A, we may assume and B = [0], 0 a1 0 a   14 that w(A) equals the largest eigenvalue of Re A. Let C =   0 a1 0 . . . . . . an−3 0 ,   and let p(z), q(z) and r(z) be the characteristic polynomials of Re A, Re B and Re C, respectively. We expand the determinant of   z −a1/2 −a1/2 z . . . . . . . . . . . . . . . z −an−1/2 −an−1/2 z − Re a   by minors on its last row to obtain p(z) = (z − Re a)q(z) − (an−12/4)r(z). Let α, β and γ be the largest eigenvalues of Re A, Re B and Re C, respectively. Then α = w(A), β = w(B) and γ = w(C). Since Re B (resp., Re C) is a principal submatrix of Re A (resp., Re B), we have β ≤ α (resp., γ ≤ β). Assume that α = β. Then the above equation yields 0 = p(α) = (α − Re a)q(β) − 1 4an−12γ(β) = − 1 4an−12γ(β). Since an−1 6= 0 and β is larger than or equal to all eigenvalues of Re C, we infer from γ(β) = 0 that β = γ or w(B) = w(C). This contradicts our induction hypothesis for B and C. Hence we must have α > β or w(A) > w(B). Proof of Theorem 2.10. By Theorem 2.5, the assumption kAk = kAkk = 1 implies that w(A ⊗ Jk+1) = w(Jk+1). Hence w(A) = kJk+1kw(A) ≥ w(A ⊗ Jk+1) = w(Jk+1) = cos π k + 2 as asserted. 15 We now prove the equivalence of (a), (b) and (c). The implications (b) ⇒ (c) and (c) ⇒ (a) are trivial. To prove (a) ⇒ (b), let x be a unit vector in Cn such that kAkxk = 1. Then kAjxk = 1 for all j, 0 ≤ j ≤ k. We now check that Ak+1x = 0. Assuming otherwise that kAk+1xk > 0, let ut = [ut1 . . . ut,k+2]T in Ck+2 ⊗ Cn, where Ak+1x if j = 1, sin (j − 1)π k + 2 Ak−j+2x if j = 2, . . . , k + 2 √1 − t2 kAk+1xk tq 2 utj =  v ≡r 2 k + 2 for any t, 0 < t < 1. Note that k + 2(cid:20)sin π k + 2 sin 2π k + 2 . . . sin (k + 1)π k + 2 (cid:21)T is a unit vector in Ck+1 with hJk+1v, vi = cos(π/(k + 2)) (cf. [8, Proposition 1 (3)]). Hence kutk = ((1 − t2) + t2kvk2)1/2 = 1, and h(Jk+2 ⊗ A)ut, uti = t√1 − t2r 2 Xj=1 = t√1 − t2r 2 = t√1 − t2r 2 π k + 2kAk+1xk jπ (j + 1)π k + 2 kAk−j+1xk2 π k + 2kAk+1xk + t2hJk+1v, vi π k + 2kAk+1xk + t2 cos k + 2 k k + 2 k + 2 k + 2 k + 2 k + 2 +t2 sin sin sin π . 2 sin sin To reach a contradiction, we need to find some t0, 0 < t0 < 1, such that h(Jk+2 ⊗ A)ut0, ut0i > cos(π/(k + 2)). This is the same as t0q1 − t2 0r 2 k + 2 sin π k + 2kAk+1xk > (1 − t2 0) cos π k + 2 or t0 >rk + 2 2 cot π k+2 kAk+1xk . p1 − t2 0 Since limt→1− t/√1 − t2 = ∞, the existence of such a t0 is guaranteed. On the other hand, we also have h(Jk+2 ⊗ A)ut0, ut0i ≤ w(Jk+2 ⊗ A) ≤ kJk+2kw(A) = w(A) = cos π k + 2 , 16 hence a contradiction. Thus we must have Ak+1x = 0. Let K be the subspace of Cn generated by x, Ax, . . . , Akx. Then AK ⊆ K. If A′ is the restriction of A to K, then A′k+1 = 0 and kA′jxk = kAjxk = 1 for all j, 0 ≤ j ≤ k. Hence kA′jk = 1 for all such j's. Together with A′k+1 = 0, this says that pA′ = k and thus dim K = k + 1 by Proposition 1.2 (a). Therefore, A′ is unitarily similar to a matrix of the form i,j=1 with aij = 0 for all i ≥ j. Since 1 = kA′kk = a12 · · · ak,k+1, we infer that [aij]k+1 a12 = · · · = ak,k+1 = 1, and thus all the other aij's are zero. Therefore, [aij]k+1 i,j=1, and hence A′, is unitarily similar to Jk+1. Then A is unitarily similar to a matrix of the form   Jk+1 0 0 b1 · · · bn−k−1 ∗ c1 . . . ∗ cn−k−1 .   To show that all the bj's are zero, we appeal to Lemma 2.11. 1 ≤ j ≤ n − k − 1, consider the (k + 2)-by-(k + 2) matrix Indeed, for each j, Aj = 0 ... 0 bj cj Jk+1 0   .   If bj 6= 0, then w(Aj) > w(Jk+1) = cos(π/(k + 2)) by Lemma 2.11, which contradicts w(Aj) ≤ w(A) = cos(π/(k + 2)). This proves (a) ⇒ (b). Theorem 2.10 generalizes the classical result of Williams and Crimmins [17] for k = 1. The following corollary is for k = n − 1. Part of it has been proven in [19]: the equivalence of (b) and (c) is in [19, Theorem 1] and that of (b) and (d) in [19, p. 352]. 17 Corollary 2.12. The following conditions are equivalent for an n-by-n matrix A with kAk = 1: (a) kAn−1k = 1 and w(A) = cos(π/(n + 1)), (b) A is unitarily similar to Jn, (c) W (A) = {z ∈ C : z ≤ cos(π/(n + 1))}, (d) kAn−1k = 1 and An = 0n, and (e) pA = nA = n − 1. Proof. The equivalence of (a) and (b) is by Theorem 2.10. The other implications are either in [19] or trivial. Note that, in the preceding corollary, the conditions that kAk = 1 and w(A) = cos(π/(n + 1)) for an n-by-n matrix A are not sufficient to guarantee that A be unitarily similar to Jn. One example is A = Jn−1 ⊕ [cos(π/(n + 1))]. We end this section with a characterization of matrices A satisfying pA = nA. This is related to the previous results. Theorem 2.13. Let A be an n-by-n matrix with kAk = 1. Then (a) A satisfies pA = nA (≤ ∞) if and only if either it has a unitary part or is unitarily similar to a direct sum Jk+1 ⊕ B, where k = pA < ∞ and Bk+1 = 0n−k−1, and (b) if pA = nA (≤ ∞), then w(A ⊗ A) = w(A) holds, but not conversely. Proof. (a) For the necessity, we may assume, in view of Proposition 1.2 (c), that k ≡ pA = nA < ∞ and prove that A is unitarily similar to the asserted direct sum. 18 Since Ak+1 = 0n, A is unitarily similar to a block matrix A′ of the form [Aij]k+1 Aij = 0 for 1 ≤ j ≤ i ≤ k + 1. Hence 0 · · · 0 0 Qk i=1 Ai,i+1  i,j=1 with 0 ... 0 .   A′k = . . .  Since kA′kk = kAkk = kAkk = 1, we have kQk i=1 Ai,i+1k = 1. Let x be a unit vector such that k(Qk i=1 Ai,i+1)xk = 1. Then k(Qk i=j Ai,i+1)xk = 1 for all j, 1 ≤ j ≤ k. Let {e1, . . . , ek+1} be the standard basis for Ck+1, and let xj = ej ⊗ (Qk i=j Ai,i+1)x if 1 ≤ j ≤ k, and xk+1 = ek+1 ⊗ x. Then x1, . . . , xk+1 are orthonormal vectors in Cn, and A′x1 = 0 and A′xj = xj−1 for 2 ≤ j ≤ k + 1. Thus if K is the subspace generated by x1, . . . , xk+1, then dim K = k + 1, A′K ⊆ K, and the restriction of A′ to K is unitarily similar to Jk+1. We infer from kA′k = 1 and A′∗xk+1 = 0 that K reduces A′, and thus A′ is unitarily similar to Jk+1 ⊕ B with Bk+1 = 0. For the converse, if A has a unitary part, then pA = nA = ∞ by Proposition 1.2 (c). On the other hand, if A is unitarily similar to Jk+1 ⊕ B with the asserted properties, then Ak+1 = 0 implies that pA ≤ nA ≤ k. But kAkk = kJ k k+1 ⊕ Bkk = max{kJ k k+1k,kBkk} = 1 = kAkk and kAk+1k = 0 < 1 = kAkk+1 together yield pA = nA = k. (b) If A has a unitary part, then w(A⊗A) = w(A) by Proposition 2.1. On the other hand, if A is unitarily similar to Jk+1 ⊕ B as in (a), then A⊗ A is unitarily similar to (Jk+1⊗Jk+1)⊕(Jk+1⊗B)⊕(B⊗Jk+1)⊕(B⊗B). Note that w(Jk+1⊗Jk+1) = w(Jk+1) by Proposition 2.8, and (1) w(Jk+1 ⊗ B) = w(B ⊗ Jk+1) ≤ kJk+1kw(B) = w(B) by Proposition 1.1. Since Bk+1 = 0 and kBk ≤ 1, [21, Lemma 3 (a)] implies that B can be dilated to the direct sum of rank (I − B∗B) copies of Jm for some m ≤ k + 1. 19 Thus w(B) ≤ w(Jm) ≤ w(Jk+1). Combined with (1), this yields w(Jk+1 ⊗ B) ≤ w(Jk+1). Also, w(B ⊗ B) ≤ kBkw(B) ≤ w(B) ≤ w(Jk+1). Therefore, w(A ⊗ B) = max{w(Jk+1 ⊗ Jk+1), w(Jk+1 ⊗ B), w(B ⊗ B)} = w(Jk+1) = max{w(Jk+1), w(B)} = w(A). That w(A ⊗ A) = w(A) does not imply pA = nA is seen by A = J2 ⊕ [a], where 0 < a ≤ 1/2, in which case, kAk = 1 and w(A ⊗ A) = w(A) = 1/2, but pA = 1 and nA = ∞. The final result of this section is conditions for a matrix A with pA = nA so that it be unitarily similar to a block-shift matrix (2) A′ = with kA1 · · · Akk = kAk. 0 A1 0   . . . . . . Ak 0   Proposition 2.14. Let A be an n-by-n matrix with pA = nA ≡ k < ∞. If either (a) k = 1, n − 2 or n − 1, or (b) n = 2, 3, 4 or 5, then A is unitarily similar to the block-shift matrix A′ in (2) with kA1 · · · Akk = kAk. Proof. We may assume that kAk = 1. (a) If k = nA = 1, then A2 = 0n. Hence A is unitarily similar to a block-shift matrix of the form   0 A1 0 0   with kA1k = kAk. 20 If k = pA = nA = n − 1 (resp., n − 2), then Theorem 2.13 (a) implies that A is unitarily similar to Jn (resp., Jn−1 ⊕ [0]). The latter matrix plays the role of A′ with k = n − 1 (resp., n − 2) and A1 = · · · = An−1 = [1] (resp., A1 = · · · = An−3 = [1] and An−2 = [1 0]). (b) In light of (a), we need only prove for n = 5 and k = 2. Invoking Theorem 0 2.13 to obtain the unitary similarity of A and J3⊕  matrix is permutationally similar to a block-shift matrix A′ with k = 2, A1 =  and A2 =  0  . We obviously have kA1A2k = k   0  k = 1 = kAk. 0 1 b 0  , where b ≤ 1. The latter  b   1 0 0 We remark that the preceding proposition fails for n = 6 and k = 2. Here is an example. Let A = J3 ⊕ B, where 1   B = b  0 1 1 0 0 1 0 0 0 with b = q2/(3 + √5). Then kA2k = 1 = kAk2 and A3 = 06. This shows that pA = nA = 2. Since w(B) = 2b > √2/2 = w(J3) and w(B) is not a circular disc centered at the origin (cf. [13, Theorem 4.1 (2)]), we infer that nor is W (A) (= the convex hull of W (J3) ∪ W (B)). This implies that A cannot be unitarily similar to a block-shift matrix. 3 Nonnegative Matrices Recall that a matrix A = [aij]n i,j=1 is nonnegative (resp., positive), denoted by A < 0 (resp., A ≻ 0), if aij ≥ 0 (resp., aij > 0) for all i and j. Two n-by-n matrices A and B are permutationally similar if there is an n-by-n permutation matrix P (one with 21 each row and column has exactly one 1 and all other entries 0) such that P T AP = B. A is said to be (permutationally) reducible if either A is the 1-by-1 zero matrix or n ≥ 2 and it is permutationally similar to a matrix of the form   and D are square matrices; otherwise, it is (permutationally) irreducible. It is known , where B B C 0 D   that if A is nonnegative with Re A irreducible, then it is permutationally similar to a block-shift matrix if and only if its numerical range is a circular disc centered at the [16, Theorem 1 (a)⇔(r)]). Other properties of nonnegative matrices can origin (cf. be found in [11, Section 6.2 and Chapter 8]. The main result of this section is the following theorem, which essentially gener- alizes Theorem 2.5. Theorem 3.1. Let A be an n-by-n matrix and B an m-by-m nonnegative matrix with Re B irreducible. Then the following conditions are equivalent: (a) w(A ⊗ B) = kAkw(B), (b) either pA = ∞ or nB ≤ pA < ∞ and W (B) is a circular disc centered at the origin, and (c) either pA = ∞ or nB ≤ pA < ∞ and B is permutationally similar to a block- shift matrix 0 B1 0   . . . . . . Bk 0   with k = nB. For its proof, we need the following two lemmas. Lemma 3.2. Let A = [aij]n i,j=1 be a nonnegative matrix. Then the following hold: 22 (a) The index nA is finite if and only if there is no sequence of indices i0, i1, . . . , ik−1, ik (k ≥ 1) with i0 = ik such that ai0i1, . . . , aik−1ik are all nonzero. In particular, we have nA = sup{k ≥ 1 : there are distinct ij, 0 ≤ j ≤ k, such that aij ij+1 6= 0 for all j}. (b) nA = ∞ if and only if there is a k ≥ 1 such that some diagonal entry of Ak is nonzero. (c) If aii 6= 0 for some i, 1 ≤ i ≤ n, then nA = ∞. (d) If A is irreducible, then nA = ∞. (e) If A is the block-shift matrix   0n1 A1 0n2 . . . . . . Ak 0nk+1   on Cn = Cn1 ⊕ · · · ⊕ Cnk+1 and Re A is irreducible, then k = nA. Proof. (a) Assume first that the indices i0, i1, . . . , ik−1, ik = i0 (k ≥ 1) are such that ai0i1, . . . , aik−1ik 6= 0. [11, Theorem 6.2.16] says that this is the case if and only if (Ak)i0i0, the (i0, i0)-entry of Ak, is nonzero. Hence Ak 6= 0n. Similarly, considering the sequence i0, . . . , ik, i1, . . . , ik, . . . , i1, . . . , ik of ℓk + 1 indices for any ℓ ≥ 1, we also obtain Aℓk 6= 0n. It follows that nA = ∞. Conversely, assume that nA = ∞. Then Ak 6= 0n for some k ≥ n. [11, Theorem 6.2.16] yields that, for some i and j, there are indices i0 = i, i1, . . . , ik−1, ik = j such that ai0i1, . . . , aik−1ik are all nonzero. By the pigeonhole principle, we infer that is = it for some s and t, 0 ≤ s < t ≤ k. Then is, . . . , it are such that is = it and aisis+1, . . . , ait−1it 6= 0. This proves the converse. The expression for nA is an easy consequence of [11, Theorem 6.2.16] and the above arguments. So are (b) and (c). 23 (d) Note that the irreducibility of A is equivalent to the existence, for every distinct pair i and j, of indices i0 = i, i1, . . . , ik−1, ik = j (k ≥ 1) such that ai0i1, . . . , aik−1ik are all nonzero. Combining such indices from i to j with those from j to i yields one from i to i with the corresponding entries nonzero. Thus nA = ∞ by [11, Theorem 6.2.16] and (b). (e) Since Ak+1 = 0n, we have nA ≤ k. If nA < k, then Ak = 0n, which implies that A1 · · · Ak = 0. If there are any nonzero ai0i1, ai1i2, . . . , aik−1ik , where (Pℓ j=1 nj) + 1 ≤ iℓ ≤ Pℓ+1 j=1 nj for 0 ≤ ℓ ≤ k, then the (i0, nk+1 − (n − ik))-entry of A1 · · · Ak, being larger than or equal to Qk−1 j=0 aijij+1, is nonzero, which contradicts the zeroness of the product A1 · · · Ak. Thus no such nonzero sequence exists. This results in the reducibility of Re A, a contradiction. Hence we must have nA = k. We remark that the conditions in the preceding lemma can all be expressed equiv- alently in terms of the directed graph associated with the matrix A (cf. [11, Section 6.2]). Lemma 3.3. Let A and B be n-by-n and m-by-m matrices, respectively. If B is unitarily similar to a block-shift matrix (3) on Cm = Cm1 ⊕ · · · ⊕ Cmk+1   0m1 B1 0m2 . . . . . . Bk 0mk+1   with k ≤ pA ≤ ∞, then w(A ⊗ B) = kAkw(B). Proof. We may assume that kAk = 1 and B is equal to the block-shift matrix (3). Since k ≤ pA ≤ ∞, we have kAkk = kAkk = 1. Let x be a unit vector in Cn such that kAkxk = 1, and let y = [y1 . . . yk+1]T , where yj is in Cmj , 1 ≤ j ≤ k + 1, be a unit vector in Cm such that hBy, yi = w(B). Let u = [y1⊗Akx y2⊗Ak−1x . . . yk+1⊗x]T . 24 Then u is a vector in Cm ⊗ Cn with kuk = ( = ( k+1 k+1 Xj=1 Xj=1 kyj ⊗ Ak−j+1xk2)1/2 = ( k+1 Xj=1 kyjk2kAk−j+1xk2)1/2 kyjk2)1/2 = kyk = 1. Moreover, we have h(B ⊗ A)u, ui  * 0m1n B1 ⊗ A 0m2n = . . . . . . Bk ⊗ A 0mk+1n     y1 ⊗ Akx y2 ⊗ Ak−1x ... yk+1 ⊗ x ,     y1 ⊗ Akx y2 ⊗ Ak−1x ... yk+1 ⊗ x  +  (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)  k k Xj=1 Xj=1 Xj=1 k = = h(Bjyj+1) ⊗ (Ak−j+1x), yj ⊗ (Ak−j+1x)i hBjyj+1, yjikAk−j+1xk2 hBjyj+1, yji = = hBy, yi = w(B). This shows that w(B) ≤ w(B ⊗ A) = w(A ⊗ B). But w(A ⊗ B) ≤ kAkw(B) = w(B) always holds by Proposition 1.1. Hence w(A ⊗ B) = w(B) as asserted. We are now ready to prove Theorem 3.1. Proof of Theorem 3.1. For (a) ⇒ (b), We assume that kAk = 1 and A is c.n.u. In view of Theorem 2.2 and Proposition 1.2 (c), we need only check that w(A ⊗ B) = w(B) implies nB ≤ pA (< ∞). Let B = [bij]m i,j=1, and let x be a unit vector in Cm ⊗ Cn such that w(B ⊗ A) = h(B ⊗ A)x, xi. If x = [x1 . . . xm]T , where xj is in Cn for 25 1 ≤ j ≤ m, then w(B) = w(B ⊗ A) = h[bijA]x, xi (4) (5) (6) bijhAxj, xii bijkAxjkkxik bijkxjkkxik ≤Xi,j ≤Xi,j ≤ kAkXi,j ≤ hBx′, x′i ≤ w(B), where x′ = [kx1k . . . kxmk]T is a unit vector in Cm. This shows that the above inequalities are equalities throughout. Since B < 0 and Re B is irreducible, there is a unique unit vector y in Cm with y ≻ 0 such that hBy, yi = w(B) (cf. [14, Proposition 3.3]). The equality in (6) yields that x′ = y and thus xj 6= 0 for all j. Also, the equalities in (4) and (5) imply that hAxj, xii = kAxjkkxik = kxjkkxik for all those bij's with bij > 0. Thus Axj = λijxi for some λij satisfying λij = kxjk/kxik. Assume first that k ≡ nB < ∞. Thus Bk 6= 0m. By Lemma 3.2 (a), there are distinct indices i0, . . . , ik such that bi0i1, . . . , bik−1ik > 0. It thus follows from above that Axij = λij−1ij xij−1 for 1 ≤ j ≤ k. Hence Akxik = (Qk )kxi0k = kxikk, kAkxikk = ( k Yj=1 kxijk kxij−1k j=1 λij−1ij )xi0. Since we obtain kAkk = 1 or pA ≥ k = nB. On the other hand, if nB = ∞, then the same arguments as above with k arbitrarily large yield that pA = ∞, which contradicts our assumption that A is c.n.u. This proves (a) ⇒ (b). That (b) ⇔ (c) is a consequence of [16, Theorem 1 (a)⇔(r)], and (c) ⇒ (a) is by Lemma 3.2 (e) and Lemma 3.3. Note that, in Theorem 3.1, the implication (a) ⇒ (b) or (a) ⇒ (c) is no longer true if B is nonnegative but without the irreducibility of Re B. One example is 26 A = B = J2 ⊕ [a], where 0 < a ≤ 1/2 (cf. the end of the proof of Theorem 2.13 (b)). The next example shows that the same can be said if B is not nonnegative but Re B is irreducible. Example 3.4. Let A = J3 and B =  0 −√2 1 0 0 0 1 0 √2/2 .   Then Re B is easily seen to be irreducible. We now show that W (B) = D. This is seen via [13, Corollary 2.5] by letting u = 0 and λ = √2/2 therein and checking that tr (B∗B2) = tr   0 0 0 0 0 1 0 0 √2/4   √2 4 = = λλ2 and tr (B∗B) = 9/2 ≥ 5λ2, where tr (·) denotes the trace of a matrix. We next prove that 1 is an eigenvalue of Re (A ⊗ B). Indeed, since we need to check that Re (A ⊗ B) = 03 B 03 B∗ 03 B 03 B∗ 03 1 2   ,   det  2I3 −B 03 −B∗ 2I3 −B 03 −B∗ 2I3 = 0.   By a repeated use of the Schur decomposition, the above determinant is seen to be 27 equal to 2I3 −B −B∗ 2I3 −B∗ 03 1 2   ( I3) [−B 03]  det (2I3) det    −B∗   −   2I3 − (1/2)B∗B −B  2I3 −1 −√2/2 √2/2 1 = 8 det  = 8 det (4I3 − B∗B − BB∗) = 8 det  1 −1 −√2/2 √2/2 1   = 0 as required. Since W (A ⊗ B) is a circular disc centered at the origin (by the unitary similarity of A ⊗ B and eiθ(A ⊗ B) for all real θ) and w(A ⊗ B) ≤ kAkw(B) = 1, we infer from 1 ∈ σ(Re (A⊗ B)) that W (A⊗ B) = D. Hence w(A⊗ B) = 1 = kAkw(B). But, obviously, we have nB = ∞ and pA = 2. (cid:3) The next corollary gives a more concrete equivalent condition, in terms of block- shift matrices, for w(A ⊗ B) = kAkw(B) when A = B < 0 and Re B is irreducible. Corollary 3.5. Let A be an n-by-n nonnegative matrix with Re A irreducible. Then the following conditions are equivalent: (a) w(A ⊗ A) = kAkw(A), (b) pA = nA (≤ ∞), and (c) either A is unitarily similar to [a]⊕ A′ with a ≥ kA′k, or A is permutationally similar to a block-shift matrix A′′ =   0 A1 0 28 . . . . . . Ak 0   with kA1 · · · Akk = kAk. Proof. We may assume that kAk = 1. The implication (a) ⇒ (b) is by Theorem 3.1 and Proposition 1.3 (d). For (b) ⇒ (c), if pA = nA = ∞, then A has a unitary part by Proposition 1.2 (c), and hence A is unitarily similar to [a]⊕ A′ with a = 1 ≥ kA′k as asserted. On the other hand, if pA = nA < ∞, then w(A⊗ A) = w(A) by Theorem 2.13 (b). Hence Theorem 2.2 implies that W (A) is a circular disc centered at the origin. For a nonnegative A with Re A irreducible, this is equivalent to A being permutationally similar to the block-shift matrix A′′ (cf. [16, Theorem 1 (a)⇔(r)]). As nA′′ = k by Lemma 3.2 (e), we also have pA = k. Thus kAkk = kAkk = 1, which yields that kA1 · · · Akk = 1 = kAk as required. Finally, for (c) ⇒ (a), if A is unitarily similar to [a] ⊕ A′ with a ≥ kA′k, then w(A ⊗ A) = w(A) by Lemma 2.1. On the other hand, if A is permutationally similar to the block-shift matrix A′′ with kA1 · · · Akk = 1, then kAkk = kA′′kk = kA1 · · · Akk = 1 = kAkk. Thus pA ≥ k = nA. The equality w(A ⊗ A) = w(A) then follows from Theorem 3.1. Corollary 3.6. Let A = [aij]n i,j=1, where aij ≥ 0 for all i and j, aij = 0 for i ≥ j, and ai,i+1 > 0 for all i. Then the following conditions are equivalent: (a) w(A ⊗ A) = kAkw(A), (b) pA = nA = n − 1, and (c) a12 = · · · = an−1,n and aij = 0 for all other pairs of i and j. Proof. In this case, A is nonnegative, Re A is irreducible and nA = n − 1. Con- sequently, Corollary 3.5 yields the equivalence of (a), (b) and the condition (c') that A is permutationally similar to a block-shift matrix A′′ as in Corollary 3.5 29 (c). Since k = nA′′ = nA by Lemma 3.2 (e), A′′ is necessarily equal to A with a12 · · · an−1,n = kAk and aij = 0 for all other pairs of i and j. The norm condition above yields that a12 = · · · = an−1,n = kAk. Thus (c') is the same as (c), and we have the equivalence of (a), (b) and (c). Acknowledgements This research was partially supported by the National Science Council of the Re- public of China under projects NSC-101-2115-M-008-006, NSC-101-2115-M-009-001 and NSC-101-2115-M-009-004 of the respective authors. P. Y. Wu was also supported by the MOE-ATU. This paper was presented by him at the 4th International Confer- ence on Matrix Analysis and Applications in Konya, Turkey. He thanks the organizers for their works with the conference. 30 References [1] H.-L. Gau, Numerical ranges of reducible companion matrices, Linear Algebra Appl. 432 (2010), pp. 1310 -- 1321. [2] H.-L. Gau and P. Y. Wu, Numerical range of S(φ), Linear Multilinear Algebra 45 (1998), pp. 49 -- 73. [3] H.-L. Gau and P. Y. Wu, Condition for the numerical range to contain an elliptic disc, Linear Algebra Appl. 364 (2003), pp. 213 -- 222. [4] H.-L. Gau and P. Y. Wu, Finite Blaschke products of contractions, Linear Algebra Appl. 368 (2003), pp. 359 -- 370. [5] H.-L. Gau and P. Y. Wu, Companion matrices: reducibility, numerical ranges and similarity to contractions, Linear Algebra Appl. 383 (2004), pp. 127 -- 142. [6] H.-L. Gau and P. Y. Wu, Numerical ranges of companion matrices, Linear Al- gebra Appl. 421 (2007), pp. 202 -- 218. [7] K. Gustafson and D. K. M. Rao, Numerical Range. The Field of Values of Linear Operators and Matrices, Springer, New York, 1997. [8] U. Haagerup and P. de la Harpe, The numerical radius of a nilpotent operator on a Hilbert space, Proc. Amer. Math. Soc. 115 (1992), pp. 371 -- 379. [9] P. R. Halmos, A Hilbert Space Problem Book, 2nd ed., Springer, New York, 1982. [10] J. A. R. Holbrook, Multiplicative properties of the numerical radius in operator theory, J. Reine Angew. Math. 237 (1969), pp. 166 -- 174. [11] R. A. Horn and C. R. Johnson, Matrix Analysis, Cambridge University Press, Cambridge, 1985. [12] R. A. Horn and C. R. Johnson, Topics in Matrix Analysis, Cambridge University Press, Cambridge, 1991. 31 [13] D. S. Keeler, L. Rodman and I. M. Spitkovsky, The numerical range of 3 × 3 matrices, Linear Algebra Appl. 252 (1997), pp. 115 -- 139. [14] C.-K. Li, B.-S. Tam and P. Y. Wu, The numerical range of a nonnegative matrix, Linear Algebra Appl. 350 (2002), pp. 1 -- 23. [15] V. Pt´ak, Lyapunov equations and Gram matrices, Linear Algebra Appl. 49 (1983), pp. 33 -- 55. [16] B.-S. Tam and S. Yang, On matrices whose numerical ranges have circular or weak circular symmetry, Linear Algebra Appl. 302/303 (1999), pp. 193 -- 221. [17] J. P. Williams and T. Crimmins, On the numerical radius of a linear operator, Amer. Math. Monthly 74 (1967), pp. 832 -- 833. [18] P. Y. Wu, Unitary dilations and numerical ranges, J. Operator Theory 38 (1997), pp. 25 -- 42. [19] P. Y. Wu, A numerical range characterization of Jordan blocks, Linear Multilin- ear Algebra 43 (1998), pp. 351 -- 361. [20] P. Y. Wu, Numerical ranges as circular discs, Applied Math. Lett. 24 (2011), pp. 2115 -- 2117. [21] P. Y. Wu, H.-L. Gau and M.-C. Tsai, Numerical radius inequality for C0 con- tractions, Linear Algebra Appl. 430 (2009), pp. 1509 -- 1516. 32
1804.01022	1	1804	2018-04-03T15:05:49	Green's function of the problem of bounded solutions in the case of a block triangular coefficient	[ "math.FA", "math.CA", "math.DS", "math.OA", "math.SP" ]	It is known that the equation $x'(t)=Ax(t)+f(t)$, where $A$ is a bounded linear operator, has a unique bounded solution $x$ for any bounded continuous free term~$f$ if and only if the spectrum of the coefficient $A$ does not intersect the imaginary axis. The solution can be represented in the form \begin{equation} x(t)=\int_{-\infty}^{\infty}\mathcal G(s)f(t-s)\,ds. \end{equation} The kernel $\mathcal G$ is called Green's function. In this paper, the case when $A$ admits a representation by a block triangular operator matrix is considered. It is shown that the blocks of $\mathcal G$ are sums of special convolutions of Green's functions of diagonal blocks of $A$.	math.FA	math	GREEN'S FUNCTION OF THE PROBLEM OF BOUNDED SOLUTIONS IN THE CASE OF A BLOCK TRIANGULAR COEFFICIENT V.G. KURBATOV AND I.V. KURBATOVA Abstract. It is known that the equation x′(t) = Ax(t) + f (t), where A is a bounded linear operator, has a unique bounded solution x for any bounded continuous free term f if and only if the spectrum of the coefficient A does not intersect the imaginary axis. The solution can be represented in the form G(s)f (t − s) ds. x(t) = Z ∞ −∞ The kernel G is called Green's function. In this paper, the case when A admits a representation by a block triangular operator matrix is considered. It is shown that the blocks of G are sums of special convolutions of Green's functions of diagonal blocks of A. Introduction We consider the equation (1) where A is a linear bounded operator acting in a Banach space X. We assume that f is continuous. x′(t) − Ax(t) = f (t), The bounded solutions problem is the problem of finding a bounded solution x that corresponds to a bounded free term f . The bounded solutions problem is closely connected with the problem of the exponential dichotomy of solutions. For the discussion of the bounded solutions problem from different points of view and related questions, see [1, 2, 4, 9, 10, 20, 26, 27, 38, 42, 44, 46] and the references therein. It is well known (see Theorem 5) that equation (1) has a unique bounded solution x for any bounded continuous free term f if and only if the spectrum of the coefficient A is disjoint from the imaginary axis. In this case, the solution can be represented in the form The kernel G is called Green's function. x(t) =Z ∞ −∞ G(s)f (t − s) ds. In this paper, we consider the case when A admits a representation in the form of a block triangular matrix (4). The simplest 2×2 matrix representation of the coefficient A is naturally induced by the decomposition of the space X into the direct sum X−⊕X+ of two spectral subspaces related to the parts of σ(A) that lie in the left and right complex half- planes. This matrix representation is diagonal, but it can be 'bad' in the sense that the corresponding projectors have large norms; in such a case it may be convenient to replace one of the subspaces by the orthogonal (or close to orthogonal) complement of the other; as a result one will arrive at a triangular matrix representation of A. Similarly, the spectrum of A may be divided into clusters; so, it is again natural to use a diagonal or triangular matrix representation; the phenomenon of clusterization is discussed, e.g., in [21, lecture 12], [11, 37]. Representation by triangular operator matrices is also natural for causal 1991 Mathematics Subject Classification. 47A60; 47A80; 34B27; 34B40; 34D09. Key words and phrases. bounded solutions problem; Green's function; divided difference with operator argu- ments; block matrix; causal operator. 1 2 V.G. KURBATOV AND I.V. KURBATOVA operators; in their turn, causal operators are widely used in control theory [13, 16, 53] and functional differential equations [34, 35, 36]. For other aspects of the theory of triangular operator matrices, see [7, 8, 19, 22, 23, 28, 31, 33, 45] and the references therein. The main results of this paper are Theorems 19 and 23; see also Theorem 5. From these theorems, it follows that Green's function is also induced by a triangular matrix and its blocks can be represented as the sums of special convolutions of Green's functions of the diagonal blocks of A; see Example 2. Similar representations and related formulas for the fundamental solution of equa- tion (1) were proposed, discussed, and applied by many authors [6, 12, 14, 17, 24, 32, 41, 43, 47, 50, 51, 52]; such formulas are widely used in numerical methods and other applications. We repeat some of these results in this paper (i) for the convenience of their comparison with our results connected with Green's function, and because (ii) we propose a new proof for them, (iii) and discuss the infinite-dimensional case, which requires some additional considerations in the proof; see Section 3. The paper is organized as follows. In Section 1, we recall the definition of an analytic function with an operator argument. In Section 2, we describe the representation of the fundamental solution of initial value problem and Green's function of bounded solutions problem in the form of the analytic functions exp±,t and gt, respectively, of the coeffi- cient A. In Section 3, we discuss the subalgebra of operators induced by block triangular matrices. This subalgebra is not full, which leads to some technical difficulties in the sub- sequent presentation. In Section 4, we describe (Theorem 10) a representation of blocks of an analytic function f of a triangular matrix via contour integrals. In Section 5, the main terms of the formula from Theorem 10 are represented as divided differences of f with operator arguments (Theorem 17). In Section 6, we show that divided differences of exp±,t and gt can be represented as convolutions with respect to the variable t of functions of one variable (Theorem 22). In Section 7, we describe a representation of divided differences of exp±,t and gt with operator arguments (Theorem 23). The combination of Theorems 19 and 23 allows one to represent the blocks of the fundamental solution of initial value problem and Green's function of the bounded solutions problem as special convolutions of the functions exp±,t and gt applied to the diagonal blocks of A (Examples 1 and 2). 1. Functions of operators Let X and Y be non-zero complex Banach spaces. We denote by B(X, Y ) the set of all bounded linear operators A : X → Y . If X = Y , we use the brief notation B(X). The symbol 1 = 1X stands for the identity operator from B(X). Let B be a non-zero complex Banach algebra [5, 29, 48] with the unit 1 (unital algebra). The main example of a unital Banach algebra is the algebra B(X); another important example is the algebra of all n × n matrices, n ∈ N. A subset R of an algebra B is called a subalgebra if A + B, λA, AB ∈ R for all A, B ∈ R and λ ∈ C. If the unit 1 of an algebra B belongs to its subalgebra R, then R is called a subalgebra with a unit or a unital subalgebra. A unital subalgebra R of a unital algebra B is called [5, ch. 1, § 3.6] full if it possesses the property: if for B ∈ R there exists B−1 ∈ B such that BB−1 = B−1B = 1, then B−1 ∈ R. Below (Remark 1) we will see that the subalgebra of all block triangular matrices is not always full. Let B be a (nonzero) unital algebra and A ∈ B. The set of all λ ∈ C such that the element λ1 − A is not invertible is called the spectrum of the element A (in the algebra B) and is denoted by the symbol σ(A) or σB(A). The complement ρ(A) = ρB(A) = C \ σ(A) GREEN'S FUNCTION IN THE CASE OF A TRIANGULAR COEFFICIENT 3 is called the resolvent set of A. The function Rλ = (λ1 − A)−1 is called the resolvent of the element A. The spectral radius r(A) = rB(A) is the radius of the smallest closed circle in C with center at 0 that contains σ(A). Proposition 1 ([5, Chap. 1, Sec. 4, Theorem 3], [48, Theorem 10.18]). Let R be a closed unital subalgebra of a unital algebra B. Then the spectrum σR(A) of an element A ∈ R in the algebra R is the union of the spectrum σ(A) of A in the algebra B and (possibly empty) collection of bounded connected components of the resolvent set ρ(A). In particular, the spectral radii rR(A) and rB(A) coincide. Let A ∈ B and let U ⊆ C be an open set that contains the spectrum σ(A). The set U must not be connected. Let f : U → C be an analytic function. The function f of the element A is defined [29, ch. V, § 1], [10, p. 17] by the formula f (A) = f (λ)(λ1 − A)−1 dλ, (2) 1 2πiZΓ where the contour Γ surrounds the set σB(A) in the counterclockwise direction and the function f is analytic inside Γ. Proposition 2 ([29, Theorem 5.2.5], [48, Theorem 10.27]). The mapping f 7→ f (A) preserves algebraic operations, i. e., (f + g)(A) = f (A) + g(A), (αf )(A) = αf (A), (f g)(A) = f (A)g(A), where f + g, αf , and f g are defined pointwise. Corollary 3. For the function rλ0(λ) = 1 λ0−λ , λ0 ∈ ρ(A), we have rλ0(A) = (λ01 − A)−1. Proof. The proof follows from Proposition 2. (cid:3) 2. The differential equation with a constant coefficient In this Section, we describe three analytic functions that are closely related to the representation of solutions of linear differential equations with constant coefficients. For λ ∈ C and t ∈ R, we consider the functions exp+, t(λ) =(eλt, exp−, t(λ) =(0, 0, −eλt, if t > 0, if t < 0, if t > 0, if t < 0, gt(λ) =(exp−, t(λ), exp+, t(λ), if Re λ > 0, if Re λ < 0. These functions are undefined for t = 0. The function gt is also undefined for Re λ = 0. For any fixed t 6= 0, all three functions are analytic on their domains. Let X be a Banach space and A ∈ B(X). We consider the differential equation (3) We recall two well-known theorems. The first theorem shows that exp±, (·)(A) are x′(t) = Ax(t) + f (t), t ∈ R. fundamental solutions of the initial value problems. 4 V.G. KURBATOV AND I.V. KURBATOVA Theorem 4 ([10, ch. 1, § 4], [26, ch. IV, Corollary 2.1]). Let f : R → X be a continuous function. The solution of the initial value problem x′(t) = Ax(t) + f (t), x(0) = 0 t > 0, is the function x(t) =Z t 0 exp+, s(A) f (t − s) ds, t > 0. The solution of the initial value problem is the function x′(t) = Ax(t) + f (t), x(0) = 0 t < 0, x(t) =Z 0 t exp−, s(A) f (t − s) ds, t < 0. The function t 7→ exp+, t(A) is called [26] the fundamental solution for equation (3). Now we turn to the bounded solutions problem, i.e. the problem of seeking bounded solution x : R → X under the assumption that the free term f : R → X is a bounded function. Theorem 5 ([10, Theorem 4.1, p. 81]). Let A ∈ B(X). Equation (3) has a unique bounded on R solution x for any bounded continuous function f if and only if the spectrum σ(A) of A does not intersect the imaginary axis. This solution admits the representation where x(t) =Z ∞ −∞ G(s)f (t − s) ds, G(t) = gt(A), t 6= 0. The function G is called [10] the Green's function of the bounded solutions problem for equation (3). 3. Causal spectrum of a block triangular matrix Let a Banach space X be represented as the direct sum of its closed nonzero subspaces Xi, i = 1, . . . , n: This means that every x ∈ X can be uniquely represented in the form X = X1 ⊕ X2 ⊕ · · · ⊕ Xn. x = x1 + x2 + · · · + xn, where xi ∈ Xi, i = 1, . . . , n. It is easy to prove that the norm on X is equivalent to the norm We denote by M = M(X1, X2, . . . , Xn) the set of all operator matrices kxk = kx1k + kx2k + · · · + kxnk. { Tij ∈ B(Xj, Xi) : i, j = 1, . . . , n }. We endow M with the norm k{ Tij }k = maxjPn i=1kTijk. It is easy to show that M is a unital Banach algebra with respect to the usual matrix multiplication, and the Banach algebra M is isomorphic (not isometrically) to the algebra B(X). As usual, we do not distinguish very carefully matrices and operators induced by them. GREEN'S FUNCTION IN THE CASE OF A TRIANGULAR COEFFICIENT 5 We denote by M+ = M+(X1, X2, . . . , Xn) the set of all lower triangular matrices . (4) A11 A21 . . . 0 A22 . . . An−1,1 An−1,2 An,1 An,2   0 0 . . . . . . . . . . . . . . . An−1,n−1 . . . An,n−1 An,n 0 0 . . . 0   We denote by B+(X) the class of operators induces by M+. Clearly, M+ is a closed subalgebra of the algebra M. Therefore, B+(X) is a closed subalgebra of the algebra B(X). We call operators from the class B+(X) causal in analogy with a similar class of operators in the control theory [13, 16, 53] and in the theory of functional differential equations [34, 35, 36], see also the references therein. Namely, if one interprets the indices i = 1, . . . , n as successive instants of time, then the triangularity of a matrix A means that the value (Ax)i of the 'output' Ax at any instant i may depend only on values xj of the 'input' x at the previous instants j ≤ i. Remark 1. The subalgebra M+ (and consequently, the subalgebra B+(X)) may be not full if the space X is infinite-dimensional. We give a corresponding example. Let X be the space Lp(R), 1 ≤ p ≤ ∞. We represent X = Lp(R) as Lp(−∞, 0] ⊕ Lp[0, ∞), where Lp(−∞, 0] and Lp[0, ∞) are the subspaces of functions that are equal to zero out- is induced by a lower triangular matrix (thus it is causal), but the inverse operator side (−∞, 0] and [0, ∞) respectively. Clearly, the operator of delay (cid:0)Sx(cid:1)(t) = x(t − 1) (cid:0)S−1x(cid:1)(t) = x(t + 1) is induced by an upper triangular matrix (thus S−1 is not causal). Consequently, in contrast to the finite-dimensional case, the (ordinary) spectrum of a triangular matrix may be not the union of the spectra of its diagonal blocks, see Propo- sition 7. See a more detailed discussion of this phenomenon in [25]. If an operator T ∈ B+(X) is invertible and the inverse operator belongs to B+(X), we say that T is causally invertible. We call the spectrum of T ∈ B+(X) in the algebra B+(X) the causal spectrum and denote it by σ+(T ). Clearly, σ(T ) ⊆ σ+(T ). We denote by ρ+(T ) the causal resolvent set C \ σ+(T ). The same terminology and notation will be used for matrices M ∈ M+. We recall that an open set D ⊆ C is called simply-connected if any simple closed curve in D can be shrunk continuously to a point. Proposition 6. Let the domain D ⊆ C of an analytic function f be simply-connected (examples of such functions are exp±, t and gt). Let T ∈ B+(X). Then σ+(T ) ⊂ D provided σ(T ) ⊂ D. Thus the function f (T ) of a causal operator T is defined in algebras B(X) and B+(X) simultaneously. Proof. A possible difficulty can occur when the spectrum σ(T ) is contained in the domain D of the definition of f , but σ+(T ) * D. Therefore the resolvent (λ1−A)−1 in integral (2) is defined in B(X), but it may not exist in B+(X). By Proposition 1, the causal spectrum σ+(T ) is the union of the ordinary spectrum σ(T ) and (possibly) some bounded components of the resolvent set ρ(A). Since the domain D of f is simply-connected, bounded components of the resolvent set ρ(A) are contained in the domain D, provided the spectrum σ(T ) itself is contained in the domain D. (cid:3) 6 V.G. KURBATOV AND I.V. KURBATOVA Proposition 7 ([34], [36, Proposition 2.1.7]). The causal spectrum of a lower triangular matrix { Tij } (and the causal spectrum of the corresponding operator) is the union of the (ordinary) spectra σ(Tii) of the diagonal blocks Tii. Proof. It suffices to prove that a lower triangular matrix has a lower triangular inverse if and only if all diagonal blocks Tii are invertible. Let the lower triangular matrix { Bij } be the inverse of the lower triangular ma- trix { Tij }. Then it follows from the matrix multiplication rule that Bii are inverses of Tii. Conversely, let the diagonal blocks Tii be invertible. Then from the Gaussian elimination (cid:3) algorithm it easily follows that the inverse matrix exists and is triangular. 4. Functions of block triangular matrices Theorem 8. Let a causal matrix be causally invertible. Then the elements of the inverse matrix T =  B =  T1,1 T2,1 . . . 0 T2,2 . . . Tn−1,1 Tn−1,2 Tn,1 Tn,2 0 0 . . . . . . . . . . . . . . . Tn−1,n−1 . . . Tn,n−1 0 0 . . . 0 Tn,n B1,1 B2,1 . . . 0 B2,2 . . . Bn−1,1 Bn−1,2 Bn,1 Bn,2 0 0 . . . . . . . . . . . . . . . Bn−1,n−1 . . . Bn,n−1 Bn,n 0 0 . . . 0     have the form Bi,j = Xi=i1>i2···>im=j In particular, (−1)m+1T −1 i1,i1 Ti1,i2T −1 i2,i2 Ti2,i3 . . . Tim−1,imT −1 im,im , i ≥ j. Bi,i = T −1 i,i , Bi+1,i = −T −1 Bi+2,i = −T −1 i+1,i+1Ti+1,iT −1 i,i , i+2,i+2Ti+2,iT −1 i,i + T −1 i+2,i+2Ti+2,i+1T −1 i+1,i+1Ti+1,iT −1 i,i . Proof. Let us verify that T B = 1. Clearly, (T B)ii = 1. We calculate, for example, (T B)n,1: (T B)n,1 = Tn,1T −1 11 − Tn,2T −1 33 T31T −1 11 + T −1 33 T32T −1 22 T21T −1 11 ) + . . . 22 T21T −1 (−1)m+1T −1 11 + Tn,3(−T −1 n,nTn,i2T −1 i2,i2 Ti2,i3 . . . Tim−1,1T −1 1,1 22 T21T −1 11 + Tn,3(−T −1 33 T31T −1 11 + T −1 33 T32T −1 22 T21T −1 11 ) + . . . (−1)m+1Tn,i2T −1 i2,i2Ti2,i3 . . . Tim−1,1T −1 1,1 = 0. = Tn,1T −1 + Tn,n Xn=i1>i2···>im=1 + Xn=i1>i2···>im=1 11 − Tn,2T −1 In a similar way one establishes that BT = 1. (cid:3) GREEN'S FUNCTION IN THE CASE OF A TRIANGULAR COEFFICIENT 7 Theorem 9. Let A ∈ M+. Then the causal resolvent (i.e., the resolvent (λ1 − A)−1 at the points λ ∈ ρ+(A)) of A has the form (λ1 − A)−1 = R1,1 R2,1 . . . 0 R2,2 . . . Rn−1,1 Rn−1,2 Rn,2 Rn,1   0 0 . . . . . . . . . . . . . . . Rn−1,n−1 . . . Rn,n−1 Rn,n 0 0 . . . 0 ,   where Rij for i ≥ j are defined by the formula Rij = Xi=i1>i2···>im=j In particular, (λ1 − Ai1,i1)−1Ai1,i2(λ1 − Ai2,i2)−1Ai2,i3 . . . Aim−1,im(λ1 − Aim,im)−1. Rii = (λ1 − Aii)−1, Ri+1,i = (λ1 − Ai+1,i+1)−1Ai+1,i(λ1 − Aii)−1, Ri+2,i = (λ1 − Ai+2,i+2)−1Ai+2,i(λ1 − Aii)−1 + (λ1 − Ai+2,i+2)−1Ai+2,i+1(λ1 − Ai+1,i+1)−1Ai+1,i(λ1 − Aii)−1. Proof. The proof follows from Theorem 8. The sign (−1)m+1 disappears because (λ1 − A)ij = −Aij for i > j. (cid:3) Theorem 10. Let a function f be analytic in a neighborhood of the causal spectrum σ+(A) of a matrix A ∈ M+. Then the matrix F = f (A) has the form F = 0 F2,2 . . . 0 0 . . . . . . . . . . . . . . . Fn−1,n−1 . . . Fn,n−1 Fn−1,1 Fn−1,2 Fn,1 Fn,2 0 0 . . . 0 Fn,n ,   F1,1 F2,1 . . .   2πiZΓ 1 where Fij for i ≥ j are defined by the formula Fij = Xi=i1>i2···>im=j × Aim−1,im(λ1 − Aim,im)−1 dλ, f (λ)(λ1 − Ai1,i1)−1Ai1,i2(λ1 − Ai2,i2)−1Ai2,i3 . . . where Γ surrounds the causal spectrum σ+(A) of the matrix A. In particular, Fii =ZΓ Fi+1,i =ZΓ Fi+2,i =ZΓ +ZΓ f (λ)(λ1 − Aii)−1 dλ, f (λ)(λ1 − Ai+1,i+1)−1Ai+1,i(λ1 − Aii)−1 dλ, f (λ)(λ1 − Ai+2,i+2)−1Ai+2,i(λ1 − Aii)−1 dλ f (λ)(λ1 − Ai+2,i+2)−1Ai+2,i+1(λ1 − Ai+1,i+1)−1Ai+1,i(λ1 − Aii)−1 dλ. Proof. Substituting into formula (2) the representation of (λ1 − A)−1 from Theorem 9, we obtain the desired result. (cid:3) 8 V.G. KURBATOV AND I.V. KURBATOVA Remark 2. Let a matrix A ∈ M+ has only two non-zero diagonals: .   0 A1,1 A2,1 A2,2 . . . . . . 0 0 0 0 0 0 . . . . . . . . . . . . . . . An−1,n−1 . . . An,n−1 An,n 0 0 . . . 0   Let a function f be analytic in a neighborhood of the causal spectrum σ+(A) of the matrix A. Then it follows from Theorem 10 that the elements Fij for i ≥ j of the matrix F = f (A) consist of exactly one summand: Fij = 1 2πiZΓ f (λ)(λ1 − Ai,i)−1Ai,i+1(λ1 − Ai+1,i+1)−1Ai+1,i+2 . . . Aj−1,j(λ1 − Aj,j)−1 dλ. For the function f = exp+, t this phenomenon was described in [6] and applied in [24]. Corollary 11. Let the domain D ⊆ C of an analytic function f be simply-connected (examples of such functions are exp±, t and gt). Then the conclusion of Theorem 10 is true if the function f is analytic in a neighborhood of the ordinary spectrum σ(A) of the matrix A ∈ M+. Proof. The proof follows from Proposition 6. (cid:3) Remark 3. For scalar matrices, Theorem 10 goes back to [47]. For matrices consisting of finite-dimensional blocks, it was published in [12, Theorem 2]. More precisely, in [12] it was considered only the case of polynomial functions f ; but from the case of polynomials, it follows the case of general analytic functions, since (if A is a scalar matrix) any analytic function can be replaced by its interpolating polynomial. 5. Divided differences Let µ1, µ2, . . . , µn be given complex numbers (some of them may coincide with others) called points of interpolation. Let a complex-valued function f be defined and analytic in a neighborhood of these points. Divided differences of the function f with respect to the points µ1, µ2, . . . , µn are defined (see, e.g., [18, 30]) by the recurrent relations f [0](µi) = f (µi), f [1](µi, µi+1) = f [0](µi+1) − f [0](µi) µi+1 − µi , f [m](µi, . . . , µi+m) = f [m−1](µi+1, . . . , µi+m) − f [m−1](µi, . . . , µi+m−1) µi+m − µi (5) . In these formulas, if the denominator vanishes, then the quotient is understood as the derivative with respect to one of the arguments of the previous divided difference (the naturalness of this agreement can be derived by continuity from Corollary 13). Proposition 12 ([18, ch. 1, formula (54)]). Let a function f be analytic in an open neighbourhood of the points of interpolation µ1, µ2, . . . , µn. Then divided differences admit the representation f [n−1](µ1, . . . , µn) = 1 2πiZΓ f (z) Ω(z) dz, GREEN'S FUNCTION IN THE CASE OF A TRIANGULAR COEFFICIENT 9 where the contour Γ encloses all the points of interpolation and n Corollary 13. Divided differences are differentiable functions of their arguments. Ω(z) = (z − µk). Yk=1 Proof. The proof follows from Proposition 12. Corollary 14. If D ⊆ C is the domain of an analytic function f , then f [n−1] is defined in Dn. (cid:3) Proof. The proof follows from Proposition 12. Corollary 15. Divided differences f [n−1](µ1, . . . , µn) are symmetric function, i.e., they do not depend on the order of their arguments µ1, . . . , µn. (cid:3) Proof. The proof follows from Proposition 12. (cid:3) Proposition 16 ([18, ch. 1, formula (48)], [30, p. 19, formula (1)]). Let the points of interpolation µ1, . . . , µn be distinct. Then divided differences admit the representation f [n−1](µ1, . . . , µn) = n Xj=1 Proof. The proof follows from Proposition 12. f (µj) n . (µj − µk) Qk=1 k6=j (cid:3) For a function f analytic in a neighborhood of the causal spectrum of a matrix A ∈ M+, we denote by f [m−1](A; i1, i2, . . . , im) the summands from Theorem 10: f [m−1](A; i1, i2, . . . , im) = 1 2πiZΓ f (λ)(λ1 − Ai1,i1)−1Ai1,i2(λ1 − Ai2,i2)−1Ai2,i3 . . . × Aim−1,im(λ1 − Aim,im)−1 dλ, (6) where Γ surrounds the causal spectrum σ+(A) of the matrix A. Theorem 17. Let a function f be analytic in a neighborhood of the causal spectrum σ+(A) of a matrix A ∈ M+. Then f [m−1](A; i1, i2, . . . , im) = 1 (2πi)m ZΓi1 · · ·ZΓim f [m−1](λi1, . . . , λim)(λi11 − Ai1,i1)−1 (7) × Ai1,i2(λi21 − Ai2,i2)−1Ai2,i3 . . . Aim−1,im(λim1 − Aim,im)−1 dλi1 . . . dλim, where Γik surrounds the spectrum of Aik,ik . Proof. Since f is analytic in a neighborhood of σ+(A) =Sn we may assume without loss of generality that Γik in (7) surrounds the whole Sn and, moreover, Γik surrounds Γik+1, see Figure 1. i=1 σ(Aii) (see Proposition 7), i=1 σ(Aii) Since the contours Γik are disjoint, the points λi1, . . . , λim in the integrand of (7) are distinct. Therefore we can substitute the representation of divided differences from Proposition 16 into definition (7): f [m−1](A; i1, i2, . . . , im) = 1 (2πi)m ZΓi1 · · ·ZΓim m Xj=1 f (λij ) m (λij − λik) (8) × (λi11 − Ai1,i1)−1Ai1,i2 . . . Aim−1,im(λim1 − Aim,im)−1 dλi1 . . . dλim. Qk=1 k6=j 10 V.G. KURBATOV AND I.V. KURBATOVA ✬ ✫ ✬ ⑦⑦⑦✬ ✫ ✫ σ+(A) ✻Γi3 ✩ ✪ ✻Γi2 ✩ ✩ ✻Γi1 ✪ ✪ f Figure 1. The choice of the contours Γik in the proof of Theorem 17. The symbol f means the localization of singularities of f Let us begin with the first summand. We have f (λi1)(λi11 − Ai1,i1)−1Ai1,i2 1 2πiZΓii1 ×(cid:20). . . ×(cid:20) 1 ×(cid:20) 1 1 2πiZΓim−2 2πiZΓim−1 2πiZΓim 1 1 λi1 − λim−2 × (λim−21 − Aim−2,im−2)−1Aim−2,im−1 1 λi1 − λim−1 (λim−11 − Aim−1,im−1)−1Aim−1,im λi1 − λim (λim1 − Aim,im)−1 dλim(cid:21) dλim−1(cid:21) dλim−2 . . .(cid:21) dλi1. By Corollary 3, for the internal integral, we have 1 2πiZΓim 1 λi1 − λim (λim1 − Aim,im)−1 dλim = (λi11 − Aim,im)−1. Now we can calculate the next internal integral (again using Corollary 3): 1 2πiZΓim−1 =(cid:20) 1 2πiZΓim−1 1 λi1 − λim−1 (λim−11 − Aim−1,im−1)−1Aim−1,im(λi11 − Aim,im)−1 dλim−1 1 λi1 − λim−1 (λim−11 − Aim−1,im−1)−1 dλim−1(cid:21)Aim−1,im(λi11 − Aim,im)−1 = (λi11 − Aim−1,im−1)−1Aim−1,im(λi11 − Aim,im)−1. And so on. Finally, we arrive at the representation (for the first summand in (8)) 1 2πiZΓi1 f (λi1)(λi11 − Ai1,i1)−1Ai1,i2(λi11 − Ai2,i2)−1Ai2,i3 . . . Aim−1,im(λi11 − Aim,im)−1 dλi1, which coincides with formula (7). Next we show that the other summands in (8) are zero. Let us consider, for example, the second summand 1 (2πi)m ZΓi1 · · ·ZΓim f (λi2) m (λi2 − λik ) Qk=1 k6=2 × (λi11 − Ai1,i1)−1Ai1,i2 . . . Aim−1,im(λim1 − Aim,im)−1 dλi1 . . . dλim. GREEN'S FUNCTION IN THE CASE OF A TRIANGULAR COEFFICIENT 11 Proceeding as above (i.e. successively calculating integrals over all variables except λi1), at the final stage, we arrive at the integral 1 2πiZΓi1 1 λi2 − λi1 (λi11 − Ai1,i1)−1 dλi1. We notice that the singularity of the function λi1 7→ 1 lies inside the contour Γi1. Hence the integrand λi1 7→ 1 outside the contour Γi1 and decreases at infinity as zero. λi2 −λi1 1 λ2 i1 λi2 −λi1 (i.e., the point λi2 ∈ Γi2) (λi11 − Ai1,i1)−1 is analytic . Therefore the integral equals (cid:3) Remark 4. For the case of the first divided difference, a more detailed discussion of formula (7) can be found in [39, Theorem 41], see also the references therein. Corollary 18. Let a function f be analytic in a neighborhood of the causal spectrum σ+(A) of a matrix A ∈ M+. Then the matrix F = f (A) has the form where Fij, i ≥ j, admits the representation F11 F21 . . . 0 F22 . . . Fn−1,1 Fn−1,2 Fn,1 Fn,2 F =  Fij = Xi=i1>i2···>im=j 0 0 . . . . . . . . . . . . . . . Fn−1,n−1 . . . Fn,n−1 0 0 . . . 0 Fn,n ,   f [m−1](A; i1, i2, . . . , im). Proof. The proof follows from Theorems 10 and 17. (cid:3) Theorem 19. Let A ∈ M+. Then E±,1,1 E±,2,1 where E±,i,j, i ≥ j, admits the representation 0 E±,2,2 . . . . . . exp±, t(A) =  E±,i,j = Xi=i1>i2···>im=j E±,n−1,1 E±,n−1,2 E±,n,2 E±,n,1 0 0 . . . . . . . . . . . . . . . E±,n−1,n−1 . . . E±,n,n−1 E±,n,n 0 0 . . . 0 ,   exp[m−1] ±, t (A; i1, i2, . . . , im); and (provided the spectrum σ(A) does not intersect the imaginary axis) gt(A) = 0 G2,2 . . . Gn−1,1 Gn−1,2 Gn,2 Gn,1 G1,1 G2,1 . . .   Gi,j = Xi=i1>i2···>im=j 0 0 . . . . . . . . . . . . . . . Gn−1,n−1 . . . Gn,n−1 Gn,n 0 0 . . . 0 ,   g[m−1] t (A; i1, i2, . . . , im). where Gi,j, i ≥ j, admits the representation Proof. The proof follows from Corollary 11 and Theorem 17. (cid:3) 12 V.G. KURBATOV AND I.V. KURBATOVA 6. Divided differences of the functions exp±, t and gt Lemma 20. Let the points of interpolation λj be distinct. Then the divided differences of the function rλ(ν) = 1 λ−ν admit the representation Proof. The proof is by induction on n. For n = 1 we have r[n−1] λ (λ1, . . . , λn) = j=1(λ − λj) . 1 Qn r[1] λ (λ1, λ2) = λ−λ1 1 − 1 λ1 − λ2 λ−λ2 = 1 (λ − λ1)(λ − λ2) . Assuming that the formula holds for n − 2, we prove it for n − 1. We have r[n−1] λ (λ1, . . . , λn) = r[n−1] λ (λ1, . . . , λn−2, λn−1) − r[n−1] λ λn−1 − λn (λ1, . . . , λn−2, λn) We recall [15, 40, 49] that the bilateral (or two-sided) Laplace transform of a function f : R → C is the function The value (cid:0)Bf(cid:1)(λ) at the point λ ∈ C is defined if the integral converges absolutely. If f equals zero on (−∞, 0) (as the function exp+, (·)), this definition takes the form 1 1 Qn−2 j=1 (λ−λj ) λ−λn−1 − 1 1 Qn−2 j=1 (λ−λj ) λ−λn λn−1 − λn 1 1 λ−λn−1 − 1 λ−λn λn−1 − λn = = = . (cid:3) 1 j=1(λ − λj) j=1 (λ − λj) Qn−2 Qn (cid:0)Bf(cid:1)(λ) =Z ∞ −∞ e−λtf (t) dt. (cid:0)Bf(cid:1)(λ) =Z ∞ 0 e−λtf (t) dt. (cid:0)Bf(cid:1)(λ) =Z 0 −∞ e−λtf (t) dt. Usually in this case, the integral converges absolutely for Re λ sufficiently large. If f equals zero on (0, ∞) (as the function exp−, (·)), the definition of the bilateral Laplace transform takes the form In this case, we assume that the integral converges absolutely for Re λ sufficiently small. We recall the following statement. Lemma 21. Let λ0 ∈ C. (a) The bilateral Laplace transform of the function t 7→ exp+, t(λ0) is the function (cid:0)B exp+, (·)(λ0)(cid:1)(λ) = 1 λ − λ0 , Re λ > Re λ0. (b) The bilateral Laplace transform of the function t 7→ exp−, t(λ0) is the function (cid:0)B exp−, (·)(λ0)(cid:1)(λ) = 1 λ − λ0 , Re λ < Re λ0. GREEN'S FUNCTION IN THE CASE OF A TRIANGULAR COEFFICIENT 13 (c) The bilateral Laplace transform of the function t 7→ gt(λ0), Re λ0 6= 0, is the function (the complete domain of the function of B g(·)(λ0) is Re λ > Re λ0 if Re λ0 < 0 and is Re λ < Re λ0 if Re λ0 > 0) (cid:0)B g(·)(λ0)(cid:1)(λ) = 1 λ − λ0 , Re λ < Re λ0. Proof. Assertion (a) is widely known [40, pp. 300, 305]. The proofs of all assertions are reduced to straightforward calculations. The proof of assertion (c) can also be obtained from the definition of gt and (a) and (b). (cid:3) We recall [48, ch. 6] that the convolution of two summable functions f, g : R → C is the function If f (t) = 0 and g(t) = 0 for t < 0, then this formula takes the form (cid:0)f ∗ g(cid:1)(t) =Z ∞ −∞ f (s)g(t − s) ds. If f (t) = 0 and g(t) = 0 for t > 0, then the definition of convolution takes the form (cid:0)f ∗ g(cid:1)(t) =(R t (cid:0)f ∗ g(cid:1)(t) =(0 R 0 0 f (s)g(t − s) ds 0 for t > 0, for t < 0. t f (s)g(t − s) ds for t > 0, for t < 0. Theorem 22. (a) The divided differences of the function t 7→ exp+, t admit the representation exp[n−1] +, (·) (λ1, . . . , λn) = exp+, (·)(λ1) ∗ · · · ∗ exp+, (·)(λn). For example, for t > 0, we have exp[1] +, t(λ1, λ2) =Z t +, t(λ1, λ2, λ3) =Z t 0 Z r 0 0 exp[2] exp+, s(λ1) exp+, t−s(λ2) ds, exp+, s(λ1) exp+, r−s(λ2) exp+, t−r(λ3) ds dr. (b) The divided differences of the function t 7→ exp−, t admit the representation exp[n−1] −, (·) (λ1, . . . , λn) = exp−, (·)(λ1) ∗ · · · ∗ exp−, (·)(λn). For example, for t < 0, we have exp[1] −, t(λ1, λ2) =Z 0 −, t(λ1, λ2, λ3) =Z 0 t Z 0 r t exp[2] exp−, s(λ1) exp−, t−s(λ2) ds, exp−, s(λ1) exp−, r−s(λ2) exp−, t−r(λ3) ds dr. (c) The divided differences of the function t 7→ gt admit the representation g[n−1] (·) (λ1, . . . , λn) = g(·)(λ1) ∗ · · · ∗ g(·)(λn), Re λ1, . . . , Re λn 6= 0. 14 V.G. KURBATOV AND I.V. KURBATOVA For example, for t 6= 0, we have gs(λ1)gt−s(λ2) ds, g[1] t (λ1, λ2) =Z ∞ t (λ1, λ2, λ3) =Z ∞ −∞ g[2] −∞Z ∞ −∞ gs(λ1)gr−s(λ2)gt−r(λ3) ds dr. Remark 5. Assertion (a) is established in [51]. Assertion (b) is proved in a similar way. Assertion (c) is proved in [20, p. 53] for other aims. For completeness, we give here an independent proof of (c). Proof. Suppose that the points of interpolation λj are distinct. By Proposition 16, we have g[n−1] t (λ1, . . . , λn) = Xj=1 n gt(λj) n . (λj − λk) Qk=1 k6=j Form this representation and Lemma 21, it easily follows that the bilateral Laplace trans- form of the function t 7→ g[n−1] (λ1, . . . , λn) is t (·) (cid:0)B g[n−1] (λ1, . . . , λn)(cid:1)(λ) = 1 λ − λj n Xj=1 1 n . (λj − λk) Qk=1 k6=j By Proposition 16, the last expression is the (n − 1)-th divided difference of the function rλ(ν) = 1 λ−ν . By Lemma 20, r[n−1] λ (λ1, . . . , λn) = 1 j=1(λ − λj) . Qn We apply the inverse bilatiral Laplace transform to tht function λ 7→ r[n−1] (λ1, . . . , λn). Clearly, the restriction of the bilatiral Laplace transform to the imaginary axis is the Fourier transform. The Fourier transform maps the convolution of functions to the prod- uct of their images [40, p. 337], which implies assertion (c). λ The case of coinciding points of interpolation λj follows from continuity. (cid:3) 7. The divided differences exp[m] t with operator arguments In this Section, we apply previous results to the calculation of the functions exp[m] g[m] t with operator arguments. Theorem 23. Let A ∈ M+. Then for t > 0, we have ±, t and g[m] ±, t and exp[m−1] +, t (A; i1, i2, . . . , im−1, im) =Z t 0 Z sm−1 0 · · ·Z s2 0 exp+, s1(Ai1,i1)Ai1,i2 exp+, s2−s1(Ai2,i2) × Ai2,i3 . . . exp+, sm−1−sm−2(Aim−1,im−1)Aim−1,im exp+, t−sm−1(Aim,im) ds1 . . . dsm−1; for t < 0, we have exp[m−1] −, t (A; i1, i2, . . . , im−1, im) =Z 0 t Z 0 sm−1 · · ·Z 0 s2 × Ai2,i3 . . . exp−, sm−1−sm−2(Aim−1,im−1)Aim−1,im exp−, t−sm−1(Aim,im) ds1 . . . dsm−1; exp−, s1(Ai1,i1)Ai1,i2 exp−, s2−s1(Ai2,i2) GREEN'S FUNCTION IN THE CASE OF A TRIANGULAR COEFFICIENT 15 and (if the spectrum σ(A) is disjoint from the imaginary axis) for t 6= 0, we have g[m−1] t (A; i1, i2, . . . , im−1, im) =Z ∞ −∞ · · ·Z ∞ −∞ gs1(Ai1,i1)Ai1,i2 gs2−s1(Ai2,i2) × Ai2,i3 . . . gsm−1−sm−2(Aim−1,im−1)Aim−1,im gt−sm−1(Aim,im) ds1 . . . dsm−1. Proof. For simplicity of notation, we prove only the formula g[2] t (A; 3, 2, 1) =Z ∞ −∞Z ∞ −∞ By Theorem 17, we have gs1(A33)A32 gs2−s1(A22)A21 gt−s2(A11) ds1 ds2. g[2] t (A; 3, 2, 1) = 1 (2πi)3 ZΓ1ZΓ2ZΓ3 g[2] t (λ1, λ2, λ3) × (λ31 − A33)−1A32(λ21 − A22)−1A21(λ11 − A11)−1 dλ1 dλ2 dλ3. By Theorem 22, we have g[2] t (λ1, λ2, λ3) =Z ∞ −∞Z ∞ −∞ gs(λ1) gr−s(λ2) gt−r(λ3) ds dr. Substituting the latter formula into the former one and performing the integration over λ1, λ2, and λ3, we arrive at the desired formula. (cid:3) Combining Theorems 19 and 23, we obtain the following examples. Example 1. Let A be the block matrix Then for t > 0, we have 0 0 A11 0 A21 A22 A31 A32 A33 A =    . exp+, t(A) =  where exp+, t(A11) exp[1] +, t(A; 2, 1) +, t(A; 3, 1) + exp[2] exp[1] 0 exp+, t(A22) 0 0 +, t(A; 3, 2, 1) exp[1] +, t(A; 3, 2) exp+, t(A33) (9) (10) (11)   , exp[1] +, t(A; i1, i2) =Z t +, t(A; 3, 2, 1) =Z t 0 Z r 0 0 exp[2] exp+, s(Ai,1i1)Ai1,i2 exp+, t−s(Ai2,i2) ds, exp+, s(A33)A32 exp+, r−s(A22)A21 exp+, t−r(A11) ds dr. Remark 6. Integral (11) was first obtained in [3, ch. 10, § 14, formula (5)] in a different context. Formula (10) for a triangular block matrix (with blocks consisting of scalars) of the size less than or equal to 4 × 4 was appeared in [51, theorem 1] and for a triangular block matrix of any size in [6]. Example 2. Let A be the block matrix (9), whose spectrum is disjoint from the imaginary axis. Then for t 6= 0, we have gt(A) =  gt(A11) g[1] t (A; 2, 1) g[1] t (A; 3, 1) + g[2] t (A; 3, 2, 1) g[1] 0 gt(A22) t (A; 3, 2) gt(A33) 0 0   , 16 where V.G. KURBATOV AND I.V. KURBATOVA gs(Ai1,i1)Ai1,i2gt−s(Ai2,i2) ds, g[1] t (A; i1, i2) =Z ∞ t (A; 3, 2, 1) =Z ∞ −∞ −∞Z ∞ −∞ g[2] gs(A33)A32gr−s(A22)A21gt−r(A11) ds dr. Acknowledgments The first author was supported by the Ministry of Education and Science of the Russian Federation under state order No. 3.1761.2017/4.6. The second author was supported by the Russian Foundation for Basic Research under research projects No. 16-01-00197. References [1] R. R. Akhmerov and V. G. Kurbatov, Exponential dichotomy and stability of neutral type equations, J. Differential Equations 76 (1988), no. 1, 1 -- 25. MR 964610 [2] A. G. Baskakov, Estimates for the Green's function and parameters of exponential dichotomy of a hyperbolic operator semigroup and linear relations, Mat. Sb. 206 (2015), no. 8, 23 -- 62, (in Russian); English translation in Sb. Math., 206 (2015), no. 8, 1049 -- 1086. MR 3438589 [3] R. Bellman, Introduction to matrix analysis, McGraw-Hill Book Co., New York -- Toronto -- London, 1960. MR 0122820 [4] A. A. Boichuk and A. A. Pokutnii, Bounded solutions of linear differential equations in a Banach space, Nonlinear Oscillations 9 (2006), no. 1, 3 -- 14. MR 2369770 [5] N. Bourbaki, ´El´ements de math´ematique. Fascicule XXXII. Th´eories spectrales. Chapitre I: Alg`ebres norm´ees. Chapitre II: Groupes localement compacts commutatifs, Actualit´es Scientifiques et Industrielles, No. 1332, Hermann, Paris, 1967, (in French). MR 0213871 [6] F. Carbonell, J. C. J´ımenez, and L. M. Pedroso, Computing multiple integrals involving matrix exponentials, J. Comput. Appl. Math. 213 (2008), no. 1, 300 -- 305. MR 2382698 [7] M. Ceballos, J. N´unez, and A. F. Tenorio, Complete triangular structures and Lie algebras, Int. J. Comput. Math. 88 (2011), no. 9, 1839 -- 1851. MR 2810866 [8] Wai-Shun Cheung, Lie derivations of triangular algebras, Linear and Multilinear Algebra 51 (2003), no. 3, 299 -- 310. MR 1995661 [9] C. Chicone and Y. Latushkin, Evolution semigroups in dynamical systems and differential equations, Mathematical Surveys and Monographs, vol. 70, American Mathematical Society, Providence, RI, 1999. MR 1707332 [10] Ju. L. Daleckiı and M. G. Kreın, Stability of solutions of differential equations in Banach space, Translations of Mathematical Monographs, vol. 43, American Mathematical Society, Providence, RI, 1974. MR 0352639 [11] Ph. I. Davies and N. J. Higham, A Schur-Parlett algorithm for computing matrix functions, SIAM J. Matrix Anal. Appl. 25 (2003), no. 2, 464 -- 485 (electronic). MR 2047429 [12] Ch. Davis, Explicit functional calculus, Linear Algebra and Appl. 6 (1973), 193 -- 199. MR 0327792 [13] C.A. Desoer and M. Vidyasagar, Feedback systems: input-output properties, Academic Press, New York -- London, 1975. MR 0490289 [14] L. Dieci and A. Papini, Pad´e approximation for the exponential of a block triangular matrix, Linear Algebra Appl. 308 (2000), no. 1-3, 183 -- 202. MR 1751139 [15] M. A. Evgrafov, Analiticheskie funktsii [Analytic functions], Izdat. "Nauka", Moscow, 1965, (in Russian). MR 0188404 [16] A. Feintuch and R. Saeks, System theory: A Hilbert space approach, Pure and Applied Mathematics, vol. 102, Academic Press, Inc. [Harcourt Brace Jovanovich, Publishers], New York -- London, 1982. MR 663906 [17] Richard P. Feynman, An operator calculus having applications in quantum electrodynamics, Physical Rev. (2) 84 (1951), 108 -- 128. MR 0044379 [18] A. O. Gel′fond, Calculus of finite differences, International Monographs on Advanced Mathematics and Physics, Hindustan Publishing Corp., Delhi, 1971, Translation of the third Russian edition. MR 0342890 [19] H. Ghahramani, Zero product determined triangular algebras, Linear and Multilinear Algebra 61 (2013), no. 6, 741 -- 757. MR 3005653 [20] S. K. Godunov, Ordinary differential equations with constant coefficient, Translations of Mathematical Mono- graphs, vol. 169, American Mathematical Society, Providence, RI, 1997, Translated from the 1994 Russian original by Tamara Rozhkovskaya. MR 1465434 [21] , Lectures on modern aspects of linear algebra, University series, vol. 12, Science book, Novosibirsk, Russia, 2002, (in Russian). GREEN'S FUNCTION IN THE CASE OF A TRIANGULAR COEFFICIENT 17 [22] I. Gohberg, S. Goldberg, and M. A. Kaashoek, Classes of linear operators. Vol. II, Operator Theory: Ad- vances and Applications, vol. 63, Birkhauser Verlag, Basel, 1993. MR 1246332 [23] I. Gohberg and M. G. Kreın, Theory and applications of Volterra operators in Hilbert space, Translations of Mathematical Monographs, vol. 24, American Mathematical Society, Providence, RI, 1970. MR 0264447 [24] D. L. Goodwin and I. Kuprov, Auxiliary matrix formalism for interaction representation transformations, optimal control, and spin relaxation theories, The Journal of Chemical Physics 143 (2015), 084113 -- 1 -- 084113 -- 7. [25] Jin Kyu Han, Hong Youl Lee, and Woo Young Lee, Invertible completions of 2 × 2 upper triangular operator matrices, Proc. Amer. Math. Soc. 128 (2000), no. 1, 119 -- 123. MR 1618686 [26] Ph. Hartman, Ordinary differential equations, second ed., Classics in Applied Mathematics, Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 2002. MR 1929104 [27] D. Henry, Geometric theory of semilinear parabolic equations, Lecture Notes in Mathematics, vol. 840, Springer-Verlag, Berlin -- New York, 1981. MR 610244 [28] N. J. Higham, A survey of condition number estimation for triangular matrices, SIAM Rev. 29 (1987), no. 4, 575 -- 596. MR 917696 [29] E. Hille and R. S. Phillips, Functional analysis and semi-groups, American Mathematical Society Colloquium Publications, vol. 31, Amer. Math. Soc., Providence, Rhode Island, 1957. MR 0089373 [30] Ch. Jordan, Calculus of finite differences, third ed., Chelsea Publishing Co., New York, 1965. MR 0183987 [31] R. V. Kadison and I. M. Singer, Triangular operator algebras. Fundamentals and hyperreducible theory, Amer. J. Math. 82 (1960), 227 -- 259. MR 0121675 [32] C. S. Kenney and A. J. Laub, A Schur-Fr´echet algorithm for computing the logarithm and exponential of a matrix, SIAM J. Matrix Anal. Appl. 19 (1998), no. 3, 640 -- 663 (electronic). MR 1611163 [33] D. Kressner, R. Luce, and F. Statti, Incremental computation of block triangular matrix exponentials with application to option pricing, arXiv:1703.00182 (2017). [34] V. G. Kurbatov, Linear functional-differential equations of neutral type and the retarded spectrum, Sibirskii Matematicheskii Zhurnal 16 (1975), no. 3, 538 -- 550, (in Russian); English translation in Siberian Mathemat- ical Journal, 16 (1975), no. 3, 412 -- 422. MR 0402226 [35] [36] , On the stability of functional-diffferential equations, Differencial′nye Uravnenija 17 (1981), no. 6, 963 -- 972, (in Russian); English translation in Differential Equations, 17 (1981), no. 6, 611 -- 618. MR 620094 , Functional differential operators and equations, Mathematics and its Applications, vol. 473, Kluwer Academic Publishers, Dordrecht, 1999. MR 1702280 [37] V. G. Kurbatov and I. V. Kurbatova, Computation of a function of a matrix with close eigenvalues by means of the Newton interpolating polynomial, Linear and Multilinear Algebra 64 (2016), no. 2, 111 -- 122 (English). MR 3434507 [38] , Computation of Green's function of the bounded solutions problem, Computational Methods in Ap- plied Mathematics (Published Online 2017-10-12, DOI: 10.1515/cmam-2017-0042). [39] V. G. Kurbatov, I. V. Kurbatova, and M. N. Oreshina, Analytic functional calculus for two operators, arXiv:1604.07393v (2016). [40] W. R. LePage, Complex variables and the Laplace transform for engineers, Dover Publications, Inc., New York, 1980, Corrected reprint of the 1961 original. MR 616824 [41] M. Lutzky, Parameter differentiation of exponential operators and the Baker -- Campbell -- Hausdorff formula, J. Mathematical Phys. 9 (1968), 1125 -- 1128. [42] J. L. Massera and J. J. Schaffer, Linear differential equations and function spaces, Pure and Applied Math- ematics, Vol. 21, Academic Press, New York -- London, 1966. MR 0212324 [43] I. Najfeld and T. F. Havel, Derivatives of the matrix exponential and their computation, Adv. in Appl. Math. 16 (1995), no. 3, 321 -- 375. MR 1342832 [44] A. A. Pankov, Bounded and almost periodic solutions of nonlinear operator differential equations, Mathemat- ics and its Applications (Soviet Series), vol. 55, Kluwer Academic Publishers, Dordrecht, 1990, Translated from the 1985 Russian edition. MR 1120781 [45] B. N. Parlett, A recurrence among the elements of functions of triangular matrices, Linear Algebra and Appl. 14 (1976), no. 2, 117 -- 121. MR 0448846 [46] A. V. Pechkurov, Bisectorial operator pencils and the problem of bounded solutions, Izv. Vyssh. Uchebn. Zaved. Mat. (2012), no. 3, 31 -- 41, (in Russian); English translation in Russian Math. (Iz. VUZ), 56 (2012), no. 3, 26 -- 35. MR 3076516 [47] P. C. Rosenbloom, Bounds on functions of matrices, Amer. Math. Monthly 74 (1967), 920 -- 926. MR 0222102 [48] W. Rudin, Functional analysis, first ed., McGraw-Hill Series in Higher Mathematics, McGraw-Hill Book Co., New York -- Dusseldorf -- Johannesburg, 1973. MR 0365062 [49] B. van der Pol and H. Bremmer, Operational calculus. Based on the two-sided Laplace integral, third ed., Chelsea Publishing Co., New York, 1987. MR 904873 [50] Ch. F. Van Loan, The sensitivity of the matrix exponential, SIAM J. Numer. Anal. 14 (1977), no. 6, 971 -- 981. MR 0468137 18 [51] V.G. KURBATOV AND I.V. KURBATOVA , Computing integrals involving the matrix exponential, IEEE Trans. Automat. Control 23 (1978), no. 3, 395 -- 404. MR 0494865 [52] R. M. Wilcox, Exponential operators and parameter differentiation in quantum physics, J. Mathematical Phys. 8 (1967), 962 -- 982. MR 0234689 [53] J. C. Willems, The analysis of feedback systems, The MIT Press, Cambridge, 1971. E-mail address: [email protected] Department of Mathematical Physics, Voronezh State University, 1, Universitetskaya Square, Voronezh 394018, Russia E-mail address: la [email protected] Department of Software Development and Information Systems Administration, Voronezh State University, 1, Universitetskaya Square, Voronezh 394018, Russia
1606.02125	1	1606	2016-06-07T12:57:25	Uncertainty Principles of Ingham and Paley-Wiener on Semisimple Lie Groups	[ "math.FA" ]	Classical results due to Ingham and Paley-Wiener characterize the existence of nonzero functions supported on certain subsets of the real line in terms of the pointwise decay of the Fourier transforms. Viewing these results as uncertainty principles for Fourier transforms, we prove certain analogues of these results on connected, noncompact, semisimple Lie groups with finite center. We also use these results to show unique continuation property of solutions to the initial value problem for time-dependent Schr\"odinger equations on Riemmanian symmetric spaces of noncompact type.	math.FA	math	UNCERTAINTY PRINCIPLES OF INGHAM AND PALEY-WIENER ON SEMISIMPLE LIE GROUPS MITHUN BHOWMIK AND SUPARNA SEN Abstract. Classical results due to Ingham and Paley-Wiener characterize the existence of nonzero functions supported on certain subsets of the real line in terms of the pointwise decay of the Fourier transforms. Viewing these results as uncertainty principles for Fourier transforms, we prove certain analogues of these results on connected, noncompact, semisimple Lie groups with finite center. We also use these results to show unique continuation property of solutions to the initial value problem for time-dependent Schrodinger equations on Riemmanian symmetric spaces of noncompact type. 1. Introduction It is a well known fact in harmonic analysis that if the Fourier transform of an integrable function on R is very rapidly decreasing then the function can not be compactly supported unless it vanishes identically. A manifestation of this fact is as follows: Let f ∈ L1(R) and a > 0 be such that bf (ξ) ≤ Ce−aξ, for all ξ ∈ R. If f is compactly supported then f is identically zero. This holds due to the fact that the very rapid decay of the Fourier transform imposes real analyticity on the function. In fact, f extends to a holomorphic function on an open subset of C. This initial observation motivates one to endeavour for a more optimal decay of the Fourier transform bf for such a conclusion. For instance we may ask: if bf decays faster than 1/(1 + · )n for all n ∈ N but slower than the function e−a., can f be compactly supported without being identically zero? A more precise question could be: is there a nonzero integrable compactly supported function f on R with its Fourier transform satisfying (1.1) log ξ , for large ξ? bf (ξ) ≤ Ce− ξ The answer to the above question is in the negative and follows from classical results due to Paley-Wiener ([28], Theorem II; [27], P. 16, Theorem XII) and Ingham [21] (see also Theorem 2010 Mathematics Subject Classification. Primary 22E30; Secondary 22E46, 43A80. Key words and phrases. uncertainty principle, semisimple Lie group, symmetric space, Schrodinger equation. This work was supported by Indian Statistical Institute, India (Research fellowship to Mithun Bhowmik) and Department of Science and Technology, India (INSPIRE Faculty Award to Suparna Sen). 1 2 MITHUN BHOWMIK AND SUPARNA SEN 1.1). These results may be viewed as instances of uncertainty principles in harmonic analysis. We refer the reader to [14, 16, 35] for the literature on uncertainty principles. Despite being results of same genre the treatment of Paley-Wiener and Ingham mentioned above have intriguing differences: while Paley-Wiener's method is complex analytic, Ingham relies on the notion of quasianalytic functions and the celebrated Denjoy-Carleman theorem ([30], Theorem 19.11). We would like to point out that the result of Paley and Wiener leads to a proof of the Denjoy-Carleman theorem. In some sense the result of Ingham is stronger than that of Paley and Wiener. In fact, a close examination of the proof of Ingham's result (as given in [21]) reveals that if bf satisfies (1.1) and f vanishes on an open set then f actually vanishes identically. Different versions of the result of Ingham was later proved by Levinson and Beurling [24, 25, 23]. All these results mentioned above are obtained only for the circle group and real line. Only very recently we have obtained the following analogues of the results of Paley-Wiener and Ingham in the context of n-dimensional Euclidean spaces (see Theorem 2.3 and Theorem 2.2 of [2] respectively). For f ∈ L1(Rn), we shall define its Fourier transform bf by f (x) e−ix·ξ dx, for ξ ∈ Rn. Theorem 1.1. Let ψ : [0, ∞) → [0, ∞) be a locally integrable function and f ∈ L1(Rn) be such that (1.2) Let (1.3) bf (ξ) =ZRn bf (ξ) ≤ Ce−ψ(kξk), I =Z ∞ 0 for all ξ ∈ Rn. ψ(r) 1 + r2 dr. (a) If f ∈ Cc(Rn) and I = ∞ then f is identically zero on Rn. Conversely, if I < ∞ and ψ is non-decreasing then there exists a nonzero function f ∈ C ∞ c (Rn) satisfying (1.2). (b) Let ψ(r) = rθ(r), for r ∈ [0, ∞), where θ : [0, ∞) → [0, ∞) is a decreasing function with limr→∞ θ(r) = 0. If f vanishes on an open set in Rn and I = ∞ then f is zero. c (Rn) satisfying (1.2). Conversely, if I < ∞ then there exists a nonzero function f ∈ C ∞ Part (a) of the above theorem corresponds to the result of Paley-Wiener and part (b) corresponds to that of Ingham. Viewing Rn as a noncompact Riemannian symmetric space one can naturally ask about analogues of the above results for Riemannian symmetric spaces of noncompact type or more generally for connected, noncompact, semisimple Lie groups with finite centre. In these cases nice parametrization of relevant representations are available and consequently the natural domain of the Fourier transform turns out to be R or Rn. This enables us to formulate analogous questions for these spaces. In this paper we will first prove an analogue of the result of Paley-Wiener for connected, noncompact semisimple Lie groups with finite centre and arbitrary rank (Theorem UNCERTAINTY PRINCIPLES OF INGHAM AND PALEY-WIENER ON SEMISIMPLE LIE GROUPS 3 3.5). The main ingredient here is Lemma 3.3 whose proof was inspired by a related result of Hirschmann [20]. However, an analogue of Ingham's result poses much greater difficulty. This is due to the unavailability of a suitable analogue of the Denjoy-Carleman theorem for Riemannian manifolds. Certain analogues of the Denjoy-Carleman theorem for Riemannian manifolds were proved by Bochner and Taylor [4, 5] but these results don't seem to be suitable for our purpose. We therefore restrict ourselves to bi-K-invariant functions on complex semisimple Lie groups and obtain an analogue of Ingham's theorem (Theorem 3.9). The main idea here is to use the explicit description of elementary spherical functions and reduce matters to Euclidean spaces. Our next set of results deal with Riemannian symmetric spaces of noncompact type. Our aim here is to relate the results of Paley-Wiener and Ingham to the problem of unique continuation of solution to the Schrodinger equation in the spirit of [12, 13, 7, 29]. These results present some pleasant surprise in the sense that the expected results demand more decay than that is required in Theorem 3.5 and Theorem 3.9 (see the beginning of §4.2.1 and Remark 4.4). This can be attributed to the exponential volume growth of the invariant measure on the symmetric space. Taking this into account we could obtain unique continuation properties of the solution to the Schrodinger equation on symmetric spaces (Theorem 4.3, Theorem 4.5). This paper is organized as follows: In the next section we describe the required preliminaries on connected, noncompact, real semisimple Lie groups with finite centre and the associated symmetric spaces. In Section 3 we shall prove an analogue of Theorem 1.1 (a) for f ∈ Cc(G) where G is a connected, noncompact, real semisimple Lie group with finite centre. We also prove an analogue of Theorem 1.1 (b) on a noncompact, complex semisimple Lie group. In the last section we consider unique continuation properties of solutions to the initial value problem for time-dependent Schrodinger equation on Riemannian symmetric spaces of the noncompact type. First we prove such a result for bi-K-invariant initial value on noncompact, complex semisimple Lie groups related to Theorem 1.1 (b). We also prove such a result on Riemannian symmetric spaces of the noncompact type related to Theorem 1.1 (a). 2. Notation and Preliminaries In this section, we shall discuss the preliminaries and notation related to the noncompact semisimple Lie groups and the associated symmetric spaces. These are standard and can be found, for example, in [15, 17, 18, 19, 22]. To make the article self-contained, we shall gather only those results which will be used throughout this paper. We shall use the following notation in this paper: Cc(X) denotes the set of compactly sup- ported continuous functions on X, C ∞ c (X) denotes the set of compactly supported smooth functions on X and C denotes a constant whose value may vary. For x, y ∈ Rn, we shall use kxk to denote the Euclidean norm of the vector x and x · y to denote the Euclidean inner product 4 MITHUN BHOWMIK AND SUPARNA SEN of the vectors x and y. We shall also denote the Euclidean norm in Cn by k · k. ℑ(λ) denotes the imaginary part of λ ∈ Cn. Let G be a connected, noncompact, real semisimple Lie group with finite centre and K be a fixed maximal compact subgroup of G. Let g and k denote the Lie algebras of G and K respectively. Suppose that B is the Cartan Killing form on g and g = k ⊕ p is a Cartan decomposition of g. It is known that B restricted to p is positive definite, thus it gives an inner product and hence a norm k · kB on p. Fix a maximal abelian subspace a of p. If the dimension of a is l, then G is said to be of real rank l. We can identify a with Rl endowed with the inner product induced from p. Let R denote the set of nonzero roots for the adjoint action of a on g and W denote the Weyl group corresponding to R. Fix a Weyl chamber a+ of a and let R+ be the corresponding set of positive roots. Let A = exp a and A+ = exp a+. If A+ denotes the closure of A+ in G, then one gets the polar decomposition G = KA+K, that is, each g ∈ G can be uniquely written as g = k1ak2, for k1, k2 ∈ K and a ∈ A+. Using the polar decomposition k · kB on G is defined by kgkB = kk1ak2kB = k log akB, where log a denotes the unique element in a such that exp(log a) = a. Let gα denote the root space corresponding to α ∈ R with mα = dim gα. The Haar measure dg on G relative to the polar decomposition is given by dg = J(a) dk1 da dk2, that is, for any suitable function f on G, we have f (k1ak2) J(a) dk1 da dk2, f (g)dg =ZKZA+ZK ZG J(a) = Yα∈R+(cid:16)eα(log a) − e−α(log a)(cid:17)mα where dk is the normalised Haar measure on K, da is the Lebesgue measure on A ∼= Rl and , for a ∈ A+. We define the half-sum of the elements of R+ counted with their multiplicities by ρ given by ρ = 1 2 Xα∈R+ mαα. Thus we get the trivial estimate (2.1) J(a) ≤ Ce2ρ(log a), for a ∈ A+. For an element λ in the dual a∗ of a, let Hλ be the unique element in a such that λ(H) = B(H, Hλ), for all H ∈ a. UNCERTAINTY PRINCIPLES OF INGHAM AND PALEY-WIENER ON SEMISIMPLE LIE GROUPS 5 The map λ 7→ Hλ identifies a∗ with a, and we use it to define the dual inner product on a∗, also denoted B(·, ·), by the formula B(λ, µ) = B(Hλ, Hµ), for λ, µ ∈ a∗. The elements of the Weyl group W are orthogonal transformations of a∗ which correspond to orthogonal transformations of a by the formula (2.2) sHλ = Hsλ, for s ∈ W and λ ∈ a∗. The bilinear extension of B(·, ·) to a∗ by B(·, ·). C, the dual of the complexification aC of a, is also denoted We now describe the principal series representations of G. The Iwasawa decomposition G = KAN gives rise to the projection mappings κ : G → K, H : G → a and η : G → N such that g = κ(g) exp H(g)η(g). dimensional Hilbert space on which ξ is realised. We define the Hilbert space Hξ by Let M be the centraliser of A in K. Given ξ in the unitary dual cM of M let H be the finite kφ(k)k2dk < ∞(cid:27) , (cid:26)φ : K → H measurable, φ(km) = ξ(m−1)φ(k) for k ∈ K, m ∈ M and ZK where k · k denotes the norm on H induced from the inner product h·, ·i on H. The Hilbert space Hξ is equipped with the inner product hφ, ψiξ =ZK hφ(k), ψ(k)i dk. C we have a representation πξ,λ acting on the Hilbert space Hξ given by It is known that πξ,λ is unitary if λ ∈ a∗ and πξ,λ, πξ,µ are unitarily equivalent if and only if (πξ,λ(g)φ) (k) = e(iλ−ρ)H(g−1k)φ(cid:0)κ(g−1k)(cid:1) , For ξ ∈ cM and λ ∈ a∗ λ = sµ where s ∈ W . Moreover, given ξ ∈ cM there exists a dense open subset Oξ ⊂ a∗ such that for λ ∈ Oξ, πξ,λ is irreducible. We now define the group Fourier transform of f ∈ L1(G) given by the operator valued integral for g ∈ G, k ∈ K, φ ∈ Hξ. πξ,λ(f ) =ZG f (g)πξ,λ(g)dg, for (ξ, λ) ∈ cM × a∗. For f ∈ L1∩L2(G), πξ,λ(f ) is a Hilbert-Schmidt operator and we shall denote its Hilbert-Schmidt norm by kπξ,λ(f )kHS. If ξ is the trivial representation in cM , then we denote πξ,λ by πλ and the set of representations {πλ}λ∈a∗ , realized on the Hilbert space L2(K/M ), are called the class one principal series repre- sentations of G. We observe that πλK are given by left translations on L2(K/M ), in particular, the K-fixed vectors are given by constant functions. The right-K-invariant functions on G can be viewed as functions on the symmetric space X = G/K and vice-versa. For the harmonic 6 MITHUN BHOWMIK AND SUPARNA SEN analysis of a function f on X, only the class one principal series representations πλ are relevant and πλ(f ) is completely determined by πλ(f )e0 where e0 denotes the constant function 1 on K/M . In this case, the group theoretic Fourier transform can be reinterpreted as the Helgason- Fourier transform on the symmetric space X, as introduced by Helgason. For a sufficiently nice C × K/M given by function f on X, its Helgason-Fourier transform ef is a function defined on a∗ f (g)e(iλ−ρ)H(g−1k)dg, (2.3) C, kM ∈ K/M whenever this integral exists (see [18], Ch. III, §1). Thus it follows that for λ ∈ a∗ (2.4) ZK/M (πλ(f )e0)(kM ) = ef (λ, kM ) =ZG ef (λ, kM )2dk = kπλ(f )e0k2 +ZK/M ZX f (x)2dx =Za∗ The Plancherel theorem states that the Helgason Fourier transform extends to an isometry of L2(X) onto L2(a∗ + × K/M, c(λ)−2dk dλ) where c(λ) is Harish-Chandra's c-function, that is, L2(K/M ) = kπλ(f )k2 HS , for λ ∈ a∗. ef (λ, kM )2c(λ)−2dk dλ. Using the polar decomposition of G we may view a bi-K-invariant function f on G as that depending only on its values on A+, or by using the inverse exponential map we may also view f as a function on a solely determined by its values on a+. Henceforth we shall denote the set of bi-K-invariant functions in L1(G) by L1(K\G/K). If f ∈ L1(K\G/K) then πλ(f ) is determined by hπλ(f )e0, e0i and the spherical transform ef (λ) of f is defined by f (g)φλ(g)dg = hπλ(f )e0, e0i, for λ ∈ a∗, where φλ is the elementary spherical function corresponding to λ ∈ a∗ given by (2.5) e(iλ−ρ)H(g−1k)dk, for g ∈ G. ef (λ) =ZG φλ(g) = hπλ(g)e0, e0i =ZK It is clear that the elementary spherical function φλ is a bi-K-invariant function on G for any λ ∈ a∗. The following estimates regarding φλ are well known (see [15], §4.6). (2.6) e−ρ(log a) ≤ φ0(a) ≤ C(1 + k log akB)me−ρ(log a), for a ∈ A+, some m, C > 0, (2.7) 0 < φiλ(a) ≤ eλ+(log a)φ0(a), for a ∈ A+, λ ∈ a∗, where λ+ is the element in the fundamental Weyl chamber corresponding to λ. In this case (2.8) kπλ(f )kHS = hπλ(f )e0, e0i = ef (λ), for λ ∈ a∗. Apart from Fourier transform, we will also need the notion of Radon transform on Riemannian symmetric space X. In the following, we shall describe the properties of Radon transform which UNCERTAINTY PRINCIPLES OF INGHAM AND PALEY-WIENER ON SEMISIMPLE LIE GROUPS 7 will be used in this paper. By identifying the space G/M N of horocycles on X with (K/M ) × A, we define the Radon transform of a sufficiently regular function f : X → C by (2.9) Rf (kM, a) = eρ(log a)ZN f (kan · o) dn for all kM ∈ K/M , a ∈ A whenever this integral exists (see [18], P. 220, see also [31]) and o = eK denotes the identity coset in X. It is known that the Helgason-Fourier transform of a sufficiently regular function f is the Euclidean Fourier transform of the Radon transform (see [18], P. 219) given by (2.10) ef (λ, kM ) = FA(Rf (kM, ·))(λ), where FA denotes the Euclidean Fourier transform of h ∈ L1(A) defined by for almost all kM ∈ K/M and all λ ∈ a∗, (FAh)(λ) =ZA h(a)e−iλ(log a)da, for λ ∈ a∗ ∼= Rl. It is well known that Radon transform is injective on L1(X) (see [18], Ch. II, Theorem 3.2). For f ∈ L1(K\G/K), we define the Abel transform by the following integral Af (a) = eρ(log a)ZN f (an · o) dn, for a ∈ A, (see [18], P. 381). Note that the Radon transform of a bi-K-invariant function reduces to the Abel transform. It follows from (2.10) that the spherical transform of f ∈ L1(K\G/K) is the Euclidean Fourier transform of the Abel transform Af given by (2.11) ef (λ) = FA(Af )(λ), for λ ∈ a∗ ∼= Rl. This property of Abel transform is crucial for reducing some questions on bi-K-invariant func- tions defined on a semisimple Lie group G to related questions on Rl. Moreover, it is known that the Abel transform A induces a bijection between C ∞ c (K\G/K) and the set of W -invariant functions in C ∞ c (A) (see [18], Ch. IV, Theorem 4.1). 3. Uncertainty Principles In this section we shall present analogues of the results of Paley-Wiener and Ingham, namely Theorem 1.1 (a) and Theorem 1.1 (b) on connected, noncompact, real semisimple Lie groups with finite centre and noncompact, complex semisimple Lie groups respectively. 3.1. Real Semisimple Lie Groups. To prove the first theorem we shall prove a lemma on entire functions on Cn. The proof of this lemma for the one-variable case is given in [3]. Since the paper is yet to be accepted, we reproduce the proof here for the sake of completeness. To prove this lemma, we shall need two results. The first one is regarding upper semicontinuous functions. 8 MITHUN BHOWMIK AND SUPARNA SEN Theorem 3.1 ([8], P. 218, Theorem 3.6). Let (X, d) be a metric space, v : X → [−∞, ∞) be upper semi-continuous and v ≤ M < ∞ on X. Then there exists a decreasing sequence of uniformly continuous functions {fn} on X such that fn ≤ M and for every x ∈ X, fn(x) decreases to v(x). The second result we need is an analogue of the maximum modulus principle on unbounded domain for subharmonic functions. We briefly recall the definition of subharmonic functions. Let D be an open subset of C. A function u : D → [−∞, ∞) is called subharmonic if u is upper semi-continuous and satisfies the local submean inequality, that is, for any w ∈ D there exists ρ > 0 such that for all r ∈ (0, ρ) the following holds u(w) ≤ 1 2πZ 2π 0 u(w + reit)dt. It is well known that if f is a holomorphic function then g(z) = log(f (z)) is a subharmonic function (see [30], P. 336). Theorem 3.2 ([6], P. 224, Theorem 7.15). Let Ω be a region (not necessarily bounded) and u : Ω → R be a subharmonic function which is bounded above. Let A be a proper, countable subset of the boundary ∂Ω of Ω and M a finite constant such that lim u(z) ≤ M for all ξ ∈ ∂Ω r A. z→ξ Then u ≤ M throughout Ω. We shall now present the required lemma on entire functions. In the following we shall interpret a radial function on Rn as an even function on R. Lemma 3.3. Let f be an entire function on Cn and ψ be a non-negative, radial, locally integrable function on Rn such that for positive constants a and C, we have (3.1) (3.2) If f (z) ≤ Ceakzk, f (x) ≤ Ce−ψ(x), for all z ∈ Cn, for all x ∈ Rn. ZR ψ(r) 1 + r2 dr = ∞, then f is identically zero on Cn. Proof. First we shall prove the result for n = 1. We consider the function g(z) = 1 C eiazf (z), for all z ∈ C, and observe that g is an entire function. We want to apply Phragm´en-Lindelof theorem to show that for all z in the closed upper half plane H = {z ∈ C : ℑz ≥ 0} (3.3) g(z) ≤ 1, UNCERTAINTY PRINCIPLES OF INGHAM AND PALEY-WIENER ON SEMISIMPLE LIE GROUPS 9 (see [32], P. 124, Theorem 3.4). Let Q1 = {z = x + iy ∈ C : x > 0, y > 0}. It follows from the estimate (3.1) that g(iy) = 1 C e−ayf (iy) ≤ e−ayeay = 1, for all y > 0. It is also immediate from (3.2) that g(x) = 1 C f (x) ≤ e−ψ(x) ≤ 1, for x ∈ R. In particular, g is bounded by 1 on the positive real and positive imaginary axes. As g satisfies the estimate (3.1) we can apply the Phragm´en-Lindelof theorem to the sector Q1 to obtain (3.3). A similar argument for the quadrant Q2 = {z = x + iy ∈ C : x < 0, y > 0} proves the estimate (3.3) for all z ∈ H. Since g is an entire function, log g is a subharmonic function on C and (3.4) log g(z) ≤ 0, for all z ∈ H. Now, we apply Theorem 3.1 for X = H, v = log g, and M = 0. Then there exists a decreasing sequence of uniformly continuous functions fn on H such that fn ≤ 0 and fn(z) decreases to v(z) for every z ∈ H. We define vn(z) = max{fn(z), −n}, for z ∈ H, n ∈ N. It is clear that {vn} is a decreasing sequence of continuous functions. As fn takes only negative values it follows that vn(z) ∈ [−n, 0] for all z ∈ H. In particular, vn is bounded for each n ∈ N. We now claim that {vn} converges pointwise to v on H. We first assume that v(z) = −∞. Then {fn(z)} and {−n} both converge to −∞ and hence so does {vn(z)}. Now assume that v(z) is finite. In this case fn(z) ∈ (v(z) − 1, v(z) + 1) for all large n and hence vn(z) = fn(z) for all large n ∈ N. It follows that {vn(z)} converges to v(z) for all z ∈ H. Let Un be the Poisson integral of the restriction of the function vn on R given by Since vn ∈ L∞(R), the above integral exists and defines a harmonic function on the open upper half plane H = {z ∈ C : ℑz > 0}. Moreover, since vn is continuous, we can extend Un to H as a continuous function by letting Un(x) = vn(x) for x ∈ R ([33], P. 47, Theorem 2.1(b)). We now define (3.5) Un(x + iy) = yvn(t) y2 + (x − t)2 dt, for x ∈ R, y > 0. 1 πZR (3.6) Vn(z) = log g(z) − Un(z), for z ∈ H. Since Un is harmonic, we get that Vn is subharmonic on H. Since vn(t) ≥ −n, for all t ∈ R, it follows from the definition of Un given in (3.5) that Un(z) ≥ −n, for z ∈ H. 10 MITHUN BHOWMIK AND SUPARNA SEN Using (3.4) we get that In particular, Vn is bounded above for each n ∈ N. Since Vn(z) ≤ n, for all z ∈ H. v(z) = log g(z) ≤ vn(z), for all z ∈ H, it follows that lim y→0 Vn(x + iy) = log g(x) − vn(x) ≤ 0, for all x ∈ R. We now apply Theorem 3.2 for Ω = H, u = Vn and A = φ, the empty set to conclude that Vn(z) ≤ 0, for all z ∈ H. It follows from (3.6) that log g(x + iy) ≤ 1 πZR yvn(t) y2 + (x − t)2 dt, for all y > 0, x ∈ R. Since {vn} is a decreasing sequence, by using monotone convergence theorem and taking limit as n → ∞ in the inequality above we get log g(x + iy) ≤ 1 πZR y log g(t) y2 + (x − t)2 dt. The estimate (3.2) now implies that log g(x + iy) ≤ − 1 πZR yψ(t) y2 + (x − t)2 dt ≤ −Cx,yZR ψ(t) 1 + t2 dt = −∞, where Cx,y is a positive constant which depends on x and y. So, for each x ∈ R and y positive, it follows that g(x + iy) = 0. As f is an entire function it follows that f (z) = 0 for all z ∈ C. This proves the lemma for n = 1. Now, we shall prove the case n > 1. Let ξ ∈ Rn−1 be fixed and we define g(z) = f (ξ, z), for z ∈ C, which is clearly an entire function on C. Using (3.1) we have for any z ∈ C g(z) = f (ξ, z) ≤ Ceak(ξ,z)k = Ceak(ξ,0)+(0,z)k ≤ Cξeaz. Using (3.2) we get that for any x ∈ R g(x) = f (ξ, x) ≤ e−ψ(ξ,x) = e−ψξ(x), where the function ψξ on R is defined as ψξ(x) = ψ(ξ, x), for x ∈ R. UNCERTAINTY PRINCIPLES OF INGHAM AND PALEY-WIENER ON SEMISIMPLE LIE GROUPS 11 Since ψ is radial, ψξ is an even function on R. Moreover using the radiality of ψ and the change of variable kξk2 + x2 = r2 we have ZR ψξ(x) 1 + x2 dx = ZR = 2Z ∞ ≥ 2Z ∞ kξk kξk = ∞. ψ(pkξk2 + x2) 1 + x2 dx ψ(r) 1 + r2 − kξk2 ψ(r) 1 + r2 dr dr r pr2 − kξk2 Applying the lemma for the case n = 1 on g, we get that g is zero on C. So it follows that f (ξ, z) is zero for all z ∈ C. Since this is true for all ξ ∈ Rn−1 we obtain that f is identically zero on Rn. Since f is an entire function on Cn which vanishes on Rn, we can conclude that f vanishes on Cn. (cid:3) (3.7) We shall need few other facts to prove our first main theorem. Let G be a connected, non- compact, semisimple Lie group with finite centre. We shall continue to assume the notation j : j ∈ N} be an orthonormal basis of Hξ introduced in the previous section. For ξ ∈ cM let {eξ consisting of K-finite vectors. For g ∈ G, λ ∈ a∗ and m, n ∈ N we define Φm,n ξ,λ (g) =Dπξ,λ(g)eξ m, eξ nEξ =ZK e(iλ−ρ)H(g−1k)Deξ m(cid:0)κ(g−1k)(cid:1) , eξ n(k)E dk. m, eξ The basis vectors eξ n ∈ Hξ being K-finite, actually belong to C ∞(K, H) and hence are bounded as functions into H. Therefore it follows that for each g ∈ G the integral defining Φm,n ξ,λ (g) is continuous and the function λ 7→ Φm,n ∼= Cl. From (3.7) and (2.5) we get the following easy estimate ξ,λ (g) extends as an entire function of λ ∈ a∗ C. Moreover the function g 7→ Φm,n ξ,λ (g) makes sense for λ ∈ a∗ C Using the bi-K-invariance of φiℑ(λ) and (2.7) we get that e−(ℑ(λ)+ρ)H(g−1k)dk = Cφiℑ(λ)(g), ξ,λ (g)(cid:12)(cid:12)(cid:12) ≤ CZK (cid:12)(cid:12)(cid:12)Φm,n (cid:12)(cid:12)(cid:12)Φm,n ξ,λ (g)(cid:12)(cid:12)(cid:12) ≤ Ceℑ(λ)+(log a)φ0(a) ≤ Cekℑ(λ)kB k log akB φ0(a), for g ∈ G. for g ∈ G, where g = k1ak2; k1, k2 ∈ k and a ∈ A+. (3.8) We have already defined πξ,λ for all λ ∈ a∗ C and noted that πξ,λ is a unitary operator on the Hilbert space Hξ for λ ∈ a∗. In general for λ ∈ a∗ C \ a∗, the representation πξ,λ may not be unitary. However for f ∈ Cc(G), it is easy to see that πξ,λ(f ) makes sense for all λ ∈ a∗ C and ξ ∈ cM . We shall also need the following lemma which is essentially proved in [10] using Harish-Chandra's subquotient theorem. 12 MITHUN BHOWMIK AND SUPARNA SEN Lemma 3.4. Let f ∈ Cc(G). If πξ,λ(f ) = 0 for all λ ∈ a∗ Now we shall present the first main result, which is an analogue of Theorem 1.1 (a). Theorem 3.5. Let ψ : [0, ∞) → [0, ∞) be a locally integrable function such that C and ξ ∈ cM then f is zero. I =Z ∞ 0 ψ(r) 1 + r2 dr. (a) Suppose f ∈ Cc(G) satisfies the estimate kπξ,λ(f )kHS ≤ Cξe−ψ(kλkB ), (3.9) for ξ ∈ cM , λ ∈ a∗ ∼= Rl, where Cξ is a positive constant depending on ξ. If I = ∞ then f is zero on G. (b) If I is finite and ψ is nondecreasing then there exists a nontrivial f ∈ C ∞ c (K\G/K) satisfying the estimate (3.9). Proof. First we shall prove (a). For fixed ξ ∈ cM and m, n ∈ N we define f (g) Φm,n ξ,λ (g)dg, F m,n ξ m, eξ (3.10) for λ ∈ a∗. (λ) =Dπξ,λ(f )eξ nEξ =ZG C, it follows that F m,n Since f ∈ Cc(G) and Φm,n of λ ∈ a∗ is enough to prove that for each fixed m, n ∈ N, F m,n decomposition, (3.10) and (3.8), it follows that ξ,λ (g) is a continuous function of g ∈ G as well as an entire function C. From Lemma 3.4, it C. Using polar extends as an entire function of λ ∈ a∗ (λ) is zero for all λ ∈ a∗ ξ ξ (3.11) F m,n ξ (λ) ≤ZKZa+ZK f (k1(exp H)k2) ekℑ(λ)kB kHkB φ0(exp H) J(exp H) dk1 dH dk2. Since f ∈ Cc(G), it follows from (3.11), (2.6) and (2.1) that there exists a constant γ > 0 such that (3.12) F m,n ξ (λ) ≤ Cf eγkℑ(λ)kB ≤ Cf eγkλkB , for all λ ∈ a∗ C ∼= Cl. Now by (3.9) and the fact that we get that (3.13) m, eξ (cid:12)(cid:12)(cid:12)hπξ,λ(f )eξ niξ(cid:12)(cid:12)(cid:12) ≤ kπξ,λ(f )kHS , for λ ∈ a∗, F m,n ξ (λ) ≤ Cξe−ψ(kλkB ), for all λ ∈ a∗ ∼= Rl. ξ Since all norms are equivalent on finite dimensional spaces, (3.12), (3.13) and Lemma 3.3 implies that F m,n vanishes on a∗ C 3.4 that f vanishes on G. ∼= Cl. By varying m, n over N and ξ over cM it now follows from Lemma Now we shall prove (b). Since I is finite and ψ is nondecreasing, by Theorem 1.1 (a) there c (Rl) satisfying (1.2). Since f0 is radial on Rl, it can c (K\G/K) such that exists a nontrivial radial function f0 ∈ C ∞ be thought of as a W -invariant function on A ∼= Rl. So there exists f ∈ C ∞ Af = f0. Using (2.8) and (2.11) it follows that f ∈ C ∞ c (K\G/K) satisfies the estimate (3.9). UNCERTAINTY PRINCIPLES OF INGHAM AND PALEY-WIENER ON SEMISIMPLE LIE GROUPS 13 (cid:3) Remark 3.6. If we think of a function on X = G/K as a right-K-invariant function on G then the following theorem is, in view of the relation (2.4), an easy corollary of Theorem 3.5. Theorem 3.7. Let ψ and I be as in Theorem 3.5. (a) Let f ∈ Cc(X) and the Helgason Fourier transform ef of f satisfies the estimate for λ ∈ a∗, kM ∈ K/M. (3.14) If I = ∞ then f = 0. ef (λ, kM ) ≤ Ce−ψ(kλkB ), (b) If I is finite and ψ is nondecreasing then there exists a nontrivial K-invariant f ∈ C ∞ c (X) satisfying the estimate (3.14). 3.2. Complex Semisimple Lie Groups. We shall prove an analogue of Theorem 1.1 (b) for bi-K-invariant functions on a noncompact, complex semisimple Lie group using the explicit expression of the elementary spherical functions φλ available in this case. First we shall recall the following result proved in [2] which is needed in our proof. Lemma 3.8. Let P be a polynomial on Rn and θ : Rn → [0, ∞) be a decreasing radial function with limkξk→∞ θ(ξ) = 0 and (a) Let f ∈ L1(Rn) be a nontrivial function satisfying the estimate θ(ξ) kξkn dξ. I =Zkξk≥1 bf (ξ) ≤ CP (ξ)e−kξkθ(ξ), for all ξ ∈ Rn. If f vanishes on a nonempty open set then I is finite. (b) If I is finite then there exists a nontrivial f ∈ C ∞ c (Rn) satisfying (3.15). If G is noncompact, complex semisimple, it is known that the Haar measure on A correspond- ing to the polar decomposition is given by φ(H)2dH for H ∈ a where φ(H) is defined by the formula (3.15) (3.16) (3.17) φ(H) = Xs∈W det(s) esρ(H), for H ∈ a, (see [9], P. 907 and P. 910). It is also known that on a noncompact, complex semisimple Lie group, the elementary spherical functions are given by the expression φλ(H) = c(λ)Ps∈W det(s)e−isλ(H) φ(H) , for H ∈ a, λ ∈ a∗, and the function λ 7→ c(λ)−1, for λ ∈ a∗, is of polynomial growth ([19], P. 432, Theorem 5.7). Thus for f ∈ L1(K\G/K) the spherical transform ef of f is given by the integral f (H)φλ(H)φ(H)2dH, for λ ∈ a∗. (3.18) ef (λ) =Za 14 MITHUN BHOWMIK AND SUPARNA SEN Theorem 3.9. Let G be a connected, noncompact, complex semisimple Lie group and θ : [0, ∞) → [0, ∞) be a decreasing function with limr→∞ θ(r) = 0 and (a) Let f ∈ L1(K\G/K) satisfy the estimate I =Z ∞ 1 θ(r) r dr. (3.19) ef (λ) ≤ Ce−kλkB θ(kλkB ), for λ ∈ a∗ ∼= Rl. If f vanishes on a nonempty open set in G and I = ∞ then f = 0. (b) If I is finite then there exists a nontrivial f ∈ C ∞ c (K\G/K) satisfying the estimate (3.19). Proof. First we shall prove (a). Using (3.17) and (2.2) the spherical transform of f can be written as ef (λ) = c(λ)Xs∈W = c(λ)Xs∈W = c(λ)Xs∈W where the function g on a is defined as det(s)Za det(s)Za det sbg(sHλ), e−isλ(H)f (H)φ(H)dH e−iB(H,sHλ)f (H)φ(H)dH for λ ∈ a∗, g(H) = f (H)φ(H), for H ∈ a, andbg denotes the Euclidean Fourier transform of g. As f ∈ L1(K\G/K) it follows that Consequently bg is well defined. Since f is bi-K-invariant and φ is odd under the action of the g(H)dH =Za Weyl group, that is, f (H)φ(H)dH < ∞. Za φ(sH) = det(s)φ(H), for H ∈ a, s ∈ W, we get that g(sH) = det(s)g(H), for H ∈ a, s ∈ W. It follows thatbg is also odd under the action of the Weyl group, that is, for λ ∈ a∗, s ∈ W. Consequently the spherical transform of f can be written as bg(sHλ) = det(s)bg(Hλ), ef (λ) = c(λ)W bg(Hλ), for λ ∈ a∗, UNCERTAINTY PRINCIPLES OF INGHAM AND PALEY-WIENER ON SEMISIMPLE LIE GROUPS 15 where W denotes the number of elements in the Weyl group W . From the estimate of ef given in (3.19) it follows that bg(Hλ) ≤ Cc(λ)−1e−θ(kλkB )kλkB , for λ ∈ a∗ ∼= Rl. Since c(λ)−1 is of polynomial growth, using the equivalence of all norms on finite dimensional spaces we can apply Lemma 3.8 to the function g on a to get that g is zero. Hence it follows that f is zero. To prove (b), first we note that by Theorem 1.1 (b) we get a nontrivial radial function f0 ∈ c (Rl) satisfying (1.2). Then we proceed as in Theorem 3.5 (b) to construct f ∈ C ∞ c (K\G/K) (cid:3) C ∞ satisfying the estimate (3.19). 4. Unique Continuation Property of Solutions to the Schrodinger Equation We consider the initial value problem for the time-dependent Schrodinger equation on Rn given by ∂u ∂t (x, t) − i∆u(x, t) = 0, for (x, t) ∈ Rn × R, u(x, 0) = f (x), for x ∈ Rn. Our aim is to obtain sufficient conditions on the behaviour of the solution u at two different times t = 0 and t = t0 which guarantee that u ≡ 0 is the unique solution of the above equation. It has recently been observed that uncertainty principles can be used to obtain such sufficient conditions. We refer the reader to [13] and the references therein for results in this regard. These results were further generalized in the context of noncommutative groups in [7, 29, 1, 26, 2, 3]. In this section we wish to relate the theorems of Ingham and Paley-Wiener to the above mentioned problem in the context of symmetric spaces. We first deduce one such result for the damped Schrodinger Equation on Rn. 4.1. Euclidean Space. We consider the initial value problem for the time-dependent damped Schrodinger Equation on Rn given by  (4.1) ∂u ∂t (x, t) − i(∆ − c)u(x, t) = 0, u(x, 0) = f (x), for (x, t) ∈ Rn × R, for x ∈ Rn, where ∆ denotes the Laplacian on Rn, c ∈ R is the damping parameter and f ∈ L1(Rn). It follows that the unique solution ut(x) = u(x, t) of (4.1) is characterized via the Euclidean Fourier transform of ut given by the equation (4.2) but(ξ) = e−it(kξk2+c)bf (ξ), for ξ ∈ Rn. 16 MITHUN BHOWMIK AND SUPARNA SEN Theorem 4.1. Let u be the solution to the equation (4.1) given by (4.2), with initial value f ∈ Cc(Rn). Suppose there exists t0 > 0 such that (4.3) u(x, t0) ≤ Ce−ψ(kxk), for x ∈ Rn, where ψ : [0, ∞) → [0, ∞) is a locally integrable function. If Z ∞ 1 ψ(r) 1 + r2 dt = ∞, then u is zero. Proof. The solution ut(x) = u(x, t) of (4.1) is given by ut(x) = γc,t ∗ f (x), for x ∈ Rn, t ∈ R, where the kernel γc,t is defined by (4.4) γc,t(x) = (4πt)−n/2e−icte−iπsign(t)n/4eikxk2/4t, for x ∈ Rn, t ∈ R, so that its Fourier transform is cγc,t(ξ) = e−i(kξk2+c)t, for ξ ∈ Rn, t ∈ R, (see [29], P. 869). Here sign(t) = t/t denotes the sign of t. Using the expression for γc,t in (4.4) we can express the solution u(x, t0) at time t = t0 > 0 as u(x, t0) = (4πt0)−n/2e−ict0e−iπn/4ei kxk2 (4.5) 4t0 bh(cid:18) x 2t0(cid:19) where the function h on Rn is defined by (4.6) h(y) = ei kyk2 4t0 f (y), for y ∈ Rn. Using (4.3) and (4.5) and applying Theorem 1.1 (a) to the function h ∈ Cc(Rn) we get that h is zero. Hence f is zero and so is u. (cid:3) Remark 4.2. It is easy to see from the relations (4.5) and (4.6) that the result corresponding to Theorem 1.1 (b) can also be proved similarly if we assume the corresponding conditions on the function f and the solution u(·, t0). 4.2. Riemannian Symmetric Space. Let X = G/K be a Riemannian symmetric space of noncompact type where G is a noncompact, connected semisimple Lie group with finite center and K is a maximal compact subgroup of G. We have a G-invariant Riemannian metric on X induced by the Killing form B restricted to p and we can form a Laplace-Beltrami operator UNCERTAINTY PRINCIPLES OF INGHAM AND PALEY-WIENER ON SEMISIMPLE LIE GROUPS 17 ∆ using this metric (see [19]). We consider the initial value problem for the time-dependent Schrodinger equation on X given by (4.7) ∂u ∂t (x, t) − i∆u(x, t) = 0, for (x, t) ∈ X × R, u(x, 0) = f (x), for x ∈ X.  For g ∈ G, we define σ(g) = d(gK, K) where d is the canonical distance function on X = G/K coming from the Riemannian metric on X induced by the Killing form B. The function σ is bi-K-invariant and continuous. If g = k1 exp Hk2 with g ∈ G, H ∈ a+ and k1, k2 ∈ K, then σ(g) = σ(exp H) = kHkB. 4.2.1. Complex Semisimple Lie Groups. We shall first prove a unique continuation property of solution to the Schrodinger equation (4.7) in the context of complex semisimple Lie groups with bi-K-invariant initial data corresponding to Theorem 1.1 (b). In this case, the solution u(·, t) is also bi-K-invariant and can be considered as a function on a. It turns out that the situation here is slightly different from that of Euclidean spaces. The reason is roughly speaking the exponential growth of the G-invariant measure on G/K. Consequently, to prove this unique continuation property in the context of complex semisimple Lie groups we need to impose more decay on the solution u(x, t). We will also show that the unique continuation property does not hold in the absence of such extra decay (see Remark 4.4). Our method of proof will follow that in [7] and will use some of their notation and calculations. Theorem 4.3. Let θ : [0, ∞) → [0, ∞) be a decreasing function with limr→∞ θ(r) = 0 and u be a solution to the equation (4.7) with initial value f ∈ L1(K\G/K). Suppose there exists t0 > 0 such that (4.8) u(H, t0) ≤ C φ0(H) e−kHkB θ(kHkB ), for H ∈ a. If f vanishes on an open set in a ∼= Rl and (4.9) then u is identically zero. I =Z ∞ 1 θ(r) r dr = ∞ Proof. It is proved in [7] that the solution u(H, t) of (4.7) at t = t0 can be written as (4.10) u(H, t0)φ(H) = CW 2t−l/2 0 −i(cid:18)t0kρk2 B − e kHk2 B 4t0 (cid:19) where the function g on a is defined by bgf(cid:18) H 2t0(cid:19) , for H ∈ a, (4.11) gf (H) = ei kHk2 B 4t0 f (H)φ(H), for H ∈ a, 18 MITHUN BHOWMIK AND SUPARNA SEN where φ is given by (3.16). Since f vanishes on an open set in a ∼= Rl, g also vanishes on the same open set in a. From the estimate of the Jacobian J given in (2.1) we get that (4.12) φ(H) =pJ(exp H) ≤ Ceρ(H), for H ∈ a. It follows from (4.8), (4.10), (4.12) and (2.6) that for H ∈ a (4.13) 2t0(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) ≤ Cφ0(H)e−kHkB θ(kHkB )eρ(H) ≤ C(1 + kHkB)me−kHkB θ(kHkB ). (cid:12)(cid:12)(cid:12)(cid:12)bgf(cid:18) H As gf vanishes on an open set and bgf satisfies (4.13) it follows from Lemma 3.8 that gf is zero. Therefore f is zero and hence so is u. (cid:3) Remark 4.4. We give an example to show that if 0 ≤ α < 1 then u(H, t0) ≤ C φ0(H)α e−kHkB θ(kHkB ), for H ∈ a ∼= R, with θ as in Theorem 4.3 satisfying I = ∞ does not imply that u ≡ 0. We consider the group G = SL(2, C) (see [19], P. 433; [34], P. 313). In this case A =( eH 0 0 e−H! : H ∈ R) and a =(AH = H 0 0 −H! : H ∈ R) . It is known that the Jacobian of the Haar measure is sinh2 2H so that we have and the distance function is given by φ(H) = sinh 2H, for H ∈ R, kHkB = 4H, for H ∈ R. The elementary spherical functions are given by sin λH φλ(H) = C , for λ, H ∈ R, λ sinh 2H φ0(H) = C H sinh 2H , for H ∈ R. We fix 0 ≤ α < 1 and t0 > 0. We will give an example of a non-zero initial value f ∈ Cc(K\G/K) such that the corresponding solution u to the Schrodinger equation (4.7) satisfies (4.14) u(H, t0) ≤ C φ0(H)α e−kHkB θ(kHkB ), for H ∈ a ∼= R, where θ is as in Theorem 4.3 satisfying (4.9). We consider η ∈ R such that 0 < η < 1 − α and define β = 1 − α − η. For any h ∈ C ∞ c (R) we have (4.15) for all H ∈ R, bh(H) ≤ CeβH, We consider such a non-zero function h which is supported in [β′, β] ⊂ [−β, β] ⊂ R where 0 < β′ < β. Since θ is a function decreasing to zero, there exists M1 > 1 such that (4.16) θ(4H) < η 4 , for H > M1, UNCERTAINTY PRINCIPLES OF INGHAM AND PALEY-WIENER ON SEMISIMPLE LIE GROUPS 19 and there exists M2 > 1 such that (4.17) eH ≤ C sinh 2H, for H > M2. From (4.15), (4.16) and (4.17) it follows that for M = max{M1, M2} > 1 and H > M ≤ CHα(sinh 2H)1−αe−4Hθ(4H) bh(H) ≤ Ce(1−α)He−ηH sinh 2H(cid:19)α = C(cid:18) H (sinh 2H)e−4Hθ(4H) (4.18) = Cφ(H)φ0(H)αe−4Hθ(4H). We define a function f ∈ C ∞ c (R) by (4.19) f (H) = 1 2t0 e−4i H2 t0 φ(H)−1h(cid:18) H 2t0(cid:19) , for H ∈ R. This is well defined since φ vanishes only at 0 and h is supported away from zero. We consider the solution u(H, t) of the system (4.7) with the initial value f . It follows from (4.11) and (4.19) that So by (4.10) Hence by (4.18) we get that 1 2t0 gf (H) = h(cid:18) H 2t0(cid:19) , u(H, t0)φ(H) = Cbh(H), for H ∈ R. for H ∈ R. u(H, t0)φ(H) ≤ Cφ(H)φ0(H)αe−4Hθ(4H), for H > M. Since by continuity the function H 7→ u(H, t0)e4Hθ(4H)(φ0(H))−α is bounded on the compact set [0, M ] it follows that the solution u satisfies (4.14) with non-zero initial data f ∈ C ∞ c (R). 4.2.2. Symmetric Space of Noncompact Type. We shall now prove a unique continuation prop- erty of solution to the Schrodinger equation (4.7) in the context of Riemannian symmetric spaces of noncompact type using Theorem 4.1. As in the case of complex semisimple Lie groups, here also we need to impose extra decay on the solution u(x, t) and using a similar example we will show that such decay is necessary (see Remark 1). However, here we shall prove our result by reducing the problem to the Euclidean case via the Radon transform. It can be seen that for the initial data f ∈ L2(X) there exists unique solution u(·, t) ∈ L2(X) satisfying the Schrodinger equation (4.7) in the sense of distributions. However the Radon transform is not defined in gen- eral for functions in L2(X). So we need to consider our initial data f in the L2-Schwartz space 20 MITHUN BHOWMIK AND SUPARNA SEN S(X) consisting of smooth rapidly decreasing functions on X, on which the Radon transform happens to be defined. We shall start with the definition of S(X). Let DL(G) and DR(G) denote the algebra of left invariant and that of right invariant differen- tial operators on G respectively. The L2-Schwartz space S(G) is defined as the space of smooth functions f on G such that for each D ∈ DL(G), E ∈ DR(G) and l ∈ N (4.20) (1 + σ(g))lφ0(g)−1(DEf )(g) < ∞. sup g∈G The L2-Schwartz space S(X) is then defined as the space of f ∈ S(G) which are right invariant under K (see [18], P. 214). For λ ∈ a∗ C and kM ∈ K/M we define the function eλ,k : X → C by eλ,k(x) = e(iλ−ρ)H(x−1k), for x = gK ∈ X, g ∈ G. It is known that these functions eλ,k appearing in the definition of the Helgason-Fourier transform (2.3) are eigenfunctions of the Laplace-Beltrami operator ∆ given by ∆eλ,k = −(kλk2 B + kρk2 B)eλ,k, for λ ∈ a∗ C, kM ∈ K/M, (see [18], P. 99 and [19], Ch. II). It follows that for f ∈ L2(X) the unique solution u(·, t) ∈ L2(X) satisfying the Schrodinger equation (4.7) is characterized by (4.21) for kM ∈ K/M and λ ∈ a∗. eut(λ, kM ) = e−it(kλk2 B +kρk2 B)ef (λ, kM ), However we are interested in the special case where the initial data f ∈ S(X). In this case it turns out that the solution ut(·) = u(t, ·) ∈ S(X) (see [29], P. 872; [11], Theorem 4.1.1). It follows that that the analysis of solutions of the Schrodinger equation on Rn carries out to X when the Fourier transform on Rn is replaced by the Helgason-Fourier transform. For any f ∈ S(X) the Radon transform Rf is well defined (see [18], P. 218, Theorem 1.15). From (2.10) and (4.21) it follows that for λ ∈ a∗ and fixed kM ∈ K/M we have (4.22) FA((Rut)(kM, ·)(λ) = e−it(kλk2 B +kρk2 B)FA((Rf )(kM, ·)(λ). We shall recall few facts which will be needed. Firstly, the function σ satisfies the following inequality (4.23) σ(an) ≥ σ(a), for all a ∈ A, n ∈ N (see [15], Lemma 6.2.7). Moreover, there exists a non-negative integer m ≥ 0 such that for some constant C0 > 0, (4.24) φ0(an) (1 + σ(an))−mdn ≤ C0, for all a ∈ A, eρ log aZN (see [15], P. 264, Theorem 6.2.3). We shall now present the unique continuation property for solutions to the Schrodinger equation (4.7) for symmetric spaces. UNCERTAINTY PRINCIPLES OF INGHAM AND PALEY-WIENER ON SEMISIMPLE LIE GROUPS 21 Theorem 4.5. Let u(x, t) be a solution to the equation (4.7) with initial value f ∈ C ∞ Suppose there exists t0 > 0 such that c (X). (4.25) u(x, t0) ≤ C φ0(x) e−ψ(σ(x)), for x ∈ X, where ψ : [0, ∞) → [0, ∞) is a non-decreasing function. If then f is zero on X. I =Z ∞ 1 ψ(r) 1 + r2 dr = ∞, Proof. Since f ∈ C ∞ function on G, from the definition of L2-Schwartz space (4.20) we get that for each l ∈ N c (X) ⊂ S(X), it follows that ut ∈ S(X). Viewing ut as a right K-invariant (4.26) ut(g) ≤ Cφ0(g)(1 + σ(g))−2l, for g ∈ G. By multiplying the inequalities in (4.25) and (4.26) it follows that for each l ∈ N (4.27) u(x, t0) ≤ C φ0(x) (1 + σ(x))−le− 1 2 ψ(σ(x)), for x ∈ X. Now, from the expression of ut given in (4.2) and the relation (4.22) it follows that for fixed kM ∈ K/M , (Rut)(kM, ·) is a solution to the system (4.1) with initial value (Rf )(kM, ·) and damping parameter kρk2 B. We wish to apply Theorem 4.1 to this particular solution. Let us fix kM ∈ K/M . Since f ∈ C ∞ c (A). We need to prove that (Rut)(kM, ·) satisfies the estimate (4.3) at time t = t0. Indeed, for a ∈ A and l = m using (2.9) and (4.27) we get that c (X) it is easy to see that Rf (kM, ·) ∈ C ∞ (Rut0 )(kM, a) = eρ log aZN ≤ C eρ log aZN ut0 (kan · o) dn φ0(an) (1 + σ(an))−me− 1 2 ψ(σ(an)) dn. Since ψ is non-decreasing using (4.23) it follows that (4.28) ψ(σ(an)) ≥ ψ(σ(a)), for all a ∈ A, n ∈ N. Using (4.28) and (4.24) we get that (Rut0 )(kM, a) ≤ Ce− 1 2 ψ(σ(a)), for all a ∈ A. Using the equivalence of norms in finite dimensional spaces, we can invoke Theorem 4.1 to get that (Rf )(kM, ·) is zero on A. Since this is true for all kM ∈ K/M we get that Rf is zero on K/M × A. By the injectivity of Radon transform ([18], P. 220, Corollary 1.6) we conclude that f is zero. Hence so is u. (cid:3) 22 MITHUN BHOWMIK AND SUPARNA SEN Remark 4.6. (1) We give an example to show that if 0 ≤ α < 1 then there exists a non- decreasing function ψ (depending on α) and a non-zero initial value f ∈ Cc(K\G/K) such that the corresponding solution u to the Schrodinger equation (4.7) satisfies (4.29) u(H, t0) ≤ C φ0(H)α e−ψ(kHkB ), for H ∈ a ∼= R, where ψ is as in Theorem 4.5 with I = ∞. We again consider the group G = SL(2, C) and as in Remark 4.4 we get h ∈ C ∞ c (R) supported in [β′, β] ⊂ [−β, β] with 0 < β′ < β such that where β = 1 − α − η and 0 < η < 1 − α. We now define bh(H) ≤ CeβH, for all H ∈ R, (4.30) ψ(H) = ηH, for H ∈ R. Using (4.17) and (4.30) in (4.18) we get that for H > M2 bh(H) ≤ Cφ(H)φ0(H)αe−ψ(H). c (R) by (4.19) we conclude as in Remark 4.4 that u satisfies (4.29) with Defining f ∈ C ∞ non-zero initial data f ∈ C ∞ c (R). (2) It seems to be an interesting problem to see whether Theorem 3.9 and Theorem 4.3 can be generalized to general Riemannian symmetric spaces of noncompact type. Acknowledgement. We would like to thank Swagato K. Ray for suggesting this problem and for the many useful discussions during the course of this work. We also thank Rudra P Sarkar for his valuable comments regarding this work. References [1] Ben Saıd, S.; Thangavelu, S.; Dogga, V. N. Uniqueness of solutions to Schrodinger equations on H-type groups, J. Aust. Math. Soc. 95 (2013) no. 3, 297-314. MR3164504 [2] Bhowmik, M.; Ray, S. K.; Sen, S. Around Uncertainty Principles of Ingham-type on Rn, Tn and two step nilpotent Lie Groups, preprint, arXiv:1605.09616 [3] Bhowmik, M.; Sen, S. An Uncertainty Principle of Paley and Wiener on Euclidean Motion Group, preprint, arXiv:1606.01704 [4] Bochner, S. Quasi-analytic functions, Laplace operator, positive kernels, Ann. of Math. (2) 51, (1950). 68-91. MR0032708 (11,334g) [5] Bochner, S.; Taylor, A. E. Some Theorems on Quasi-Analyticity for Functions of Several Variables, Amer. J. Math. 61 (1939), no. 2, 303-329. MR1507378 [6] Burckel, Robert B. An introduction to classical complex analysis, Pure and Applied Mathematics, 82, Aca- demic Press, Inc [Harcourt Brace Jovanovich, Publishers], New York-London, 1979. MR0555733 (81d:30001) [7] Chanillo, S. Uniqueness of solutions to Schrodinger equations on complex semi-simple Lie groups, Proc. Indian Acad. Sci. Math. Sci. 117 (2007), no. 3, 325-331. MR2352052 (2008h:22010) [8] Conway, John B. Functions of one complex variable II, Graduate Texts in Mathematics, 159, Springer-Verlag, New York, 1995. MR1344449 (96i:30001) UNCERTAINTY PRINCIPLES OF INGHAM AND PALEY-WIENER ON SEMISIMPLE LIE GROUPS 23 [9] Cowling, M.; Nevo, A. Uniform estimates for spherical functions on complex semisimple Lie groups, Geom. Funct. Anal. 11 (2001), no. 5, 900-932. MR1873133 (2002k:43005) [10] Cowling, M.; Sitaram, A.; Sundari, M. Hardy's uncertainty principle on semisimple groups, Pacific J. Math. 192 (2000), no. 2, 293-296. MR1744570 (2001c:22007) [11] Eguchi, M. Asymptotic expansions of Eisenstein integrals and Fourier transform on symmetric spaces, J. Funct. Anal. 34 (1979), no. 2, 167-216. MR0552702 (81e:43022) [12] Escauriaza, L.; Kenig, C. E.; Ponce, G.; Vega, L. On uniqueness properties of solutions of Schrodinger equations, Comm. Partial Differential Equations 31 (2006), no. 10-12, 18111823. MR2273975 (2009a:35198) [13] Escauriaza, L.; Kenig, C. E.; Ponce, G.; Vega, L. Uniqueness properties of solutions to Schrodinger equations, Bull. Amer. Math. Soc. (N.S.) 49 (2012), no. 3, 415-442. MR2917065 [14] Folland, G, B.; Sitaram, A. The uncertainty principle: A mathematical survey, J. Four. Anal. Appl. 3 (1997), no. 3, 207-238. MR1448337 (98f:42006) [15] Gangolli, R.; Varadarajan V. S. Harmonic Analysis of Spherical Functions on Real Reductive Groups, Results in Mathematics and Related Areas, 101, Springer-Verlag, Berlin, 1988. MR954385 (89m:22015) [16] Havin, V.; Joricke, B. The uncertainty principle in harmonic analysis, Ergebnisse der Mathematik und ihrer Grenzgebiete, 3, Folge, 28, Berlin, Springer-Verlag, 1994. MR1303780 (96c:42001) [17] Helgason, S. Differential geometry, Lie groups, and symmetric spaces, Graduate Studies in Mathematics, 34, American Mathematical Society, Providence, RI, 2001. MR1834454 (2002b:53081) [18] Helgason, S. Geometric Analysis on Symmetric Spaces, Mathematical Surveys and Monographs 39. American Mathematical Society, Providence, RI, 1994. MR1280714 (96h:43009) [19] Helgason, S. Groups and geometric analysis, Integral geometry, invariant differential operators, and spherical functions, Mathematical Surveys and Monographs, 83. American Mathematical Society, Providence, RI, 2000. MR1790156 (2001h:22001) [20] Hirschman, I. I. On the behaviour of Fourier transforms at infinity and on quasi-analytic classes of functions Amer. J. Math. 72 (1950), 200-213. MR0032816 (11,350f) [21] Ingham, A. E. A Note on Fourier Transforms, J. London Math. Soc. S1-9 no. 1, 29. MR1574706 [22] Knapp, A., Representation theory of semisimple groups, An overview based on examples, Princeton Mathe- matical Series, 36. Princeton University Press, Princeton, NJ, 1986. MR855239 (87j:22022) [23] Koosis, P. The logarithmic integral I (Corrected reprint of the 1988 original) Cambridge Studies in Advanced Mathematics, 12. Cambridge University Press, Cambridge, 1998. xviii+606 pp. MR1670244 (99j:30001) [24] Levinson, N. Gap and Density Theorems American Mathematical Society Colloquium Publications, v. 26. American Mathematical Society, New York, 1940. MR0003208 (2,180d) [25] Levinson, N. On a Class of Non-Vanishing Functions Proc. London Math. Soc. S2-41 no. 5, 393. MR1576177 [26] Ludwig, J.; Muller, D. Uniqueness of solutions to Schrodinger equations on 2-step nilpotent Lie groups, Proc. Amer. Math. Soc. 142 (2014) no. 6, 2101-2118. MR3182028 [27] Paley, R. E. A. C.; Wiener, N. Fourier transforms in the complex domain, American Mathematical Society Colloquium Publications, 19. American Mathematical Society, Providence, RI, 1987. MR1451142 (98a:01023) [28] Paley, R. E. A. C.; Wiener, N. Notes on the theory and application of Fourier transforms. I, II, Trans. Amer. Math. Soc. 35 (1933), no. 2, 348-355. MR1501688 [29] Pasquale, A.; Sundari, M. Uncertainty principles for the Schrodinger equation on Riemannian symmetric spaces of the noncompact type, Ann. Inst. Fourier (Grenoble) 62 (2012), no. 3, 859-886. MR3013810 [30] Rudin, W. Real and Complex Analysis, McGraw-Hill Book Co., New York, 1987. MR924157 (88k:00002) [31] Sarkar, R. P.; Sengupta, J. Beurling's theorem and characterization of heat kernel for Riemannian symmetric spaces of noncompact type, Canad. Math. Bull. 50 (2007), no. 2, 291-312. MR2317450 (2008e:43014) 24 MITHUN BHOWMIK AND SUPARNA SEN [32] Stein, E. M.; Shakarchi, R. Complex analysis, Princeton Lectures in Analysis, II. Princeton University Press, Princeton, NJ, 2003. MR1976398 (2004d:30002) [33] Stein, E. M.; Weiss, G. Introduction to Fourier analysis on Euclidean spaces, Princeton Mathematical Series, No. 32. Princeton University Press, Princeton, N.J., 1971. MR0304972 (46 # 4102) [34] Terras, A. Harmonic analysis on symmetric spaces and applications II, Springer-Verlag, Berlin, 1988. MR0955271 (89k:22017) [35] Thangavelu, S. An Introduction to the Uncertainty Principle. Hardy's theorem on Lie groups, Progress in Mathematics, 217. Birkhauser Boston, Inc., Boston, MA, 2004. MR2008480 (2004j:43007) [36] Thangavelu, S. On Theorems of Hardy, Gelfand-Shilov and Beurling for Semisimple Lie Groups, Publ. Res. Inst. Math. Sci. 40 (2004), no. 2, 311-344. MR2049638 (2005e:43014) Stat-Math Unit, Indian Statistical Institute, 203 B. T. Road, Kolkata - 700108, India. E-mail address: [email protected], [email protected]
1604.05485	4	1604	2017-04-19T07:02:51	Contractions with Polynomial Characteristic Functions II. Analytic Approach	[ "math.FA", "math.OA" ]	The simplest and most natural examples of completely nonunitary contractions on separable complex Hilbert spaces which have polynomial characteristic functions are the nilpotent operators. The main purpose of this paper is to prove the following theorem: Let $T$ be a completely nonunitary contraction on a Hilbert space $\mathcal{H}$. If the characteristic function $\Theta_T$ of $T$ is a polynomial of degree $m$, then there exist a Hilbert space $\mathcal{M}$, a nilpotent operator $N$ of order $m$, a coisometry $V_1 \in \mathcal{L}(\overline{ran} (I - N N^) \oplus \mathcal{M}, \overline{ran} (I - T T^))$, and an isometry $V_2 \in \mathcal{L}(\overline{ran} (I - T^* T), \overline{ran} (I - N^* N) \oplus \mathcal{M})$, such that \[ \Theta_T = V_1 \begin{bmatrix} \Theta_N & 0 0 & I_{\mathcal{M}} \end{bmatrix} V_2. \]	math.FA	math	CONTRACTIONS WITH POLYNOMIAL CHARACTERISTIC FUNCTIONS II. ANALYTIC APPROACH CIPRIAN FOIAS, CARL PEARCY, AND JAYDEB SARKAR This paper is dedicated to our lifelong friend Ron Douglas on the occasion of his upcoming 80th birthday Abstract. The simplest and most natural examples of completely nonunitary contractions on separable complex Hilbert spaces which have polynomial characteristic functions are the nilpotent operators. The main purpose of this paper is to prove the following theorem: Let T be a completely nonunitary contraction on a Hilbert space H. If the characteristic function ΘT of T is a polynomial of degree m, then there exist a Hilbert space M, a nilpotent operator N of order m, a coisometry V1 ∈ L(ran(I − N N ∗) ⊕ M, ran(I − T T ∗)), and an isometry V2 ∈ L(ran(I − T ∗T ), ran(I − N ∗N ) ⊕ M), such that ΘT = V1(cid:20)ΘN 0 0 IM(cid:21) V2. 1. Introduction This is a sequel to our paper [4], where we identified the structure of the completely nonunitary contractions on a Hilbert space that have a polynomial characteristic function. Namely, we proved that the characteristic function ΘT of a completely nonunitary contraction T on a separable, infinite dimensional, complex Hilbert space H is a polynomial if and only if there exist three closed subspaces H1, H0, H−1 of H with H = H1 ⊕H0 ⊕H−1, a pure isometry S in L(H1), a nilpotent N in L(H0), and a pure co-isometry C in L(H−1), such that T has the matrix representation T =   S ∗ ∗ 0 N ∗ 0 0 C   . Moreover, the multiplicities of S and C, in other words, dim ker S∗ and dim ker C are unitary invariants of T , and the nilpotent operator is uniquely determined by T up to a quasi-similarity. For earlier results on contractions with constant characteristic functions see [1], [9] and [10]. Recall that a pure isometry is a unilateral shift of some multiplicity and a pure coisometry is the adjoint of a pure isometry. Recall also that a contraction T on a Hilbert space H, (i.e., kT hk ≤ khk for all h in H) is completely nonunitary (c.n.u.) if there is no nontrivial reducing subspace M of H for T such that T M is unitary. 2000 Mathematics Subject Classification. 47A45, 47A20, 47A48, 47A56. Key words and phrases. Characteristic function, model, nilpotent operators, operator valued polynomials. 1 2 FOIAS, PEARCY, AND SARKAR In this paper we shall adopt a second approach to prove the theorem stated in the abstract, based essentially on new factorizations of characteristic functions of upper triangular 3 × 3 block contractions (see Theorem 2.3). Before we continue we recall the notion of the characteristic function of a contraction. Consider a contraction T on a Hilbert space H. The defect operators DT and DT ∗ and the defect spaces DT and DT ∗ of T are defined by DT = (IH − T ∗T ) 1 2 , and DT ∗ = (IH − T T ∗) 1 2 , and DT = RanDT , and DT ∗ = RanDT ∗, respectively. Then the characteristic function of the contraction T is the L(DT , DT ∗)-valued contractive analytic function defined by ΘT (z) = [−T + zDT ∗(IH − zT ∗)−1DT ]DT (z ∈ D). In particular, ΘT is a L(DT , DT ∗)-valued bounded analytic function on D (see [6]). Moreover, the characteristic function ΘT is purely contractive, that is, kΘT (0)ηk < kηk (η ∈ DT , η 6= 0). Let Θ : D → L(M, M∗) and Ψ : D → L(N , N∗) be two operator valued analytic functions on D. We say that Θ and Ψ coincide and write Θ ∼= Ψ if there exist two unitary operators τ : M → N and τ∗ : M∗ → N∗ such that Θ(z) = τ −1 (z ∈ D), ∗ Ψ(z)τ or, equivalently, for all z ∈ D the following diagram commutes: M Θ(z) −−−→ M∗ τy N τ∗y Ψ(z) −−−→ N∗ The characteristic function is a complete unitary invariant in the following sense (see [6], Theorem 3.4): Two c.n.u. contractions T on H and R on K are unitarily equivalent (that is, there is a unitary operator U from H to K such that T = U ∗RU) if and only if ΘT ∼= ΘR. Moreover, for a given L(E , E∗)-valued purely contractive analytic function Θ defined on D, there exists a c.n.u. contraction T on some Hilbert space, explicitly determined by Θ, such that ΘT coincides with Θ. Contractive operator valued analytic functions play an important role in operator theory and serve as a bridge between operator theory and function theory in terms of systems theory and interpolation theory (cf. [2], [6], [8]). The class of nilpotent contractions yields a natural set of examples of operators that have polynomial characteristic functions. Indeed, let N be a contraction and a nilpotent operator CONTRACTIONS WITH POLYNOMIAL CHARACTERISTIC FUNCTIONS 3 of order m, m ≥ 1, that is, kNk ≤ 1, N m = 0, and N m−1 6= 0. The characteristic function ΘN of N is given by ΘN (z) = [−N + zDN ∗(IH − zN ∗)−1DN ]DN = [−N + = [−N + ∞ Xp=0 Xp=0 m−1 zp+1DN ∗N ∗pDN ]DN zp+1DN ∗N ∗pDN ]DN , for all z ∈ D. Therefore ΘN is a polynomial in z of degree at most m with operator coefficients. From this viewpoint, it is important to understand, up to unitary equivalance, the analytic structure of polynomial characteristic functions of contractions. The main goal of the present paper is to address this issue. More specifically, in Theorem 3.2 we prove: If the characteristic function ΘT of a c.n.u. contraction T is a polynomial of degree m, then there exist a Hilbert space M, a nilpotent operator N of order m, a co-isometry V1 ∈ L(DN ∗ ⊕ M, DT ∗), and an isometry V2 ∈ L(DT , DN ⊕ M), such that ΘT = V1(cid:20)ΘN 0 0 IM(cid:21) V2. Along the way we prove the following factorization result for characteristic functions (see Theorem 2.3): Let H1, H0, H−1 be three Hilbert spaces and set H = H1 ⊕ H0 ⊕ H−1. Let T =   S ∗ ∗ 0 N ∗ 0 C 0   be any contraction on H with the above matricial form. Then the characteristic function ΘT of T and (cid:20)ΘC 0 0 IE1(cid:21) U1(cid:20)ΘN 0 0 IM(cid:21) U2(cid:20)ΘS 0 0 IE2(cid:21) . coincide, where E1, E2 and M are Hilbert spaces, and U1 ∈ L(DN ∗ ⊕ M, DC ⊕ E1) and U2 ∈ L(DS ∗ ⊕ E2, DN ⊕ M) are unitary operators. Our results rely on the upper triangular representation of operators with polynomial char- acteristic functions (see Theorem 3.1) and a factorization of characteristic functions of upper triangular 2 × 2 block contractions due to Sz.-Nagy and the first author (see Theorem 2.2). The rest of this paper is organized as follows:. In Section 2, we give the factorization of the characteristic function of an upper triangular 3 × 3 block contraction on Hilbert space. Our main result is given in Section 3, and provides a complete analytic characterization of polynomial characteristic functions for c.n.u. contractions on Hilbert space. 2. Factorizations of characteristic functions We start by recalling some known facts about upper triangular 2 × 2 block contractions, since they will be frequently used in what follows. 4 FOIAS, PEARCY, AND SARKAR The first is a classification of 2 × 2 block contractions. This is the content of Theorem 1 in [7] (also see Chapter IV, Lemma 2.1 in [3]). Theorem 2.1. Let H1 and H2 be Hilbert spaces and let T = (cid:20)T1 X 0 T2(cid:21) be a bounded linear operator on H1 ⊕ H2. Then T is a contraction if and only if T1 and T2 are contractions and for some contraction Γ from DT2 to DT ∗ 1 . X = DT ∗ 1 ΓDT2, The second key tool used in our development is the factorization of characteristic functions of 2 × 2 block contractions (see Theorem 2 in [7]). Theorem 2.2. Let H1 and H2 be Hilbert spaces, let T = (cid:20)T1 X 0 T2(cid:21) be a contraction on H1 ⊕ H2, and let X = DT ∗ unitary operators τ ∈ L(DT , DT1 ⊕ DΓ) and τ∗ ∈ L(DT ∗, DT ∗ 1 ΓDT2 for some contraction Γ ∈ L(DT ∗ 2 ⊕ DΓ∗) such that 1 , DT2). Then there exist ΘT (z) = τ −1 ∗ (cid:20)ΘT2(z) 0 0 IDΓ∗(cid:21) J[Γ](cid:20)ΘT1(z) 0 0 IDΓ(cid:21) τ (z ∈ D), where J[Γ] = (cid:20) Γ∗ DΓ DΓ∗ −Γ(cid:21) ∈ L(DT ∗ 1 ⊕ DΓ, DT2 ⊕ DΓ∗). Recall that if A is a contraction from H to K then (2.1) J[A] = (cid:20) A∗ DA DA∗ −A(cid:21) is a unitary operator from K ⊕ DA to H ⊕ DA∗ (see Halmos [5]). We are now ready to prove our first factorization result. Theorem 2.3. Let H1, H0, H−1 be Hilbert spaces, and let H = H1 ⊕ H0 ⊕ H−1. Let T =   S ∗ ∗ 0 N ∗ 0 C 0   , be a contraction on H. Then there exist three Hilbert spaces E1, E2 and M and two unitary operators U1 ∈ L(DN ∗ ⊕ M, DC ⊕ E1) and U2 ∈ L(DS ∗ ⊕ E2, DN ⊕ M) such that ΘT ∼= (cid:20)ΘC 0 Proof. Set 0 IE1(cid:21) U1(cid:20)ΘN 0 0 IM(cid:21) U2(cid:20)ΘS 0 0 IE2(cid:21) . K1 = H1 ⊕ H0, T = (cid:20)T1 X1 0 C (cid:21) ∈ L(K1 ⊕ H−1), CONTRACTIONS WITH POLYNOMIAL CHARACTERISTIC FUNCTIONS 5 and T1 = (cid:20)S X 0 N(cid:21) ∈ L(H1 ⊕ H0), where X1 ∈ L(H−1, K1) and X ∈ L(H0, H1). Theorem 2.1 implies that there exist contrac- tions Γ1 ∈ L(DC, DT ∗ 1 ) and Γ ∈ L(DN , DS ∗) such that 1 Γ1DC, X1 = DT ∗ and By Theorem 2.2 there exist unitary operators X = DS ∗ΓDN . (2.2) and (2.3) such that and for all z ∈ D, where and τ1 : DT → DT1 ⊕ DΓ1, τ1∗ : DT ∗ → DC ∗ ⊕ DΓ∗ 1 , τ : DT1 → DS ⊕ DΓ, τ∗ : DT ∗ 1 → DN ∗ ⊕ DΓ∗, ΘT (z) = τ −1 0 1∗ (cid:20)ΘC(z) ∗ (cid:20)ΘN (z) 0 IDΓ∗ 0 0 1(cid:21) J[Γ1](cid:20)ΘT1(z) IDΓ∗(cid:21) J[Γ](cid:20)ΘS(z) 0 0 0 IDΓ1(cid:21) τ1, IDΓ(cid:21) τ, 0 ΘT1(z) = τ −1 J[Γ1] = (cid:20) Γ∗ DΓ∗ 1 DΓ1 1 −Γ1(cid:21) ∈ L(DT ∗ 1 ⊕ DΓ1, DC ⊕ DΓ∗ 1), J[Γ] = (cid:20) Γ∗ DΓ DΓ∗ −Γ(cid:21) ∈ L(DS ∗ ⊕ DΓ, DN ⊕ DΓ∗), are unitary operators (see (2.1)). Now setting ΦS(z) = (cid:20)ΘS(z) 0 for all z ∈ D, we get 0 IDΓ(cid:21) , ΦN (z) = (cid:20)ΘN (z) 0 0 IDΓ∗(cid:21) , and ΦC(z) = (cid:20)ΘC(z) 0 0 1(cid:21) , IDΓ∗ ΘT (z) = τ −1 0 IDΓ1(cid:21) τ1 0 1∗ ΦC(z)J[Γ1](cid:20)ΘT1(z) 1∗ ΦC(z)J[Γ1](cid:20)τ −1 1∗ ΦC(z)(cid:16)J[Γ1](cid:20)τ −1 1∗ ΦC(z)U1(cid:20)ΦN (z)J[Γ]ΦS(z)τ ∗ 0 0 0 ∗ ΦN (z)J[Γ]ΦS(z)τ 0 IDΓ1(cid:21) τ1 IDΓ1(cid:21)(cid:17)(cid:20)ΦN (z)J[Γ]ΦS (z)τ 0 0 0 IDΓ1(cid:21) τ1, = τ −1 = τ −1 = τ −1 0 IDΓ1(cid:21) τ1 for all z ∈ D, where U1 ∈ L((DN ∗ ⊕ DΓ∗) ⊕ DΓ1, DC ⊕ DΓ∗ 1) is the unitary operator defined by 6 FOIAS, PEARCY, AND SARKAR Hence U1 = J[Γ1](cid:20)τ −1 ∗ 0 0 IDΓ1(cid:21) . ΘT (z) = τ −1 0 1∗ ΦC(z)U1(cid:20)ΦN (z)J[Γ]ΦS(z)τ IDΓ1(cid:21)(cid:20)J[Γ] 1∗ ΦC(z)U1(cid:20)ΦN (z) IDΓ∗ ⊕DΓ1(cid:21)(cid:20)J[Γ] 1∗ ΦC(z)U1(cid:20)ΘN (z) 0 0 0 0 0 0 = τ −1 = τ −1 0 IDΓ1(cid:21) τ1 IDΓ1(cid:21)(cid:20)ΦS(z) 0 0 0 IDΓ1(cid:21)(cid:20)τ 0 0 IDΓ1(cid:21) τ1 IDΓ⊕DΓ1(cid:21)(cid:20)τ 0 0 0 IDΓ1(cid:21) τ1, 0 IDΓ1(cid:21)(cid:20)ΘS(z) 0 for all z ∈ D. Let U2 ∈ L((DS ∗ ⊕DΓ)⊕DΓ1 , (DN ⊕DΓ∗)⊕DΓ1) and τ1 ∈ L(DT , (DS ⊕DΓ)⊕DΓ1) be unitary operators defined by U2 = (cid:20)J[Γ] 0 0 IDΓ1(cid:21) , and respectively. Hence we obtain τ1 = (cid:20)τ 0 IDΓ1(cid:21) τ1, 0 (2.4) ΘT (z) = τ −1 1∗ (cid:16)(cid:20)ΘC(z) 0 0 1(cid:21) U1(cid:20)ΘN (z) 0 IDΓ∗ 0 IDΓ∗ ⊕DΓ1(cid:21) U2(cid:20)ΘS(z) 0 0 IDΓ⊕DΓ1(cid:21)(cid:17)τ1, for all z ∈ D, and therefore ΘT ∼= (cid:20)ΘC 0 0 1(cid:21) U1(cid:20)ΘN 0 IDΓ∗ 0 IDΓ∗ ⊕DΓ1(cid:21) U2(cid:20)ΘS 0 0 IDΓ⊕DΓ1(cid:21) , holds. Setting E1 = DΓ∗ proof of the theorem. 1, M = DΓ∗ ⊕ DΓ1 and E2 = DΓ ⊕ DΓ1 in the above, we conclude the Of particular interest is the case when S and C ∗ are pure isometries. Corollary 2.4. With the hypotheses of Theorem 2.3, let also assume that S and C ∗ are pure isometries. Then there exist a Hilbert space M, a co-isometry V1 ∈ L(DN ∗ ⊕ M, DT ∗), and an isometry V2 ∈ L(DT , DN ⊕ M), such that ΘT = V1(cid:20)ΘN 0 0 IM(cid:21) V2. Proof. Notice that since DC ∗ = {0H−1} and DS = {0H1}, the characteristic functions ΘC : D → L(DC, DC ∗) of C and ΘS : D → L(DS, DS ∗) of S are identically zero, that is, 0C := ΘC ≡ 0 : DC → {0H−1}, and 0S := ΘS ≡ 0 : {0H1} → DS ∗. Furthermore, the unitary operators in (2.2) and (2.3) become (2.5) τ1 : DT → DT1 ⊕ DΓ1, τ1∗ : DT ∗ → {0H−1} ⊕ DΓ∗ 1 , CONTRACTIONS WITH POLYNOMIAL CHARACTERISTIC FUNCTIONS 7 and (2.6) τ : DT1 → {0H1} ⊕ DΓ, τ∗ : DT ∗ 1 → DN ∗ ⊕ DΓ∗. This along with (2.4) yields ΘT = τ −1 0 1∗ (cid:20)0C = V1(cid:20)ΘN 0 0 1(cid:21) U1(cid:20)ΘN 0 IDΓ∗ 0 IDΓ∗ ⊕DΓ1(cid:21) U2(cid:20)0S 0 0 IDΓ⊕DΓ1(cid:21) τ1 0 IM(cid:21) V2, M = DΓ∗ ⊕ DΓ1, V1 = τ −1 1∗ (cid:20)0C 0 0 1(cid:21) U1 ∈ L((DN ∗ ⊕ DΓ∗) ⊕ DΓ1, DT ∗) IDΓ∗ V2 = U2(cid:20)0S 0 0 IDΓ⊕DΓ1(cid:21) τ1 ∈ L(DT , (DN ⊕ DΓ∗) ⊕ DΓ1). where and Now using 0C0∗ readily see that V1V ∗ 1 = IDT ∗ and V ∗ 2 V2 = IDT . This completes the proof of the corollary. C = IDC∗ = I{0H−1 } and 0∗ S0S = IDS = I{0H1 } along with (2.5) and (2.6) we 3. Polynomial characteristic functions For the readers convenience, we first state the main result of [4]. Theorem 3.1. Let T be a c.n.u. contraction on a Hilbert space H. Then the characteristic function ΘT of T is a polynomial of degree m if and only if there exist three closed subspaces H1, H0, H−1 of H with H = H1⊕H0⊕H−1, a pure isometry S in L(H1), a nilpotent N of order m in L(H0), and a pure co-isometry C in L(H−1), such that T has the matrix representation T =   S ∗ ∗ 0 N ∗ 0 C 0   . We are now ready for the main theorem on analytic description of contractions which have polynomial characteristic functions. Theorem 3.2. Let T be a c.n.u. contraction on a Hilbert space H. If the characteris- tic function ΘT of T is a polynomial of degree m, then there exist a Hilbert space M, a nilpotent operator N of order m, a co-isometry V1 ∈ L(DN ∗ ⊕ M, DT ∗), and an isometry V2 ∈ L(DT , DN ⊕ M), such that ΘT = V1(cid:20)ΘN 0 0 IM(cid:21) V2. 8 FOIAS, PEARCY, AND SARKAR Proof. Let T be a c.n.u. contraction such that the characteristic function ΘT of T is a polynomial of degree m. According to Theorem 3.1 there exist closed subspaces H1, H0, H−1 of H such that H = H1 ⊕ H0 ⊕ H−1 and such that with respect to that decomposition, T admits the matrix representation T =   S ∗ ∗ 0 N ∗ 0 C 0   , where S ∈ L(H1) is a pure isometry, N ∈ L(H0) is a nilpotent operator of order m, and C ∈ L(H−1) is a pure coisometry. The result now follows from Corollary 2.4. Remark 3.3. The converse of the above theorem is not true in full generality: Let M be an infinite dimensional separable Hilbert space, and let T be a c.n.u. contraction with infinite dimensional defect spaces (for example, one can consider T = S ⊕ S∗ on H 2 E(D), where E is an infinite dimensional Hilbert space, H 2 E(D) is the E-valued Hardy space, and S is the shift operator on H 2 E(D)). Let N be a nilpotent operator of order m and let E(D) ⊕ H 2 be an isometry, where V2 = (cid:20)V21 V22(cid:21) : DT → DN ⊕ M, kV21ηk = kV22ηk (η ∈ DT ). Also, let V1 : DN ∗ ⊕ M → DT ∗ be a coisometry with ker V1 = DN ∗. If ΘT = V1(cid:20)ΘN 0 0 IM(cid:21) V2, then ΘT is a polynomial of degree 0. However, it is easy to see that the following weak converse of Theorem 3.2 is true: Let T be a c.n.u. contraction on a Hilbert space H. Let ΘT = V1(cid:20)ΘN 0 0 IM(cid:21) V2, for some Hilbert space M, nilpotent operator N of order m, a co-isometry V1 ∈ L(DN ∗ ⊕ M, DT ∗), and an isometry V2 ∈ L(DT , DN ⊕ M). Then the characteristic function ΘT of T is a polynomial of degree less than or equal to m. It is important to note that the conclusion of Theorem 3.2 depends explicitly on the de- composition of T as used in the proof of Theorem 2.3. With the same setting as in Theorem 3.2, below we will show that the same conclusion holds for the following decomposition of T : T = (cid:20)S X−1 0 T−1(cid:21) = (cid:20)S DS ∗Γ−1DT−1 T−1 0 (cid:21) ∈ L(H0 ⊕ K−1), where K−1 = H0 ⊕ H−1, T−1 = (cid:20)N X 0 C(cid:21) = (cid:20)N DN ∗ΓDC C 0 (cid:21) ∈ L(H1 ⊕ H0), CONTRACTIONS WITH POLYNOMIAL CHARACTERISTIC FUNCTIONS 9 and X−1 = DS ∗Γ−1DT1, X = DN ∗ΓDC, and Γ−1 in L(DT−1, DS ∗) and Γ in L(DC, DN ∗) are a pair of contractions. In this case, again by Theorem 2.2, we have (3.1) and (3.2) where (3.3) and (3.4) ΘT = τ −1 −1∗(cid:20)ΘT−1 0 0 IDΓ∗ −1(cid:21) J[Γ−1](cid:20)0S 0 ΘT−1 = τ −1 ∗ (cid:20)0C 0 0 IDΓ∗(cid:21) J[Γ](cid:20)ΘN 0 0 IDΓ−1(cid:21) τ−1, 0 IDΓ(cid:21) τ, τ−1 : DT → {0H1} ⊕ DΓ−1, τ−1∗ : DT ∗ → DT ∗ −1 ⊕ DΓ∗ −1, τ : DT−1 → DN ⊕ DΓ, τ∗ : DT ∗ −1 → {0H−1} ⊕ DΓ∗, are unitary operators. Moreover J[Γ−1] = (cid:20) Γ∗ DΓ∗ −1 DΓ−1 −1 −Γ−1(cid:21) ∈ L(DS ∗ ⊕ DΓ−1, DT−1 ⊕ DΓ∗ −1), and By setting J[Γ] = (cid:20) Γ∗ DΓ DΓ∗ −Γ(cid:21) ∈ L(DN ∗ ⊕ DΓ, DC ⊕ DΓ∗). Ψ0 = (cid:20)0C 0 and using (3.1) and (3.2) we obtain 0 IDΓ∗(cid:21) and ΨN = (cid:20)ΘN 0 0 IDΓ(cid:21) , ΘT = τ −1 0 IDΓ∗ −1(cid:21) J[Γ−1](cid:20)0S 0 0 IDΓ∗ 0 IDΓ−1(cid:21) τ−1 −1(cid:21) J[Γ−1](cid:20)0S −1(cid:21)(cid:20)J[Γ] 0 IDΓ∗ 0 0 0 IDΓ∗ 0 IDΓ−1(cid:21) τ−1 −1(cid:21)(cid:20)ΨN 0 0 IDΓ∗ −1(cid:21)(cid:20)τ 0 0 IDΓ∗ −1(cid:21) J[Γ−1](cid:20)0S 0 0 IDΓ−1(cid:21) τ−1 0 = τ −1 = τ −1 −1∗(cid:20)ΘT−1 −1∗(cid:20)τ −1 −1∗(cid:20)τ −1 = V1(cid:20)ΨN = V1(cid:20)ΘN ∗ 0 0 0 ∗ Ψ0J[Γ]ΨN τ 0 0 0 IDΓ∗ −1(cid:21)(cid:20)Ψ0 −1(cid:21) V2 0 IDΓ∗ 0 −1(cid:21) V2, IDΓ⊕DΓ∗ 10 where and Hence FOIAS, PEARCY, AND SARKAR V1 = τ −1 ∗ 0 −1∗(cid:20)τ −1 −1∗(cid:20)τ −1 ∗ 0 = τ −1 0 IDΓ∗ 0 IDΓ∗ 0 −1(cid:21)(cid:20)Ψ0 −1(cid:21)(cid:20)0C 0 0 IDΓ∗ −1(cid:21)(cid:20)J[Γ] 0 0 IDΓ∗ −1(cid:21) 0 −1(cid:21)(cid:20)J[Γ] 0 0 IDΓ∗ −1(cid:21) , IDΓ∗ ⊕DΓ∗ V2 = (cid:20)τ 0 0 IDΓ∗ −1(cid:21) J[Γ−1](cid:20)0S 0 0 IDΓ−1(cid:21) τ−1. 0 ΘT = V1(cid:20)ΘN = V1(cid:20)ΘN 0 −1(cid:21) V2 , 0 IDΓ⊕DΓ∗ 0 I M(cid:21) V2 where M = DΓ ⊕ DΓ∗ isometric operators, that is, V1 V ∗ −1. Finally, by virtue of (3.3) and (3.4), we have that V ∗ V2 = IDT . Yet, we do not know if 1 = IDT ∗ and V ∗ 2 1 and V2 are where M is as in Theorem 3.2. dim M = dim M, Acknowledgement: The authors are grateful to the referee for pointing out that the formula- tion of Theorem 3.2 in the submitted version was inappropriate. The example in Remark 3.3 is due to the referee. The third author is supported in part by NBHM (National Board of Higher Mathematics, India) grant NBHM/R.P.64/2014. The third author is also grateful for hospitality of Texas A&M University, USA, during July 2016 and July-August 2015. References [1] B. Bagchi and G. Misra, Constant characteristic functions and homogeneous operators, J. Operator Theory 37 (1997), 51 -- 65. [2] L. de Branges and J. Rovnyak, Canonical models in quantum scattering theory, Perturbation Theory and its Applications in Quantum Mechanics, C.H. Wilcox (ed.), 295-392. Wiley, New York, 1966. [3] C. Foias and A. Frazho, The commutant lifting approach to interpolation problems, Operator Theory: Advances and Applications, 44. Birkhauser Verlag, Basel, 1990. [4] C. Foias and J. Sarkar, Contractions with polynomial characteristic functions I. Geometric approach, Transaction of American Math. Society, 364 (2012), 4127 -- 4153. [5] P. Halmos, A Hilbert space problem book, Graduate Texts in Mathematics, 19, Springer-Verlag, New York-Berlin, 1982. [6] B. Sz.-Nagy and C. Foias, Harmonic Analysis of Operators on Hilbert Space, North Holland, Amsterdam, 1970. [7] B. Sz.-Nagy and C. Foias, Forme triangulaire d'une contraction et factorisation de la fonction car- act´eristique, Acta Sci. Math. (Szeged) 28 (1967) 201-212. CONTRACTIONS WITH POLYNOMIAL CHARACTERISTIC FUNCTIONS 11 [8] N. Nikolskii and V. Vasyunin, Notes on two function models, in: The Bieberbach Conjecture, West Lafayette, IN, 1985, in: Math. Surveys Monogr., vol. 21, Amer. Math. Soc., Providence, RI, 1986, pp. 113-141. [9] R. Teodorescu, Fonctions caract´eristiques constantes, Acta Sci. Math. (Szeged) 38 (1976), no. 1-2, 183 -- 185. [10] P. Y. Wu, Contractions with constant characteristic function are reflexive, J. London Math. Soc. (2) 29 (1984), 533 -- 544. (Ciprian Foias) Department of Mathematics, Texas A&M University, College Station, Texas 77843, USA (Carl Pearcy) Department of Mathematics, Texas A&M University, College Station, Texas 77843, USA E-mail address: [email protected] (Jaydeb Sarkar) Indian Statistical Institute, Statistics and Mathematics Unit, 8th Mile, Mysore Road, Bangalore, 560059, India E-mail address: [email protected], [email protected]
1003.3382	1	1003	2010-03-17T14:41:58	Arens Regularity And The Topological Centers Of Module Actions	[ "math.FA" ]	In this article, for Banach left and right module actions, we will extend some propositions from Lau and $\ddot{U}lger$ into general situations and we establish the relationships between topological centers of module actions. We also introduce the new concepts as $Lw^w$-property and $Rw^w$-property for Banach $A-bimodule$ $B$ and we investigate the relations between them and topological center of module actions. We have some applications in dual groups.	math.FA	math	THE TOPOLOGICAL CENTERS OF MODULE ACTIONS KAZEM HAGHNEJAD AZAR AND ABDOLHAMID RIAZI Abstract. In this article, for Banach left and right module actions, we will extend some propositions from Lau and U lger into general situations and we establish the relationships between topological centers of module actions. We also introduce the new concepts as Lw∗w-property and Rw∗w-property for Banach A − bimodule B and we investigate the relations between them and topological center of module actions. We have some applications in dual groups. 1.Introduction and Preliminaries As is well-known [1], the second dual A∗∗ of A endowed with the either Arens mul- tiplications is a Banach algebra. The constructions of the two Arens multiplications in A∗∗ lead us to definition of topological centers for A∗∗ with respect both Arens multiplications. The topological centers of Banach algebras, module actions and ap- plications of them were introduced and discussed in [6, 8, 13, 14, 15, 16, 17, 21, 22], and they have attracted by some attentions. Now we introduce some notations and definitions that we used throughout this paper. Let A be a Banach algebra. We say that a net (eα)α∈I in A is a left approximate identity (= LAI) [resp. right approximate identity (= RAI)] if, for each a ∈ A, eαa −→ a [resp. aeα −→ a]. For a ∈ A and a′ ∈ A∗, we denote by a′a and aa′ respectively, the functionals on A∗ defined by < a′a, b >=< a′, ab >= a′(ab) and < aa′, b >=< a′, ba >= a′(ba) for all b ∈ A. The Banach algebra A is embedded in its second dual via the identification < a, a′ > - < a′, a > for every a ∈ A and a′ ∈ A∗. We denote the set {a′a : a ∈ A and a′ ∈ A∗} and {aa′ : a ∈ A and a′ ∈ A∗} by A∗A and AA∗, respectively, clearly these two sets are subsets of A∗. Let A has a BAI. If the equality A∗A = A∗, (AA∗ = A∗) holds, then we say that A∗ factors on the left (right). If both equalities A∗A = AA∗ = A∗ hold, then we say that A∗ factors on both sides. Let X, Y, Z be normed spaces and m : X × Y → Z be a bounded bilinear mapping. Arens in [1] offers two natural extensions m∗∗∗ and mt∗∗∗t of m from X ∗∗ × Y ∗∗ into Z ∗∗ as following: 1. m∗ : Z ∗ × X → Y ∗, given by < m∗(z ′, x), y >=< z ′, m(x, y) > where x ∈ X, y ∈ Y , z ′ ∈ Z ∗, 2. m∗∗ : Y ∗∗ × Z ∗ → X ∗, given by < m∗∗(y′′, z ′), x >=< y′′, m∗(z ′, x) > where x ∈ X, y′′ ∈ Y ∗∗, z ′ ∈ Z ∗, 3. m∗∗∗ : X ∗∗ × Y ∗∗ → Z ∗∗, given by < m∗∗∗(x′′, y′′), z ′ > =< x′′, m∗∗(y′′, z ′) > where x′′ ∈ X ∗∗, y′′ ∈ Y ∗∗, z ′ ∈ Z ∗. The mapping m∗∗∗ is the unique extension of m such that x′′ → m∗∗∗(x′′, y′′) from X ∗∗ into Z ∗∗ is weak∗ − to − weak∗ continuous for every y′′ ∈ Y ∗∗, but the mapping 2000 Mathematics Subject Classification. 46L06; 46L07; 46L10; 47L25. Key words and phrases. Arens regularity, bilinear mappings, Topological center, Second dual, Module action. 1 2 y′′ → m∗∗∗(x′′, y′′) is not in general weak∗ − to − weak∗ continuous from Y ∗∗ into Z ∗∗ unless x′′ ∈ X. Hence the first topological center of m may be defined as following Z1(m) = {x′′ ∈ X ∗∗ : y′′ → m∗∗∗(x′′, y′′) is weak∗ − to − weak∗ − continuous}. Let now mt : Y × X → Z be the transpose of m defined by mt(y, x) = m(x, y) for every x ∈ X and y ∈ Y . Then mt is a continuous bilinear map from Y × X to Z, and so it may be extended as above to mt∗∗∗ : Y ∗∗ × X ∗∗ → Z ∗∗. The mapping mt∗∗∗t : X ∗∗ × Y ∗∗ → Z ∗∗ in general is not equal to m∗∗∗, see [1], if m∗∗∗ = mt∗∗∗t, then m is called Arens regular. The mapping y′′ → mt∗∗∗t(x′′, y′′) is weak∗ − to − weak∗ continuous for every y′′ ∈ Y ∗∗, but the mapping x′′ → mt∗∗∗t(x′′, y′′) from X ∗∗ into Z ∗∗ is not in general weak∗ − to − weak∗ continuous for every y′′ ∈ Y ∗∗. So we define the second topological center of m as Z2(m) = {y′′ ∈ Y ∗∗ : x′′ → mt∗∗∗t(x′′, y′′) is weak∗ − to − weak∗ − continuous}. It is clear that m is Arens regular if and only if Z1(m) = X ∗∗ or Z2(m) = Y ∗∗. Arens regularity of m is equivalent to the following lim i lim j < z ′, m(xi, yj) >= lim j lim i < z ′, m(xi, yj) >, whenever both limits exist for all bounded sequences (xi)i ⊆ X , (yi)i ⊆ Y and z ′ ∈ Z ∗, see [6, 18]. The regularity of a normed algebra A is defined to be the regularity of its algebra multiplication when considered as a bilinear mapping. Let a′′ and b′′ be elements of A∗∗, the second dual of A. By Goldstin,s Theorem [6, P.424-425], there are nets (aα)α and (bβ)β in A such that a′′ = weak∗ − limα aα and b′′ = weak∗ − limβ bβ. So it is easy to see that for all a′ ∈ A∗, and lim α lim β < a′, m(aα, bβ) >=< a′′b′′, a′ > lim β lim α < a′, m(aα, bβ) >=< a′′ob′′, a′ >, where a′′b′′ and a′′ob′′ are the first and second Arens products of A∗∗, respectively, see [6, 14, 18]. The mapping m is left strongly Arens irregular if Z1(m) = X and m is right strongly Arens irregular if Z2(m) = Y . This paper is organized as follows. a) In section two, for a Banach A − bimodule, we have (1) a′′ ∈ ZB∗∗(A∗∗) if and only if π∗∗∗∗ (2) F ∈ ZB∗∗((A∗A)∗) if and only if π∗∗∗∗ (3) G ∈ Z(A∗A)∗(B ∗∗) if and only if π∗∗∗∗ (4) Let B has a BAI (eα)α ⊆ A such that eα r ℓ ℓ (b′, a′′) ∈ B ∗ for all b′ ∈ B ∗. (g, F ) ∈ B ∗ for all g ∈ B ∗. (g, G) ∈ A∗A for all g ∈ B ∗. e∗∗(B ∗∗) = B ∗∗ [ resp. Ze∗∗(B ∗∗) = B ∗∗] and B ∗ factors on the left [resp. right], but not on the right [resp. left], then ZB∗∗(A∗∗) 6= Z t B∗∗(A∗∗). (5) B ∗A ⊆ wapℓ(B) if and only if AA∗∗ ⊆ ZB∗∗(A∗∗). (6) Let b′ ∈ B ∗. Then b′ ∈ wapℓ(B) if and only if the adjoint of the mapping w∗ → e′′. Then if Z t π∗ ℓ (b′, ) : A → B ∗ is weak∗ − to − weak continuous. 3 b) In section three, for a Banach A − bimodule B, we define Lef t − weak∗ − to − weak property [=Rw∗w− property] and Right − weak∗ − to − weak property [=Rw∗w− property] for Banach algebra A and we show that (1) If A∗∗ = a0A∗∗ [resp. A∗∗ = A∗∗a0] for some a0 ∈ A and a0 has Rw∗w− property [resp. Lw∗w− property], then ZB∗∗ (A∗∗) = A∗∗. (2) If B ∗∗ = a0B ∗∗ [resp. B ∗∗ = B ∗∗a0] for some a0 ∈ A and a0 has Rw∗w− property [resp. Lw∗w− property] with respect to B, then ZA∗∗(B ∗∗) = B ∗∗. (3) If B ∗ factors on the left [resp. right] with respect to A and A has Rw∗w− property [resp. Lw∗w− property], then ZB∗∗ (A∗∗) = A∗∗. (4) If B ∗ factors on the left [resp. right] with respect to A and A has Rw∗w− property [resp. Lw∗w− property] with respect B, then ZA∗∗(B ∗∗) = B ∗∗. (5) If a0 ∈ A has Rw∗w− property with respect to B, then a0A∗∗ ⊆ ZB∗∗(A∗∗) and a0B ∗ ⊆ wapℓ(B). (6) Assume that AB ∗ ⊆ wapℓB. If B ∗ strong factors on the left [resp. right], then A has Lw∗w− property [resp. Rw∗w− property ] with respect to B. (7) Assume that AB ∗ ⊆ wapℓB. If B ∗ strong factors on the left [resp. right], then A has Lw∗w− property [resp. Rw∗w− property ] with respect to B. 2. The topological centers of module actions Let B be a Banach A − bimodule, and let πℓ : A × B → B and πr : B × A → B. be the left and right module actions of A on B. Then B ∗∗ is a Banach A∗∗ − bimodule with module actions π∗∗∗ ℓ : A∗∗ × B ∗∗ → B ∗∗ and π∗∗∗ r : B ∗∗ × A∗∗ → B ∗∗. Similarly, B ∗∗ is a Banach A∗∗ − bimodule with module actions πt∗∗∗t ℓ : A∗∗ × B ∗∗ → B ∗∗ and πt∗∗∗t r : B ∗∗ × A∗∗ → B ∗∗. We may therefore define the topological centers of the right and left module actions of A on B as follows: ZA∗∗(B ∗∗) = Z(πr) = {b′′ ∈ B ∗∗ : the map a′′ → π∗∗∗ r (b′′, a′′) : A∗∗ → B ∗∗ is weak∗ − to − weak∗ continuous} ZB∗∗(A∗∗) = Z(πℓ) = {a′′ ∈ A∗∗ : the map b′′ → π∗∗∗ ℓ (a′′, b′′) : B ∗∗ → B ∗∗ is weak∗ − to − weak∗ continuous} Z t A∗∗(B ∗∗) = Z(πt ℓ) = {b′′ ∈ B ∗∗ : the map a′′ → πt∗∗∗ ℓ (b′′, a′′) : A∗∗ → B ∗∗ is weak∗ − to − weak∗ continuous} Z t B∗∗(A∗∗) = Z(πt r) = {a′′ ∈ A∗∗ : the map b′′ → πt∗∗∗ r (a′′, b′′) : B ∗∗ → B ∗∗ is weak∗ − to − weak∗ continuous} We note also that if B is a left(resp. right) Banach A − module and πℓ : A × B → B (resp. πr : B × A → B) is left (resp. right) module action of A on B, then B ∗ is a right (resp. left) Banach A − module. 4 We write ab = πℓ(a, b), ba = πr(b, a), πℓ(a1a2, b) = πℓ(a1, a2b), πr(b, a1a2) = πr(ba1, a2), π∗ ℓ (a1b′, a2) = π∗ b′ ∈ B ∗ when there is no confusion. r (b′, ab), for all a1, a2, a ∈ A, b ∈ B and ℓ (b′, a2a1), π∗ r (b′a, b) = π∗ Theorem 2-1. We have the following assertions. (1) Assume that B is a Left Banach A − module. Then, a′′ ∈ ZB∗∗(A∗∗) if and only if π∗∗∗∗ ℓ (b′, a′′) ∈ B ∗ for all b′ ∈ B ∗. (2) Assume that B is a right Banach A − module. Then, b′′ ∈ ZA∗∗(B ∗∗) if and only if π∗∗∗∗ r (b′, b′′) ∈ A∗ for all b′ ∈ B ∗. Proof. (1) Let b′′ ∈ B ∗∗. Then, for every a′′ ∈ ZB∗∗ (A∗∗), we have < π∗∗∗∗ ℓ =< πt∗∗∗t ℓ (b′, a′′), b′′ >=< b′, π∗∗∗ ℓ (a′′, b′′), b′ >=< πt∗∗∗ (b′, a′′) = πt∗∗ ℓ It follow that π∗∗∗∗ Conversely, let a′′ ∈ A∗∗ and let π∗∗∗∗ all b′′ ∈ B ∗∗, we have ℓ ℓ ℓ (a′′, b′′) >=< π∗∗∗ ℓ (a′′, b′′), b′ > (b′′, a′′), b′ >=< b′′, πt∗∗ (a′′, b′) ∈ B ∗. ℓ (a′′, b′) > . (a′′, b′) ∈ B ∗ for all b′ ∈ B ∗. Then for < π∗∗∗ ℓ (a′′, b′′), b′ >=< b′, π∗∗∗ ℓ (a′′, b′′) >=< π∗∗∗∗ ℓ (b′, a′′), b′′ > =< πt∗∗ ℓ (a′′, b′), b′′ >=< b′′, πt∗∗ (a′′, b′) >=< πt∗∗∗ ℓ ℓ =< πt∗∗∗t (a′′, b′′), b′ > . ℓ Consequently a′′ ∈ ZB∗∗(A∗∗). (2) Prof is similar to (1). (b′′, a′′), b′ > (cid:3) Theorem 2-2. Assume that B is a Banach A−bimodule. Then we have the following assertions. (1) F ∈ ZB∗∗((A∗A)∗) if and only if π∗∗∗∗ (2) G ∈ Z(A∗A)∗(B ∗∗) if and only if π∗∗∗∗ ℓ r (g, F ) ∈ B ∗ for all g ∈ B ∗. (g, G) ∈ A∗A for all g ∈ B ∗. Proof. (1) Let F ∈ ZB∗∗((A∗A)∗) and (b′′ α)α ⊆ B ∗∗ such that b′′ α w∗ → b′′. Then for all g ∈ B ∗, we have < π∗∗∗∗ ℓ (g, F ), b′′ α >=< g, π∗∗∗ ℓ (F, b′′ α) >=< π∗∗∗ ℓ (F, b′′ α), g > →< π∗∗∗ ℓ (F, b′′), g >=< π∗∗∗∗ ℓ (g, F ), b′′ > . (g, F ) ∈ (B ∗∗, weak∗)∗ = B ∗. (g, F ) ∈ B ∗ for F ∈ (A∗A)∗ and g ∈ B ∗. Assume that Thus, we conclude that π∗∗∗∗ Conversely, let π∗∗∗∗ ℓ ℓ α)α ⊆ B ∗∗ such that b′′ b′′ ∈ B ∗∗ and (b′′ α α) >=< π∗∗∗∗ α), g >=< g, π∗∗∗ (F, b′′ (g, F ) >=< π∗∗∗∗ (g, F ) >→< b′′, π∗∗∗∗ w∗ → b′′. Then ℓ α, π∗∗∗∗ < π∗∗∗ =< b′′ (F, b′′ ℓ ℓ ℓ ℓ =< π∗∗∗ It follow that F ∈ ZB∗∗((A∗A)∗). ℓ (2) Proof is similar to (1). ℓ (F, b′′), g > . (g, F ), b′′ α > (g, F ), b′′ > (cid:3) 5 In the proceeding theorems, if we take B = A, we obtain some parts of Lemma 3.1 from [14]. An element e′′ of A∗∗ is said to be a mixed unit if e′′ is a right unit for the first Arens multiplication and a left unit for the second Arens multiplication. That is, e′′ is a mixed unit if and only if, for each a′′ ∈ A∗∗, a′′e′′ = e′′oa′′ = a′′. By [4, p.146], an element e′′ of A∗∗ is mixed unit if and only if it is a weak∗ cluster point of some BAI (eα)α∈I in A. Let B be a Banach A − bimodule and a′′ ∈ A∗∗. We define the locally topological center of the left and right module actions of a′′ on B, respectively, as follows a′′ (B ∗∗) = Z t Z t a′′(πt Za′′ (B ∗∗) = Za′′ (πt ℓ) = {b′′ ∈ B ∗∗ : πt∗∗∗t r) = {b′′ ∈ B ∗∗ : πt∗∗∗t r ℓ (a′′, b′′) = π∗∗∗ (b′′, a′′) = π∗∗∗ ℓ r (a′′, b′′)}, (b′′, a′′)}. Thus we have \ a′′ ∈A∗∗ \ a′′ ∈A∗∗ Z t a′′(B ∗∗) = Z t A(B ∗∗) = Z(πt r), Za′′(B ∗∗) = ZA(B ∗∗) = Z(πr). Definition 2-3. Let B be a left Banach A − module and e′′ ∈ A∗∗ be a mixed unit for A∗∗. We say that e′′ is a left mixed unit for B ∗∗, if π∗∗∗ ℓ (e′′, b′′) = πt∗∗∗t ℓ (e′′, b′′) = b′′, for all b′′ ∈ B ∗∗. The definition of right mixed unit for B ∗∗ is similar. B ∗∗ has a mixed unit if it has left and right mixed unit that are equal. It is clear that if e′′ ∈ A∗∗ is a left (resp. right) unit for B ∗∗ and Ze′′ (B ∗∗) = B ∗∗, then e′′ is left (resp. right) mixed unit for B ∗∗. w∗ → e′′. e∗∗ (B ∗∗) = B ∗∗ [ resp. Ze∗∗ (B ∗∗) = B ∗∗] and B ∗ factors on the left [resp. Theorem 2-4. Let B be a Banach A−bimodule with a BAI (eα)α such that eα Then if Z t right], but not on the right [resp. left], then ZB∗∗(A∗∗) 6= Z t B∗∗(A∗∗). Proof. Suppose that B ∗ factors on the left with respect to A, but not on the right. w∗ → e′′. Thus for all b′ ∈ B ∗ there are Let (eα)α ⊆ A be a BAI for A such that eα a ∈ A and x′ ∈ B ∗ such that x′a = b′. Then for all b′′ ∈ B ∗∗ we have < π∗∗∗ ℓ (e′′, b′′), b′ >=< e′′, π∗∗ ℓ (b′′, b′) >= lim α < π∗∗ ℓ (b′′, b′), eα > = lim α = lim α ℓ (b′, eα) >= lim α ℓ (x′, aeα) >= lim α < b′′, π∗ ℓ (x′a, eα) > < π∗∗ ℓ (b′′, x′), aeα > < b′′, π∗ < b′′, π∗ =< π∗∗ ℓ (b′′, x′), a >=< b′′, b′ > . Thus π∗∗∗ e′′ ∈ ZB∗∗(A∗∗). If we take ZB∗∗(A∗∗) = Z t (e′′, b′′) = b′′ consequently B ∗∗ has left unit A∗∗ − module. It follows that B∗∗(A∗∗). Then the B∗∗(A∗∗), then e′′ ∈ Z t ℓ 6 mapping b′′ → πt∗∗∗t w∗ → e′′, πt∗∗∗t Since eα r (b′′, e′′) is weak∗ − to − weak∗ continuous from B ∗∗ into B ∗∗. (b′′, eα) w∗ (b′′, e′′). Let b′ ∈ B ∗ and (bβ)β ⊆ B such that → πt∗∗∗t r w∗ → b′′. Since Z t e∗∗(B ∗∗) = B ∗∗, we have the following quality r bβ < πt∗∗∗t r (b′′, e′′), b′ >= lim α < πt∗∗∗t r (b′′, eα), b′ >= lim α < πt∗∗∗ r (eα, b′′), b′ > = lim α lim β < πt∗∗∗ r (eα, bβ), b′ >= lim α lim β < πr(bβ, eα), b′ > = lim α lim β < b′, πr(bβ, eα) >= lim β lim α < b′, πr(bβ, eα) > = lim β < b′, bβ >=< b′′, b′ > . Thus πt∗∗∗t that b′′ ∈ B ∗∗ and (bβ)β ⊆ B such that bβ (b′′, e′′) = π∗∗∗ r r (b′′, e′′) = b′′. It follows that B ′′ has a right unit. Suppose w∗ → b′′. Then for all b′ ∈ B ∗ we have < b′′, b′ >=< π∗∗∗ r (b′′, e′′), b′ >=< b′′, π∗∗ r (e′′, b′) >= lim β < π∗∗ r (e′′, b′), bβ > = lim β < e′′, π∗ r (b′, bβ) >= lim β lim α < π∗ r (b′, bβ), eα > = lim β lim α < π∗ r (b′, bβ), eα >= lim β lim α < b′, πr(bβ, eα) > = lim α lim β < π∗∗∗ r (bβ, eα), b′ >= lim α lim β < bβ, π∗∗ r (eα, b′) > = lim α < b′′, π∗∗ r (eα, b′) > . It follows that weak − limα π∗∗ factors on the right that is contradiction. r (eα, b′) = b′. So by Cohen Factorization Theorem, B ∗ (cid:3) Corollary 2-5. Let B be a Banach A − bimodule and e′′ ∈ A∗∗ be a left mixed unit for B ∗∗. If B ∗ factors on the left, but not on the right, then ZB∗∗(A∗∗) 6= Z t B∗∗(A∗∗). In the proceeding corollary, if we take B = A, then it is clear Z t so we obtain Proposition 2.10 from [14]. e∗∗ (A∗∗) = A∗∗, and Theorem 2-6. Suppose that B is a weakly complete Banach space. Then we have the following assertions. (1) Let B be a Left Banach A − module and e′′ be a left mixed unit for B ∗∗. If AB ∗∗ ⊆ B, then B is reflexive. (2) Let B be a right Banach A − module and e′′ be a right mixed unit for B ∗∗. If ZA∗∗(B ∗∗)A ⊆ B, then ZA∗∗(B ∗∗) = B. Proof. (1) Assume that b′′ ∈ B ∗∗. Since e′′ is also mixed unit for A∗∗, there is a BAI (eα)α ⊆ A for A such that eα b′′ in B ∗∗. Since AB ∗∗ ⊆ B, we have π∗∗∗ π∗∗∗ and so B is reflexive. (e′′, b′′) = (eα, b′′) ∈ B. Consequently (e′′, b′′) = b′′ in B. Since B is a weakly complete, b′′ ∈ B, (eα, b′′) w→ π∗∗∗ → π∗∗∗ ℓ ℓ ℓ w∗ → e′′. Then π∗∗∗ (eα, b′′) w∗ ℓ ℓ 7 (2) Since b′′ ∈ ZA∗∗(B ∗∗), we have π∗∗∗ ZA∗∗(B ∗∗)A ⊆ B, π∗∗∗ π∗∗∗ (b′′, e′′) = b′′ in B ∗∗. Since (b′′, eα) w→ (b′′, e′′) = b′′ in B. It follows that b′′ ∈ B, since B is a weakly complete. (cid:3) (b′′, eα) ∈ B. Consequently we have π∗∗∗ → π∗∗∗ r r r r r (b′′, eα) w∗ A functional a′ in A∗ is said to be wap (weakly almost periodic) on A if the mapping a → a′a from A into A∗ is weakly compact. The procceding definition to the equiva- lent following condition, see [6, 14, 18]. For any two net (aα)α and (bβ)β in {a ∈ A : k a k≤ 1}, we have limαlimβ < a′, aαbβ >= limβlimα < a′, aαbβ >, whenever both iterated limits exist. The collection of all wap functionals on A is de- noted by wap(A). Also we have a′ ∈ wap(A) if and only if < a′′b′′, a′ >=< a′′ob′′, a′ > for every a′′, b′′ ∈ A∗∗. Definition 2-7. Let B be a left Banach A − module. Then, b′ ∈ B ∗ is said to be left weakly almost periodic functional if the set {πℓ(b′, a) : a ∈ A, k a k≤ 1} is relatively weakly compact. We denote by wapℓ(B) the closed subspace of B ∗ consisting of all the left weakly almost periodic functionals in B ∗. The definition of the right weakly almost periodic functional (= wapr(B)) is the same. By [18], the definition of wapℓ(B) is equivalent to the following < π∗∗∗ ℓ (a′′, b′′), b′ >=< πt∗∗∗t ℓ (a′′, b′′), b′ > for all a′′ ∈ A∗∗ and b′′ ∈ B ∗∗. Thus, we can write wapℓ(B) = {b′ ∈ B ∗ : < π∗∗∗ ℓ (a′′, b′′), b′ >=< πt∗∗∗t ℓ (a′′, b′′), b′ > f or all a′′ ∈ A∗∗, b′′ ∈ B ∗∗}. Theorem 2-8. Suppose that B is a left Banach A − module. Consider the following statements. (1) B ∗A ⊆ wapℓ(B). (2) AA∗∗ ⊆ ZB∗∗(A∗∗). (3) AA∗∗ ⊆ AZB∗∗((A∗A)∗). Then, we have (1) ⇔ (2) ⇐ (3). Proof. (1) ⇒ (2) Let (b′′ α)α ⊆ B ∗∗ such that b′′ α w∗ → b′′. Then for all a ∈ A and a′′ ∈ A∗∗, we have < π∗∗∗ ℓ (aa′′, b′′ =< a′′, π∗∗ ℓ (b′′ α, b′)a > ℓ (b′′ (a′′, b′′), b′a) > ℓ (b′′ α), b′ >=< aa′′, π∗∗ (a′′, b′′ (aa′′, b′′), b′) > . α, b′a) >=< π∗∗∗ =< π∗∗∗ ℓ ℓ α, b′) >=< a′′, π∗∗ α), b′a) >→< π∗∗∗ ℓ Hence aa′′ ∈ ZB∗∗ (A∗∗). (2) ⇒ (1) Let a ∈ A and b′ ∈ B ∗. Then < π∗∗∗ ℓ (a′′, b′′ α), b′a >=< aπ∗∗∗ ℓ (a′′, b′′ α), b′ >=< π∗∗∗ ℓ (aa′′, b′′ α), b′ > 8 =< πt∗∗∗t It follow that b′a ∈ wapℓ(B). (3) ⇒ (2) Since AZB∗∗((A∗A)∗) ⊆ ZB∗∗(A∗∗), proof is hold. α), b′ >=< πt∗∗∗t (aa′′, b′′ ℓ ℓ (a′′, b′′ α), b′a > . (cid:3) In the proceeding theorem, if we take B = A, then we obtain Theorem 3.6 from [14] and the same as proceeding theorem, we can claim the following assertions: If B is a right Banach A − module, then for the following statements we have (1) ⇔ (2) ⇐ (3). (1) AB ∗ ⊆ wapr(B). (2) A∗∗A ⊆ ZB∗∗(A∗∗). (3) A∗∗A ⊆ ZB∗∗((A∗A)∗)A. The proof of the this assertion is similar to proof of Theorem 2-8. Corollary 2-9. Suppose that B is a Banach A − bimodule. Then if A is a left [resp. right] ideal in A∗∗, then B ∗A ⊆ wapℓ(B) [resp. AB ∗ ⊆ wapr(B)]. Example 2-10. Suppose that 1 ≤ p ≤ ∞ and q is conjugate of p. We know that if G is compact, then L1(G) is a two-sided ideal in its second dual of it. By proceeding Theorem we have Lq(G) ∗ L1(G) ⊆ wapℓ(Lp(G)) and L1(G) ∗ Lq(G) ⊆ wapr(Lp(G)). Also if G is finite, then Lq(G) ⊆ wapℓ(Lp(G))∩wapr(Lp(G)). Hence we conclude that ZL1(G)∗∗(Lp(G)∗∗) = Lp(G) and ZLp(G)∗∗(L1(G)∗∗) = L1(G). Theorem 2-11. We have the following assertions. (1) Suppose that B is a left Banach A − module and b′ ∈ B ∗. Then b′ ∈ wapℓ(B) ℓ (b′, ) : A → B ∗ is weak∗−to−weak if and only if the adjoint of the mapping π∗ continuous. (2) Suppose that B is a right Banach A−module and b′ ∈ B ∗. Then b′ ∈ wapr(B) r (b′, ) : B → A∗ is weak∗−to−weak if and only if the adjoint of the mapping π∗ continuous. Proof. (1) Assume that b′ ∈ wapℓ(B) and π∗ ℓ (b′, )∗ : B ∗∗ → A∗ is the adjoint of ℓ (b′, ). Then for every b′′ ∈ B ∗∗ and a ∈ A, we have π∗ ℓ (b′, a) > . ℓ (b′, )∗b′′, a >=< b′′, π∗ < π∗ Suppose (b′′ α)α ⊆ B ∗∗ such that b′′ α w∗ → b′′ and a′′ ∈ A∗∗ and (aβ)β ⊆ A such that aβ w∗ → a′′. By easy calculation, for all y′′ ∈ B ∗∗ and y′ ∈ B ∗, we have < π∗ ℓ (y′, )∗, y′′ >= π∗∗ ℓ (y′′, y′). Since b′ ∈ wapℓ(B), < π∗∗∗ ℓ (a′′, b′′ α), b′ >→< π∗∗∗ ℓ (a′′, b′′), b′ > . Then we have the following statements lim α < a′′, π∗ ℓ (b′, )∗b′′ α >= lim α < a′′, π∗∗ ℓ (b′′ α, b′) > = lim α < π∗∗∗ ℓ (a′′, b′′ α), b′ >=< π∗∗∗ ℓ 9 (a′′, b′′), b′ > =< a′′, π∗ ℓ (b′, )∗b′′ > . It follow that the adjoint of the mapping π∗ weak continuous. Conversely, let the adjoint of the mapping π∗ weak continuous. Suppose (b′′ for every a′′ ∈ A∗∗, we have ℓ (b′, ) : A → B ∗ is weak∗ − to − α)α ⊆ B ∗∗ such that b′′ α ℓ (b′, ) : A → B ∗ is weak∗ − to − w∗ → b′′ and b′ ∈ B ∗. Then lim α < π∗∗∗ ℓ (a′′, b′′ α), b′ >= lim α < a′′, π∗∗ ℓ (b′′ α, b′) > = lim α < a′′, π∗ ℓ (b′, )∗b′′ α >=< a′′, π∗ ℓ (b′, )∗b′′ >=< π∗∗∗ ℓ It follow that b′ ∈ wapℓ(B). (2) proof is similar to (1). (a′′, b′′), b′ > . (cid:3) Corollary 2-12. Let A be a Banach algebra. Assume that a′ ∈ A∗ and Ta′ is the linear operator from A into A∗ defined by Ta′ a = a′a. Then, a′ ∈ wap(A) if and only if the adjoint of Ta′ is weak∗ − to − weak continuous. So A is Arens regular if and only if the adjoint of the mapping Ta′ a = a′a is weak∗ − to − weak continuous for every a′ ∈ A∗. 3. Lw∗w-property and Rw∗w-property In this section, we introduce the new definition as Lef t − weak∗ − to − weak property and Right−weak∗−to−weak property for Banach algebra A and make some relations between these concepts and topological centers of module actions. As some conclu- sion, we have ZL1(G)∗∗(M (G)∗∗) 6= M (G)∗∗ where G is a locally compact group. If G is finite, we have ZM(G)∗∗ (L1(G)∗∗) = L1(G)∗∗ and ZL1(G)∗∗(M (G)∗∗) = M (G)∗∗. Definition 3-1. Let B be a left Banach A − module. We say that a ∈ A has Lef t − weak∗ − to − weak property (= Lw∗w− property) with respect to B, if for w∗ → 0 implies ab′ α w→ 0. If every a ∈ A has Lw∗w− property all (bα)α ⊆ B ∗, ab′ α with respect to B, then we say that A has Lw∗w− property with respect to B. The definition of the Right − weak∗ − to − weak property (= Rw∗w− property) is the same. We say that a ∈ A has weak∗ − to − weak property (= w∗w− property) with respect to B if it has Lw∗w− property and Rw∗w− property with respect to B. If a ∈ A has Lw∗w− property with respect to itself, then we say that a ∈ A has Lw∗w− property. For proceeding definition, we have some examples and remarks as follows. a) If B is Banach A-bimodule and reflexive, then A has w∗w−property with respect to B. Then i) L1(G), M (G) and A(G) have w∗w−property when G is finite. ii) Let G be locally compact group. L1(G) [resp. M (G)] has w∗w−property [resp. 10 Lw∗w− property ]with respect to Lp(G) whenever p > 1. b) Suppose that B is a left Banach A − module and e is left unit element of A such that eb = b for all b ∈ B. If e has Lw∗w− property, then B is reflexive. c) If S is a compact semigroup, then C+(S) = {f ∈ C(S) : f > 0} has w∗w−property. Theorem 3-2. Suppose that B is a Banach A − bimodule. Then we have the following assertions. (1) If A∗∗ = a0A∗∗ [resp. A∗∗ = A∗∗a0] for some a0 ∈ A and a0 has Rw∗w− property [resp. Lw∗w− property], then ZB∗∗ (A∗∗) = A∗∗. (2) If B ∗∗ = a0B ∗∗ [resp. B ∗∗ = B ∗∗a0] for some a0 ∈ A and a0 has Rw∗w− property [resp. Lw∗w− property] with respect to B, then ZA∗∗(B ∗∗) = B ∗∗. Proof. (1) Suppose that A∗∗ = a0A∗∗ for some a0 ∈ A and a0 has Rw∗w− w∗ → b′′. Then for all a ∈ A and α)α ⊆ B ∗∗ such that b′′ α property. Let (b′′ b′ ∈ B ∗, we have α, π∗ α, b′) w∗ < π∗∗ ℓ (b′′ α, b′), a >=< b′′ ℓ (b′, a) >→< b′′, π∗ ℓ (b′, a) >=< π∗∗ ℓ (b′′, b′), a >, ℓ (b′′ w∗ α, b′)a0 it follow that π∗∗ → ℓ (b′′, b′)a0. ℓ (b′′, b′)a0. Since a0 has Rw∗w− property, π∗∗ π∗∗ Now let a′′ ∈ A∗∗. Then there is x′′ ∈ A∗∗ such that a′′ = a0x′′ consequently we have ℓ (b′′, b′). Also we can write π∗∗ ℓ (b′′ w→ π∗∗ α, b′)a0 → π∗∗ ℓ (b′′ < π∗∗∗ ℓ (a′′, b′′ α), b′ >=< a′′, π∗∗ ℓ (b′′ α, b′) >=< x′′, π∗∗ α, b′)a0 > →< x′′, π∗∗ ℓ (b′′, b′)a0 >=< π∗∗∗ ℓ (a′′, b′′ ℓ (b′′ α), b′ > . We conclude that a′′ ∈ ZB∗∗(A∗∗). Proof of the next part is the same as the proceeding proof. (2) Let B ∗∗ = a0B ∗∗ for some a0 ∈ A and a0 has Rw∗w− property with respect w∗ → a′′. Then for all b ∈ B, we α)α ⊆ A∗∗ such that a′′ α to B. Assume that (a′′ have < π∗∗ r (a′′ α, b′), b >=< a′′ α, π∗∗ r (b′, b) >→< a′′, π∗∗ r (b′, b) >=< π∗∗ r (a′′, b′), b > . α, b′) w∗ r (a′′ → π∗∗ We conclude that π∗∗ π∗∗ r (a′′, b′)a0. Since a0 has Rw∗w− property with respect to B, π∗∗ π∗∗ r (a′′, b′)a0. Now let b′′ ∈ B ∗∗. Then there is x′′ ∈ B ∗∗ such that b′′ = a0x′′. Hence, we have r (a′′, b′) then we have π∗∗ α, b′)a0 α, b′)a0 r (a′′ r (a′′ w∗ → w→ < π∗∗∗ r (b′′, a′′ =< x′′, π∗∗ r (a′′ r (a′′ α), b′ >=< b′′, π∗∗ α, b′)a0 >→< x′′, π∗∗ α, b′) >=< a0x′′, π∗∗ r (a′′ r (a′′, b′)a0 >=< b′′, π∗∗ α, b′) > r (a′′, b′) > =< π∗∗∗ r (b′′, a′′), b′ > . It follow that b′′ ∈ ZA∗∗(B ∗∗). The next part is similar to the proceeding proof. (cid:3) 11 i) Let G be a locally compact group. Since M (G) is a Banach Example 3-3. L1(G)-bimodule and the unit element of M (G) has not Lw∗w− property or Rw∗w− property, by Theorem 2-3, ZL1(G)∗∗(M (G)∗∗) 6= M (G)∗∗. ii) If G is finite, then by Theorem 2-3, we have ZM(G)∗∗(L1(G)∗∗) = L1(G)∗∗ and ZL1(G)∗∗(M (G)∗∗) = M (G)∗∗. Assume that B is a Banach A − bimodule. We say that B factors on the left (right) with respect to A if B = BA (B = AB). We say that B factors on both sides, if B = BA = AB. Theorem 3-4. Suppose that B is a Banach A − bimodule and A has a BAI. Then we have the following assertions. (1) If B ∗ factors on the left [resp. right] with respect to A and A has Rw∗w− property [resp. Lw∗w− property], then ZB∗∗ (A∗∗) = A∗∗. (2) If B ∗ factors on the left [resp. right] with respect to A and A has Rw∗w− property [resp. Lw∗w− property] with respect B, then ZA∗∗(B ∗∗) = B ∗∗. Proof. (1) Assume that B ∗ factors on the left and A has Rw∗w− property. Let α)α ⊆ B ∗∗ such that b′′ α w∗ → b′′. Since B ∗A = B ∗, for all b′ ∈ B ∗ there are (b′′ x ∈ A and y′ ∈ B ∗ such that b′ = y′x. Then for all a ∈ A, we have < π∗∗ =< b′′ ℓ (b′′ α, π∗ α, y′)x, a >=< b′′ α, π∗ ℓ (b′, a) >→< b′′, π∗ ℓ (y′, a)x >=< π∗∗ ℓ (b′, a) >=< π∗∗ α, b′), a > ℓ (b′′, y′)x, a > . ℓ (b′′ Thus, we conclude that π∗∗ property, π∗∗ ℓ (b′′ w→< π∗∗ ℓ (b′′ (a′′, b′′ α, y′)x α), b′ >=< a′′, π∗∗ ℓ (b′′, y′)x. Now let b′′ ∈ A∗∗. Then α, y′)x > α, b′) >=< a′′, π∗∗ ℓ (b′′ < π∗∗∗ ℓ α, y′)x w∗ →< π∗∗ ℓ (b′′, y′)x. Since A has Rw∗w− →< a′′, π∗∗ ℓ (b′′, y′)x >=< π∗∗∗ ℓ ℓ (b′′ (a′′, b′′), b′ > . It follow that a′′ ∈ ZB∗∗(A∗∗) = A∗∗. If B ∗ factors on the right and A has Lw∗w− property, then proof is the same as preceding proof. (2) Let B ∗ factors on the left with respect to A and A has Rw∗w− property with w∗ respect to B. Assume that (a′′ → a′′. Since B ∗A = B, for all b′ ∈ B ∗ there are x ∈ A and y′ ∈ B ∗ such that b′ = y′x. Then for all b ∈ B, we have α)α ⊆ A∗∗ such that a′′ α < π∗∗ r (a′′ α, y′)x, b >=< π∗∗ =< a′′, π∗ r (a′′ r (b′, b) >=< π∗∗ α, b′), b >=< a′′ α, π∗ r (a′′, y′)x, b > . r (b′, b) > Consequently π∗∗ respect to B, π∗∗ have r (a′′ r (a′′ α, y′)x α, y′)x w∗ → π∗∗ w→ π∗∗ r (a′′, y′)x. Since A has Rw∗w− property with r (a′′, y′)x. It follow that for all b′′ ∈ B ∗∗, we < π∗∗∗ r (b′′, a′′ α), b′ >=< b′′, π∗∗ =< π∗∗∗ r (a′′ (b′′, a′′), b′ > . r α, y′)x >→< b′′, π∗∗ r (a′′, y′)x > 12 Thus we conclude that b′′ ∈ ZA∗∗(B ∗∗). The proof of the next assertions is the same as proceeding proof. (cid:3) Theorem 3-5. Suppose that B is a Banach A−bimodule. Then we have the following assertions. (1) If a0 ∈ A has Rw∗w− property with respect to B, then a0A∗∗ ⊆ ZB∗∗(A∗∗) and a0B ∗ ⊆ wapℓ(B). (2) If a0 ∈ A has Lw∗w− property with respect to B, then A∗∗a0 ⊆ ZB∗∗(A∗∗) and B ∗a0 ⊆ wapℓ(B). (3) If a0 ∈ A has Rw∗w− property with respect to B, then a0B ∗∗ ⊆ ZA∗∗(B ∗∗) and B ∗a0 ⊆ wapr(B). (4) If a0 ∈ A has Lw∗w− property with respect to B, then B ∗∗a0 ⊆ ZA∗∗(B ∗∗) and a0B ∗ ⊆ wapr(B). Proof. (1) Let (b′′ α)α ⊆ B ∗∗ such that b′′ α w∗ → b′′. Then for all a ∈ A and b′ ∈ B ∗, we have < π∗∗ ℓ (b′′ α, b′)a0, a >=< π∗∗ ℓ (b′′ α, b′), a0a >=< b′′ α, π∗ ℓ (b′, a0a) > →< b′′, π∗ ℓ (b′, a0a) >=< π∗∗ ℓ (b′′, b′)a0, a > . w∗ → π∗∗ ℓ (b′′ It follow that π∗∗ with respect to B, π∗∗ We conclude that a0a′′ ∈ ZB∗∗(A∗∗) so that a0A∗∗ ∈ ZB∗∗(A∗∗). Since ℓ (b′′, b′)a0 = π∗∗ π∗∗ ℓ (b′′, b′)a0. Since a0 has Rw∗w− property w→ π∗∗ ℓ (b′′, b′a0), a0B ∗ ⊆ wapℓ(B). α, b′)a0 ℓ (b′′ ℓ (b′′, b′)a0. α, b′)a0 (2) proof is similar to (1). (3) Assume that (a′′ α)α ⊆ A∗∗ such that a′′ α w∗ → a′′. Let b ∈ B and b′ ∈ B ∗. Then we have < π∗∗ r (a′′ α, b′)a0, b >=< π∗∗ r (a′′ α, b′), a0b >=< a′′ α, π∗ r (b′, a0b) > →< a′′, π∗ r (b′, a0b) >=< π∗∗ r (a′′, b′)a0, b > . Thus we conclude π∗∗ erty with respect to B, π∗∗ have r (a′′ α, b′)a0 r (a′′ r (a′′, b′)a0. Since a0 has Rw∗w− prop- r (a′′, b′)a0. If b′′ ∈ B ∗∗, then we w→ π∗∗ w∗ → π∗∗ α, b′)a0 < π∗∗∗ r (a0b′′, a′′ α), b′ >=< a0b′′, π∗∗∗ r (a′′ α, b′) >=< b′′, π∗∗ r (a′′ α, b′)a0 > =< b′′, π∗∗ r (a′′ α, b′)a0 >=< π∗∗∗ r (a0b′′, a′′), b′ > . It follow that a0b′′ ∈ ZA∗∗(B ∗∗). Consequently we have a0B ∗∗ ∈ ZA∗∗(B ∗∗). The proof of the next assertion is clear. (4) Proof is similar to (3). (cid:3) 13 Theorem 3-6. Let B be a Banach A − bimodule. Then we have the following assertions. (1) Suppose lim α lim β < b′ β, bα >= lim β lim α < b′ β, bα >, for every (bα)α ⊆ B and (b′ Rw∗w− property with respect to B. β)β ⊆ B ∗. Then A has Lw∗w− property and (2) If for some a ∈ A, lim α lim β < ab′ β, bα >= lim β lim α < ab′ β, bα >, for every (bα)α ⊆ B and (b′ respect to B. Also if for some a ∈ A, β)β ⊆ B ∗, then a has Rw∗w− property with lim α lim β < b′ βa, bα >= lim β lim α < b′ βa, bα >, for every (bα)α ⊆ B and (b′ respect to B. β)β ⊆ B ∗, then a has Lw∗w− property with Proof. (1) Assume that a ∈ A such that ab′ β w∗ → 0 where (b′ β)β ⊆ B ∗. Let b′′ ∈ B ∗∗ and (bα)α ⊆ B such that bα w∗ → b′′. Then lim β < b′′, ab′ β >= lim β lim α < bα, ab′ β >= lim β lim α < ab′ β, bα > = lim α lim β < ab′ β, bα >= 0. We conclude that ab′ β has Rw∗w− property. w→ 0, so A has Lw∗w− property. It also easy that A (2) Proof is easy and is the same as (1). (cid:3) α = b′aα [resp. b′ Definition 3-7. Let B be a left Banach A − module. We say that B ∗ strong factors on the left [resp. right] if for all (b′ α)α ⊆ B ∗ there are (aα)α ⊆ A and b′ ∈ B ∗ such that b′ If B ∗ strong factors on the left and right, then we say that B ∗ strong factors on the both side. It is clear that if B ∗ strong factors on the left [resp. right], then B ∗ factors on the left [resp. right]. α = aαb′] where (aα)α has limit the weak∗ topology in A∗∗. Theorem 3-8. Suppose that B is a Banach A − bimodule. Assume that AB ∗ ⊆ wapℓB. If B ∗ strong factors on the left [resp. right], then A has Lw∗w− property [resp. Rw∗w− property ] with respect to B. Proof. Let (b′ are (aα)α ⊆ A and b′ ∈ B ∗ such that b′ α)α ⊆ B ∗ such that ab′ α w∗ → 0. Since B ∗ strong factors on the left, there α = b′aα. Let b′′ ∈ B ∗∗ and (bβ)β ⊆ B such 14 that bβ w∗ → b′′. Then we have lim α < b′′, ab′ α >= lim α lim β < bβ, ab′ α >= lim α lim β < ab′ α, bβ > = lim α lim β < ab′aα, bβ >= lim α lim β < ab′, aαbβ > = lim β lim α < ab′, aαbβ >= lim β lim α < ab′ α, bβ >= 0 It follow that ab′ α w→ 0. Problems . (cid:3) (1) Suppose that B is a Banach A − bimodule. If B is left or right factors with respect to A, dose A has Lw∗w−property or Rw∗w−property, respectively? (2) Suppose that B is a Banach A − bimodule. Let A has Lw∗w−property with respect to B. Dose ZB∗∗(A∗∗) = A∗∗? References 1. R. E. Arens, The adjoint of a bilinear operation, Proc. Amer. Math. Soc. 2 (1951), 839-848. 2. N. Arikan, A simple condition ensuring the Arens regularity of bilinear mappings, Proc. Amer. Math. Soc. 84 (4) (1982), 525-532. 3. J. Baker, A.T. Lau, J.S. Pym Module homomorphism and topological centers associated with weakly sequentially compact Banach algebras, Journal of Functional Analysis. 158 (1998), 186- 208. 4. F. F. Bonsall, J. Duncan, Complete normed algebras, Springer-Verlag, Berlin 1973. 5. H. G. Dales, A. Rodrigues-Palacios, M.V. Velasco, The second transpose of a derivation, J. London. Math. Soc. 2 64 (2001) 707-721. 6. H. G. Dales, Banach algebra and automatic continuity, Oxford 2000. 7. N. Dunford, J. T. Schwartz, Linear operators.I, Wiley, New york 1958. 8. M. Eshaghi Gordji, M. Filali, Arens regularity of module actions, Studia Math. 181 3 (2007), 237-254. 9. M. Eshaghi Gordji, M. Filali, Weak amenability of the second dual of a Banach algebra, Studia Math. 182 3 (2007), 205-213. 10. K. Haghnejad Azar, A. Riazi, Arens regularity of bilinear forms and unital Banach module space, (Submitted). 11. E. Hewitt, K. A. Ross, Abstract harmonic analysis, Springer, Berlin, Vol I 1963. 12. E. Hewitt, K. A. Ross, Abstract harmonic analysis, Springer, Berlin, Vol II 1970. 13. A. T. Lau, V. Losert, On the second Conjugate Algebra of locally compact groups, J. London Math. Soc. 37 (2)(1988), 464-480. 14. A. T. Lau, A. U lger, Topological center of certain dual algebras, Trans. Amer. Math. Soc. 348 (1996), 1191-1212. 15. S. Mohamadzadih, H. R. E. Vishki, Arens regularity of module actions and the second adjoint of a derivation, Bulletin of the Australian Mathematical Society 77 (2008), 465-476. 16. M. Neufang, Solution to a conjecture by Hofmeier-Wittstock, Journal of Functional Analysis. 217 (2004), 171-180. 17. M. Neufang, On a conjecture by Ghahramani-Lau and related problem concerning topological center, Journal of Functional Analysis. 224 (2005), 217-229. 18. J. S. Pym, The convolution of functionals on spaces of bounded functions, Proc. London Math Soc. 15 (1965), 84-104. 19. A. U lger, Arens regularity sometimes implies the RNP, Pacific Journal of Math. 143 (1990), 377-399. 20. A. U lger, Some stability properties of Arens regular bilinear operators, Proc. Amer. Math. Soc. (1991) 34, 443-454. 15 21. A. U lger, Arens regularity of weakly sequentialy compact Banach algebras, Proc. Amer. Math. Soc. 127 (11) (1999), 3221-3227. 22. A. U lger, Arens regularity of the algebra A ⊗B, Trans. Amer. Math. Soc. 305 (2) (1988) 623-639. 23. P. K. Wong, The second conjugate algebras of Banach algebras, J. Math. Sci. 17 (1) (1994), 15-18. 24. N. J. Young Theirregularity of multiplication in group algebra Quart. J. Math. Soc., 24 (2) (1973), 59-62. 25. Y. Zhing, Weak amenability of module extentions of Banach algebras Trans. Amer. Math. Soc. 354 (10) (2002), 4131-4151. Department of Mathematics, Amirkabir University of Technology, Tehran, Iran Email address: [email protected] Department of Mathematics, Amirkabir University of Technology, Tehran, Iran Email address: [email protected]
1509.03644	1	1509	2015-09-11T20:23:26	Fundamental function for Grand Lebesgue Spaces	[ "math.FA" ]	We investigate in this short article the fundamental function for the so-called Grand Lebesgue Spaces (GLS) and show in particular a one-to-one and mutually continuous accordance between its fundamental and generating function.	math.FA	math	FUNDAMENTAL FUNCTION FOR GRAND LEBESGUE SPACES. E.Ostrovsky, L.Sirota Department of Mathematic, Bar-Ilan University, Ramat Gan, 52900, Israel, e-mails: [email protected], [email protected] Abstract. We investigate in this short article the fundamental function for the so-called Grand Lebesgue Spaces (GLS) and show in particular a one-to-one and mutually continuous accordance between its fundamental and generating function. Key words and phrases: Young-Orlicz function, ordinary and Grand Lebesgue Spaces (GLS); Orlicz, GLS norms, rearrangement invariant spaces, fundamental and generating function, Young-Fenchel, or Legendre transform, theorem of Fenchel- Moraux, inverse function, Exponential Orlicz function (EOF) and Spaces (EOS). Mathematics Subject Classification (2000): primary 60G17; secondary 60E07; 60G70. 1 Notations. Statement of problem. A. A triplet (X, B, µ), where X = {x} is arbitrary set, B is non-trivial certain sigma-algebra of subsets X and µ is probabilistic: µ(X) = 1 diffuse non-negative completely additive measure defined on the B. The non-probabilistic case µ(X) = ∞ will be consider further. Recall that the measure µ is said to be diffuse iff for arbitrary measurable set A1 ∈ B with positive measure: µ(A1) > 0 there exists it subset A2 ⊂ A1 such that µ(A2) = µ(A1)/2. We denote as usually for any arbitrary measurable function f : X → R f p = (cid:20)ZX f (x)p µ(dx)(cid:21)1/p , p ≥ 1; Lp = {f, f p < ∞}. B. The so-called Grand Lebesgue Space (GLS) Gψ with norm · Gψ is defined ( not only in this article) as follows: Gψ = {f, f Gψ < ∞}, f Gψ def= sup p≥1 " f p ψ(p)# . (1.1) Here ψ = ψ(p), 1 ≤ p < ∞ is some continuous strictly increasing function such that limp→∞ ψ(p) = ∞. 1 The detail investigation of this spaces (and more general spaces) see in [14], [19]. See also [5], [6], [8], [9], [10] etc. The case when in (1.1) supremum is calculated over finite interval is investigated in [14], [19], [20]: Gbψ = {f, f Gbψ < ∞}, f Gbψ def= sup 1≤p<b" f p ψ(p)# , b = const > 1, (1.2) but in (1.2) ψ = ψ(p) is continuous function in the semi-open interval 1 ≤ p < b such that inf p∈(1,b) ψ(p) > 0. We will denote or simple b = b(ψ) := supp ψ(·), including the case b = ∞. [1, b) := supp ψ(·), Definition 1.1. The function = ψ(p) which appeared in (1.1) and (1.2), will be named as generating function for the correspondent Banach space Gψ. An used further example: ψ(β,b)(p) = (b − p)−β, 1 ≤ p < b, β = const > 0; b = const > 1, Gβ,b(p) := Gbψ(β,b)(p). C. We denote as ordinary for any measurable set A, A ∈ B it indicator function by I(A) = IA(ω). D. The Grand Lebesgue Spaces {Gψ} are rearrangement invariant in the [2], chapter 1. Therefore, its fundamental function classical definition, see e.g. φG(ψ)(δ), δ ≥ 0 is correctly defined in the considered case as follows: φG(ψ)(δ) def= sup p∈supp ψ" δ1/p ψ(p)# , φG(ψ)(1) = 1 inf p∈supp ψ ψ(p) . see [2], chapters 2 and 5. For instance, Note also (1.3) (1.3a) φG(C·ψ) = φG(ψ)/C, C = const > 0. (1.3b) This notion play a very important role in the functional analysis, [2], [22], [23]; in the theory of interpolation of operators, [2], [5], [7], in the theory of probability [13], [15], [16], [17]; in the theory of Partial Differential equations [7], [9]; in the 2 theory of martingales [20]; in the theory of approximation, in the theory of random processes etc. E. Let g = g(p), p ∈ (a, b), 1 ≤ a < b ≤ ∞ be some numerical valued continuous strictly increasing (or decreasing) function. The inverse function will be denoted by g(−1)(z), g(a) ≤ z ≤ g(b), in contradistinction to the usually notation g−1(p) = 1/g(p). F. The Young-Fenchel, or Legendre transform g∗(q) for the function g = g(p) one can to define g∗(q) def= sup p∈supp g (p q − g(p)). (1.4) Our goal in this short report is to establish a one-to-one and mutually continuous connection between the fundamental and generating functions for the Grand Lebesgue Spaces. In some previous works: [6], [12], chapter 8; [14], [19], [18], [22] these function was evaluated and applied in many practical cases. 2 Main result. Problem A. Let the generating function ψ be a given: ψ ∈ GΨ(a, b), 1 ≤ a < b ≤ ∞. Find the fundamental function for the correspondent Grand Lebesgue Space Gψ. Suppose the function p → p ψ(p) , p ∈ (a, b) is strictly increasing; and define therefore the function ν(p) = νψ(p) = " p ψ(p)#(−1) , p ∈ supp ψ, (2.1) and ν(p) = +∞ otherwise. Introduce also the following Young-Orlicz function and define finally N(u) = Nψ(u) := exp(cid:16)ν ∗ ψ(u)(cid:17) − exp(cid:16)ν ∗ ψ(0)(cid:17) , θ(δ) = θψ(δ) def= 1 N (−1) ψ (1/δ) , δ > 0. (2.2) (2.3) Proposition 2.1. We propose under formulated above conditions, for instance, µ(X) = 1, diffuseness of the measure µ, and in the case when b = ∞ 3 φG(ψ)(δ) = θψ(δ), δ > 0. (2.4) Remark 2.1. The equality (2.4) is more convenient than source definition (1.3). In particular, it allows for a relatively simple inversion. Proof is very simple; it based on the computation of the fundamental function for Orlicz spaces, see the book of Krasnosel'skii M.A. and Rutickii Ya.B. [11], chapter 3; see also the classical monographs [26], [27]. In detail, it is proved in particular in the articles [14], [18], [19] that under our conditions the Grand Lebesgue Space Gψ coincides with certain Orlicz space over source probability triplet (X, B, µ) relative the Young-Orlicz function Nψ(u). We deduce reducing considered case to the well-known calculation of fundamen- tal function for Orlicz space, [11], chapter 3 φG(ψ)(δ) = Q.E.D. 1 N (−1) ψ (1/δ) = θψ(δ), δ > 0, (2.4) An inverse problem B. Let the fundamental function φGψ(δ) = φ(δ) be a given. Find the correspondent generating function ψ(p). A first restrictions: the function φ = φ(δ) is strictly increasing and continuous; in particular φ(0+) = φ(0) = 0. We find from the equality (2.4) or equivalently Nψ(1/δ) = 1 φ(δ)!(−1) , Nψ(z) = 1 φ(δ)!(−1) .δ:=1/z. . (2.5) (2.5a) A second restriction: the function V ∗(z) = ln(C + Nψ(z)), z ≥ 0, where V ∗(0) = ln C, is continuous and upward convex. It follows immediately from (2.6) by virtue of theorem of Fenchel-Moraux Since we derive finally V (z) = {ln(C + Nψ(z))}∗ , z ≥ 0. V (p) = " p ψ(p)#(−1) , (2.6) (2.7) Proposition 2.2. We conclude under formulated in this pilcrow conditions 4 ψ(p) = p V (−1)(p) . (2.8) 3 The case of infinite measure. The case when µ(X) = ∞ is more complicated. Recall first of all definition and some facts about the so-called Exponential Orlicz Spaces (EOS), see for example [21]. Let N = N(u) be an N − Young-Orlicz's function, i.e. downward convex, even, continuous function differentiable for all sufficiently greatest values u, u ≥ u0, u0 = const > 0, strongly increasing along the right semi-axis and such that N(u) = 0 ⇔ u = 0; u → ∞ ⇒ dN(u)/du → ∞. We can say that N(·) is an exponential Orlicz function, briefly, N(·) ∈ EOF, if N(u) has a form of a continuous differentiable strongly increasing downward convex function W = W (u) in the domain [2, ∞] such that u → ∞ ⇒ W /(u) → ∞ and N(u) = N(W, u) = exp(W (log u)), u ≥ e2. For the values u ∈ [−e2, e2] we define N(W, u) arbitrarily, but so that the function N(W, u) is even, continuous, convex, strictly increasing along the right semi-axis and so that N(u) = 0 ⇔ u = 0. We denote the correspondent Orlicz space on (X, µ) with a measure µ and with N − function of the form N(W, u) as L(N) = EOS(W ); EOS = ∪W {EOS(W )} (exponential Orlicz's space). For example, let m = const > 0, r = const ∈ R1, Nm,r(u) = exphum (cid:16)log−mr(C1(r) + u)(cid:17)i − 1, C1(r) = e, r ≤ 0; C1(r) = exp(r), r > 0. Then Nm,r(·) ∈ EOS. In the case r = 0 we can write Nm = Nm,0. Recall that the Orlicz's norm on the arbitrarily measurable space (X, A, µ) f L(N) = f L(N, X, µ) can be calculated by the following formula (see, for example, [11], p. 66; [26], p. 73 ) f L(N) = inf v>0(cid:26)v−1(cid:18)1 +ZX N(vf (x)) µ(dx)(cid:19)(cid:27) . Let α be arbitrary number, α = const ≥ 1, and N(·) ∈ EOS(W ) for some W = W (·). For such a function N = N(W, u) we denote by N (α)(W ; u) = N (α)(u) a new Young-Orlicz's function N (α)(u) such that N (α)(u) = C1 uα, N (α)(u) = C3 + C4u, u ∈ [0, C2]; u ∈ (C2, C5]; N (α)(u) = N(u), u > C5, 0 < C2 < C5 < ∞, (3.1) C1,2,3,4,5 = C1,2,3,4,5(α, N(·)). 5 In the case of α = m(j +1), m > 0, j = 0, 1, 2, . . . the function N (α) to the following Trudinger's function: m (u) is equivalent N (α) m (u) ∼ N (α) [m] (u) = exp (um) − uml/l!. j Xl=0 This method is described in [28], p. 42-47. These Orlicz spaces are applicable to the theory of non-linear partial differential equations. We denote hereinafter generally by Ck = Ck(·), k = 1, 2, . . . some positive finite essentially constructive constants, and by C, C0 non-essentially constants, also con- structive. We proved the existence of constants C1,2,3,4,5 = C1,2,3,4,5(α, N(·)) such that N (α) is a new exponential N Orlicz's function in [21]. We denote classical absolute constants by the symbols Kj. Now we introduce some new Grand Lebesgue Spaces. Let ψ = ψ(p), p ≥ α, α = const ≥ 1 be a continuous positive ψ(α) > 0 finite strictly increasing function such that the function p → p log ψ(p) is downward convex, and We denote the set of all these functions by Ψ; Ψ = {ψ}. A particular case lim p→∞ ψ(p) = ∞. where ψ(p) = ψ(W ; p) = exp(W ∗(p)/p), W ∗(p) = sup z≥α (pz − W (z)) is so-called Young-Fenchel, or Legendre transform of W (·). It follows from the the- orem of Fenchel-Moraux that in this case W (p) = [p log ψ(W ; p)]∗ , p ≥ p0 = const ≥ 2, and, consequently, for all ψ(·) ∈ Ψ we introduce a correspondent Young-Orlicz N − function by the equality: N([ψ]) = N([ψ], u) = exp {[p log ψ(p)]∗ (log u)} , u ≥ e2. Definition 3.1. We introduce for such arbitrary function ψ(·) ∈ Ψ the so- called G(α; ψ) norms and correspondent Banach GLS space G(α; ψ) as a set of all measurable (complex) functions with finite norms: f G(α; ψ) = sup p≥α (f p/ψ(p)). (3.2) For instance, ψ(p) may be ψ(p) = ψm(p) = p1/m, m = const > 0; in this case, we can write G(α, ψm) = G(α, m) and f G(α, m) = sup p≥α(cid:16)f p p−1/m(cid:17) . 6 Theorem A, see [21]. Let the measure µ be diffuse, µ(X) = ∞, α = const ≥ 1 and ψ ∈ Ψ. We assert that the norms Orlicz-Luxemburg norm · L(N (α), [ψ]) and Grand Lebesgue Space norm · G(α, ψ), α ≥ 1 are equivalent. Arguing similarly to the second section, we obtain the following result. Proposition 3.1. We propose under conditions of theorem A φG(ψ)(δ) ≍ θψ(δ), δ > 0. Remark 3.1. Note that this case δ ∈ (0, ∞), in contradiction to the proposition 2.1, where it in naturally to take δ ∈ (0, 1). 4 Concluding remarks. Open problems. It is interest by our opinion to investigate the notion of fundamental function and also its relation with generating function for the so-called mixed, or equally anisotropic Grand Lebesgue Spaces. Recall that the definition of mixed, or equivalently anisotropic ordinary Lebesgue Spaces appeared at first in the article [1] and was investigated in detail in the classical books [3], [4]. The anisotropic Grand Lebesgue Spaces as a slight generalization of Lp spaces arises in turn in [23] with the correspondent fundamental function; in the preprint [24] both this notions was applied in the operator's theory. References [1] Benedek A. and Panzone R. The space L(p) with mixed norm. Duke Math. [2] Bennet C., Sharpley R. Interpolation of operators. Orlando, Academic Press Inc., (1988). [3] Besov O.V., Ilin V.P., Nikolskii S.M. Integral representation of functions and imbedding theorems. Vol.1; Scripta Series in Math., V.H.Winston and Sons, (1979), New York, Toronto, Ontario, London. [4] Besov O.V., Ilin V.P., Nikolskii S.M. Integral representation of functions and imbedding theorems. Vol.2; Scripta Series in Math., V.H.Winston and Sons, (1980), New York, Toronto, Ontario, London. [5] Capone C., Fiorenza A., Krbec M. On the Extrapolation Blowups in the Lp Scale. Collectanea Mathematica, 48, 2, (1998), 71-88. [6] Fiorenza A. Duality and reflexivity in grand Lebesgue spaces. Collectanea Mathematica (electronic version), 51, 2, (2000), 131-148. 7 [7] Fiorenza A., and Karadzhov G.E. Grand and small Lebesgue spaces and their analogs. Consiglio Nationale Delle Ricerche, Instituto per le Applicazioni del Calcoto Mauro Picine, Sezione di Napoli, Rapporto tecnico n. 272/03, (2005). [8] Iwaniec T., and Sbordone C. On the integrability of the Jacobian under minimal hypotheses. Arch. Rat.Mech. Anal., 119, (1992), 129143. [9] Iwaniec T., P. Koskela P., and Onninen J. Mapping of finite distortion: Monotonicity and Continuity. Invent. Math. 144 (2001), 507-531. [10] Kozachenko Yu. V., Ostrovsky E.I. (1985). The Banach Spaces of ran- dom Variables of subgaussian type. Theory of Probab. and Math. Stat. (in Russian). Kiev, KSU, 32, 43-57. [11] Krasnosel'skii, M.A.; Rutickii, Ya.B. (1961). Convex Functions and Or- licz Spaces. Groningen: P.Noordhoff Ltd [12] A. Kufner, O. John and S. Fucik. Function Spaces. Noordhoff Interna- tional Publishingr, Leyden, 1977. [13] Ledoux M., Talagrand M. (1991) Probability in Banach Spaces. Springer, Berlin, MR 1102015. [14] Liflyand E., Ostrovsky E., Sirota L. Structural Properties of Bilateral Grand Lebesgue Spaces. Turk. Journal of Math., 34, (2010), 207-219. [15] Ostrovsky E., Rogover E. Exact exponential Bounds for the random field Maximum Distribution via the majorizing Measures (Generic Chaining). arX- iv:0802.0349v1 [math.PR] 4 Feb 2008. [16] Ostrovsky E.I. (1999). Exponential estimations for random Fields and its applications. (in Russian). Moscow-Obninsk, OINPE. [17] Ostrovsky E.I. (1994.) Exponential Bounds in the Law of Iterated Logarithm in Banach Space. Math. Notes, 56, 5, p. 98-107. [18] Ostrovsky E., Sirota L. Fouier Transforms in exponential rearrangement invaiant Spaces. arXiv:040639v1 [math.FA], 20 Jun 2004. [19] Ostrovsky E. and Sirota L. Moment Banach spaces: theory and applica- tions. HIAT Journal of Science and Engineering, C, Volume 4, Issues 1-2, pp. 233-262, (2007). [20] Ostrovsky E. and Sirota L. Moment and Tail Inequalities for polynomial Martimgales. The case of heavy tails. arXiv: 1112.2768v1 [math.PR] 13 Dez 2011. [21] Ostrovsky E. and Sirota L. Fourier transforms in exponential rearrange- ment invariant spaces. arXiv:math/0406391v1 [math.FA] 20 Jun 2004 8 [22] Ostrovsky E. Exponential Orlicz Spaces: New Norms and Applications. arXiv:math/0406534v1 [math.FA] 25 Jul 2004 [23] Ostrovsky E. and Sirota L. Exact norm estimates for multivariate di- lation operators between two Bilateral Weight Grand Lebesgue Spaces. arX- iv:1503.05235v1 [math.FA] 17 Mar 2015 [24] Ostrovsky E. and Sirota L. Central Limit Theorem and exponential tail estimation in mixed (anisotropic) Grand Lebesgue Spaces. arXiv:1308.5606v1 [math.PR] 26 Aug 2013 [25] Pizier G. Condition d entropic assupant la continuite de certains processus et applications a lanalyse harmonique. Seminaire d analyse fonctionalle. (1980), Exp.13, p. 23-34. [26] Rao M.M., Ren Z.D. Theory of Orlicz Spaces. Marcel Dekker Inc., 1991. New York, Basel, Hong Kong. [27] Rao M.M., Ren Z.D. Applications of Orlicz Spaces. Marcel Dekker Inc., 2002. New York, Basel, Hong Kong. [28] Taylor M.E. Partial Differential Equations. v.3, Nonlinear Equations. Springer Verlag, 1991. Berlin, Heidelberg, New York. [29] Talagrand M. (1996). Majorizing measure: The generic chaining. Ann. Probab., 24 1049-1103. MR1825156 9
1602.08790	2	1602	2016-05-14T15:33:04	Product operators on mixed norm spaces	[ "math.FA" ]	Inequalities for product operators on mixed norm Lebesgue spaces and permuted mixed norm Lebesgue spaces are established. They depend only on inequalities for the factors and on the Lebesgue indices involved. Inequalities for the bivariate Laplace transform are given to illustrate the method. Also, an elementary proof is presented for an $n$-variable Young's inequality in mixed norm spaces.	math.FA	math	PRODUCT OPERATORS ON MIXED NORM SPACES WAYNE GREY AND GORD SINNAMON Abstract. Inequalities for product operators on mixed norm Lebesgue spaces and permuted mixed norm Lebesgue spaces are established. They depend only on inequalities for the factors and on the Lebesgue indices involved. Inequal- ities for the bivariate Laplace transform are given to illustrate the method. Also, an elementary proof is presented for an n-variable Young's inequality in mixed norm spaces. 1. Introduction The techniques used to study embeddings of mixed norm spaces in [6] and [8] can be extended to work with operators other than the identity. Here we begin this work by considering a select class of operators. Mixed norm spaces are spaces of multivariable functions in which the norm takes advantage of the product structure in the domain. They have a long informal history but were first named and formally studied by Benedek and Panzone in [4]. Permuted mixed norms only arise when studying more than a single space, since they occur when the order in which the factor norms are taken is different in different spaces. The importance of permuted mixed norms was noted in [5] and [11, §XI.1]. They also appear in [3], a study of the Hardy operator in mixed-norm Lebesgue spaces, and in the papers [10] and [2], in which partial integral operators are considered between mixed-norm Banach function spaces, with Lebegue and Orlicz norms as special cases. A systematic study of continuous inclusions between mixed-norm Lebesgue spaces was undertaken in the first author's thesis, [6]. It will be convenient to introduce the two-variable mixed norm spaces needed in Section 2 first, postponing the introduction of n-variable spaces until Section 3. Let (T1, λ1) and (T2, λ2) be σ-finite measure spaces and let L0 denote the collection of (λ1 × λ2)-measurable functions on T1 × T2. In particular, either or both of λ1 and λ2 may be "counting" measure. Thus, the definitions of this section and the results of the next apply to weighted or unweighted sums as special cases. λ1×λ2 Fix indices p1, p2 ∈ (0, ∞). For any f ∈ L0 λ1×λ2 , kf kL(p1,p2) λ1×λ2 = ZT2(cid:18)ZT1 f (t1, t2)p1 dλ1(t1)(cid:19)p2/p1 dλ2(t2)!1/p2 . The first variable of the function f is always in the λ1 measure space, and the order of the indices and measures indicates which variable is the "inner" one. So for any 2010 Mathematics Subject Classification. Primary 46E30, Secondary 44A35, 26D15. Key words and phrases. mixed norm, permuted mixed norm, product operator, Young's in- equality, bivariate Laplace transform. Supported by the Natural Sciences and Engineering Research Council of Canada. 1 2 WAYNE GREY AND GORD SINNAMON f ∈ L0 λ1×λ2 , kf kL(p2,p1) λ2×λ1 = ZT1(cid:18)ZT2 f (t1, t2)p2 dλ2(t2)(cid:19)p1/p2 dλ1(t1)!1/p1 . Although these are genuine norms only when p1 ≥ 1 and p2 ≥ 1 we will refer to them as mixed norms even when some indices are less than 1. Also, the results we present will often extend to the case when one or both of the indices is infinite, indicating that the supremum norm is to be taken in that factor. The product operators that we consider will be introduced in Section 2. As mentioned above, we restrict our attention to the two-variable case. Although the extension from bivariate operators to multivariate operators may seem straightfor- ward, the delicate arguments needed in the embedding case in [6] and [8] and the advanced techniques introduced in [7] put this extension beyond the scope of the present article. Convolution operators have some product structure but not enough to make them product operators in general. Convolution is considered in Section 3, where it leads to an elementary proof of a multivariate mixed norm Young's inequality. The Hardy operators considered in [3] are product operators (and also convolu- tion operators) with the added advantage that weighted norm inequalities for the factors have been characterized. The results may be compared with Theorem 2.1, below. Hardy operators are not the only operators with these useful properties, see for example, the Riemann-Liouville operators studied in [13]. We will make frequent use of Minkowski's (integral) inequality in the following mixed norm form: If 0 < p1 ≤ p2 < ∞, then, ≤ kf kL(p2,p1) kf kL(p1,p2) λ1×λ2 λ2×λ1 , f ∈ L+ λ1×λ2 . For easy recognition, we will enclose the function in square brackets when applying Minkowski's inequality in integral estimates. It becomes, ZT2(cid:18)ZT1 [f (t1, t2)]p1 dλ1(t1)(cid:19)p2/p1 ≤ ZT1(cid:18)ZT2 dλ2(t2)!1/p2 [f (t1, t2)]p2 dλ2(t2)(cid:19)p1/p2 dλ1(t1)!1/p1 . 2. Product operators Let (T1, λ1), (T2, λ2), (X1, µ1), and (X2, µ2) be σ-finite measure spaces. Let denote the collection of non-negative (λ1 × λ2)-measurable functions. → L+ λ1×λ2 µ1×µ2 will be called a product operator provided it L+ λ1×λ2 An operator K : L+ can be expressed in the form, Kf (x1, x2) =ZZT1×T2 k1(x1, t1)k2(x2, t2)f (t1, t2) d(λ1 × λ2)(t1, t2), where kj is a non-negative (µj × λj )-measurable function for j = 1, 2. In this case we define Kj : L+ λj → L+ µj by Kjf (x) =ZTj kj(x, t)f (t) dλj (t), j = 1, 2. PRODUCT OPERATORS ON MIXED NORM SPACES 3 Note that, by Tonelli's theorem, Kf (x1, x2) =ZT1 =ZT2 k1(x1, t1)ZT2 k2(x2, t2)ZT1 k2(x2, t2)f (t1, t2) dλ2(t2) dλ1(t1) k1(x1, t1)f (t1, t2) dλ1(t1) dλ2(t2). Suppose p1, p2, r1, and r2 are positive and let Cj be the least constant, finite or infinite, such that kKjf kL rj µj ≤ Cjkf kL , pj λj f ∈ L+ λj . Theorem 2.1. Fix positive indices p1, p2, r1, and r2. Suppose K is a product operator, with kj , Kj, Cj as above for j = 1, 2. (a) If r1 ≥ 1 then K satisfies the mixed norm inequality, kKf kL(r1,r2 ) µ1×µ2 ≤ C1C2kf kL(p1,p2) λ1×λ2 , f ∈ L+ λ1×λ2 . (b) If r1 ≥ 1, r2 ≥ 1 and min(p1, r1) ≤ max(p2, r2) then K satisfies the permuted mixed norm inequality, kKf kL(r1,r2 ) µ1×µ2 ≤ C1C2kf kL(p2,p1) λ2×λ1 , f ∈ L+ λ1×λ2 . Proof. In a slight abuse of notation we write, K1f (x1, t2) =ZT1 K2f (t1, x2) =ZT2 k1(x1, t1)f (t1, t2) dλ1(t1) and k2(x2, t2)f (t1, t2) dλ2(t2). To prove part (a), use the hypothesis r1 ≥ 1 to apply Minkowski's inequality and then invoke the definitions of C1 and C2. This yields, Kf (x1, x2)r1 dµ1(x1)(cid:19)r2/r1 dµ2(x2)!1/r2 ZX2(cid:18)ZX1 = ZX2(cid:18)ZX1(cid:18)ZT2hk2(x2, t2)K1f (x1, t2)i dλ2(t2)(cid:19)r1 ≤ ZX2 ZT2(cid:18)ZX1hk2(x2, t2)K1f (x1, t2)ir1 = ZX2 ZT2 ≤ C1 ZX2 ZT2 ≤ C1C2 ZT2(cid:18)ZT1 dµ1(x1)(cid:19)1/r1 K1f (x1, t2)r1 dµ1(x1)(cid:19)1/r1 f (t1, t2)p1 dλ1(t1)(cid:19)1/p1 dλ2(t2)!1/p2 k2(x2, t2)(cid:18)ZX1 k2(x2, t2)(cid:18)ZT1 f (t1, t2)p1 dλ1(t1)(cid:19)p2/p1 . 1/r2 dµ1(x1)(cid:19)r2/r1 dλ2(t2)!r2 dλ2(t2)!r2 dλ2(t2)!r2 dµ2(x2)!1/r2 dµ2(x2)! dµ2(x2)! dµ2(x2)! 1/r2 1/r2 Part (b) will be done in four cases: They arise from the observation that the condition min(p1, r1) ≤ max(p2, r2) is satisfied if and only if one of, p1 ≤ p2, r1 ≤ r2, 4 WAYNE GREY AND GORD SINNAMON p1 ≤ r2, or r1 ≤ p2 holds. inequality, yields If p1 ≤ p2 then part (a), followed by Minkowski's kKf kL(r1,r2) µ1×µ2 ≤ C1C2kf kL(p1,p2) λ1×λ2 ≤ C1C2kf kL(p2,p1) λ2×λ1 , f ∈ L+ λ1×λ2 . If r1 ≤ r2 then we begin by using Minkowski's inequality, and follow with part (a) to get kKf kL(r1,r2 ) µ1×µ2 ≤ kKf kL(r2,r1 ) µ2×µ1 ≤ C2C1kf kL(p2,p1) λ2×λ1 , f ∈ L+ λ1×λ2 . If p1 ≤ r2 then we apply the definition of C1 followed by Minkowski's inequality and then the definition of C2 to get Kf (x1, x2)r1 dµ1(x1)(cid:19)r2/r1 dµ2(x2)!1/r2 k1(x1, t1)K2f (t1, x2) dλ1(t1)(cid:19)r1 ZX2(cid:18)ZX1 = ZX2(cid:18)ZX1(cid:18)ZT1 ≤ C1 ZX2(cid:18)ZT1hK2f (t1, x2)ip1 ≤ C1 ZT1(cid:18)ZX2hK2f (t1, x2)ir2 ≤ C1C2 ZT1(cid:18)ZT2 dλ1(t1)(cid:19)r2/p1 dµ2(x2)(cid:19)p1/r2 f (t1, t2)p2 dλ2(t2)(cid:19)p1/p2 dµ2(x2)!1/r2 dλ1(t1)!1/p1 dλ1(t1)!1/p1 . dµ1(x1)(cid:19)r2/r1 dµ2(x2)!1/r2 If r1 ≤ p2 the process is somewhat lengthy, using the definitions of C1 and C2 as well as three applications of Minkowski's inequality. First, apply Minkowski's inequality with r1 ≥ 1, followed by the definition of C2 to get Kf (x1, x2)r1 dµ1(x1)(cid:19)r2/r1 dµ2(x2)!1/r2 ZX2(cid:18)ZX1 = ZX2(cid:18)ZX1(cid:18)ZT2hk2(x2, t2)K1f (x1, t2)i dλ2(t2)(cid:19)r1 ≤ ZX2 ZT2(cid:18)ZX1hk2(x2, t2)K1f (x1, t2)ir1 = ZX2 ZT2 k2(x2, t2)(cid:18)ZX1 ≤ C2 ZT2(cid:18)ZX1hK1f (x1, t2)ir1 dµ1(x1)(cid:19)1/r1 K1f (x1, t2)r1 dµ1(x1)(cid:19)1/r1 dµ1(x1)(cid:19)p2/r1 dλ2(t2)!1/p2 . dµ1(x1)(cid:19)r2/r1 dλ2(t2)!r2 dλ2(t2)!r2 dµ2(x2)!1/r2 dµ2(x2)! dµ2(x2)! 1/r2 1/r2 PRODUCT OPERATORS ON MIXED NORM SPACES 5 Now Minkowski's inequality with p2 ≥ r1 shows that the last expression is no larger than C2 ZX1(cid:18)ZT2hK1f (x1, t2)ip2 dλ2(t2)(cid:19)r1/p2 dµ1(x1)!1/r1 . After expanding K1f (x1, t2) in this expression we may apply Minkowski's inequality with p2 ≥ 1 to get C2 ZX1(cid:18)ZT2(cid:18)ZT1hk1(x1, t1)f (t1, t2)i dλ1(t1)(cid:19)p2 ≤ C2 ZX1 ZT1(cid:18)ZT2hk1(x1, t1)f (t1, t2)ip2 = C2 ZX1 ZT1 ≤ C1C2 ZT1(cid:18)ZT2 k1(x1, t1)(cid:18)ZT2 f (t1, t2)p2 dλ2(t2)(cid:19)p1/p2 dλ2(t2)(cid:19)1/p2 f (t1, t2)p2 dλ2(t2)(cid:19)1/p2 dλ1(t1)!1/p1 , dλ2(t2)(cid:19)r1/p2 dλ1(t1)!r1 dλ1(t1)!r1 dµ1(x1)!1/r1 dµ1(x1)! dµ1(x1)! 1/r1 1/r1 where the last inequality uses the definition of C1. These estimates complete the proof of the fourth and last case. (cid:3) The index conditions in Theorem 2.1(b) are somewhat stronger than needed. It is enough to assume that one (or more) of the following four conditions holds: p1 ≤ r2; 1 ≤ r1 ≤ p2; 1 ≤ r1 and p1 ≤ p2; or 1 ≤ r2 and r1 ≤ r2. The statement of the theorem remains valid when some indices are infinite, provided the index conditions are met, but the proofs would have to be modified to accommodate occurrences of the supremum norm. The inequalities of Theorem 2.1 are stated for non-negative functions only but it is routine to extend the operator K to all functions for which the right hand side is finite, in a way that preserves the norm inequalities. Also, the results of Theorem 2.1 may be seen to hold for a more general class of positive operators. Instead of supposing that the factors K1 and K2 are integral operators with non-negative kernels, it would suffice to assume that they map positive functions to positive functions and possess formal adjoints. See [12, Lemma 2.4] for properties of such operators and [9, Section 4] for connections with positive integral operators. One advantage of such an extension is that the identity operator is not an integral operator (in general) but it does have a formal adjoint. Corollary 2.2. Suppose p1, p2 ∈ (1, 2]. The bivariate Laplace transform L2, de- fined by L2f (x1, x2) =Z ∞ 0 Z ∞ 0 satisfies the permuted mixed norm inequality e−x1t1−x2t2 f (t1, t2) dt1 dt2, x1 > 0, x2 > 0, kL2f kL(p′ 1 = 1/p2 + 1/p′ 1 ) ≤ C(p1)C(p2)kf kL(p2,p1) . ,p′ 2 Here 1/p1 + 1/p′ Laplace transform as a map from Lp to Lp′ . 2 = 1, and C(p) is the norm of the single-variable 6 WAYNE GREY AND GORD SINNAMON Proof. The bivariate Laplace transform is a product operator, in the above sense, whose factors are both the single-variable Laplace transform, L , given by L f (x) =Z ∞ 0 e−xtf (t) dt, x > 0. It is well known that L is bounded as a map from Lp to Lp′ when 1 < p ≤ 2. Since p′ the second statement of Theorem 2.1 gives the conclusion. 2 ≥ 1, and min(p1, p′ 1 ≥ 1, p′ 1) = p1 ≤ 2 ≤ p′ , i.e. C(p) is finite, 2) (cid:3) 2 = max(p2, p′ Note that if 1 < p2 < p1 ≤ 2, the permuted mixed norm inequality given above can be used to refine the unpermuted one given by Theorem 2.1(a) in two different ways. Corollary 2.2 and Minkowski's inequality give kL2f kL(p′ 1 ,p′ 2 ) ≤ C(p1)C(p2)kf kL(p2,p1) ≤ C(p1)C(p2)kf kL(p1,p2) , and also, kL2f k L(p′ 1,p′ 2) ≤ kL2f k L(p′ 2,p′ 1) ≤ C(p1)C(p2)kf kL(p1,p2) . Using Theorem 2.1 it is easy to generate additional permuted mixed norm in- equalities, by beginning with known single-variable inequalities. See, for example, [1, Corollary 1] for more single-variable inequalities involving the Laplace transform. 3. Young's inequality Fix a positive integer n and let L+ be the collection of non-negative Lebesgue measurable functions on Rn. For P = (p1, . . . , pn) ∈ [1, ∞)n define kf kLP = ZR . . . ZR(cid:18)ZR f (t1, . . . , tn)p1 dt1(cid:19)p2/p1 dt2!p3/p2 1/pn . . . . dtn  The convolution of two (real-valued) functions on Rn is defined by, f ∗ g(x) =ZRn whenever the integral exists. f (x − t)g(t) dt, For a fixed function g the map f 7→ f ∗ g has a kind of product structure. But, even in the two-variable case, it is not a product operator of the sort considered in the previous section unless g factors as g(t1, t2) = g1(t1)g2(t2). This will keep us from establishing permuted mixed norm inequalities. Nevertheless, exploiting the existing product structure provides an elementary proof of a n-variable mixed norm Young's inequality. Recall the single-variable Young's inequality: If p, q, r ∈ [1, ∞] satisfy 1/p+1/q = 1/r + 1, f ∈ Lp and g ∈ Lq, then f ∗ g is well defined and kf ∗ gkLr ≤ kf kLpkgkLq. Theorem 3.1. Suppose P = (p1, . . . , pn), Q = (q1, . . . , qn), and R = (r1, . . . rn) satisfy, 1 pj + 1 qj = 1 rj + 1, j = 1, . . . , n. If f ∈ LP and g ∈ LQ then f ∗ g is well defined, and kf ∗ gkLR ≤ kf kLP kgkLQ. PRODUCT OPERATORS ON MIXED NORM SPACES 7 Proof. We begin by proving the inequality when f and g are non-negative, so that f ∗ g is well-defined as a function taking values in [0, ∞]. As the first step in a recursive argument, let f1 = f and g1 = g. Define x and t so that (x1, x2, . . . , xn) = (x1, x) and (t1, t2, . . . , tn) = (t1, t). Also, let dt denote dt2, . . . dtn. Then Young's inequality gives, for each x and t, (cid:18)ZR(cid:18)ZR f1(x1 − t1, x − t)g1(t1, t) dt1(cid:19)r1 dx1(cid:19) 1/r1 ≤(cid:18)ZR f1(y1, x − t)p1 dy1(cid:19)1/p1(cid:18)ZR g1(t1, t)q1 dt1(cid:19)1/q1 . So, applying Minkowski's inequality, we get f1 ∗ g1(x1, x)r1 dx1(cid:19)1/r1 (cid:18)ZR =(cid:18)ZR(cid:18)ZRn−1(cid:20)ZR ≤ZRn−1(cid:18)ZR(cid:20)ZR ≤ZRn−1(cid:18)ZR f1(x1 − t1, x − t)g1(t1, t) dt1(cid:21) dt(cid:19)r1 f1(x1 − t1, x − t)g1(t1, t) dt1(cid:21)r1 dx1(cid:19) dx1(cid:19) 1/r1 dt 1/r1 f1(y1, x − t)p1 dy1(cid:19)1/p1(cid:18)ZR g1(t1, t)q1 dt1(cid:19)1/q1 dt. Now define f2, g2 : Rn−1 → [0, ∞] by We have just shown that, Applying the above argument to f2 and g2 gives f1(y1, y2, . . . , yn)p1 dy1(cid:19)1/p1 g1(t1, t2, . . . , tn)q1 dt1(cid:19)1/q1 . f2(y2, . . . , yn) =(cid:18)ZR g2(t2, . . . , tn) =(cid:18)ZR f1 ∗ g1(x1, x2, . . . , xn)r1 dx1(cid:19)1/r1 f2 ∗ g2(x2, x3, . . . , xn)r2 dx2(cid:19)1/r2 f3(y3, . . . , yn) =(cid:18)ZR g3(t3, . . . , tn) =(cid:18)ZR fn−1 ∗ gn−1(xn−1, xn)rn−1 dxn−1(cid:19)1/rn−1 f2(y2, y3, . . . , yn)p2 dy2(cid:19)1/p2 g2(t2, t3, . . . , tn)q2 dt2(cid:19)1/q2 . and (3.1) (cid:18)ZR (cid:18)ZR (3.2) where and (3.3) (cid:18)ZR ≤ f2 ∗ g2(x2, . . . , xn). ≤ f3 ∗ g3(x3, . . . , xn), We continue in this way until reaching ≤ fn ∗ gn(xn), 8 where and WAYNE GREY AND GORD SINNAMON fn(yn) =(cid:18)ZR gn(tn) =(cid:18)ZR fn−1(yn−1, yn)pn−1 dyn−1(cid:19)1/pn−1 gn−1(tn−1, tn)qn−1 dtn−1(cid:19)1/qn−1 . Then Young's inequality gives, (3.4) (cid:18)ZR fn ∗ gn(xn)rn dxn(cid:19)1/rn ≤(cid:18)ZR fn(tn)pn dtn(cid:19)1/pn(cid:18)ZR gn(tn)qn dtn(cid:19)1/qn . Recursively applying the definitions of fj and gj shows that the right hand side of (3.4) is just kf kLP kgkLQ. The convolution estimates (3.1),(3.2),...,(3.3) concatenate to give a lower bound for the left hand side of (3.4) and we have kf ∗ gkLR ≤ kf kLP kgkLQ. Now we drop the assumption of positivity on f ∈ LP and g ∈ LQ. For bounded, integrable functions the convolution exists and is finite everywhere. And it is routine to express f and g as pointwise limits of bounded, integrable functions fk and gk, respectively, satisfying fk ≤ f and gk ≤ g. Since f ∈ LP and g ∈ LQ we have shown f ∗ g ∈ LR and hence f ∗ g(x) < ∞ almost everywhere. So for almost every x the dominated convergence theorem proves that f ∗ g(x) exists. Also, kf ∗ gkLR ≤ kf ∗ gkLR ≤ kf kLP kgkLQ = kf kLP kgkLQ. This completes the proof. (cid:3) Once again, the statement of the theorem remains valid when some indices are infinite, but the proof would have to be modified to accommodate occurrences of the supremum norm. The same straightforward procedure may be applied to prove a mixed norm Young's inequality over any finite product of locally compact unimodular groups. References [1] Kenneth F. Andersen, Weighted inequalities for convolutions, Proc. Amer. Math. Soc. 123 (1995), no. 4, 1129–1136, DOI 10.2307/2160710. [2] J. Appell, A. S. Kalitvin, and P. P. Zabreıko, Partial integral operators in Orlicz spaces with mixed norm, Colloq. Math. 78 (1998), no. 2, 293–306. [3] Jurgen Appell and Alois Kufner, On the two-dimensional Hardy operator in Lebesgue spaces with mixed norms, Analysis 15 (1995), no. 1, 91–98, DOI 10.1524/anly.1995.15.1.91. [4] A. Benedek and R. Panzone, The space Lp, with mixed norm, Duke Math. J. 28 (1961), 301–324. [5] John J. F. Fournier, Mixed norms and rearrangements: Sobolev's inequality and Littlewood's inequality, Ann. Mat. Pura Appl. (4) 148 (1987), 51–76, DOI 10.1007/BF01774283. [6] Wayne Grey, Inclusions Among Mixed-Norm Lebesgue Spaces, University of Western Ontario, London, Canada, 2015. Ph.D. Thesis. [7] , Mixed-norm estimates and symmetric geometric means, To appear in Positivity. arXiv:1507.08327 [math.FA] (July 29, 2015). [8] Wayne Grey and Gord Sinnamon, The inclusion problem for mixed-norm spaces, Trans. Amer. Math. Soc., posted on January 26, 2016, DOI 10.1090/tran6665, (to appear in print). PRODUCT OPERATORS ON MIXED NORM SPACES 9 [9] Ralph Howard and Anton R. Schep, Norms of positive operators on Lp-spaces, Proc. Amer. Math. Soc. 109 (1990), no. 1, 135–146, DOI 10.2307/2048373. [10] A. S. Kalitvin and P. P. Zabrejko, On the theory of partial integral operators, J. Integral Equations Appl. 3 (1991), no. 3, 351–382, DOI 10.1216/jiea/1181075630. [11] L. V. Kantorovich and G. P. Akilov, Functional analysis, 2nd ed., Pergamon Press, Oxford- Elmsford, N.Y., 1982. Translated from the Russian by Howard L. Silcock. [12] Gord Sinnamon, Weighted inequalities for positive operators, Math. Inequal. Appl. 8 (2005), no. 3, 419–440, DOI 10.7153/mia-08-39. [13] V. D. Stepanov, Two-weight estimates for Riemann-Liouville integrals, Izv. Akad. Nauk SSSR Ser. Mat. 54 (1990), no. 3, 645–656 (Russian); English transl., Math. USSR-Izv. 36 (1991), no. 3, 669–681. Department of Mathematics, University of Western Ontario, London, Canada E-mail address: [email protected] Department of Mathematics, University of Western Ontario, London, Canada E-mail address: [email protected]
1909.00996	1	1909	2019-09-03T08:22:56	Which Topologies induced by order convergences	[ "math.FA" ]	In this paper, we will study on some topologies induced by order convergences in a vector lattice. We will investigate the relationships of them.	math.FA	math	WHICH TOPOLOGIES INDUCED BY ORDER CONVERGENCES KAZEM HAGHNEJAD AZAR∗ Abstract. In this paper, we will study on some topologies induced by order convergences in a vector lattice. We will investigate the relationships of them. 1. Introduction Recall that a net (xα)α∈A in a Riesz space E is order convergent to x ∈ E, o−→ x whenever there exists another net (yβ)β∈B in E such that denoted by xα yβ ↓ 0 and that for every β ∈ B, there exists α0 ∈ A such that xα − x ≤ yβ for all α ≥ α0. If there exists a net (yα)α∈A (with the same index set) in a Riesz space o−→ x. Conversely, E such that yα ↓ 0 and xα − x ≤ yα for each α ∈ A, then xα if E is a Dedekind complete Riesz space and (xα)α∈A is order bounded, then o−→ x in E implies that there exists a net (yα)α∈A (with the same index set) xα such that yα ↓ 0 and xα − x ≤ yα for each α ∈ A. For sequences in a Riesz o−→ x if and only if there exists a sequence (yn) such that yn ↓ 0 and space E, xn xn − x ≤ yn for each n ∈ N (cf. [1, P.17 and P.18]). We adopt [2] as standard reference for basic notions on Riesz spaces and Banach lattices. Recall that a real vector space E (with elements x,y,...) is called an ordered vector space if E is partially ordered in such a manner that the vector space structure and order structure are compatible, that is to say, x ≤ y implies x + z ≤ y + z for every z ∈ E and x ≥ y implies αx ≥ αy for every α ≥ 0 in R. A Riesz space E is an order vector space in which sup(x, y) ( it is customary to write sometimes x ∨ y instead of sup(x, y) and x ∧ y instead of inf(x, y) ) exists for every x, y ∈ E. Let E be a Riesz space, for each x, y ∈ E with x ≤ y, the set [x, y] = {z ∈ E : x ≤ z ≤ y} is called an order interval. A subset of E is said to be order bounded if it is included in some order interval. A Riesz space is said to be Dedekind complete (resp. σ-Dedekind complete) if every order bounded above subset (resp. countable subset) has a supremum. A subset A of a Riesz space E is said to be solid if it follows from y ≤ x whit x ∈ A and y ∈ E that y ∈ A . An order ideal of E is a solid subspace. A band of E is an order closed order ideal. A Banach lattice E is a Banach space (E, k.k) such that E is a Riesz space and its norm satisfies the following property: for each x, y ∈ E such that x ≤ y, we have kxk ≤ kyk. A Banach lattice E has order continuous norm if kxαk → 0 Copyright 2016 by the Tusi Mathematical Research Group. Date: Received: xxxxxx; Revised: yyyyyy; Accepted: zzzzzz. ∗Corresponding author. 2010 Mathematics Subject Classification. Primary 46B42; Secondary 47B60. Key words and phrases. Riesz space, unbounded order convergence, uo-continuous operator, uσo-continuous operator. 1 2 K. HAGHNEJAD AZAR for every decreasing net (xα)α with inf α xα = 0. A vector x > 0 in a Riesz space E is called an atom if Ex = {y ∈ E : ∃λ > 0, y ≤ λx}, the ideal generated by x, is one-dimensional if and only if u, v ∈ [0, x] with u ∧ v = 0 implies u = 0 or v = 0. A Riesz space E is said to be atomic if the linear span of all atoms is order dense in E if and only if it is the band generated by its atoms. For example c, c0, ℓp(1 ≤ p ≤ ∞) are atomic Banach lattices and C[0, 1], L1[0, 1] are atomless Banach lattices. Let E, F be Riesz spaces. An operator T : E → F is said to be order bounded if it maps each order bounded subset of E into order bounded subset of F . The collection of all order bounded operators from a Riesz space E into a Riesz space F will be denoted by Lb(E, F ). The collection of all order bounded linear functionals on a Riesz space E will be denoted by E ∼, that is E ∼ = Lb(E, R) . A functional on a Riesz space is order continuous (resp. σ-order continuous) if it maps order null nets (resp. sequences) to order null nets (resp. sequences). The collection of all order continuous (resp. σ-order continuous) linear functionals on a Riesz space E will be denoted by E ∼ c ). For unexplained terminology and facts on Banach lattices and positive operators, we refer the reader to the excellent book of [2]. n (resp. E ∼ 2. Order Topology Let E be a vector lattice. A subset A of a E is said to be quasi-order closed whenever for every (xα) ⊆ A with xα ↑ x or xα ↓ x implies x ∈ A. We observe that a solid subset A ⊆ E is a quasi-order closed if and only if A is order closed. θ ⊆ E is called order open if and only if E \ θ is quasi-order closed. Now consider the following topologies: (1) First topology is called quasi-order topology which we define as follows. τo = {θ ⊆ E : E \ θ is quasi-order closed} It is clear, τo is a topology for E. (2) Assume that τe be a topology for E with following basis {(a, b) : a, b ∈ E and a < b}. We call this topology as order topology. In the following proposition, we show that (E, τo) and (E, τe) are both vector topologies. Proposition 2.1. Let E be a Dedekind complete vector lattice. Then τo and τe both are vector topology. Proof. Obvious that τe is a vector topology. We only show that τo is vector topology. First, we prove that the operation x → tx for each t ∈ R is continuous. Let θ ⊂ E be an order open subset of E, then we must show that tθ is an order open subset of E for each t ∈ R. Since θ is order open, it follows that θc = F is quasi-order closed. Put (tθ)c = G and (xα) ⊆ G with xα ↑ x. Then we have xα /∈ tθ iff t−1xα /∈ θ iff t−1xα ∈ F for each α and since t−1xα ↑ t−1x, follows that t−1x ∈ F , implies that x ∈ tF . Then we have t−1x ∈ F iff t−1x /∈ θ iff x /∈ θ iff x ∈ G, which follows that G is quasi order closed, and so tθ is an order open subset of E. WHICH TOPOLOGIES INDUCED BY ORDER CONVERGENCES 3 Now we show that the operation (x, y) → x + y is continuous. Set θ1 and θ2 order open subsets of E, we show that θ1 + θ2 is an order open subset of E. Let a ∈ θ1. First we prove that a + θ2 is an order open subset of E. Put θc 2 = F and (a + θ2)c = G. We show that G is quasi-order closed. Let (xα) ⊆ G and xα ↑ x in G. Then we have xα ∈ G iff xα /∈ (a + θ2) iff (xα − a) /∈ θ2. Since (xα − a) ↑ (x − a), follows that (x − a) ∈ F , and so (x − a) /∈ θ2 iff x /∈ (a + θ2) iff x ∈ G. Thus G is quasi-order closed, and so a + θ2 is an order open subset of E. Now by θ1 + θ2 = Sa∈θ1 (a + θ2), the proof follows. (cid:3) Lemma 2.2. Let E be a Dedekind complete vector lattice and τo be a order topology for E. Then for each c ∈ E and neighborhood Uc of c, there are a, b ∈ E such that c ∈ (a, b) ⊂ Uc. Proof. let c ∈ E and Uc be an neighbourhood of c in order topology. First we show that there is a ∈ E such that (a, c) ⊂ Uc. By contradiction, let (a, c) ∩ U c c 6= ∅. Then for each a < c there is ca ∈ (a, c) ∩ U c c . It follows that sup{ca : ca ∈ (a, c) ∩ U c c } = c. For each a < b < c, we can set ca < cb. It follows that for each a < c, there exists cα(a) ∈ (a, c) ∩ U c c , which is not possible. Thus there is a < x such that (a, c) ⊂ Uc. In the similar way there is a c < b such that (c, b) ⊂ Uc and proof follows. (cid:3) c with cα(a) ↑ c. It follows that c ∈ U c The preceding lemma shows that τo ⊆ τe, but as following example, in general two topologies not coincide. Example 2.3. Consider E = ℓ∞ and e1 = (1, 0, 0, 0...). Then (−e1, e1) is member of τe, but is not belong to τo. Consider xn ∈ ℓ∞ which first n terms are zero and others are 1. Obviously xn ↓ 0, but xn /∈ (−e1, e1) for each n. This example shows that the sequence (xn) is order convergent to zero, but is not topological convergence to zero. On the other hand, since (−e1, e1) /∈ τo, two topologies not coincide. Theorem 2.4. Let E be a Dedekind complete vector lattice with topology τe and (xα) ⊂ E. If xα τe−→ 0, then (xα) is order convergence to zero. τe−→ x, there exists Proof. Assume that a, b ∈ E with x ∈ (a, b) ⊆ E. Since xα α(a,b) such that xα ∈ (a, b) for each α ≥ α(a,b). Put yα(a,b) = b − a. On the other hands, (α(a,b))x∈(a,b) is a directed set with the following order relation α(a,b) ≤ α(c,d) iff (c, d) ⊆ (a, b). It follows that Thus xα o−→ x. xα − x = (xα ∨ x) − (xα ∧ x) ≤ b − a = yα(a,b) ↓ 0. (cid:3) By Example 2.3, the converse of Theorem 2.4 in general not holds. 4 K. HAGHNEJAD AZAR Proposition 2.5. Let E be a Dedekind complete vector lattice and τo be an order topology for E. If B is an ideal and quasi-order closed subset of E, then B is a band in E. o−→ x, we show that x ∈ B. Obversely sup{xα ∧ x} = Proof. Let (xα) ⊆ B and xα x. Set yβ = (Wα6β xα) ∧ x, then yβ ↑ x. Since (yβ) ⊆ B and B is quasi-order closed, follows that x ∈ B and the result follows. (cid:3) References 1. Y. A. Abramovich, C. D. Aliprantis, An Invitation to Operator Theory, Graduate Studies in Mathematics, vol. 50, American Mathematical Society, Providence, RI, 2002. 2. C.D. Aliprantis, and O. Burkinshaw, Positive Operators, Springer, Berlin, 2006. 3. R. Demarr, Partially ordered linear spaces and locally convex linear topological spaces, Illi- nois J. Math., 8, 1964, 601-606. Department of Mathematics, University of Mohaghegh Ardabili, Ardabil, Iran. E-mail address: [email protected]
1010.2342	3	1010	2012-10-25T03:06:37	Fourier transform and rigidity of certain distributions	[ "math.FA" ]	Let $E$ be a finite dimensional vector space over a local field, and $F$ be its dual. For a closed subset $X$ of $E$, and $Y$ of $F$, consider the space $D^{-\xi}(E;X,Y)$ of tempered distributions on $E$ whose support are contained in $X$ and support of whose Fourier transform are contained in $Y$. We show that $D^{-\xi}(E;X,Y)$ possesses a certain rigidity property, for $X$, $Y$ which are some finite unions of affine subspaces.	math.FA	math	FOURIER TRANSFORM AND RIGIDITY OF CERTAIN DISTRIBUTIONS BINYONG SUN AND CHEN-BO ZHU Abstract. Let E be a finite dimensional vector space over a local field, and F be its dual. For a closed subset X of E, and Y of F , consider the space D−ξ(E; X, Y ) of tempered distributions on E whose support are contained in X and support of whose Fourier transform are contained in Y . We show that D−ξ(E; X, Y ) possesses a certain rigidity property, for X, Y which are some finite unions of affine subspaces. 1. Introduction and main result One of the most fundamental results in Euclidean harmonic analysis is the un- certainty principle. As a meta-theorem, it states that a nonzero function and its Fourier transform cannot both be sharply localized. There are various concrete for- malisations of this principle, most famously the Heisenberg uncertainty principle. For a lively and insightful discussion of this topic, see [Tao]. We shall consider distributions ([Sch]). Fix a finite dimensional vector space E and its dual F . For a closed subset X of E, and Y of F , consider the space D−ξ(E; X, Y ) of tempered distributions on E whose support are contained in X and support of whose Fourier transform are contained in Y . The general thrust of the current note is to examine to what extent one is able to separate the support and support of the Fourier transform for distributions in D−ξ(E; ∪X , ∪Y), where X and Y are finite sets of affine subspaces. This is the meaning of rigidity in the title, which the authors consider as another instance of the uncertainty principle. One key observation, which surely has been noted by others before us, is the general importance of relative position of the pair (X, Y ), now assumed to be affine subspaces. As it turns out, there will be three different circumstances, which we respectively call thin, perfect, thick (Equation (2)). As an indication of the relevance of these concepts, we have: (a) D−ξ(E; X, Y ) = 0 if (X, Y ) is a thin pair; (b) D−ξ(E; X, Y ) can be explicitly described in terms of a countable linear basis if (X, Y ) is a perfect pair. This is similar to the classical result of L. Schwartz on the structure of distributions supported on a single point; (c) If (X, Y ) is a thick pair, then D−ξ(E; X, Y ) contains (in a non-canonical fashion) the space of 2000 Mathematics Subject Classification. 42B35, 46F05 (Primary). Key words and phrases. Fourier transform, tempered distributions, affine subspaces, rigidity. 1 2 BINYONG SUN AND CHEN-BO ZHU tempered distributions on a nonzero subspace of E and is thus not rigid in any reasonable sense. Our main result (Theorem A in this section) is that D−ξ(E; ∪X , ∪Y) possesses the afore-mentioned rigidity property as long as there is no pair (X, Y ) ∈ X × Y which is thick. Very roughly the idea goes as follows: By applying a good multiplier operator (a suitable function which vanishes on a part of the support of the Fourier transform), one may cut off that part of the support in the Fourier transform side. At the distribution side, the process will generally yield a distribution with additional support. If no thick pairs are involved, then this process can be carried out in such a way that the additional support is very much controlled. We will have more to say on the detailed strategy later. The result of this note was motivated by certain representation-theoretic issues arising from the proof of archimedean multiplicity-one theorems [SZ]. More specifi- cally the rigidity statement of Theorem A allows us to establish the semi-simplicity and non-negativity of an Euler vector field, crucial in certain descent step involving the so-called distinguished nilpotent orbits. For applications to representation the- ory of algebraic groups over non-archimedean local fields, we will prove our results over an arbitrary local field, rather than over R or C. We note that in Euclidean harmonic analysis and integral geometry, it is of substantial interest to calculate the Fourier (or Radon) transform of distributions with support in an algebraic set. We refer the interested reader to the classical book "Generalized functions" by Gel'fand and Shilov [GS, vol. 1 and 4]. We would also like to point out that for a real quadratic space E and X = Y the null cone, the space D−ξ(E; X, Y ) is of special interest for a variety of reasons. Some related topics are explored in [HT, Chapter 4]. We now introduce some necessary notation for this note. Let k be an arbitrary local field, and let ψ : k → C× be a fixed nontrivial unitary character. Let E be a finite dimensional k-vector space. Denote by C ς (E) ⊂ C−ξ(E), D ς(E) ⊂ D−ξ(E) the (complex) spaces of Schwartz functions, tempered generalized functions, Schwartz densities, and tempered distributions on E, respectively. Thus C−ξ(E) (resp., D−ξ(E)) is the dual of D ς(E) (resp., C ς(E)). Let F be a finite-dimensional k-vector space which is dual to E, i.e., a non- degenerate bilinear map is given. The Fourier transform h , i : E × F → k D ς(F ) → C ς(E) ω 7→ ω FOURIER TRANSFORM AND DISTRIBUTIONS 3 is the linear isomorphism given by Dually for every D ∈ D−ξ(E), its Fourier transform bD ∈ C−ξ(F ) is given by For every closed subset X of E, and Y of F , denote ψ(hx, yi) ω(y), x ∈ E. ω(x) :=ZF bD(ω) := D(ω), ω ∈ D ς (F ). D−ξ(E; X) := {D ∈ D−ξ(E) supp(D) ⊂ X}, and (1) Let X be an affine subspace of E, and Y an affine subspace of F . Denote D−ξ(E; X, Y ) := {D ∈ D−ξ(E; X) supp(bD) ⊂ Y }. L(X) := {u − v u, v ∈ X} the subspace associated to X, and likewise L(Y ) for Y . We say that the pair (X, Y ) is thick, perfect or thin according as (2) L(X)⊥ ( L(Y ), L(X)⊥ = L(Y ), or L(X)⊥ * L(Y ), or what is the same, L(Y )⊥ ( L(X), L(Y )⊥ = L(X), or L(Y )⊥ * L(X), respectively. Example: Take F to be a non-degenerate quadratic space. Suppose that F0 is a non-degenerate nonzero subspace of F and (F0)⊥ = F + ⊕ F − is a decomposition into totally isotropic subspaces F + and F −. Then the pairs (F +, F +), (F + ⊕ F0, F +), and (F + ⊕ F0, F + ⊕ F0) are thin, perfect and thick, respectively. Remark: If (X, Y ) is a perfect pair, then D−ξ(E; X, Y ) =(cid:26) C µX,y0, (C[X] µX,y0) ⊗ D−ξ(L′; {0}), if k is nonarchimedean, if k is archimedean. Here y0 ∈ Y is an arbitrary element, µX,y0 ∈ D−ξ(X) is the product of the func- tion ψ(h·, −y0i) with a fixed L(X)-invariant positive Borel measure on X. In the archimedean case, L′ is an arbitrary subspace of E such that L(X) ⊕ L′ = E. Through addition we have a decomposition E = X × L′. The space C[X] is then the algebra of (complex valued) polynomial functions on X, viewed as a real affine space. 4 BINYONG SUN AND CHEN-BO ZHU We now state the main result of this note. Theorem A. Let X be a finite set of affine subspaces of E, and Y a finite set of affine subspaces of F . Assume that there is no pair (X, Y ) ∈ X × Y which is thick. Then D−ξ(E; ∪X , ∪Y) = D−ξ(E; X, Y ). M (X,Y )∈X ×Y that is perfect Remarks: (i) Theorem A asserts in particular that D−ξ(E; ∪X , ∪Y) = 0, if every (ii) In the archimedean case, D−ξ(E; X, Y ) is a pair (X, Y ) ∈ X × Y is thin. module for the Weyl algebra of E (consisting of (complex) polynomial coefficient differential operators on E). We note that for a perfect pair (X, Y ), the Weyl algebra module D−ξ(E; X, Y ) is irreducible. The strategy to prove Theorem A is to control support and it goes as follows. For every vector u ∈ E, define the following function on F : (3) φu := ψ(hu, ·i). Take a distribution D ∈ D−ξ(E; ∪X , ∪Y). For any Y ∈ Y, pick a nonzero uY ∈ L(Y )⊥. The function φuY takes a constant value on Y , which we denote by cY . Thus φuY − cY will vanish on Y , and multiplying a high power of φuY − cY will cut Y out of the support of the Fourier transform of D. The result is the Fourier transform of a new distribution which is a linear combination of D and translates of D by multiples of uY . Doing this consecutively for different Y 's in Y will thus yield a distribution which is a linear combination of D and translates of D by elements of the lattice in L(Y )⊥ generated by uY 's, and which has a significantly reduced support for its Fourier transform. If X ∈ X is thin with respect to some Y 's, then one could arrange the uY 's (c.f. Lemma 2.1) so that the lattice generated by uY 's is in a favorable position relative to X, resulting in an excellent control on the support of the new distribution. Here are some words on the organization of this note. In Section 2, we show that if X ∈ X has the property that (X, Y ) is thin for every Y ∈ Y, then X in fact does not appear in the support of any D ∈ D−ξ(E; ∪X , ∪Y). This is a form of the uncertainty principle. In Section 3, we show that the rigidity property as claimed in Theorem A holds in the special case, when X (resp. Y) is a finite set of translations of a subspace X0 of E (resp. a subspace Y0 of F ). Section 4 is devoted to the general case. We prove an inductive step in Proposition 4.1. Theorem A will then follow quickly from the results of Sections 2 and 3, and the induction result just alluded to. FOURIER TRANSFORM AND DISTRIBUTIONS 5 Acknowledgements. The authors thank the referee for helpful comments. Biny- ong Sun is supported by NSFC grants 10801126 and 10931006. Chen-Bo Zhu is supported by MOE2010-T2-2-113. 2. Thin pairs and elimination of support Let X be a finite set of affine subspaces of E, and let Y be a finite set of affine subspaces of F , as in the Introduction. Assume that both X and Y are nonempty. Note that an integer may also be considered as an element of k. For any family a = {aY }Y ∈Y ∈ ZY , denote [a] the corresponding element of kY . Lemma 2.1. Assume that X1 ∈ X and that (X1, Y ) is thin for every Y ∈ Y. Let x1 ∈ X1 \ ∪(X \ {X1}). Then there is a family {uY ∈ L(Y )⊥}Y ∈Y of vectors in E with the following property: for every a = {aY }Y ∈Y ∈ ZY with [a] 6= 0, we have Proof. For every X ∈ X and every a = {aY }Y ∈Y ∈ ZY with [a] 6= 0, put XY ∈Y SX,a :=({uY } ∈ YY ∈Y aY uY /∈ ∪X − x1. L(Y )⊥ XY ∈Y aY uY ∈ X − x1) . If X 6= X1, then 0 /∈ SX,a, and SX,a is a proper affine subspace of QY ∈Y L(Y )⊥. If X = X1, then SX,a is a subspace of QY ∈Y L(Y )⊥, and is proper due to the measure zero set ofQY ∈Y L(Y )⊥, and so is the (countable) union ∪X∈X , [a]6=0SX,a. hypothesis that (X1, Y ) is thin for every Y ∈ Y. We finish the proof by taking a vector in In any case each SX,a is a L(Y )⊥) \ (∪X∈X , [a]6=0SX,a). (YY ∈Y (cid:3) For every vector u ∈ E, denote by Tu : D−ξ(E) → D−ξ(E) the push forward of the translation by u, and write (Tuf )(ω) := f (Tuω), f ∈ C−ξ(E), ω ∈ D ς (E). Similar notation applies for v ∈ F . The following is a form of the uncertainty principle. Proposition 2.2. Assume that X1 ∈ X and that (X1, Y ) is thin for every Y ∈ Y. Then we have D−ξ(E; ∪X , ∪Y) = D−ξ(E; ∪(X \ {X1}), ∪Y). Consequently we have D−ξ(E; ∪X , ∪Y) = D−ξ(E; ∪X ′, ∪Y ′), YY ∈Y (φuY − cY )mbD = 0, (TuY − cY )mD = 0, YY ∈Y Xa={aY }∈{0,1,··· ,m}Y ca := YY ∈Y(cid:18) m ua :=XY ∈Y caTuaD = 0, aY(cid:19)(−cY )m−aY , aY uY . or what is the same, (4) where and 6 where BINYONG SUN AND CHEN-BO ZHU X ′ := {X ∈ X (X, Y ) is not thin for some Y ∈ Y}, Y ′ := {Y ∈ Y (X, Y ) is not thin for some X ∈ X }. Proof. Let x1 ∈ X1 \ ∪(X \ {X1}) and {uY ∈ L(Y )⊥}Y ∈Y be as in Lemma 2.1. Let φuY be as in (3), and cY be its common value on Y , as in the Introduction. Let D ∈ D−ξ(E; ∪X , ∪Y). Then where m = 1 if k is nonarchimedean, and m is a sufficiently large positive integer if k is archimedean. The above equality is equivalent to The choice of {uY } ensures that −ua /∈ ∪X − x1 whenever [a] is nonzero. Let U be an open neighborhood of 0 in E, small enough so that (−ua + U) ∩ (∪X − x1) = ∅, for all a ∈ {0, 1, · · · , m}Y such that [a] 6= 0. Since D is supported in ∪X , this implies that (TuaD)x1+U = 0, for all a ∈ {0, 1, · · · , m}Y such that [a] 6= 0. Together with (4), this implies that Dx1+U = 0. Since x1 is arbitrary, we conclude that D is supported in ∪(X \ {X1}). (cid:3) 3. A special case Proposition 3.1. If X is a finite set of translations of a subspace X0 of E, and Y is a finite set of translations of a subspace Y0 of F , then (5) D−ξ(E; ∪X , ∪Y) = M(X,Y )∈X ×Y D−ξ(E; X, Y ). FOURIER TRANSFORM AND DISTRIBUTIONS 7 Proof. Assume that X ⊥ sides of (5) are 0 by Proposition 2.2, and there is nothing to prove. 0 ⊆ Y0, i.e., (X0, Y0) is not a thin pair. Otherwise both Let D ∈ D−ξ(E; ∪X , ∪Y). For every X ∈ X , denote by DX ∈ D−ξ(E; X) the distribution which coincides with D on a neighborhood of X. Then D = XX∈X DX. Note that ∪X is a disjoint union, by our assumption on X . Assume that k is archimedean. Let P be a real polynomial function whose zero locus is ∪Y. Then there is a positive integer k such that P kbD = 0, or equivalently, (eP )kD = 0, where eP is a certain constant coefficient differential operator, acting on the space D−ξ(E). Therefore for all X ∈ X , (eP )kDX = 0, which implies that DX ∈ D−ξ(E; X, ∪Y). This proves that D−ξ(E; X, ∪Y). (6) D−ξ(E; ∪X , ∪Y) = MX∈X Now assume that k is nonarchimedean. Fix X1 ∈ X . For any X ∈ X \ {X1}, choose a vector vX ∈ X ⊥ 0 such that ψ(hX, vXi) 6= ψ(hX1, vXi). Here ψ(hX, vXi) stands for ψ(hu, vXi), which is independent of u ∈ X, and ψ(hX1, vXi) is defined similarly. Then we have that  YX∈X \{X1} (φvX − ψ(hX, vXi)) D = YX∈X \{X1} (This is not true when k is archimedean.) (ψ(hX1, vXi − ψ(hX, vXi)) DX1. The Fourier transform of the left hand side of the above equality is  YX∈X \{X1} (TvX − ψ(hX, vXi))bD. Since vX ∈ X ⊥ 0 ⊆ Y0, and since ∪Y is invariant under translation by elements of Y0, the above generalized function is again supported in ∪Y. Therefore the Fourier transform of DX1 is also supported in ∪Y. This proves (6) in the nonarchimedean case. Applying (6) to the pair Y and {X}, we have that (7) D−ξ(E; X, ∪Y) =MY ∈Y D−ξ(E; X, Y ). We finish the proof by combining (6) and (7). (cid:3) 8 BINYONG SUN AND CHEN-BO ZHU 4. The general case As before, let X be a finite set of affine subspaces of E, and let Y be a finite set of affine subspaces of F . We start with the following Proposition 4.1. Let X0 be a subspace of E and Y0 a subspace of F . Write X0 := {X ∈ X L(X) = X0} and Y0 := {Y ∈ Y L(Y ) = Y0}. Assume that (X ′, Y0) and (X0, Y ′) are thin for all X ′ ∈ X ′ := X \ X0 and all Y ′ ∈ Y ′ := Y \ Y0. Then (8) D−ξ(E; ∪X , ∪Y) = D−ξ(E; ∪X0, ∪Y0) ⊕ D−ξ(E; ∪X ′, ∪Y ′). Proof. Proposition 2.2 implies that the right hand side of (8) is a direct sum. Let {uY ′ ∈ L(Y ′)⊥}Y ′∈Y ′ be a family of vectors in E such that for every family a = {aY ′}Y ′∈Y ′ ∈ ZY ′ with [a] 6= 0, we have XY ′∈Y ′ aY ′uY ′ /∈ [X1,X2∈X0 (X1 − X2). The existence of such a family is proved along the same line as that of Lemma 2.1. Let φuY ′ be as in (3), and cY ′ be its common value on Y ′, as in the Introduction. Let D ∈ D−ξ(E; ∪X , ∪Y). We take m to be a sufficiently larger positive integer if k is archimedean and m = 1 if k is non-archimedean. Then the generalized function is supported in ∪Y0. It is the Fourier transform of the distribution YY ′∈Y ′ D′ := YY ′∈Y ′ (φuY ′ − cY ′)m!bD (TuY ′ − cY ′)m! D. D′ = Xa={aY ′ }∈{0,1,··· ,m}Y′ caTuaD, ca := YY ′∈Y ′(cid:18) m ua := XY ′∈Y ′ aY ′(cid:19)(−cY ′)m−aY ′ , aY ′uY ′. By expansion, we have (9) where and FOURIER TRANSFORM AND DISTRIBUTIONS 9 For every set Z of affine subspaces of E, we put eZ :=nua + Z a ∈ {0, 1, · · · , m}Y ′ , Z ∈ Zo . since (X ′, Y0) is thin for all X ′ ∈ X ′, Proposition 2.2 implies that it is supported Then D′ is clearly supported in ∪eX . Its Fourier transform is supported in ∪Y0, and in ∪eX0. Now the choice of {uY ′} ensures that (a disjoint union), fX0 = X0 ⊔ X1 where X1 :=nua + X a ∈ {0, 1, · · · , m}Y ′ , [a] 6= 0, X ∈ X0o . By Proposition 3.1, we may write (10) where D0 ∈ D−ξ(E; ∪X0, ∪Y0) and D1 ∈ D−ξ(E; ∪X1, ∪Y0). D′ = D0 + D1, The disjointness of X0 and X1 allows us to choose an open neighborhood U of 0 in E, small enough so that Let and (11) x0 ∈ ∪X0 \ ∪fX ′. (x0 + U) ∩ (∪fX ′) = ∅, (x0 + U) ∩ (∪X1) = ∅. For all nonzero [a], we thus have and therefore (x0 + U − ua) ∩ (∪X ) = ∅ (TuaD)x0+U = 0. Then (9) implies that D′x0+U = c0Dx0+U , with c0 = YY ′∈Y ′ (−cY ′)m, and (10) and (11) implies that (D′ − D0)x0+U = D1x0+U = 0. We thus conclude from the last two equalities that D − D0/c0 vanishes on x0 + U. Since x0 is arbitrary, we see that D − D0/c0 is supported in where (∪X ) \ (∪X0 \ ∪fX ′) ⊂ (∪X ′) ∪ (∪(X0 ∧fX ′)). X0 ∧fX ′ = {X0 ∩ X ′ X0 ∈ X0, X ′ ∈fX ′}. 10 BINYONG SUN AND CHEN-BO ZHU see that D − D0/c0 actually belongs to Since every pair in (X0 ∧fX ′) × Y is thin, and every pair in X ′ × Y0 is thin, we D−ξ(E; (∪X ′) ∪ (∪(X0 ∧fX ′)), ∪Y) = D−ξ(E; ∪X ′, ∪Y) = D−ξ(E; ∪X ′, ∪Y ′), by two applications of Proposition 2.2. This finishes the proof of the current proposition. (cid:3) We are now ready to prove Theorem A. Assume that no pair in X × Y is thick. Put For every L ∈ L, put L := {L(X) (X, Y ) ∈ X × Y is perfect}. XL := {X ∈ X L(X) = L} and YL⊥ := {Y ∈ Y L(Y ) = L⊥}. Then we have D−ξ(E; ∪X , ∪Y) (by Proposition 2.2) (by Proposition 4.1) = D−ξ(E; ∪(∪L∈LXL), ∪(∪L∈LYL⊥)) D−ξ(E; ∪XL, ∪YL⊥) = ML∈L ML∈L, (X,Y )∈XL×YL⊥ = D−ξ(E; X, Y )) (by Proposition 3.1) D−ξ(E; X, Y ). = M (X,Y )∈X ×Y that is perfect This finishes the proof of Theorem A. References [GS] I.M. Gel'fand and G.E. Shilov, Generalized functions, Vol. 1–5, Academic Press, 1966–1968. [HT] R. Howe and E.C. Tan, Non-abelian harmonic analysis: applications of SL(2, R), Univer- sitext, Springer Verlag, 1992. [Sch] L. Schwartz, Theorie des distributions, Hermann, 1966. [SZ] B. Sun and C.-B. Zhu, Multiplicity one theorems: the Archimedean case, Annals Math. 175, (2012), no. 1, 23–44. [Tao] T. Tao, The uncertainty principle, http://terrytao.wordpress.com/2010/06/25/the-uncertainty-principle/. Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Beijing, 100190, P.R. China E-mail address: [email protected] Department of Mathematics, National University of Singapore, Block S17, 10 Lower Kent Ridge Road, Singapore 119076 E-mail address: [email protected]
1804.08518	1	1804	2018-04-23T15:44:47	The Bezout equation on the right half plane in a Wiener space setting	[ "math.FA" ]	This paper deals with the Bezout equation $G(s)X(s)=I_m$, $\Re s \geq 0$, in the Wiener space of analytic matrix-valued functions on the right half plane. In particular, $G$ is an $m\times p$ matrix-valued analytic Wiener function, where $p\geq m$, and the solution $X$ is required to be an analytic Wiener function of size $p\times m$. The set of all solutions is described explicitly in terms of a $p\times p$ matrix-valued analytic Wiener function $Y$, which has an inverse in the analytic Wiener space, and an associated inner function $\Theta$ defined by $Y$ and the value of $G$ at infinity. Among the solutions, one is identified that minimizes the $H^2$-norm. A Wiener space version of Tolokonnikov's lemma plays an important role in the proofs. The results presented are natural analogs of those obtained for the discrete case in [11].	math.FA	math	THE B´EZOUT EQUATION ON THE RIGHT HALF PLANE IN A WIENER SPACE SETTING G.J. GROENEWALD, S. TER HORST, AND M.A. KAASHOEK Abstract. This paper deals with the B´ezout equation G(s)X(s) = Im, ℜs ≥ 0, in the Wiener space of analytic matrix-valued functions on the right half plane. In particular, G is an m × p matrix-valued analytic Wiener function, where p ≥ m, and the solution X is required to be an analytic Wiener function of size p × m. The set of all solutions is described explicitly in terms of a p × p matrix-valued analytic Wiener function Y , which has an inverse in the analytic Wiener space, and an associated inner function Θ defined by Y and the value of G at infinity. Among the solutions, one is identified that minimizes the H 2- norm. A Wiener space version of Tolokonnikov's lemma plays an important role in the proofs. The results presented are natural analogs of those obtained for the discrete case in [11]. 1. Introduction and main results In this paper we deal with the B´ezout equation G(s)X(s) = Im on the closed right half plane ℜs ≥ 0, assuming that the given function G is of the form G(s) = D +Z ∞ 0 e−stg(t) dt (ℜs ≥ 0), g ∈ L1 where m×p(R+) ∩ L2 (1.1) In particular, G belongs to the analytic Wiener space W m×p solutions X ∈ W p×m (1.2) X(s) = DX +Z ∞ e−stx(t) dt , that is, + + 0 (ℜs ≥ 0), where x ∈ L1 m×p(R+). m×p(R+). . We are interested in Throughout p ≥ m. We refer to the final paragraph of this introduction for a further explanation of the notation. With G given by (1.1) we associate the Wiener-Hopf operator TG mapping p(R+) into L2 m(R+) which is defined by L2 (1.3) (TGh)(t) = Dh(t) +Z ∞ 0 g(t − τ )h(τ )dτ, t ≥ 0 (h ∈ L2 m(R+)). 2010 Mathematics Subject Classification. Primary 47A56; Secondary 47A57, 47B35, 46E40, 46E15. Key words and phrases. B´ezout equation, corona problem, Wiener space on the line, matrix-valued functions, minimal norm solutions, Tolokonnikov's lemma. The third author gratefully thanks the mathematics department of North-West University, Potchefstroom campus, South Africa, for the hospitality and support during his visit from Sep- tember 21 -- October 15, 2015. This work is based on the research supported in part by the National Research Foundation of South Africa (Grant Number 93406). 1 2 G.J. GROENEWALD, S. TER HORST, AND M.A. KAASHOEK For X as in (1.2) we define the Wiener-Hopf operator TX mapping L2 L2 p(R+) in a similar way, replacing D by DX and g by x. If the B´ezout equation m(R+) into (1.4) G(s)X(s) = Im, ℜs ≥ 0. has a solution X as in (1.2), then (using the analyticity of G and X) the theory of Wiener-Hopf operators (see [7, Section XII.2] or [2, Section 9]) tells us that TGTX = TGX = I, where I stands for the identity operator on L2 m(R+). Thus for the B´ezout equation (1.4) to be solvable the operator TG must be surjective or, equivalently, TGT ∗ G must be strictly positive. We shall see that this condition is also sufficient. To state our main results, we assume that TGT ∗ G is strictly positive. Then D = G(∞) is surjective, and hence DD∗ is strictly positive too. We introduce two matrices D+ and E, of sizes p× m and p× (p− m), respectively, and a p × p matrix function Y in W p×p + , as follows: (i) D+ = D∗(DD∗)−1, where D = G(∞); (ii) E is an isometry mapping Cp−m into Cp such that Im E = Ker D; (iii) Y is the p × p matrix function given by Y (s) = Ip −Z ∞ 0 e−sty(t)dt, ℜs ≥ 0, where y = T ∗ G(TGT ∗ G)−1g. From the definitions of D+ and E it follows that the p × p matrix (cid:2)D+ E(cid:3) is non-singular. In fact (1.5) (1.6) (1.8) (cid:20) D E ∗(cid:21)(cid:2)D+ E(cid:3) =(cid:20)Im 0 0 Ip−m(cid:21) . As we shall see (Proposition 2.1 in Section 2 below), the fact that the given function g ∈ L1 m×p(R+) ∩ L2 In particular, Y ∈ W p×p (1.7) m×p(R+) implies that a similar result holds true for y. + . In what follows Ξ and Θ are the functions defined by 0 Ξ(s) =(cid:16)Ip −Z ∞ Θ(s) =(cid:16)Ip −Z ∞ + , we have Ξ ∈ W p×m + 0 e−sty(t)dt(cid:17)D+ = Y (s)D+, ℜs > 0; e−sty(t)dt(cid:17)E = Y (s)E, , ℜs > 0. + + and Θ ∈ W p×(p−m) . Finally, recall that a is inner whenever Ω(s) is an isometry Since Y ∈ W p×p function Ω in the analytic Wiener space W k×r for each s ∈ iR. We now state our main results. Theorem 1.1. Let G be the m × p matrix-valued function given by (1.1). Then the equation G(s)X(s) = Im, ℜs > 0, has a solution X ∈ W p×m if and only if TG is right invertible. In that case the function Ξ defined by (1.7) is a particular solution and the set of all solutions X ∈ W p×m (1.9) is given by X(s) = Ξ(s) + Θ(s)Z(s), ℜs > 0, + + where Ξ and Θ are defined by (1.7) and (1.8), respectively, and the free parameter . Moreover, the function Θ belongs to and is inner. Furthermore, the solution Ξ is the minimal H 2 solution Z is an arbitrary function in W (p−m)×m W p×(p−m) + + THE B´EZOUT EQUATION ON THE RIGHT HALF PLANE IN A WIENER SPACE SETTING3 in the following sense kX(·)uk2 H 2 p (1.10) = kΞ(·)uk2 H 2 p + kZ(·)uk2 H 2 p−m , where u ∈ Cm and Z ∈ fW (p−m)×m + . In the above theorem, for any positive integer k, H 2 k = H 2 space of Ck-valued functions on the right half plane given by H 2 where J is the unitary operator defined by 1 √2πF : L2 k(R) → L2 k(iR) (1.11) J = k (iR) is the Hardy k (iR) = J L2 k(R+), k(R) onto L2 k(iR). Moreover, Z ∈ + means that with F being the Fourier transform mapping L2 fW (p−m)×m Z(s) = DZ +Z ∞ (ℜs ≥ 0), where DZ is a (p − m) × p matrix and z ∈ L1 e−stz(t) dt 0 (p−m)×p(R+). See the final part of this introduction for further information about the used nota- tion, in particular, see (1.14) for the definition of the Fourier transform F . setting. The result emphasizes the central role of the function Y . The second theorem is a variant of the Tolokonnikov lemma [20] in the present (p−m)×p(R+) ∩ L2 G is strictly positive, and let Y be the matrix function + , det Y (s) 6= 0 whenever + . Furthermore, the p× p matrix function Theorem 1.2. Assume TGT ∗ defined by (1.5). Then Y belongs to the Wiener space W p×p ℜs ≥ 0, and hence Y is invertible in W p×p (1.12) (cid:20) G(s) E ∗Y (s)−1(cid:21) , ℜs ≥ 0, is invertible in the Wiener algebra W p×p (1.13) (cid:20) G(s) E ∗Y (s)−1(cid:21)−1 + and its inverse is given by = Y (s)(cid:2)D+ E(cid:3) =(cid:2)Ξ(s) Θ(s)(cid:3) , ℜs ≥ 0. The literature on the B´ezout equation and the related corona problem is exten- sive, starting with Carleson's corona theorem [3] (for the case when m = 1) and Fuhrmann's extension to the matrix-valued case [6], both in a H ∞ setting. The topic has beautiful connections with operator theory (see the books [14], [16], [17], [18], and the more recent papers [21], [22], [23]). Rational matrix equations of the form (1.4) play an important role in solving systems and control theory problems, in particularly, in problems involving coprime factorization, see, e.g., [24, Section 4.1], [10, Section A.2], [25, Chapter 21]. For more recent work see [12] and [13], and [15, page 3] where it is proved that the scalar analytic Wiener algebra is a pre-B´ezout ring. For matrix polynomials, the equation (1.4) is closely related to the Sylvester resultant; see, e.g., Section 3 in [9] and the references in that paper. The present paper is inspired by [5] and [11]. The paper [5] deals with equation (1.4) assuming the matrix function G to be a stable rational matrix function, and the solutions are required to be stable rational matrix functions as well. The com- ment in the final paragraph of [5, Section 2] was the starting point for our analysis. The paper [11] deals with the discrete case (when the right half plane is replaced by the open unit disc). Theorems 1.1 and 1.2 are the continuous analogue of Theorem 4 G.J. GROENEWALD, S. TER HORST, AND M.A. KAASHOEK 1.1 in [11]. The absence of an explicit formula for the function Y −1 in the present setting makes the proofs more complicated than those in [11]. The paper consists of five sections, including the present introduction and an appendix. Section 2, which deals with the right invertibility of the operator TG, has an auxiliary character. Theorem 1.2 is proved in Section 3, and Theorem 1.1 in Section 4. The Appendix, Section A, contains a number of auxiliary results involving the Lebesgue space L1(R) ∩ L2(R) and its vector-valued counterpart, which are collected together simply for the convenience of the reader and contains no significantly new material. Notation and terminology. We conclude this section with some notation and termi- nology. Throughout, a linear map A : Cr → Ck is identified with the k×r matrix of A relative to the standard orthonormal bases in Cr and Ck. The space of all k × r k×r(R). As usual bf denotes the matrices with entries in L1(R) will be denoted by L1 Fourier transform of f ∈ L1 (1.14) k×r(R), that is, e−stf (t)dt, s ∈ iR. bf (s) = (F f )(s) =Z ∞ −∞ Note that bf is continuous on the extended imaginary axis iR∪ {± i∞}, and is zero at ± i∞ by the Riemann-Lebesgue lemma. By W k×r we denote the Wiener space consisting of all k × r matrix functions F on the imaginary axis of the form (1.15) F (s) = DF + bf (s), s ∈ iR, where f ∈ L1 k×r(R) and DF is a constant matrix. Since bf is continuous on the extended imaginary axis and is zero at ±i∞, the function F given by (1.15) is also continuous on the extended imaginary axis and the constant matrix DF is equal to the value of F at infinity. We write W k×r for the space of all F of the form (1.15) with the additional property that f has its support in R+ = [0,∞), that is, f is equal to zero on (−∞, 0). Any function F ∈ W k×r is analytic and bounded on the open right half plane. Thus any F ∈ W k×r + is a matrix- valued H ∞ function. Finally, by W k×r −,0 we denote the Wiener space consisting of all F of the form (1.15) with the additional property that DF = 0 and f has its support in (−∞, 0]. Thus we have the following direct sum decomposition: (1.16) + + W k×r = W k×r + +W k×r −,0 . + if the function f in (1.15) belongs to L1 We write F ∈ fW k×r Similarly, F ∈ fW k×r Let F ∈ W k×r be given by (1.15). With F we associate the Wiener-Hopf op- k(R+). This operator (see [7, Section XII.2]) is k×r(R−) and DF = 0. −,0 if f ∈ L1 k×r(R+) ∩ L2 k×r(R−) ∩ L2 r(R+) into L2 k×r(R+). erator TF mapping L2 defined by (1.17) (TF h)(t) = DF h(t) +Z ∞ f (t − τ )h(τ )dτ, t ≥ 0 (h ∈ L2 r(R+)). The orthogonal complement of H 2 will be denoted by K 2 belongs to H 2 k(iR). If F ∈ fW k×r k (iR). Similarly, F (·)u belongs to K 2 k(R+), with J as in (1.11), in L2 k(iR) + , then for each u ∈ Cr the function F (·)u k(iR) if F ∈ fW k×r −,0 . 0 k (iR) = J L2 THE B´EZOUT EQUATION ON THE RIGHT HALF PLANE IN A WIENER SPACE SETTING5 Finally, for f ∈ L1 function in L1 k×m(R), see [19, Section 7.13], given by r×m(R) the convolution product f ⋆ g is the k×r(R) and g ∈ L1 (f ⋆ g)(t) =Z ∞ −∞ (1.18) f (t − τ )g(τ ) dτ a.e. on R. 2. Right invertibility of TG In this section G ∈ W m×p know that the B´ezout equation (1.4) having a solution X in W p×m TG is right invertible or, equivalently, TGT ∗ containing formula (1.4). , where G is given by (1.1) and p ≥ m. We already implies that G is strictly positive; see the paragraph + + In this section we present an auxiliary result that will be used to prove our main theorems. For this purpose we need the m × m matrix-valued function R on the imaginary axis defined by R(s) = G(s)G(s)∗, s ∈ iR. It follows that R ∈ W m×m. By TR we denote the corresponding Wiener-Hopf operator acting on L2 m(R+). Thus (TRf )(t) = DD∗f (t) +Z ∞ r(t − τ )f (τ ) dτ, with r(t) = Dg∗(t) + g(t)D∗ +Z ∞ 0 −∞ 0 ≤ t < ∞, g(t − τ )g∗(τ ) dτ, t ∈ R. Here g∗(t) = g(−t)∗ for t ∈ R. It is well-known (see, e.g., formula (24) in Section XII.2 of [7]) that (2.1) Here HG is the Hankel operator mapping L2 is, TR = TGT ∗ G + HGH ∗ G. (2.2) (HGf )(t) =Z ∞ 0 g(t + τ )f (τ )dτ, f ∈ L2 p(R+). p(R+) into L2 m(R+) defined by G, that We shall prove the following proposition. For the case when G is a rational matrix function, the first part (of the "if and only if" part) of the proposition is covered by Lemma 2.3 in [5]. The proof given in [5] can also be used in the present setting. For the sake of completeness we include a proof of the first part. Proposition 2.1. Let G be given by (1.1). Then the operator TG is right invertible if and only if TR and I − H ∗ R HG are both invertible operators. In that case the inverse of TGT ∗ G is given by (TGT ∗ G)−1 = T −1 R HG)−1H ∗ R + T −1 GT −1 R . GT −1 GT −1 (2.3) R HG(I − H ∗ Furthermore, m(R+) ∩ L2 (a) (TGT ∗ (b) the function y defined by y = T ∗ G)−1 maps L1 m(R+) in a one-to-one way onto itself; p×p(R+)∩L2 G(TGT ∗ G)−1g belongs to L1 + . p×p(R+), in particular, the function Y given by (1.5) is in fW p×p Proof. We split the proof into four parts. In the first part we assume that TG is right invertible, and we show that TR and I − H ∗ R HG are both invertible operators and that the inverse of TGT ∗ G is given by (2.3). The second part deals with the reverse implication. Items (a) and (b) are proved in the last two parts. Part 1. Assume TG is right invertible. Then the operator TGT ∗ According to (2.1) we have TR ≥ TGT ∗ G is strictly positive. G, and hence TR is also strictly positive. GT −1 6 G.J. GROENEWALD, S. TER HORST, AND M.A. KAASHOEK In particular, TR is invertible. Rewriting (2.1) as TGT ∗ multiplying the latter identity from the left and from the right by T that G = TR − HGH ∗ −1/2 R G, and shows (2.4) T −1/2 R GT −1/2 R −1/2 R HGH ∗ GT −1/2 R . TGT ∗ −1/2 R = I − T R HG = (cid:0)H ∗ −1/2 R HGH ∗ GT is strictly positive which shows that H ∗ GT −1 is a Hence I − T strict contraction. But then H ∗ is also a strict contraction, and thus the operator I − H ∗ R HG is strictly positive. In particular, I − H ∗ G, a usual Schur complement type of argument (see, e.g., Section 2.2 in [1]), including the well- known inversion formula (cid:1)∗ G = TR − HGH ∗ R HG is invertible. Finally, since TGT ∗ (cid:1)(cid:0)H ∗ GT −1 GT −1 −1/2 R −1/2 R GT GT GT −1/2 R (A − BC)−1 = A−1 + A−1B(I − CA−1B)−1CA−1, GT GT −1 G)−1 is given by (2.3). G = TR − HGH ∗ then shows that (TGT ∗ Part 2. In this part we assume that TR and I − H ∗ R HG are both invertible operators, and we show that TG is right invertible. According to (2.1) the operator TR is positive. Since we assume TR to be invertible, we conclude that TR is strictly positive. Rewriting (2.1) as TGT ∗ G, and multiplying the latter we obtain the identity (2.4). identity from the left and from the right by T −1/2 R HGH ∗ is a contraction. Hence I−T GT −1 But then H ∗ is also a contraction, and thus the GT −1 operator I − H ∗ R HG is invertible. −1/2 It follows that I − H ∗ R HG is a strict −1/2 . This implies contraction. But then the same holds true for T R is strictly positive, and (2.4) shows that TG is right that I − T invertible. Part 3. In this part we prove item (a). Observe that g ∈ L1 TG maps L1 R HG is strictly positive, and hence T GT is positive which shows that H ∗ −1/2 R R HG is positive. By assumption I − H ∗ R HG =(cid:0)H ∗ m×p(R) implies that −1/2 R HGH ∗ −1/2 R HGH ∗ (cid:1)(cid:0)H ∗ p(R+) into L1 p×m(R) and GT −1 GT −1 −1/2 R −1/2 R −1/2 R −1/2 R −1/2 R (cid:1)∗ GT m(R+) ∩ L2 m(R+). Since g∗ ∈ L1 GT GT GT p(R+) ∩ L2 (T ∗ Gf )(t) = D∗f (t) +Z ∞ m(R+)∩ L2 g∗(t − τ )f (τ ) dτ, 0 m(R+) into L1 0 ≤ t < ∞, p(R+). Thus TGT ∗ G maps m(R+) into itself. We have to show that the same holds true for its p(R+)∩ L2 G maps L1 the operator T ∗ L1 m(R+) ∩ L2 inverse. To do this we apply Lemmas A.3 and A.4. Lemma A.3 tells us that T −1 m(R+)∩ L2 itself. This allows us to apply Lemma A.4 with R maps L1 m(R+) in a one-to-one way onto Q = T −1 R , H = HG and H = H ∗ G. Recall that HG is a Hankel operator, see (2.2), and H ∗ in fact G is also a Hankel operator, (H ∗ Gf )(t) =Z ∞ 0 g(t + τ )∗f (τ ) dτ, 0 ≤ t < ∞. Since I − H ∗ L1 p(R+)∩ L2 its inverse (I − H ∗ m(R+)∩L2 maps L1 m(R+) ∩ L2 into L1 R HG is invertible, Lemma A.4 then shows that I − H ∗ GT −1 R HG maps p(R+) in a one-to-one way onto itself, and hence the same holds true for GT −1 R p(R+) m(R+) ∩ R HG)−1. To complete the proof of item (a) note that H ∗ p(R+)∩L2 GT −1 m(R+) into L1 m(R+). But then (2.3) shows that (TGT ∗ p(R+), and T −1 R HG maps L1 G)−1 maps L1 p(R+)∩L2 GT −1 THE B´EZOUT EQUATION ON THE RIGHT HALF PLANE IN A WIENER SPACE SETTING7 m(R+) ∩ L2 m(R+) ∩ L2 G)−1 is one-to-one on L1 m(R+) into itself. To see that (TGT ∗ m(R+) and m(R+) onto itself, one can follow the same argumentation as in L2 maps L1 the last part of the proof of Lemma A.3. Part 4. In this part we prove item (b). Since g belongs to L1 G)−1g also belongs to L1 item (a) tells us that f := (TGT ∗ We already have seen (in the first paragraph of the previous part) that T ∗ L1 m(R+) ∩ L2 L1 p×p(R+) ∩ L2 m×p(R+), m×p(R+). G maps Gf belongs to (cid:3) m(R+) into L1 p×p(R+), as desired. m×p(R+)∩ L2 m×p(R+) ∩ L2 It follows that y = T ∗ p(R+) ∩ L2 p(R+). 3. The functions Y and Θ, and proof of Theorem 1.2 We begin with three lemmas involving the functions Y and Θ defined by (1.5) and (1.8), respectively. From Proposition 2.1, item (b), and (1.8) we know that Y ∈ W p×p Lemma 3.1. Assume that TG is right invertible, and let Y ∈ W p×p defined by (1.5). Then ; see also the paragraph preceding Theorem 1.1. and Θ ∈ W p×(p−m) be the function + + + (3.1) G(s)Y (s) = D, ℜs > 0. G)−1g = g. Since the functions Proof. To prove (3.1) note that TGy = TGT ∗ g and y both have their support in R+, the identity TGy = g can be rewritten as Dy + g ⋆ y = g, where ⋆ is the convolution product of matrix-valued functions with entries in L1(R); see (1.18). Thus G(TGT ∗ (3.2) Dy(t) + (g ⋆ y)(t) = Dy(t) +Z ∞ −∞ g(t − τ )y(τ )dt = g(t), t ∈ R. Next use that the Fourier transform of a convolution product is just the product of the Fourier transforms of the functions in the convolution product. Thus taking Fourier transforms in (3.2) yields Dby +bgby =bg. The latter identity can be rewritten as Gby =bg. Hence, using the definition of Y in (1.5), we obtain G(s)Y (s) = G(s)(cid:16)Ip −by(s)(cid:17) = G(s) −bg(s) = D. This proves (3.1). (cid:3) + Lemma 3.2. Assume that TG is right invertible. Then the function Θ defined by (1.8) belongs to W p×(p−m) and is an inner function, that is, Θ(s) is an isometry for each s ∈ iR and at infinity. Proof. We already know that Θ ∈ W p×(p−m) G)−1g as in (1.5), and put f = (TGT ∗ y = T ∗ Proposition 2.1 (b), and y = T ∗ . To prove that Θ is inner, let G)−1g. Thus f ∈ L1 m×m(R+), by Gf . The latter can be rewritten as G(TGT ∗ + g∗(t − τ )f (τ ) dτ, t ≥ 0. y(t) = D∗f (t) +Z ∞ y(t) = D∗f (t) +Z ∞ −∞ 0 Note that g∗(t) = g(−t)∗, and hence g∗ has its support in (−∞, 0]. Therefore (3.3) g∗(t − τ )f (τ ) dτ, t ∈ R. Put (3.4) ρ(t) =(cid:26) 0 when t ≥ 0, (g∗ ⋆ f )(t) when t < 0. 8 G.J. GROENEWALD, S. TER HORST, AND M.A. KAASHOEK Using the definition of the convolution product ⋆, see (1.18), we can rewrite (3.3) as Taking Fourier transforms we obtain y(t) = D∗f (t) + (g∗ ⋆ f )(t) − ρ(t) t ∈ R. by(s) = D∗bf (s) + bg∗(s)bf (s) −bρ(s) = G(s)∗bf (s) −bρ(s), s ∈ iR. Hence, we have s ∈ iR. E ∗Y (s)∗Y (s)E = Y (s) = I − G(s)∗bf (s) +bρ(s), (3.5) Now let us compute Θ(s)∗Θ(s) = E ∗Y (s)∗Y (s)E for s ∈ iR. We have = E ∗Y (s)∗E − E ∗Y (s)∗G(s)∗bf (s)E + E ∗Y (s)∗bρ(s)E = E ∗Y (s)∗E + E ∗Y (s)∗bρ(s)E = E ∗E − E ∗by(s)∗E + E ∗Y (s)∗bρ(s)E = Ip−m − Ω(s). −,0 , and thus Ω belongs to W (p−m)×(p−m) Here Ω(s) = E ∗by(s)∗E−E ∗Y (s)∗bρ(s)E. Note that the functionsby(·)∗ and Y (·)∗bρ(·) belong to W p×p . On the other hand, the function E ∗Y (·)∗Y (·)E is hermitian on the imaginary axis, and hence the same is true for Ω. But for any positive integer k we have (because G(s)Y (s)E = 0 by (3.1) and DE = 0) −,0 W k×k −,0 ∩ (W k×k −,0 )∗ = {0}. + be the function Thus Ω is identically zero, and thus Θ(s)∗Θ(s) = E ∗Y (s)∗Y (s)E = Ip−m for any s ∈ iR. Moreover, Θ(∞)∗Θ(∞) = E ∗E = I. This proves that Θ is inner. (cid:3) Lemma 3.3. Assume that TG is right invertible, and let Y ∈ W p×p defined by (1.5). Then Y is invertible in W p×p + . Proof. Fix s ∈ iR, and assume u ∈ Cp such that Y (s)u = 0. Then G(s)Y (s) = D implies that Du = 0. By definition of E, u = Ev for some v ∈ Cp−m. Next use Θ(s) = Y (s)E. It follows that Θ(s)v = Y (s)Ev = Y (s)u = 0. However, Θ(s) is an isometry, by Lemma 3.2. So v = 0, and hence u = 0. We see that det Y (s) 6= 0. Also Y (∞) = Ip. We conclude that TY is a Fredholm operator; see [7, Theorem XII.3.1]. Next we prove that Ker TY = {0}. Take h ∈ Ker TY . Then TY h = 0, and hence Y (s)bh(s) = 0 for each s ∈ iR. But det Y (s) 6= 0 for each s ∈ iR. Hence bh = 0, and We want to prove that TY is invertible. Given the results of the preceding first two paragraphs it suffices to show that ind TY = 0. This will be done in the next step by an approximation argument, using the fact, from [5], that we know the result is true for rational matrix functions. therefore h = 0. Let g be as in (1.1). Note that g is the limit in L1 of a sequence g1, g2, . . . such of [7]. Since TG is right invertible, TGn will also be right invertible for n sufficiently )−1gn. large. In fact, TGnT ∗ that Gn(s) = D+cgn(s) is a stable rational matrix function; cf., Part (v) on page 229 Then yn → y in the L1-norm. Put Yn(s) = I −cyn(s). Then TYn → TY in operator G in operator norm. Put yn = T ∗ Gn Gn → TGT ∗ (TGnT ∗ Gn THE B´EZOUT EQUATION ON THE RIGHT HALF PLANE IN A WIENER SPACE SETTING9 norm. For n sufficiently large the operator TYn is invertible (see the paragraph preceding Theorem 1.2 in [5] and formula (2.17) in [5]). In particular, the Fredholm index of TYn is zero. But ind TY = limn→∞ ind TYn = 0. Thus TY is invertible, and hence Y is invertible in W p×p + . (cid:3) Proof of Theorem 1.2. From Lemma 3.3 we know that Y ∈ W p×p and that Y is invertible in W p×p + . Thus we only have to prove the second part of the theorem. Since Y is invertible in W p×p + , the p × p matrix function given by (1.12) is well- defined and belongs to W p×p + . Furthermore, from (1.6) we know that the p × p matrix (cid:2)D+ E(cid:3) is invertible. Hence the function defined by the right hand side of (1.13) belongs to W p×p and is invertible in W p×p (1.6) we see that (cid:20) G(s) E ∗Y (s)−1(cid:21) Y (s)(cid:2)D+ E(cid:3) =(cid:20)G(s)Y (s) + . Using (3.1) and the identity E ∗ + + (cid:21)(cid:2)D+ E(cid:3) Ip−m(cid:21) , ℜs ≥ 0. 0 =(cid:20) D E ∗(cid:21)(cid:2)D+ E(cid:3) =(cid:20)Im 0 This proves the first identity (1.13). The second identity is an immediate conse- quence of the definitions of Ξ and Θ in (1.7) and (1.8), respectively. (cid:3) 4. Proof of Theorem 1.1 We begin with a lemma concerning the functions Ξ and Θ. Lemma 4.1. Assume that TG is right invertible, and let Ξ and Θ be the functions defined by (1.7) and (1.8), respectively. Then (4.1) Ker TG = TΘL2 p−m(R+) and Θ∗Ξ ∈ fW (p−m)×m 0,− . Proof. We split the proof into two parts. Part 1. In this part we prove the inclusion of (4.1). Take s ∈ iR. Note that in Proposition 3.2 it was shown for s ∈ iR that Y (s) = I − G(s)∗bf (s) +bρ(s), G)−1g and ρ is defined by (3.4); see (3.5). From (1.7) and (1.8) we where f = (TGT ∗ then see that Θ(s)∗Ξ(s) = E ∗Y (s)∗Y (s)D+ = E ∗Y (s)∗(cid:16)I − G(s)∗bf (s) +bρ(s)(cid:17)D+. Now use that G(s)Y (s)E = DE = 0, and hence E ∗Y (s)∗G(s)∗ = 0. The latter identity and the fact that E ∗D+ = 0 and Y = I −by imply that Θ(s)∗Ξ(s) = −E ∗by(s)∗D+ + E ∗bρ(s)D+ − E ∗by(s)∗bρ(s)D+ = −A(s) + B(s) − C(s). p×p(R+) ∩ L2 and by(∞) = 0, that is, by ∈ fW p×p (4.2) From item (b) in Proposition 4.1 we know that y ∈ L1 thus by ∈ fW p×p A(·) := E ∗by(·)∗D+ ∈ fW (p−m)×p 0,+ . It follows that (4.3) 0,− + . Recall that ρ is given by (3.4) with f = (TGT ∗ to L1 G)−1g. Since the function g belongs m×p(R+), item (b) in Proposition 2.1 tells us that the same holds m×p(R+)∩ L2 p×p(R+), and 10 G.J. GROENEWALD, S. TER HORST, AND M.A. KAASHOEK . 0,− p×p(R) ∩ L2 true for f . It follows that g∗ ⋆ f ∈ L1 ρ ∈ L1 (4.4) p×p(R−) ∩ L2 p×p(R−). We conclude that B(·) := E ∗bρ(·)D+ ∈ fW (p−m)×p Finally, note that y∗(t) = y(−t)∗ for t ∈ R and (cid:0)by(s)(cid:1)∗ by(s)∗bρ(s) =(cid:16)\y∗ ⋆ ρ(cid:17)(s), s ∈ iR, y∗(t − τ )ρ(τ ) dτ =Z 0 p×p(R−) ∩ L2 Since both y∗ and ρ belong to L1 Section 2 in [4]) that the same holds true for y∗ ⋆ ρ. But then (y∗ ⋆ ρ)(t) =Z ∞ and −∞ −∞ p×p(R). The latter implies that = by∗(s) for s ∈ iR. Thus y∗(t − τ )ρ(τ ) dτ. p×p(R−), it is well known (see, e.g., (4.5) C(·) := E ∗by(·)∗bρ(·)D+ ∈ fW (p−m)×p 0,− . Part 2. In this part we prove the identity of (4.1). Using (3.1) we see that From (4.3), (4.4), (4.5) and (4.2) it follows that Θ∗Ξ ∈ fW (p−m)×m s ∈ iR. G(s)Θ(s) = G(s)Y (s)E = DE = 0, 0,− . 0 This implies that TGTΘ = 0, and hence Im TΘ ⊂ Ker TG. To prove the reverse inclusion, take h ∈ Ker TG. Thus h ∈ L2 It follows that G(s)bh(s) = 0 for ℜs > 0. Put H(s) =bh(s). Then H(·) belongs to H 2 m(iR). Next we apply Theorem 1.2. Using the identities in (1.13) we see that p(R+) and TGh = 0. =(cid:2)Ξ(s) Θ(s)(cid:3)(cid:20) H(s) =(cid:2)Ξ(s) Θ(s)(cid:3)(cid:20) G(s) E ∗Y (s)−1(cid:21) H(s) E ∗Y (s)−1H(s)(cid:21) = Θ(s)E ∗Y (s)−1H(s). Hence bh(s) = Θ(s)Ψ(s), where Ψ(s) = E ∗Y (s)−1bh(s). Since h ∈ L2 Y (·)−1 is a matrix function with H ∞ entries, we conclude that Ψ ∈ H 2 hence Ψ = bu for some u ∈ L2 bh(s) = Θ(s)bu(s). This shows that h = TΘu, and hence Ker TG ⊂ Im TΘ. W (p−m)×m Part 1. In this part we show that the equation G(s)X(s) = Im, ℜs > 0, has a solution X ∈ W p×m if and only if TG is right invertible. Furthermore, we show that in that case the function Ξ defined by (1.7) is a particular solution. From the one but last sentence of the paragraph containing (1.4) we know that it suffices to prove the "if part" only. Therefore, in what follows we assume that TG is right invertible. Since Ξ(s) = Y (s)D+ and Y ∈ W p×p . Moreover, using the identity (4.1) we have p−m(R+). The identity bh(s) = Θ(s)Ψ(s) then yields Proof of Theorem 1.1. From Lemma 3.2 we know that Θ is an inner function in . The proof of the other statements is split into three parts. + , we have Ξ ∈ W p×m p(R+) and p−m, and (cid:3) + + + G(s)Ξ(s) = G(s)Y (s)D+ = DD+ = Im. Thus Ξ is a particular solution. THE B´EZOUT EQUATION ON THE RIGHT HALF PLANE IN A WIENER SPACE SETTING11 Part 2. This second part deals with the description of all in solutions in W p×m . Let Z be an arbitrary function in W p×m be defined by (1.9). Then , and let X ∈ W p×m + + + G(s)X(s) = G(s)Ξ(s) + G(s)Θ(s)Z(s) = Im + G(s)Θ(s)Z(s), ℜs ≥ 0. Recall that G(s)Θ(s) = G(s)Y (s)E = DE = 0. Thus G(s)X(s) = Im, ℜs ≥ 0, and thus X is a solution. be a solution of the equation and G(s)H(s) = 0. Using the To prove the converse implication, let X ∈ W p×m G(s)X(s) = Im. Put H = X − Ξ. Then H ∈ W p×m identities in (1.13), we obtain E ∗Y (s)−1(cid:21) H(s) E ∗Y (s)−1H(s)(cid:21) = Θ(s)E ∗Y (s)−1H(s). H(s) =(cid:2)Ξ(s) Θ(s)(cid:3)(cid:20) G(s) =(cid:2)Ξ(s) Θ(s)(cid:3)(cid:20) 0 + + Thus H(s) = Θ(s)Z(s), where Z(s) = E ∗Y (s)−1H(s). Since Y is invertible in W p×p . Together with the fact that H ∈ W p×m + , the function Y (·)−1 is in W p×p , this yields Z ∈ W (p−m)×m Part 3. In this part we prove the identity (1.10). Assume Z ∈ fW (p−m)×m, and let X be the function defined by (1.9). Fix u ∈ Cm. Then Z(·)u ∈ H 2 (4.6) . It follows X has the desired representation (1.9). m(iR), and + + + p−m(iR). Θ(·)Z(·)u = MΘZ(·)u ∈ H 2 Here MΘ is the operator of multiplication by Θ(·) mapping H 2 Furthermore, since MΘ is an isometry, we also see that m(iR) into H 2 p−m(iR). (4.7) kZ(·)uk = kΘ(·)Z(·)uk. p×p(R+) implies that Y ∈ fW p×p p×p(R+) ∩ L2 The fact that y ∈ L1 + . But then Ξ(s) = Y (s)D+ yields Ξ(·)u ∈ H 2 m(iR). Using the identity (1.9) we conclude that Ξ(·)u also belongs to H 2 m(iR). It follows that all norms in (1.10) are well defined, and in order to prove the identity (1.10) it suffices that to show that in H 2 m(iR) the function Θ(·)Z(·)u is orthogonal to the function Ξ(·)v for any v ∈ Cm. The latter fact follows from the inclusion in the second part of (4.1). Indeed, this inclusion tells us that M ∗ ΘΞ(·)v = 0, and hence hΞ(·)v, Θ(·)Z(·)uiH 2 m (iR) = hΞ(·)v, MΘZ(·)uiH 2 ΘΞ(·)v, Z(·)uiH 2 = hM ∗ m(iR) m (iR) = 0. This completes the proof. (cid:3) Appendix A. The Lebesgue space L1(R) ∩ L2(R) The material in this section is standard and is presented for the convenience of the reader. Throughout we deal with the Lebesgue spaces of complex-valued functions on the real line L1(R) and L2(R), their vector-valued counterparts L1 m(R) and L2 m(R). The m(R), and the intersection of the latter two spaces: L1 m(R) ∩ L2 12 G.J. GROENEWALD, S. TER HORST, AND M.A. KAASHOEK norms on these spaces are given by for f ∈ L2(R), for f ∈ L1(R), for f = (f1, . . . , fm)⊤ ∈ L1 −∞ f (t) dt −∞ f (t)2 dt(cid:17)1/2 kfik2 kfk1 =Z ∞ kfk2 =(cid:16)Z ∞ 1(cid:17)1/2 kfk1 =(cid:16) mXi=1 2(cid:17)1/2 kfk2 =(cid:16) mXi=1 kfk0 = max{kfk1,kfk2} m×p(R). Thus k is an m× p matrix function of which the (i, j)-th entry kij ∈ L1(R). For each ϕ ∈ L1 p(R) the convolution products k ⋆ ϕ and k ⋆ ψ, see (1.18), are well defined, k ⋆ ϕ belongs to L1 m(R) and k ⋆ ψ belongs to L2 m(R). In particular, if f ∈ L1 m(R). It follows that for a given k ∈ L1 m×p(R) the convolution product induces linear maps from the space L1 p(R) into L1 m(R), from the space L2 m(R), and from the space L1 m(R) ∩ L2 m(R). The resulting operators will be denoted by K1, K2 and K0, respectively. The proof of the following lemma is standard (see, e.g., page 216 in [6]) and therefore it is omitted. for f = (f1, . . . , fm)⊤ ∈ L2 for f ∈ L1 m(R). p(R) and ψ ∈ L2 p(R) ∩ L2 p(R), then k ⋆ f belongs to L1 Let k ∈ L1 p(R) ∩ L2 p(R) into L1 m(R), m(R), m(R) ∩ L2 m(R) ∩ L2 p(R) into L2 kfik2 Lemma A.1. The operators K1, K2 and K0 are bounded linear operators, and (A1) kKνk ≤ κ (ν = 1, 2, 0), where κ =(cid:16) mXi=1 pXj=1 kkijk2 1(cid:17)1/2 . With k ∈ L1 Hankel operator H defined by m×p(R+) we also associate the Wiener-Hopf operator W and the (W f )(t) =Z ∞ (Hf )(t) =Z ∞ 0 0 k(t − τ )f (τ ) dτ, 0 ≤ t < ∞, k(t + τ )f (τ ) dτ, 0 ≤ t < ∞. Using the classical relation between the convolution operator defined by k and the operators W and H (see, e.g., Section XII.2 in [7]) it is easy to see that W and H map the space L1 m(R+), and the space L1 p(R+) ∩ L2 m(R+). We denote the resulting operators by W1, W2, W0, and H1, H2, H0, respectively. Lemma A.1 shows that these operators are bounded and m(R+), the space L2 m(R+) ∩ L2 p(R+) into L1 p(R+) into L1 p(R+) into L2 (A2) kWνk ≤ κ and kHνk ≤ κ (ν = 1, 2, 0), where κ =(cid:16) mXi=1 pXj=1 kkijk2 1(cid:17)1/2 . Furthermore, using the line of reasoning in Lemma XX.2.4 in [7], we have the following corollary. Corollary A.2. The Hankel operators H1, H2, and H0 are the limit in operator norm of finite rank operators, and hence compact. THE B´EZOUT EQUATION ON THE RIGHT HALF PLANE IN A WIENER SPACE SETTING13 Next we present an auxiliary result that is used in the proof of Proposition 2.1. Put (A3) R(s) = DR +Z ∞ −∞ e−str(t) dt where r ∈ L1 m×m(R). By TR we denote the Wiener-Hopf operator on L2 m(R+) defined by R, that is, (A4) (TRf )(t) = DRf (t) +Z ∞ 0 r(t − τ )f (τ ) dτ, 0 ≤ t < ∞. As we know from the first paragraph of this section, the fact that r ∈ L1 implies that TR maps L1 m(R+) into itself. m×m(R) m(R+) ∩ L2 Lemma A.3. If TR is invertible as an operator on L2 space L1 m(R+) in a one-to-one way onto itself. m(R+) ∩ L2 m(R+), then T −1 R maps the Proof. Since TR is invertible, R admits a canonical factorization (see Section XXX.10 in [8]), and hence we can write T −1 R = LU , where L and U are Wiener- Hopf operators on L2 m(R+), (A5) (A6) (Lf )(t) = DLf (t) +Z t (U f )(t) = DU f (t) +Z ∞ 0 t ℓ(t − τ )f (t) dt, 0 ≤ t < ∞, u(t − τ )f (t) dt, 0 ≤ t < ∞. R maps L1 m(R+)∩ L2 m(R+) ∩ L2 m(R+) and f = T −1 m(R+) into L2 p(R+) and from L2 m(R+) into itself. Hence T −1 m(R+) ∩ L2 m(R+) ∩ L2 Here ℓ and u both belong to L1 in R−. The fact that both ℓ and u belong to L1 U map L1 T −1 R is one-to-one on L2 f ∈ L1 R g. This shows that T −1 T −1 m×m(R), with support of ℓ in R+ and support of u m×m(R) implies that both L and R has the same property. Since m(R+). For R TRf = m(R+). (cid:3) m(R+), it is also one-to-one on L1 m(R+)∩ L2 m(R+) onto L1 m(R+), respectively. Let Q be any operator on L2 m(R+), we have g = TRf ∈ L1 m(R+) ∩ L2 p×m(R+) and k ∈ L1 Lemma A.4. Let k ∈ L1 corresponding Hankel operators acting from L2 into L2 m(R+) mapping L1 m(R+) into itself, and assume that the restricted operator Q0 acting on L1 L2 m(R+) is bounded. If the operator I − HQH is invertible on L2 L2 HQH maps the space L1 m×p(R+), and let H and H be the p(R+) p(R+) ∩ m(R+)∩ m(R+), then I − m(R+) in a one-to-one way onto itself. p(R+) ∩ Lp m(R+) ∩ L2 Proof. We know that H maps L1 m(R+). Fur- thermore the same holds true for H with the role of p and m interchanged. Hence our hypothesis on Q implies that I − HQH maps the space L1 m(R+) into itself. Let M0 be the corresponding restricted operator. We have to prove that M0 is invertible. Note that Corollary A.2 implies that M0 is equal to the identity operator minus a compact operator, and hence M0 is a Fredholm operator of index zero. Therefore, in order to prove that M0 is invertible, it suffices to show that Ker M0 consists of the zero element only. Assume not. Then there exists a non-zero f in L1 m(R+) such that M0f = 0. The fact that f belongs to m(R+) shows that 0 = M0f = (I − HQH)f . But I − HQH m(R+)∩ L2 L1 is assumed to be invertible. Hence f must be zero. Thus M0 is invertible. m(R+) ⊂ L2 m(R+) ∩ L2 m(R+) ∩ L2 2(R+) into L1 m(R+) ∩ L2 (cid:3) 14 G.J. GROENEWALD, S. TER HORST, AND M.A. KAASHOEK Acknowledgement This work is based on the research supported in part by the National Research Foundation. Any opinion, finding and conclusion or recommendation expressed in this material is that of the authors and the NRF does not accept any liability in this regard. References [1] H. Bart, I. Gohberg, M.A. Kaashoek, and A.C.M. Ran, Factorization of matrix and operator functions: the state space method, Oper. Theory Adv. Appl. 178, Birkhauser Verlag, Basel, 2008. [2] A. Bottcher and B. Silbermann, Analysis of Toeplitz operators, 2nd Edition, Springer Mono- graphs in Mathematics, Springer Verlag, Berlin Heidelberg, 2006. [3] L. Carleson, Interpolation by bounded analytic functions and the corona problem, Ann. Math. 76 (1962), 547 -- 559. [4] H. Dym and I. Gohberg, On an extension problem, generalized Fourier analysis, and an entropy formula, Integr. Equ. Oper. Theory 3 (1980), 144 -- 215. [5] A.E. Frazho, M.A. Kaashoek, and A.C.M. Ran, Rational Matrix Solutions of a Bezout Type Equation on the Half-plane, in: Advances in Structured Operator Theory and Related Areas. The Leonid Lerer Anniversary Volume, Oper. Theory Adv. Appl. 237 (2012), 145 -- 160. [6] P. Fuhrmann, On the corona theorem and its applications to spectral problems in Hilbert space, Trans. Amer. Math. Soc. 132 (1968), 55 -- 66. [7] I. Gohberg, S. Goldberg, and M.A. Kaashoek, Classes of Linear Operators, Volume I, Oper. Theory Adv. Appl. 49, Birkhauser Verlag, Basel, 1990. [8] I. Gohberg, S. Goldberg, and M.A. Kaashoek, Classes of Linear Operators, Volume II, Oper. Theory Adv. Appl. 63, Birkhauser Verlag, Basel, 1993. [9] I. Gohberg, M.A. Kaashoek, and L. Lerer, The resultant for regular matrix polynomials and quasi commutativity, Indiana Univ. Math. J., 57 (2008), 2783 -- 2813. [10] M. Green and D.J.N. Limebeer, Linear Robust Control, Prentice Hall, Englewood Cliffs, NJ, 1995. [11] G.J. Groenewald, S. ter Horst and M.A. Kaashoek, The Bezout-Corona problem revisited: Wiener space setting, Complex Anal. Oper. Theory 10 (2016), 115 -- 139. [12] G. Gu and E.F. Badran, Optimal design for channel equalization via the filterbank approach, IEEE Trans. Signal Proc. 52 (2004), 536 -- 545. [13] G. Gu and L. Li, Worst-case design for optimal channel equalization in filterbank transceivers, IEEE Trans. Signal Proc. 51 (2003), 2424 -- 2435. [14] J.W. Helton, Operator Theory, analytic functions, matrices and electrical engineering, Re- gional Conference Series in Mathematics 68, Amer. Math. Soc., Providence, RI, 1987. [15] R. Mortini and A. Sasane, On the pre-B´ezout property of Wiener algebras on the disc and the half-plane, New Zealand J. Math. 38 (2008), 45 -- 55. [16] N.K. Nikol'skii, Treatise on the shift operator, Grundlehren 273, Springer Verlag, Berlin 1986. [17] N.K. Nikol'skii, Operators, Functions and Systems, Math. Surveys Monographs 92, Amer. Math. Soc., Providence, RI, 2002. [18] V.V. Peller, Hankel Operators and their Applications, Springer Monographs in Mathematics, Springer, 2003. [19] W. Rudin, Real and Complex Analysis, McGraw-Hill, New York, 1966. [20] V.A. Tolokonnikov, Estimates in Carleson's corona theorem. Ideals of the algebra H∞, the problem of Szekefalvi-Nagy, Zap. Naucn. Sem. Leningrad. Otdel. Mat. Inst. Steklov. (LOMI) 113 (1981), 178 -- 198 (Russian). [21] S. Treil, Lower bounds in the matrix corona theorem and the codimension one conjecture, GAFA 14 (2004), 1118 -- 1133. [22] S. Treil and B.D. Wick, The matrix-valued H p corona problem in the disk and polydisk, J. Funct. Anal. 226 (2005), 138 -- 172. [23] T.T. Trent and X. Zhang, A matricial corona theorem, Proc. Amer. Math. Soc. 134 (2006), 2549 -- 2558. THE B´EZOUT EQUATION ON THE RIGHT HALF PLANE IN A WIENER SPACE SETTING15 [24] M. Vidyasagar, Control system synthesis: a factorization approach, The MIT Press, Cam- bridge, MA, 1985. [25] K. Zhou with J.C. Doyle and K. Glover, Robust and optimal control, Prentice Hall, NJ, 1996. G.J. Groenewald, Department of Mathematics, Unit for BMI, North-West Univer- sity, Private Bag X6001-209, Potchefstroom 2520, South Africa E-mail address: [email protected] S. ter Horst, Department of Mathematics, Unit for BMI, North-West University, Private Bag X6001-209, Potchefstroom 2520, South Africa E-mail address: [email protected] M.A. Kaashoek, Department of Mathematics, VU University Amsterdam, De Boele- laan 1081a, 1081 HV Amsterdam, The Netherlands E-mail address: [email protected]
1810.07624	1	1810	2018-10-17T15:38:11	Best proximity point results for almost contraction and application to nonlinear differential equation	[ "math.FA" ]	Brinde [Approximating fixed points of weak contractions using the Picard itration, Nonlinear Anal. Forum 9 (2004), 43-53] introduced almost contraction mappings and proved Banach contraction principle for such mappings. The aim of this paper is to introduce the notion of multivalued almost $\Theta$- contraction mappings and present some best proximity point results for this new class of mappings. As applications, best proximity point and fixed point results for weak single valued $\Theta$-contraction mappings are obtained. An example is presented to support the results presented herein. An application to a nonlinear differential equation is also provided.	math.FA	math	Best proximity point results for almost contraction and application to nonlinear differential equation Azhar Hussain ∗Mujahid Abbas †, Muhammad Adeel‡Tanzeela Kanwal§. Abstract Brinde [Approximating fixed points of weak contractions using the Picard itration, Nonlinear Anal. Forum 9 (2004), 43-53] introduced almost contrac- tion mappings and proved Banach contraction principle for such mappings. The aim of this paper is to introduce the notion of multivalued almost Θ- contraction mappings and present some best proximity point results for this new class of mappings. As applications, best proximity point and fixed point results for weak single valued Θ-contraction mappings are obtained. An ex- ample is presented to support the results presented herein. An application to a nonlinear differential equation is also provided. Mathematics Subject Classification 2010: 55M20, 47H10 Keywords: Almost contraction, Θ-contraction, best proximity points. 1 Introduction and preliminaries The following concept was introduced by Berinde as 'weak contraction' in [9]. But in [10], Berinde renamed 'weak contraction' as 'almost contraction' which is appropriate. ∗Department of Mathematics, University of Sargodha, Sargodha-40100, Pakistan. Email: [email protected], †Department of Mathematics, Government College University, Lahore 54000, Pakistan. Email: [email protected] ‡Department of Mathematics, University of Sargodha, Sargodha-40100, Pakistan. Email: [email protected], §Department of Mathematics, University of Sargodha, Sargodha-40100, Pakistan. Email: [email protected] 1 Definition 1.1. Let (X, d) be a metric space. A mapping F : X → X is called almost contraction or (δ, L)-contraction if there exist a constant δ ∈ (0, 1) and some L ≥ 0 such that for any x, y ∈ X, we have d(F x, F y) ≤ δ.d(x, y) + Ld(y, F x), (1) Von Neumann [28] considered fixed points of multivalued mappings in the study of game theory. Indeed, the fixed point results for multivalued mappings play a significant role in study of control theory and in solving many problems of economics and game theory. Nadler [25] used the concept of the Hausdorff metric to obtain fixed points of multivalued contraction mappings and obtained the Banach con- traction principle as a special case. Here, we recall that a Hausdorff metric H induced by a metric d on a set X is given by H(A, B) = max{sup x∈A d(x, B), sup y∈B d(y, A)} (2) for every A, B ∈ CB(X), where CB(X) is the collection of the closed and bounded subsets of X. M. Berinde and V. Berinde [11] introduced the notion of multivalued almost contraction as follows: Let (X, d) be a metric space. A mapping F : X → CB(X) is called almost contraction if there exist two constants δ ∈ (0, 1) and L ≥ 0 such that for any x, y ∈ X, we have H(F x, F y) ≤ δd(x, y) + LD(y, F x). (3) Berinde [11] proved Nadler's fixed point theorem in ( [25]): Theorem 1.1. Let (X, d) be a complete metric space and F : X → CB(X) a almost contraction. Then F has a fixed point.. Jleli et al. [22] defined Θ-contraction mapping as follows: A mapping F : X → X is called Θ-contraction if for any x, y ∈ X Θ(d(F x, F y)) ≤ [Θ(d(x, y))]k (4) where, k ∈ (0, 1) and Θ : (0, ∞) → (1, ∞) is a mapping which satisfies the following conditions (Θ1) Θ is nondecreasing; 2 (Θ2) for each sequence {αn} ⊆ R+, 0; lim n→∞ Θ(αn) = 1 if and only if lim n→∞ (αn) = (Θ3) there exist 0 < k < 1 and l ∈ (0, ∞) such that lim α→0+ Θ(α)−1 αk = l; Denote Ω = {Θ : (0, ∞) → (1, ∞) : Θ satisf ies Θ1 − Θ3}. (5) Theorem 1.2. ( [22]) Let (X, d) be a complete metric space and F : X → X a Θ-contraction, then F has a unique fixed point. Hancer et al. [20] introduced the notion of multi-valued Θ-contraction mapping as follows: Let (X, d) be a complete metric space and T : X → CB(X) a multi- valued mapping. Suppose that there exists Θ ∈ Ω and 0 < k < 1 such that Θ(H(T x, T y)) ≤ [Θ(d(x, y))]k (6) for any x, y ∈ X provided that H(T x, T y) > 0, where CB(X) is a collection of all nonempty closed and bounded subsets of X. Theorem 1.3. Let (X, d) be a complete metric space and F : X → K(X) a multi-valued Θ-contraction, then F has a fixed point. Let A and B be two nonempty subsets of a metric space (X, d) and F : A → CB(B). A point x∗ ∈ A is called a best proximity point of F if D(x∗, F x∗) = inf{d(x∗, y) : y ∈ F x∗} = dist(A, B), where dist(A, B) = inf{d(a, b) : a ∈ A, b ∈ B}. If A ∩ B 6= φ, then x∗ is a fixed point of F. If A ∩ B = φ, then D(x, F x) > 0 for all x ∈ A and F has no fixed point. Consider the following optimization problem: min{D(x, F x) : x ∈ A}. (7) It is then important to study necessary conditions so that the above mini- mization problem has at least one solution. Since d(A, B) ≤ D(x, F x) (8) 3 for all x ∈ A. Hence the optimal solution to the problem min{D(x, F x) : x ∈ A} (9) for which the value d(A, B) is attained is indeed a best proximity point of multivalued mapping F. For more results in this direction, we refer to [1, 2, 4, 5, 7, 8, 13, 14, 17, 18, 21, 24, 29, 30] and references mentioned therein. Let A and B two nonempty subsets of X. Denote A0 = {a ∈ A : d(a, b) = d(A, B) for some b ∈ B} B0 = {b ∈ B : d(a, b) = d(A, B) for some a ∈ A}. Definition 1.2. the pair (A, B) has the P -property if [27] Let (X, d) be a metric space and A0 6= φ, we say that d(x1, y1) = d(A, B) d(x2, y2) = d(A, B) (cid:27) implies that d(x1, x2) = d(y1, y2), (10) where x1, x2 ∈ A and y1, y2 ∈ B. Definition 1.3. the pair (A, B) has the weak P -property if [31] Let (X, d) be a metric space and A0 6= φ, we say that d(x1, y1) = d(A, B) d(x2, y2) = d(A, B) (cid:27) implies that d(x1, x2) ≤ d(y1, y2), (11) where x1, x2 ∈ A and y1, y2 ∈ B. Definition 1.4. [12] Let (X, d) be a metric space, A, B two subsets of X and α : A × A → [0, ∞). A mapping F : A → 2B\{φ} is called α-proximal admissible if α(x1, x2) ≥ 1, d(u1, y1) = d(A, B), d(u2, y2) = d(A, B)   implies that α(u1, u2) ≥ 1 (12) for all x1, x2, u1, u2 ∈ A, y1 ∈ F x1 and y2 ∈ F x2. Definition 1.5. [12] Let F : X → CB(Y ) be a multi-valued mapping, where (X, d1), (Y, d2) are two metric spaces. A mapping F is said to be continuous at x ∈ X if H(F x, F xn) → 0 whenever d1(x, xn) → 0 as n → ∞. The aim of this paper is to obtain some best proximity point results for multivalued almost Θ-contraction mappings. We also present some best proximity point and fixed point results for single valued mappings. More- over, an example to prove the validity and application to nonlinear differen- tial equation for the usability of our results is presented. Our results extend, unify and generalize the comparable results in the literature. 4 2 Best proximity points of multivalued almost Θ- contraction We begin with the following definition: Definition 2.1. Let A, B be two nonempty subsets of a metric space (X, d) and α : A × A → [0, ∞). Let Θ ∈ Ω be a continuous function. A multivalued mapping F : A → 2B\{φ} is called almost Θ-contraction if for any x, y ∈ A, we have α(x, y)Θ[H(F x, F y)] ≤ [Θ(d(x, y) + λD(y, F x)))]k (13) where k ∈ (0, 1) and λ ≥ 0. Theorem 2.1. Let (X, d) be a complete metric space and A, B nonempty closed subsets of X such that A0 6= φ. Suppose that F : A → K(B) is a continuous mapping such that (i) F x ⊆ B0 for each x ∈ A0 and (A, B) satisfies the weak P -property; (ii) F is α-proximal admissible mapping; (iii) there exists x0, x1 ∈ A0 and y0 ∈ F x0 ⊆ B0 such that d(x1, y0) = d(A, B) and α(x0, x1) ≥ 1; (iv) F is multivalued almost Θ-contraction. Then F has a best proximity point in A. Proof. Let x0, x1 be two given points in A0 and y0 ∈ F x0 ⊆ B0 such that d(x1, y0) = d(A, B) and α(x0, x1) ≥ 1. If y0 ∈ F x1, then d(A, B) ≤ D(x1, F x1) ≤ d(x1, y0) = d(A, B) implies that D(x1, F x1) = d(A, B) and x1 is a best proximity point of F . If y0 /∈ F x1 then, 0 < D(y0, F x1) ≤ H(F x0, F x1). Since F (x1) ∈ K(B), we can choose y1 ∈ F x1 such that 1 < Θ[d(y0, y1)] ≤ Θ[H(F x0, F x1)]. As F is multivalued almost Θ-contraction mapping, we have 1 < Θ[d(y0, y1)] ≤ α(x0, x1)Θ[H(F x0, F x1)] ≤ [Θ(d(x0, x1) + λD(x1, F x0))]k = [Θ(d(x0, x1))]k. (14) 5 Since y1 ∈ F x1 ⊆ B0, there exists x2 ∈ A0 such that d(x2, y1) = d(A, B) and α(x1, x2) ≥ 1. By weak P -property of the pair (A, B) we obtain that d(x2, x1) ≤ d(y0, y1). If y1 ∈ F x2, then x2 is a best proximity point of F . If y1 /∈ F x2, then D(y1, F x2) ≤ H(F x1, F x2). We now choose y2 ∈ F x2 such that 1 < Θ[d(y1, y2)] ≤ Θ[H(F x1, F x2)] ≤ α(x1, x2)Θ[H(F x1, F x2)] ≤ [Θ(d(x1, x2) + λD(x2, F x1))]k = [Θ(d(x1, x2))]k. (15) Continuing this process, we can obtain two sequences {xn} and {yn} in A0 ⊆ A and B0 ⊆ B, respectively such that yn ∈ F xn and it satisfies d(xn+1, yn) = d(A, B) with α(xn, xn+1) ≥ 1 where n = 0, 1, 2, .... Also, 1 < Θ[d(yn, yn+1)] ≤ Θ[H(F xn, F xn+1)] ≤ α(xn, xn+1)Θ[H(F xn, F xn+1)] ≤ [Θ(d(xn, xn+1) + λD(xn, F xn+1))]k = [Θ(d(xn, xn+1))]k. implies that Since and 1 < Θ[d(yn, yn+1)] ≤ (Θ(d(xn, xn+1)))k. d(xn+1, yn) = d(A, B) d(xn, yn−1) = d(A, B) for all n ≥ 1, it follows by the weak P -property of the pair (A, B) that d(xn, xn+1) ≤ d(yn−1, yn) (16) (17) (18) (19) (20) for all n ∈ N. Now by repeated application of (17), (20) and the monotone 6 property of Θ, we have 1 < Θ[d(xn, xn+1)] ≤ Θ(d(yn−1, yn)) ≤ Θ(H(F xn−1, F xn)) ≤ α(xn−1, xn)Θ(H(F xn−1, F xn)) ≤ [Θ(d(xn−1, xn) + λD(xn, F xn−1))]k = (Θ(d(xn−1, xn)))k ≤ (Θ(d(yn−2, yn−1)))k ≤ (Θ(H(F xn−2, F xn−1)))k ≤ (α(xn−2, xn−1)Θ(H(F xn−2, F xn−1))k ≤ [Θ(d(xn−2, xn−1) + λD(xn−1, F xn−2))]k2 = (Θ(d(xn−2, xn−1)))k2 . . . ≤ (Θ(d(x0, x1)))kn . (21) Θ(d(xn, xn+1)) = 1 and hence for all n ∈ N ∪ {0}. This shows that lim n→∞ limn→∞ d(xn, xn+1) = 0. From (Θ3), there exist 0 < r < 1 and 0 < l ≤ ∞ such that lim n→∞ Θ(d(xn, xn+1)) − 1 [d(xn, xn+1)]r = l. (22) Assume that l < ∞ and β = l/2. From the definition of the limit there exists n0 ∈ N such that Θ(d(xn, xn+1)) − 1 [d(xn, xn+1)]r (cid:12)(cid:12)(cid:12)(cid:12) which implies that ≤ B, for all n ≥ n0 − l(cid:12)(cid:12)(cid:12)(cid:12) Θ(d(xn, xn+1) − 1 [d(xn, xn+1)]r ≥ l − β = β for all n ≥ n0. Hence n[d(xn, xn+1)]r ≤ nα[Θ(d(xn, xn+1) − 1] for all n ≥ n0, where α = 1/β. Assume that l = ∞. Let β > 0 be a given real number. From the definition of the limit there exists n0 ∈ N such that Θ(d(xn, xn+1) − 1 [d(xn, xn+1)]r ≥ β for all n ≥ n0 7 implies that n[d(xn, xn+1)]r ≤ nα[Θ(d(xn, xn+1) − 1] for all n ≥ n0, where α = 1/β. Hence, in all cases there exist α > 0 and n0 ∈ N such that n[d(xn, xn+1)]r ≤ nα[Θ(d(xn, xn+1) − 1] for all n ≥ n0. From (21), we have n[d(xn, xn+1)]r ≤ nα[Θ(d(xn, xn+1) − 1] for all n ≥ n0. On taking the limit as n → ∞ on both sides of the above inequality, we have lim n→∞ n[d(xn, xn+1)]r = 0. (23) It follows from (23) that there exists n1 ∈ N such that n[d(xn, xn+1)]r ≤ 1 for all n > n1. This implies that d(xn, xn+1) ≤ 1 n1/r for all n > n1. Now, for m > n > n1, we have d(xn, xm) ≤ m−1 Xi=n d(xi, xi+1) ≤ m−1 Xi=n 1 1 r i . Since 0 < r < 1,P∞ converges. Therefore d(xn, xm) → 0 as m, n → ∞. This shows that {xn} and {yn} are Cauchy sequences in A and B, respec- tively. Next, we assume that there exists elements u, v ∈ A such that i=n 1 1 i r xn → u and yn → v as n → ∞. Taking limit as n → ∞ in (18), we obtain that d(u, v) = d(A, B). (24) Now , we claim that v ∈ T u. Since yn ∈ F xn, we have D(yn, F u) ≤ H(F xn, F u). 8 Taking limit as n → ∞ on both sides above sides of above inequality, we have D(v, F u) = lim n→∞ D(yn, F u) ≤ lim n→∞ H(F xn, F u) = 0. As F u ∈ K(B), D(v, F u) = 0 implies that v ∈ F u. By (24), we have D(u, F u) ≤ d(u, v) = dist(A, B) ≤ D(u, F u), which implies that D(u, F u) = dist(A, B) and hence u is a best proximity point of F in A. Remark 2.1. In the next theorem, we replace the continuity assumption on F with the following condition: If {xn} is a sequence in A such that α(xn, xn+1) ≥ 1 for all n and xn → x ∈ A as n → ∞, then there exists a subsequence {xn(k)} of {xn} such that α(xn(k), x) ≥ 1 for all k. If the above condition is satisfied then we say that the set A satisfies α− subsequential property. Theorem 2.2. Let (X, d) be a complete metric space and (A, B) a pair of nonempty closed subsets of X such that A0 6= φ. Let F : A → K(B) be a multivalued mapping such that conditions (i)-(iv) of Theorem 2.1 are satisfied. Then F has a best proximity point in A provided that A satisfies α− subsequential property. Proof. Following Arguments similar to those in the proof of Theorem 2.1, we obtain two sequences {xn} and {yn} in A and B, respectively such that α(xn, xn+1) ≥ 1, xn → u ∈ A and yn → v ∈ B as n → ∞ and d(u, v) = dist(A, B) (25) (26) (27) By given assumption, there exists a subsequence {xn(k)} of {xn} such that α(xn(k), u) ≥ 1 for all k. Since yn(k) ∈ F xn(k) for all k ≥ 1, applying condition (iv) of Theorem 2.1, we obtain that 1 < Θ(D(yn(k), F u)) ≤ Θ(H(F xn(k), F u)) ≤ α(xn(k), u)Θ(H(F xn(k), F u)) = (Θ(d(xn(k), u)))k. (28) On taking limit as k → ∞ in (28) and using the continuity of Θ, we have Θ(D(v, F u) = 1. Therefore, by (Θ2) we obtain that D(v, F u) = 0. As shown in the proof of Theorem 2.1, we have D(v, F u) = dist(A, B) and hence u is a best proximity point of F in A. 9 Remark 2.2. To obtain the uniqueness of the best proximity point of mul- tivalued almost Θ-contraction mappings, we propose the following H condi- tion: H : for any best proximity points x1, x2 of mapping F, we have α(x1, x2) ≥ 1. Theorem 2.3. Let A and B be two nonempty closed subsets of a com- plete metric space (X, d) such that A0 6= φ and F : A → K(B) continuous multivalued mapping satisfying the conditions of Theorem 2.1 (respectively in Theorem 2.2). Then the mapping F has a unique best proximity point provided that it satisfies the condition H. Proof. Let x1, x2 be two best proximity points of F such that x1 6= x2, then by the given hypothesis H we have α(x1, x2) ≥ 1 and D(x1, F x1) = D(A, B) = D(x2, F x2). Since F x1 and F x2 are compact sets, there exists elements y0 ∈ F x1 and y1 ∈ F x2 such that d(x1, y0) = dist(A, B) and d(x2, y1) = dist(A, B). As F satisfies the weak P -property, we have d(x1, x2) ≤ d(y0, y1), Since F is multivalued almost Θ-contraction mapping, we obtain that Θ(d(x1, x2)) ≤ Θ(d(y0, y1)) ≤ Θ(H(F x1, F x2)) ≤ α(x1, x2)Θ(H(F x1, F x2)) ≤ [Θ(d(x1, x2) + λD(x2, F x1))]k = (Θ(d(x1, x2)))k < Θ(d(x1, x2)), a contradiction. Hence d(x1, x2) = 0, and x1 = x2. If the pair (A, B) satisfies the weak P -property, then it satisfies the P - property, we have the following corollaries: Corollary 2.1. Let (X, d) be a complete metric space and (A, B) a pair of nonempty closed subsets of X such that A0 6= φ. Suppose that a continuous mapping F : A → K(B) satisfies the following properties: (i) F x ⊆ B0 for each x ∈ A0 and (A, B) satisfies the P -property; 10 (ii) F is multivalued α-proximal admissible mapping; (iii) there exists x0, x1 ∈ A0 and y0 ∈ F x0 ⊆ B0 such that d(x1, y0) = d(A, B) and α(x0, x1) ≥ 1; (iv) F is multivalued almost Θ-contraction. Then F has a best proximity point in A. Corollary 2.2. Let (X, d) be a complete metric space and (A, B) a pair of nonempty closed subsets of X such that A0 6= φ. let F : A → K(B) be a multi-valued mapping such that conditions (i)-(iv) of Corollary 2.1 are satisfied. Then F has a best proximity point in A provided that A has α− subsequential property. Now we give an example to support Theorem 2.1. Example 2.1. Let X = R2 be a usual metric space. Let A = {(−2, 2), (2, 2), (0, 4)} (29) and B = {(−8, γ) : γ ∈ [−8, 0]} ∪ {(8, γ) : γ ∈ [−8, 0]} ∪ {(β, −8) : β ∈ (−8, 8)}. (30) Then d(A, B) = 8, A0 = {(−2, 2), (2, 2)} and B0 = {(−8, 0), (8, 0)}. Define the mapping F : A → K(B) by {(−8, 0)} {(8, 0)} {(β, −8) : β ∈ (−8, 8)} if x = (−2, 2) if x = (2, 2) if x = (0, 4). F x =  and α : A × A → [0, ∞) by α((x, y), (u, v)) = 11 10 . (31) Clearly, F (A0) ⊆ B0. For (−2, 2), (2, 2) ∈ A and (−8, 0), (8, 0) ∈ B, we have (cid:26) d((−2, 2), (−8, 0)) = d(A, B) = 8, d((2, 2), (8, 0)) = d(A, B) = 8. Note that d((−2, 2), (2, 2)) < d((−8, 0), (8, 0)). (32) 11 that is, the pair (A, B) has weak P -property. Also, F is α-proximal admis- sible mapping. Now we show that F is multivalued almost Θ-contraction where Θ : (0, ∞) → (1, ∞) is given by Θ(t) = 5t. Note that α((−2, 2), (2, 2))Θ[H(F (−2, 2), F (2, 2))] = 11 10 (516) (33) and If we take k ∈ ( 717 [Θ(d((−2, 2), (2, 2)) + λD((2, 2), (−8, 0)))]k =(cid:0) 528(cid:1)k 1250 , 1) and λ = 2 in (34), we have Similarly, 11 10 (516) <(cid:0) 528(cid:1)k . . (34) (35) α(x, y)Θ[H(F x, F y)] ≤ [Θ(d(x, y) + λD(y, F x)))]k (36) holds for the remaining pairs. Hence all the conditions of Theorem 2.1 are satisfied. Moreover, (−2, 2), (2, 2) are best proximity points of F in A. Remark 2.3. Note that mapping F in the above example does not hold for the case of Nadler [25] as well as for Hancer et al. [20]. For if, take x = (−2, 2), y = (2, 2) ∈ A, we have Θ(H(T x, T y)) = 516 > 54 > (54)k = [Θ(d(x, y))]k for k ∈ (0, 1). Also H(T x, T y) = 16 > 4 = d(x, y) > αd(x, y) for α ∈ (0, 1). Remark 2.4. In above example 2.1, the pair (A, B) does not satisfy the P -property and hence the Corollary 2.1 is not applicable in this case. 3 Application to single valued mappings In this section, we obtain some best proximity point results for single- valued mappings as applications of our results obtain in section 2. 12 Definition 3.1. X, a nonself mapping F : A → B is called α-proximal admissible if [22] Let (X, d) be a metric space and A, B two subsets of α(x1, x2) ≥ 1, d(u1, F x1) = d(A, B), d(u2, F x2) = d(A, B). implies α(u1, u2) ≥ 1 (37)   for all x1, x2, u1, u2 ∈ A where α : A × A → [0, ∞). Definition 3.2. Let α : A × A → [0, ∞) and Θ : (0, ∞) → (1, ∞) a nonde- creasing and continuous function. A mapping F : A → B is called almost Θ-contraction if for any x, y ∈ A, we have α(x, y)Θ(d(F x, F y)) ≤ [Θ(d(x, y) + λd(y, F x)))]k (38) where, k ∈ (0, 1) and λ ≥ 0. Theorem 3.1. Let (X, d) be a complete metric space and (A, B) a pair of nonempty closed subsets of X such that A0 is nonempty. If F : A → B is a continuous mapping such that (i) F (A0) ⊆ B0 and (A, B) satisfies the weak P -property; (ii) F is α-proximal admissible mapping; (iii) there exists x0, x1 ∈ A0 such that d(x1, F x0) = d(A, B) and α(x0, x1) ≥ 1; (iv) F is almost Θ-contraction. Then F has a best proximity point in A. Proof. As for every x ∈ X, {x} is compact in X. Define a multivalued map- ping T : A → K(B) by T x = {F x} for x ∈ A. The continuity of F implies that T is continuous. Now F (A0) ⊆ B0 implies that T x = {F x} ⊆ B0 for each x ∈ A0. If x1, x2, v1, v2 ∈ A, y1 ∈ T x1 = {F x1} and y2 ∈ T x2 = {F x2} are such that α(x1, x2) ≥ 1, d(v1, y1) = dist(A, B) and d(v2, y2) = dist(A, B). (39) That is, α(x1, x2) ≥ 1, d(v1, F x1) = dist(A, B) and d(v2, F x2) = dist(A, B). (40) 13 Then we have α(v1, v2) ≥ 1 as F is α-proximal admissible mapping. Hence T is α-proximal admissible mapping. Suppose there exist x0, x1 ∈ A0 such that d(x1, F x0) = dist(A, B) and α(x0, x1) ≥ 1. Let y0 ∈ T x0 = {F x0} ⊆ B0. Then d(x1, F x0) = dist(A, B) gives that d(x1, y0) = dist(A, B). By condition (iii), there exist x0, x1 ∈ A0 and y0 ∈ T x0 ⊆ B0 such that d(x1, y0) = dist(A, B) and α(x0, x1) ≥ 1. Since F is almost Θ-contraction, we have α(x, y)Θ(H(T x, T y)) = α(x, y)Θ[d(F x, F y)] ≤ [Θ(d(x, y) + λD(y, F x)))]k, (41) for any x, y ∈ A which implies that T is multivalued almost Θ-contraction. Thus, all the conditions of Theorem 2.1 are satisfied and hence T has a best proximity point x∗ in A. Thus we have D(x∗, T x∗) = dist(A, B) and hence d(x∗, F x∗) = dist(A, B), that is x∗ is a best proximity point of F in A. Theorem 3.2. Let (X, d) be a complete metric space and (A, B) a pair of nonempty closed subsets of X such that A0 is nonempty. If F : A → B is a mapping such that conditions (i)-(iv) of Theorem 3.1 are satisfied. Then F has a best proximity point in A provided that A satisfies α-subsequential property. Proof. Let T : A → K(B) be as given in proof of Theorem 3.1. Following arguments similar to those in the proof of Theorem 3.1, we obtain that (i) T x ⊆ B0 for each x0 ∈ A0; (ii) T is multi-valued α-proximal admissible mapping; (iii) there exists x0, x1 ∈ A0 and y0 ∈ T x0 ⊆ B0 such that d(x1, y0) = d(A, B) and α(x0, x1) ≥ 1; (iv) T is multivalued almost Θ-contraction. Thus, all the conditions of Theorem 2.2 are satisfied and hence T has a best proximity point x∗ in A, that is, D(x∗, T x∗) = dist(A, B) Consequently, d(x∗, F x∗) = dist(A, B), and x∗ is a best proximity point of F in A. Corollary 3.1. Let (X, d) be a complete metric space and (A, B) a pair of nonempty closed subsets of X such that A0 is nonempty. If F : A → B is a continuous mapping such that 14 (i) F (A0) ⊆ B0 and (A, B) satisfies the P -property; (ii) F is α-proximal admissible mapping; (iii) there exists x0, x1 ∈ A0 such that d(x1, T x0) = d(A, B) and α(x0, x1) ≥ 1; (iv) F is almost Θ-contraction. Then F has a best proximity point in A. Proof. Replace the condition of weak P-property with P-property in Theo- rem 3.1. Corollary 3.2. Let (X, d) be a complete metric space and (A, B) be a pair of nonempty closed subsets of X such that A0 is nonempty. If F : A → B is a mapping such that conditions (i)-(iv) of Corollary 3.1 are satisfied. Then F has a best proximity point in A provided that A satisfies α− subsequential property. Proof. Replace the condition of weak P-property with P-property in Theo- rem 3.2. 4 Fixed point results for single and multi-valued mappings In this section, fixed points of singlevalued and multivalued almost Θ- contraction mappings are obtained. Taking A = B = X in Theorem 2.1 (Theorem 2.2), we obtain corre- sponding fixed point results for almost Θ-contraction mappings. Theorem 4.1. Let (X, d) be a complete metric space. If F : X → K(X) is a continuous mapping satisfying (i) F is α admissible mapping; (ii) there exists x0 ∈ X such that α(x0, F x0) ≥ 1; (iii) F is multivalued almost Θ-contraction. Then F has a fixed point in X. 15 Theorem 4.2. Let (X, d) be a complete metric space. Let F : X → K(X) be a multi-valued mapping such that conditions (i)-(iii) of Theorem 4.1 are satisfied. Then F has a fixed point in X provided that X satisfies α− sub- sequential property. Taking A = B = X in Theorem 3.1 (in Theorem 3.2), we obtain the corresponding fixed point results of almost Θ-contraction mappings. Theorem 4.3. Let (X, d) be a complete metric space. Let F : X → X be a continuous mapping satisfying (i) F is α admissible mapping; (ii) there exists x0 ∈ X such that α(x0, F x0) ≥ 1; (iii) F is almost Θ-contraction. Then F has a fixed point in X. Theorem 4.4. Let (X, d) be a complete metric space. Let F : X → X be a multi-valued mapping such that conditions (i)-(iii) of Theorem 4.1 are satis- fied. Then F has a fixed point in X provided that X has a α−subsequential property. Remark 4.1. In Theorem 4.1 (respectively in 4.3) (i). If we take α(x, y) = 1, we obtain the main results of Durmaz [16] and Altun [3]. (ii). Taking λ = 0 and α(x, y) = 1, we obtain the main result of Hancer et al. [20] and Jelli [22] respectively. (iii). Taking α(x, y) = 1 and Θ(t) = et, we obtain the main result of Berinde [11] and [9]. (iv). Taking α(x, y) = 1, λ = 0 and Θ(t) = et, we obtain the main result of Nadler [25] and Banach [6]. 5 Application to Nonlinear Differential Equations Let C([0, 1]) be the set of all continuous functions defined on [0, 1] and d : C([0, 1]) × C([0, 1]) → R be the metric defined by d(x, y) = x − y∞ = max t∈[0,1] x(t) − y(t). (42) It is known that (C([0, 1]), d) is a complete metric space. 16 Let us consider the two-point boundary value problem of the second- order differential equation: − d2x dt2 = f (t, x(t)) x(0) = x(1) = 0 t ∈ [0, 1]; (cid:27) where f : [0, 1] × R → R is a continuous mapping. The Green function associated with (43) is defined by G(t, s) =(cid:26) t(1 − s) s(1 − t) if 0 ≤ t ≤ s ≤ 1, if 0 ≤ s ≤ t ≤ 1. Let φ : R × R → R be a given function. Assume that the following conditions hold: (43) (44) (i) f (t, a) − f (t, b) ≤ max a,b∈R φ(a, b) ≥ 0; a − b for all t ∈ [0, 1] and a, b ∈ R with (ii) there exists x0 ∈ C[0, 1] such that φ(x0(t), F x0(t)) ≥ 0 for all t ∈ [0, 1] where F : C[0, 1] → C[0, 1]; (iii) for each t ∈ [0, 1] and x, y ∈ C[0, 1], φ(x(t), y(t)) ≥ 0 implies φ(F x(t), F y(t)) ≥ 0; (iv) for each t ∈ [0, 1], if {xn} is a sequence in C[0, 1] such that xn → x in C[0, 1] and φ(xn(t), xn+1(t)) ≥ 0 for all n ∈ N , then φ(xn(t), x(t)) ≥ 0 for all n ∈ N . We now prove the existence of a solution of the second order differential equation (43). Theorem 5.1. Under the assumptions (i)-(iv), (43) has a solution in C 2([0, 1]). Proof. It is well known that x ∈ C 2([0, 1]) is a solution of (43) is equivalent to x ∈ C([0, 1]) is a solution of the integral equation x(t) =Z 1 0 G(t, s)f (s, x(s))ds, t ∈ [0, 1]. (45) Let F : C[0, 1] → C[0, 1] be a mapping defined by F x(t) =Z 1 0 G(t, s)f (s, x(s))ds. (46) 17 Suppose that x, y ∈ C([0, 1]) such that φ(x(t), y(t)) ≥ 0 for all t ∈ [0, 1]. By applying (i), we obtain that 0 F u(x) − F v(x) = Z 1 ≤ Z 1 ≤ Z 1 ≤ Z 1 0 0 0 G(t, s)f (s, x(s))ds −Z 1 0 G(t, s)f (s, y(s))ds G(t, s)[f (s, x(s)) − f (s, y(s))]ds G(t, s)f (s, x(s)) − f (s, y(s))ds G(t, s) · (max x(s) − y(s))ds ≤ x − y∞ · sup t∈[0,1] Z 1 0 G(t, s)ds!. Since 1 R0 G(t, s)ds = −(t2/2) + (t/2), for all t ∈ [0, 1], we have sup t∈[0,1] Z 1 0 G(t, s)ds! = 1 8 . It follows that F x − F y∞ ≤ 1 8 (x − y∞ Taking exponential on the both sides, we have eF x−F y∞ ≤ e 1 8 (x−y∞) = [e(x−y∞)] 1 8 , (47) (48) for all x, y ∈ C[0, 1]. Now consider a function Θ : (0, ∞) → (1, ∞) by Θ(t) = et. Define α(x, y) =(cid:26) 1 if φ(x(t), y(t)) ≥ 0, t ∈ [0, 1], otherwise. 0 Then from (48) with k = 1 8 , we obtain that α(x, y)Θ(F x − F y∞) ≤ [Θ(d(x, y))]k ≤ [Θ(d(x, y) + λd(y, F x))]k. Therefore the mapping F is multivalued almost Θ-contraction. 18 From (ii) there exists x0 ∈ C[0, 1] such that α(x0, F x0) ≥ 1. Next, for any x, y ∈ C[0, 1] with α(x, y) ≥ 1, we have φ(x(t), y(t)) ≥ 0 for all t ∈ [0, 1] ⇒ φ(F x(t), F y(t)) ≥ 0 for all t ∈ [0, 1] ⇒ α(F x, F y) ≥ 1, and hence F is α-admissible. It follows from Theorem 4.3 that F has a fixed point x in C([0, 1]) which in turns is the solution of (43). 6 Conclusion This paper is concerned with the existence and uniqueness of the best proximity point results for Brinde type contractive conditions via auxiliary function Θ ∈ Ω in the framework of complete metric spaces. Also, some fixed point results as a special cases of our best proximity point results in the relevant contractive conditions are studied. Moreover, the correspond- ing fixed point results are obtained. An example is discussed to show the significance of the investigation of this paper. An application to a nonlinear differential equation is presented to illustrate the usability of the new theory. Acknowledgments This paper was funded by the University of Sargodha, Sargodha funded research project No. UOS/ORIC/2016/54. The first and third author, therefore, acknowledges with thanks UOS for financial support. References [1] A. Abkar, M. Gabeleh, The existence of best proximity points for multi- valued non-selfmappings, Rev. R. Acad. Cienc. Exactas F´ıs. Nat., Ser. A Mat., RACSAM, 107 (2013), 319 -- 325. [2] A. Abkar, M. Gabeleh, Global optimal solutions of noncyclic mappings in metric spaces, J. Optim. Theory Appl., 153 (2012), 298 -- 305. [3] I. Altun, H.A. Hnacer, G. Minak, On a general class of weakly picard operators, Miskolc Mathematical Notes, Vol. 16 (2015), No. 1, pp. 25 -- 32. 19 [4] M.A. Al-Thagafi, N. Shahzad, Convergence and existence results for best proximity points, Nonlinear Anal., Theory Methods Appl.,70 (2009), 3665 -- 3671. [5] A. Amini-Harandi, Best proximity points for proximal generalized con- tractions in metric spaces, Optim. Lett., 7 (2013), 913 -- 921. [6] S. Banach, Sur les op´erations dans les ensembles abstraits et leur appli- cation aux equations int´egrales, Fundamenta Mathematicae, 3, 133 -- 181, (1922). [7] S.S. Basha, Best proximity points: Global optimal approximate solu- tions, J. Glob. Optim., 49 (2011), 15 -- 21. [8] S.S. Basha, P. Veeramani, Best proximity pair theorems for multifunc- tions with open fibres, J. Approx. Theory, 103 (2000), 119 -- 129. [9] V. Berinde, Approximating fixed points of weak contractions using the Picard itration, Nonlinear Anal. Forum 9 (2004), 43 -- 53. [10] V. Berinde, General constructive fixed point theorems for ´Ciri´c-type almost contractions in metric spaces, Carpathian J. Math. 24 (2008), 10 -- 19. [11] M. Berinde, V. Berinde, On a general class of multivalued weakly picard mappings, J. Math. Anal. 326 (2007), 772 -- 782. [12] BS. Choudhurya, P. Maitya, N. Metiya, Best proximity point results in set-valued analysis, Nonlinear Analysis: Modelling and Control, 21(3) 2016, 293 -- 305. [13] B.S. Choudhury, P. Maity, P. Konar, A global optimality result using nonself mappings, Opsearch, 51 (2014), 312 -- 320. [14] R.C. Dimri, P. Semwal, Best proximity results for multivalued map- pings, Int. J. Math. Anal., 7 (2013), 1355 -- 1362. [15] G. Durmaz, Some theorems for a new type of multivalued contractive maps on metric space, Turkish Journal of Mathematics, 41 (2017), 1092 -- 1100. [16] G. Durmaz, I. Altun, A new perspective for multivalued weakly picard operators, Publications De L'institut Math´ematique, 101(115) (2017), 197 -- 204. 20 [17] A.A. Eldred, J. Anuradha, P. Veeramani, On equivalence of general- ized multivalued contractions and Nadler's fixed point theorem, J. Math. Anal. Appl., 336 (2007), 751 -- 757. [18] A.A. Eldred, P. Veeramani, Existence and convergence of best proximity points, J. Math. Anal. Appl., 323 (2006), 1001 -- 1006. [19] M. Gabeleh, Best proximity points: Global minimization of multivalued non-self mappings, Optim Lett., 8 (2014) 1101 -- 1112. [20] H.A. Hancer, G. Mmak, I. Altun, On a broad category of multivalued weakly Picard operators, Fixed Point Theory 18 (2017), 229 -- 236. [21] A. Hussain, M. Adeel, T. Kanwal, N. Sultana, Set valued contraction of Suzuki-Edelstein-Wardowski type and best proximity point results, Bulletin of Mathematical Analysis and Applications, 10(2) (2018), 53- 67. [22] M. Jleli, B. Samet, A new generalization of the Banach contraction principle, J. Inequal. Appl., 2014(2014), 8 pages. [23] Z. Kadelburg, S. Radenovi, A note on some recent best proximity point results for non-self mappings, Gulf J. Math., 1 (2013), 36 -- 41. [24] A. Latif, M. Abbas, A. Husain, Coincidence best proximity point of Fg- weak contractive mappings in partially ordered metric spaces, Journal of Nonlinear Science and Application, 9 (2016), 2448 -- 2457. [25] S. B. Nadler Jr, Multivalued contraction mappings, Pacific J. Math. 30 (1969), 475 -- 488. [26] E. Nazari, Best proximity points for generalized muktivalued contrac- tions in metric spaces, Miskolc Mathematical Notes, 16 (2015), 1055 -- 1062. [27] V. Sankar Raj, A best proximity point theorem for weakly contractive non-self-mappings, Nonlinear Anal., Theory Methods Appl., 74 (2011), 4804 -- 4808. [28] J. Von Neuman, Uber ein okonomisches Gleichungssystem und eine Verallgemeinerung des Brouwerschen Fixpunktsatzes, Ergebn. Math. Kolloq. 8 (1937), 73-83. 21 [29] K. Wlodarczyk, R. Plebaniak, A. Banach, Best proximity points for cyclic and noncyclic setvalued relatively quasi-asymptotic contractions in uniform spaces, Nonlinear Anal., Theory Methods Appl., 70 (2009), 3332 -- 3341. [30] K. Wlodarczyk, R. Plebaniak, C. Obczynski, Convergence theorems, best approximation and best proximity for set-valued dynamic systems of rel- atively quasi-asymptotic contractions in cone uniform spaces, Nonlinear Anal., Theory Methods Appl., 72 (2010), 794 -- 805. [31] J. Zhang, Y. Su, Q. Cheng, A note on 'A best proximity point theorem for Geraghty-contractions', Fixed Point Theory Appl., 2013:99 (2013). 22
1302.0517	1	1302	2013-02-03T18:01:35	On the optimal multilinear Bohnenblust--Hille constants	[ "math.FA" ]	The upper estimates for the optimal constants of the multilinear Bohnenblust--Hille inequality obtained in [J. Funct. Anal. 264 (2013), 429--463] are here improved to: {0.1cm} {enumerate} For real scalars: $K_{n}\leq\sqrt{2}(n-1)^{0.526322}$. For complex scalars: $K_{n}\leq\frac{2}{\sqrt{\pi}}(n-1)^{0.304975}$.{enumerate} {0.1cm} \noindent We also obtain sharper estimates for higher values of $n$. For instance, \[ K_{n}<1.30379(n-1) ^{0.526322}\] for real scalars and $n>2^{8}$ and \[ K_{n}<0.99137(n-1) ^{0.304975}\] for complex scalars and $n > 2^{15}.$	math.FA	math	ON THE OPTIMAL MULTILINEAR BOHNENBLUST -- HILLE CONSTANTS D. NU NEZ-ALARC ´ON, D. PELLEGRINO, J.B. SEOANE-SEP ´ULVEDA*, AND D. M. SERRANO-RODR´IGUEZ Abstract. The upper estimates for the optimal constants of the multilinear Bohnenblust -- Hille inequality obtained in [2] are here improved to: (1) For real scalars: Kn ≤ (2) For complex scalars: Kn ≤ 2 We also obtain sharper estimates for higher values of n. For instance, √2 (n − 1)0.526322 . √π (n − 1)0.304975. for real scalars and n > 28 and for complex scalars and n > 215. Kn < 1.30379 (n − 1)0.526322 Kn < 0.99137 (n − 1)0.304975 1. Preliminaries. First estimates Let K be the real or complex scalar field. The multilinear Bohnenblust -- Hille inequality asserts that for every positive integer n ≥ 1 there exists a sequence of positive scalars (Cn)∞n=1 in [1,∞) such that   N Xi1,...,in=1(cid:12)(cid:12)U (ei1 , . . . , ein)(cid:12)(cid:12) n+1 2n 2n n+1  ≤ Cn sup z1,...,zn∈DN U (z1, . . . , zn) for all n-linear forms U : KN × ··· × KN → K and every positive integer N , where (ei)N i=1 denotes the canonical basis of KN and DN represents the open unit polydisc in KN . In the last years a considerable effort was spent in the searching of optimal values for the constants Cn; for details and the state-of-the-art we refer to [2]. The notation and terminology used in this note are the same as those from [2], where it is proved that the optimal multilinear Bohnenblust -- Hille constants (Kn)∞n=2 satisfy (1.1) and Kn < 1.65 (n − 1)0.526322 + 0.13 (real scalars) Kn < 1.41 (n − 1)0.304975 − 0.04 (complex scalars). The proof of the above estimates is achieved by following a series of technical steps. In the case of real scalars, using some previous lemmata, it is observed that the sequence (cid:0)√2(cid:1)n−1 DM n 2 DM n+1 2 if n = 1, 2 if n is even, and if n is odd Mn =   Supported by Capes. Supported by CNPq Grant 301237/2009-3. **Supported by grant MTM2012-34341. 1 2 D. NU NEZ ET AL. satisfies the multilinear Bohnenblust -- Hille inequality, where D = e . Then, using a "uniform approximation" argument, the estimate (1.1) is achieved. In this section we remark that this final step of the proof, i.e., the uniform approximation argument, can be dropped and a quite simple argument provides even better constants. In fact, from [2] we know that, for all k ≥ 1 and n ≥ 2, we have 1 1− 2 γ √2 Mn = √2Dk−1 whenever n ∈ Bk = {2k−1 + 1, . . . , 2k}. Thus, k − 1 ≤ log2 (n − 1) and, hence, ≤ Using a similar argument (for complex scalars) it follows that √2Dlog2(n−1) = √2 (n − 1) Mn ≤ √2 (n − 1)0.526322 . log2 e 1− 1 2 γ √2 ! Mn ≤ 2 √π (n − 1) log2(cid:16)e 1 2 − 1 2 γ(cid:17) ≤ 2 √π (n − 1)0.304975 for the complex scalar field. Of course, the other estimates of [2] related to the above results can be straightforwardly improved by using these new estimates. 2. Sharper estimates for big values of n In this section we improve our previous estimates for large values of n. 2.1. Real case. If (Cn)∞n=1 denotes the sequence in [2, (4.3)], if we fix any k0, it is obvious that 1 with D = e1− 2 γ √2 such that Then and Jn =   Cn DJ n 2 2 (cid:17) D(cid:16)J n−1 n+1 2n n−1 2n (cid:16)J n+1 2 (cid:17) if n ≤ 2k0 , if n > 2k0 is even, and if n > 2k0 is odd , satisfies the multilinear Bohnenblust -- Hille inequality. For n > 2k0, let k1 > k0 be 2k1−1 + 1 ≤ n ≤ 2k1 . k1 − k0 ≤ log2(cid:18) n − 1 2k0−1(cid:19) Kn ≤ J2k1 = Dk1−k0 C2k0 log2(cid:16) n−1 2k0−1(cid:17) ≤ C2k0 D C2k0 Dk0−1 (n − 1)log2 D . = We thus have Kn ≤ From [3, Theorem 3.1] we know that C2k0 Dk0−1 (n − 1) log2 e 1− 1 2 γ √2 ! . (2.1) C2k0 ≤ 4Dk0−4 ON THE OPTIMAL MULTILINEAR BOHNENBLUST -- HILLE CONSTANTS 3 whenever k0 ≥ 4. Thus, Kn ≤ 4 √2 (cid:19)3 (n − 1) (cid:18) e1− 1 2 γ log2 e 1− 1 2 γ √2 ! < 1.338887 (n − 1)0.526322 . Summarizing, we have: Theorem 2.1. If n > 16, then Kn ≤ (cid:18) e log2 e 1− 1 2 γ √2 ! . 4 1− 1 2 γ √2 (cid:19)3 (n − 1) Numerically, (2.2) Kn < 1.338887 (n − 1)0.526322 . If we use the exact value of C2k0 instead of estimate (2.1) we can improve (2.2) as n grows. For example, n > 26 ⇒ Kn < 1.310883 (n − 1)0.526322 n > 27 ⇒ Kn < 1.306156 (n − 1)0.526322 n > 28 ⇒ Kn < 1.303787 (n − 1)0.526322 . 2.2. Complex case. Let (Cn)∞n=1 denote the sequence in [1, Theorem 2.3]. If we fix any k0, and as the authors did in [2], we can show that Jn =   Cn DJ n 2 2 (cid:17) D(cid:16)J n−1 n+1 2n n−1 2 (cid:17) 2n (cid:16)J n+1 if n ≤ 2k0, if n > 2k0 is even, and if n > 2k0 is odd, 1 1 with D = e the real case we obtain 2− 2 γ, satisfies the multilinear Bohnenblust -- Hille inequality. For n > 2k0, by mimicking Thus, using the values of C2k from [1] we have Kn ≤ C2k0 Dk0−1 (n − 1) log2(cid:16)e 1 2 − 1 2 γ(cid:17) . n > 23 ⇒ Kn < 1.02960973695 (n − 1)0.304975 , n > 24 ⇒ Kn < 1.01089344604 (n − 1)0.304975 , n > 25 ⇒ Kn < 1.00123230777 (n − 1)0.304975 , n > 26 ⇒ Kn < 0.99632125476 (n − 1)0.304975 , . . . n > 214 ⇒ Kn < 0.99137409768 (n − 1)0.304975 , n > 215 ⇒ Kn < 0.99136434217 (n − 1)0.304975 , . . . n > 225 ⇒ Kn < 0.99135459597 (n − 1)0.304975 , . . . n > 250 ⇒ Kn < 0.99135458644 (n − 1)0.304975 . 4 D. NU NEZ ET AL. References [1] D. Nunez-Alarc´on, D. Pellegrino, J.B. Seoane-Sep´ulveda, On the Bohnenblust -- Hille inequality and a variant of Littlewoods 4/3 inequality, J. Funct. Anal. 264 (2013), 326 -- 336. [2] D. Nunez-Alarc´on, D. Pellegrino, J.B. Seoane-Sep´ulveda, D.M. Serrano-Rodr´ıguez, There exist multilinear Bohnenblust -- Hille constants (Cn)∞n=1 with limn→∞ (Cn+1 − Cn) = 0, J. Funct. Anal. 264 (2013), 429 -- 463. [3] D.M. Serrano-Rodr´ıguez, Improving the closed formula for subpolynomial constants in the multilinear Bohnenblust -- Hille inequalities, Linear Alg. Appl., in press. Departamento de Matem´atica, Universidade Federal da Para´ıba, 58.051-900 - Joao Pessoa, Brazil. E-mail address: [email protected] Departamento de Matem´atica, Universidade Federal da Para´ıba, 58.051-900 - Joao Pessoa, Brazil. E-mail address: [email protected] and [email protected] Departamento de An´alisis Matem´atico, Facultad de Ciencias Matem´aticas, Plaza de Ciencias 3, Universidad Complutense de Madrid, Madrid, 28040, Spain. E-mail address: [email protected] Departamento de Matem´atica, Universidade Federal da Para´ıba, 58.051-900 - Joao Pessoa, Brazil. E-mail address: [email protected]
1703.02909	1	1703	2017-03-08T16:48:00	Weak separation properties for closed subgroups of locally compact groups	[ "math.FA", "math.OA" ]	Three separation properties for a closed subgroup $H$ of a locally compact group $G$ are studied: (1) the existence of a bounded approximate indicator for $H$, (2) the existence of a completely bounded invariant projection of $VN\left(G\right)$ onto $VN_{H}\left(G\right)$, and (3) the approximability of the characteristic function $\chi_{H}$ by functions in $M_{cb}A\left(G\right)$ with respect to the weak$^{*}$ topology of $M_{cb}A\left(G_{d}\right)$. We show that the $H$-separation property of Kaniuth and Lau is characterized by the existence of certain bounded approximate indicators for $H$ and that a discretized analogue of the $H$-separation property is equivalent to (3). Moreover, we give a related characterization of amenability of $H$ in terms of any group $G$ containing $H$ as a closed subgroup. The weak amenability of $G$ or that $G_{d}$ satisfies the approximation property, in combination with the existence of a natural projection (in the sense of Lau and \"Ulger), are shown to suffice to conclude (3). Several consequences of (2) involving the cb-multiplier completion of $A\left(G\right)$ are given. Finally, a convolution technique for averaging over the closed subgroup $H$ is developed and used to weaken a condition for the existence of a bounded approximate indicator for $H$.	math.FA	math	WEAK SEPARATION PROPERTIES FOR CLOSED SUBGROUPS OF LOCALLY COMPACT GROUPS ZSOLT TANKO Abstract. Three separation properties for a closed subgroup H of a locally compact group G are studied: (1) the existence of a bounded approximate indicator for H, (2) the existence of a completely bounded invariant projection V N (G) → V NH (G), and (3) the approximability of the characteristic function χH by functions in McbA (G) with respect to the weak∗ topology of McbA (Gd). We show that the H-separation property of Kaniuth and Lau is characterized by the existence of certain bounded approximate indicators for H and that a discretized analogue of the H-separation property is equivalent to (3). Moreover, we give a related characterization of amenability of H in terms of any group G containing H as a closed subgroup. The weak amenability of G or that Gd satisfies the approximation property, in combination with the existence of a natural projection (in the sense of Lau and Ülger), are shown to suffice to conclude (3). Several consequences of (2) involving the cb-multiplier completion of A (G) are given. Finally, a convolution technique for averaging over the closed subgroup H is developed and used to weaken a condition for the existence of a bounded approximate indicator for H. 1. Introduction Our objective is to study connections between various forms of amenability for a locally compact group G and certain separation properties for closed subgroups, and moreover to establish relationships between these separa- tion properties. Following the influential work of Ruan [34], much work on the homology of the Fourier algebra A (G) as a completely contractive Banach algebra has affirmed this as the appropriate category in which to consider A (G) and the related algebras of abstract harmonic analysis (e.g. [1, 9, 17, 18]). Motivated by the success of this perspective, we focus on completely bounded projections, operator amenability, and the completely bounded multiplier algebra of A (G). We consider the following separation properties for a closed subgroup H of G: (1) The existence of a bounded approximate indicator for H. (2) The existence of a completely bounded A (G)-bimodule projection of V N (G) onto V NH (G). 2010 Mathematics Subject Classification. Primary 43A15, Secondary 43A22, 43A30, 46L07. Key words and phrases. locally compact group, group von Neumann algebra, Fourier algebra, completely bounded multiplier, invariant projection. 1 WEAK SEPARATION PROPERTIES FOR CLOSED SUBGROUPS 2 (3) When the characteristic function χH may be approximated by func- tions in B (G) or McbA (G) in the weak∗ topology on the correspond- ing algebra of the discretized group. Condition (3) is also considered for subsets of G that are not necessarily closed subgroups. Bounded approximate indicators for closed subgroups were introduced in [1] as a means of obtaining invariant projections. They have subsequently been shown to have an intimate connection with homological properties of A (G) and its completion in the cb-multipliers Acb (G) [9]. A result of Granirer and Leinert [19, Theorem B2] yields bounded approximate indi- cators in B (G) from weaker conditions than given in Definition 2.2. This useful tool is unavailable for nets in McbA (G) and Section 6 develops a convolution technique relative to the closed subgroup H that recovers the weakened condition in the cb-multiplier setting. The existence of invariant projections has been studied by several authors in connection with other separation properties and the existence of approx- imate identities for ideals in A (G) [1, 9, 10, 15, 17, 24]. In particular, in [24] it is shown that if G has the H-separation property, then an invariant projection V N (G) → V NH (G) exists. We show in Section 4 that, in fact, the H-separation property is equivalent to the existence of a bounded ap- proximate indicator for H consisting of positive definite functions that are identically one on H. An analogue of the H-separation property is more- over shown to characterize the B (G)-approximability of χH. Condition (3) was first studied in [1], where it was claimed that a bounded approximate indicator for H exists whenever χH is B (G)-approximable. This argument was later found to contain a gap [2]. We give examples in Section 3 showing that the cb-multiplier analogue is false. Conditions (1) to (3) are related to amenability properties of G and to homological properties of A (G), and it is this connection that is the main focus of the present article. We show in Section 4 that H is already amenable when χH is A (G)-approximable for any locally compact group G contain- ing H as a closed subgroup. In the case that G is amenable, the algebra A (G) has a bounded approximate identity and [15, Proposition 6.4] then as- serts that an invariant projection V N (G) → V NH (G) exists exactly when the ideal IA(G) (H) in A (G) has a bounded approximate identity. By [17], the latter occurs for every closed subgroup of G and it follows that an ap- proximate indicator exists for every closed subgroup, since (1G − eα)α is an approximate indicator for H when (eα)α is a bounded approximate identity for IA(G) (H). Thus all closed subgroups of an amenable locally compact group are separated in the strongest sense that we consider. For generic locally compact groups the situation is more complicated, although some strong connections are known to hold in general. For example, using the identity A (G) b⊗A (G) = A (G × G), it is routine to show that an approx- imate indicator for the diagonal G∆ in A (G × G) (and bounded there) is exactly a bounded approximate diagonal for A (G), the existence of which WEAK SEPARATION PROPERTIES FOR CLOSED SUBGROUPS 3 characterizes amenability of G [34]. Moreover, the existence of an invariant projection V N (G × G) → V NG∆ (G × G) characterizes the operator biflat- ness of A (G), a weaker homological condition than operator amenability. Contractive operator biflatness, which asks that this invariant projection to be a complete contraction, was recently shown to be equivalent to the existence of a contractive approximate indicator for G∆ in B (G × G) [9]. A bounded approximate indicator for H always yields the approximability of χH in the corresponding algebra, however it is unclear when the latter follows from the existence of an invariant projection V N (G) → V NH (G) alone. In Section 3 we show that if H satisfies certain weak forms of amenability, then the existence of a bounded map V N (G) → V NH (G) sat- isfying λ (s) 7→ χH (s) λ (s) -- a weaker condition than the existence of an invariant projection onto V NH (G) -- implies the McbA (G)-approximability of χH. We give an example in which the former condition fails while the latter holds. Establishing relations amongst the conditions (1) to (3) in the presence of amenability type conditions on H or G is the second goal of this article. 2. Preliminaries For a locally compact group G, the following algebras were defined by Eymard in [13], who established the basic properties we outline below. The space of coefficient functions of strongly continuous unitary representations of G, (2.1) B (G) = {hπ (·) ξηi : π : G → B (H) is a representation, ξ, η ∈ H} , forms a commutative completely contractive Banach algebra, the Fourier -- Stieltjes algebra of G, under pointwise multiplication and norm kukB(G) = inf {kξk kηk : u = hπ (·) ξηi with π, ξ, η as in (2.1)} . The operator space structure on B (G) arises from its identification with the dual space of the universal enveloping C ∗-algebra of L1 (G) -- the group C ∗-algebra C ∗ (G) of G -- via the duality hu, π (f )iB(G),C ∗(G) =ZG f u (cid:16)u ∈ B (G) , f ∈ L1 (G)(cid:17) . We refer to [12] for the theory of operator spaces and completely contractive Banach algebras. The positive definite functions in B (G) are those that correspond to positive functionals on C ∗ (G) and are denoted P (G). For u ∈ P (G) we have kukB(G) = kukL∞(G) = u (e). The adjoint on B (G) is given by u∗ (s) = u (s−1) (u ∈ B (G) , s ∈ G) and the self-adjoint functions in B (G) correspond to the self-adjoint bounded functionals on C ∗ (G). Consequently, given u ∈ B (G) self-adjoint, there ex- ist u± ∈ P (G) such that u = u+ − u− and kukB(G) = ku+kB(G) + ku−kB(G). WEAK SEPARATION PROPERTIES FOR CLOSED SUBGROUPS 4 The Fourier algebra of G is the closed ideal A (G) of B (G) given by the coefficients of the left regular representation λ : G → B (L2 (G)), which is defined by λ (s) ξ (t) = ξ(cid:16)s−1t(cid:17) (cid:16)s, t ∈ G, ξ ∈ L2 (G)(cid:17) . When necessary, we denote this representation of G by λG. The Fourier algebra coincides with the closure of the compactly supported functions in B (G), is closed under the adjoint, and is regular in the sense that for any K ⊂ G compact and U ⊃ K open, there is a function v ∈ A (G) with v (K) = 1 and v (G \ U) = 0. The group von Neumann algebra of G is the weak operator topology closure V N (G) of spanλ (G) in B (L2 (G)) and is identified with the dual of the Fourier algebra via hλ (s) , uiV N (G),A(G) = u (s) (s ∈ G, u ∈ A (G)) . Given an algebra A of functions on G and a subset E of G, we denote by IA (E) the ideal of functions in A vanishing on E. For a closed subgroup H of G, the annihilator IA(G) (H)⊥ coincides with the von Neumann algebra V NH (G) generated by λG (H) [39, Theorem 6], which is identified with V N (H) via the normal ∗-isomorphism defined by λH (h) 7→ λG (h) for h ∈ H. The preadjoint of this normal ∗-isomorphism is the restriction map rH : A (G) → A (H), which is thus a complete quotient. The closed subgroups of G are sets of spectral synthesis for A (G) [20], meaning that the ideal IA(G) (H) of A (G) coincides with the closure of the functions in A (G) that have compact support disjoint from H. The completely bounded (cb-) multiplier algebra McbA (G) of A (G) is the algebra of functions m on G for which mA (G) ⊂ A (G) and the map Mm : V N (G) → V N (G) is completely bounded, where Mm is the adjoint of the multiplication map u 7→ mu on A (G). Such functions are continuous and bounded, and form a completely contractive Banach algebra under pointwise multiplication and norm kmkMcbA(G) = kMmkcb. The cb-multipliers of A (G) admit the following representation theorem [22]. Theorem 2.1. (Gilbert's representation theorem) Let G be a locally com- pact group. A function m on G is in McbA (G) if and only if there exists a Hilbert space H and bounded continuous maps P, Q : G → H such that m(cid:16)s−1t(cid:17) = hP (t) Q (s)i (s, t ∈ G) . The norm kmkMcbA(G) is the infimum of the quantities kP k∞ kQk∞ taken over all such maps P and Q and Hilbert spaces H. It follows from Gilbert's representation theorem that B (G) ⊂ McbA (G) and that k·kMcbA(G) ≤ k·kB(G) on B (G). The restriction rH : McbA (G) → McbA (H) is a well defined complete contraction [11, Proposition 1.12]. The norm closure of A (G) in the potentially smaller cb-multiplier norm is de- noted Acb (G) and forms a completely contractive Banach subalgebra of the cb-multipliers with spectrum G [18, Proposition 2.2]. WEAK SEPARATION PROPERTIES FOR CLOSED SUBGROUPS 5 Since cb-multipliers of A (G) lie in L∞ (G), we may consider L1 (G) as a subspace of the dual of McbA (G). Taking the completion of L1 (G) with respect to the norm given, for f ∈ L1 (G), by kf kQ(G) = sup(cid:26)(cid:12)(cid:12)(cid:12)(cid:12)ZG f m(cid:12)(cid:12)(cid:12)(cid:12) : m ∈ McbA (G) with kmkMcbA(G) ≤ 1(cid:27) yields a predual Q (G) for McbA (G) [11, Proposition 1.10]. With this pre- dual, the cb-multipliers McbA (G) form a completely contractive dual Ba- nach algebra in the sense of Runde [35]. It follows from Theorem 2.1 that k·kL∞(G) ≤ k·kMcbA(G) and consequently k·kQ(G) ≤ k·kL1(G) on L1 (G). We let Cc (G) and Cb (G) denote respectively the continuous compactly supported and continuous bounded functions of G. We have A (G) ⊂ Acb (G) and A (G) ⊂ B (G) ⊂ McbA (G) ⊂ Cb (G) . The second containment is strict unless G is compact. An unpublished result of Ruan asserts that A (G) is closed in McbA (G) exactly when G is amenable, so that that the first and third containments are strict unless G is amenable. For u ∈ A (G) and v ∈ B (G), kuk∞ ≤ kukAcb(G) = kukMcbA(G) ≤ kukA(G) = kukB(G) , kvk∞ ≤ kvkMcbA(G) ≤ kvkB(G) , so the first, third, and fourth inclusions are in general contractive while the second is isometric. The locally compact group G equipped with the discrete topology is de- noted Gd. The inclusions B (G) ⊂ B (Gd) and McbA (G) ⊂ McbA (Gd) are complete isometries ([13] and [38, Corollary 6.3], respectively). For each s ∈ G the point mass δs ∈ ℓ1 (Gd) is contained in C ∗ (Gd) and in Q (Gd) as the evaluation functional at s, from which it follows that convergence in the weak∗ topology of B (Gd) or McbA (Gd) implies pointwise convergence. It will be important for us that, on bounded sets, the converse holds (see [13] regarding B (Gd) and [18, Lemma 2.6] or the useful Appendix A of [27] regarding McbA (Gd)). The separation properties for closed subgroups that we discuss are defined as follows. Definition 2.2. Let G be a locally compact group and H a closed subgroup. (1) A bounded approximate indicator for H is a bounded net (mα)α in McbA (G) satisfying (a) kurH (mα) − ukA(H) → 0 for all u ∈ A (H), and (b) kumαkA(G) → 0 for all u ∈ IA(G) (H). If (mα)α is in B (G) and is also bounded there, then we refer to a bounded approximate indicator in B (G). (2) A completely bounded projection V N (G) → V NH (G) is invariant if it is an A (G)-bimodule map. WEAK SEPARATION PROPERTIES FOR CLOSED SUBGROUPS 6 (3) Given a norm closed subalgebra A of B (G) or McbA (G) and E ⊂ G, the characteristic function χE is called A-approximable if χE is in the weak∗ closure of A in B (Gd) or McbA (Gd), respectively. The operator amenability of the Fourier algebra asserts the existence of a bounded approximate diagonal in the operator space projective tensor prod- uct A (G) b⊗A (G). This is a bounded net (dα)α in the tensor product which satisfies the norm convergence u · dα − dα · u → 0 and ∆ (dα) u → u (u ∈ A (G)) , where ∆ : A (G) b⊗A (G) → A (G) is the completely bounded linearization of the multiplication and the A (G) action on the tensor product is given by u·(v ⊗ w) = uv⊗w and (v ⊗ w)·u = v⊗wu, for u, v, w ∈ A (G). The product map ∆ is a complete quotient and the operator biflatness of A (G) asserts the existence of a completely bounded A (G)-bimodule left inverse to its adjoint ∆∗ : V N (G) → V N (G) ⊗V N (G), where the A (G) action on V N (G) is the dual action. The identification A (G) b⊗A (G) = A (G × G) (see [12, Theo- rem 7.2.4]) yields ker ∆ = IA(G×G) (G∆), from which it follows that such a left inverse is exactly an invariant projection V N (G × G) → V NG∆ (G × G). A thorough account of the homological conditions we consider is given in [36]. Leptin's classical result states that amenability of a locally compact group G is characterized by the existence of a bounded approximate identity in A (G) [29]. The locally compact group G is called weakly amenable when Acb (G) has a bounded approximate identity. This weaker notion was intro- duced by de Cannière and Haagerup [11], who showed that the free group on two generators is weakly amenable. The weak amenability of G is equivalent to the assertion that every cb-multiplier is the weak∗ limit of a bounded net in A (G). When A (G) is merely weak∗ dense in McbA (G), the group G is said to have the approximation property (see [21]). 3. Approximability of characteristic functions In this section, we investigate when characteristic functions of subsets of a locally compact group G are approximable. For a closed subgroup H of G, the assertion that χH is approximable may be viewed as a very weak form of subgroup separation. In [1], the discretized Fourier -- Stieltjes algebra Bd (G) is defined to be the weak∗ closure of B (G) in B (Gd). We make the analogous definition for the cb-multipliers of G. Definition 3.1. The discretized cb-multiplier algebra M d cbA (G) of a locally compact group G is the weak∗ closure of McbA (G) in McbA (Gd). The pred- ual Q (Gd) /M d cb (G) denote the weak∗ closures of A (G) in B (Gd) and in McbA (Gd), respectively. cbA (G) is denoted Qd (G). Let Ad (G) and Ad cb (G)⊥ of M d Given E ⊂ G, for χE to be approximable, it must already be that χE ∈ McbA (Gd), and the subsets of G for which this occurs are not well understood when Gd is not amenable. In the amenable case, the algebras WEAK SEPARATION PROPERTIES FOR CLOSED SUBGROUPS 7 McbA (Gd) and B (Gd) coincide [31] and the Cohen-Host idempotent theo- rem provides a complete description of the subsets of G with characteristic function in B (Gd). For discrete groups G, determining the approximable characteristic functions is exactly the problem of determining the sets with characteristic function in the cb-multipliers. When G is moreover weakly amenable, Corollary 5.4 of [9] together with Corollary 3.5 of [18] implies that such sets E are exactly those for which the ideal IA(G) (E) has a cb- multiplier bounded approximate identity. Example 3.2. Let G be a locally compact group and H a closed subgroup for which a bounded approximate indicator (mα)α exists. We show that χH is approximable. If s ∈ H, then we may find u ∈ A (H) with u (s) = 1, in which case mα (s) − 1 ≤ kurH (mα) − ukL∞(H) ≤ kurH (mα) − ukA(H) → 0. If s ∈ G \ H, then [13, Lemme 3.2] asserts that we may find w ∈ IA(G) (H) with w (s) = 1, and mα (s) ≤ kwmαkL∞(G) ≤ kwmαkA(G) → 0. Since McbA (G) is contained in McbA (Gd) and weak∗ and pointwise conver- gence coincide on bounded subsets of McbA (Gd), it follows that (mα)α has weak∗ limit χH in McbA (Gd). The algebra M d cbA (G) is a weak∗ closed subalgebra of McbA (Gd) and thus has separately weak∗ continuous multiplication. If we impose a rather weak condition on the discrete group Gd, then every function in McbA (Gd) may be approximated in the weak∗ topology by functions in A (G). Proposition 3.3. Let G be a locally compact group. The inclusion Acb (Gd) ⊂ Ad cb (G) always holds. Consequently, if Gd has the approxima- tion property, then McbA (Gd) = Ad cb (G). Proof. By Proposition 1 of [4] we have A (Gd) ⊂ Ad (G) and, because weak∗ convergence in B (Gd) implies weak∗ convergence in McbA (Gd), it follows that A (Gd) ⊂ Ad cb (G). Taking norm closures in McbA (Gd) yields Acb (Gd) ⊂ cb (G). If (eα)α is a net in A (Gd) converging weak∗ to 1G in McbA (Gd) Ad w∗ → m in and if m ∈ McbA (Gd), then meα ∈ A (Gd) for each α and meα McbA (Gd), implying that m ∈ Ad (cid:3) cb (G). In Section 3 of [1], that the Fourier -- Stieltjes algebra is the dual of a C ∗- algebra is noted to imply that the unique weak∗ continuous extension of the inclusion B (G) ⊂ Bd (G) is a quotient map B (G)∗∗ → Bd (G). We provide a more concrete construction of the analogous canonical map McbA (G)∗∗ → M d cbA (G). Let ιd : McbA (G) → McbA (Gd) and κQ : Q (Gd) → Q (Gd)∗∗ denote the inclusion maps. By the bipolar theorem M d cbA (G) that exploits the relation between McbA (G) and M d cbA (G)⊥ = McbA (G)⊥ and we have κQ (McbA (G)⊥) ⊂ McbA (G)⊥ = im (ιd)⊥ = ker (ι∗ d) , WEAK SEPARATION PROPERTIES FOR CLOSED SUBGROUPS 8 which together imply that the composition κQ→ Q (Gd)∗∗ ι∗ Q (Gd) d→ Q (G)∗∗ induces a map Qd (G) → Q (G)∗∗. Denote the adjoint of this induced map by τ : McbA (G)∗∗ → M d cbA (G). It is straightforward to verify that τ extends inclusion McbA (G) ⊂ M d cbA (G), so that τ (m) (s) = m (s) for m ∈ McbA (G) and s ∈ G. It follows that if a net (mα)α in McbA (G) converges weak∗ to ω ∈ McbA (G)∗∗, then τ (ω) (s) = lim α mα (s) (s ∈ G) , so that the map τ extracts pointwise limits from nets in McbA (G) that are weak∗ convergent in the bidual. Moreover, the range of τ consists of exactly those functions in M d cbA (G) that are limits of bounded nets in McbA (G). Proposition 3.4. Let G be a locally compact group and E ⊂ G. If there is a bounded map Ψ : V N (G) → V N (G) satisfying Ψ (λ (s)) = χE (s) λ (s) for all s ∈ G, then χEA (G) ⊂ Ad cb (G). If, moreover, χE ∈ McbA (Gd) and 1G is Acb (G)-approximable, then χE is Acb (G)-approximable. Proof. Let κA : A (G) → A (G)∗∗ and ιA : A (G) → McbA (G) be the inclu- sions and let σ denote the composition A (G) κA−→ A (G)∗∗ Ψ∗ −→ A (G)∗∗ ι∗∗ A−→ McbA (G)∗∗ τ−→ M d cbA (G) . For u ∈ A (G) and s ∈ G, with δs denoting the point mass at s in ℓ1 (Gd) ⊂ Q (Gd) ⊂ Qd (G), σ (u) (s) = hσ (u) , δsiM d cbA(G),Qd(G) dκQ (δs)iMcbA(G)∗∗,McbA(G)∗ A Ψ∗κA (u) , ι∗ = hι∗∗ = hΨ∗κA (u) , λ (s)iA(G)∗∗,V N (G) = hΨ (λ (s)) , uiV N (G),A(G) = χE (s) u (s) . cbA (G), it maps A (G) into Ad Thus χEu = σ (u) ∈ M d inclusion of McbA (G) into M d together with the weak∗ continuity of τ ι∗∗ in Ad then that χE ∈ McbA (Gd) implies χEeα multiplication in McbA (Gd). Since χEeα ∈ Ad Acb (G)-approximable. cbA (G) for all u ∈ A (G). Since τ extends the cb (G), which A implies that σ in fact has range cb (G). If (eα)α is a net in A (G) converging weak∗ to 1G in McbA (Gd), w∗ → χE, by weak∗ continuity of cb (G), we conclude that χE is (cid:3) For a subgroup H of a locally compact group G, it is straight for- ward to verify that χH is a positive definite function on Gd, so that χH ∈ B (Gd) ⊂ McbA (Gd) and the second part of Proposition 3.4 is ap- plicable to characteristic functions of subgroups. It is shown in Lemma 3.8 below that, when χHA (G) ⊂ Ad cb (G), we need only require 1H to be Acb (H)-approximable to deduce that χH is Acb (G)-approximable. WEAK SEPARATION PROPERTIES FOR CLOSED SUBGROUPS 9 In [30], Lau and Ülger define a projection Ψ on V N (G) to be natural if Ψ (λ (s)) = χE (s) λ (s) for some subset E of G. We may interpret Propo- sition 3.4 as imposing restrictions on which subsets of G can arise from a natural projection. Example 3.5. Let G be a locally compact group and H a closed subgroup. We show that an invariant projection Ψ : V N (G) → V NH (G) is natural. It is clear that Ψ (λ (s)) = λ (s) for s ∈ H. Let s ∈ G \ H and let Tα ∈ spanλ (H) converge weak∗ to Ψ (λ (s)) in V N (G). If u ∈ A (G) with u (s) = 1 and uH = 0, then hu · λ (s) , vi = hλ (s) , vui = v (s) u (s) = v (s) = hλ (s) , vi (v ∈ A (G)) , so u · λ (s) = λ (s). If S =Pj αjλ (sj) ∈ spanλ (H), then hu · S, vi =Pj αjv (sj) u (sj) = 0 so that u · S = 0. Thus 0 = u · Tα and Ψ (λ (s)) = 0. (v ∈ A (G)) , w∗ → u · Ψ (λ (s)) = Ψ (u · λ (s)) = Ψ (λ (s)) Let G be a locally compact group. A bounded net (eα)α in A (G) or Acb (G) is called a ∆-weak bounded approximate identity if it converges pointwise to 1G. This notion was introduced in [23] and shown in [30] to be closely related to the existence of natural projections. Reasoning as in Example 3.2 shows that a bounded approximate identity for Acb (G) is a ∆-weak one, so that Acb (G) has ∆-weak bounded approximate identity whenever G is weakly amenable. The function 1G is Acb (G)-approximable when Acb (G) has a ∆-weak bounded approximate identity. The construction of the map τ above and the proof of Proposition 3.4 may be carried out with McbA (G) replaced by B (G), but, to conclude that χH is B (G)-approximable using this result, we require 1G to be in the weak∗ closure of A (G) in B (Gd). The proof of Theorem 4.6 shows that the canoni- cal map A (G)∗∗ → Ad (G) extending inclusion A (G) ⊂ Ad (G) is surjective, so that 1G is then the weak∗ limit of a bounded net, which is then a ∆- weak bounded approximate identity for A (G), implying that G is already amenable [25, Theorem 5.1]. It is the availability of ∆-weak bounded ap- proximate identities in Acb (G) for a larger class of groups -- containing at least the weakly amenable groups -- that is responsible for the utility of Proposition 3.4. Whether the existence of a ∆-weak bounded approximate identity for Acb (G) implies weak amenability of G appears to be an open question. For a locally compact group G and closed subgroup H, let rH : McbA (Gd) → McbA (Hd) and eH : McbA (Hd) → McbA (Gd) denote respectively the restriction map and the extension by zero. For a ◦ function f on H let f denote its extension by zero to G. The restriction rH is a complete quotient and extension eH a complete isometry ([38, Corollary 6.3] or [37, Proposition 4.1]), and it is clear that rHeH = idMcbA(Hd) and eHrH = MχH , the multiplication by χH. WEAK SEPARATION PROPERTIES FOR CLOSED SUBGROUPS 10 Lemma 3.6. Let G be a locally compact group and H a closed subgroup. The maps rH and eH are weak∗ continuous. Proof. If f =Pn hr∗ H (f ) , mi = j=1 αjδxj ∈ Cc (Hd), then ◦ f ∈ Cc (Gd) and αjm (xj) =(cid:28) ◦ f , m(cid:29) (m ∈ McbA (Gd)) , nXj=1 showing that r∗ follows that rH is weak∗ continuous. H (Cc (Hd)) ⊂ Q (Gd). Since Cc (Hd) is dense in Q (Hd), it Now if f =Pn H (f ) , mi =D ◦ he∗ j=1 αjδxj ∈ Cc (Gd), then, for m ∈ McbA (Hd), m, fE = αjχH (xj) m (xj) =* nXj=1 nXj=1 αjχH (xj) δxj , m+ , j=1 αjχH (xj) δxj ∈ Cc (Hd). Therefore e∗ the claim follows by density, as above. H (Cc (Gd)) ⊂ Q (Hd) and (cid:3) and so Pn Lemma 3.7. Let G be a locally compact group and H a closed subgroup. The restriction rH maps M d cbA (H). In addition, the following are equivalent: cbA (G) into M d cb (G). (1) χH A (G) ⊂ Ad (2) eH(cid:16)Ad (3) Ad cb (H)(cid:17) ⊂ Ad cb (G) = IAd cb (G). cb(G) (G \ H) ⊕ IAd cb(G) (H) (algebraic direct sum). Proof. Since the restriction of a cb-multiplier of G to the closed subgroup H yields a cb-multiplier of H [11, Proposition 1.12], the first claim follows from weak∗ continuity of rH. (1) implies (2): If χHA (G) ⊂ Ad cb (G), then, because A (H) = rH (A (G)), eH (A (H)) = eH (rH (A (G))) = χHA (G) ⊂ Ad cb (G) and (2) follows by weak∗ continuity of eH. (2) implies (3): If eH(cid:16)Ad cb (H)(cid:17) ⊂ Ad weak∗ continuity of rH implies rH (m) ∈ Ad eHrH (m) ∈ Ad m = χHm + χG\H m ∈ IAd have trivial intersection. cb(G) (G \ H) + IAd cb (G), whence χG\H m = m − χHm ∈ Ad cb (G), then given m ∈ Ad cb (G), the cb (H) and it follows that χHm = cb (G) and therefore cb(G) (H). These ideals clearly (3)implies (1): Write m ∈ A (G) as m = m1 + m2 for m1 ∈ IAd and m2 ∈ IAd cb(G) (H), in which case χH m = m1 ∈ Ad cb (G). cb(G) (G \ H) (cid:3) When the equivalent conditions of Lemma 3.7 hold, condition (2) implies cb(G) (G \ H) and, because eH is isometric, condition that eH(cid:16)Ad cb (H)(cid:17) = IAd (3) asserts that Ad cb (G) = Ad cb (H) ⊕ IAd cb(G) (H). Lemma 3.8. Let G be a locally compact group and H a closed subgroup. If χHA (G) ⊂ Ad cb (G) and 1H is Acb (H)-approximable, then χH is Acb (G)- approximable. Proof. It follows from Lemma 3.7(2) that χH = eH (1H) ∈ Ad cb (G). (cid:3) WEAK SEPARATION PROPERTIES FOR CLOSED SUBGROUPS 11 Combining the results of this section, we obtain the following. Theorem 3.9. The characteristic function of a closed subgroup H of a locally compact group G is Acb (G)-approximable when either of the following conditions is satisfied: (1) Gd has the approximation property. (2) There is a bounded map Ψ : V N (G) → V N (G) such that Ψ (λ (s)) = χH (s) λ (s) for s ∈ G, which is satisfied if Ψ is a natural or invari- ant projection onto V NH (G), and 1H is Acb (H)-approximable, which occurs when H is weakly amenable or Hd has the approximation prop- erty. Proof. (1) When Gd has the approximation property, Proposition 3.3 as- serts that Ad cb (G) = McbA (Gd). Since McbA (Gd) contains the characteristic functions of all subgroups of G, we have χH ∈ Ad cb (G). (2) If Ψ : V N (G) → V NH (G) is a bounded map such that Ψ (λ (s)) = χH (s) λ (s) for s ∈ G and 1H is Acb (H)-approximable, then χHA (G) ⊂ Ad cb (G) by Proposition 3.4. Lemma 3.8 then asserts that χH is Acb (G)- approximable. Example 3.5 shows that invariant projections V N (G) → V NH (G) are natural, and the latter are by definition bounded map satis- fying the condition of (2). It was noted above that weak amenability of H implies 1H ∈ Ad cb (H) when Hd has the approximation property. (cid:3) cb (H), while Proposition 3.3 implies 1H ∈ Ad Theorem 3.7 of [1] claims that a bounded approximate indicator in B (G) for a closed subgroup H of the locally compact group G exists whenever χH is B (G)-approximable. The argument establishing this result was found to contain an error [2] and it is not known whether the claim holds. The following examples show that χH may be McbA (G)-approximable even when no invariant projection onto V NH (G) exists, in which case no bounded approximate indicator for H exists, either, by Proposition 5.1. Example 3.10. The locally compact group G = SL (2, R) contains H = F2 as a closed subgroup. It has recently been shown that Gd is weakly amenable [28], so that χH is Acb (G)-approximable by Proposition 3.3. Since G is con- nected, its group von Neumann algebra is injective [6, Corollary 6.9(c)], so there exists a completely bounded projection B (L2 (G)) → V N (G). If a completely bounded projection V N (G) → V NH (G) existed, then composi- tion would yield a completely bounded projection B (L2 (G)) → V NH (G), implying that V NH (G) = V N (H) is an injective von Neumann algebra [5]. It would follow that the discrete group H is amenable [32, (2.35)], which is false. Example 3.11. Let G = SL (2, R) and consider the diagonal subgroup G∆ of G × G. The Fourier algebra A (G) is not operator biflat by Corollary 3.7 of [9], meaning that no invariant projection V N (G × G) → V NG∆ (G × G) exists. But the weak amenability of Gd, noted in the preceding example, WEAK SEPARATION PROPERTIES FOR CLOSED SUBGROUPS 12 implies the weak amenability of (G × G)d, so that χG∆ is Acb (G × G)- approximable, again by Proposition 3.3. 4. The discretized H-separation property In this section, we characterize the approximability of the characteristic function of a closed subgroup H of a locally compact group G in the spirit of the H-separation property of Kaniuth and Lau. For a closed subgroup H of G, let PH (G) denote the norm closed convex set {u ∈ P (G) : u (H) = 1}. Definition 4.1. ([24]) A locally compact group G is said to have the H- separation property for a closed subgroup H if, for each s ∈ G \ H, there exists u ∈ PH (G) such that u (s) 6= 1. It is routine to verify that G has the H-separation property for any open, compact, or normal subgroup H, and it was shown by Forrest [16] that if G is a SIN group, then G has the H-separation property for every closed subgroup H. In [24], a fixed point argument is used to show that an invariant projection V N (G) → V NH (G) exists when the locally compact group G has the H-separation property (it is noted in Proposition 5.1 below that the projections arising this way are completely positive, in particular completely bounded). In fact, the following stronger result holds. Proposition 4.2. Let G be a locally compact group and H a closed subgroup. Then G has the H-separation property if and only if there exists a bounded approximate indicator for H in PH (G). Proof. Suppose that G has the H-separation property. The proof of [24, Proposition 3.1] constructs an invariant projection P : V N (G) → V NH (G) that is the weak∗ operator topology limit of a net (Muα)α, where uα ∈ PH (G) and Muα : V N (G) → V N (G) is the adjoint of the multiplication map u 7→ uαu on A (G). Let rH : A (G) → A (H) be the restriction map, which is a surjection satisfying r∗ H (V N (H)) ⊂ V NH (G). Given u ∈ A (H), let u ∈ A (G) with rH (u) = u, so that for T ∈ V N (H), hT, urH (uα)i = hT, rH (uuα)i H (T )) , ui H (T )) , ui = hMuα (r∗ → hP (r∗ = hr∗ = hT, ui . H (T ) , ui If w ∈ IA(G) (H), then hT, wuαi = hMuα (T ) , wi → hP (T ) , wi = 0 (T ∈ V N (G)) since P (T ) ∈ V NH (G) = IA(G) (H)⊥. Therefore urH (uα) → u weakly in A (H) for all u ∈ A (H) and wuα → 0 weakly in A (G) for all w ∈ IA(G) (H). Passing to convex combinations yields a bounded approximate indicator for H which remains in the convex set PH (G). WEAK SEPARATION PROPERTIES FOR CLOSED SUBGROUPS 13 Conversely, if (uα)α is a bounded approximate indicator for H in PH (G), then, given s ∈ G \ H, choose w ∈ IA(G) (H) with w (s) = 1, in which case uα (s) = uα (s) w (s) ≤ kuαwkL∞(G) ≤ kuαwkA(G) → 0 implies uα (s) 6= 1 eventually. (cid:3) For a closed subgroup H of a locally compact group G, we now show that a weaker form of the H-separation property, replacing the algebra B (G) with Bd (G), characterizes when χH is B (G)-approximable. Definition 4.3. Let G be a locally compact group and H a closed subgroup. The group G is said to have the discretized H-separation property if, for any s ∈ G \ H, there exists u ∈ Bd (G) ∩ PH (Gd) such that u (s) 6= 1. Proposition 4.4. Let G be a locally compact group and H a closed subgroup. Then G has the discretized H-separation property if and only if χH is B (G)- approximable. s kB(Gd) = un sH = 1 and u0 s (e) = 1 and thus has a weak∗ cluster point u0 Proof. Suppose that G has the discretized H-separation property and for each s ∈ G \ H let us ∈ Bd (G) ∩ PH (Gd) with us (s) 6= 1. Replac- ing us by 1 2 (1G + us), which remains in Bd (G) ∩ PH (Gd), we may assume s )n≥1 is in Bd (G) ∩ PH (Gd) with that us (s) < 1. Then the sequence (un kun s in the unit s (s) ≤ lim supn us (s)n = 0, so ball of Bd (G). Then u0 u0 s (s) = 0. Let F be the collection of finite subsets of G and for each s. Ordering F by inclusion, we have uF H = 1 and ptw → χH and by bounded- w∗ → χH in B (Gd), whence χH ∈ Bd (G). The converse is clear, given (cid:3) uF (s) = 0 eventually for each s ∈ G \ H, so that uF ness uF that the characteristic function of a subgroup is always in PH (Gd). F ∈ F let uF = Qs∈F u0 When the locally compact group G is second countable, the H-separation property may also be characterized in terms of a single function on G. Theorem 4.5. Let G be a second countable locally compact group. For a closed subgroup H, the following are equivalent: (1) G has the H-separation property. (2) There is u ∈ B (G) of norm one with {s ∈ G : u (s) = 1} = H. (3) There is u ∈ P (G) with {s ∈ G : u (s) = 1} = H. Proof. (1) implies (2): For s ∈ G \ H, let us ∈ PH (G) with us (s) 6= 1 and choose an open neighborhood Us of s with 1 /∈ us (Us). Then (Us)s∈G\H is an open cover of G \ H, so has a countable subcover (Usn)n≥1 by σ-compactness closed convex set PH (G), so that kukB(G) = u (e) = 1. Given s ∈ G \ H, choose n such that s ∈ Usn, so usn (s) 6= 1. Since kusnkL∞(G) = 1, we have of the open set G \ H in G. The function u = Pn≥1 2−nusn is in the norm Reusn (s) < 1, implying that Reu (s) = Pn≥1 2−nReusn (s) < 1 and hence that u (s) 6= 1. WEAK SEPARATION PROPERTIES FOR CLOSED SUBGROUPS 14 (2) implies (3): Replacing u by 1 2 (u + u∗), where u∗ (s) = u (s−1), we obtain function of B (G)-norm one (because the B (G) norm dominates the L∞ (G) norm) for which we may write u = u+ − u− with u± ∈ P (G) satisfying kukB(G) = ku+kB(G) + ku−kB(G). Then u+ (e)−u− (e) = u (e) = 1 = kukB(G) =(cid:13)(cid:13)(cid:13)u+(cid:13)(cid:13)(cid:13)B(G) and consequently ku−kB(G) = u− (e) = 0, so that u = u+ ∈ PH (G). +(cid:13)(cid:13)(cid:13)u−(cid:13)(cid:13)(cid:13)B(G) = u+ (e)+u− (e) (3) implies (1): This is clear. (cid:3) The amenability of H is known to imply the existence of an invariant projection V N (G) → V NH (G) for any locally compact group G containing H as a closed subgroup [8, Corollary 3.7] (see also [10]). We now show that the former condition may be characterized in terms of a separation property relative to any such G. Theorem 4.6. A locally compact group H is amenable if and only if χH is A (G)-approximable for some (equivalently, any) locally compact group G containing H as a closed subgroup. Proof. Fix a locally compact group G that contains H as a closed subgroup. Suppose that H is amenable. Let (eα)α be a bounded approximate identity for A (H) and let Ψ : V N (G) → V NH (G) be an an invariant projection. Example 3.5 shows that Ψ (λG (s)) = χH (s) λG (s) for all s ∈ G and the ptw → 1H. Recall that the adjoint of argument of Example 3.2 shows that eα the restriction rH : A (G) → A (H) is a ∗-isomorphism τ = r∗ H : V N (H) → V NH (G) taking λH (s) to λG (s) for all s ∈ H. The composition V N (H)∗ (τ ∗)−1 → V NH (G)∗ Ψ∗ → V N (G)∗ satisfies, for s ∈ G, DΨ∗ (τ ∗)−1 (eα) , λG (s)E = D(τ ∗)−1 (eα) , χH (s) λG (s)E = Deα, τ −1 (χH (s) λG (s))E =  heα, λH (s)i , 0, s ∈ H s /∈ H → χH (s) . Let E be a weak∗ cluster point of the bounded net (cid:16)Ψ∗ (τ ∗)−1 (eα)(cid:17)α in V N (G)∗, so that hE, λG (s)i = χH (s) for all s ∈ G by the above computa- tion. Letting (uα)α be a bounded net in A (G) converging weak∗ to E, we have uα w∗ → χH in B (Gd) by boundedness. Conversely, suppose that (uα)α is a net in A (G) such that uα w∗ → χH in B (Gd). Analogous to the proof of Lemma 3.6, the restriction rH : B (Gd) → B (Hd) is weak∗ continuous, so that rH (uα) w∗ → 1H and 1H is in the weak∗ clo- sure of A (H) in B (Hd). Viewing λH as a representation of Hd, the universal ptw → χH and therefore uα WEAK SEPARATION PROPERTIES FOR CLOSED SUBGROUPS 15 property of C ∗ (Hd) yields a quotient ∗-homomorphism C ∗ (Hd) → C ∗ δ (H) : ωHd (s) 7→ λH (s), where ωHd denotes the universal representation of Hd and C ∗ δ (H) the C ∗-algebra generated by λH (H) in B (L2 (H)), a subalgebra of V N (H). Composing with the inclusion, we obtain a ∗-homomorphism Ψ : C ∗ (Hd) → V N (H). The adjoint Ψ∗ : A (H)∗∗ → B (Hd) is the weak∗ continuous extension of the inclusion A (H) ⊂ B (Hd), so that ker Ψ = A (H)⊥ (preannihilator taken with respect to the B (Hd) -- C ∗ (Hd) duality). Then A (H)⊥ = Ad (H)⊥ is an ideal in C ∗ (Hd) and Ψ drops to an injective ∗-homomorphism C ∗ (Hd) /Ad (H)⊥ → V N (H). This injec- tive ∗-homomorphism is isometric, so that its adjoint A (H)∗∗ → Ad (H) is a quotient map and given ǫ > 0, the function 1H is the weak∗ limit in B (Hd) of a net (vα)α in A (H) of bound 1 + ǫ. The adjoint on B (Hd) is a weak∗ continuous isometry preserving A (H), so that we may replace vα by 1 α ∈ P (H) ∩ A (H) with α kB(H). Passing to subnets, let v± be weak∗ limits kvαkB(H) = kv+ of (v± α) and assume that vα = v+ α kB(H) + kv− 2 (vα + v∗ α )α in B (Hd). Then α − v− α for v± 1 = lim α vα (e) = lim α v+ α (e) − lim α v− α (e) = v+ (e) − v− (e) and so v− (e) = v+ (e) − 1 = lim α Let ωǫ be a weak∗ cluster point of (v+ V N (H) satisfying, for s ∈ H, v+ α (e) − 1 = lim − 1 ≤ ǫ. α )α in A (H)∗∗, so that ωǫ is a state on α(cid:13)(cid:13)(cid:13)B(H) α (cid:13)(cid:13)(cid:13)v+ vα (s)(cid:12)(cid:12)(cid:12)(cid:12) α (e) = v− (e) ≤ ǫ. v− α hωǫ, λH (s)i − 1 = (cid:12)(cid:12)(cid:12)(cid:12)lim α (s)(cid:12)(cid:12)(cid:12) ≤ lim α (cid:12)(cid:12)(cid:12)v− v+ α (s) − lim α = lim α Letting ω be a weak∗ cluster point of the states (ωǫ)ǫ>0 on V N (H), we have hω, λH (s)i = 1 for all s ∈ H, and any extension of ω to a state on B (L2 (H)) still takes the value 1 on the unitaries λH (s) for s ∈ H. The amenability of H follows: by a Cauchy-Schwarz argument, any extension of ω to a state on B (L2 (H)) is invariant under the conjugation action of the unitaries λH (s) for s ∈ H, whence λH is an amenable representation of H (see [3]). (cid:3) Corollary 4.7. Let G be a locally compact group and H a closed subgroup. Then H is amenable if and only if G has the discretized H-separation prop- erty witnessed by functions in Ad (G) ∩ PH (Gd). Proof. In the argument establishing Proposition 4.4, substituting 1 for the function 1 if and only if χH is A (G)-approximable. s + us) 2 (1G + us) yields a proof that G has the desired property (cid:3) 2 (u2 5. Invariant projections and bounded approximate indicators In this section we establish some consequences of the existence of a bounded approximate indicator for a closed subgroup of a locally compact WEAK SEPARATION PROPERTIES FOR CLOSED SUBGROUPS 16 group. We first provide the well known argument that this stronger separa- tion property indeed yields invariant projections. For a commutative com- pletely contractive Banach algebra A, let CBA (A∗) denote the completely bounded A-bimodule maps on A∗. This space has compact unit ball when given the weak∗ operator topology, which is determined by the seminorms Ψ 7→ hΨ (ϕ) , ai (ϕ ∈ A∗, a ∈ A) . Proposition 5.1. Let G be a locally compact group and H a closed subgroup. If there is a bounded approximate indicator for H, then there is a completely bounded invariant projection V N (G) → V NH (G). If there is a bounded approximate indicator for H consisting of positive definite functions, then there is a completely positive invariant projection V N (G) → V NH (G). Proof. Let (mα)α a bounded approximate indicator for H, so that the net of multiplication maps (Mmα)α in CBA(G) (V N (G)) is then bounded and thus has a weak∗ operator topology cluster point Ψ ∈ CB (V N (G)). Passing to a subnet if necessary, we may assume that Ψ is the limit of this net. For u, v ∈ A (G) and T ∈ V N (G), hΨ (v · T ) , ui = lim α hT, mαuvi = hΨ (T ) , uvi = hv · Ψ (T ) , ui , showing that Ψ is invariant. Given T ∈ V NH (G), so that T = r∗ some S ∈ V N (H), we have for u ∈ A (G) that H (S) for hΨ (T ) , ui = lim α = lim α hS, rH (uuα)i hS, rH (u) rH (uα)i = hS, rH (u)i = hT, ui , whence Ψ is the identity on V NH (G). For T ∈ V N (G) and u ∈ IA(G) (H) we have hΨ (T ) , ui = limα hS, uuαi = 0, so that Ψ maps into IA(G) (H)⊥ = V NH (G) and is thus a projection onto V NH (G). If the functions mα are in P (G), then the maps Mmα are completely positive [11, Proposition 4.2] and by [33, Theorem 7.4] their weak∗ operator topology cluster point Ψ is also completely positive. (cid:3) From the preceding we obtain an analogous result for Acb (G), at least when G is a weakly amenable locally compact group. Proposition 5.2. Let G be a weakly amenable locally compact group and H a closed subgroup. If there is a bounded approximate indicator for H, then there is a completely bounded invariant projection Acb (G)∗ → IAcb(G) (H)⊥. Proof. Let (mα)α an approximate indicator for H. Since A (G) is an ideal in McbA (G), so too is its closure Acb (G), so that multiplication by mα is a completely bounded Acb (G)-module map on Acb (G). Denote its adjoint by Mmα. Passing to a subnet, we may assume that (Mmα)α has a weak∗ operator topology limit Ψ ∈ CB (Acb (G)∗), and passing to a further subnet we may assume that the net of maps (Mmα)α in CBA(G) (V N (G)) also has a weak∗ WEAK SEPARATION PROPERTIES FOR CLOSED SUBGROUPS 17 operator topology limit ΨA, which is an invariant projection V N (G) → V NH (G) by the argument of Proposition 5.1. For u, v ∈ Acb (G) and T ∈ Acb (G)∗, we have hΨ (v · T ) , ui = lim α hT, mαuvi = hΨ (T ) , uvi = hv · Ψ (T ) , ui , so Ψ is invariant. Let ι : A (G) → Acb (G) be the inclusion. If T ∈ Acb (G)∗ and u ∈ A (G), then hΨAι∗ (T ) , ui = lim α = lim α hι∗ (T ) , umαi hT, ι (u) mαi = hΨ (T ) , ι (u)i = hι∗Ψ (T ) , ui and ΨAι∗ = ι∗Ψ by density of A (G) in Acb (G). It follows that ι∗Ψ2 = ΨAι∗Ψ = Ψ2 which, together with injectivity of ι∗, IAcb(G) (H)⊥, then ι∗ (T ) ∈ IA(G) (H)⊥ and Aι∗ = ΨAι∗ = ι∗Ψ, implies that Ψ2 = Ψ. If T ∈ hΨ (T ) , ι (u)i = hΨAι∗ (T ) , ui = hι∗ (T ) , ui = hT, ι (u)i (u ∈ A (G)) , whence Ψ (T ) = T , again by density of A (G). Therefore IAcb(G) (H)⊥ is contained in the range of Ψ. Finally, for any T ∈ Acb (G)∗, if u ∈ IA(G) (H), then hΨ (T ) , ι (u)i = hΨAι∗ (T ) , ui = 0, so Ψ (T ) ∈ IA(G) (H)⊥ = (cid:18)IA(G) (H)Acb(G)(cid:19)⊥ . That Acb (G) has bounded approximate identity im- plies every set of synthesis for A (G) is one for Acb (G) [18, Proposition 3.1] and, because compactly supported functions in Acb (G) are in A (G), that H is of spectral synthesis for Acb (G) is exactly the assertion that IA(G) (H)Acb(G) = IAcb(G) (H). Therefore Ψ has range in IAcb(G) (H)⊥. (cid:3) Note that the arguments of the preceding two propositions yield projec- tions of completely bounded norm at most the bound on an approximate indicator for the subgroup. Let ΛG denotes the Cowling -- Haagerup constant, that is, the infimum of bounds on approximate identities for Acb (G). Corollary 5.3. Let G be a weakly amenable locally compact group and H a closed subgroup for which an approximate indicator of bound C exists. The following hold: (1) IAcb(G) (H) has an approximate identity of bound (1 + C) ΛG. (2) An approximate indicator for H of bound 1 + (1 + C) ΛG exists that takes the value one on H. (3) An approximate indicator for H of bound CΛG exists in Acb (G). Proof. (1) The argument of Proposition 5.2 yields an invariant projection Acb (G)∗ → IAcb(G) (H)⊥ of norm at most C and, because the Banach alge- bra Acb (G) has an approximate identity of bound ΛG, it follows from [15, WEAK SEPARATION PROPERTIES FOR CLOSED SUBGROUPS 18 Proposition 6.4] and its proof that the ideal IAcb(G) (H) has an approximate identity of bound (1 + C) ΛG. (2) If (eα)α is an approximate identity for IAcb(G) (H) of bound (1 + C) ΛG, then (1G − eα)α is an approximate indicator for H with the claimed norm bound. (3) Let (eα)α∈A be a bounded approximate identity for Acb (G), let (mβ)β∈B a bounded approximate indicator for H, and for γ = (cid:16)β, (αβ ′)β ′∈B(cid:17) ∈ B × AB set uγ = mβeαβ , which is in the ideal Acb (G) of McbA (G). Giving B × AB the product order, w ∈ IA(G) (H) we have the norm limits for u ∈ A (H) and lim γ∈B×AB rH (uγ) u = lim β lim α rH (mβeα) u = lim β rH (mβ) u = u and lim γ∈B×AB uγw = lim β lim α mβeαw = lim β mβw = 0 by [26, p. 69], hence (uγ)γ∈B×AB is a bounded approximate indicator for H of norm bound supα keαkMcbA(G) supβ kmβkMcbA(G). (cid:3) 6. Convergence of cb-multipliers and averaging over closed subgroups Fix a locally compact group G and a closed subgroup H. It is folklore that the convergence properties of nets of cb-multipliers can be improved by convolving them with probability measures in Cc (G). For example, Knudby recently recorded the following, the second part of which originates in an argument of Cowling and Haagerup [7, Proposition 1.1]. If a net (mα)α of functions on a topological space converges uniformly on compact sets to a function m, we write mα Theorem 6.1. ([27, Lemma B.2]) Let (mα)α be a bounded net in McbA (G), m ∈ McbA (G), and let f ∈ Cc (G) be such that f ≥ 0 and RG f = 1. Convolution on the left with f is a contraction on McbA (G) and the following hold: ucs→ m. (1) If mα (2) If mα A (G). w∗ → m in McbA (G), then f ∗ mα ucs→ m, then k(f ∗ mα) u − (f ∗ m) ukA(G) → 0 for all u ∈ ucs→ f ∗ m. In this section, we develop an analogue of the convolution technique rel- ative to a closed subgroup. Fix a function f ∈ Cc (H) such that f ≥ 0 and RH f = 1. For any function f on G and s, t ∈ G, let sf (t) = f (st). Definition 6.2. For u ∈ Cb (G), define a function Ωf (u) on G by the formula We will show that Ωf defines a bounded map on McbA (G). For a Hilbert space H, let Cc (G, H) and Cb (G, H) denote the continuous functions G → H that are compactly supported and bounded, respectively. Ωf (u) (s) =ZH f (h) u(cid:16)h−1s(cid:17) dh (s ∈ G) . WEAK SEPARATION PROPERTIES FOR CLOSED SUBGROUPS 19 If u ∈ Cc (G, H) then for Lemma 6.3. Let H be a Hilbert space. any ǫ > 0 there is an open neighborhood U of the identity e such that supt∈G ku (st) − u (t)k < ǫ for all s ∈ U. Proof. The standard proof in the case that H = C, for example [14, Propo- sition 2.6], works for any Hilbert space. (cid:3) If u ∈ Cb (G, H), s0 ∈ G, Lemma 6.4. Let H be a Hilbert space. and ǫ > 0, then there is an open neighborhood U of s0 in G such that suph∈H kf (h) u (sh) − f (h) u (s0h)k < ǫ for all s ∈ U. Proof. If u = 0, then the claim trivially holds, so assume u 6= 0. Since H is closed in G, the function f extends to a continuous compactly supported function f ′ on G. Assume that s0 = e. Since f ′u is compactly supported, Lemma 6.3 yields an open neighborhood U of e such that kf ′u (st) − f ′u (t)k < sup t∈G ǫ 2 and sup t∈G f ′ (st) − f ′ (t) < ǫ 2 kuk∞ for all s ∈ U. Then sup h∈H kf (h) u (sh) − f (h) u (h)k ≤ sup t∈G ≤ sup t∈G kf ′ (t) u (st) − f ′ (t) u (t)k (kf ′ (t) u (st) − f ′u (st)k + kf ′u (st) − f ′ (t) u (t)k) ǫ f ′ (st) − f ′ (t) + 2 < kuk∞ sup t∈G < ǫ, for all s ∈ U. For s0 6= e, the above argument with u replaced by s0u yields a neighborhood U of e and s0U is then the desired neighborhood of s0. (cid:3) Proposition 6.5. If u ∈ McbA (G), kΩf (u)kMcbA(G) ≤ kukMcbA(G) and rHΩf (u) = f ∗ rH (u). then Ωf (u) ∈ McbA (G) with Proof. Let u ∈ McbA (G) and apply Gilbert's theorem to obtain a Hilbert space H and functions P, Q ∈ Cb (G, H) such that u (s−1t) = hP (t) Q (s)i for all s, t ∈ G. Then, for s, t ∈ G, f (h) Q (sh) dh(cid:29). Ωf (u)(cid:16)s−1t(cid:17) =ZH We show that q (s) =RH f (h) Q (sh) dh defines a bounded continuous func- f (h) u(cid:16)h−1s−1t(cid:17) dh =(cid:28)P (t)(cid:12)(cid:12)(cid:12)(cid:12)ZH tion on G, from which it will follow that Ωf (u) is in McbA (G), again by Gilbert's theorem. Define Q′ : G → L1 (H, H) by Q′ (s) = f (sQ), which maps into L1 (H, H) since f has compact support. Set K = suppf and let K denote the Haar measure of K, which is nonzero and finite by continuity of the nonzero, compactly supported function f . Given s0 ∈ G and ǫ > 0, Lemma 6.4 yields an open neighborhood U of s0 in G such that kQ′ (s) − Q′ (s0)kL∞(H,H) = sup h∈H kf (h) Q (sh) − f (h) Q (s0h)k < ǫ K WEAK SEPARATION PROPERTIES FOR CLOSED SUBGROUPS 20 for all s ∈ U. Since Q′ (s) is supported in K for every s ∈ G, it follows that kQ′ (s) − Q′ (s0)kL1(H,H) = kχK (Q′ (s) − Q′ (s0))kL1(H,H) ≤ kχKkL1(H,H) kQ′ (s) − Q′ (s0)kL∞(H,H) < ǫ for all s ∈ U. Thus Q′ is continuous and so too is q, the latter being the Using that f is nonnegative with mass one, if s ∈ G, then kq (s)k ≤ composition of Q′ with the bounded map L1 (H, H) → H : g 7→RH g. RH f (h) kQ (sh)k dh ≤ kQk∞, so q is bounded with kqk∞ ≤ kQk∞. By the norm characterization of Gilbert's theorem, kΩf (u)kMcbA(G) ≤ kP k∞ kqk∞ ≤ kP k∞ kQk∞ and since P , Q, and H are an arbitrary rep- resentation of u, we conclude that kΩf (u)kMcbA(G) ≤ kukMcbA(G). Finally, we have for s ∈ H that rHΩf (u) (s) = ZH = ZH f (h) u(cid:16)h−1s(cid:17) dh f (h) rH (u)(cid:16)h−1s(cid:17) dh = f ∗ rh (u) (s) . (cid:3) Theorem 6.6. Let (mα)α be a bounded net in McbA (G) and let m ∈ McbA (H). The following hold: (1) If rH (mα) w∗ (2) If rH (mα) ucs→ m, then krHΩf (mα) u − (f ∗ m) ukA(H) → 0 for all → m in McbA (H), then rHΩf (mα) ucs→ f ∗ m. u ∈ A (H). Proof. These follow immediately from Theorem 6.1 and Proposition 6.5. (cid:3) In our applications, the preceding theorem will be applied with m = 1H, which is fixed under convolution with f on the left. We list some additional properties that the map Ωf enjoys. (1) An argument very similar to that establishing Proposition 6.5 shows that Ωf (u) is bounded and continuous for any bounded continuous function u on G. (2) If u = hπ (·) ξηi is a coefficient of the unitary representation π of G, then Ωf (u) (s) = ZH f (h) hπ (s) ξπ (h) ηi dh so Ωf (u) ∈ = (cid:28)π (s) ξ(cid:12)(cid:12)(cid:12)(cid:12)(cid:18)ZH B (G), and f (h) π (h) dh(cid:19) η(cid:29) , from kRH f (h) π (h) dhk RH f (h) kπ (h)k dh = 1 it follows that kΩf (u)kB(G) ≤ kukB(G). Thus Ωf restricts to a contraction on B (G) and moreover restricts to a contraction on A (G), since Ωf (u) is a coefficient of the same representation as u. ≤ WEAK SEPARATION PROPERTIES FOR CLOSED SUBGROUPS 21 (3) An argument similar to that establishing the weak∗ continuity of the map Φf in the proof of [21, Lemma 1.16] shows that Ωf is weak∗ continuous on McbA (G) with preadjoint mapping g ∈ L1 (G) to the L1 (G) function s 7→RH f (h) g (hs) dh. Say that a net (mα)α of functions on a topological space X converges locally le→ 0 if for any compact subset eventually to zero on A ⊂ X and write mα K of A there is an index α0 such that mαK = 0 for all α ≥ α0. Proposition 6.7. If (mα)α is a bounded net in McbA (G) and m′ Ωf (mα), then the net (m′ α)α has the same norm bound as (mα)α and α = (1) if rH (mα) ucs→ 1H, then ku · rH (m′ α) − ukA(H) → 0 for all u ∈ A (H), and (2) if mα le→ 0 on G \ H, then m′ α le→ 0 on G \ H. If the bounded net (mα)α satisfies the hypotheses of both (1) and (2), then (m′ α)α is a bounded approximate indicator for H. Proof. The claim regarding norm bounds holds since the map Ωf of Section 6 is a contraction on McbA (G). (1) If rH (mα) ucs→ 1H, then, since restriction is a contraction from McbA (G) into McbA (H), the net (rH (mα))α is bounded and (1) follows from Theorem 6.6. (2) Suppose that mα le→ 0 on G \ H. Let K ⊂ G \ H be compact and choose α0 such that α ≥ α0 implies mα = 0 on the compact set (supp (f ))−1 K. For α ≥ α0, if s ∈ K and h ∈ H, then f (h) mα (h−1s) = 0 since either h /∈ supp (f ) or h−1s ∈ (supp (f ))−1 K, implying that m′ α (s) = RH f (h) mα (h−1s) dh = 0. Therefore m′ αu = 0 eventually by (2), so certainly kum′ If (mα)α satisfies the hypotheses of both (1) and (2), then (m′ α)α satisfies the first condition of Definition 2.2. If u ∈ IA(G) (H) has compact support, then m′ αkA(G) → 0. Using that H is of spectral synthesis in A (G), if u ∈ IA(G) (H) is arbitrary, then given ǫ > 0 choose u0 ∈ IA(G) (H) of compact support with ku − u0kA(G) < ǫ. For sufficiently large α, α = 0 on K, for all α ≥ α0. kum′ αkA(G) ≤ ku0m′ αkA(G) + ku0m′ α − um′ αkA(G) < ǫ km′ αkMcbA(G) , and thus kum′ αkA(G) → 0 by boundedness of (m′ α)α. (cid:3) Proposition 6.7 allows one to obtain approximate indicators consisting of cb-multipliers by verifying the same conditions that yielded approximate indicators in [1]. References [1] O. Y. Aristov, V. Runde, N. Spronk, Operator biflatness of the Fourier algebra and approximate indicators for subgroups. J. Funct. Anal. 209(2) (2004), 367 -- 387. [2] O. Y. Aristov, V. Runde, N. Spronk, Z. Tanko, Corrigendum to: "Operator biflatness of the Fourier algebra and approximate indicators for subgroups" [J. Funct. Anal. 209(2) (2004) 367 -- 387]. J. Funct. Anal. 270(6) (2016), 2381 -- 2382. WEAK SEPARATION PROPERTIES FOR CLOSED SUBGROUPS 22 [3] M. B. Bekka, Amenable unitary representations of locally compact groups. Invent. Math. 100(2) (1990), 383 -- 401. [4] M. B. Bekka, E. Kaniuth, A. T. Lau, G. Schlichting, On C ∗-algebras associated with locally compact groups. Proc. Amer. Math. Soc. 124(10) (1996), 3151 -- 3158. [5] E. Christensen, A. M. Sinclair, On von Neumann algebras which are complemented subspaces of B (H). J. Funct. Anal. 122 (1994), 91 -- 102. [6] A. Connes, Classification of injective factors. Cases II1, II∞, III∞, λ 6= 1. Ann. of Math. (2) 104(1) (1976), 73 -- 115. [7] M. Cowling, U. Haagerup, Completely bounded multipliers of the Fourier algebra of a simple Lie group of real rank one. Invent. Math. 96(3) (1989), 507 -- 549. [8] J. Crann, On hereditary properties of quantum group amenability. arXiv:1603.03842 [math.OA] [9] J. Crann, Z. Tanko, On the operator homology of the Fourier algebra and its cb- multiplier completion. arXiv:1602.05259 [math.OA] [10] A. Derighetti, Conditional expectations on CVp (G). Applications. J. Funct. Anal. 247(1) (2007), 231 -- 251. [11] J. de Cannière, U. Haagerup, Multipliers of the Fourier algebras of some simple Lie groups and their discrete subgroups. Amer. J. Math. 107(2) (1985), 455 -- 500. [12] E. G. Effros, Z.-J. Ruan, Operator Spaces. London Mathematical Society Mono- graphs, New Series 23, Oxford University Press, New York, 2000. [13] P. Eymard, L'algèbre de Fourier d'un groupe localement compact. Bull. Soc. Math. France 92(1964), 181 -- 236. [14] G. B. Folland, A course in abstract harmonic analysis. Studies in Advanced Mathe- matics. CRC Press, Boca Raton, Florida, 1995. [15] B. Forrest, Amenability and bounded approximate identities in ideals of A (G). Illinois J. Math. 34(1) (1990), 1 -- 25. [16] B. Forrest, Amenability and ideals in A (G). J. Austral. Math. Soc. Ser. A 53(2) (1992), 143 -- 155. [17] B. Forrest, E. Kaniuth, A. T. Lau, N. Spronk, Ideals with bounded approximate identities in Fourier algebras. J. Funct. Anal. 203(1) (2003), 286 -- 304. [18] B. Forrest, V. Runde, N. Spronk, Operator amenability of the Fourier algebra in the cb-multiplier norm. Canad. J. Math. 59(5) (2007), 966 -- 980. [19] E. E. Granirer, M. Leinert, On some topologies which coincide on the unit sphere of the Fourier -- Stieltjes algebra B(G) and of the measure algebra M (G). Rocky Moun- tain J. Math. 11(1981), 459 -- 472. [20] C. Herz, Harmonic synthesis for subgroups. Ann. Inst. Fourier (Grenoble) 23(3) (1973), 91 -- 123. [21] U. Haagerup, J. Kraus, Approximation properties for group C ∗-algebras and group von Neumann algebras. Trans. Amer. Math. Soc. 344(2) (1994), 667 -- 699. [22] P. Jolissaint, A characterization of completely bounded multipliers of Fourier algebras. Colloq. Math. 63(2) (1992), 311 -- 313. [23] C. A. Jones, C. D. Lahr, Weak and norm approximate identities are different, Pacific J. Math. 72 (1977), no. 1, 99 -- 104. [24] E. Kaniuth, A. T. Lau, A separation property of positive definite functions on locally compact groups and applications to Fourier algebras. J. Funct. Anal. 175(1) (2000), 89 -- 110. [25] E. Kaniuth, A. Ülger, The Bochner-Schoenberg-Eberlein property for commutative Banach algebras, especially Fourier and Fourier-Stieltjes algebras. Trans. Amer. Math. Soc. 362(8) (2010), 4331 -- 4356. [26] J. L. Kelley, General Topology. Graduate Textbooks in Mathematics 27, Springer -- Verlag, New York -- Berlin -- Heidelberg, 1955. [27] S. Knudby, The weak Haagerup property. Trans. Amer. Math. Soc. 368(5) (2016), 3469 -- 3508. WEAK SEPARATION PROPERTIES FOR CLOSED SUBGROUPS 23 [28] S. Knudby, K. Li, Approximation properties of simple Lie groups made discrete. J. Lie Theory 25(4) (2015), 985 -- 1001. [29] H. Leptin, Sur l'algèbre de Fourier d'un groupe localement compact. C. R. Acad. Sci. Paris Sér. A-B 266 (1968), A1180 -- A1182. [30] A. T. Lau, A. Ülger, Characterization of closed ideals with bounded approximate identities in commutative Banach algebras, complemented subspaces of the group von Neumann algebras and applications. Trans. Amer. Math. Soc. 366(8) (2014), 4151 -- 4171. [31] V. Losert, Properties of the Fourier algebra that are equivalent to amenability. Proc. Amer. Math. Soc. 92(3) (1984), 347 -- 354. [32] A. L. T. Paterson, Amenability. Mathematical Surveys and Monographs, 29. Amer- ican Mathematical Society, Providence, RI, 1988. [33] V. Paulsen, Completely bounded maps and operator algebras. Cambridge Studies in Advanced Mathematics, 78. Cambridge University Press, Cambridge, 2002. [34] Z.-J. Ruan, The operator amenability of A (G). Amer. J. Math. 117 (1995), 1449 -- 1474. [35] V. Runde, Amenability for dual Banach algebras. Studia Math. 148 (2001), 47 -- 66. [36] V. Runde, Lectures on amenability. Lecture Notes in Mathematics, 1774. Springer- Verlag, Berlin, 2002. [37] A. P. Stan, On idempotents of completely bounded multipliers of the Fourier algebra A (G). Indiana Univ. Math. J. 58(2) (2009), 523 -- 535. [38] N. Spronk, Measurable Schur multipliers and completely bounded multipliers of the Fourier algebras. Proc. London Math. Soc. (3) 89(1) (2004), 161 -- 192. [39] M. Takesaki, N. Tatsuuma, Duality and subgroups. Ann. of Math. 93(2) (1971) 344 -- 364. Department of Pure Mathematics, University of Waterloo, Waterloo, Ontario, N2L 3G1, Canada E-mail address: [email protected]
1809.06583	1	1809	2018-09-18T08:25:57	Carleson measures and Toeplitz operators on small Bergman spaces on the ball	[ "math.FA" ]	We study the Carleson measures and the Toeplitz operators on the class of so-called small weighted Bergman spaces, introduced recently by Seip. A characterization of Carleson measures is obtained which extends Seip's results from the unit disc of $\mathbb C$ to the unit ball of $\mathbb C^n$. We use this characterization to give necessary and sufficient conditions for the boundedness and compactness of Toeplitz operators. Finally, we study the Schatten $p$ classes membership of Toeplitz operators for $1<p<\infty$.	math.FA	math	CARLESON MEASURES AND TOEPLITZ OPERATORS ON SMALL BERGMAN SPACES ON THE BALL VAN AN LE Abstract. We study the Carleson measures and the Toeplitz operators on the class of so-called small weighted Bergman spaces, introduced recently by Seip. A characterization of Carleson measures is obtained which extends Seip's results from the unit disc of C to the unit ball of Cn. We use this characterization to give necessary and sufficient conditions for the boundedness and compactness of Toeplitz operators. Finally, we study the Schatten p classes membership of Toeplitz operators for 1 < p < ∞. 1. Introduction Let Cn denote the n−dimensional complex Euclidean space, Bn = {z ∈ Cn : z < 1} be the unit ball and Sn = {z ∈ Cn : z = 1} be the unit sphere in Cn. Denote by H(Bn) the space of all holomorphic functions on the unit ball Bn. Let dv be the normalized volume measure on Bn. The normalized surface measure on Sn will be denoted by dσ. Let ρ be a positive continuous and integrable function on [0, 1). We extend it to Bn by ρ(z) = ρ(z), and call such ρ a weight function. The weighted Bergman space A2 ρ is the space of functions f in H(Bn) such that ρ =ZBnf (z)2ρ(z)dv(z) < ∞. kfk2 Note that A2 endowed with the inner product ρ is a closed subspace of L2(Bn, ρdv) and hence is a Hilbert space hf, giρ =ZBn f (z)g(z)ρ(z)dv(z), f, g ∈ A2 ρ. When ρ(r) = (1 − r2)α, α > −1, we obtain the standard Bergman spaces A2 α. We impose a normalization condition on ρ: Consider the points rk ∈ [0, 1) determined by the relation 0 Z 1 Z 1 rk x2n−1ρ(x)dx = 1. ρ(x)dx = 2−k. 2010 Mathematics Subject Classification. 30H20; 47B35. Key words and phrases. Bergman spaces, Carleson measures, Toeplitz operators, Schat- ten classes. 1 CARLESON MEASURES AND TOEPLITZ OPERATORS 2 Denote by S the class of weights ρ such that inf k 1 − rk 1 − rk+1 > 1. (1.1) Since the function r 7→RSnf (rξ)2dσ(ξ) is non-decreasing, we also have the equivalent norm kfk2 ρ ≍ 2−kZSnf (rkξ)2dσ(ξ), ∞Xk=1 f ∈ A2 ρ. (1.2) The class S was introduced by Kristian Seip in [13]. It is easy to see that the functions and belong to S. ρ(x) = (1 − x)−β, 1 ρ(x) = (1 − x)−1(cid:18)log 1 − x(cid:19)−α 0 < β < 1, , 1 < α < ∞, Bergman spaces A2 properties of Toeplitz operators on these spaces. In this paper we prove a characterization of Carleson measure for weighted ρ, with ρ ∈ S. This result is then used to study spectral Let µ be a finite positive Borel measure on Bn. We say that µ is a Carleson measure for a Hilbert space X of analytic functions in Bn if there exists a positive constant C such that ZBnf (z)2dµ(z) ≤ Ckfk2 X, f ∈ X. It is clear that µ is a Carleson measure for A2 and the identity operator Id : A2 constant of µ, denoted by Cµ(A2 Suppose that µ is a Carleson measure for A2 Carleson measure for A2 is, ρ ⊂ L2(Bn, dµ) ρ → L2(Bn, dµ) is bounded. The Carleson ρ), is the norm of this identity operator Id. ρ. We say that µ is a vanishing ρ if the above identity operator Id is compact. That ρ if and only if A2 k→∞ZBnfk(z)2dµ(z) = 0 lim whenever {fk} is a bounded sequence in A2 on compact subsets of Bn. ρ which converges to 0 uniformly The concept of a Carleson measure was first introduced by L. Carleson [2, 3] in order to study interpolating sequences and the corona problem on the algebra H ∞ of all bounded analytic functions on the unit disk. It quickly became a powerful tool for the study of function spaces and operators acting on them. The Carleson measures on Bergman spaces were studied by Hast- ings [4], and later on by Luecking [6], and many others. Recently, Pau and Zhao [8] gave a characterization for Carleson measures and vanishing Car- leson measures on the unit ball by using the products of functions in weighted Bergman spaces. In [9], Pel´aez and Rattya gave a description of Carleson measures for A2 r ρ(t)dt is either equivalent to 1 or tends to ∞, and in [10] they then got a criterion for A2 (1 − r)ρ(r)R 1 ρ on unit disk when ρ is such that ρ(s)ds. ρ on unit disk when ρ ∈ bD, which meansR 1 r ρ(s)ds .R 1 r+1 2 1 CARLESON MEASURES AND TOEPLITZ OPERATORS 3 In [13], Seip gave a characterization of Carleson measures for A2 ρ with ρ ∈ S in the case n = 1. One of our main results, Theorem 2.1, extends this result to the case n > 1. ρ with Given a function ϕ ∈ L∞(Bn), the Toeplitz operator Tϕ on A2 symbol ϕ is defined by Tϕf = P (ϕf ), f ∈ A2 ρ, where P : L2(Bn, ρdv) → A2 the integral representation of P , we can write Tϕ as ρ is the orthogonal projection onto A2 ρ. Using Tϕf (z) =ZBn Kρ(z, w)f (w)ϕ(w)ρ(w)dv(w), z ∈ Bn, where Kρ(z, w) is the reproducing kernel for A2 ρ. The Toeplitz operators can also be defined for unbounded symbols or for finite measures on Bn. In fact, given a finite positive Borel measure µ on Bn, the Toeplitz operator Tµ : A2 ρ is defined as follows: ρ → A2 Note that Tµf (z) =ZBn hTµf, giρ =ZBn Kρ(z, w)f (w)dµ(w), z ∈ Bn. f (z)g(z)dµ(z), f, g ∈ A2 ρ. ρ, with ρ ∈ S. The Toeplitz operators acting on various spaces of holomorphic functions have been extensively studied by many authors, and the theory is especially well understood in the case of Hardy spaces or standard Bergman spaces (see [14], [15] and the references therein). Luecking [7] was the first to study Toeplitz operators on Bergman spaces with measures as symbols, and some interesting results about Toeplitz acting on large Bergman spaces were ob- tained by Lin and Rochberg [5]. In this paper, we will study the boundedness and compactness of Tµ on A2 Next we study when our Toeplitz operators belong to the Schatten class. We refer to [15, Chapter 1] for a brief account on the Schatten classes. A description for the standard Bergman spaces on the unit disk was given (see [15, Chapter 7]), and a description for the case of large Bergman spaces on the disk was obtained in 2015 by H. Arroussi, I. Park, and J. Pau [1]. In 2016, Pel´aez and Rattya [11] gave an interesting characterization for the case of small Bergman spaces on unit disk, where the weight ρ ∈ bD. Note that S $ bD, but(cid:8)A2 that ρ∗(r) . ρ(r) for r ∈ (0, 1), where 1 We introduce a subclass S∗ of weights in S determined by the condition ρ : ρ ∈ S(cid:9) =nA2 ρ∗(r) = ρ(t)dt. For example, the weights ρ(x) = (1 − x)−β(cid:18)log 0 < β < 1, α ∈ R ρ : ρ ∈ bDo. 1 − rZ 1 1 − x(cid:19)α 1 , r CARLESON MEASURES AND TOEPLITZ OPERATORS 4 belong to S∗, but the weights ρ(x) = (1 − x)−1(cid:18)log 1 1 − x(cid:19)α 1 − x(cid:19)−1(cid:18)log log 1 ρ(x) = (1 − x)−1(cid:18)log , α < −1, 1 1 − x(cid:19)α , α < −1, do not belong to S∗. For weights ρ in S∗, we obtain a characterization of the symbols of the Toeplitz operators in the Schatten classes Sp. In [12], Pel´aez, Rattya and Sierra gave a characterization for the case of dimension n = 1 when the weight is regular, that is ρ∗(r) ≍ ρ(r). As an easy observation, our result is equivalent to their result when n = 1. We point out that our approach is completely different from that of [12], which does not seem to work in higher dimensions. On the other hand, for regular weights ρ in S \ S∗, this characterization fails. A counterexample was given in [12]. The paper is organized as follows: The main results are stated in Section 2 and their proofs are given in Section 3 -- 5. 2. Main results Throughout this text, we use the following notation. For every nonnega- tive integer k, set Ωk = {z ∈ Bn : rk ≤ z < rk+1} , and let µk be the measure defined by µk = χΩk µ whenever a nonnegative Borel measure µ on Bn is given. The notation U (z) . V (z) (or equivalently V (z) & U (z)) means that there is a positive constant C such that U (z) ≤ CV (z) holds for all z in the set in question, which may be a space of functions or a set of numbers. If both U (z) . V (z) and V (z) . U (z), then we write U (z) ≍ V (z). Our results are following: Theorem 2.1. Let ρ ∈ S, and let µ be a finite positive Borel measure on Bn. Then (i) µ is a Carleson measure for A2 ρ if and only if each µk is a Carleson measure for the Hardy space H 2 with Carleson constant Cµk (H 2) . 2−k, k ≥ 0. ρ if and only if (ii) µ is a vanishing Carleson measure for A2 2kCµk (H 2) = 0. lim k→∞ Theorem 2.1 (i) for the case n = 1 was obtained by Seip in [13]. Theorem 2.2. Let ρ ∈ S, and let µ be a finite positive Borel measure on Bn. Then (i) The Toeplitz operator Tµ is bounded on A2 ρ if and only if µ is a Carleson measure for A2 ρ. (ii) The Toeplitz operator Tµ is compact on A2 ρ if and only if µ is a vanishing Carleson measure for A2 ρ. CARLESON MEASURES AND TOEPLITZ OPERATORS 5 Given z ∈ Bn and 0 < α < 1, we consider the Bergman metric ball where β(z, w) is the Bergman metric given by E(z, α) = {w ∈ Bn : β(z, w) < α} , β(z, w) = 1 2 log 1 + ϕz(w) 1 − ϕz(w) , z, w ∈ Bn. Here, ϕz is the Mobius transformation on Bn that interchanges 0 and z. We know that E(0, α) is actually a Euclidean ball of radius R = tanh α, centered at the origin, and E(z, α) = ϕz(cid:0)E(0, α)(cid:1). more details. Moreover, for fixed α, v(cid:0)E(z, α)(cid:1) ≍ (1 − z)n+1. See [14, Chapter 1] for For a measure µ on Bn and α > 0, we define the function bµα by Denote by fTµ the Berezin transform of Tµ, and set 2kµ(cid:0)E(z, α)(cid:1) bµα(z) = (1 − z)n 2kρ(z)dv(z) z ∈ Ωk. dλρ(z) = , (1 − z)n , z ∈ Ωk. Theorem 2.3. Let ρ be in S∗, µ be a finite positive Borel measure and 1 < p < ∞. The following conditions are equivalent: (a) The Toeplitz operator Tµ is in the Schatten class Sp. (b) The function fTµ is in Lp(Bn, dλρ). (c) The function bµα is in Lp(Bn, dλρ) for sufficiently small α > 0. Given a ∈ Bn \ {0} and r > 0. Let δ(a) =p2(1 − a). Define Q(a, r) ⊂ Bn and O(a, r) ⊂ Sn as follows: 3. Proof of Theorem 2.1 Q(a, r) = {z ∈ Bn :p1 − ha/a, zi < r}, O(a, r) = {ζ ∈ Sn :p1 − ha/a, ζi < r}. For simplicity of notation, we write Qa instead of Q(cid:0)a, δ(a)(cid:1), Oa instead of O(cid:0)a, δ(a)(cid:1). We recall a well known characterization of Carleson measures for the Hardy space (see [14]): A positive Borel measure µ on Bn is a Carleson measure for H 2 if and only if µ(Qa) . (1 − a)n for all a ∈ Bn \ {0}. Furthermore, Cµ(H 2) ≍ supa∈Bn\{0} µ(Qa)(1 − a)−n. We use the following covering lemma from [14, Lemma 4.7]. Lemma 3.1. Suppose N is a natural number, al ∈ Bn \ {0} , 1 ≤ l ≤ N , E = Oal . N[l=1 There exists a subsequence {li}, 1 ≤ i ≤ M , such that (a) Oali , 1 ≤ i ≤ M , are disjoint. CARLESON MEASURES AND TOEPLITZ OPERATORS 6 (b) O(cid:0)ali, 3δ(ali )(cid:1), 1 ≤ i ≤ M , cover E. Lemma 3.2. Let µ be a finite positive measure on Bn. Then µk is a Car- leson measure for H 2 if and only if µk(Qa) . (1 − a)n for all a ∈ Ωk. Furthermore, Cµk (H 2) ≍ supa∈Ωk (1 − a)−nµk(Qa). Proof. Let a ∈ Bn \ {0}. Then a ∈ Ωl for some l ≥ 1. If l > k, then µk(Qa) = 0 and there is nothing to prove. When a ∈ Ωl, l ≤ k, we can cover Qa \ rkBn by a finite family Λ of Qal with al ∈ Ωk−1. Denote by Λ0 the sub-family of Λ obtained by Lemma 3.1. Then µk(Qa) = µk(Qa \ rkBn) ≤ Xl∈Λ0 µk(cid:16)Q(cid:0)al, 3δ(al)(cid:1)(cid:17). Finally, σ(Oal) = σ(cid:16)[l∈Λ0 Since al ∈ Ωk−1, we have µk(cid:16)Q(cid:0)al, 3δ(al)(cid:1)(cid:17) . (1 − al)n ≍ σ(Oal ). Hence µk(Qa) . Xl∈Λ0 σ(cid:16)[l∈Λ0 Oal(cid:17). Oal(cid:17) . σ(Oa) ≍ (1 − a)n. Therefore µk(Qa) . (1 − a)n. This completes the proof. 3.1. Proof of Part (i). (⇐=) Since µk are Carleson measures for H 2 with Carleson constants . 2−k, the same holds for H 2 on the smaller ball rk+2Bn. This means that (cid:3) ZΩkf (z)2dµ(z) . 2−kZSnf (rk+2ξ)2dσ(ξ) for an arbitrary function f in A2 all k ≥ 1 we get ρ and for all k. Summing this estimate over ZBnf (z)2dµ(z) . ∞Xk=1 2−kZSnf (rk+2ξ)2dσ(ξ) ≍ kfk2 ρ. (=⇒) We just need to check that µk(Qa) . 2−k(1 − a)n when a is in Ωk, k ≥ 0. We use the test function with large γ. By (1.2), we have kfak2 ρ ≍ 1 − ha, rjξi2γ dσ(ξ) 1 fa(z) = (1 − ha, zi)−γ 2−jZSn ∞Xj=1 ∞Xj=1 (1 − rja)2γ−n . ρ ≍ 2−k(1 − a)−2γ+n. kfak2 2−j ≍ Since a ∈ Ωk, by Property (1.1) we obtain (3.1) (3.2) CARLESON MEASURES AND TOEPLITZ OPERATORS 7 On the other hand, for every z in Qa we have 1 − ha, zi = (1 − a) + a(1 − ha/a, zi) ≤ (1 − a) + a1 − ha/a, zi < (1 − a) + 2a(1 − a) ≤ 3(1 − a). fa(z) & (1 − a)−γ, z ∈ Qa. (3.3) Hence, Thus, ZBnfa(z)2dµ(z) & (1 − a)−2γµ(Qa ∩ Ωk). Since µ is a Carleson measure for A2 ρ, we get µ(Qa ∩ Ωk) . 2−k(1 − a)n. This implies that µk is a Carleson measure for Hardy space H 2 with Carleson constant Cµk (H 2) . 2−k. 3.2. Proof of Part (ii). Suppose that µ is a vanishing Carleson measure for A2 ρ. Given a in Ωk, consider the function fa defined by (3.1). By (3.2), kfak2 ρ ≍ 2−k(1 − a)−2γ+n. Set ha(z) = (1 − ha, zi)−γ (3.4) ✷ . 2−k/2(1 − a)−γ+n/2 z ∈ Qa. ρ and ha tends to 0 uniformly Then khak2 ρ ≍ 1 and by (3.3), ha(z)2 & 2k (1 − a)n , Since µ is a vanishing Carleson measure for A2 on compact subsets of the unit ball as a → 1, we have lim a→1ZBnha(z)2dµ(z) = 0. 2kµk(Qa ∩ Ωk) (1 − a)n → 0 as k → ∞. Hence, Thus, sup a∈Ωk Conversely, let µr = µBn\rBn (i) of Theorem 2.1 implies that , where rBn = {z ∈ Bn : z < r}. Then part lim k→∞ 2kCµk (H 2) = 0. where ZBnh(z)2dµr(z) ≤ Crkhk2 k:rk>r Cµk (H 2), and ρ, Cr = sup h ∈ A2 ρ, lim r→1 Cr = 0. (3.5) Let {fk} be a bounded sequence in A2 ρ converging uniformly to 0 on compact subsets of Bn. Let ε > 0. By (3.5), there exists r0 ∈ (0, 1) such that Cr < ε for all r ≥ r0. Moreover, by the uniform convergence on compact subsets, CARLESON MEASURES AND TOEPLITZ OPERATORS 8 we may choose k0 ∈ N such that fk(z)2 < ε for all k ≥ k0 and z ∈ r0Bn. It follows that ZBnfk(z)2dµ(z) =Zr0Bnfk(z)2dµ(z) +ZBn\r0Bnfk(z)2dµ(z) < εµ(r0Bn) +ZBnfk(z)2dµr0(z) ≤ εµ(r0Bn) + Cr0kfkk2 ≤ εC, k ≥ k0, ρ for some positive constant C. Hence, µ is a vanishing Carleson measure for A2 ρ. ✷ 4. Proof of Theorem 2.2 4.1. Proof of Part (i). (=⇒) Given a in Ωk, we define ha by (3.4). Then khak2 ρ ≍ 1 and ha(z)2 & 2k(1 − a)−n, z ∈ Qa. Consider the function fTµ(a) = hTµha, haiρ =ZBnha2dµ(z). Since Tµ is bounded, A := supa∈BnfTµ(a) < ∞. Then A ≥ZBnha(z)2dµ(z) ≥ZBnha(z)2dµk(z) ≥ZQaha(z)2dµk(z) & 2k(1 − a)−nµk(Qa). (4.1) (4.2) Hence, µk(Qa) . 2−k(1 − a)n for every a ∈ Ωk. By Theorem 2.1 and Lemma 3.2, µ is Carleson measure for A2 ρ. and ρ, ZBnf (z)2dµ(z) ≤ Ckfk2 ZBng(z)2dµ(z) ≤ Ckgk2 ρ. (⇐=) For every f, g ∈ A2 ρ we have hTµf, giρ =ZBn Then by Cauchy -- Schwarz inequality, we get f (z)g(z)dµ(z). hTµf, giρ ≤ZBnf (z)g(z)dµ(z) ≤(cid:18)ZBnf (z)2dµ(z)(cid:19) 1 2(cid:18)ZBng(z)2dµ(z)(cid:19) 1 2 . Since µ is a Carleson measure for A2 that ρ, there exists a positive constant C such CARLESON MEASURES AND TOEPLITZ OPERATORS Hence, hTµf, giρ ≤ Ckfkρkgkρ Thus, Tµ is bounded on A2 ρ. for all f, g ∈ A2 ρ. 4.2. Proof of part (ii). We need the following auxiliary results. Proposition 4.1. Suppose that f ∈ A2 ρ with ρ ∈ S. Then (1 − z)nkfk2 ρ, z ∈ Ωk, k ≥ 0, f (z)2 ≤ C2k where C is a positive constant independent of k and z. 9 ✷ (4.3) Proof. Let z ∈ Ωk. Applying [14, Corollary 4.5] to the function g(z) = f (rk+2z) at the point , we obtain z rk+2 f (z)2 ≤ZSnf (rk+2ζ)2 (1 − z/rk+22)n By (1.1), 1−hz/rk+2, ζi ≥ 1−hz/rk+2, ζi ≥ 1− zζ for z ∈ Ωk, ζ ∈ Sn. Thus, f (z)2 .ZSnf (rk+2ζ)2 (1 − z2)n rk+2 1 − hz/rk+2, ζi2n dσ(ζ). = 1−z/rk+2 & 1−z (1 + z)n (1 − z)2n dσ(ζ) (1 − z)nZSnf (rk+2ζ)2dσ(ζ) (1 − z)n 2−kZSnf (rk+2ζ)2dσ(ζ) 2−jZSnf (rj+2ζ)2dσ(ζ) 2k 2k ∞Xj=1 (1 − z)n (1 − z)nkfk2 2k ρ, ≤ . ≤ . (cid:3) with constants independent of k and z. Corollary 4.2. A sequence of functions {fk} ⊂ A2 ρ if and only if it is bounded in A2 A2 compact subset of Bn. Proof of part (ii) of Theorem 2.2. Suppose that Tµ is compact on A2 ρ converges to 0 weakly in ρ and converges to 0 uniformly on each ρ. We ρ ≍ 1 and ha converges uniformly to 0 on compact subsets of Bn as a → 1. Since Tµ is compact, define ha, a ∈ Bn by (3.4) andfTµ by (4.1). Then khak2 fTµ(a) → 0 as a → 1. By (4.2) this implies that 2kµk(Qa) (1 − a)n → 0 as k → ∞. sup a∈Ωk Hence, lim k→∞ 2kCµk (H 2) = 0. By part (ii) of Theorem 2.1, µ is a vanishing Carleson measure for A2 ρ. CARLESON MEASURES AND TOEPLITZ OPERATORS 10 Conversely, assume that µ is a vanishing Carleson measure for A2 ρ. For every h ∈ A2 ρ we have kTµhkρ = sup hTµh, giρ. g∈A2 ρ kgkρ≤1 Furthermore, hTµh, giρ =(cid:12)(cid:12)(cid:12)(cid:12)ZBn h(z)g(z)dµ(z)(cid:12)(cid:12)(cid:12)(cid:12) ≤ZBnh(z)g(z)dµ(z) ≤(cid:18)ZBnh(z)2dµ(z)(cid:19)1/2(cid:18)ZBng(z)2dµ(z)(cid:19)1/2 .(cid:18)ZBnh(z)2dµ(z)(cid:19)1/2 kgkρ. The last inequality follows from the fact that µ is a Carleson measure for A2 ρ. Therefore, kTµhkρ .(cid:18)ZBnh(z)2dµ(z)(cid:19)1/2 , h ∈ A2 ρ. Now, let {fk} ⊂ A2 subsets of Bn. Since µ is a vanishing Carleson measure for A2 ρ, ρ be bounded and converge uniformly to 0 on compact k→∞ZBnfk(z)2dµ(z) = 0. lim It follows that kTµfkkρ → 0 and hence Tµ is compact. 5. Proof of Theorem 2.3 Proposition 5.1. Let Kρ(z, w) be the reproducing kernel of A2 ρ. (a) Let k ≥ 1, z ∈ Ωk. Then Kρ(z, z) ≍ 2k (1 − z)n . (b) There exists α = α(ρ) > 0 such that for every z ∈ Bn, Kρ(z, w)2 ≍ Kρ(z, z)Kρ(w, w) whenever w ∈ E(z, α). (cid:3) (5.1) (5.2) Proof. (a) Fix k ≥ 1. Given z ∈ Ωk, let Lz be the point evaluation at z on A2 ρ. It is well known that By Proposition 4.1, Kρ(z, z) = kLzk2. Furthermore, choosing hz by (3.4), we have khzkρ ≍ 1 and kLzk2 . 2k (1 − z)n . hz(z)2 & 2k (1 − z)n . CARLESON MEASURES AND TOEPLITZ OPERATORS 11 kLzk2 & 2k (1 − z)n . Hence, Thus Kρ(z, z) ≍ (b) It is well known that 2k (1 − z)n , z ∈ Ωk. for all z, w ∈ Bn. For any fixed z0 ∈ Ωk, consider the subspace A2 as ρ(z0) defined Denote by Lz0 the one-dimensional subspace spanned by the function A2 Kρ(z, w)2 ≤ Kρ(z, z)Kρ(w, w) ρ : f (z0) = 0(cid:9) . ρ(z0) =(cid:8)f ∈ A2 pKρ(z0, z0) kρ,z0(z) = Kρ(z, z0) . Then we have the orthogonal decomposition A2 ρ = A2 ρ(z0) ⊕ Lz0. Hence Kρ(z, w) = Kρ,z0(z, w) + kρ,z0(w)kρ,z0(z), where Kρ,z0 is the repro- ducing kernel of A2 ρ(z0). Therefore, Kρ(z0, w) = kρ,z0(w)kρ,z0(z0) and Kρ(w, w) = Kρ,z0(w, w) + kρ,z0(w)2. We are going to prove that there exist α > 0 such that Kρ,z0(w, w) < 1 2 Kρ(w, w), w ∈ E(z0, α). (5.3) (5.4) By (1.1), there exists α1 > 0 such that E(z0, α) ⊂ Ωk−1 ∪ Ωk ∪ Ωk+1, 0 < α < α1. Hence, for every f ∈ A2 ρ(z0) such that kfkρ = 1, by Proposition 4.1 we have f (w)2 . 2k (1 − w)n ≍ 2k (1 − z0)n as whenever w ∈ E(z0, α). Since E(z0, α) = ϕz0(cid:0)E(0, α)(cid:1), we can rewrite (5.5) whenever η ∈ E(0, α). Note that f (z0) = f(cid:0)ϕz0(0)(cid:1) = 0. Therefore, by the f(cid:0)ϕz0(η)(cid:1)2 . Schwarz lemma, we get (1 − z0)n (5.6) 2k (5.5) whenever η ∈ E(0, α). This implies that there is a constant C > 0 such that f(cid:0)ϕz0(η)(cid:1)2 . η2 f(cid:0)ϕz0(η)(cid:1)2 ≤ Cη2 2k (1 − z0)n ≍ η2 2k (1 − ϕz0 (η))n 2k (1 − ϕz0(η))n , η ∈ E(0, α). CARLESON MEASURES AND TOEPLITZ OPERATORS 12 Therefore, we can choose α so small that f(cid:0)ϕz0(η)(cid:1)2 < This proves (5.4). 1 2 Kρ(cid:0)ϕz0(η), ϕz0 (η)(cid:1), η ∈ E(0, α). Now, from (5.3) and (5.4), we obtain that kρ,z0(w)2 > ever w ∈ E(z0, α). This means that 1 Kρ(w, z0)2 > 2 Kρ(z0, z0)Kρ(w, w) whenever w ∈ E(z0, α), which completes the proof. Lemma 5.2. Let T be a positive operator on A2 transform of T , defined by 1 2 Kρ(w, w) when- (cid:3) ρ, and let eT be the Berezin eT (z) = hT kz, kziρ, z ∈ Bn. if z ∈ Ωk. Here, dλρ(z) = 2kρ(z)dv(z) (1 − z)n (a) Let 0 < p ≤ 1. If eT ∈ Lp(Bn, dλρ), then T is in Sp. (b) Let p ≥ 1. If T is in Sp, then eT ∈ Lp(Bn, dλρ). Proof. Note that dλρ(z) ≍ K(z, z)ρ(z)dv(z) = kKzk2ρ(z)dv(z). The proof is similar to the proof of [1, Lemma 4.2]. The positive operator T is in Sp if and only if T p is in the trace class S1. Fix an orthonormal basis ρ. Since T p is positive, it is in S1 if and only ifPkhT pek, ekiρ < ∞. {ek} of A2 Let U = √T p. By Fubini's theorem, the reproducing property of Kz, and Parseval's identity, we have Xk hT pek, ekiρ =Xk kU ekk2 ρ =Xk ZBnU ek(z)2ρ(z)dv(z) U ek(z)2! ρ(z)dv(z) hU ek, Kziρ2! ρ(z)dv(z) hek, U Kziρ2! ρ(z)dv(z) =ZBn Xk =ZBn Xk =ZBn Xk =ZBnkU Kzk2 =ZBnhT pKz, Kziρρ(z)dv(z) =ZBnhT pkz, kziρkKzk2 ≍ZBnhT pkz, kziρdλρ(z). ρρ(z)dv(z) ρρ(z)dv(z) CARLESON MEASURES AND TOEPLITZ OPERATORS 13 Hence, both (a) and (b) are the consequences of the well known inequalities (see [15, Proposition 1.31]) hT pkz, kziρ ≤ hT kz, kzip hT pkz, kziρ ≥ hT kz, kzip ρ =(cid:0)eT (z)(cid:1)p, ρ =(cid:0)eT (z)(cid:1)p, (cid:3) Lemma 5.3. Let ρ ∈ S∗ and z ∈ Ωk. Then there exists α0 > 0 such that for every α ∈ (0, α0) we have 0 < p ≤ 1, p ≥ 1. f (z)2 . 2k (1 − z)nZE(z,α)f (w)2ρ(w)dv(w) for all f ∈ H(Bn). Proof. Let z ∈ Ωk. For each f ∈ H(Bn), by the subharmonicity of the function w 7→ f (w)2 and the estimate v(cid:0)E(z, α)(cid:1) ≍ (1 − z)n+1, we have 1 It is easy to see that 1 − z ≍ 1 − w for w ∈ E(z, α). Hence, dv(w) 1 f (z)2 . f (z)2 . = (1 − z)n+1ZE(z,α)f (w)2dv(w). (1 − z)nZE(z,α)f (w)2 (1 − z)nZE(z,α)f (w)2 2−k 1 − w 1 − w 2k 1 dv(w). (5.7) By (1.1), for small α0 we have E(z, α0) ⊂ Ωk−1 ∪ Ωk ∪ Ωk+1. Therefore, for every α ∈ (0, α0), we have rk−1 < w < rk+2 for w ∈ E(z, α). Since w ρ(t)dt for every w ∈ E(z, α), α ∈ rk+2 (0, α0). Plugging this into (5.7) and using that ρ∗(w) . ρ(w), we get ρ(t)dt = 2−k−2, we obtain 2−k .R 1 R 1 f (z)2 . . 2k (1 − z)nZE(z,α)f (w)2ρ∗(w)dv(w) (1 − z)nZE(z,α)f (w)2ρ(w)dv(w). 2k fTµ(z) =ZBnkz(w)2dµ(w) = kKzk−2 ρ ZE(z,α)Kz(w)2dµ(w) ρdµ(w) ≍bµα(z). ≥ kKzk−2 ≍ZE(z,α)kKwk2 Since fTµ is in Lp(Bn, dλρ), bµα is also in Lp(Bn, dλρ). This completes the proof. (cid:3) Proof of Theorem 2.3. (a) ⇒ (b). This follows from Lemma 5.2 (b). (b) ⇒ (c). By Proposition 5.1 (b), for sufficiently small α > 0, we have ρkKwk2 ρ, Then by Proposition 5.1 (a), we get Kz(w)2 ≍ kKzk2 w ∈ E(z, α), z ∈ Bn. ρ ZBnKz(w)2dµ(w) where 1 p + 1 q = 1. Thus, (5.8) implies that Xl hTµel, elip , ≤(cid:18)ZBnel(w)2bµα(w)pρ(w)dv(w)(cid:19)1/p ×(cid:18)ZBnel(w)2ρ(w)dv(w)(cid:19)1/q =(cid:18)ZBnel(w)2bµα(w)pρ(w)dv(w)(cid:19)1/p ρ .ZBn Xl el(w)2!bµα(w)pρ(w)dv(w) =ZBnkKwk2 ρbµα(w)pρ(w)dv(w) ≍ZBnbµα(w)pdλρ(w) < ∞. CARLESON MEASURES AND TOEPLITZ OPERATORS 14 (c) ⇒ (a). For every orthonormal basis {el} of A2 ρ, we have By Lemma 5.3, el(z)2 . 2k ρ =Xl (cid:18)ZBnel(z)2dµ(z)(cid:19)p Xl hTµel, elip (1 − z)nZE(z,α)el(w)2ρ(w)dv(w), ZBnel(z)2dµ(z) .ZBnel(w)2bµα(w)ρ(w)dv(w) By Fubini's theorem and Holder's inequality, we have . (5.8) z ∈ Ωk. This proves (a). (cid:3) Remark 5.4. Let 1 < p < ∞. spaces, Arroussi, Park and Pau proved in [1, Theorem 4.6] that In the case of large weighted Bergman µ(cid:0)B(z, ε)(cid:1) (1 − z)2n is in the corresponding weighted Lp, Tµ ∈ Sp ⇐⇒eµε(z) = where B(z, ε) is the Euclidean ball with center z and radius ε(1−z). When the dimension n = 1, we can see that eµε is in Lp if and only if bµε is in Lp. However, for n > 1, this equivalence is not true anymore. Let us verify this. Choose zk ∈ Bn such that zk tend to 1 sufficiently rapidly as k → ∞. Consider µ = ckχB(zk,ε) and µ∗ = ∞Xk=1 ∞Xk=1 ckχB(zk,3ε), where ck > 0 will be chosen later. We have µ .eµε . µ∗ CARLESON MEASURES AND TOEPLITZ OPERATORS 15 and Hence ∞Xk=1 ck ck ∞Xk=1 v(cid:0)B(zk, ε)(cid:1) v(cid:0)E(zk, ε)(cid:1) χE(zk,3ε). cp v(cid:0)B(zk, ε)(cid:1) v(cid:0)E(zk, ε)(cid:1) χE(zk,ε) .bµε . eµε ∈ Lp ⇐⇒ bµε ∈ Lp ⇐⇒ kv(cid:0)B(zk, ε)(cid:1) cp ∞Xk=1 ∞Xk=1 k(cid:0)v(cid:0)B(zk, ε)(cid:1)(cid:1)p(cid:0)v(cid:0)E(zk, ε)(cid:1)(cid:1)1−p kv(cid:0)B(zk, ε)(cid:1) < ∞, k (cid:0)v(cid:0)B(zk, ε)(cid:1)(cid:1)p (cid:0)v(cid:0)E(zk, ε)(cid:1)(cid:1)p−1 < ∞. v(cid:0)E(zk, ε)(cid:1)!p−1 = v(cid:0)B(zk, ε)(cid:1) as k → ∞, we can choose ck such that bµε ∈ Lp but eµε /∈ Lp. On the other hand, one can easily see that eµε ∈ Lp implies bµε ∈ Lp. Acknowledgments. I am deeply grateful to my advisors Professors Alexander Borichev and El Hassan Youssfi for their help and many sug- gestions during the preparation of this paper. ≍ (1 − zk)(n−1)(p−1) −→ 0 and Since cp cp References [1] H. Arroussi, I. Park and J. Pau, Schatten class Toeplitz operators acting on large weighted Bergman spaces, Studia Mathematica 229 (2015), no. 3 , 203 -- 221. [2] L. Carleson, An interpolation problem for bounded analytic functions, Amer. J. Math. 80 (1958), 921 -- 930. [3] L. Carleson, Interpolations by bounded analytic functions and the corona problem, Ann. of Math. 76 (1962), 547 -- 559. [4] W. Hastings, A Carleson measure theorem for Bergman spaces, Proceedings of the American Mathematical Society 52 (1975), 237 -- 241. [5] P. Lin and R. Rochberg, Trace ideal criteria for Toeplitz and Hankel operators on the weighted Bergman spaces with exponential type weights, Pacific Journal of Mathematics 173 (1996), no. 1, 127 -- 146. [6] D. Luecking, A technique for characterizing Carleson measures on Bergman spaces, Proceedings of the American Mathematical Society 87 (1983), 656 -- 660. [7] D. Luecking, Trace ideal criteria for Toeplitz operators, Journal of Functional Analysis 73 (1987), 345 -- 368. [8] J. Pau and R. Zhao, Carleson measures and Teoplitz operators for weighted Bergman spaces on the unit ball, Michigan Math. J. 64 (2015), no. 4, 759 -- 796. [9] J. A. Pel´aez and J. Rattya, Weighted Bergman spaces induced by rapidly increasing weights, Mem. Amer. Math. Soc. 227 (2014), no. 1066, 124 pp. [10] J. A. Pel´aez and J. Rattya, Embedding theorems for Bergman spaces via harmonic analysis, Mathematische Annalen 362 (2015), no. 1, 205 -- 239. [11] J. A. Pel´aez and J. Rattya, Trace class criteria for Toeplitz and composition operators on small Bergman spaces, Advances in Mathematics 293 (2016), 606-643. [12] J. A. Pel´aez, J. Rattya and K. Sierra, Berezin Transform and Toeplitz Operators on Bergman Spaces Induced by Regular Weights, The Journal of Geometric Analysis 28 (2018), 656-687. CARLESON MEASURES AND TOEPLITZ OPERATORS 16 [13] K. Seip, Interpolation and sampling in small Bergman spaces, Collectanea Mathe- matica 64 (2013), no. 1, 61 -- 72. [14] K. Zhu, Spaces of Holomorphic Functions in the Unit Ball, Graduate Texts in Math- ematics, vol. 226, Springer-Verlag, New York, 2005. [15] K. Zhu, Operator Theory in Function Spaces, 2nd ed., Mathematical surveys and monographs, vol. 138, American Mathematical Society, Providence, RI, 2007. Aix -- Marseille University, CNRS, Centrale Marseille, I2M, Marseille, France University of Quynhon, Department of Mathematics, 170 An Duong Vuong, Quy Nhon, Vietnam E-mail address: [email protected]
1202.3963	1	1202	2012-02-17T16:55:14	On the numerical radius of the truncated adjoint Shift	[ "math.FA" ]	A celebrated thorem of Fejer (1915) asserts that for a given positive trigonometric polynomial $\sum_{j=-n+1}^{n-1}c_{j}e^{ijt}$, we have $\lvert c_{1}\lvert\leqslant c_{0}\cos\frac{\pi}{n+1}$. A more recent inequality due to U. Haagerup and P. de la Harpe asserts that, for any contraction $T$ such that $T^{n}=0$, for some $n\geq2$, the inequality $\omega_{2}(T)\leqslant\cos\frac{\pi}{n+1}$ holds, and $\omega_{2}(T)=\cos\frac{\pi}{n+1}$ when T is unitarily equivalent to the extremal operator ${S}^{\ast}_{n}={\bbs}_{\lvert{\C}^{n}}={\bbs}_{\lvert Ker (u_{n}(\bbs))}$ where $u_{n}(z)=z^{n}$ and $\bbs$ is the adjoint of the shift operator on the Hilbert space of all square summable sequences. Apparently there is no relationship between them. In this mathematical note, we show that there is a connection between Taylor coefficients of positive rational functions on the torus and numerical radius of the extremal operator $\bbs(\phi)=\bbs_{\lvert Ker(\phi(\bbs))}$ for a precise inner function $\phi$. This result completes a line of investigation begun in 2002 by C. Badea and G. Cassier \cite{Cassier}. An upper and lower bound of the numerical radius of $\bbs(\phi)$ are given where $\phi$ is a finite Blashke product with unique zero.	math.FA	math	ON THE NUMERICAL RADIUS OF THE TRUNCATED ADJOINT SHIFT HAYKEL GAAYA j=−n+1 cj eijt, we have c16 c0 cos π n+1 holds, and ω2(T ) = cos π CI n = S∗ Abstract. A celebrated thorem of Fejer (1915) asserts that for a given pos- itive trigonometric polynomial Pn−1 n+1 . A more recent inequality due to U. Haagerup and P. de la Harpe [9] asserts that, for any contraction T such that T n = 0, for some n ≥ 2, the inequality ω2(T ) 6 cos π n+1 when T is unitarily equivalent to the extremal operator S∗ Ker(un(S ∗)) where un(z) = zn and S∗ is the adjoint of the shift operator on the Hilbert space of all square summable sequences. Apparently there is no relationship between them. In this mathematical note, we show that there is a connection between Taylor coefficients of positive rational functions on the torus and numerical radius of the extremal operator S∗(φ) = S∗ Ker(φ(S ∗)) for a precise inner function φ. This result completes a line of investigation begun in 2002 by C. Badea and G. Cassier [1]. An upper and lower bound of the numerical radius of S∗(φ) are given where φ is a finite Blashke product with unique zero. n = S∗ 1. Introduction Let H be a complex separable Hilbert space and B(H) the collection of all bounded linear operators on H. The numerical range of an operators T in B(H) is the subset W (T ) = {< T x, x >∈ CI ; x ∈ H,kxk= 1} of the plane, where < ., . > denotes the inner product in H and the numerical radius of T is defined by Re(T ) is the self-adjoint operator defined by ω2(T ) = sup{z; z ∈ W (T )} . Re(T ) = 1 2 (T + T ∗). We denote by [x] the integer part of x and by S the unilateral shift acting on the Hardy space HI 2 of the square summable analytic functions and by S∗ its adjoint: S : HI 2 → HI 2 7→ zf (z) f S∗ : HI 2 → 7→ f HI 2 f (z) − f (0) z . 2000 Mathematics Subject Classification. 47A12, 47B35. Key words and phrases. Numerical radius, Numerical range, Truncated shift, Eigenvalues, Toeplitz forms, Inequalities for positive trigonometric polynomials. 1 2 HAYKEL GAAYA Beurling's theorem implies that the non zero invariant subspaces of S are of the forme φ HI 2, where φ is some inner function . Let S(φ) denote the compression of S to the space H(φ) = HI 2 ⊖ φ HI 2 : S(φ)f (z) = P (zf (z)), where P denotes the ortogonal projection from HI 2 onto H(φ). We denote by S∗(φ) the adjoint of S(φ): S∗(φ) = S(φ)∗ = S∗ H(φ) = S∗ Ker(φ(S)∗) . The model operator S(φ) has many properties and it was studied intensively in the 1960s and '70s. For exemple, it has norm 1 (for dim H(φ) > 1) and it is cyclic. The function φ is the minimal function of S(φ) meaning that φ(S(φ)) = 0 and φ divides any function ψ in H ∞ with ψ(S(φ)) = 0. The space H(φ) is finite- dimensional exactly when φ is a finite Blaschke product : φ(z) = Yj=1 In this case the polynomial p(z) = Qn j=1(z − αj) is both the minimal and char- acteristic polynomial of S(φ) and (αj )16j6n are its eigenvalues. In particular, if φ(z) = zn then S(φ) is unitarily equivalent to Sn where . n z − αj 1 − αjz 0 1 . . . . . . . . . 1 0 .   Sn =   n+1o and In 1992 U. Haagerup and P. de la Harpe proved that W (Sn) is the disc Dn =nz ∈ CI ;z6 cos π (1.1) ω2(Sn) = cos π n + 1 and more generally a natural connection between Fejer's inequality and the numer- ical radius of a nilpotent matrix was established by Haagerup and de la Harpe. They proved, using solely elementary methods (positive definite kernels) that: Theorem 1.1 ([9]). Let T be an operator on H such that T n = 0 for some n ≥ 2. One has: ω2(T ) 6 kTkcos n + 1 π and ω2(T ) = kTkcos π n+1 when T is unitarily equivalent to kTkSn. The reader may consult [10] chapter 22 for properties of numerical ranges of operators in general, [11] chapter 1 for those of finite dimensional operators and particulary [6] for the geometric properties of the numerical range of S(φ). Organisation of the paper: In [1], C. Badea and G. Cassier showed: ON THE NUMERICAL RADIUS OF THE TRUNCATED ADJOINT SHIFT 3 Theorem 1.2. ([1]) Let F = P/Q be a rational function with no principal part (d◦P < d◦Q) which is positive on the torus. Then the Taylor coefficient ck of order k satisfies the following inequalitiy where R = S∗ Ker(Q(S ∗)). ck6 c0 ω2(Rk), In Section 2, our main theorem is the Theorem 1.2. We give an extension of the result of C. Badea and G. Cassier for Taylor coefficients of all rational functions which are positive on the torus. We make no extra assumptions about P and Q. We do not, for exemple require them to obey any degree restrictions, they need only be coprime. The theorem has many applications and will explain how we can easily recover the remarkable Egerváry and Szász inequality. See Corollary 2.3. Toeplitz matrices are found in several areas of mathematics and its applications such as complex and harmonic analysis. The KMS Toeplitz matrix Kn(α) = (αr−s)n r,s=1 associated with the Poisson kernel introduced by Kac, Murdokh and Szegö [12] is of particular interst in these areas. In order to formulate our problem we first review in Section 3.1 some of the known results on the spectra of these matrices and we give a better upper bound of λ(n) is the largest eigenvalue of the KMS matrix. In Section 3.2, our main theorem is the Theorem 3.7 which gives un upper and lower bound of the numerical radius of the truncated shift S∗(φ) where 1 , where λ(n) 1 1 − αz(cid:19)n φ(z) = (cid:18) z − α with 0 6 α < 1 is a finite Blashke product with unique zero. Our preoccupation with this particular case 0 6 α < 1 to the exclusion to any α in the disc is explained by the fact that the numerical radius is independent with the argument of α. This formula is expressed in terms of eigenvalues of the KMS matrices and provides an easy proof for the Haagerup and de la Harpe result (1.1). 2. Main Theorem There are many classical inequalities for coefficients of positive trigonometric polynomials. The next result shows the links between the numerical radius of the truncated adjoint shift and the Taylor coefficients of rational functions positive on the torus. Theorem 2.1. Let F = P/Q be a rational function which is positive on the torus, where P and Q are coprime. Denote by and p q φ(z) = 1 − αj z(cid:19)mj 1 − βjz(cid:19)dj Yj=1(cid:18) z − αj Yj=1(cid:18) z − βj j=1 mj and d = Pq ψ(z) = the respectively finite Blashke products formed by the nonzero roots of P and Q in j=1 dj . Then the Taylor coefficient ck the open disc, let m = Pp of order k of F satisfies the following inequality: ck6 c0 ω2(S∗k(ϕ)), where ϕ(z) = zmax(0,m−d+1)ψ(z). 4 HAYKEL GAAYA Lemma 2.2 ([1] lemma 3.2). Let u be a inner function and let f be a positive function in the subspace u HI 1 0 of L1(TI ). Then there exists a function h in H(u) = HI 2 ⊖ u HI 2 such that f = h2. Proof. First, note that by continuity we may assume that F is strictly positive on the torus. Let F = P/Q and assume that F (z) > 0 for every z ∈ TI . Now, let z(cid:19). G (z) = F(cid:18) 1 We see that G is analytic, except a finite set of complex numbers. Since F is real on the torus, we have G(eit) = F (eit) = F (eit) for every t ∈ RI and the analytic extension principle implies that F (z) = P (z) Q (z) = G(z) = , P (cid:0) 1 z(cid:1)Q(cid:0) 1 z(cid:1) z(cid:19)Q (z) P (z) Q(cid:18) 1 z(cid:19) = P(cid:18) 1 α(cid:1) = 0. Then P can be written as necessarily P(cid:0) 1 except for a finite set in CI . (2.1) implies that if P (α) = 0, with α 6= 0 then P (z) = c1 zm0 (z − α1)m1 ... (z − αp)mp (1 − α1z)m1 ... (1 − αpz)mp with a constant c1. With the same argument as before, Q can be written as Q (z) = c2 zd0 (z − β1)d1 ... (z − βq)dq (cid:0)1 − β1z(cid:1)d1 ...(cid:0)1 − βqz(cid:1)dq with a constant c2. Since P and Q are coprime, we must have m0 = 0 or d0 = 0. thus (2.1) Then where P1 (z) =Qp P1(cid:0)eit(cid:1) Q1 (eit)2 j=1 (z − βj)dj therefore we have F(cid:0)eit(cid:1) = F(cid:0)eit(cid:1) = c j=1 (z − αj)mj and Q1 (z) =Qq j=1(cid:0)eit − αj(cid:1)mj(cid:0)e−it − αj(cid:1)mj F (cid:0)eit(cid:1) = c Qp j=1 (eit − βj)dj(cid:0)e−it − βj(cid:1)dj Qq with a constant c. Let m = m1 + ··· + mp, d = d1 + ··· + dq and ϕ(z) = zrψ(z) where r = max(0, m − d + 1). Now, q = c eirt j=1(cid:0)eit − αj(cid:1)mj(cid:0)e−it − αj(cid:1)mj ϕ(cid:0)eit(cid:1) F (cid:0)eit(cid:1) = c eirtψ(cid:0)eit(cid:1)Qp j=1 (eit − βj)dj(cid:0)e−it − βj(cid:1)dj Qq j=1(cid:0)eit − αj(cid:1)mj(cid:0)1 − αjeit(cid:1)mj e−imt 1 − βjeit(cid:19)dj Qp j=1 (eit − βj)dj(cid:0)1 − βjeit(cid:1)dj e−idt Qq j=1(cid:0)eit − αj(cid:1)mj(cid:0)1 − αjeit(cid:1)mj j=1(cid:0)1 − βjeit(cid:1)2dj Qq j=1(cid:0)eit − αj(cid:1)mj(cid:0)1 − αjeit(cid:1)mj j=1(cid:0)1 − βjeit(cid:1)2dj Qq Yj=1(cid:18) eit − βj = c ei(d−m)teirt Qp = c eit max(d−m,1) Qp ON THE NUMERICAL RADIUS OF THE TRUNCATED ADJOINT SHIFT 5 0. It follows from Lemma 2.2 that we have F = f2 which implies that ϕF ∈ HI 1 with a suitable f ∈ H(ϕ). Then for any integer k, we get ck = < F, eikt > = < f f , eikt > = < f e−ikt, f > = < (S∗(ϕ))kf, f > = < S∗k(ϕ)f, f > . Therefore ck ≤ kfk2 2 ω2(S∗k(ϕ)) = kFk1 ω2(S∗k(ϕ)) = c0 ω2(S∗k(ϕ)). The proof is now complete. (cid:3) Corollary 2.3 (Egerváry and Szász [4]). Let F (eit) =Pn−1 trigonometric polynomial (n ≥ 2). Then k (cid:3) + 2! f or 1 6 k 6 n − 1. (cid:2) n−1 ck6 c0 cos π j=−n+1 cjeijt be a positive Proof. We have : F (eit) = c−n+1ei(−n+1)t + ··· + c0 + ··· + cn−1ei(n−1)t = e(−n+1)it(cid:16)c−n+1 + ··· + c0ei(n−1)t + ··· + cn−1e2i(n−1)t(cid:17) = P (eit) Q(eit) where P (eit) = c−n+1 + ··· + c0ei(n−1)t + ··· + cn−1e2i(n−1)t and Q(eit) = ei(n−1)t. In this case we have m = n − 1, d = 0 and ϕ(z) = zn. Then Theorem 2.1 implies that ck ≤ c0 ω2(S∗ n k). But generally S∗ n dimension, the largest dimension being s(k, n) + 1 where s(k, n) = [ n−1 ω2(S∗ n k is unitarily equivalent to an orthogonal sum of shifts of smaller k ]. Therefore . The same computation follows from [8]. k) = ω2(S∗ s(k,n)+1) = cos π s(k, n) + 2 Finally, this implies that π s(k, n) + 2 ck 6 c0 cos = c0 cos π k (cid:3) + 2! . (cid:2) n−1 (cid:3) Remark 2.1. The bound for c1 is due to Fejer (1915). 6 HAYKEL GAAYA 3. The numerical radius of the shift compression 3.1. Preliminaries. The spectral decomposition of the KMS matrix Kn(α) =  1 α ... αn−1 α . . . . . . ··· ··· αn−1 ... . . . . . . α α 1   = (αr−s)n r,s=1 ; 0 6 α < 1 is very well understood in the computational sense. For this reason, these matrices are often used as test matrices. It's shown in [7] page 69 -- 72 that Kn(α) is a Toeplitz matrix associated with the Poisson kernel Pα(eit) = 1 − α2 1 − αeit2 and its eigenvalues are: where t(n) k are the solutions of λ(n) k = Pα(eit (n) k ) , 1 ≤ k ≤ n (3.1) pn(cos t) = sin(n + 1)t − 2α sin nt + α2 sin(n − 1)t sin t = 0. The expression pn(cos t) is a polynomial of degree n in cos t and it has n real distinct zeros cos t(n) k for 1 6 k 6 n where : 1 < t(n) 0 < t(n) 2 < t(n) 3 < ··· < t(n) n < π . This implies that 1 + α 1 − α > λ(n) 1 > λ(n) 2 > λ(n) 3 > ··· > λ(n) n > 1 − α 1 + α . The evaluation of the zeros t(n) k it is easy to show that they are separated by the quantities xk = in explicit terms seems to be out of end. However, , 1 ≤ k ≤ n. n + 1 kπ Indeed, for 1 6 k 6 n and pn(cos xk) = (−1)k2α(1 − α cos xk) Also we see by direct substitution that the latter equation holds for k = 0, so that sgn pn(cos xk) = (−1)k . 0 < t(n) 1 6 x1 < t(n) 2 6 x2 < ··· < t(n) n 6 xn < π . Remark 3.1. In the case where α = 0 in (3.1) we have t(n) k = xk. In the next proposition we give a better lower and upper bound for tn 1 . Proposition 3.1. For each integer n ≥ 2; (3.2) arccos(α) 6 t(n) 1 6 arccos(α) . 2 n + 1 Proof. First, we note that (3.3) pn(t) = 2 Since for all 0 < t 6 π sin t(cid:16)sin (n+1)t n+1 , we have 2 − α sin (n−1)t 2 (cid:17)(cid:16)cos (n+1)t 2 (cid:17) . 2 − α cos (n−1)t (3.4) 0 < (n−1)t 2 < (n+1)t 2 6 π 2 ON THE NUMERICAL RADIUS OF THE TRUNCATED ADJOINT SHIFT 7 this implies that then t(n) 1 is zero of (3.5) α sin (n−1)t 2 < sin (n−1)t 2 < sin (n+1)t 2 cos (n+1)t . 2 2 = α cos (n−1)t n+1 arccos(α) then (n+1)t which contradicts(3.5). 2 (n) 1 < arccos(α) and Now if we suppose that t(n) cos (n+1)t(n) Hence t(n) > α ≥ α cos (n−1)t(n) n+1 arccos(α) holds. 2 2 1 ≥ 2 1 1 1 < 2 From (3.5), we have cos t cos (n−1)t 2 − sin t sin (n−1)t 2 = α cos (n−1)t 2 which implies that while from (3.4), sin (n−1)t positive, which completes the proof. 2 (cos t − α) cos (n−1)t and cos (n−1)t 2 2 = sin t sin (n−1)t 2 are both positive, therefore cos t − α is (cid:3) Remark 3.2. For a fixed 0 6 α < 1, since Pα(eit) is positive and decreasing on the interval [0, π], then it is easy to obtain the sharp lower and upper bound of the largest eigenvalues of Kn(α): 1 6 λ(n) 1 6 1 − 2α cos(cid:16) 2 1 − α2 n+1 arccos(α)(cid:17) + α2 . Note that λ(n) is also the numerical radius of Kn(α). This is due to the fact that the norm and numerical radius of a symmetric matrix coincides with its largest eigenvalue. 1 For 0 6 α < 1, we denote by 1 − α)t 1 + α2 ; −α(cos t(n) 1 − 2α cos t(n) n + α2) ; n − α) 1 − 2α 1 + α2 1 + 2α3 1 + α2 , −(1 + α2) cos t(n) 1 − 2α cos t(n) n + 2α n + α2 ) , −(1 + α2) cos t(n) 1 − 2α cos t(n) n + 2α n + α2 ) . and sn(α) = max( α(cos t(n) 1 − 2α cos t(n) mn(α) = max((1 + α2) cos t(n) 1 − 2α cos t(n) Mn(α) = max( (1 − 3α2) cos t(n) 1 − 2α cos t(n) Jn(α) =  0 ... ... 0 α . . . . . . Proposition 3.2. For each integer n ≥ 2 and 0 6 α < 1, let ··· αn−1 ... . . . . . . . . . α 0 ;   then we have ω2(Jn(α)) = ω2(Re(Jn(α))) = sn(α). 8 HAYKEL GAAYA Proof. First observe that ω2(Re(Jn(α))) = = = = = sup ,kuk=1< Re(Jn(α))u, u > αl−mulum(cid:12)(cid:12)(cid:12) αl−mulum αl−mulum 1 2 1 2 u=(u0,··· ,un−1)∈ CI n 1 2 sup sup Pn−1 Pn−1 l=0 ul2=1(cid:12)(cid:12)(cid:12) X1≤m6=l≤n−1 l=0 ul2=1 X1≤m6=l≤n−1 2 X1≤m<l≤n−1 l=0 ul2=1(cid:12)(cid:12)(cid:12) X1≤m<l≤n−1 l=0 ul2=1 Pn−1 sup sup Pn−1 αl−mulum(cid:12)(cid:12)(cid:12) α(cos t − α) 1 − 2α cos t + α2 = g(t). = ω2(Jn(α)). We note that Re(Jn(α)) is the Toeplitz matrix associated with the Toeplitz form 1 2 (Pα(eit) − 1) = To complete the proof of the proposition, we can easily observe that if a and b are arbitrary real number and f (x) a Toeplitz form with γn k as eigenvalues then the eigenvalues of a + bf (x) will be a + bγn k . This shows that the eigenvalues of Re(Jn(α)) are λ′(n) k = 1 2 (λ(n) k − 1) = k − α) α(cos t(n) 1 − 2α cos t(n) 1 + α2 , 1 ≤ k ≤ n. Now, since g(t) is decreasing on the interval [0, π] and Re(Jn(α) is symmetric then ω2(Jn(α)) = ω2(Re(Jn(α))) k ; 1 6 k 6 no = maxnλ′(n) = maxnλ′(n) n o ;−λ′(n) = max( α(cos t(n) 1 − α) 1 − 2α cos t(n) 1 1 + α2 = sn(α) (3.6) ; −α(cos t(n) 1 − 2α cos t(n) n + α2) n − α) where (3.6) is due to the fact that cos t(n) 3.1. n is nonpositive and by using Proposition (cid:3) Corollary 3.3. For 0 6 α < 1, we have ω2(J2(α)) = α 2 . Proof. This result is known, but it is interesting to notice that this result can also easily be obtained by using Proposition 3.2. Indeed p2(t) = sin(3t) − 2α sin(2t) + α2 sin t sin t = 4 cos2 t − 4α cos t + α2 − 1 Therefore, we obtain: cos t(2) then we have λ′(2) 1 = 2 = − α 2 α + 1 2 and λ′(2) 1 = α 2 . Now since T r(J2(α)) = 0 . This completes the proof. (cid:3) ON THE NUMERICAL RADIUS OF THE TRUNCATED ADJOINT SHIFT 9 Corollary 3.4. For 0 6 α < 1, we have ω2(J3(α)) = α(√α2 + 8 − 3α) 4 + 2α2 − 2α√α2 + 8 . Proof. We have p3(t) = 2 sin t(cid:0) sin(2t) − α sin t(cid:1)(cid:0) cos(2t) − α cos t(cid:1) α + √α2 + 8 , cos t(3) this implies that cos t(3) then 1 = 4 λ′(3) 1 = 1 − α) 1 + α2 α(cos t(3) 1 − 2α cos t(3) = −α(cos t(3) 1 − 2α cos t(3) 3 − α) 3 + α2 = = and λ′(3) 3 The proof is complete. 3 = α 2 2 = and cos t(3) α(√α2 + 8 − 3α) 4 + 2α2 − 2α√α2 + 8 α(√α2 + 8 + 3α) 4 + 2α2 + 2α√α2 + 8 α − √α2 + 8 4 . (cid:3) Corollary 3.5. For each integer n ≥ 4 and α 6qcos 2π n+1 , we have ω2(Jn(α)) = 1 − α) α(cos t(n) 1 − 2α cos t(n) 1 + α2 . Proof. It follows from the hypothesis of Corollary 3.5 that t(n) 1 ≤ (3.7) then and π n + 1 ≤ arccosr 1 + α2 2 ≤ arccos(α) 6 t(n) n 6 nπ n + 1 λ′(n) 1 = 1 − α) α(cos t(n) 1 − 2α cos t(n) 1 + α2 λ′(n) n = −λ′(n) n = −α(cos t(n) 1 − 2α cos t(n) n − α) n + α2 In view of the inequality (3.7), g(arccos α) = 0 and the fact that g is decreasing in [0, π], it suffices to prove that g( π g( π n + 1 ) − g( nπ n + 1 ) = = = n+1 ) ≥ g( nπ α(cos π 1 − 2α cos π n+1 ). We have n+1 − α) n+1 + α2 − 2α2(2 cos2 π α(cos π n+1 + α) n+1 + α2 1 + 2α cos π n+1 − α2 − 1) n+1 + α2)(1 + 2α cos π 2α2(cos 2π n+1 + α2)(1 + 2α cos π n+1 − α2) n+1 + α2) n+1 + α2) (1 − 2α cos π (1 − 2α cos π ≥ 0. The result follows. (cid:3) 10 HAYKEL GAAYA 3.2. The numerical radius of the compressed shift. In this section, we will focus on the particular case where First, we notice some properties for the general case where φ is a finite Blashke product : φ(z) = Qn evaluation functional kλ ∈ HI 2 by the requirement that f (λ) =< f, kλ >. Thus . For each λ in the unit disc, we define the j=1 . 1 − αz(cid:19)n φ(z) =(cid:18) z − α z − αj 1 − αjz kλ(z) = 1 1 − λz and {e1, . . . , en} the collection of functions of H(φ) defined as follows : and e1(z) =(cid:0)1 − α12(cid:1) ek(z) =(cid:0)1 − αk2(cid:1) for any k = 2, ..., n. 1 2 1 k−1 1 − α1z Yj=1 1 2 1 1 − αkz z − αj 1 − αjz It is known that {e1, . . . , en} is an orthonormal basis of H(φ) and with respect to this basis the matrix of S∗(φ) is given by [alk], where αl σlσl+1 σlσkQk−1 0 j=l+1(−αj) if l = k if k = l + 1 if k > l + 1 unless alk =  1 2 , for each 1 6 k 6 n. 0 Indeed, for k > l + 1, we have and σk =(cid:0)1 − αk2(cid:1) < S∗(φ)ek, el > = σkσlZ 2π = σkσlZ 2π = σkσlZ 2π Yj=l+1 = σkσl k−1 0 0 (−αj). 1 e−iθ 1 − αkeiθ 1 − αle−iθ 1 e−iθ 1 − αkeiθ 1 − αle−iθ 1 1 1 − αkeiθ 1 − αleiθ dθ 2π k−1 k−1 eiθ − αj 1 − αjeiθ Yj=l+1 eiθ − αj 1 − αjeiθ Yj=l eiθ − αl 1 − αleiθ Yj=l+1 k−1 eiθ − αj 1 − αjeiθ dθ 2π dθ 2π Using the same scheme as before,we prove easily that < S∗(φ)ek+1, ek >= σkσk+1 and < S∗(ek), el >= 0 if k < l. Finally < S∗(φ)ek, ek > = σ2 kZ 2π 0 1 e−iθ 1 − αkeiθ z 1 − αkz 1 − αke−iθ > dθ 2π = σ2 k < Kαk , = αk ON THE NUMERICAL RADIUS OF THE TRUNCATED ADJOINT SHIFT 11 In the sequel of the paper, φ denotes the finite Blashke product with unique zero α: S∗(φ) gets the following matricial representation: . 1 − αz(cid:19)n φ(z) =(cid:18) z − α ··· σ(−α)n−2 σ . . . . . . σ −σα ··· . . . α . . . . . . . . . . . . . . . . . . . . . . . . . . . α 0 α 0 ... ... ... 0   ... ... −σα σ α   where σ = 1 − α2. Proposition 3.6. For 0 6 α < 1, we have ω2(Re(S∗(φ))) = mn(α). Proof. First, notice that where α = 0, then Re(S∗(φ)) = 1 2 1 0 1 0 . . . 0 1 0 1 . . . 0 0 1 0 . . . . . . . . . . . . . . . . . . .     0 1 0 0 . . . kπ n + 1 In this case the eigenvalues are cos , for k = 1, . . . , n. For the proof there are many references, we refer the reader for example to [7] page 67 or [2] page 35, therefore ω2(Re(S∗(φ))) = cos = mn(0). Then we can limit our study to the case α 6= 0. Now, notice that n + 1 π Re(S∗(φ)) = 1 − α2 2α − 2α2 1 − α2 α ... αn−1   α . . . . . . ··· ··· . . . . . . α − αn−1 ... α 2α2 1 − α2 .   Here Re(S∗(φ)) is the Toeplitz matrix associated with the Toeplitz form: 1 − α2 2α (Pα(eit) − 1 − 2α2 1 − α2 ) = (1 + α2) cos t − 2α 1 − 2α cos t + α2 = h(t). Since h(t) is monotonic on [0, π], thus with the same argument that in the proof of Proposition 3.2, we may assume that: ω2(Re(S∗(φ))) = maxn (1+α2) cos tn 1−2α cos tn = mn(α). 1 −2α 1 +α2 , −(1+α2) cos tn 1−2α cos tn n+2α n+α2 o This completes the proof. (cid:3) 12 HAYKEL GAAYA 1 − αz(cid:19)n Theorem 3.7. Let φ(z) =(cid:18) z − α with α ∈ CI and α< 1. (1) The numerical radius of S∗(φ) is independent from the argument of α and for 0 6 α < 1 the numerical range of S∗(φ) is symmetric with respect to the real axis. (2) For n ≥ 2 , we have mn(α) 6 ω2(S∗(φ)) 6 Mn(α). Proof. For α 6= 0, and for t = arg(α), we have: ω2(S∗(φ)) = = = = = (3.8) sup kuk2=1 < S∗(φ)u, u > 1 − α2 Pn−1 Pn−1 Pn−1 sup sup α − l=0 ul2=1(cid:12)(cid:12)(cid:12) l=0 ul2=1(cid:12)(cid:12)(cid:12)α − l=0 ul2=1(cid:12)(cid:12)(cid:12)α − l=0 vl2=1(cid:12)(cid:12)(cid:12)α − sup sup Pn−1 α 1 − α2 α 1 − α2 α 1 − α2 α X1≤m<l≤n−1 X1≤m<l≤n−1 X1≤m<l≤n−1 X1≤m<l≤n−1 (−α)l−m ulum(cid:12)(cid:12)(cid:12) (−α)l−m ulum(cid:12)(cid:12)(cid:12) (−α)l−m eiltuleimtum(cid:12)(cid:12)(cid:12) (−α)l−m vlvm(cid:12)(cid:12)(cid:12) . The last equality implies that the numerical radius of S∗(φ) is independant from the argument of α. Hence we can suppose that 0 < α < 1. Now assume that z is in W (S∗(φ)), then there is u = (u0, . . . , un−1) a unit vector in CI n such that z = < S∗(φ)u, u > = α − 1 − α2 α X1≤m<l≤n−1 (−α)l−m ulum and z = α − 1 − α2 α X1≤m<l≤n−1 (−α)l−m umul = < S∗(φ)u, u > . This implies that z is in W (S∗(φ)) and (1) holds. As remarked before, we can restrict our study to the case where 0 < α < 1 and from (3.8) we have: ω2(S∗(φ)) ≤ α + 1 − α2 α 1 − α2 = α + α = Mn(α). Pn−1 sup l=0 ul2=1(cid:12)(cid:12)(cid:12) X1≤m<l≤n−1 αl−mulum(cid:12)(cid:12)(cid:12) ω2(Jn(α)) On the other hand, it is obvious to note that ω2(ReT ) 6 ω2(T ) for each bounded operator T , an application of the Proposition 3.6 completes the proof of (2). (cid:3) ON THE NUMERICAL RADIUS OF THE TRUNCATED ADJOINT SHIFT 13 Corollary 3.8. For n = 2 and 0 6 α < 1; we have: 1 + 2α − α2 ω2(S∗(φ)) = 2 . Corollary 3.9. For each integer n ≥ 2; ω2(S∗ n) = cos π n + 1 . In the last two corollaries, the results are due to the fact that both quantities mn and Mn coincide. Acknowledgements: The author would like to express his gratitude to Gilles Cassier for his help and his good advices and Alfonso Montes Rodriguez as well for advices in the translation of this work.. References [1] C. Badea and G. Cassier, Constrained von Neumann inequalities, Adv. Math. 166 (2002), no. 2, 260 -- 297. [2] A. Böttcher and S. Grudsky, Spectral properties of banded Toeplitz matrices. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 2005. [3] G. Cassier, I. Chalendar, and B. Chevreau, New examples of contractions illustring member- ship and non membership in The classe An,m, Acta Sci. Math (Szeged) 64 (1998), 701 -- 731. [4] E.V. Egerváry and O. Szász, Einige Extremalprobleme in Bereiche der trigonometrischen Polynomen, Math. Z. 27 (1928), 641-652. [5] H. L. Gau and P. Y. Wu, Numerical range and Poncelet property, Taiwanese J. Math. 7 (2003), no. 2, 173 -- 193. [6] H. L. Gau and P. Y. Wu, Numerical range of S(φ). Linear and Multilinear Algebra 45 (1998), no. 1, 49 -- 73. [7] U. Grenander and G. Szegö, Toeplitz forms and their applications. California Monographs in Mathematical Sciences University of California Press, Berkeley-Los Angeles 1958. [8] K.E. Gustafson, D.K.M. Rao: Numerical Range, Springer, New York, 1997. [9] U. Haagerup and P. de la Harpe, The numerical radius of a nilpotent operator on a Hilbert space, Proc. Amer. Math. Soc. 115(1992), 371 -- 379. [10] P. R. Halmos, A Hilbert space problem book. Second edition. Graduate Texts in Mathematics, 19. Encyclopedia of Mathematics and its Applications, 17. Springer-Verlag, New York-Berlin, 1982. [11] R. A. Horn and C. R. Johnson, Topics in matrix analysis. Cambridge University Press, Cambridge, 1991. [12] M. Kac, W. L. Murdock and G. Szegö, On the eigenvalues of certain Hermitian forms. J. Rational Mech. Anal. 2, (1953). 767 -- 800. [13] B. Sz.-Nagy and C. Foias, Harmonic analysis of operator on Hilbert space, North-Holland Publishing Co., Amesterdam-London; Americain Elsevier Publishing Co., Inc., New York; Akadémiai Kiado, Budapest, 1970. [14] W. F. Trench, Interlacement of the even and odd spectra of real symmetric Toeplitz matrices. Linear Algebra Appl. 195 (1993), 59 -- 68. ‡.Institute Camille Jordan, Office 107 University of Lyon1, 43 Bd November 11, 1918, 69622-Villeurbanne, France. E-mail address: ‡[email protected]
1001.1516	1	1001	2010-01-10T13:46:35	Continuous Shearlet Tight Frames	[ "math.FA" ]	Based on the shearlet transform we present a general construction of continuous tight frames for $L^2(\mathbb{R}^2)$ from any sufficiently smooth function with anisotropic moments. This includes for example compactly supported systems, piecewise polynomial systems, or both. From our earlier results it follows that these systems enjoy the same desirable approximation properties for directional data as the previous bandlimited and very specific constructions due to Kutyniok and Labate. We also show that the representation formulas we derive are in a sense optimal for the shearlet transform.	math.FA	math	Continuous Shearlet Tight Frames Philipp Grohs∗ Abstract Based on the shearlet transform we present a general construction of continuous tight frames for L2(R2) from any sufficiently smooth function with anisotropic moments. This includes for example compactly sup- ported systems, piecewise polynomial systems, or both. From our earlier results in [5] it follows that these systems enjoy the same desirable ap- proximation properties for directional data as the previous bandlimited and very specific constructions due to Kutyniok and Labate; [8]. We also show that the representation formulas we derive are in a sense optimal for the shearlet transform. Contents 1 Introduction 1.1 Shearlets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 The construction Ingredients 2.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Main Result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 Concluding Remarks 4 Acknowledgments 1 Introduction 1 2 3 4 7 9 10 The purpose of this short paper is to give a construction of a system of bivariate functions which has the following desirable properties: directionality. The geometry of the set of singularities of a tempered distribution f can be accurately described in terms of the interaction be- tween f and the elements of the system. tightness. The system forms a tight frame of L2(R2). locality. The representation is local. By this we mean that the represen- tation can also be interpreted as a representation with respect to a non tight frame and its dual frame such that both of these frames only consist of compactly supported functions. ∗TU Wien, strasse http://www.dmg.tuwien.ac.at/grohs, phone: +43 58801 11318. 8-10, Institute of Discrete Mathematics and Geometry, Wiedner Haupt- 1040 Wien, Austria. [email protected], web: Email: 1 The construction is based on the shearlet transform which has been introduced in [9] and has become popular in Computational Harmonic Analysis for its ability to sparsely represent bivariate functions. A related important result is that the coefficients of a general tempered distribution with respect to this transform exactly characterize the Wavefront Set (see e.g. [7] for the definition) of this distribution [8]. Notation. We shall use the symbol · indiscriminately for the absolute value on R, R2, C and C2. We usually denote vectors in R2 by x, t, ξ, ω and their elements by x1, x2, t1, t2, . . . . In general it should always be clear to which space a variable belongs. The symbol k · k is reserved for various function space and operator norms. For two vectors s, t ∈ R2 we denote by st their Euclidean inner product. We use the symbol f . g for two functions f, g if there exists a constant C such that f (x) ≤ Cg(x) for large values of x. We will often speak of frames. By this we mean continuous frames as defined in [1]. We define f (ω) = R f (x) exp(2πiωx)dx to be the Fourier transform for a function f ∈ L1 ∩ L2 and continuously extend this notion to tempered distributions. 1.1 Shearlets We start by defining what a shearlet is and what the shearlet transform is: Definition 1.1. A function ψ ∈ L2(R2) is called shearlet if it possesses M ≥ 1 vanishing moments in x1-direction, meaning that ZR2 ψ(ω)2 ω12M dω < ∞. Let f ∈ L2(R2). The shearlet transform of f with respect to a shearlet ψ maps f to where SHψf (a, s, t) := hf, ψasti, ψast(x) := ψ(cid:0) x1 − t1 − s(x2 − t2) a , x2 − t2 a1/2 (cid:1). The following reproduction formula holds [4]: Theorem 1.2. For all f ∈ L2(R2) and ψ a shearlet f (x) = ZR2 ZRZR+ SHψ(a, s, t)ψast(x)a−3dadsdt, (1) where equality is understood in a weak sense. The shearlet transform captures local, scale- and directional information via the parameters t, a, s respectively. A significant drawback of this representation is the fact that the directional parameter runs over the non compact set R. Also it is easy to see that the distribution of directions becomes infinitely dense as s grows. These problems led to the construction of shearlets on the cone [8]. The idea is to restrict the shear parameter s to a compact interval. Since this only allows to caption a certain subset of all possible directions, the function f is split into f = P f + P νf , where P is the frequency projection onto the cone with slope ≤ 1 and only P f is analyzed with the shearlet ψ while P νf 2 is analyzed with ψν(x1, x2) := ψ(x2, x1). For very specific choices of ψ Labate and Kutyniok proved a representation formula 2 = ZR2 hf, TtWi2dt +ZR2 Z 2 kfk2 −2Z 1 0 SHψP f (a, s, t)2a−3dadsdt +ZR2 Z 2 −2Z 1 0 SHψν P νf (a, s, t)2a−3dadsdt, (2) Tt denoting the translation operator by t. The main weakness of this decompo- sition is its lack of locality: indeed, first of all the need to perform the frequency projection P to f destroys any locality. But also the functions ψ which have been considered in [8] are very specific bandlimited functions which do not have compact support. As a matter of fact no useful local representation via compactly supported functions which is able to capture directional smoothness properties has been found to date. The present paper aims at providing a step towards finding such a representation using two crucial observations: First, in [5] we were able to show that the description of directional smoothness via the decay rate of the shearlet coefficients for a → 0 essentially works for any func- tion which is sufficiently smooth and has sufficiently many vanishing moments in the first direction, hence also for compactly supported functions. The second observation is that actually a full frequency projection P is not necessary to ar- rive at a useful representation similar to (2). Instead of the operator P we will use a 'localized frequency projection' which is given by Fourier multiplication with a function p0 to be defined later. Our main result Theorem 2.4 will prove a representation formula similar to (2) where ψ is allowed to be compactly supported and the frequency projection is replaced with a local variant. We also show that this local variant of the frequency projection is the best we can do -- without it, no useful continuous tight frames (or continuous frames with a structured dual) can be constructed within the scope of the shearlet transform. This is shown in Theorem 2.6. 2 The construction The goal of this section is to derive a representation formula kfk2 2 = 0 hf, q0 ∗ ψasti2a−3dadsdt 1 Cψ(cid:0)ZR2 Z 2 −2Z 1 −2Z 1 0 hf, q1 ∗ ψν + ZR2 Z 2 + ZR2 hf, Ttϕi2dt(cid:1) asti2a−3dadsdt (3) for L2(R2)-functions f and with some localized frequency projections q0, q1 and a window function ϕ to be defined later. In order to guarantee that the shearlet- part contains the high-frequency part of f and the rest contains low frequencies, it is necessary to ensure that the window function ϕ in this formula is sufficiently smooth. 3 2.1 Ingredients Here we first state the definitions and assumptions that we use in the construc- tion. Then we collect some auxiliary results which we later combine to prove our main results Theorems 2.4 and 2.6. The large part of the results will concern the construction of a useful window function and to ensure its smoothness. We say that a bivariate function f has Fourier decay of order Li in the i-th variable (i ∈ {1, 2}) if f (ξ) . ξi−Li. We start with a shearlet ψ which has M vanishing moments in x1-direction and Fourier decay of order L1 in the first variable. It is clear from the definition θ with θ ∈ L2(R2). We assume that θ has Fourier decay of order L2 in the second variable so that the following relation holds: of vanishing moments that ψ = (cid:0) ∂ ∂x1(cid:1)M We also set and 2M − 1/2 > L2 > M ≥ 1. N := 2 min(L2 − M, L1), Cψ := ZR2 ψ(ω)2 ω12 dω, ∆ψ(ξ) := Z 2 −2Z 1 0 ψ(aξ1, a1/2(ξ2 − sξ1))2a−3/2dads. We define functions ϕ0, ϕ1 via (4) (5) (6) (7) (8) ϕ0(ξ)2 = Cψ − ∆ψ(ξ) and ϕ1(ξ)2 = Cψ − ∆ψν (ξ), where ψν (x1, x2) := ψ(x2, x1). We write χC for the indicator function of the cone C = {ξ = (ξ1, ξ2) ∈ R2 : ξ2 ≤ ξ1}. Finally, we pick a smooth and compactly supported bump function Φ with Φ(0) = 1 and define functions p0, p1 via Clearly, p0 and p1 are both compactly supported tempered distributions. p0 = Φ ∗ χC and p1 = 1 − p0. Lemma 2.1. We have p0(ξ) . ξ−N for ξ1/ξ2 ≥ 3/2 p1(ξ) . ξ−N for ξ2/ξ1 ≥ 3/2 (9) (10) Proof. Assume that ξ = te with e a unit vector with e1/e2 ≥ 3/2 and t > 0. There exists a uniform δ > 0 such that for all η with η < δt we have ξ− η ∈ Cc, and hence χC(ξ − η) = 0. It follows that we can write p0(ξ) = ZR2 χC(ξ − η) Φ(η)dη ≤ Zη>δt Φ(η)dη . t−N = ξ−N 4 if Φ is sufficiently smooth. On the other hand, let ξ = te with e a unit vector with e2/e1 ≥ 3/2 and t > 0. There exists a uniform δ > 0 such that for all η with η < δt we have ξ − η ∈ C and hence χC(ξ − η) = 1. Now we can estimate p1(ξ) = 1 − p0(ξ) = ZR2 Φ(η)(1 − χC(ξ − η))dη = Zη>δt Φ(η)(1 − χC(ξ − η))dη . t−N = ξ−N again for Φ smooth. Note that in the first equality we have used that Φ(0) = 1. This proves the statement. Lemma 2.2. We have ϕ0(ξ)2 . ξ−N for ξ1/ξ2 ≤ 3/2 ϕ1(ξ)2 . ξ−N for ξ2/ξ1 ≤ 3/2 (11) Proof. This follows from [5, Lemma 4.7]. For the convenience of the reader we present a proof here as well. We only prove the assertion for ϕ0 since the proof of the corresponding statement for ϕ1 is the same. By definition we have ϕ0(ξ)2 = (cid:0)Za∈R, s>2 ψ(cid:0)aξ1,√a(ξ2 − sξ1)(cid:1)2a−3/2dads +Za>1, s<2 ψ(cid:0)aξ1,√a(ξ2 − sξ1)(cid:1)2a−3/2dads(cid:1). We start by estimating the second integral using the Fourier decay in the first variable: Za>1, s<2 ψ(cid:0)aξ1,√a(ξ2 − sξ1)(cid:1)2a−3/2dads . 4Za>1 (aξ1)−2L1 a−3/2da . ξ1−2L1 . ξ−2L1 for all ξ in the cone with slope 3/2. To estimate the other term we need the θ(ξ) moment condition and the decay in the second variable. We write ψ(ξ) = ξM 1 and ξ = (ξ1, rξ1), r ≤ 3/2. We start with the high frequency part: Za<1, s>2 ψ(cid:0)aξ1,√a(ξ2 − sξ1)(cid:1)2a−3/2da = Za<1, s>2 aξ12Mθ(cid:0)aξ1,√a(ξ2 − sξ1)(cid:1)2a−3/2dads . Za<1, s>2 aξ12M√a(ξ2 − sξ1)−2L2a−3/2dads = Za<1, s>2 a2M−L2−3/2ξ12M−2L2r − s−2L2dads we now use that r − s is always strictly away from zero. By assumption L2 = 2M − 1/2 − ε for some ε > 0. Hence we can estimate further . . . = ξ1−2(L2−M)Za<1, s>2 a−1+εr − s−2L2 dads . ξ−2(L2−M). 5 The low-frequency part can simply be estimated as follows: Za>1, s>2 ψ(cid:0)aξ1,√a(ξ2 − sξ1)(cid:1)2a−3/2dads a−3/2−L2r − s−2L2dads . ξ1−2L2 Za>1, s>2 . ξ−2L2. Putting these estimates together proves the statement. Lemma 2.3. Assume that ψ is compactly supported with support in the ball BA := {ξ : ξ ≤ A}. Then the tempered distributions ( ϕi2)∨, i = 0, 1 are both of compact support with support in the ball 2p3 + √5BA. Proof. Since the inverse Fourier transform of the constant function Cψ is a Dirac, this follows if we can establish that the functions ∆ψ and ∆ψν are Fourier transforms of distributions of compact support. We show this only for ∆ψ, the In what follows we will use the notation f−(x) := other case being similar. f (−x) for a function f . Since ψ has M anisotropic moments we can write ψ = ∂x1(cid:1)M (cid:0) ∂ θ for some θ ∈ L2(R2) with the same support as ψ. Consider the function Θ(ξ) := θ(ξ)2. It is easy to see that this is the Fourier transform of the so-called Autocorrellation function θ∗ θ− of θ which is compactly supported with support in B2A. Therefore, by the theorem of Paley-Wiener, the function Θ possesses an analytic extension to C2 which we will henceforth call F . Furthermore, by the same theorem F is of exponential type: F (ζ) . (1 + ζ)K exp(2Aℑζ), (12) where ζ ∈ C2 and K ∈ N. We now consider the analytic extension Γ of the function ∆ψ which is given by Γ(ζ) = Z 2 −2Z 1 0 (aζ1)2M F (aζ1, a1/2(ζ2 − sζ1))a−3/2dads, ζ ∈ C2. a 0 Since M ≥ 1 by (4) the above integral is locally integrable which implies that Γ is actually an entire function. It is also of exponential type: Writing Mas := (cid:18) −sa1/2 a1/2 (cid:19) a short computation reveals that kMaskL2(C2)→L2(C2) ≤ a1/2(cid:0)1 + )1/2(cid:1)1/2 =: a1/2C(s). + (s2 + s2 2 s2 4 Now we estimate ∆ψ(ζ) = Z 2 −2Z 1 0 (aζ1)2M F (Masζ)a−3/2dads . ζ2M (1 + Masζ)K exp(Masℑζ) (1 + ζ)2M+K exp(2AC(s)ℑζ) . (1 + ζ)2M+K exp(2AC(2)ℑζ). . sup s∈[−2,2] By the Theorem of Paley-Wiener-Schwartz [7] it follows that ∆ψ is of compact support with support in B2√3+√5A . 6 2.2 Main Result We are now ready to prove our main result, the local representation formula in Theorem 2.4 which is similar to (2) only with local frequency projections (given by convolution with p0, p1) and with possibly compactly supported shearlets. In addition, in Theorem 2.6 we show that in a way this is the simplest repre- sentation that one can achieve with shearlets -- the local frequency projections are necessary in order to wind up with useful systems. From now on we shall assume that 0 ≤ Φ(ξ) ≤ 1 for all ξ ∈ R2 and define qi(ξ) := pi(ξ)1/2. Furthermore, we define ϕ(ξ) := (p0(ξ)ϕ0(ξ)2 + p1(ξ)ϕ1(ξ)2)1/2. Observe that by the positivity assumption above the radicands in the previous definitions are nonnegative and real. Theorem 2.4. We have the representation formulas kfk2 2 = 0 hf, q0 ∗ ψasti2a−3dadsdt 1 Cψ(cid:0)ZR2 Z 2 −2Z 1 −2Z 1 0 hf, q1 ∗ ψν + ZR2 Z 2 + ZR2 hf, Ttϕi2dt(cid:1) asti2a−3dadsdt and f = f high + f low := + ZR2 Z 2 −2Z 1 0 hf, ψν 0 hf, ψastip0 ∗ ψasta−3dadsdt 1 −2Z 1 Cψ(cid:0)ZR2 Z 2 astip1 ∗ ψasta−3dadsdt(cid:1) + 1 Cψ(cid:0)ZR2hf, Ttϕ0ip0 ∗ Ttϕ0dt +ZR2hf, Ttϕ1ip1 ∗ Ttϕ1dt(cid:1). The function f low satisfies (f low)∧(ξ) . ξ−N (13) (14) (15) for any f . We have the following locality property: If ψ is compactly supported in BA and Φ has support in BB, then f low(t) only depends on f restricted to t + B2√3+√5A+B Proof. We first prove (14). Taking the Fourier transform of the right hand side yields . 1 Cψ(cid:0)p0(ξ)(∆ψ(ξ) + ϕ0(ξ)2) + p1(ξ)(∆ψν (ξ) + ϕ1(ξ)2)(cid:1) f (ξ) = f (ξ). 7 The proof of (13) is similar. We prove the equation (15): Again taking the Fourier transform of f low gives 1 Cψ(cid:0)p0(ξ) ϕ0(ξ)2 + p1(ξ) ϕ1(ξ)2(cid:1) f (ξ). Now, the desired estimate follows from Lemmas 2.1 and 2.2. The last statement follows from the observation that f low can be written as f low = 1 Cψ f ∗(cid:0)p0 ∗ ( ϕ02)∨ + p1 ∗ ( ϕ12)∨(cid:1) together with Lemma 2.3. Remark 2.5. While we could show that the functions ϕ−i ∗ ϕi and pi, i = 1, 2 are compactly supported, it would be desirable to show that also the functions ϕi, qi, i = 1, 2 have compact support. We currently do not know how to do this. By a result of Boas and Kac [2], in the univariate case there always exists for any compactly supported function g with positive Fourier transform a compactly supported function f with f ∗ f− = g. However, in dimensions ≥ 2 this holds no longer true and things become considerably more difficult. The previous theorem for the first time gives a completely local represen- tation for square integrable functions which also allows to handle directional phenomena efficiently: by the results of [5] it follows that the decay rate of the coefficients hf, ψasti for a → 0 accurately describes the microlocal smoothness of f at t in the direction with slope s. It is interesting to ask if the frequency projections given by convolution with p0, p1 are really necessary, or in other words if it is possible to construct tight frame systems (Ttϕ)t∈R2 ∪ (ψast)a∈[0,1],s∈[−2,2],t∈R2 ∪ (ψν ast)a∈[0,1],s∈[−2,2],t∈R2 for L2(R2). We show that this is actually impossible, meaning that in a sense Theorem 2.4 is the best we can do. We remark that for any shearlet ψ, the constant Cψ can also be computed as Cψ = Z ∞ −∞Z ∞ 0 ψ(aξ1, a1/2(ξ2 − sξ1))2a−3/2dads, as a short computation reveals. This fact is related to the inherent group struc- ture of the shearlet transform [4]. Theorem 2.6. Assume that ψ = (cid:0) ∂ θ is a shearlet with M ≥ 1 vanishing moments in x1-direction such that either M > 1 or M = 1 and θ ∈ L∞(R2). Furthermore we assume and L1 > 0, L2 > M with L1, L2 defined as in Lemma 2.2. Then there does not exist a window function ϕ such that ∂x1(cid:1)M ϕ(ξ) = 0 lim ξ→∞ and such that the system (Ttϕ)t∈R2∪(ψast)a∈[0,1],s∈[−2,2],t∈R2∪(ψν constitutes a tight frame for L2(R2), which means that a representation formula ast)a∈[0,1],s∈[−2,2],t∈R2 kfk2 2 = 1 C(cid:0)ZR2 Z 2 −2Z 1 −2Z 1 0 hf, ψν + ZR2 Z 2 0 hf, ψasti2a−3dadsdt asti2a−3dadsdt +ZR2 hf, Ttϕi2dt(cid:1) (16) 8 holds for all f ∈ L2(R2) and some constant C. Proof. In terms of the Fourier transform, (16) translates to ∆ψ(ξ) + ∆ψν (ξ) + ϕ(ξ)2 = C. If we assume that ϕ has Fourier decay limξ→∞ ϕ(ξ) = 0 this would imply that lim ξ→∞(cid:0)∆ψ(ξ) + ∆ψν (ξ)(cid:1) = C. Lemma 2.2 implies that limξ→∞ ∆ψ(ξ) = limξ→∞ ∆ψν (ξ) = Cψ for 2 ξ2 ≤ 3 2 . It follows that C = 2Cψ. On the other hand, using the moment condition θ for M = 1 and θ ∈ L∞(R2), we have the following estimate for ∆ψ ψ = ξM 1 and ξ in the strip Sδ := {ξ : ξ1 ≤ δ}: −2Z 1 a2kθk∞a−3/2dads ≤ 8kθk∞δ2. 3 ≤ ξ1 ∆ψ(ξ) ≤ δ2Z 2 0 If we assume that M > 1, we can write ψ(ξ) = ξM−1 shearlet. We get a similar estimate as above: 1 µ(ξ) where µ is still a ∆ψ(ξ) ≤ δ2(M−1)Z 2 0 µ(aξ1, a1/2(ξ2 − sξ1))2a−3/2dads ≤ Cµδ2(M−1). At any rate, by choosing δ appropriately small, this implies that for ξ ∈ Sδ we have −2Z 1 C = lim ξ→∞(cid:0)∆ψ(ξ) + ∆ψν (ξ)(cid:1) < C, which gives a contradiction. It is easy to extend this argument to show that there do not exist shearlet frames such that there exists a dual frame which also has the structure of a shearlet system, see [1] for more information on frames in general and [6] for more information on shearlet frames in particular. 3 Concluding Remarks In future work we would like to address the problem of constructing continuous tight frames for the so-called 'Hart Smith Transform' [3, 11] where the shear operation is replaced with a rotation. We think that in this case the results might become simpler. One reason for this is that in this case no (smoothed) projection onto a frequency cone is needed. In view of constructing discrete tight frames we think that a simple discretization of a continuous tight frame will not work for non-bandlimited shearlets. The approach that we are currently pursuing in this direction is to construct so-called Shearlet MRA's via specific scaling functions and to try to generalize the 'unitary extension principle' of Ron and Shen [10] to the shearlet setting [6]. 9 4 Acknowledgments The research for this paper has been carried out while the author was working at the Center for Geometric Modeling and Scientific Visualization at KAUST, Saudi Arabia. References [1] S. T. Ali, J. P. Antoine, and J. P. Gazeau. Continuous frames in Hilbert space. Annals of Physics, 222:1 -- 37, 1993. [2] R. Boas Jr. and M. Kac. Inequalities for Fourier transforms of positive functions. Duke Mathematical Journal, 12(1):189 -- 206, 1945. [3] E. J. Candes and D. L. Donoho. Continuous curvelet transform: I. resolu- tion of the wavefront set. Applied and Computational Harmonic Analysis, 19:162 -- 197, 2003. [4] S. Dahlke, G. Kutyniok, P. Maass, C. Sagiv, H.-G. Stark, and G. Teschke. The uncertainty principle associated with the continuous shearlet trans- form. International Journal of Wavelets, Multiresolution and Information Processing, 6:157 -- 181, 2008. [5] P. Grohs. set. Continuous shearlet frames and resolution of the wave- from Technical front http://www.dmg.tuwien.ac.at/grohs/papers/shres.pdf. report, KAUST, 2009. available [6] P. Grohs. Refinable functions for composite dilation systems. 2009. manuscript in preparation. [7] L. Hormander. The Analysis of linear Partial Differential Operators. Springer, 1983. [8] G. Kutyniok and D. Labate. Resolution of the wavefront set using con- tinuous shearlets. Transactions of the American Mathematical Society, 361:2719 -- 2754, 2009. [9] D. Labate, G. Kutyniok, W.-Q. Lim, and G. Weiss. Sparse multidimen- In Wavelets XI (San Diego, CA, sional representation using shearlets. 2005), 254-262, SPIE Proc. 5914, SPIE, Bellingham, WA, 2005. [10] A. Ron and Z. Shen. Affine systems in L2(Rd): The analysis of the analysis operator. Journal of Functional Analysis, 148(2):408 -- 447, 1997. [11] H. F. Smith. A Hardy space for Fourier integral operators. Journal of Geometic Analysis, 8:629 -- 653, 1998. 10
1004.2103	1	1004	2010-04-13T03:48:04	On dominant contractions and a generalization of the zero-two law	[ "math.FA" ]	Zaharopol proved the following result: let $T,S:L^1(X,{\cf},\m)\to L^1(X,{\cf},\m)$ be two positive contractions such that $T\leq S$. If $\\|S-T\\|<1$ then $\\|S^n-T^n\\|<1$ for all $n\in\bn$. In the present paper we generalize this result to multi-parameter contractions acting on $L^1$. As an application of that result we prove a generalization of the "zero-two" law.	math.FA	math	ON DOMINANT CONTRACTIONS AND A GENERALIZATION OF THE ZERO-TWO LAW FARRUKH MUKHAMEDOV Abstract. Zaharopol proved the following result: let T, S : L1(X,F , µ) → L1(X,F , µ) be two positive contractions such that T ≤ S. If kS − Tk < 1 then kS n − T nk < 1 for all n ∈ N. In the present paper we generalize this result to multi-parameter contractions acting on L1. As an application of that result we prove a generalization of the "zero-two" law. Keywords: dominant contraction, positive operator, "zero-two" law. AMS Subject Classification: 47A35, 17C65, 46L70, 46L52, 28D05. 0 1 0 2 r p A 3 1 ] . A F h t a m [ 1 v 3 0 1 2 . 4 0 0 1 : v i X r a 1. Introduction Let (X,F , µ) be a measure space with a positive σ-additive measure µ. In what follows for the sake of shortness by L1 we denote the usual L1(X,F , µ) space associated with (X,F , µ). A linear operator T : L1 → L1 is called a positive contraction if T f ≥ 0 whenever f ≥ 0 and kTk ≤ 1. In [9] it was proved so called "zero-two" law for positive contractions of L1-spaces: Theorem 1.1. Let T : L1 → L1 be a positive contraction. If for some m ∈ N ∪ {0} one has kT m+1 − T mk < 2, then n→∞kT n+1 − T nk = 0. lim In [2] it was proved a "zero-two" law for Markov processes, which allowed to study random walks on locally compact groups. Other extensions and gen- eralizations of the formulated law have been investigated by many authors [7, 4, 5]. Using certain properties of L1-spaces Zaharopol [10] by means of the follow- ing theorem reproved Theorem 1.1. Theorem 1.2. Let T, S : L1 → L1 be two positive contractions such that T ≤ S. If kS − Tk < 1 then kSn − T nk < 1 for all n ∈ N In the paper we provide an example (see Example 2) for which the formu- lated theorem 1.2 can not be applied. Therefore, we prove a generalization of Theorem 1.2 for multi-parameter contractions acting on L1. As a consequence 1 2 FARRUKH MUKHAMEDOV of that result we shall provide a generalization of the "zero-two" law. Similar generalization has been considered in [5]. 2. Dominant operators Let T, S : L1 → L1 be two positive contractions. We write T ≤ S if S − T is a positive operator. In this case we have (2.1) for every x ≥ 0. Moreover, for positive operator T : L1 → L1 one can prove the following equality kSx − T xk = kSxk − kT xk, (2.2) kxk=1kT xk = sup The main result of this section is the following kTk = sup kxk=1,x≥0kT xk. 2 k < 1. Then kS1Sn Theorem 2.1. Let T1, T2, S1, S2 : L1 → L1 be positive contractions such that Ti ≤ Si, i = 1, 2 and S1S2 = S2S1. If there is an n0 ∈ N such that kS1Sn0 2 − T1T n0 2 − T1T n Proof. Let us assume that kS1Sn denote 2 k < 1 for every n ≥ n0. 2 − T1T n m = min{n ∈ N : kS1Sn0+n 2 k = 1 for some n > n0. Therefore, It is clear that m ≥ 1. The inequalities T1 ≤ S1, T2 ≤ S2 imply that S1Sn0+n − T1T n0+n is a positive operator. Then according to (2.2) there exists a sequence {xn} ∈ L1 such that xn ≥ 0, kxnk = 1,∀n ∈ N and (2.3) − T1T n0+n k = 1}. 2 2 2 2 Positivity of S1Sn0+n 2 n→∞k(S1Sn0+n lim − T1T n0+n 2 2 2 )xnk = 1. − T1T n0+n and xn ≥ 0 together with (2.1) imply that k(S1Sn0+n (2.4) for every n ∈ N. It then follows from (2.3),(2.4) that )xnk = kS1Sn0+m − T1T n0+n 2 2 2 xnk − kT1T n0+m 2 xnk (2.5) (2.6) 2 n→∞kS1Sn0+m lim n→∞kT1T n0+m lim 2 xnk = 1, xnk = 0. Thanks to the contractivity of S, Z and S1S2 = S2S1 one gets kS1Sn0+m 2 xnk = kS2(S1Sn0+m−1 2 xn)k ≤ kS1Sn0+m−1 2 xnk ≤ kSm 2 xnk which with (2.5) yields (2.7) n→∞kS1Sn0+m−1 lim 2 xnk = 1, n→∞kSm lim 2 xnk = 1. Moreover, the contractivity of Si, Ti (i = 1, 2) implies that kT1T n0+m−1 xnk ≤ 2 T mxnk ≤ 1 for every n ∈ N. Therefore, we may 2 xnk ≤ 1 and kS1Sn0 1, kT m 2 ON DOMINANT CONTRACTIONS choose a subsequence {yk} of {xn} such that the sequences {kT1T n0+m−1 {kT m (2.8) 2 T mxkk} converge. Put 2 ykk}, {kS1Sn0 α = lim 2 3 ykk}, (2.9) (2.10) β = lim 2 ykk, 2 T mykk, k→∞kT1T n0+m−1 k→∞kS1Sn0 k→∞kT m 2 ykk. − T1T n0+m−1 γ = lim 2 The inequality kS1Sn0+m−1 Hence we may choose a subsequence {zk} of {yk} such that kT1T n0+m−1 for all k ∈ N. and hence γ > 0. 2 zkk together with (2.8), (2.10) we find α ≤ γ, k < 1 with (2.7) implies that α > 0. zkk 6= 0 zkk ≤ kT m 2 2 2 From kT1T n0+m−1 Using (2.1) one gets kS1Sn0 2 T m 2 2 zkk = kS1Sn0+m = kS1Sn0+m ≥ kS1Sn0+m = kS1Sn0+m 2 2 2 2 zk − (S1Sn0+m zkk − kS1Sn0+m zkk − kSm zkk − kSm zk − S1Sn0 2 T m 2 zk)k zk − S1Sn0 2 T m 2 zkk 2 zk − T m 2 zkk 2 zkk + kT m 2 zkk 2 (2.11) Due to (2.5),(2.7) we have k→∞kS1Sn0+m lim 2 zkk − kSm 2 zkk = 0; which with (2.11) implies that k→∞kS1ZSn0 lim 2 T m 2 zkk ≥ lim k→∞kT m 2 zkk, therefore, β ≥ γ. On the other hand, by kS1Sn0 Now set 2 T m 2 zkk one gets γ ≥ β, hence γ = β. 2 zkk ≤ kT m T m 2 zk kT m 2 zkk uk = , k ∈ N. Then using the equality γ = β and (2.6) one has kS1Sn0 2 T m kT m 2 zkk 2 ukk = lim k→∞ 2 zkk k→∞kS1Sn0 lim k→∞kT1T n0 lim 2 ukk = lim k→∞ So, owing to (2.1) and positivity of S1Sn0 2 , we get = 1, = 0. kT1T n0+mzkk kT m 2 zkk 2 − T1T n0 2 )zkk = 1. k→∞k(S1Sn0 lim 2 − T1T n0 Since kukk = 1, uk ≥ 0,∀k ∈ N from (2.2) one finds kS1Sn0 which is a contradiction. This completes the proof. 2 − T1T n0 2 k = 1, (cid:3) 4 FARRUKH MUKHAMEDOV Corollary 2.2. Let Z, T, S : L1 → L1 be positive contractions such that T ≤ S and ZS = SZ. If there is an n0 ∈ N such that kZ(Sn0 − T n0)k < 1. Then kZ(Sn − T n)k < 1 for every n ≥ n0. Assume that Z = Id. If n0 = 1, then from Corollary 2.2 we immediately get the Zaharopol's result (see Theorem 1.2). If n0 > 1 then we obtain a main result of [8]. Let us provide an example of Z, S, T positive contractions for which state- ment of Corollary 2.2 is satisfied. Example 1. Consider R2 with a norm kxk = x1 +x2, where x = (x1, x2). An order in R2 is defined as usual, namely x ≥ 0 if and only if x1 ≥ 0, x2 ≥ 0. Now define mappings Z : R2 → R2,T : R2 → R2 and S : R2 → R2, respectively, by (2.12) (2.13) Z(x1, x2) = (ux1 + vx2, ux2), S(x1, x2) = (cid:18)x1 + x2 2 , x2 2 (cid:19), T (x1, x2) = (λx2, 0). (2.14) The positivity of Z,S and T implies that u, v, λ ≥ 0. It is easy to check that T ≤ S holds if and only if 2λ ≤ 1. One can see that kZk = sup kxk=1 x≥0 kZxk = max x1+x2=1 x1,x2≥0 {ux1 + (u + v)x2} 0≤x2≤1{u + vx2} = max = u + v Hence, contractivity of Z implies that u + v = 1. Similarly, we find that kSk = 1 and kTk = λ. From (2.12) and (2.13) one gets that ZS = SZ. By means of (2.12),(2.13),(2.14) one finds Similarly, one gets kZ(S − T )k = sup kxk=1 x≥0 kZ(S − T )xk = max (2.15) = 2(cid:0)ux1 + x2 + ux2 − 2λux2(cid:1)(cid:27) x1+x2=1 x1 ,x2≥0 (cid:26) 1 1 + u(1 − 2λ) . 2 The condition 2λ ≤ 1 yields that kZ(S − T )k < 1. Consequently, Corollary 2.2 implies kZ(Sn − T n)k < 1 for all n ∈ N. Now let us formulate a multi-parametric version of Theorem 1.1. Theorem 2.3. Let Ti, Si : L1 → L1, i = 1, . . . , N be positive contractions such that Ti ≤ Si with (2.16) for every i, j = 1, . . . , N. TiTj = TjTi, SiSj = SjSi ON DOMINANT CONTRACTIONS 5 If there are ni,0 ∈ N, i = 1, . . . , N such that N − T n1,0 (2.17) · · · SnN,0 kSn1,0 1 1 · · · T nN,0 N k < 1. Then (2.18) kSm1 1 · · · SmN N − T m1 1 · · · T mN N k < 1 for all mi ≥ ni,0, i = 1, . . . , N . Proof. Let us fix the first N − 1 operators in (2.17), i.e. denote for a moment we (2.19) SN −1 = Sn1,0 1 · · · SnN−1,0 N −1 TN −1 = T n1,0 1 · · · T nN−1,0 N −1 , then (2.17) can be written as follows kSN −1SnN,0 N − TN −1T nN,0 N k < 1. After applying Theorem 2.1 to the last inequality we find (2.20) kSN −1SmN N − TN −1T mN N k < 1 for all mN ≥ nN,0. Now taking into account (2.19) and (2.16) we rewrite (2.20) as follows (2.21) N Sn1,0 kSmN 1 · · · SnN−1,0 N −1 − T mN N T n1,0 1 · · · T nN−1,0 N −1 k < 1. Now again applying the same idea as above to (2.21) we get N Sn1,0 kSmN 1 · · · SmN−1 N −1 − T mN N T n1,0 1 · · · T mN−1 N −1 k < 1, for all mN −1 ≥ nN −1,0, mN ≥ nN,0. Hence, continuing this procedure N − 2 times we obtain the desired inequality. (cid:3) Remark 3.1. It should be noted the following: (i) Since the dual of L1 is L∞ then due to the duality theory the proved Theorems 2.1 and 2.3 holds true if we replace L1-space with L∞. (ii) Unfortunately, that the proved theorems and its corollaries are not longer true if one replaces L1-space by an Lp-space, 1 < p < ∞. Indeed, consider X = {1, 2}, F = P({1, 2}) and the measure µ is given by µ({1}) = µ({2}) = 1/2. In this case, Lp is isomorphic to the Banach lattice R2 (here an order is defined as usual, namely x ≥ 0 if and only if x1 ≥ 0, x2 ≥ 0) with the norm kxkp = (cid:0)x1p + x2p(cid:1)1/p/2, where x = (x1, x2). Define two operators by S(x1, x2) = (cid:18)x1 + x2 2 , x1 + x2 2 (cid:19), T (x1, x2) = (cid:18)0, x1 2 (cid:19) Then it is shown (see [10]) that kS − Tk < 1, but kS2 − T 2k = 1. 6 FARRUKH MUKHAMEDOV (iii) It would be better to note that certain ergodic properties of dominant positive operators has been studied in [3]. In general, a monograph [6] is devoted to dominant operators. Let us give another example, for which conditions of Theorem 1.2 does not hold, but Theorem 2.1 can be applied. Example 2. Let us consider R2 as in Example 1. Now define mappings T : R2 → R2 and S : R2 → R2 as follows (2.22) x1 + S(x1, x2) = (cid:18)1 T (x1, x2) = (cid:18)1 2 4 x2, 0(cid:19). 1 3 x2, 1 2 x1 + 1 3 x2(cid:19), (2.23) (2.24) (2.25) It is clear that S and T are positive and T ≤ S. One can see that kSk = 1, kTk = 1/4. From (2.22),(2.23) one gets kS − Tk = sup kxk=1 x≥0 k(S − T )xk = max 0≤x1≤1(cid:26) 7x1 + 5 12 (cid:27) = 1 kS2 − T 2k = sup kxk=1 x≥0 k(S2 − T 2)xk = max 0≤x1≤1(cid:26) 5x1 + 10 18 (cid:27) = 15 18 Consequently, we have positive contractions T and S with S ≥ T such that kS − Tk = 1,kS2 − T 2k < 1. This shows that the condition of Theorem 1.2 is not satisfied, but due to Corollary 2.2 with Z = id we have kSn − T nk < 1 for all n ≥ 2. Therefore the proved Theorem 2.2 is an extension of the Zaharopol's result. 3. A generalization of the zero-two law In this section we are going to prove a generalization of the zero-two law for positive contractions on L1. Before formulate the main result we prove some auxiliary facts. First note that for any x, y ∈ L1 one defines (3.1) x ∧ y = (x + y − x − y). 1 2 It is well known (see [1]) that for any mapping S of L1 one can define its modulus by (3.2) Sx = sup{Sy : y ≤ x}, x ∈ L1, x ≥ 0. Hence, similarly to (3.1) for given two mappings S, T of L1 we define (3.3) (S ∧ T )x = 1 2 (Sx + T x − S − Tx), x ∈ L1. ON DOMINANT CONTRACTIONS 7 A linear operator Z : L1 → L1 is called a lattice homomorphism whenever (3.4) Z(x ∨ y) = Zx ∨ Zy holds for all x, y ∈ L1. One can see that such an operator is positive. Note that such homomorphisms were studied in [1]. Recall that a net {xα} in L1 is order convergent to x, denoted xα →o x whenever there exists another net {yα} with the same index set satisfying xα − x ≤ yα ↓ 0. An operator T : L1 → L1 is said to be order continuous, if xα →o 0 implies T xα →o 0. Lemma 3.1. Let S, T be positive contractions of L1, and Z be an order continuous lattice homomorphism of L1. Then one has (3.5) Moreover, we have (3.6) ZS − T = Z(S − T ). Z(S ∧ T ) = ZS ∧ ZT. Proof. From (3.2) we find that ZS − Tx = Z(sup{(S − T )y : = sup{Z(S − T )y : = Z(S − T )x, y ≤ x}) y ≤ x}) (3.7) for every x ∈ L1, x ≥ 0. The equality (3.3) yields that (3.8) ZS − T = ZS + ZT − 2Z(S ∧ T ), Z(S − T ) = ZS + ZT − 2(ZS ∧ ZT ), which with (3.7) imply that Z(S ∧ T ) = ZS ∧ ZT. (cid:3) In what follows, an order continuous lattice homomorphism Z : L1 → L1 Now we have the following with kZk ≤ 1, is called a lattice contraction. Lemma 3.2. Let Z be a lattice contraction and T be a positive contraction of L1 such that ZT = T Z. If for some m ∈ N ∪ {0}, k ∈ N one has kZ(T m+k − T m)k < 2, then kZ(T m+k − T m+k ∧ T m)k < 1. 8 FARRUKH MUKHAMEDOV Proof. According to the assumption there is δ > 0 such that kZ(T m+k−T m)k = 2(1 − δ). Let us suppose that kZ(T m+k − T m+k ∧ T m)k = 1. Then thanks to (2.2) there exists x ∈ L1 with x ≥ 0, kxk = 1 such that δ kZ(T m+k − T m+k ∧ T m)xk > 1 − 4 , which with (2.1) implies that kZT m+kxk > 1 − δ/4 and kZ(T m+k ∧ T m)xk < δ/4. The commutativity T and Z yields that kZT mxk > 1 − δ/4. Now using (3.8) and (3.6) one finds (cid:13)(cid:13)Z(T m+k − T m)x(cid:13)(cid:13) = kZT m+kxk + kZT mxk − 2kZ(T m+k ∧ T m)xk δ 4 − 2 · δ 4 δ 4 > 1 − = 2(cid:18)1 − + 1 − 2(cid:19). δ This with the equality (cid:13)(cid:13)Z(T m+k − T m)(cid:13)(cid:13) = kZ(T m+k − T m)k, contradicts to kZ(T m+k − T m)k = 2(1 − δ/2). Lemma 3.3. Let Z be a lattice contraction and T be a positive contraction of L1 such that ZT = T Z. If for some m ∈ N ∪ {0}, k ∈ N one has kZ(T m+k − T m+k ∧ T m)k < 1, then for any ε > 0 there are d, n0 ∈ N such that (cid:3) kZ d(T n+k − T n)k < ε for all n ≥ n0 Proof. It is known that (see [11], p. 310) for any contraction T on L1 there is γ > 0 such that (3.9) ≤ γ √ℓ . 2 (cid:19)ℓ(cid:13)(cid:13)(cid:13)(cid:13) − T(cid:18)I + T 2 (cid:19)ℓ(cid:13)(cid:13)(cid:13)(cid:13) − T k(cid:18) I + T ≤ Let ε > 0 and fix ℓ ∈ N such that kγ/√ℓ < ε/4. Then for given k ∈ N, using (3.9) one easily finds that kγ √ℓ 2 (cid:19)ℓ (cid:13)(cid:13)(cid:13)(cid:13) (cid:18) I + T 2 (cid:19)ℓ (cid:13)(cid:13)(cid:13)(cid:13) (cid:18)I + T . (3.10) Then according to Corollary 2.2 from the assumption of the lemma we have (3.11) (cid:13)(cid:13)Z(T ℓ(m+k) − (T m+k ∧ T m)ℓ)(cid:13)(cid:13) < 1. ON DOMINANT CONTRACTIONS 9 Hence, 2 (cid:19)ℓ (cid:13)(cid:13)(cid:13)(cid:13) Z(cid:18)T ℓ(m+k) − (cid:18)I + T Z(cid:18)T ℓ(m+k) − (T m+k ∧ T m)ℓ(cid:19)(cid:13)(cid:13)(cid:13)(cid:13) ℓ C i = ℓ C i ℓ 1 2ℓ Xi=0 ℓT i(T m+k ∧ T m)ℓ(cid:19)(cid:13)(cid:13)(cid:13)(cid:13) 2ℓ (cid:13)(cid:13)Z(T ℓ(m+k) − T i(T m+k ∧ T m)ℓ)(cid:13)(cid:13) Xi=0 = (cid:13)(cid:13)(cid:13)(cid:13) Xi=0 2ℓ(cid:13)(cid:13)Z(T ℓ(m+k) − (T m+k ∧ T m)ℓ)(cid:13)(cid:13) + 1 2ℓ + C i ℓ 2ℓ = 1. C i ℓ 2ℓ ≤ ≤ < 1 ℓ (3.12) Define ℓ Xi=1 2 (cid:19)ℓ Qℓ := T ℓ(m+k) −(cid:18)I + T (T m+k ∧ T m)ℓ and put V (1) ℓ = (T m+k ∧ T m)ℓ. Then one can see that 2 (cid:19)ℓ T ℓ(m+k) = (cid:18)I + T V (1) ℓ + Qℓ. Now for every d ∈ N, define V (d+1) ℓ = T ℓ(m+k)V (d) ℓ + V (1) ℓ Qd ℓ . Then by induction one can establish [11] that (3.13) 2 (cid:19)ℓ T dℓ(m+k) = (cid:18)I + T V (d) ℓ + Qd ℓ for every d ∈ N. Due to Proposition 2.1 [10] one has (3.14) kV (d) ℓ k ≤ 2 for all d ∈ N. Now from (3.12) we find kZQℓk < 1, therefore there exists d ∈ N such that k(ZQℓ)dk < ε/4. So, commutativity Z and T implies that ZQℓ = QℓZ, which yields that kZ dQd ℓk < ε/4. 10 FARRUKH MUKHAMEDOV Put n0 = dℓ(m + k), then from (3.13) with (3.10),(3.14) we get kZ d(T n0+k − T n0)k = (cid:13)(cid:13)(cid:13)(cid:13) ≤ (cid:13)(cid:13)(cid:13)(cid:13) +Z d(T kQd 2 (cid:19)ℓ Z d(cid:18)T k(cid:18)I + T ℓ )(cid:13)(cid:13)(cid:13)(cid:13) ℓ − Qd 2 (cid:19)ℓ (cid:18)T k(cid:18) I + T −(cid:18) I + T +kZ dQd ℓ (T − 1)k + 2 · < ε. kγ √ℓ ε 4 ≤ 2 · −(cid:18)I + T 2 (cid:19)ℓ(cid:19)V (d) ℓ 2 (cid:19)ℓ(cid:19)V (d) ℓ (cid:13)(cid:13)(cid:13)(cid:13) Take any n ≥ n0, then from the last inequality one finds kZ d(T n+k − T n)k = kT n−n0Z d(T n0+k − T n0)k ≤ kZ d(T n0+k − T n0)k < ε which completes the proof. (cid:3) Now we are ready to formulate the main result of this section. Theorem 3.4. Let Z, T be two positive contractions of L1 such that T Z = ZT . If for some m ∈ N ∪ {0}, k ∈ N one has kZ(T m+k − T m)k < 2, then for any ε > 0 there are d, n0 ∈ N such that kZ d(T n+k − T n)k < ε for all n ≥ n0 The proof of this theorem immediately follows from Lemmas 3.2 and 3.3. Remark. Note that if we take as Z = I,k = 1 then we obtain Theorem 1.1 as a corollary of Theorem 3.4. References [1] C.D. Aliprantis, O. Burkinshaw, Positive operators, Springer, 2006. [2] Y. Derriennic, Lois "z´ero ou deux" pour les processes de Markov. Applications aux marches al´eatoires. Ann. Inst. H. Poincar´e Sec. B 12(1976), 111 -- 129. [3] E.Ya. Eemel'yanov, Non-spectral asymptotic analysis of oneparameter operator semi- groups. In: Operator Theory: Advances and Applications, 173. Basel. Birkhauser Ver- lag. 2007. [4] S. R. Foguel, On the zero-two law, Israel J. Math. 10(1971), 275 -- 280. [5] S. R. Foguel, A generalized 0-2 law, Israel J. Math. 45(1983), 219 -- 224. [6] A.G. Kusraev, Dominated operators. In: Mathematics and its Applications, V. 519. Dordrecht: Kluwer Academic Publishers 2000. [7] M. Lin, On the 0 -- 2 law for conservative Markov operators, Z. Wahrscheinlichkeitstheor. Verw. Geb. 61 (1982), 513 -- 525. [8] F. Mukhamedov, S. Temir, H. Akin, A note on dominant contractions of Jordan alge- bras, Turk. J. Math. 34,(2010), 85 -- 93. [9] D. Ornstein, L. Sucheston, An operator theorem on L1 convergence to zero with appli- cations to Markov operators, Ann. Math. Statist. 41(1970) 1631 -- 1639. ON DOMINANT CONTRACTIONS 11 [10] R. Zaharopol, On the 'zero-two' law for positive contraction, Proc. Edin. Math. Soc. 32,(1989), 363 -- 370. [11] R. Zaharopol, The modulus of a regular linear operator and the 'zero-two' law in Lp- spaces (1 < p < +∞, p 6= 2), Jour. Funct. Anal. 68,(1986), 300 -- 312. Farrukh Mukhamedov, Department of Computational & Theoretical Sci- ences, Faculty of Science, International Islamic University Malaysia, P.O. Box, 141, 25710, Kuantan, Pahang, Malaysia E-mail address: [email protected] farrukh [email protected]
1609.08612	1	1609	2016-09-27T15:40:05	Representations of $p$-convolution algebras on $L^q$-spaces	[ "math.FA", "math.OA" ]	For a nontrivial locally compact group $G$, and $p\in [1,\infty)$, consider the Banach algebras of $p$-pseudofunctions, $p$-pseudomeasures, $p$-convolvers, and the full group $L^p$-operator algebra. We show that these Banach algebras are operator algebras if and only if $p=2$. More generally, we show that for $q\in [1,\infty)$, these Banach algebras can be represented on an $L^q$-space if and only if one of the following holds: (a) $p=2$ and $G$ is abelian; or (b) $\|\frac 1p - \frac 12\|=\|\frac 1q - \frac 12\|$. This result can be interpreted as follows: for $p,q\in [1,\infty)$, the $L^p$- and $L^q$-representation theories of a group are incomparable, except in the trivial cases when they are equivalent. As an application, we show that, for distinct $p,q\in [1,\infty)$, if the $L^p$ and $L^q$ crossed products of a topological dynamical system are isomorphic, then $\frac 1p + \frac 1q=1$. In order to prove this, we study the following relevant aspects of $L^p$-crossed products: existence of approximate identities, duality with respect to $p$, and existence of canonical isometric maps from group algebras into their multiplier algebras.	math.FA	math	REPRESENTATIONS OF p-CONVOLUTION ALGEBRAS ON Lq-SPACES EUSEBIO GARDELLA AND HANNES THIEL Abstract. For a nontrivial locally compact group G, consider the Banach algebras of p-pseudofunctions, p-pseudomeasures, p-convolvers, and the full group Lp-operator algebra. We show that these Banach algebras are operator algebras if and only if p = 2. More generally, we show that for q ∈ [1, ∞), these Banach algebras can be represented on an Lq-space if and only if one (cid:12) of the following holds: (a) p = 2 and G is abelian; or (b) . (cid:12) (cid:12) This result can be interpreted as follows: for p, q ∈ [1, ∞), the Lp- and Lq- representation theories of a group are incomparable, except in the trivial cases when they are equivalent. = (cid:12) (cid:12) (cid:12) (cid:12) (cid:12) (cid:12) (cid:12) (cid:12) (cid:12) 1 2 1 2 1 q − 1 p − 1 p + 1 As an application, we show that, for distinct p, q ∈ [1, ∞), if the Lp and Lq crossed products of a topological dynamical system are isomorphic, then = 1. In order to prove this, we study the following relevant aspects of Lp-crossed products: existence of approximate identities, duality with respect to p, and existence of canonical isometric maps from group algebras into their multiplier algebras. q 1 v 2 1 6 8 0 . 9 0 6 1 : v i X r a Contents 1. Introduction 2. Preliminaries 3. F p(Z) is not isomorphic to F q(Z) 4. Nonrepresentability on Lq-spaces 5. An application to crossed products References 2 4 6 10 20 30 Date: 17 July 2018. 2010 Mathematics Subject Classification. Primary: 47L10, 43A15, Secondary: 43A65, 46E30. Key words and phrases. Locally compact group, algebra of p-pseudofunctions, algebra of p- pseudomeasures, contractive approximate identity, multiplier algebra, group amenability, crossed product. The first named author was partially supported by the D. K. Harrison Prize from the University of Oregon and by a Postdoctoral Research Fellowship from the Humboldt Foundation. The second named author was partially supported by the Deutsche Forschungsgemeinschaft (SFB 878). Part of this work was completed while the authors were taking part in the Research Program Classifi- cation of operator algebras, complexity, rigidity and dynamics, held at the Institut Mittag-Leffler, between January and April of 2016. We would like to thank the staff and organizers, and Søren Eilers in particular, for the hospitality and financial support. 1 2 EUSEBIO GARDELLA AND HANNES THIEL 1. Introduction double commutant of C ∗ λ(G) is the algebra F p We say that a Banach algebra is an operator algebra if it admits an isometric representation as bounded operators on a Hilbert space. Associated to any locally compact group G there are three fundamentally important operator algebras: its reduced group C ∗-algebra C ∗ λ(G), its full group C ∗-algebra C ∗(G), and its group von Neumann algebra L(G). These are, respectively, the Banach algebra generated by the left regular representation of G on L2(G); the universal C ∗-algebra with respect to unitary representations of G on Hilbert spaces; and the weak∗ closure (also called λ(G) in B(L2(G)). (We canonically identify B(L2(G)) with ultraweak closure) of C ∗ the dual of the projective tensor product L2(G)b⊗L2(G).) Equivalently, L(G) is the λ(G) in B(L2(G)). These operator algebras admit generalizations to representations of G on Lp- spaces, for p ∈ [1,∞). The analog of C ∗ λ (G) of p-pseudofunc- tions on G, introduced by Herz in [Her73] and originally denoted by P Fp(G). The analog of C ∗(G) is the full group Lp-operator algebra F p(G), defined by Phillips in [Phi13]. Finally, the von Neumann algebra L(G) has two analogs, at least for p 6= 1: the algebra P Mp(G) of p-pseudomeasures, which is the weak∗ closure of F p λ (G) in B(Lp(G)) (where we canonically identify B(Lp(G)) with the dual of the projective tensor product Lp(G)b⊗Lp(G)∗); and the algebra CVp(G) of p-convolvers, which is the double commutant of F p λ (G) in B(Lp(G)) (it is also the commutant of the right regular representation). Both P Mp(G) and CVp(G) were introduced in [Her73]. These objects, and related ones, have been studied by a number of authors in the last three decades. For instance, see [Cow98], [NR09], [Run05], [DS13], and the more recent papers [Phi13], [Phi14], [GT15b], and [GT14]. In [GT16a], it is shown that there is a certain quotient of F p(Z) which cannot be represented on an Lp-space (in fact, on any Lq-space for q ∈ [1,∞)), thus answering a 20-year-old question of Le Merdy. Despite the advances in the area, some basic questions remain open. One im- portant open problem is whether P Mp(G) = CVp(G) for all p ∈ (1,∞) and for all locally compact groups G. This is known to be true when p = 2 (both algebras agree with L(G)), essentially by the double commutant theorem. It is immediate that P Mp(G) ⊆ CVp(G) in general, while Herz showed in [Her73] that equality holds for all p if G is amenable, a result that was later generalized by Cowling in [Cow98] to groups with the approximation property. A less studied problem is the following. By universality of F p(G), there is a canonical contractive homomorphism κp : F p(G) → F p λ (G) with dense range. For p = 2, this map is known to be a quotient map, and for p = 1 it is an isomorphism regardless of G. On the other hand, we do not know if κp is also a quotient map for all other values of p. In fact, we do not even know whether κp is surjective. If this map is not necessarily surjective, can it be injective without the group being amenable? (By Theorem 3.7 in [GT15b], G is amenable if and only if κp is bijective for some (equivalently, for all) p ∈ (1,∞). This result was independently obtained by Phillips in [Phi13] and [Phi14], using different methods.) In this case, it would be interesting to describe precisely for what groups (and Holder exponents) the map κp is injective but not surjective. Questions of the nature described above would in principle be easier to tackle if the objects considered had a better understood structure, as is the case for operator algebras. Despite the fact that the Banach algebras F p(G), F p λ (G), P Mp(G) and REPRESENTATIONS OF p-CONVOLUTION ALGEBRAS ON Lq-SPACES 3 CVp(G) have natural representations as operators on an Lp-space, this by itself does not rule out having isometric representations on Hilbert spaces as well; see, for example, [BLM95]. It is therefore not a priori clear whether the Lp-analogs of group operator algebras can be isometrically represented on Hilbert spaces. In this paper, we settle this question negatively. Indeed, we show in Theorem 4.12 that for a nontrivial locally compact group G, and for p ∈ [1,∞) \ {2}, none of the algebras F p(G), F p λ (G), P Mp(G), or CVp(G) can be isometrically represented on a Hilbert space. This result generalizes Theorem 2.2 in [NR09], where Neufang and Runde assume that G is amenable and has a closed infinite abelian subgroup. More generally, for p, q ∈ [1,∞) we show that the algebras F p λ (G), F p(G), P Mp(G), or CVp(G) can be isometrically represented on an Lq-space if and only if one of the following holds: (1) p = 2 and G is abelian; or (2) (cid:12)(cid:12)(cid:12) 1 p − 1 2(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12) 1 q − 1 2(cid:12)(cid:12)(cid:12). (This is equivalent to either p = q or 1 p + 1 q = 1.) This result can be interpreted as asserting that the Lp- and Lq-representation theories of a nontrivial group are incomparable, whenever they are not "obviously" equivalent. As a consequence, it follows that if there is an isometric Banach algebra isomorphism F p λ (G) (or between full group algebras, pseudomeasures or convolvers) for distinct p, q ∈ [1,∞), then 1 q = 1; see Corollary 4.13. The converse also holds; see Theorem 4.11. λ (G) ∼= F q p + 1 (cid:12)(cid:12)(cid:12) 1 q − 1 2(cid:12)(cid:12)(cid:12) = (cid:12)(cid:12)(cid:12) 1 2(cid:12)(cid:12)(cid:12); see Corollary 5.11. Since we do not know whether an isomor- As an application, we show that, for p, q ∈ [1,∞), if the Lp- and Lq-crossed products of a topological dynamical system are isometrically isomorphic, then p − 1 phism F p(G, X, α) → F q(G, X, α) (or F p λ (G, X, α)) must necessar- ily respect the group action α, Corollary 4.13 is not enough to obtain the conclusion. This means that even if we are only interested in isomorphisms of crossed prod- ucts, we are forced to consider arbitrary representations of p-convolution algebras on Lq-spaces. In order to obtain these results, we need to develop the theory of Lp-crossed products further, and we do so by exploring the following fundamental aspects: existence of approximate identities, duality with respect to p, and existence of canonical isometric maps from group algebras into their multiplier algebras. λ (G, X, α) → F q We give an outline of the proof of our main result (Theorem 4.12). Let E be λ (G) → B(E) be an isometric representation. (Similar an Lq-space and let ϕ : F p arguments apply for the algebras F p(G), P Mp(G), or CVp(G).) Step 1. The case q = 2 is treated separately, since in this case one can show that F p λ (G) is a C ∗-algebra; see Theorem 4.6. So assume that q 6= 2. Step 2. Using results in [GT16c], we may assume that ϕ is non-degenerate. Step 3. Functoriality properties of F p λ with respect to subgroups ([GT14]) allow us to further reduce the problem to the case where G is a cyclic group (finite or infinite); see Lemma 4.10. Step 4. For cyclic G, we need to know that ϕ(F p λ (G)) is isometrically isomorphic to F q λ (G). This requires non-trivial results from [GT15a] on spectral con- figurations. (For instance, we use the fact that F q λ (Z) is the unique Lq- operator algebra generated by an invertible isometry whose Gelfand trans- form is not surjective.) We conclude that there is an isometric isomorphism λ (G) ∼= F q F p λ (G). 4 EUSEBIO GARDELLA AND HANNES THIEL Step 5. When G = Z, we show that (cid:12)(cid:12)(cid:12) 1 mentary computations. p − 1 2(cid:12)(cid:12)(cid:12) = (cid:12)(cid:12)(cid:12) 1 q − 1 2(cid:12)(cid:12)(cid:12) in Theorem 3.6 using ele- Step 6. When G = Zn, we use the existence, for 1 ≤ p ≤ q ≤ 2, of a canonical, contractive map γp,q : F p(G) → F q(G) with dense range (see Theorem 2.5), to show that the Gelfand transform of F p λ (G) is an isometric isomorphism, thus reducing the problem to the case in Step 1 above. one is only interested in the case (cid:12)(cid:12)(cid:12) 1 We wish to point out that the use of spectral configurations can be avoided if Indeed, this situation can be entirely dealt with the maps γp,q (see Proposition 4.1 or Theorem 4.14 for a similar argument, but in a different context). Also, the fact that we can assume representations to be non-degenerate (Step 2) is by no means obvious, and the paper [GT16c] grew out of our attempts to prove this. 2(cid:12)(cid:12)(cid:12) > (cid:12)(cid:12)(cid:12) 1 2(cid:12)(cid:12)(cid:12). p − 1 q − 1 n p + 1 1.1. Notation. We take N = {1, 2, . . .}. For n ∈ N and p ∈ [1,∞), we write ℓp in place of ℓp({1, . . . , n}), and we write ℓp in place of ℓp(Z). For a Banach space E, we write B(E) for the Banach algebra of bounded linear operators on E. For p ∈ (1,∞), we denote by p′ its conjugate (Holder) exponent, which is determined by the identity 1 p′ = 1. Consistently, for a Banach space E, we denote its dual space by E′, and for a linear map π : E → F between Banach spaces E and F , we denote by π′ : F ′ → E′ its transpose map. Locally compact groups are assumed to be Hausdorff, and will always be im- plicitly endowed with a (fixed) left Haar measure, which will be chosen to be the counting measure whenever the group is discrete. The left Haar measure of a locally compact group G will be denoted by µ, and we will denote by ν the right Haar measure on G determined by ν(U ) = µ(U −1) for all measurable sets U ⊆ G. The modular function of G will be denoted by ∆ : G → R+. We will repeatedly use the following identities: ZG valid for all f ∈ L1(G, µ). ∆(t−1)f (t−1) dµ(t) =ZG f (t−1) dν(t) =ZG f (t) dµ(t) =ZG ∆(t)f (t) dν(t), 2. Preliminaries In this section, we recall and collect the necessary definitions and theorems that will be used throughout the paper. The only thing in this section that is really new is Proposition 2.3. We begin by defining the main objects of study of this work. Definition 2.1. Let G be a locally compact group, and let p ∈ [1,∞). Denote by Repp(G) the class of all contractive representations of L1(G) on Lp-spaces. The full group Lp-operator algebra of G, denoted F p(G), is the completion of L1(G) in the norm given by kfkF p(G) = sup(cid:8)kπ(f )k : π ∈ Repp(G)(cid:9) for f ∈ L1(G). Denote by λp : L1(G) → B(Lp(G)) the left regular representation, which is given by λp(f )ξ = f ∗ ξ for f ∈ L1(G) and ξ ∈ Lp(G). The algebra of p-pseudofunctions REPRESENTATIONS OF p-CONVOLUTION ALGEBRAS ON Lq-SPACES 5 on G (sometimes also called reduced group Lp-operator algebra of G), here denoted F p λ (G), is the completion of L1(G) in the norm for f ∈ L1(G). kfkF p λ (G) = kλp(f )kB(Lp(G)) The algebra F p(G) has been defined in [Phi13] and [GT15b] as the completion of L1(G) with respect to non-degenerate, contractive representations on Lp-spaces. In the proposition below, we show that this distinction is irrelevant. We recall the following, which is a particular case of a result from [GT16c]. Theorem 2.2. ([GT16c]). Let A be a Banach algebra with a left contractive ap- proximate identity, let E be a reflexive Banach space, and let ϕ : A → B(E) be a contractive homomorphism. Denote by E0 the essential subspace of ϕ, this is, E0 = span ϕ(A)E. Then there exists a contractive idempotent e ∈ B(E) satisfying e(E) = E0. Proposition 2.3. Let p ∈ [1,∞) and let G be a locally compact group. If E is an Lp-space and π : L1(G) → B(E) is a contractive representation, then there exist an Lp-space F and a contractive, non-degenerate representation ϕ : L1(G) → B(F ) such that for all f ∈ L1(G). In particular, Definition 2.1 agrees with the definitions given in [Phi13] and [GT15b]. kπ(f )k ≤ kϕ(f )k Proof. We treat the case p = 1 first. In this case, we may take ϕ to be the left regular representation λ1. Indeed, let π : L1(G) → B(E) be a contractive representation on an L1-space, and let (aj )j∈J be a contractive approximate identity for L1(G). Then kπ(f )k ≤ kfk1 = lim j∈J kf ∗ ajk1 = lim j∈J kλ1(f )ajk1 ≤ kλ1(f )k, for all f ∈ L1(G), as desired. Assume now that p > 1. Let π : L1(G) → B(E) be a contractive representation on an Lp-space. Since E is reflexive and L1(G) has a contractive approximate identity, it follows from Theorem 2.2 that there exists a contractive projection e ∈ B(E) such that e(E) is the essential subspace E0 of π. Then E0 is an Lp-space by Theorem 6 in [Tza69]. Let ϕ : L1(G) → B(E0) be the restriction of π. Then ϕ is a non-degenerate, contractive representation, and it is clear that kπ(f )k = kϕ(f )k for all f ∈ L1(G). (cid:3) The following duality principle was established in Proposition 2.18 of [GT15b]. Proposition 2.4. Let G be a locally compact group, let p ∈ (1,∞), and let p′ be its conjugate exponent. Then the inversion map on G extends to a canonical isometric isomorphism F p(G) ∼= F p′ (G). The next result is a combination of Theorem 2.30 and Corollary 3.20 in [GT15b]. The case 2 ≤ q ≤ p < ∞ is obtained using duality, so we omit it. Theorem 2.5. Let G be a locally compact group. identity map on L1(G) extends to a contractive homomorphism If 1 ≤ p ≤ q ≤ 2, then the γp,q : F p(G) → F q(G) 6 EUSEBIO GARDELLA AND HANNES THIEL with dense range. In particular, kfkF q(G) ≤ kfkF p(G) for every f ∈ L1(G). If, moreover, G is amenable, then: (a) γp,q is injective. (b) γp,q is surjective if and only if G is finite. Remark 2.6. When p = 2, we have F p(G) = C ∗(G), the full group C ∗-algebra, and F p In particular, when G is λ(G), the reduced group C ∗-algebra. λ (G) = C ∗ abelian, then γp,2 : F p(G) → C ∗(G) ∼= C0(bG) is the Gelfand transform. It follows from universality of the norm of F p(G) that the identity map on L1(G) extends to a contractive homomorphism κp : F p(G) → F p λ (G) with dense range. The following is part of Theorem 3.7 in [GT15b]. This result has been independently obtained by Phillips, and will appear in [Phi14]. Theorem 2.7. Let G be a locally compact group, and let p ∈ (1,∞). The following are equivalent: (1) The group G is amenable. (2) The canonical map κp : F p(G) → F p In view of the above result, we will identify F p(G) and F p λ (G) is an (isometric) isomorphism. λ (G) in a canonical manner whenever G is amenable. 3. F p(Z) is not isomorphic to F q(Z) The main step in the proof of Theorem 4.12 is to use the functoriality properties studied in [GT14] to reduce the statement to the case where G is a cyclic group (finite or infinite). While the case of a finite cyclic group will be dealt with using spectral configurations introduced in [GT15a], the case of the infinite cyclic group has to be dealt with separately, and we do so in this section. Our goal here is to prove the following: for p, q ∈ [1,∞), there is an isometric isomorphism F p(Z) ∼= F q(Z) as Banach algebras if and only if (cid:12)(cid:12)(cid:12) 1 p − 1 2(cid:12)(cid:12)(cid:12) = (cid:12)(cid:12)(cid:12) 1 q − 1 2(cid:12)(cid:12)(cid:12) (which is equivalent to p and q being either equal or Holder conjugate); see Theorem 3.6. In order to prove this, we will reduce to the problem of ruling out the possibility of a canonical isomorphism for the cyclic group Z2. In this case, the norm of a certain element can be explicitly computed, yielding the result; see Proposition 3.3. It should be pointed out that the methods of this section do not seem to gen- eralize to other finite cyclic groups besides Z2. We will postpone dealing with such groups until the next section; see Theorem 4.12, and specifically the proof of Claim 2 in it. The approach we take here for the case of Z2 has the advantage of being completely elementary and it does not depend on results from other works (in particular, we do not need here anything about spectral configurations from [GT15a]). We begin by looking at the group Lp-operator algebra of a finite cyclic group, which by amenability coincides with the algebra of p-pseudofunctions (this is the presentation we actually use). REPRESENTATIONS OF p-CONVOLUTION ALGEBRAS ON Lq-SPACES 7 Example 3.1. Let n ∈ N and let p ∈ [1,∞). Then Lp-group algebra F p(Zn) is the Banach subalgebra of B(ℓp n) generated by the cyclic shift sn of order n sn =   0 1 0 . . . 1 0 . . . . . . 0 1   . (The algebra B(ℓp n) is just Mn with the Lp-operator norm.) It is easy to check that F p(Zn) is algebraically isomorphic to Cn, but the canonical embedding F p(Zn) → Mn is not as diagonal matrices. In fact, computing the norm of an element in F p(Zn) is challenging for p other than from 1 and 2, essentially because computing p-norms of matrices that are not diagonal is difficult. Indeed, let ωn = e n , and set 2πi If ξ = (ξ0, . . . , ξn−1) ∈ Cn ∼= F p(Zn), then its norm in F p(Zn) is   (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 1 ωn ω2 n ... 1 1 1 ... 1 ωn−1 n 1 ω2 n ω4 n ... ω2(n−1) n n n 1 ··· ωn−1 ··· ··· ω2(n−1) . . . ··· ω(n−1)2 ... n   ξ0 ξ1 . . . ξn−1   u−1 n .   (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)B(ℓp . n) un = 1 √n kξkF p(Zn) = un The matrix un is unitary (its inverse is the transpose of its conjugate), and hence kξkF 2(Zn) = kξk∞. Moreover, if 1 ≤ p ≤ q ≤ 2, then k · kF q(Zn) ≤ k · kF p(Zn) by Theorem 2.5. In particular, the norm k·kF p(Zn) always dominates the norm k·k∞. The automorphism group of F p(Zn) is not easy to describe when p 6= 2, since not every permutation of the coordinates of Cn is isometric. The next proposition asserts that the cyclic shift on Cn is isometric. Proposition 3.2. Let n ∈ N and let p ∈ [1,∞). Denote by τ : Cn → Cn the cyclic forward shift, this is, τ (ξ0, . . . , ξn−1) = (ξn−1, ξ0, . . . , ξn−2) for all (ξ0, . . . , ξn−1) ∈ Cn. Then τ induces an isometric isomorphism F p(Zn) → F p(Zn). Proof. We follow the notation from Example 3.1, except that we write u in place of un, and we write s in place of sn. For ξ ∈ Cn, let d(ξ) denote the diagonal n × n matrix with d(ξ)j,j = ξj for j = 0, . . . , n − 1. Denote by ρ : Cn → Mn the unital homomorphism given by ρ(ξ) = ud(ξ)u−1 for ξ ∈ Cn. Then kξkF p(Zn) = kρ(ξ)kB(ℓp n) = kud(ξ)u−1kB(ℓp n) 8 EUSEBIO GARDELLA AND HANNES THIEL Set ω = (1, ω1 for all ξ ∈ Cn. Given ξ ∈ Cn, one checks that n, . . . , ωn−1 n ) ∈ Cn, and denote by ω its (coordinate-wise) conjugate. d(τ (ξ)) = sd(ξ)s−1, us = d(ω)u, and s−1u−1 = u−1d(ω). It follows that kτ (ξ)kF p(Zn) = kud(τ (ξ))u−1kB(ℓp = kusd(ξ)s−1u−1kB(ℓp = kd(ω)ud(ξ)u−1d(ω)kB(ℓp n). n, we conclude that n) n) Since d(ω) and d(ω) are isometries of ℓp kτ (ξ)kF p(Zn) = kd(ω)uτ (ξ)u−1d(ω)kB(ℓp n) = kud(ξ)u−1kB(ℓp n) = kξkF p(Zn), as desired. (cid:3) The fact that F p(Z2) is isometrically isomorphic to F q(Z2) only in the "obvious" cases can be proved directly by computing the norm of a special element, as we show below. Proposition 3.3. Let p, q ∈ [1,∞). Then F p(Z2) is isometrically isomorphic, as Banach algebras, to F q(Z2) if and only if (cid:12)(cid:12)(cid:12) 1 p − 1 2(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12) 1 q − 1 2(cid:12)(cid:12)(cid:12). Proof. The "if" implication follows from Proposition 2.4. We proceed to show the "only if" implication. Given t ∈ [1,∞), we claim that 2. k(1, i)kF t(Z2) = 2 1 By Proposition 2.4, the quantity on the left-hand side remains unchanged if one replaces t with its conjugate exponent. Since the same holds for the quantity on the right-hand side, it follows that it is enough to prove the claim for t ∈ [1, 2]. Let a be the matrix Define a continuous function δ : [1, 2] → R by δ(t) = k(1, i)kF t(Z2) for t ∈ [1, 2]. − 1 t a = 1 2(cid:18)1 + i 1 − i 1 + i(cid:19) . 1 − i Then δ(t) = kakB(ℓt to compute, and we have δ(1) = 2 satisfy 2) for all t ∈ [1, 2]. The values of δ at t = 1 and t = 2 are easy 2 and δ(2) = 1. Fix t ∈ (1, 2) and let θ ∈ (0, 1) 1 t 1 − θ 1 θ 2 = + . 1 Using the Riesz-Thorin Interpolation Theorem in [1, 2], we conclude that δ(t) ≤ δ(1)1−θ · δ(2)θ = 2 1 2 ( 2 t −1) · 1 = 2 1 t − 1 2 . For the converse inequality, fix t ∈ [1, 2] and consider the vector ξ = ( 1 0 ) ∈ ℓt 2. 1−i ). We have 2 ( 1+i Then kξkt = 1 and ax = 1 1 − i(cid:19)(cid:13)(cid:13)(cid:13)(cid:13)t 2(cid:18)1 + i We conclude that (cid:13)(cid:13)(cid:13)(cid:13) 1 = 1 2 (1 + it + 1 − it) 1 t = 2( 1 t − 1 2 ). δ(t) = kakB(ℓt 2) ≥ kaxkt kξkt = 2( 1 t − 1 2 ). This shows that δ(t) = 2( 1 t − 1 2 ) for all t ∈ [1, 2], and the claim is proved. REPRESENTATIONS OF p-CONVOLUTION ALGEBRAS ON Lq-SPACES 9 Now let p, q ∈ [1,∞) and let ψ : F p(Z2) → F q(Z2) be an isometric isomorphism. Since ψ is an algebra isomorphism, we must have either ψ(x, y) = (x, y) or ψ(x, y) = (y, x) for all (x, y) ∈ C2. By Proposition 3.2, the flip (x, y) 7→ (y, x) is an isometric isomorphism of F q(Z2), so we may assume that ψ is the identity map on C2. It follows that k(1, i)kF p(Z2) = k(1, i)kF q(Z2), so (cid:12)(cid:12)(cid:12) 1 complete. p − 1 2(cid:12)(cid:12)(cid:12) = (cid:12)(cid:12)(cid:12) 1 q − 1 2(cid:12)(cid:12)(cid:12) and the proof is (cid:3) Remark 3.4. Adopt the notation from the proof above. Then the function δ attains the upper bound given by the Riesz-Thorin Interpolation Theorem, which is a rare situation. This fortunate coincidence makes the argument possible, but it is not clear to us how to generalize these computations to other cyclic groups, or even to Z3. However, knowing the result just for Z2 is enough to prove Theorem 3.6. The following is probably standard, but we have not been able to find a reference in the literature. Accordingly, we prove it here. Proposition 3.5. Let f : S1 → S1 be a homeomorphism. Then there exists ζ ∈ S1 such that f (−ζ) = −f (ζ). Proof. Regard f as a homeomorphism h : [0, 2π] → [0, 2π] with either h(0) = 0 (if f is orientation preserving) or h(0) = 2π (if f is orientation reversing). Without loss of generality, assume that h(0) = 0, and hence that h is strictly increasing with h(2π) = 2π. Let g : [0, π] → R be given by g(t) = h(t + π) − h(t) π for t ∈ [0, π]. We need to find s ∈ [0, π] with g(s) = 1. If h(π) = π, then g(π) = 1 and we are done. If h(π) > π, we have g(0) > 1, while g(π) < 1. Likewise, if h(π) < π, then g(0) < 1 and g(π) > 1. In either case, the conclusion then follows from the Mean Value Theorem. (cid:3) We are now ready to prove that for p, q ∈ [1,∞), the algebras F p(Z) and F q(Z) are (abstractly) isometrically isomorphic only in the trivial case when (cid:12)(cid:12)(cid:12) 1 (cid:12)(cid:12)(cid:12) 1 2(cid:12)(cid:12)(cid:12) = 2(cid:12)(cid:12)(cid:12). This result should be compared with part (2) of Corollary 3.20 in [GT15b], q − 1 here reproduced as Theorem 2.5, where only the canonical homomorphism is con- sidered. The strategy will be to use Proposition 3.5 to reduce to the case when the group is Z2, which is Proposition 3.3. The fact that the spectrum of F p(Z) is the circle is crucial in our proof, and we do not know how to generalize these methods to directly deal with, for example, Z2. Theorem 3.6. Let p, q ∈ [1,∞). Then F p(Z) and F q(Z) are isometrically isomor- p − 1 phic (as Banach algebras) if and only if (cid:12)(cid:12)(cid:12) 1 p − 1 2(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12) 1 q − 1 2(cid:12)(cid:12)(cid:12). Proof. The "if" implication follows from Proposition 2.4. Let us show the converse. By Proposition 3.13 in [GT15b], the maximal ideal spaces of F p(Z) and F q(Z) are canonically homeomorphic to S1. We let Γp : F p(Z) → C(S1) denote the Gelfand transform, which maps the canonical generating invertible isometry u ∈ F p(Z), associated to 1 ∈ Z, to the canonical inclusion ι : S1 → C. Let ϕ : F p(Z) → F q(Z) be an isometric isomorphism. Then ϕ induces a homeo- morphism f : S1 → S1 that maps z ∈ S1 to the unique point f (z) ∈ S1 satisfying evz ◦ ϕ = evf (z) : F p(Z) → C. 10 EUSEBIO GARDELLA AND HANNES THIEL Use Proposition 3.5 to choose ζ0 ∈ S1 such that f (−ζ0) = −f (ζ0). Denote by πp : F p(Z) → F p(Z2) and πq : F p(Z) → F q(Z2) the canonical homomorphisms associated with the surjective map Z → Z2, which are quotient maps by The- orem 2.5 in [GT14]. Let ωζ0 : F p(Z) → F p(Z) be the isometric isomorphism induced by multiplying the canonical generator u ∈ F p(Z) by ζ0. Analogously, let ωf (ζ0) : F q(Z) → F q(Z) be the isometric isomorphism induced by multiplying u ∈ F q(Z) by f (ζ0). Then the following diagram commutes: C(S1) f ∗ / C(S1) Γp Γq F p(Z) ωζ0 F p(Z) ϕ / F q(Z) ωf (ζ0 ) / F q(Z) πp F p(Z2) /❴❴❴❴❴❴❴❴❴❴❴❴❴❴ bψ πq F q(Z2). Define an isometric isomorphism ψ : F p(Z) → F q(Z) by ψ = ωf (ζ0) ◦ ϕ ◦ ω−1 . It follows that ψ induces an isometric One checks that ψ(ker(πp)) = ker(πq). isomorphism bψ : F p(Z2) → F q(Z2). By Proposition 3.3, this implies that(cid:12)(cid:12)(cid:12) 1 (cid:12)(cid:12)(cid:12) 1 2(cid:12)(cid:12)(cid:12), as desired. 4. Nonrepresentability on Lq-spaces q − 1 p − 1 ζ0 2(cid:12)(cid:12)(cid:12) = (cid:3) The goal of this section is to show that for a nontrivial locally compact group G and for p, q ∈ [1,∞) distinct, the Banach algebras F p(G), F p λ (G), P Mp(G) and CVp(G) can be represented on an Lq-space only in the trivial cases, namely if either p = 2 and G is abelian; or if p and q are Holder conjugate; see Theorem 4.12. The main step in the proof is to use the functoriality properties of these objects, studied in [GT14], to reduce the statement to the case where G is a cyclic group (finite or infinite). The case of a finite cyclic group will be dealt with using spectral configurations as in [GT15a], as well as the canonical maps γp,q : F p(G) → F q(G) constructed in Theorem 3.7 in [GT15b]. Proposition 4.1 below is our first preparatory result on representability of full group Lp-operator algebras on Lq-spaces. We need some notation first. Let G be a locally compact group, and denote by ∆ : G → R its modular function. For f ∈ L1(G), let f ♯ : G → C be given by f ♯(s) = ∆(s−1)f (s−1) for all s ∈ G. It is easy to check that the map ♯ : L1(G) → L1(G) is an anti-multiplicative isometric linear map of order two. Proposition 4.1. Let G be a locally compact group, and let p, q ∈ [1,∞) satisfy p − 1 space. 2(cid:12)(cid:12)(cid:12). Suppose that F p(G) is isometrically representable on an Lq- 2(cid:12)(cid:12)(cid:12) > (cid:12)(cid:12)(cid:12) 1 (1) If p, q ∈ [1, 2] or p, q ∈ [2,∞), then the identity map on L1(G) extends to an isometric isomorphism F p(G) ∼= F q(G). (2) If p ∈ [1, 2] and q ∈ [2,∞), or if q ∈ [1, 2] and p ∈ [2,∞), then the map ♯ on L1(G) induces, when composed with the transpose map B(Lp(G)) → B(Lp′ (G)), an isometric isomorphism F p(G) ∼= F q(G). q − 1 (cid:12)(cid:12)(cid:12) 1 / o o / O O O O / / REPRESENTATIONS OF p-CONVOLUTION ALGEBRAS ON Lq-SPACES 11 Proof. (1). The result is trivial when p = q. Suppose first that 1 ≤ p < q ≤ 2. Sup- pose that there exist an Lq-space E and an isometric representation ϕ : F p(G) → B(E). Let ιp : L1(G) → F p(G) be the canonical contractive inclusion with dense range. Then ψ = ϕ ◦ ιp : L1(G) → B(E) is a contractive representation of L1(G) on an Lq-space. Let f ∈ L1(G). Using Theorem 2.5 at the first step, we deduce that kfkF q(G) ≤ kfkF p(G) = kψ(f )kB(F ) ≤ kfkF q(G). Hence kfkF q(G) = kfkF p(G). Since f ∈ L1(G) is arbitrary, it follows that the identity on L1(G) extends to an isometric isomorphism F p(G) → F q(G), as desired. The case 2 ≤ q < p < ∞ follows from duality; see Proposition 2.4. (2). We can assume, without loss of generality, that 1 ≤ p ≤ 2 ≤ q < ∞. Denote by q′ ∈ [1, 2] the Holder conjugate exponent of q. By Proposition 2.4, the map ♯ : L1(G) → L1(G), when composed with the transpose map B(Lq(G)) → B(Lq′ (G). (The details are in the proof of Lemma 2.16 of [GT15b], the main point being that given a representation of G on an Lq-space, its dual representation is a representation of Gop on Lq′ which induces the same norm on L1(G). One composes this isometric anti- isomorphism with the map ♯ to get a multiplicative (isometric) isomorphism.) Since the identity map on L1(G) extends to an isometric isomorphism F p(G) → F q′ (G) by part (1), the result follows. (cid:3) (G)), extends to an isometric isomorphism F q(G) → F q′ Besides F p(G) and F p λ (G), the other two Banach algebras we will be concerned with, at least when p > 1, are the algebra P Mp(G) of p-pseudomeasures, and the algebra CVp(G) of p-convolvers. These are, respectively, the ultraweak closure, and the bicommutant, of F p λ (G) in B(Lp(G)). Algebras of pseudomeasures and of convolvers on groups have been thoroughly studied since their inception by Herz in the early 70's; see [Her73]. It is clear that P Mp(G) ⊆ CVp(G), and it is conjectured that they are equal for every locally compact group G and every Holder exponent p ∈ (1,∞). The conjecture is known to be true if p = 2, or if G is amenable ([Her73]), or, more generally, if G has the approximation property ([Cow98]). Our next goal is showing that these convolution algebras are never operator algebras when p 6= 2. We first need two results about C ∗-algebras which are interesting in their own right. The first one is well-known, and it follows, from example, from Theorem 10 of [Bon54]. Theorem 4.2. Let A be a C ∗-algebra, let B be a Banach algebra, and let ϕ : A → B be a contractive, injective homomorphism. Then ϕ is isometric. Remark 4.3. If in the theorem above ϕ is not assumed to be injective, then the conclusion is that it is a quotient map. On the other hand, we must assume that ϕ is contractive, and not merely continuous, for the result to hold; counterexamples are easy to construct. The result also fails for not necessarily self-adjoint operator algebras, even for uniform algebras. The next fact about C ∗-algebras is proved in [GT16c], and had not been noticed before, at least not in this generality. The main difficulty is dealing with non- degenerate representations, for which Theorem 2.2 is essential. 12 EUSEBIO GARDELLA AND HANNES THIEL Theorem 4.4. ([GT16c]) Let A be a C ∗-algebra. Then the following are equivalent: (1) A can be isometrically represented on an Lp-space, for some p ∈ [1,∞)\{2}; (2) A can be isometrically represented on an Lp-space, for all p ∈ [1,∞); (3) A is commutative (and hence A ∼= C0(X) for X = Max(A)). We also need to recall the following, whose proof can be found, for example, in [GT14]. For a Banach algebra A, we denote its left (respectively, right, two- sided) multiplier algebra by ML(A) (respectively, MR(A) and M (A)), and we write ιL : A → ML(A) (respectively, ιR : A → MR(A) and ι : A → M (A)) for the canon- ical inclusion. When confusion may arise, we write the product in M (A) with a dot. Theorem 4.5. Let A be a Banach algebra with a left (respectively, right, two-sided) contractive approximate identity. Let X be a Banach space and let ϕ : A → B(X) be a non-degenerate contractive representation. Then there exists a unique unital con- tractive homomorphism ψL : ML(A) → B(X) (respectively, ψR : MR(A) → B(X) and ψ : M (A) → B(X)) satisfying ψL ◦ ιL = ϕ (respectively, ψR ◦ ιR = ϕ and ψ ◦ ι = ϕ). Moreover, the map ψL is given by for all m ∈ ML(A), for all a ∈ A and for all ξ ∈ X. ψL(m)(ϕ(a)(ξ)) = ϕ(m · a)(ξ) The following theorem generalizes a result of Neufang and Runde, Theorem 2.2 in [NR09], where the authors assumed that the group in question was amenable and had a closed infinite abelian subgroup. Observe that for G = Z, Theorem 4.6 follows from the main result of [GT16a], since there it is shown that for p 6= 2, there exists a semisimple quotient AV of F p(Z) that is not representable on Lq for any 1 ≤ q < ∞. Indeed, if F p(Z) were an operator algebra, so would be any of its quotients. Recall that a unital commutative, semisimple operator algebra is isometrically isomorphic to a closed subalgebra of C(X) for a compact Hausdorff space X (that is, it is a uniform algebra). Hence, if F p(Z) were an operator algebra, then AV would be a uniform algebra, hence isometrically representable on an Lq- space for any q ∈ [1,∞). Theorem 4.6 will be used in the proof of Theorem 4.12, which gives a much more general result. In the proof below, for a locally compact group G and p ∈ [1,∞), and to emphasize the role played by G, we will denote by γG p,2 : F p(G) → C ∗(G) the map from Theorem 2.5. Theorem 4.6. Let G be a locally compact group and let p ∈ [1,∞). Then one of F p(G), F p λ (G), P Mp(G) or CVp(G) is an operator algebra if and only if either p = 2 or G is the trivial group. Proof. The "if" implication is obvious if G is the trivial group, since the associated Banach algebras are all C, while the statement is clear if p = 2. For the "only if" implication, it is clear that if either P Mp(G) or CVp(G) is an operator algebra, then so is F p λ (G), since there are isometric inclusions F p λ (G) ⊆ P Mp(G) ⊆ CVp(G). Assume now that F p(G) is an operator algebra and that p 6= 2. There is a canon- ical identification of F p(G) with C ∗(G), in view of Proposition 4.1 and Remark 2.6. Let κp : F p(G) = C ∗(G) → F p λ (G) denote the canonical contractive homomorphism REPRESENTATIONS OF p-CONVOLUTION ALGEBRAS ON Lq-SPACES 13 with dense range. It is well-known that the quotient C ∗(G)/ ker(κp) is a C ∗-algebra. The induced map cκp : C ∗(G)/ ker(κp) → F p λ (G) It is therefore enough to show the statement assuming that F p metrically isomorphic to C ∗(G)/ ker(κp). In particular, F p hence an operator algebra itself. is an injective, contractive homomorphism. Now, Theorem 4.2 shows that cκp is isometric. Since κp, and hence bκp, has dense range, it follows that F p algebra. Let H be a Hilbert space and let ϕ : F p representation. Claim: F p λ (G) is an operator λ (G) → B(H) be an isometric λ (G) is a C ∗-algebra. Note that A has a (two-sided) contractive approximate identity, since so does L1(G) and there is a contractive homomorphism ι : L1(G) → F p λ (G) is iso- λ (G) is a C ∗-algebra, and λ (G) with dense range. Moreover, since λ (G), η ∈ H} {ϕ(a)η : a ∈ F p is itself a Hilbert space, we may assume that the representation ϕ is non-degenerate. It follows from Theorem 4.5 that the algebra M (F p λ (G)) of multipliers on F p λ (G), is unitally representable on H. By Corollary 2.5 in [GT16b], there is a canonical isometric identification of the λ (G)) ⊆ B(F p multiplier algebra M (F p λ (G)) ⊆ B(F p λ (G)), with the Banach algebra C(F p λ (G)) = {x ∈ B(Lp(G)) : xa, ax ∈ F p λ (G) for all a ∈ F p λ (G)} ⊆ B(Lp(G)) λ (G). Denote by ψ : C(F p of centralizers of F p isometric representation. λ (G)) → B(H) the resulting unital, There is an obvious identification of G with a subgroup of the invertible isome- tries of C(F p λ (G)), given by letting a group element g ∈ G act on Lp(G) as the convolution operator with respect to the point mass measure δg. Now, for g ∈ G, set ug = ψ(δg). Then ug is an invertible isometry on H, that is, a uni- tary operator. Moreover, the map u : G → U(H), given by g 7→ ug, is easily seen to be a strongly-continuous unitary representation of G on H. The inte- grated form ρu : L1(G) → B(H) of u is therefore a contractive, non-degenerate ∗-homomorphism. Whence the subalgebra ρu(L1(G)) ⊆ B(H) is closed under the adjoint operation. Moreover, it is clear that the following diagram commutes: L1(G) ι ❍ ❍ ❍ ρu ❍ ❍ ❍ ❍ ❍ F p λ (G) ϕ / $❍ / B(H). It follows from Theorem 4.4 that F p We conclude that ϕ(F p B(H), that is, a C ∗-algebra. The claim is proved. mutative, and in particular F q [GT15b]. λ (G)), which equals ρu(L1(G)), is a closed ∗-subalgebra of λ (G) is commutative. Thus G is itself com- λ (G) = F q(G) for all q ∈ [1,∞), by Theorem 3.7 in p,2 : F p λ(G) is an isometric isomorphism by Proposition 4.1. p,2 is surjective implies that G is finite, by Theorem 2.5. Using that The fact that γG γG p,2 is isometric, we will show that G must be the trivial group. λ (G) → C ∗ The map γG Using finiteness of G, let g ∈ G be an element with maximum order. Set n = ord(g) ≥ 1, and let j : Zn ֒→ G be the group homomorphism determined $ 14 EUSEBIO GARDELLA AND HANNES THIEL by j(1) = g. By Proposition 2.3 in [GT14], there are natural isometric embeddings jp : F p λ(G). Naturality of the maps involved implies that the following diagram is commutative: λ (Zn) ֒→ F p λ(Zn) ֒→ C ∗ λ (G) and j2 : C ∗ F p λ (Zn) jp F p λ (G) γ Zn p,2 γG p,2 C ∗ λ(Zn) / C ∗ λ(G). j2 In particular, γ Zn really just the identity on Cn.) λ (Zn) → C ∗ p,2 : F p λ(Zn) is an isometric isomorphism. (This map is 2πi Set ω = e n ∈ S1. Using that p 6= 2 together with Proposition 2.8 in [GT15a] (see also the comments above it), we conclude that if x ∈ F p λ (Zn) is an invertible isometry, then there exist ζ ∈ S1 and k ∈ {0, . . . , n − 1} such that, under the algebraic identification of F p λ (Zn) with Cn, we have x =(cid:16)ζ, ζωk, . . . , ζωk(n−1)(cid:17) . In particular, if n > 1, then not every element in (S1)n ⊆ Cn has norm one in F p λ(Zn), we must have n = 1. By the λ (Zn). Since this certainly is the case in C ∗ choice of n, we conclude that G must be the trivial group, and the proof of the theorem is finished. (cid:3) In contrast with Theorem 4.6, we point out that some Lp-operator group al- gebras are contractively and isomorphically representable on Hilbert spaces. For example, for any finite group G, abelian or not, and for any p ∈ [1,∞), the map γp,2 : F p(G) → C ∗(G) from Theorem 2.5, is a contractive isomorphism. For a locally compact group G, we review the definitions of the Banach al- gebras M p(G) and M p λ(G) from [GT14]. Recall that M 1(G) is the (unital) Ba- nach algebra of finite complex Radon measures on G, and it can be identified with the multiplier algebra of L1(G). In particular, observe that the left regu- lar representation λp : L1(G) → B(Lp(G)) can be extended to a representation λp : M 1(G) → B(Lp(G)). Definition 4.7. Let G be a locally compact group, and let p ∈ [1,∞). Define M p(G) to be the completion of M 1(G) in the norm given by kµkM p(G) = sup(cid:8)kπ(µ)k : π : M 1(G) → B(Lp(ν)) is a contractive homomorphism(cid:9) for µ ∈ M 1(G). λ(G) is the completion of M 1(G) in the norm The algebra M p kµkM p λ (G) = kλp(µ)kB(Lp(G)) for µ ∈ M 1(G). Remark 4.8. For p ∈ [1,∞), one can show that there are natural (isometric) inclusions and λ(G) ⊆ M (F p Somewhat less easy is the existence of an inclusion M p see [GT14]. F p(G) ⊆ M p(G) ⊆ M (F p(G)) F p λ (G) ⊆ M p λ (G)). λ(G) ⊆ CVp(G) for p ∈ (1,∞); / / / REPRESENTATIONS OF p-CONVOLUTION ALGEBRAS ON Lq-SPACES 15 The following is a particular case of a result in [GT14]. Proposition 4.9. Let p ∈ [1,∞), let G be a locally compact group, and let H ⊆ G be an amenable subgroup. Then the inclusion H ֒→ G induces canonical isometric unital homomorphisms M p λ(H) → M p λ(G) and M p(H) → M p(G). λ (G), M p λ (H) → B(X). The way in which the above proposition will be used is through the following lemma. We state it and prove it in greater generality than needed here for use elsewhere. Lemma 4.10. Let p ∈ [1,∞) and let G be a non-trivial locally compact group. Denote by Ap(G) any of the following algebras: F p(G), M p(G), F p λ(G), If Ap(G) is isometri- P Mp(G) or CVp(G) (the last two in the case p > 1). cally and non-degenerately representable on a Banach space X, then there exist a cyclic subgroup H (finite or infinite) of G, and a unital isometric representation ψ : F p Proof. Choose a non-trivial element g ∈ G, and denote by H the (not necessarily closed) cyclic subgroup of G generated by g. Since either F p(G) or F p λ (G) is isometrically a subalgebra of Ap(G) (see Remark 4.8), we may assume, without loss of generality, that Ap(G) is either F p λ (G) or F p(G). Denote by Bp(G) either M p λ (G)) or M p(G) (if Ap(G) = F p(G)). By Proposition 4.9, and since H is amenable and discrete, the inclusion of H ⊆ G induces a canonical isometric embedding F p Let X be a Banach space and let ϕ : Ap(G) → B(X) be an isometric and non- degenerate representation. Note that Ap(G) has a contractive approximate identity, since so does L1(G) and there is a contractive homomorphism L1(G) → Ap(G) with dense image. By Theorem 4.5, ϕ can be extended to an isometric unital representation ψ : M (Ap(G)) → B(X). Since there are inclusions λ(G) (if Ap(G) = F p λ (H) ֒→ Bp(G). F p λ (H) ⊆ Bp(G) ⊆ M (Ap(G)), λ (H) is the desired isometric unital representation of F p λ (H) (cid:3) the restriction of ψ to F p on X. The next theorem will be needed in the proof of Theorem 4.12. Recall that the algebra P Mp(G) of p-pseudomeasures is the closure of F p λ (G) in B(Lp(G)) with respect to the weak∗ topology (also called the ultraweak topology) induced by the (G) given by the pairing determined on simple tensors by (canonical) identification of B(Lp(G)) with the dual of Lp(G)b⊗Lp′ ha, ξ ⊗ ηiB(Lp(G)),Lp(G) b⊗Lp′ (G) = haξ, ηiLp(G),Lp′ (G) (G). for all a ∈ B(Lp(G)), all ξ ∈ Lp(G) and all η ∈ Lp′ We denote by Gop the opposite group of G. With µ and ν as in subsection 1.1, the inversion map ι : (G, µ) → (Gop, ν) is in fact a measure-preserving group iso- morphism. In particular, the p-convolution algebras associated to G are canonically isometrically isomorphic to those associated to Gop. Theorem 4.11. Let G be a locally compact group, and let p ∈ (1,∞). Then there are canonical isometric isomorphisms F p(G) ∼= F p′ (G), M p(G) ∼= M p′ (G), F p λ (G) ∼= F p′ λ (G), 16 EUSEBIO GARDELLA AND HANNES THIEL M p λ(G) ∼= M p′ λ (G), P Mp(G) ∼= P Mp′(G) and CVp(G) ∼= CVp′ (G). Proof. For F p(G), this is Proposition 2.18 in [GT15b] (here recalled as Proposition 2.4), while the proof for M p(G) is analogous, using M 1(G) instead of L1(G). (Gop), M p For the reduced versions, we will show that there are canonical isometric iso- morphisms F p(G) ∼= F p′ λ (Gop), P Mp(G) ∼= P Mp′(Gop) and CVp(G) ∼= CVp′ (Gop), since, by the remarks preceding this lemma, this implies the result. From now on, and to minimize confusion, we will write (G, µ) and (Gop, ν) instead of G and Gop, to state explicitly which is the left Haar measure in each case. λ(G) ∼= M p′ Define an isometric anti-isomorphism φ : L1(G, µ) → L1(Gop, ν) by φ(f )(s) = ∆(s)f (s) for all f ∈ L1(G, µ) and all s ∈ G. Observe that the assignment ξ 7→ ξ ◦ ι defines an isometric isomorphism Lp′ (Gop, ν). We denote by ψ : B(Lp(G, µ)) → B(Lp′ (Gop, ν)) the map given by (G, µ) ∼= Lp′ ψ(a)(ξ)(s) = [a′(ξ ◦ ι)](s−1) for all a ∈ B(Lp(G, µ)), for all ξ ∈ Lp′ (Gop, µ) and for all s ∈ G. (In other words, ψ(a) is the transpose of a, once Lp′ (G, µ) is identified with Lp′ (Gop, ν) via the assignment ξ 7→ ξ ◦ ι.) It is easy to check that ψ is an isometric anti-isomorphism. We claim that the following diagram commutes: L1(G, µ) λG p φ ψ L1(Gop, ν) λGop p′ / B(Lp′ (Gop, ν)). B(Lp(G, µ)) Let f ∈ L1(G), let ξ ∈ Lp′ (Gop, ν) and let s ∈ G be given. Recall (see, for example, Proposition 3.5 in [GT15b]) that the transpose of λG p′ (φ(f ) ◦ ι). (The function φ(f ) ◦ ι ∈ L1(G, µ) is denoted by f ♯ in [GT15b].) Using this at the second step, we get p (f ) is λG ψ(λG p (f ))(ξ)(s) = [λG p (f )′(ξ ◦ ι)](s−1) ∆(t−1)f (t−1)ξ(st) dµ(t) = [(φ(f ) ◦ ι) ∗ (ξ ◦ ι)](s−1) =ZG =ZG f (t)ξ(st−1) dµ(t). On the other hand, λGop p′ which proves the claim. (φ(f ))(ξ)(s) = (φ(f ) ∗op ξ)(s) =ZG =ZG ∆(t)f (t)ξ(st−1) dν(t) f (t)ξ(st−1) dµ(t), / / / REPRESENTATIONS OF p-CONVOLUTION ALGEBRAS ON Lq-SPACES 17 It follows that there is a canonical isometric isomorphism F p λ (Gop), im- plemented by ψ. By taking double commutants, we deduce that ψ also implements an isomorphism between CVp(G) and CVp′ (G). Extending the above diagram to M 1(G, µ) and M 1(Gop, ν), we conclude that there is also a canonical isometric isomorphism M p λ (G) ∼= F p′ λ(G) ∼= M p′ λ (Gop). To show the result for the algebras of pseudomeasures, we need to identify the ul- (Gop, ν)). In order to do this, observe traweak topologies on B(Lp(G, µ)) and B(Lp′ that the canonical isometric identification (G, µ) ∼= Lp′ Lp(G, µ)b⊗Lp′ (G, µ)b⊗Lp(G, µ) induces the isometric isomorphism ψ : B(Lp(G, µ)) → B(Lp′ (G, µ)). Commutativity of the above diagram then implies that φ extends to an isometric isomorphism between the ultraweak closures of L1(G, µ) and L1(Gop, ν), that is, to an isometric isomorphism P Mp(G) ∼= P Mp′ (Gop). (cid:3) The folowing is one of the main results of this paper. It determines precisely when one of the p-convolution algebras considered in the literature can be represented on an Lq-space, for some q ∈ [1,∞). It can be interpreted as stating that the Lp- and Lq-representation theories of a nontrivial group, are incomparable whenever they are not "obviously" equivalent. We point out that this represents a significant generalization of Theorem 2.2 in [NR09], where the authors only deal with the case q = 2, and moreover assume that G is amenable and has a closed infinite abelian subgroup. The proof of Theorem 4.12 uses machinery from a number of other works. The reader is referred to the third page in the introduction for an overview of the proof. Theorem 4.12. Let G be a locally compact group, and let p, q ∈ [1,∞) with q > 1. Then one (or all) of F p(G), M p(G), F p λ(G), P Mp(G) or CVp(G) is isometrically representable on an Lq-space if and only if one of the following holds: λ (G), M p (1) G is the trivial group; 2 − 1 q(cid:12)(cid:12)(cid:12); or p(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12) 1 2 − 1 (3) p = 2 and G is abelian. (2) we have (cid:12)(cid:12)(cid:12) 1 space for any q ∈ [1,∞). On the other hand, the identity (cid:12)(cid:12)(cid:12) 1 Proof. We begin with the "if" implication. When G is the trivial group, all the Banach algebras in the statement are C, which is clearly representable on an Lq- equivalent to p and q being either equal or conjugate. The case p = q is trivial. If p and q are conjugate, then Theorem 4.11 shows that all of the algebras in the statement are (canonically) representable on Lq(G). Finally, if p = 2 and G is abelian, then all of the Banach algebras in the statement are commutative C ∗- algebras, and hence have the form C0(X) for some locally compact Hausdorff space X. It is then an easy exercise to check that for any q ∈ [1,∞) and for any such space X, there exist an Lq-space and an isometric representation of C0(X) on it. λ(G) ⊆ CVp(G), F p We turn to the "only if" implication. Since there are isometric inclusions F p λ (G) ⊆ M p λ (G) ⊆ P Mp(G) ⊆ CVp(G) and F p(G) ⊆ M p(G) (see Remark 4.8), we may assume that either F p λ (G) or F p(G) is representable on an Lq-space. By Proposition 2.4, and without loss of generality, we can assume that p, q ∈ [1, 2]. Suppose that p = 2 and p 6= q. Then F p λ(G) and λ (G) = C ∗ 2 − 1 p(cid:12)(cid:12)(cid:12) = (cid:12)(cid:12)(cid:12) 1 2 − 1 q(cid:12)(cid:12)(cid:12) is 18 EUSEBIO GARDELLA AND HANNES THIEL F p(G) = C ∗(G) are C ∗-algebras. By Theorem 4.4, these must be abelian C ∗- algebras, whence G itself must be abelian. Since the arguments for F p λ (G) and F p(G) are very similar, we will carry them out together until we have to distinguish the two cases. We therefore denote by Ap(G) either F p λ (G) or F p(G). Now suppose that p 6= 2. Then q 6= 2 by Theorem 4.6. Let E be an Lq-space and let ϕ : Ap(G) → B(E) be an isometric representation. Note that Ap(G) has a contractive approximate identity, since so does L1(G) and there is a contractive homomorphism L1(G) → Ap(G) with dense image. By Theorem 2.2, there exists a contractive idempotent e ∈ B(E) such that e(E) is the essential subspace E0 of ϕ. Then E0 is an Lq-space by Theorem 6 in [Tza69]. Denote by ϕ0 : Ap(G) → B(E0) the restriction of ϕ. Then ϕ0 is a non-degenerate isometric representation. By Lemma 4.10, there exist a cyclic subgroup H of G, and an isometric unital repre- sentation ψ : F p λ (H)) is isometrically isomorphic to F q λ (H) → B(E0). We claim that ψ(F p λ (H). We divide the proof into two cases. Case 1: H ∼= Z. Observe that ψ(F p λ (Z)) and F q λ (Z) are both Banach algebras generated by an invertible isometry of an Lq-space and its inverse. Since q 6= 2, by part (2) of Corollary 5.21 in [GT15a], F q λ (Z) is the unique, up to (isomet- ric) isomorphism, Banach algebra generated by an invertible isometry of an Lq- space and its inverse whose Gelfand transform is not an isomorphism (isometric or not). Since p 6= 2, the Gelfand transform of F p λ (Z) is not an isomorphism, so the same is true for ψ(F p λ (Z)). It thus follows that there is an isometric isomorphism ψ(F p λ (H)) ∼= F q Case 2: H ∼= Zr for r ≥ 2. By replacing H with a subgroup, we may assume that r is prime. Denote by v ∈ F p λ (Zr) the canonical invertible isometry generating F p λ (Zr), and set w = ψ(v), which is an invertible isometry of E0. Then λ (H), as desired. This proves the first case. sp(w) = sp(v) = {ζ ∈ S1 : ζ r = 1}. Thus, the only admissible spectral configurations σ = (σn)n∈N for w are • σ0 = sp(w) and σn = ∅ for n ≥ 1; or • σ0 = sp(w) = σr and σn = ∅ for n ≥ 1 with n 6= r. In the first case, there is an isometric isomorphism between F q(w, w−1) = ψ(F p λ (Zr)) and (Cr,k · k∞). This would then contradict part (5) of Theorem 3.5 in [GT15a], since p 6= 2. In the second case, by the definition of the norm k · kσ,q on F q(σ) ∼= F q(w, w−1), there is an isometric isomorphism F q λ (Zr)). This finishes the proof of the claim. For later use, we stress the fact that, in the case H = Zr, the isomorphism ψ(F p λ (H) is in fact canonical, in the sense that the spectral configurations of v and ψ(v) agree, and thus the isomorphism is induced by the identity on ℓ1(H). λ (Zr) → ψ(F p λ (H)) ∼= F q λ (H) ∼= We deduce from the claim that there is an isometric isomorphism F p λ (H). For H ∼= Z, Theorem 3.6 implies that p = q. For H ∼= Zr, we use the fact F q that the isomorphism can be chosen to be canonical to prove that p must equal q. So suppose that p 6= q and F p λ (Zr) canonically. Without loss of gener- ality, we may assume that 1 ≤ p < q < 2. We claim that the map γp,2 : F p λ (Zr) → λ(Zr) from Theorem 2.5 is an isometric isomorphism. In view of Theorem 3.7 in C ∗ λ (Zr) ∼= F q for all f ∈ ℓ1(Zr). θ ∈ (0, 1) satisfy 1 Let f ∈ ℓ1(Zr). Then kλ2(f )kB(ℓ2(Zr)) ≤ kλp(f )kB(ℓp(Zr)) by Theorem 2.5. Let 2 . By the Riesz-Thorin interpolation theorem, we have p + 1−θ q = θ kλq(f )kB(ℓq(Zr)) ≤ kλ2(f )kθ B(ℓ2(Zr))kλp(f )k1−θ B(ℓp(Zr)). Since kλp(f )kB(ℓp(Zr)) = kλq(f )kB(ℓq(Zr)), we conclude that B(ℓ2(Zr)), B(ℓp(Zr)) ≤ kλ2(f )kθ kλp(f )kθ Since γp,2 is an isometric isomorphism, F p(Zr) is an operator algebra. The result (cid:3) and hence kλp(f )kB(ℓp(Zr )) ≤ kλ2(f )kB(ℓ2(Zr)), as desired. This proves the claim. now follows from Theorem 4.6. Corollary 4.13. Let G be a locally compact group, and let p, q ∈ [1,∞). Then the following are equivalent: (1) There is an isometric isomorphism F p(G) ∼= F q(G); (2) There is an isometric isomorphism F p λ (G); (3) When p > 1, there is an isometric isomorphism P Mp(G) ∼= P Mq(G); (4) When p > 1, there is an isometric isomorphism CVp(G) ∼= CVq(G); λ (G) ∼= F q (5) (cid:12)(cid:12)(cid:12) 1 p − 1 2(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12) 1 q − 1 2(cid:12)(cid:12)(cid:12). REPRESENTATIONS OF p-CONVOLUTION ALGEBRAS ON Lq-SPACES 19 [GT15b], this is equivalent to showing that kλp(f )kB(ℓp(Zr)) = kλ2(f )kB(ℓ2(Zr)) λ (G) ∼= F q F p λ (H) In connection with [GT16b], we mention here that for the algebras of pseudofunc- tions, the result above can be improved to moreover allow isomorphisms between algebras associated to different groups, as follows: for locally compact groups G and H, and for p, q ∈ [1,∞) not both equal to 2, there is an isometric isomorphism if and only if (cid:12)(cid:12)(cid:12) 1 p − 1 2(cid:12)(cid:12)(cid:12) = (cid:12)(cid:12)(cid:12) 1 q − 1 2(cid:12)(cid:12)(cid:12) and G is isomorphic to H. (For discrete groups, λ .) the same result holds with P Mp or CVp instead of F p A variant of the techniques used to prove Theorem 4.12 can be used to rule out representability on certain QSLp-spaces. (Recall that a Banach space E is a QSLp- space if it is isometrically isomorphic to a subspace of a quotient of an Lp-space. Group representations on QSLp-spaces are studied, for example, in [NR09].) We are not able to use spectral configurations as in Theorem 4.12, since we do not have a description of the Banach algebra generated by an invertible isometry of a QSLp-space. Instead, we will use the fact (see Theorem 3.7 in [GT15b]) that F p λ (Zn) and F p λ (Z) are universal with respect to representations on QSLp-spaces, together with the maps γp,q, to obtain the result. The theorem below is only stated for p, q ∈ [1, 2], while the other exponents can be handled using duality. Theorem 4.14. Let G be a nontrivial locally compact group, and let p, q ∈ [1, 2]. λ(G), P Mp(G) and CVp(G) (1) If q ≤ p, then all of F p(G), M p(G), F p (2) If p < q, then none of F p(G), M p(G), F p can be isometrically represented on a QSLq-space. λ (G), M p λ(G), P Mp(G) and λ (G), M p CVp(G) can be isometrically represented on a QSLq-space. 20 EUSEBIO GARDELLA AND HANNES THIEL Proof. If q ≤ p, then every Lp-space is isomorphic to a closed subspace of an Lq- space. In particular, every Lp-space is a QSLq-space, and the result is immediate. Suppose that p < q. As in the proof of Theorem 4.12, it is enough to show that λ (G) are not representable on a QSLq-space. For convenience, we will λ(G), F p(G) and F p denote by Ap(G) one of these algebras, and by Bp(G) either M p(G) or M p depending on which one Ap(G) is denoting. Let E be a QSLq-space and let ϕ : Ap(G) → B(E) be an isometric representation. Since a subspace of a QSLq-space is obviously a QSLq-space, we can assume that ϕ is non-degenerate. By Lemma 4.10, there are a cyclic subgroup H of G and a unital, isometric representation ψ : F p λ (H) → B(E). Unlike in the proof of Theorem 4.12, we cannot really conclude that ψ(F p λ (H) directly. Instead, and using the notation from [GT15b], for f ∈ ℓ1(H), we have (explanations below): λ (H)) is isomorphic to F q kfkLp = kλp(f )kB(Lp(H)) = kψ(f )kB(E) ≤ kfkQSLq = kfkLq ≤ kfkLp. The first step is Theorem 3.7 in [GT15b] for F p(H) (the group H is amenable); the second one is the fact that ψ is isometric; the third one is the definition of the norm k·kQSLq; the fourth one is the fact (Theorem 3.7 in [GT15b]) that F p QS(H) = F p(H); and the last one is Theorem 2.5. It follows that the map γH λ (H) is an isometric isomorphism. In particular, F p λ (H) is representable on an Lq-space, which contradicts Theorem 4.12. The contradiction implies that Ap(G) cannot be isometrically represented on a QSLq-space, as desired. (cid:3) p,q : F p λ (H) → F q We close this section with a question. Theorem 4.12 asserts that a group C ∗- algebra can be represented on an Lp-space, for some p 6= 2, if and only if the group is commutative. More generally, Theorem 4.4 shows that a C ∗-algebra can be represented on an Lp-space, for some p 6= 2, if and only if it is commutative. This may well be a particular case of a more general fact ruling out representability of certain Lp-operator algebras on Lq-spaces for two different, nonconjugate, Holder exponents p and q. On the other hand, the examples in [BLM95] show that there are operator algebras, which are not C ∗-algebras, that can be represented on Lp-spaces for p 6= 2. Therefore, we suggest: Question 4.15. Let A be a unital Banach algebra, and suppose that the set {v ∈ A : v is invertible and kvk = kv−1k = 1} has dense linear span in A. (This guarantees that A is a C ∗-algebra if it is an operator algebra.) Let p, q ∈ [1, 2] with p 6= q. Suppose that A can be isometrically represented on an Lp-space and on an Lq-space. Is A commutative? Does it follow that A ∼= C(X) for some compact Hausdorff space X? 5. An application to crossed products Let X be a locally compact Hausdorff space, let G be a locally compact group, and let α : G → Homeo(X) be a topological action. For p ∈ [1,∞), N. Christopher Phillips defined in [Phi13] the full and the reduced Lp-operator crossed products F p(G, X, α) and F p λ (G, X, α) of the topological dynamical system (G, X, α), gener- alizing the well established constructions in C ∗-algebras, which are the case p = 2. In Question 8.2 of [Phi13], Phillips asked whether, for a topological dynamical system (G, X, α) and for distinct p, q ∈ [1,∞), there are any non-zero continuous REPRESENTATIONS OF p-CONVOLUTION ALGEBRAS ON Lq-SPACES 21 homomorphisms F p(G, X, α) → F q(G, X, α) or F p λ (G, X, α) → F q λ (G, X, α). While Theorem 2.5 shows that there may in general exist such homomorphisms (even contractive ones with dense range), in Corollary 5.11 we show that there is an isometric isomorphism F p(G, X, α) → F q(G, X, α) or F p λ (Gop, X, α) λ (G, X, α) → F q if and only if (cid:12)(cid:12)(cid:12) 1 p − 1 2(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12) 1 q − 1 2(cid:12)(cid:12)(cid:12). For the convenience of the reader, and since our notation is somewhat different, we recall below the definitions of the crossed products (for Lp-operator algebras other than C0(X)). Recall that for p ∈ [1,∞), we say that a Banach algebra A is an Lp-operator algebra if it can be isometrically represented on an Lp-space. The group of isometric automorphisms of A is denoted Aut(A), and is always endowed with the strong topology. The object we define next is the analog of the group algebra L1(G). All integrals are taken with respect to a fixed left Haar measure. Definition 5.1. Fix p ∈ [1,∞). Let α : G → Aut(A) be an action of a locally compact group G on a Banach algebra A. Denote by L1(G, A, α) the Banach algebra completion of the space of continuous compactly supported functions G → A with respect to the L1-norm, with product given by twisted convolution, that is, (a ∗ b)(g) =ZG a(h)αh(b(h−1g)) dh for a, b ∈ Cc(G, A, α) ⊆ L1(G, A, α) and g ∈ G. Next, we define (regular) covariant representations and the associated (reduced) crossed product. Definition 5.2. Adopt the notation from the previous definition. A (contractive) covariant representation of (G, A, α) on an Lp-space E is a pair (π, v) consisting of a nondegenerate contractive homomorphism π : A → B(E) and an isometric group representation v : G → B(E), satisfying the covariance condition vgπ(a)vg−1 = π(αg(a)) π(a(g))vg(ξ) dg for all g ∈ G and for all a ∈ A. Given such a covariant representation, its integrated form is the nondegenerate contractive homomorphism π ⋊ v : L1(G, A, α) → B(E) given by (π ⋊ v)(a)(ξ) =ZG for all a ∈ L1(G, A, α) and for all ξ ∈ E. space E0, its associated regular covariant representation is the pair (π, λE0 Lp(G) ⊗p E0 ∼= Lp(G, E0) given by Given a contractive nondegenerate representation π0 : A → B(E0) on an Lp- p ) on π(a)(ξ)(g) = π0(cid:0)αg−1 (a)(cid:1) (ξ(g)) and (λE0 for all a ∈ A, for all ξ ∈ Lp(G, E0), and for all g, h ∈ G. Denote by Repp(G, A, α) the class of all contractive covariant representations of (G, A, α) on Lp-spaces, and by RegRepp(G, A, α) the subclass of Repp(G, A, α) con- sisting of the regular covariant representations. The full crossed product F p(G, A, α) p )g(ξ)(h) = ξ(g−1h) 22 EUSEBIO GARDELLA AND HANNES THIEL and the reduced crossed product F p L1(G, A, α) in the following norms: λ (G, A, α) are defined as the completions of kakF p(G,A,α) = sup{k(π ⋊ v)(a)k : (π, v) ∈ Repp(G, A, α)} and kakF p λ (G,A,α) = sup(cid:8)(cid:13)(cid:13)(π ⋊ λE0 p )(a)(cid:13)(cid:13) : (π, λE0 p ) ∈ RegRepp(G, A, α)(cid:9) , for all a ∈ L1(G, A, α). Remark 5.3. In the context of the above definition, if A is an Lp-operator algebra, then so will be F p(G, A, α) and F p λ (G, A, α). Lp-operator crossed products generalize group Lp-operator algebras, since for the λ (G). one point space X = {∗} we have F p(G,∗, id) = F p(G) and F p Remark 5.4. Most results in [Phi13] assume that the algebra A is separable and that the group is second countable, and conclude that a number of Banach algebras can be represented on σ-finite Lp-spaces; see also Remark 1.18 in [Phi13]. (The purpose of using σ-finite measure spaces is to apply Lamperti's theorem [Lam58].) However, the theory can be developed without these countability assumptions, and in fact the arguments in [Phi13] go through in general, except that one gets an Lp- operator algebra that is not necessarily representable on a σ-finite Lp-space. This, in particular, applies to Remark 4.6 and Proposition 4.8 in [Phi13]. λ (G,∗, id) = F p We note that while Definition 5.2 assumes covariant representations to be non- degenerate, this assumption is unnecessary whenever A has a (left or right) contrac- tive approximate identity, as can be shown with essentially the same argument as in Proposition 2.3. In fact, our first preparatory result guarantees the existence of an approximate identity for full and reduced crossed products whenever the underlying algebra has one. Theorem 5.5. Let p ∈ [1,∞), let A be an Lp-operator algebra, let G be a lo- cally compact group, and let α : G → Aut(A) be an action. Consider the following statements. (1) A has a left (right, two-sided) contractive approximate identity; (2) F p(G, A, α) has a left (right, two-sided) contractive approximate identity; (3) F p λ (G, A, α) has a left (right, two-sided) contractive approximate identity. Then (1) implies (2), and (2) implies (3). Finally, if G is discrete, then also (3) implies (1), so they are all equivalent in this case. Proof. We prove the result for left contractive approximate identities; the proof for right contractive approximate identities is analogous. (1) implies (2). Let (aµ)µ∈Λ be a contractive approximate identity for A. Set U = {U ⊆ G open, such that e ∈ U and U is compact}. Given U ∈ U, let fU : G → [0,∞) be a continuous positive function with support contained in U satisfying RG fU (g)dg = 1. Define a partial order on U by setting U ≤ V if V ⊆ U , and give U × Λ the partial order given by (U, µ) ≤ (V, ν) if and only if U ≤ V and µ ≤ ν. Recall that there exists a canonical contractive homomorphism ι : L1(G, A, α) → F p(G, A, α) with dense range. Write k · k1 for REPRESENTATIONS OF p-CONVOLUTION ALGEBRAS ON Lq-SPACES 23 the norm on both L1(G) and L1(G, A, α), and k · k for the norm on both A and F p(G, A, α). For (U, µ) ∈ U × Λ, set x(U,µ) = fU aµ (pointwise product), which is an element in Cc(G, A, α) ⊆ L1(G, A, α). It is clear that kι(x(U,µ))k ≤ kfU aµk1 = kfUk1kaµk ≤ 1 left contractive approximate identity for F p(G, A, α), it is enough to show that for all (U, µ) ∈ U × Λ. Thus, in order to show that (cid:0)ι(x(U,µ))(cid:1)(U,µ)∈U ×Λ is a (cid:0)x(U,µ)(cid:1)(U,µ)∈U ×Λ is a left contractive approximate identity for L1(G, A, α). For this, it is enough to consider functions in L1(G) · A ⊆ L1(G, A, α), since these span a dense subalgebra. Let f ∈ L1(G) and a ∈ A be given. We claim that (cid:13)(cid:13)x(U,µ) ∗ (f a) − f a(cid:13)(cid:13) → 0 as (U, µ) → ∞. If either f or a is zero, then so is their product and there is nothing to show. Without loss of generality, we may assume that kfk1 = kak = 1. Let ε > 0. We make the following choices: (1) Using continuity of α, choose an open set U1 ⊆ G containing the unit of G such that kαg(a) − ak < ε/3 for all g ∈ U1. Since G is locally compact, we may assume that U1 has compact closure, so that U1 ∈ U. (2) Using continuity of the left regular representation on L1(G), choose an open set U2 ⊆ G containing the unit of G such thatRG(cid:12)(cid:12)f (g−1h) − f (h)(cid:12)(cid:12) dg < ε/3 for all g ∈ U2. Again, we may assume that U2 ∈ U. fU (g)kαg(a) − akdg + kfUk1kfk1kaµa − ak (3) Choose µ0 ∈ Λ such that kaµa − ak < ε/3 for all µ ≥ µ0. Set U0 = U1 ∩ U2 ∈ U. Given (U, µ) ≥ (U0, µ0), we have (cid:13)(cid:13)x(U,µ) ∗ (f a) − f a(cid:13)(cid:13) =ZG(cid:13)(cid:13)(cid:13)(cid:13) x(U,µ)(g)αg((f a)(g−1h))dg − f (h)a(cid:13)(cid:13)(cid:13)(cid:13) dh ZG =ZG(cid:13)(cid:13)(cid:13)(cid:13) fU (g)aµf (g−1h)αg(a)dg − f (h)a(cid:13)(cid:13)(cid:13)(cid:13) dh ZG =ZG(cid:13)(cid:13)(cid:13)(cid:13) fU (g)(cid:0)f (g−1h)aµαg(a) − f (h)a(cid:1) dg(cid:13)(cid:13)(cid:13)(cid:13) dh ZG ≤ZGZG fU (g)(cid:13)(cid:13)f (g−1h)aµαg(a) − f (h)a(cid:13)(cid:13) dgdh ≤ZGZG fU (g)(cid:13)(cid:13)f (g−1h)aµαg(a) − f (h)aµαg(a)(cid:13)(cid:13) dgdh +ZGZG +ZGZG ≤ kaµkkαg(a)kZGZU + kaµkkfk1ZU ≤ ε/3 + ε/3 + ε/3 = ε. fU (g)(cid:12)(cid:12)f (g−1h) − f (h)(cid:12)(cid:12) dhdg fU (g)kf (h)aµαg(a) − f (h)aµak dgdh fU (g)kf (h)aµa − f (h)ak dgdh We conclude that F p(G, A, α) has a left contractive approximate identity. 24 EUSEBIO GARDELLA AND HANNES THIEL (2) implies (3). This is immediate since there is a contractive homomorphism λ (G, A, α) with dense range; see Lemma 3.13 in [Phi13]. κp : F p(G, A, α) → F p Assume now that G is discrete, and let us prove that (3) implies (1). De- note by E : F p λ (G, A, α) → A the faithful conditional expectation constructed in Proposition 4.8 of [Phi13] (observe that separability is not necessary; see comments above). Let (xµ)µ∈Λ be a contractive left approximate identity for F p λ (G, A, α), and set yµ = E(xµ) ∈ A for µ ∈ Λ. Using Remark 4.6 in [Phi13] (again, separability is not needed there), we identify A with a subalgebra of F p λ (G, A, α). For a ∈ A, we have µ∈Λkyµa − ak = lim lim µ∈Λ kE(xµ)a − ak = lim µ∈Λ kE(xµa − a)k ≤ lim µ∈Λ kxµa − ak = 0. It follows that (yµ)µ∈Λ is an left approximate identity for A, and it is clear that kyµk ≤ 1 for all µ ∈ Λ. (cid:3) Even when G is discrete, the crossed products F p(G, A, α) and F p λ (G, A, α) from the proposition above may not contain F p(G) and F p λ (G) canonically; this happens only when A is unital. In general, however, they are canonically subalgebras of the multiplier algebras of the crossed products, as we show below. Recall our convention that the product in multiplier algebras is written with a dot. Theorem 5.6. Let p ∈ [1,∞), let A be an Lp-operator algebra with a left con- tractive approximate identity, let G be a locally compact group, and let α : G → Aut(A) be an action by isometric isomorphisms. For g ∈ G, define a linear map Lg : L1(G, A, α) → L1(G, A, α) by Lg(a)(h) = αg(a(g−1h)) for all a ∈ L1(G, A, α) and all h ∈ G. Then the assignment g 7→ Lg induces natural contractive homomorphisms ιG L : F p(G) → ML(F p(G, A, α)) and ιG L,λ : F p λ (G) → ML(F p λ (G, A, α)). When A has a right contractive approximate identity, for g ∈ G, the linear maps Rg : L1(G, A, α) → L1(G, A, α), given by Rg(a)(h) = a(hg) for all a ∈ L1(G, A, α) and all h ∈ H, define natural contractive homomoprhisms ιG R : F p(G) → MR(F p(G, A, α)) and ιG R,λ : F p λ (G) → MR(F p λ (G, A, α)). Finally, when A has a two-sided contractive approximate identity, then the above maps define natural contractive homomorphisms ιG : F p(G) → M (F p(G, A, α)) and ιG λ : F p λ (G) → M (F p λ (G, A, α)). Moreover, the maps ιG L,λ, ιG R,λ and ιG λ are isometric. Proof. We prove the theorem only for left approximate identities and left multiplier algebras, but analogous proofs apply to the right and two-sided versions. REPRESENTATIONS OF p-CONVOLUTION ALGEBRAS ON Lq-SPACES 25 We claim that Lg is a left multiplier on L1(G, A, α), that is, Lg(a∗ b) = Lg(a)∗ b for all a, b ∈ L1(G, A, α). Given a, b ∈ L1(G, A, α) and h ∈ G, we have Lg(a ∗ b)(h) = αg(cid:18)ZG a(k)αk(cid:0)b(k−1g−1h)(cid:1) dk(cid:19) αg(a(k))αgk(cid:0)b((gk)−1h)(cid:1) dk αg(a(g−1t))αt(cid:0)b(t−1h)(cid:1) dt =ZG =ZG = (Lg(a) ∗ b)(h), as desired. It is clear that g 7→ Lg defines a strongly continuous group homomor- phism L : G → ML(L1(G, A, α)). Claim 1: The homomorphism L extends to strongly continuous representations, ιG and ιG λ (G, A, α) by isometric left multipliers. Let g ∈ G be given. To show that Lg extends to isometric automorphisms ιG λ )g of F p(G, A, α) and F p g and λ (G, A, α), it is enough to show that for every covariant λ , of G on F p(G, A, α) and F p (ιG representation (π, v) of (G, A, α) on an Lp-space, one has k(π ⋊ v)(a)k = k(π ⋊ v)(Lg(a))k for all a ∈ L1(G, A, α). To prove this, let (π, v) be a covariant representation on an Lp-space E, and let a ∈ L1(G, A, α). For ξ ∈ E, we use the covariance identity π(αg(x)) = vgπ(x)vg−1 at the third step, to get (π ⋊ v)(Lg(a))(ξ) =ZG =ZG =ZG π(Lg(a)(h))vh(ξ) dh π(αg(a(g−1h)))vh(ξ) dh vgπ(a(g−1h))vg−1h(ξ) dh = vg(π ⋊ v)(a)(ξ). The fact that ϕg and (ιG Since vg is an isometry, it follows that k(π ⋊ v)(a)k = k(π ⋊ v)(Lg(a))k, as desired. L,λ)g are left multipliers follows immediately from the fact that Lg is a left multiplier, using an ε/3 argument. Finally, it is routine to check that the assignments g 7→ (ιG L,λ)g are strongly continuous actions of G. The claim is proved. L )g and g 7→ (ιG Since L1(G) is universal with respect to strongly continuous isometric actions L : L1(G) → ML(F p(G, A, α)) and λ (G, A, α)), which admit explicit descriptions as follows. For of G, there are contractive homomorphisms ιG L,λ : L1(G) → ML(F p ιG f ∈ L1(G) and a ∈ L1(G, A, α), we have L (f ) · a)(g) =ZG f (h)Lh(a)(g) dh =ZG (ιG f (h)αh(a(h−1g)) dh, for all g ∈ G. On the other hand, let κ : F p(G, A, α) → F p λ (G, A, α) denote the canonical contractive homomorphism with dense range. Since κ is the identity on L1(G, A, α), by Theorem 5.5 it maps a contractive left approximate identity of F p(G, A, α) to a contractive left approximate identity of F p λ (G, A, α). Thus there λ (G, A, α)), and we exists a unique unital extension eκ : ML(F p(G, A, α)) → ML(F p have ιG L . L,λ =eκ ◦ ιG 26 EUSEBIO GARDELLA AND HANNES THIEL It remains to show that ιG L,λ is isometric when L1(G) is endowed with the norm of F p λ (G). We begin with some general observations. Since F p(G, A, α) and F p λ (G, A, α) have left contractive approximate identities by Theorem 5.5, it follows from Theorem 4.5 that any contractive, nondegenerate representation of any these algebras extends to a contractive, unital representation of its left multiplier alge- bra. Let (π, v) be a covariant representation of (G, A, α) on an Lp-space E. By applying Theorem 2.2, we can assume that π (and hence π ⋊ v) is nondegenerate. We write ^π ⋊ v : ML(F p(G, A, α)) → B(E) for the extension of π ⋊ v to the left multiplier algebra, and similarly with ^π ⋊ v : ML(F p λ (G, A, α)) → B(E) if (π, v) is a regular covariant representation. By a slight abuse of notation, we denote also by v : L1(G) → B(E) the integrated form of v, which is given by v(f )ξ =RG f (g)vg(ξ)dg for f ∈ L1(G) and ξ ∈ E. Let f ∈ L1(G), let a ∈ L1(G, A, α), and let ξ ∈ E. Then L (f ) · a](ξ) L (f ) · a)(g)(cid:3) vg(ξ) dg f (h)αh(a(h−1g)) dh(cid:21) vg(ξ) dg π(cid:2)αh(a(h−1g))(cid:3) vg(ξ) dgdh L (f )) = v(f ) for all f ∈ L1(G). L (f )) [(π ⋊ v)(a)ξ] = (π ⋊ v)[ιG Claim 2: We have ^π ⋊ v(ιG (^π ⋊ v)(ιG π(cid:2)(ιG π(cid:20)ZG f (h)ZG f (h)ZG f (h)ZG =ZG =ZG =ZG =ZG =ZG =ZG π [αh(a(k))] vhk(ξ) dkdh vhπ [a(k)] vk(ξ) dkdh f (h)vh ((π ⋊ v)(a)ξ) dh = v(f )((π ⋊ v)(a)ξ), and the claim is proved. Now suppose that (π, v) is a regular covariant representation, so that there exists p ; see Definition 5.2. Let f ∈ L1(G), and observe an Lp-space E0 such that v = λE0 that kλE0 p (f )k = kλp(f ) ⊗ idE0k = kλp(f )k. Using Claim 2 at the second step, and the above identity at the third step, we get kιG p ) ∈ RegRepp(G, A, α)} L,λ(f )kM(F p λ (G,A,α)) = sup{k(π ⋊ λE0 L (f ))k : (π, λE0 p )(ιG p (f )k : (π, λE0 p ) ∈ RegRepp(G, A, α)} We conclude that ιG proof. L,λ : F p λ (G, A, α)) is isometric. This finishes the (cid:3) When p = 1, and regardless of whether G is amenable or not, there is a canonical λ (G) (see Proposition 2.11 in [GT15b]), and hence the identification F 1(G) = F 1 = sup{kλE0 = kλp(f )k = kfkF p λ (G). λ (G) → ML(F p REPRESENTATIONS OF p-CONVOLUTION ALGEBRAS ON Lq-SPACES 27 L , ιG R and ιG from Theorem 5.6 are isometric (because they agree with the maps ιG ones defined on F 1 λ (G)). However, when p > 1, it is not in general true that the maps defined on F p(G) are isometric, or even injective with closed range, as we explain in the following example. Example 5.7. Let G be a discrete group, and let Lt denote the action of G on c0(G) by left translation. By Theorem 4.3 in [GT15b], the canonical map F p(G, G, Lt) → F p λ (G, G, Lt) is an isometric isomorphism, regardless of G. De- note by κp : F p(G) → F p λ (G) the canonical contractive map with dense range, and recall (Theorem 2.7) that κp is a (not necessarily isometric) isomorphism if and only if G is amenable. Naturality of the maps involved implies that the following diagram of unital homomorphisms commutes: F p(G) κp F p λ (G) ιG ιG λ / M (F p(G, G, Lt)) ∼= / M (F p λ (G, G, Lt)), where the vertical map on the right is the canonical one. Now, if ιG were injective and had closed range, that is, if it were an isomorphism onto its range, then it would follow that κp is an isomorphism, and hence that G is amenable. Since an identical reasoning applies to left or right multiplier algebras, we con- L , ιG R clude that for any non-amenable group G, and for the action Lt, the maps ιG and ιG are not isomorphisms onto their ranges. Let α : G → Aut(A) be an action of a locally compact group G on a Banach algebra A. We denote by Aop the opposite Banach algebra, and write αop : G → Aut(Aop) for the action given by αop g = αg for all g ∈ G. When A is abelian, then clearly A = Aop and α = αop. We note that in L1(G, Aop, αop), convolution is performed using opposite product on A, which we denote by ·op to minimize confusion. Recall that ∆ : G → R+ denotes the modular function (see subsection 1.1). Proposition 5.8. Let p ∈ (1,∞), let A be an Lp-operator algebra, let G be a locally compact group, and let α : G → Aut(A) be a continuous action. Then the map θ : L1(G, A, α) → L1(G, Aop, αop) given by θ(f )(s) = ∆(s−1)αs(f (s−1)), for all f ∈ L1(G, A, α) and all s ∈ G, is an isometric anti-isomorphism, which moreover extends to isometric anti-isomorphisms F p(G, A, α) ∼= F p′ (G, Aop, αop) and F p λ (G, X, α) ∼= F p′ λ (G, Aop, αop). In particular, F p(G, A, α) and F p Lp′ -spaces. λ (G, A, α) are anti-isometrically representable on Proof. Be begin by showing that the map θ from the statement is an isometric anti-isomorphism. That θ is isometric follows from the definition of ∆. For f, g ∈ L1(G, A, α) and s ∈ G, we have θ(f ∗ g)(s) = ∆(s−1)αs(f ∗ g)(s−1) = ∆(s−1)αs(cid:18)ZG f (t)αt(g(t−1s−1)) dt(cid:19) . / / 28 EUSEBIO GARDELLA AND HANNES THIEL On the other hand, in the next computation we set s−1t = k at the fourth step to get (θ(g) ∗ θ(f )) (s) =ZG =ZG = ∆(s−1)αs(cid:18)ZG = ∆(s−1)αs(cid:18)ZG t (cid:0)θ(f )(t−1s)(cid:1) dt θ(g)(t) ·op αop αt(∆(st−1)αt−1s(f (s−1t))∆(t−1)αt(g(t−1)) dt f (s−1t)αs−1t(g(t−1)) dt(cid:19) f (k)αk(g(k−1s−1)) dk(cid:19) , which proves the claim. Denote by ι : G → G the inversion map, which is anti-multiplicative. Let π and u be representations of A and G, respectively, on an Lp-space E. By abuse of no- tation, we denote by π′ : A → B(E′) and u′ : G → B(E′) the homomorphisms given by π′(a) = π(a)′ and u′ g = (ug)′ for all a ∈ A and g ∈ G. Then π′ and u′ ◦ ι are representations of Aop and G, respectively, on the Lp′ -space E′ (and conversely, by reflexivity). Moreover, for g ∈ G and a ∈ A, the identity ugπ(a)ug−1 = π(αg(a)) is equivalent to u′ g = π′(αg(a)). It follows that (π, u) is a covariant rep- resentation for (G, A, α) if and only if (π′, u′ ◦ ι) is a covariant representation for (G, Aop, αop). This shows that the assignment (π, u) 7→ (π′, u′ ◦ ι) induces a nat- ural bijection between the classes Repp(G, A, α) and Repp′ (G, Aop, αop). By the definition of the norm on the full crossed product, we conclude that g−1 π′(a)u′ kfkF p(G,A,α) = kθ(f )kF p′ (G,Aop,αop). for all f ∈ L1(G, A, α). This proves the statement for full crossed products. The case of reduced crossed products follows similarly: the above bijection re- stricts to a bijection between RegRepp(G, A, α) and RegRepp′ (G, Aop, αop), since for an Lp-space E0, the transpose of the representation λE0 (cid:3) p is λE ′ 0 p′ ◦ ι. Note that if a Banach algebra B is isometrically isomorphic to its opposite, then any Banach algebra completion of B is also isomorphic to its opposite. We will use this (trivial) observation in the next corollary, with B = L1(G, A, α) and the completions being the full and reduced crossed products. Corollary 5.9. Adopt the notation of Proposition 5.8, and suppose that A is abelian. Then L1(G, A, α) is canonically isometrically isomorphic to L1(G, A, α)op, and moreover there are natural isometric isomorphisms F p(G, A, α) ∼= F p′ (G, A, α) and F p λ (G, A, α) ∼= F p′ λ (G, A, α). Proof. Observe that A is also an Lp′ -operator algebra, since it is abelian. By the first part of Proposition 5.8, and since A is abelian, the map θ is a natural isometric isomorphism L1(G, A, α) ∼= L1(G, A, α)op. Upon taking completions with respect to all covariant representations of (G, A, α), we conclude that F p(G, A, α) is isometrically isomorphic to its opposite algebra. Composing this isomorphism with the isomorphism F p(G, A, α)op ∼= F p′ (G, A, α) given by Proposition 5.8, we obtain the desired isometric isomorphism for full crossed products. The case of reduced crossed products is identical: one completes with respect to (cid:3) regular covariant representations instead. REPRESENTATIONS OF p-CONVOLUTION ALGEBRAS ON Lq-SPACES 29 The following is the main result of this section. When X is the one point space, we recover Theorem 4.12. We point out that we do not know how to prove Theorem 5.10 directly without first obtaining some form of Theorem 4.12, and that Corollary 4.13 is not strong enough to deduce Theorem 5.10 from it. q − 1 q − 1 p − 1 p − 1 Theorem 5.10. Let X be a locally compact Hausdorff space, let G be a nontrivial locally compact group, and let α : G → Homeo(X) be a topological action. Given p, q ∈ [1,∞) with q > 1, the Banach algebras F p(G, X, α) and F p λ (G, X, α) can be isometrically represented on an Lq-space if and only if one of the following holds: λ (G, X, α) and F 2(G, X, α) are both (2) G is abelian, p = 2 and the action α is trivial. dual group of G. Then one easily checks that F 2 2(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12) 1 2(cid:12)(cid:12)(cid:12) = (cid:12)(cid:12)(cid:12) 1 2(cid:12)(cid:12)(cid:12); or (1) (cid:12)(cid:12)(cid:12) 1 Proof. If (cid:12)(cid:12)(cid:12) 1 2(cid:12)(cid:12)(cid:12), then the conclusion follows from Corollary 5.9. Sup- pose that p = 2, that G is abelian and that the action is trivial. Denote by bG the isometrically isomorphic to C0(X × bG), so they are representable on an Lq-space, for any q ∈ [1,∞). This proves the "if" implication. Let us show the converse. We treat the case of reduced crossed products first. Let p, q ∈ [1,∞) with q > 1, let E be an Lq-space and let ϕ : F p λ (G, X, α) → B(E) be an isomeric isomorphism. Since C0(X) has a (two-sided) contractive approx- imate identity, so does F p λ (G, X, α) by Theorem 5.5. Hence, upon restricting to its essential subspace and using Theorem 2.2, we may assume that ϕ is non- λ (G, X, α)) → B(E) be the extension of ϕ provided by Theorem 4.5. By Theorem 5.6, there exists a canonical isometric homomorphism λ : F p ιG λ is an isometric repre- sentation of F p q − 1 or p = 2 and G is abelian. Assuming the latter, F 2(G, X, α) is a C ∗-algebra. Suppose, without loss of generality, that q 6= 2. Then F 2(G, X, α), and thus its multiplier algebra M (F 2(G, X, α)), must be abelian by Theorem 4.4. For g ∈ G, let ug ∈ M (F 2(G, X, α)) denote the canonical unitary implementing αg. For a ∈ C0(X) ⊆ M (F 2(G, X, α)), we have λ (G) on an Lq-space. By Theorem 4.12, either (cid:12)(cid:12)(cid:12) 1 λ (G, X, α)). We conclude that eϕ ◦ ιG degenerate. Let eϕ : M (F p λ (G) → M (F p 2(cid:12)(cid:12)(cid:12) = (cid:12)(cid:12)(cid:12) 1 p − 1 2(cid:12)(cid:12)(cid:12) αg(a) = ugau∗ g = a. It follows that α is trivial, as desired. This shows the statement for F p λ (G, X, α). We prove the statement for full crossed products now. Let p, q ∈ [1,∞) with q > 1, let F be an Lq-space and let ψ : F p(G, X, α) → B(F ) be an isomeric iso- morphism. As before, we may assume that ψ is nondegenerate, and we denote by eψ : M (F p(G, X, α)) → B(F ) its unital extension. Let g ∈ G \ {1}, and denote by H ≤ G the (not necessarily closed) cyclic subgroup of G generated by g. Then H is amenable and there is a commutative diagram of contractive homomorphisms (explanations follow below) F p(H) ιH,G M (F p(G)) ❘ ❘ ❘ ❘ ❘ ❘ ❘ ❘ ιH,G λ ❘ ❘ ❘ ❘ (❘ M (F p λ (G)) f ιG f ιG λ / M (F p(G, X, α)) eκ / M (F p λ (G, X, α)). / / ( / / 30 EUSEBIO GARDELLA AND HANNES THIEL In the diagram above, ιH,G and ιH,G λ are the canonical isometric inclusions provided λ are the canonical unital extensions of map κ from full to reduced crossed product. Since ιG by Remark 4.8 and Proposition 4.9;fιG andfιG the maps constructed in Theorem 5.6; andeκ is the unital extension of the canonical isfιG λ . By commutativity of the diagram,eκ◦fιG ◦ ιH,G is an isometric representation of F p(H) on the Lq-space F . By Theorem 4.12, we must have either (cid:12)(cid:12)(cid:12) 1 2(cid:12)(cid:12)(cid:12) = p − 1 2(cid:12)(cid:12)(cid:12), or p = 2. Assume that p = 2 and q 6= 2. Then F p(G, X, α) is a C ∗-algebra, (cid:12)(cid:12)(cid:12) 1 q − 1 so it must be abelian by Theorem 4.4. Since C ∗ algebra M (F 2 argument used before shows that α must be trivial. This finishes the proof. λ(G) embeds into the abelian C ∗- λ (G, X, α)), the group G itself must be abelian. Finally, the same (cid:3) λ is isometric by Theorem 5.6, so Finally, the following corollary asserts that the Lp-crossed products obtained from topological dynamical systems, for varying p, are pairwise non-isometrically isomorphic, except for conjugate exponents. Corollary 5.11. Let X be a locally compact Hausdorff space, let G be a locally compact group, and let α : G → Homeo(X) be a topological action. Given p, q ∈ [1,∞), the following conditions are equivalent: (1) F p(G, X, α) is isometrically isomorphic to F q(G, X, α); (2) F p λ (G, X, α); λ (G, X, α) is isometrically isomorphic to F q p − 1 q − 1 (3) (cid:12)(cid:12)(cid:12) 1 2(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12) 1 2(cid:12)(cid:12)(cid:12). References [BLM95] D. Blecher and C. Le Merdy, On quotients of function algebras and operator algebra structures on ℓp, J. Operator Theory 34 (1995), no. 2, 315–346. [Bon54] F. Bonsall, A minimal property of the norm in some Banach algebras, J. London Math. Soc. 29(1954), 156–164. [Cow98] M. Cowling, The predual of the space of convolutors on a locally compact group, Bull. Austral. Math. Soc. 57 (1998), 409–414. [Dal00] G. Dales, Banach Algebras and Automatic Continuity, London Mathematical Society Monographs, New Series, no. 24, The Clarendon Press, Oxford University Press, New York, 2000. [DS13] M. Daws and N. Spronk, The approximation property implies that convolvers are pseudo- measures, Preprint. (arXiv:1308.1073 [math.FA]), 2013. [GT14] , Functoriality of group algebras acting on Lp-spaces, preprint, arXiv:1408.6137, 2014. [GT15a] E. Gardella and H. Thiel, Banach algebras generated by an invertible isometry of an Lp-space. J. Funct. Anal. 269 (2015), 1796–1839. [GT15b] , Group algebras acting on Lp-spaces, J. Fourier Anal. Appl. 21 (2015), 1310– 1343. [GT16a] , Quotients of Banach algebras acting on Lp-spaces, Adv. Math. 296 (2016), 85–92. [GT16b] [GT16c] , Isomorphisms of algebras of convolution operators, in preparation, 2016. , Extensions of representations of Banach algebras to biduals, in preparation, 2016. [Her73] C. Herz, Harmonic synthesis for subgroups, Ann. Inst. Fourier (Grenoble) 23 (1973), 91–123. [Lam58] J. Lamperti, On the isometries of certain function-spaces, Pacific J. Math. 8 (1958), 459–466. [NR09] M. Neufang and V. Runde, Column and row operator spaces over QSLp-spaces and their use in abstract harmonic analysis, preprint, arxiv:0711.2057, 2007. REPRESENTATIONS OF p-CONVOLUTION ALGEBRAS ON Lq-SPACES 31 [Phi13] N. C. Phillips, Crossed products of Lp operator algebras and the K-theory of Cuntz alge- bras on Lp spaces, preprint, arXiv:1309.6406, 2013. [Phi14] , Multiplicative domain for Lp operator algebras, in preparation. Draft of May 2014. [Run05] V. Runde, Representations of locally compact groups on QSLp-spaces and a p-analog of the Fourier-Stieltjes algebra, Pacific J. Math. 221 (2005), 379–397. [Spa12] P. Spain, Representations of C ∗-algebras in dual & right dual Banach algebras. Houston J. Math. 41 (2015), no. 1, 231–263. [Tza69] L. Tzafriri, Remarks on contractive projections in Lp-spaces. Israel J. Math. 7 1969 9–15. Eusebio Gardella Mathematisches Institut, Fachbereich Mathematik und Informatik der Universitat Munster, Einsteinstrasse 62, 48149 Munster, Germany. E-mail address: [email protected] URL: www.math.uni-muenster.de/u/gardella/ Hannes Thiel Mathematisches Institut, Fachbereich Mathematik und Informatik der Universitat Munster, Einsteinstrasse 62, 48149 Munster, Germany. E-mail address: [email protected] URL: www.math.uni-muenster.de/u/hannes.thiel/
1206.1764	2	1206	2012-06-26T16:44:22	Constructive Analysis in Infinitely many variables	[ "math.FA", "math-ph", "math-ph" ]	In this paper we investigate the foundations for analysis in infinitely-many (independent) variables. We give a topological approach to the construction of the regular $\s$-finite Kirtadze-Pantsulaia measure on $\R^\iy$ (the usual completion of the Yamasaki-Kharazishvili measure), which is an infinite dimensional version of the classical method of constructing Lebesgue measure on $\R^n$ (see \cite{YA1}, \cite{KH} and \cite{KP2}). First we show that von Neumann's theory of infinite tensor product Hilbert spaces already implies that a natural version of Lebesgue measure must exist on $\R^{\iy}$. Using this insight, we define the canonical version of $L^2[\R^{\iy}, \la_{\iy}]$, which allows us to construct Lebesgue measure on $\R^{\iy}$ and analogues of Lebesgue and Gaussian measure for every separable Banach space with a Schauder basis. When $\mcH$ is a Hilbert space and $\la_{\mcH}$ is Lebesgue measure restricted to $\mcH$, we define sums and products of unbounded operators and the Gaussian density for $L^2[\mcH, \la_{\mcH}]$. We show that the Fourier transform induces two different versions of the Pontryagin duality theory. An interesting new result is that the character group changes on infinite dimensional spaces when the Fourier transform is treated as an operator. Since our construction provides a complete $\s$-finite measure space, the abstract version of Fubini's theorem allows us to extend Young's inequality to every separable Banach space with a Schauder basis. We also give constructive examples of partial differential operators in infinitely many variables and briefly discuss the famous partial differential equation derived by Phillip Duncan Thompson \cite{PDT}, on infinite-dimensional phase space to represent an ensemble of randomly forced two-dimensional viscous flows.	math.FA	math	CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES TEPPER L. GILL, G. R. PANTSULAIA, AND W. W. ZACHARY* 2 1 0 2 n u J 6 2 ] . A F h t a m [ 2 v 4 6 7 1 . 6 0 2 1 : v i X r a 1991 Mathematics Subject Classification. Primary (45) Secondary(46) . Key words and phrases. infinite-dimensional Lebesgue measure, Gaussian measure, Fourier transforms, Banach space, Pontryagin duality theory, partial differential operators. *deceased. 1 2 GILL, PANTSULAIA, AND ZACHARY Abstract. In this paper we investigate the foundations for analysis in infinitely-many (independent) variables. We give a topological approach to the construction of the regular σ-finite Kirtadze-Pantsulaia measure on R∞ (the usual completion of the Yamasaki-Kharazishvili measure), which is an infinite dimensional version of the classical method of con- structing Lebesgue measure on Rn (see [YA1], [KH] and [KP2]). First we show that von Neumann's theory of infinite tensor product Hilbert spaces already implies that a natural version of Lebesgue measure must exist on R∞. Using this insight, we define the canonical version of L2[R∞, λ∞], which allows us to construct Lebesgue measure on R∞ and analogues of Lebesgue and Gaussian measure for every separable Ba- nach space with a Schauder basis. When H is a Hilbert space and λH is Lebesgue measure restricted to H, we define sums and products of unbounded operators and the Gaussian density for L2[H, λH]. We show that the Fourier transform induces two different versions of the Pon- tryagin duality theory. An interesting new result is that the character group changes on infinite dimensional spaces when the Fourier transform is treated as an operator. Since our construction provides a complete σ-finite measure space, the abstract version of Fubini's theorem allows us to extend Young's inequality to every separable Banach space with a Schauder basis. We also give constructive examples of partial dif- ferential operators in infinitely many variables and briefly discuss the famous partial differential equation derived by Phillip Duncan Thomp- son [PDT], on infinite-dimensional phase space to represent an ensemble of randomly forced two-dimensional viscous flows. CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES Contents Introduction Historical Background Purpose Summary 1. Why λ∞ Must Exist 2. Lebesgue Measure on R∞I 2.1. The Construction 2.2. The Extension to R∞I 2.3. Separable Banach Spaces 2.4. Translations 2.5. Gaussian measure 2.6. Rotational Invariance Discussion 3. Operators 3.1. Bounded Operators on H2 ⊗ 3.2. Unbounded Operators on H2 ⊗ 4. Function Spaces 4.1. L1-Theory 5. Fourier Transform Theory Background 3 4 6 10 10 11 14 14 18 25 28 29 32 35 35 35 38 43 44 46 46 4 GILL, PANTSULAIA, AND ZACHARY 5.1. Pontryagin Duality Theory I 5.2. L2-Theory 5.3. Pontryagin Duality Theory II 5.4. Lp-Theory 6. Partial Differential Operators (Examples) 6.1. Discussion 7. Conclusion Acknowledgments References 47 49 50 52 56 60 61 61 61 Introduction On finite-dimensional space it is useful to think of Lebesgue measure in terms of geometric objects (e.g.,volume, surface area, etc.). Thus, it is nat- ural to expect that this measure will leave these objects invariant under translations and rotations, so that rotational and translational invariance is an intrinsic property of Lebesgue measure. However, we then find ourselves disappointed when we try to use this property to help define Lebesgue mea- sure on R∞. A more fundamental problem is that the natural Borel algebra for R∞, B[R∞], does not allow an outer measure (since the measure of any open set is infinite). CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES 5 The lack of any definitive understanding of the cause for this lack of invari- ance on R∞ has led some researchers to believe that it is not possible to have a reasonable version of Lebesgue measure on R∞ (see, for example, DaPrato [DP] or Bakhtin and Mattingly [BM]). In many applications, the study of infinite dimensional analysis is restricted to separable Hilbert spaces, using Gaussian measure as a replacement for (the supposed nonexistent) Lebesgue measure. In some cases the Hilbert space structure arises as a natural state space for the modeling of systems. In other cases, both the Hilbert spaces and probability measures are imposed for mathematical convenience and are physically artificial and limiting. However, all reasonable models of in- finite dimensional (physical) systems require some functional constraint on the effects of all but a finite number of variables. Thus, what is needed, in general, is the imposition of constraints on the functions while preserving the modeling freedom associated with infinitely-many independent variables (in some well-defined sense). Any attempt to solve this problem necessarily implies a theory of Lebsegue measure on R∞. Even if a reasonable theory of Lebsegue measure on R∞ exists, this is not sufficient to make it useful in engineering and science. In addition, all the tools developed for finite-dimensional analysis, differential operators, Fourier transforms, etc are also required. Furthermore, researchers need operational control over the convergence properties of these tools. In par- ticular, one must be able to approximate an infinite-dimensional problem as 6 GILL, PANTSULAIA, AND ZACHARY a natural limit of the finite-dimensional case in a manner that lends itself to computational implementation. This implies that a useful approach also has a well-developed theory of convergence for infinite sums and products of unbounded linear operators. Historical Background Research into the general problem of Lebesgue measure on infinite- dimensional vector spaces and R∞ in particular, has a long and varied past, with participants living in a number of different countries, during times when scientific communication was constrained by war, isolation and/or na- tional competition. These conditions allowed quite a bit of misinformation and folklore to grow up around the subject, so that even experts may have a limited view of the history. Our own experience suggest that at least a brief survey of some important events is in order. (We do not claim com- pleteness and apologize in advance if we fail to mention equally important contributions.) Early studies in infinite dimensional analysis focused on the foundations of probability theory and had a broad base of participation. However, the major inputs were made by researchers in Poland, Russia, and France, with later contributions from the US. The first important advance of the general theory was made in 1933 when Haar [HA] proved the following theorem: CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES 7 Theorem 0.1. On every locally compact abelian group G there exists a non- negative regular measure m (Haar measure) on G, which is not identically zero and is translation invariant. That is, m(A+x) = m(A) for every x ∈ G and every Borel set A in G. This theorem stimulated interest in the subject and von Neumann [VN1] proved that it is the only locally finite left-invariant Borel measure on the group (uniqueness up to a mulitplicative constant). Weil [WE] developed an axiomatic approach to the subject, made a number of important refinements and, proved the "Inverse Weil theorem" (in moderm terms): Theorem 0.2. If G is a (separable) topological group and m is a transla- tion invariant Borel measure on G, then it is always possible to define an equivalent locally compact topology on G. In 1946, Oxtoby [OX] initiated the study of translation-invariant Borel measures on Polish groups (i.e., complete separable metric groups). In this paper, Oxtoby provides a proof of the following result which he attributes to Ulam: Theorem 0.3. Let G be any complete separable metric group which is not locally compact, and let m be any left-invariant Borel measure in G. Then ev- ery neighborhood contains an uncountable number of disjoint mutually con- gruent sets of equal finite positive measure. Stated another way, he proved that 8 GILL, PANTSULAIA, AND ZACHARY Theorem 0.4. There always exists a left-invariant Borel measure on any Polish group which assigns positive finite measure to at least one set and vanishes on singletons. However, a locally finite measure is possible if and only if the group is locally compact. (In 1967, Vershik [V] proved a related result for probability measures.) Apparently uninformed of Oxtoby's work, In 1959 Sudakov [SU] indepen- dently proved a special case of Theorem 0.4: If R∞ is regarded as a linear topological space, then there does not exist a σ-finite translation-invariant Borel measure for R∞. In 1964, Elliott and Morse [EM] developed a general theory of translation invariant product measures (non-σ-finite) and, in 1965, C. C. Moore [MO] initiated the study of measures that are translation in- variant with respect to vectors in R∞0 (i.e., the set of sequences that are zero except for a finite number of terms). This work was extended and refined by Hill [HI] in 1971. Motivated by Kakutani's work on infinite product measures [KA], a num- ber of young Japanese researchers entered the field. In 1973, Hamachi [HA] made major improvements on Hill's work which, indirectly suggested the problem of identifying the largest group T, of admissible translations in the sense of invariance for any σ-finite Borel measure µ on R∞ which assigns 2 , 1 2 ]ℵo and is metrically transitivity with respect to the value of one to [− 1 R∞0 (equivalently, for each A with µ(A) > 0, there is a sequence (hk) ∈ R∞0 such that µ(R∞ \ ∪∞k=1(A + hk)) = 0). CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES 9 Yamasaki [YA1] solved this problem in 1980. Unaware of the Yamasaki's proof, Kharazishvili independently solved the same problem in 1984. In 1991 Kir- tadze and Pantsulaia [KP1] provided yet another solution (see also Pantsu- laia [PA]). Finally, In 2007, Kirtadze and Pantsulaia proved that, if µ is the completion of the measure µ, then: (see [KP2]) Theorem 0.5. The measure µ is the unique regular σ-finite measure on R∞ (uniqueness up to a mulitplicative constant), which is assigns the value one 2 , 1 to the set [− 1 has the metrically transitivity property with respect to ℓ1. 2 ]ℵo, is invariant under translations from the group ℓ1 and In the mean time, in 1991 Baker [BA1], unaware of the Elliott-Morse measures, dropped the requirement that the measure be σ-finite and con- structed a translation invariant measure, ν, on R∞ (see also Baker (2004), [BA2]). In 1992, Ritter and Hewitt [RH] constructed a translation invariant measure related to that of Elliott Morse. Starting in 2007, A. M. Vershik (see [V1], [V2], [V3] and references con- tained therein) started an investigation of an infinite-dimensional analogue of Lebesgue measure that is constructed in a different manner than that studied in the previous papers. Roughly stated, he considers the weak limit as n → ∞ of invariant measures on certain homogeneous spaces (hypersur- faces of high dimension) of the Cartan subgroup of the Lie groups SL(n, R) (i.e., the subgroups of diagonal matrices with unit determinant). Vershik's 10 GILL, PANTSULAIA, AND ZACHARY measure is also unique and invariant under the multiplicative group of pos- itive functions, suggesting that a logarithmic transformation may lead to a version of the measure in this paper. (The paper of Vandev [VA] should also be consulted.) Purpose. The purpose of this paper is to show that a minor change in the way we represent R∞ makes it possible to construct a σ-finite regular version of Lebesgue measure using basic methods of measure theory from Rn. Since the measure is regular, it turns out to be the Kirtadze and Pantsulaia [KP1] measure, which is unique (see Theorem 0.5). Using our approach, we construct an analogue of both Lebesgue and Gaussian measure (countably additive) on every (classical) separable Banach space with a Schauder basis. The version of Gaussian measure constructed is also rotationally invariant (a property not shared by Wiener measure). This approach also allows us to satisfy all the requirements of a useful infinite dimensional theory. Summary. In the first section, we show how von Neumann's infinite ten- sor product Hilbert space theory implies that a natural version of Lebesgue measure must exist on R∞ and points to a possible approach. In the first part of Section 2, we show that a slight change in thinking about the cause for problems with unbounded measures on R∞ makes the construction of Lebesgue measure not only possible, but no more difficult then the same construction on Rn. (We denote it by R∞I , for reasons that are discussed in this section.) We also provide natural analogues of Lebesgue and Gaussian CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES 11 measure for every separable Banach space with a Schauder basis and show that ℓ1 is the maximal translation invariant subspace. In the last part of Section 2, we show that ℓ2 is the maximal rotation invariant subspace. In Section 3, we study the convergence properties of infinite sums and products of bounded and unbounded linear operators. In Section 4, we investigate some of the function spaces over R∞I and in Section 5, we discuss Fourier transforms and Pontryagin duality theory for Banach spaces. A major result is that there are two different extensions of the Pontrjagin Duality theory for infinite dimensional spaces. In this section, we also show that our the- ory allows us to extend Young's inequality to ever separable Banach space with a Schauder basis. In Section 6, we give some constructive examples of partial differential operators in infinitely many variables. This allows us to briefly discuss the famous partial differential equation derived by Phillip Duncan Thompson [PDT], on infinite-dimensional phase space to represent an ensemble of randomly forced two-dimensional viscous flows. 1. Why λ∞ Must Exist In order to see that some reasonable version of Lebesgue measure must exist, we need to review von Neumann's infinite tensor product Hilbert space theory [VN2]. To do this, we first define infinite products of complex numbers. (There are a number of other possibilities, see [GU] and [PA], pg. 272-274.) In order to avoid trivialities, we always assume that, in any product, all terms are nonzero. 12 GILL, PANTSULAIA, AND ZACHARY Definition 1.1. If {zi} is a sequence of complex numbers indexed by i ∈ N (the natural numbers), every ε > 0, there is a finite set J(ε) such that, for all finite sets (1) We say that the product Qi∈N zi is convergent with limit z if, for J ⊂ N, with J(ε) ⊂ J, we have (cid:12)(cid:12)Qi∈J zi − z(cid:12)(cid:12) < ε. (2) We say that the product Qi∈N zi is quasi-convergent if Qi∈N zi is convergent. (If the product is quasi-convergent, but not convergent, we assign it the value zero.) We note that (1.1) 0 <(cid:12)(cid:12)(cid:12)Yi∈N zi(cid:12)(cid:12)(cid:12) < ∞ if and only if Xi∈N 1 − zi < ∞. Let Hi = L2[R, λ] for each i ∈ N and let H2 tensor product of von Neumann. To see what this object looks like: ⊗ = ⊗∞i=1L2[R, λ] be the infinite Definition 1.2. Let g = ⊗i∈N gi and h = ⊗i∈N hi be in H2 ⊗. (1) We say that g is strongly equivalent to h (g ≡s h) if and only if Pi∈N 1 − hgi, hiii < ∞ . (2) We say that g is weakly equivalent to h (g ≡w h) if and only if Pi∈N 1 − hgi, hiii < ∞. Proofs of the following may be found in von Neumann [VN2] (see also [GZ], [GZ1]). CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES 13 Lemma 1.3. We have g ≡w h if and only if there exist zi, zi = 1, such that ⊗i∈N zigi ≡s ⊗i∈N hi. Theorem 1.4. The relations defined above are equivalence relations on ⊗, which decomposes H2 H2 spaces). ⊗ into disjoint equivalence classes (orthogonal sub- Definition 1.5. For g = ⊗i∈N ⊗(g) to be the closed subspace generated by the span of all h ≡s g and we call it the strong partial tensor product space generated by the vector g. (von Neumann called it an ⊗, we define H2 gi ∈ H2 incomplete tensor product space.) Theorem 1.6. For the partial tensor product spaces, we have the following: (1) If hi 6= gi occurs for at most a finite number of i, then h = ⊗i∈N hi ≡s gi. g = ⊗i∈N (2) The space H2 ⊗(g) is the closure of the linear span of h = ⊗i∈N hi such that hi 6= gi occurs for at most a finite number of i. (3) If g = ⊗i∈Ngi and h = ⊗i∈Nhi are in different equivalence classes of H2 ⊗, then (g, h)⊗ =Qi∈N hgi, hiii = 0. ⊗(g)w = ⊕h≡wg (cid:2)H2 (4) H2 (5) For each g, H2 (6) For each g, H2 ⊗(g)s is a separable Hilbert space. ⊗(g)w is not a separable Hilbert space. ⊗(h)s(cid:3) . It follows from (6) that H2 ⊗ = ⊗∞i=1L2[R, λ] is not a separable Hilbert space. 14 GILL, PANTSULAIA, AND ZACHARY From (5), we see that it is reasonable to define L2[R∞, λ∞] = H2 ⊗(h)s, for some h = ⊗∞i=1hi. This definition is ambiguous, but, in most applications, the particular version does not matter. To remove the ambiguity, we should identify a canonical version of h = ⊗∞i=1hi. Any reasonable version of λ∞ should satisfy λ∞(I0) = 1, where I = [−1 2 ] and I0 = ×∞i=1I. 2 , 1 Definition 1.7. If χI is the indicator function for I and hi = χI , we set h = ⊗∞i=1hi. We define the canonical version of L2[R∞, λ∞] = L2[R∞, λ∞](h)s. 2. Lebesgue Measure on R∞I 2.1. The Construction. We now have the problem of identifying the mea- sure space associated with L2[R∞, λ∞](h)s. In the historical approach to the construction of infinite products of measures {µk, k ∈ N} on R∞, the cho- sen topology defines open sets to be the (cartesian) product of an arbitrary finite number of open sets in R, while the remaining infinite number are copies of R (cylindrical sets). The success of Kolmogorov's work on the foundations of probability theory naturally led to the condition that µk(R) be finite for all but a finite number of k (see [KO]). Thus, any attempt to construct Lebesgue measure via this approach starts out a failure in the beginning. However, Kolmogorov's approach is not the only way to induce a total measure of one for the spaces under consideration. Our definition of the canonical version of L2[R∞, λ∞] offers another ap- proach. To see how, consider a simple extension of the theory on R. Let CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES 15 2 ] and define RI = R × I1, where I1 = ∞×i=2 2 , 1 I = [− 1 Borel σ-algebra for R, let B(RI ) be the Borel σ-algebra for RI. For each I. If B(R) is the set A ∈ B(R) with λ(A) < ∞, let AI be the corresponding set in B(RI ), AI = A × I1. We define λ∞(AI ) by: λ∞(AI ) = λ(A) × λ(I) = λ(A). ∞ Yi=2 We can construct a theory of Lebesgue measure on RI that completely par- allels that on R. This suggests that we use Lebesgue measure and replace the (tail end of the) infinite product of copies of R by infinite products of copies of I. The purpose of this section is to provide such a construction. Since we will be studying unbounded measures, for consistency, we use the following conventions: 0 · ∞ = 0 and 0 · ∞∞ = ∞. Recall that R∞ is the set of all x = (x1, x2,···), where xi ∈ R. This is a linear space which is not a Banach space. However, it is a complete metric space with metric given by: d(x, y) =X∞ n=1 1 2n xn − yn 1 + xn − yn . Remark 2.1. R∞ is a special case of a Polish space, which Banach called a Fr´echet space i.e., a Polish space with a translation invariant metric (see Banach [BA]). The topology generated by d(·,·) is generally known as the Tychonoff topology. For each n, define Rn I = Rn × In, where In = ∞×i=n+1 I. 16 GILL, PANTSULAIA, AND ZACHARY Definition 2.2. If An = A × In, Bn = B × In are any sets in Rn define: I , then we (1) An ∪ Bn = A ∪ B × In, (2) An ∩ Bn = A ∩ B × In, and (3) Bc n = Bc × In. In order to avoid confusion, we always assume that I0 = ×∞i=1I ⊂ R1 I via the following class of open sets: can now define the topology for Rn I . We On = {Un : Un = U × In, U open in Rn} . 2.1.1. Definition of R∞I . It is easy to see that Rn I ⊂ Rn+1 I . Since this is an increasing sequence, we can define R′∞I by: R′∞I = limn→∞Rn I = ∞∪k=1 Rk I . Let τ1 be the topology on R′∞I = X1 induced by the class of open sets O defined by: O = ∞ [n=1 On = ∞ [n=1 {Un : Un = U × In, U open in Rn}, and let τ2 be topology on R∞ \ R′∞I = X2 induced by the metric d2, for which d2(x, y) = 1, x 6= y and d2(x, y) = 0, x = y, for all x, y ∈ X2. Definition 2.3. We define (R∞I , τ ) to be the sum (X1, τ1) and (X2, τ2), so that every open set in (R∞I , τ ) is union of two disjoint sets G1 ∪ G2, where G1 is open in (X1, τ1) and G2 is open in (X2, τ2). CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES 17 It now follows from the above construction that R∞I = R∞ as sets. (How- ever, they are not equal as topological spaces.) The following result shows that convergence in the τ -topology always implies convergence in the Ty- chonoff topology. Theorem 2.4. If yk converges to x in the τ -topology, then yk converges to x in the Tychonoff topology. Proof. Case 1. If x ∈ R∞ \ R′∞I k > N . Indeed, for a neighborhood of diameter 1 then there is N such that yk = x for all 2 about x, there is a N such that d2(x, yk) < 1/2 for all k > N . This means that yk = x for k > N ({z : d2(x, z) < 1/2} only contains x), so that yk converges to x in the Tychonoff topology. Case 2. If x ∈ R′∞I and yk converges to x, then for any neighborhood Un ⊂ On, there is N such that or all k > N, yk ∈ Un. This means that yk ∈ R′∞I for k > N , so that yk converges to x in the Tychonoff topology. (cid:3) 2.1.2. Definition of B(R∞I ). In a similar manner, if B(Rn I ) is the Borel σ-algebra for Rn I (i.e., the smallest σ-algebra generated by the On), then B(Rn I ) ⊂ B(Rn+1 I ), so we can define B′(R∞I ) by: B′(R∞I ) = limn→∞B(Rn I ) = ∞∪k=1 B(Rk I ). If P(·) denotes a powerset of a set (i.e., P(A) = {X : X ⊆ A}), let B(R∞I ) be the smallest σ-algebra containing B′(R∞I ) ∪ P(R∞I \ ∪∞n=1Rn (It is I ). obvious that the class B(R∞I ) coincides with Borel σ-algebra generated by 18 GILL, PANTSULAIA, AND ZACHARY the τ -topology on R∞.) From our definition of B(R∞I ) we see that B(R∞) ⊂ B(R∞I ) and the containment is proper. Theorem 2.5. λ∞(·) is a measure on B(Rn Lebesgue measure on Rn. I ), equivalent to n-dimensional Ai ∈ B(Rn Proof. If A = ∞×i=1 λ∞(A) =Q∞i=1 λ(Ai) always converges. Furthermore, I ), then λ(Ai) = 1 for i > n so that the series (2.1) 0 < λ∞(A) =Y∞ i=1 Since sets of the type A = n i=1 λ(Ai) = λn( λ(Ai) =Yn Ai generate B(Rn), we see that λ∞(·), re- ×i=1 ×i=1 Ai). n stricted to Rn I , is equivalent to λn(·). (cid:3) Corollary 2.6. The measure λ∞(·) is both translationally and rotationally invariant on (Rn I , B[Rn I ]). 2.2. The Extension to R∞I . It is not obvious that λ∞(·) can be extended to a countably additive measure on B(R∞I ). Definition 2.7. Let I ) ⊂ B(R∞I ) : n ∈ N, K is compact and 0 < λ∞(Kn) < ∞}, Kni, N ∈ N; Kni ∈ ∆0 and λ∞(Knl ∩ Knm) = 0, l 6= m}. i=1 ∆0 = {Kn = K×In ∈ B(Rn ∆ = {PN =[N Definition 2.8. If PN ∈ ∆, we define λ∞(PN ) =XN λ∞(Kni). i=1 CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES 19 Since PN ∈ B(Rn next result follows: I ) for some n, and λ∞(·) is a measure on B(Rn I ), the Lemma 2.9. If PN1, PN2 ∈ ∆ then: (1) If PN1 ⊂ PN2, then λ∞(PN1) ≤ λ∞(PN2 ). (2) If λ∞(PN1 ∩ PN2) = 0, then λ∞(PN1 ∪ PN2) = λ∞(PN2) + λ∞(PN2). Definition 2.10. If G ⊂ R∞I is any open set, we define: λ∞(G) = lim N→∞ sup{λ∞(PN ) : PN ∈ ∆, PN ⊂ G, } . Theorem 2.11. If O is the class of open sets in B(R∞I ), we have: (1) λ∞(R∞I ) = ∞. (2) If G1, G2 ∈ O, G1 ⊂ G2, then λ∞(G1) ≤ λ∞(G2). (3) If {Gk} ⊂ O, then λ∞([∞ Gk) ≤X∞ λ∞(Gk). k=1 k=1 (4) If the Gk are disjoint, then λ∞([∞ k=1 Gk) =X∞ k=1 λ∞(Gk). Proof. The proof of (1) is standard. To prove (2), observe that {PN : PN ⊂ G1} ⊂(cid:8)P ′N : P ′N ⊂ G2(cid:9) , so that λ∞(G1) ≤ λ∞(G2). To prove (3), let PN ⊂ S∞k=1 Gk. Since PN is compact, there is a finite number of the Gk which cover PN , so that k=1 Gk. Now, for each Gk, there is a PNk ⊂ Gk. Furthermore, as k=1 PNk . Since there is I ), we may also assume that λ∞(PNl ∩ PNm) = 20 GILL, PANTSULAIA, AND ZACHARY PN ⊂ SL PN is arbitrary, we can assume that PN = P ′N =SL an n such that all PNk ∈ B(Rn 0, l 6= m. We now have that Xk=1 λ∞(PNk ) 6 λ∞(PN ) = λ∞(Gk) 6 L L Xk=1 ∞ Xk=1 λ∞(Gk). It follows that λ∞([∞ k=1 Gk) ≤X∞ k=1 λ∞(Gk). If the Gk are disjoint, observe that if PN ⊂ P ′M , Xk=1 λ∞(P ′M ) ≥ λ∞(PN ) = L λ∞(PNk ). It follows that λ∞([∞ k=1 Gk) ≥XL k=1 λ∞(Gk). This is true for all L so that this, combined with (3), gives our result. (cid:3) If F is an arbitrary compact set in B(R∞I ), we define (2.2) λ∞(F ) = inf {λ∞(G) : F ⊂ G, G open} . Remark 2.12. At this point we see the power of B(R∞I ). Unlike B(R∞), equation (2.2) is well-defined for B(R∞I ) because it has a sufficient number of open sets of finite measure. Theorem 2.13. Equation (2.2) is consistent with Definition 2.6 and the results of Theorem 2.11. CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES 21 Definition 2.14. Let A be an arbitrary set in R∞I . (1) The outer measure (on R∞I ) is defined by: λ∗ ∞(A) = inf {λ∞(G) : A ⊂ G, G open} . We let L0 be the class of all A with λ∗∞(A) < ∞. (2) If A ∈ L0, we define the inner measure of A by λ∞,(∗)(A) = sup{λ∞(F ) : F ⊂ A, F compact} . (3) We say that A is a bounded measurable set if λ∗∞(A) = λ∞,(∗)(A), and define the measure of A, λ∞(A), by λ∞(A) = λ∗∞A). Theorem 2.15. Let A, B and {Ak} be arbitrary sets in R∞I with finite outer measure. (1) λ∞,(∗)(A) ≤ λ∗∞(A). (2) If A ⊂ B then λ∗∞(A) ≤ λ∗∞(B) and λ∞,(∗)(A) ≤ λ∞,(∗)(B). (3) λ∗∞(S∞k=1 Ak) ≤P∞k=1 λ∗∞(Ak). (4) If the {Ak} are disjoint, λ∞,(∗)(S∞k=1 Ak) ≥P∞k=1 λ∞,(∗)(Ak). Proof. The proofs of (1) and (2) are straightforward. To prove (3), let ε > 0 be given. Then, for each k, there exists an open set Gk such that Ak ⊂ Gk and λ∞(Gk) < λ∗∞(Ak) + ε2−k. Since (S∞k=1 Ak) ⊂ (S∞k=1 Gk), we have λ∗ ∞(cid:16)[∞ k=1 k=1 Ak(cid:17) 6 λ∞(cid:16)[∞ <X∞ Gk(cid:17) 6X∞ ∞(Ak) + ε2−k] =X∞ [λ∗ k=1 λ∞(Gk) k=1 λ∗ ∞(Ak) + ε. k=1 Since ε is arbitrary, we are done. 22 GILL, PANTSULAIA, AND ZACHARY To prove (4), let F1, F2, . . . , FN be compact subsets of A1, A2, . . . , AN , respectively. Since the Ak are disjoint, λ∞,(∗)(cid:16)[∞ k=1 Ak(cid:17) > λ∞(cid:18)[N k=1 Fk(cid:19) =XN k=1 λ∞(Fk). Thus, λ∞,(∗)(cid:16)[∞ k=1 Ak(cid:17) ≥XN k=1 λ∞,(∗)(Ak). Since N is arbitrary, we are done. (cid:3) The next two important theorems follow from the last one. Theorem 2.16. (Regularity) If A has finite measure, then for every ε > 0 there exist a compact set F and an open set G such that F ⊂ A ⊂ G, with λ∞(G \ F ) < ε. Proof. Let ε > 0 be given. Since A has finite measure, it follows from our definitions of λ∞,(∗) and λ∗∞ that there is a compact set F ⊂ A and an open set G ⊃ A such that λ∞(G) < λ∗ ∞(A) + ∈2 and λ∞(F ) > λ∞,(∗)(A) − ∈2 . Since λ∞(G) = λ∞(F ) + λ∞(G \ F ), we have: λ∞(G \ F ) = λ∞(G) − λ∞(F ) < (λ∞(A) + ε 2 ) − (λ∞(A) − ε 2 ) = ε. (cid:3) CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES 23 Theorem 2.17. (Countable Additivity) If the family {Ak} consists of dis- joint sets with bounded measure and A = S∞k=1 Ak, with λ∗∞(A) < ∞. then λ∞(A) =P∞k=1 λ∞(Ak). Proof. Since λ∗∞(A) < ∞, we have: ∞(Ak) =X∞ λ∞,(∗)(Ak) 6 λ∞,(∗)(A) 6 λ∗ ∞(A) 6X∞ ∞(A). λ∗ λ∗ k=1 k=1 It follows that λ∞(A) = λ∗∞(A) = λ∞,(∗)(A), so that Ak(cid:17) =X∞ λ∞(A) = λ∞(cid:16)[∞ k=1 k=1 λ∞(Ak). (cid:3) Definition 2.18. Let A be an arbitrary set in R∞I . We say that A is mea- surable if A ∩ M ∈ L0 for all M ∈ L0. by: In this case, we define λ∞(A) λ∞(A) = sup{λ∞(A ∩ M ) : M ⊂ L0} . We let L∞I be the class of all measurable sets A. Proofs of the following results are standard (see Jones [J], pages 48-52). Theorem 2.19. Let A and {Ak} be arbitrary sets in L∞I . (1) If λ∗∞(A) < ∞, then A ∈ L0 if and only if A ∈ L∞I . In this case, λ∞(A) = λ∗∞(A). (2) L∞I is closed under countable unions, countable intersections, differ- ences and complements. 24 GILL, PANTSULAIA, AND ZACHARY (3) k=1 λ∞([∞ (4) If {Ak} are disjoint, λ∞([∞ k=1 Ak) ≤X∞ k=1 λ∞(Ak). Ak) =X∞ k=1 λ∞(Ak). (5) If Ak ⊂ Ak+1 for all k, then λ∞([∞ k=1 Ak) = lim k→∞ λ∞(Ak). (6) If Ak+1 ⊂ Ak for all k and λ∞(A1) < ∞, then λ∞(\∞ k=1 Ak) = lim k→∞ λ∞(Ak). We end this section with an important result that relates Borel sets to L∞I -measurable sets (Lebesgue). Theorem 2.20. Let A be a L∞I -measurable set. Then there exists a Borel set F and a set N with λ∞(N ) = 0 such that A = F ∪ N . Thus, we see that λ∞(·) is a regular countably additive σ-finite Borel measure on R∞I = R∞ (as sets). More important is the fact that the de- velopment is no more difficult than the corresponding theory for Lebesgue measure on Rn. Throughout the remainder of the paper we will also use B[R∞I ] for its completion L∞I when convenient. This should cause no confusion since the given context will always be clear. CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES 25 2.3. Separable Banach Spaces. In order to see what other advantages our construction of (R∞I , B[R∞I ], λ∞(·)) offers, in this section we study sep- arable Banach spaces. Let B be any separable Banach space. Recall that (see Diestel [DI], page 32): Definition 2.21. A sequence (un) is called a Schauder basis for B if kunkB = 1 and, for each f ∈ B, there is a unique sequence (an) of scalars such that f = limn→∞Xn k=1 akuk. Definition 2.22. A sequence (vn) is called an absolutely convergent Schauder basis for B if P∞n=1 kvnkB < ∞ and, for each f ∈ B, there is a unique sequence (bn) of scalars such that f = limn→∞Xn k=1 bkvk. Lemma 2.23. Let (un) be a Schauder basis for B, then there exists an absolutely convergent Schauder basis for B. Proof. Let (vn) = ( un 2n ). Then ∞ Xn=1 kvnkB = kunkB 2n = ∞ Xn=1 ∞ Xn=1 1 2n = 1 < ∞. To see that (vn) is a Schauder basis for B, let f ∈ B. By definition, there is a unique sequence (an) of scalars such that f = limn→∞Xn k=1 akuk. 26 GILL, PANTSULAIA, AND ZACHARY If we take the sequence (bn) = (2nan), then limn→∞Xn k=1 bkvk = limn→∞Xn k=1 akuk = f. (cid:3) It is known that most of the natural separable Banach spaces, and all that have any use for applications in analysis, have a Schauder basis. In partic- ular, it is easy to see from the definition of a Schauder basis that, for any sequence (an) ∈ R∞I representing a function f ∈ B, we have limn→∞ an = 0. It follows that every separable Banach space (with a Schauder basis) is iso- morphic to a subspace of R∞I . Let BI be the set of all sequences (an) for which limn→∞Pn in B. Define k=1 akuk exists Lemma 2.24. An operator n (cid:13)(cid:13)(cid:13)Xn k=1 k(an)kBI = sup . akuk(cid:13)(cid:13)(cid:13)B T : (B, · B) → (BI , · BI ), defined by T (f ) = (ak) for f = limn→∞Pn from B onto BI . k=1 akuk ∈ B, is an isomorphism Let B be a separable Banach space with a Schauder basis and let BI = T [B]. If B(BI ) = BI ∩ B[R∞I ], we define the σ−algebra generated on B, and associated with B(BI ) by: BI [B] =(cid:8)T −1(A) A ∈ B[BI ](cid:9) =: T −1 {B [BI ]} . CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES 27 Note that, just as B[R∞] ⊂ B[R∞I ], we also have B[B] ⊂ BI [B] (with the containment proper). Theorem 2.25. Let A ∈ BI (B) and set λB(A) = λ∞[T (A)]. Let λB be the completion of λB, then λB is a non-zero σ-finite Borel measure on B. Proof. Let {vk} be an absolutely convergent Schauder basis. We first prove that, for any L > 0 and any sequence (ak) ∈ [−L, L]ℵo, the function f = P∞k=1 akvk ∈ B. We then prove that λB is nonzero. Part 1 Let L be given. Since (vn) is an absolutely convergent Schauder basis, given ε > 0 we can choose N such that P∞k=N vk < ǫ N ≤ m ≤ n, we have L . It follows that, for n 6 Xk=m Thus, the sequence {fn}, defined by fn =Pn in B. Since B is a Banach space, the sequence converges. kvkk < ε. n (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xk=m akvk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) k=1 akvk, is a Cauchy sequence Part 2 To prove that λB is nonzero, it suffices to show that λB(cid:2)T −1 (I0)(cid:3) 6= 0, where (I0) = [− 1 R∞I , so that B = T −1(I0) ∈ BI (B). Thus, 2 ]ℵo. First, we note that T is an injective linear map into 2 , 1 λB(B) = λ∞(cid:2)T (cid:0)T −1(I0)(cid:1)(cid:3) = λ∞(I0) = 1. (cid:3) 28 GILL, PANTSULAIA, AND ZACHARY 2.4. Translations. In the theorem below, we will provide a new proof that ℓ1 is the largest (dense) group of admissible translations for R∞I , so neces- sarily ℓ1 is the largest group of admissible translations for every separable Banach space B. Recall that h(x) = ⊗∞k=1hk(xk), where hk(xk) = 1, for xk ∈ [− 1 2 , 1 2 ]. It follows from dν = hdλ∞, that ν is absolutely continuous with respect to λ∞. Thus, ν is equivalent to λ∞. Let Tλ∞ be the set of admissible translations for R∞I (i.e., λ∞[A − x] = λ∞[A] for all A ∈ B[R∞I ] and x ∈ Tλ∞). Theorem 2.26. If A ∈ B[R∞I ] then λ∞[A − x] = λ∞[A] if and only if Tλ∞ = ℓ1. Proof. Suppose that x ∈ ℓ1. Since ν ∼ λ∞, we have that Tν = Tλ∞ (see Yamasaki [YA1]). Thus, it suffices to prove that ν[A − x] = ν[A]. By Kakutani's Theorem ([KA], see also [HHK] pg. 116), ν[A − x] ∼ ν[A] if and only if ∞ Yk=1Z ∞ −∞phk (yk) hk (yk − xk)dλ(yk) > 0. (2.3) Now, 1 2 , 1 2 ]∩[− Z ∞ −∞phk (yk) hk (yk − xk)dλ(yk) =Z[− dλ(yk) = (1−xk)+, where r+ = max(0, r). Since x ∈ ℓ1, Q∞k=n (1 − xk)+ > 0 for n large enough. Thus, equation (2.3) will be satisfied for every x ∈ ℓ1, so that ℓ1 ⊂ Tν. 1 2 +xk, 1 2 +xk] CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES 29 Now, suppose that x ∈ Tλ∞, so that λ∞[A − x] = λ∞[A] for all A ∈ B[R∞I ]. Thus, for A ∈ B[Rn I ], we have λ∞ [A − x] = λn [An − xn] · ∞ Yk=n+1 λ(cid:8)(cid:2)− 1 2 , 1 λ(cid:8)(cid:2)− 1 2(cid:3) ∩(cid:2)− 1 2 , 1 2(cid:3) ∩(cid:2)− 1 2 − xk, 1 = λn [An] · ∞ Yk=n+1 2 − xk, 1 2 − xk(cid:3)(cid:9) 2 − xk(cid:3)(cid:9) = λn [An] · ∞ Yk=n+1 (1 − xk)+. 2 , 1 k=1[− 1 If An = In = ×n that P∞k=1 xk < ∞, so that x ∈ ℓ1. 2 ], we have 1 = lim n→∞ ∞ Qk=n+1 (1 − xk)+. It follows (cid:3) In closing, we note that, since λ∞ is complete and regular, it is metrically It follows from Theorem 0.5 that λ∞ is transitivity with respect to R∞0 . unique (this comment also applies to λB). 2.5. Gaussian measure. If we replace Lebesgue measure by the infinite product Gaussian measure, µ∞, on R∞, we get countable additivity but lose rotational invariance. Furthermore, the µ∞ measure of l2 is zero. On the other hand, another approach is to use the standard projection method onto finite dimensional subspaces to construct a probability measure directly on l2. In this case, we recover rotational invariance but not translation invariance (and lose countable additivity). The resolution of this problem led to the development of the Wiener measure [WSRM] and this is where we are today. A nice discussion of this and related issues can be found in Dunford and Schwartz [DS] (see pg. 402). 30 GILL, PANTSULAIA, AND ZACHARY We now turn to take a look at infinite product Gaussian measure from our new perspective. The canonical Gaussian measure on R is defined by: dµ(x) = 1 √2π exp(−x2 2 ) dλ(x). Recall that µ∞ = ⊗∞k=1µ is countably additive on R∞, but its measure of ℓ2 is zero. If we introduce a scaled version of Gaussian measure on R∞I , we can resolve this difficulty. We seek a family of variances {σ2 k} such that µB(B) = ∞⊗k=1 µk(T [B] = 1, where µk is a linear Gaussian measure on R with parameters (0, σk) for k ∈ N and µB is defined by: µB(B) = ∞⊗k=1 µk(T [B]), for any Borel subset B of B. Lemma 2.27. Let (cid:8)σ2 k(cid:9) be a family of variances such that ∞ Xk=1 σ2 k < ∞, then µB(cid:0)T −1([−L, L]ℵo )(cid:1) > 0 for every positive number L. Proof. Let {Xk} be the family of independent Gaussian random variables defined on some common probability space, (Ω, B, P [·]), with law µk. If X = (X1, X2, . . . ), then P hnω ∈ Ω X(ω) ∈ [−L, L]ℵooi = P h\∞ =Y∞ P [{ω ∈ Ω Xk(ω) 6 L]}] >Y∞ k=1 k=1 {ω ∈ Ω Xk(ω) ∈ [−L, L]}i k=1(cid:18)1 − L2(cid:19), σ2 k by Chebyshev's inequality. CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES 31 Clearly the product is positive. We are done since B = T −1([−L, L]ℵo) ∈ B(B) and µB (B) = (⊗∞k=1µk)(cid:16)T [T −1([−L, L]ℵo )](cid:17) = P hnω ∈ Ω X(ω) ∈ [−L, L]ℵooi . (cid:3) Theorem 2.28. If the family of variances (cid:8)σ2 k(cid:9) satisfies the stronger con- dition (2.4) ∞ Xk=1 σ2 k xk < ∞ for some sequence (xk) ∈ ℓ1, then µB([B]) = 1. Proof. By definition, if f ∈ B and (un) is a Schauder basis for B, then there is a sequence of scalars (ak) such that f = limn→∞Pn k=1 akuk. Since T (f ) = (ak), k(an)kBI = sup n n (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xk=1 6" ∞ Xk=1 ak# , akuk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)B so that, if (an) ∈ ℓ1, then (an) ∈ T (B) = BI. Suppose that there is a sequence (xk) ∈ ℓ1 such that such that the in- equality (2.3) is satisfied. As in Lemma 2.24, by Chebyshev's inequality and inequality (2.3) we have µB(cid:26)T −1(cid:18) ∞×k=1h−xk1/2 , xk1/2i(cid:19)(cid:27) > 0. 32 GILL, PANTSULAIA, AND ZACHARY If An = Rn × (×∞k=n+1[−xk1/2 ,xk1/2]), then An ⊂ An+1 and An ⊆ BI for all natural n. Thus, we have µB[T −1(An)] µB[T −1(BI )] ≥ lim n→∞ µk([−xk1/2 ,xk1/2]) ≥ lim n→∞ = lim n→∞ ∞ Yk=n+1 ∞ Yk=n+1(cid:18)1 − σ2 k xk(cid:19) = 1. (cid:3) Definition 2.29. We call µB a scaled version of Gaussian measure for B. Theorem 2.30. The measure µB is a countably additive version of Gaussian measure on B. In particular, observe that we obtain a countably additive version of Gaussian measure for both ℓ2 and C0[0, 1] (the continuous functions x(t) on [0, 1] with x(0) = 0). 2.6. Rotational Invariance. In this section we study rotational invariance on subspaces of (R∞I , BI [R∞] λ∞). First, we need a little more information about Gaussian measures on vector spaces. (See Yamasaki [YA], pg. 151, for a proof of the next Theorem). Let F be a a real vector space, let F a be its algebraic dual space, and let BF be the smallest σ-algebra such that L(x) is measurable for each functional L ∈ F a and all x ∈ F. Theorem 2.31. If µ is a measure on (F a, BF ), then the following are equivalent. CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES 33 (1) The Fourier transform of µ, µ, is of the form: µ(x) = exp(cid:8)− 1 2 hx, xi(cid:9) , for some inner product on F. (2) For every x ∈ F, the distribution of L(x) is a one-dimensional Gaussian measure. In this general setting, a measure µ is said to be Gaussian on (F a, BF ) if it satisfies either of the above conditions. Example 2.32. Let F = R∞0 , the set of sequences that are zero except for a finite number of terms and let h·, ·i be the inner product on R∞0 . It is easy to show that the corresponding measure on F a = R∞ (satisfying either (1) or (2) above) is the infinite product Gaussian measure. To understand the importance of this example, let (an) be any sequence of positive numbers and let (2.5) Ha =( x ∈ R∞ a2 nx2 n < ∞) . ∞ Xn=1 The proof of the following is due to Yamasaki ([YA], pg. 153). Lemma 2.33. If a ∈ ℓ2, µ[Ha] = 1, and if a /∈ ℓ2, µ[Ha] = 0. Now, let us note that the standard one-dimensional Gaussian density, which is normally written as fX(x) = [√2π]−1exp{− 1 2 x2}, may also be written as fX(x) = exp{−π x2} with no factors of √2π if we scale x → 34 GILL, PANTSULAIA, AND ZACHARY . With this convention, we can write the infinite dimensional version for x√2π L2[H, λH] as the derivative of the Gaussian distribution µH with respect to the Lebesgue measure on H: (2.6) f (x) = exp{−π x2 H} = dµH(x) dλH(x) . This shows that, with the appropriate definition of Lebesgue measure, there is a corresponding density for a Gaussian distribution on Hilbert space. Remark 2.34. In the general case (see DePrato [DP]), when Q is a (positive definite) trace-class operator and x is a Gaussian random variable with mean m and covariance Q, we can write equation (2.6) as: f (x) = [det Q]−1/2 exp(cid:8)−π(cid:10)Q−1(x − m), (x − m)(cid:11)H(cid:9) dµH(x) dλH(x) . Definition 2.35. A rotation on H is a bijective isometry U : H → H. It is well-known that µH is invariant under rotations over (H, BH) (see Yamasaki [YA], pg. 163). Theorem 2.36. The measure, λH, is invariant under rotations and R =: (T −1(ℓ2)) is dense in H and the maximal rotation invariance subspace for λH. Proof. Let any measurable set A ∈ BH. If U is any rotation on H, then µH(U A) = µH(A) and U x2 H. It follows from equation (2.6) that λH(U A) = λH(A). H = x2 CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES 35 It follows from R∞0 ⊂ R ⊂ H, that R is dense, and from Lemma 2.33 that R is maximal. (cid:3) Discussion. In this section, we have shown that what appears to be a minor change in the way we represent R∞ makes it possible to define an analogue of both Lebesgue and Gaussian measure (countably additive) on every (classical) separable Banach space with a Schauder basis. Further- more, our version of Gaussian measure is rotationally invariant, a property not shared by Wiener measure. (What is more important, we have obtained our core results using basic methods of Lebesgue measure theory from Rn.) 3. Operators This section provides the background to understand the relationship be- ⊗ (which is nonseparable), and their restriction ⊗(h). We also obtain general conditions that allow us to define infinite tween operators defined on H2 to H2 sums and products of linear operators on H2 ⊗(h) for a given h. 3.1. Bounded Operators on H2 ⊗. In this section we review the class of bounded operators on H2 ⊗ and their relationship to those on each Hi. Many of the results are originally due to von Neumann [VN2]. However, the proofs are new or simplified versions (some from the literature). 36 GILL, PANTSULAIA, AND ZACHARY Let L[H2 ⊗] be the set of bounded operators on H2 ⊗. For each fixed i0 ∈ N i N N i ) = Ai0gk ⊗i∈Ngk and Ai0 ∈ L(Hi0), define Ai0 ∈ L(H2 ⊗) by: Xk=1 k=1 ⊗i∈Ngk i0 ⊗ (⊗i6=i0gk i ) Xk=1 Ai0( for PN in H2 ⊗ and N finite but arbitrary. Extending to all of H2 ⊗ produces an isometric isomorphism of L[Hi0] into L[H2 ⊗], which we denote by L[H(i0)], so that the relationship L[Hi] ↔ L[H(i)] is an isometric isomorphism of algebras. Let L#[H2 ⊗] be the uniform closure of the algebra It is clear that L#[H2 generated by {L[H(i)], i ∈ N}. ⊗]. von Neumann has shown that the inclusion becomes equality if and only if N is ⊗] ⊂ L[H2 ⊗] clearly consists of all ⊗ that are generated directly from the family {L[H(i)], i ∈ replaced by a finite set. On the other hand, L#[H2 operators on H2 N} by algebraic and topological processes. g denote the projection from H2 ⊗(g)s, and let Pw ⊗ onto H2 g denote Let Ps the projection from H2 ⊗ onto H2 ⊗(g)w. Theorem 3.1. If T ∈ L#(H2 ⊗), then Ps gT = TPs g and Pw g T = TPw g . Proof. The weak case follows from the strong case, so we prove that Ps gT = TPs i , with gk i = gi for all k=1 ⊗i∈Ngk g. Since vectors of the form G = PL ⊗(g)s; it suffices to show that Tf ∈ but a finite number of i, are dense in H2 H2 ⊗(g)s. Now, T ∈ L#(H2 ⊗) implies that there exists a sequence of operators Tn such that kT − Tnk⊗ → 0 as n → ∞, where each Tn is of the form: Tn = PNn k = k a complex scalar, Nn < ∞, and each T n k , with an k=1 an k T n CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES 37 ⊗i∈N\Mk Ii for some finite set of i-values Mk, where Ii is the identity ki ⊗i∈Mk T n operator on Hi. Hence, Tnf =XL l=1XNn k=1 kigl an k ⊗i∈Mk T n i ⊗i∈N\Mk gl i. i ⊗i∈N\Mk gl ⊗(g)s for each n, so that Tn ∈ L[H2 Now, it is easy to see that, for each l, ⊗i∈MkT n kigl It follows that Tnf ∈ H2 L[H2 gT = TPs Ps g. ⊗(g)s] is a norm closed algebra, T ∈ L[H2 i ≡s ⊗i∈Ngi. ⊗(g)s]. Since ⊗(g)s] and it follows that (cid:3) Let zi ∈ C, zi = 1, and define U [z] by: U [z] ⊗i∈N gi = ⊗i∈Nzigi. Theorem 3.2. The operator U [z] has a unique extension to a unitary op- erator on H2 ⊗, which we also denote by U [z], such that: ⊗(g)w, so that Pw ⊗(g)w → H2 is quasi-convergent but not convergent, g U [z] = U [z]Pw g . then U [z] : (1) U [z] : H2 (2) If Qν zν ⊗(g)s → H2 H2 ⊗(h)s, for some h ∈ H2 ⊗(g)w with g⊥h. (3) U [z] : H2 ⊗(g)s → H2 ⊗(g)s if and only if Qi zi converges and U [z] = ⊗. This implies (Qi zi)I⊗, where I⊗ is the identity operator on H2 that Ps gU [z] = U [z]Ps g. Proof. For (1), let h = PN trary and 1 6 k 6 N . Then k=1 ⊗i∈Nhk i , where ⊗i∈Nhk i ≡w ⊗i∈Ngi, N is arbi- U∗[z]U [z]h = N Xk=1 ⊗i∈Nz∗i zihk i = h = U [z]U∗[z]h. 38 GILL, PANTSULAIA, AND ZACHARY Thus, we see that U [z] is a unitary operator, and since h of the above k=1 ⊗i∈Nhk g U [z] = U [z]Pw ⊗(g)w if PN k=1 ⊗i∈Nzihk ⊗(g)w → H2 ⊗. By def- ⊗(g)w, so that g . To prove (2), use The- ν ∈ H2 i ∈ H2 ⊗(g)w and Pw form are dense, U [z] extends to a unitary operator on H2 inition, PN U [z] : H2 orem 1.6 (3) and (4) to note that Qi zi = 0 and ⊗i∈Nhk that ⊗i∈Nzihk 0 < Qi zi < ∞, then U [z] = [(Qi zi)I⊗], so that U [z] : H2 Now suppose that U [z] : H2 and so Qi zi must converge. Therefore, U [z]h = [(Qi zi)I⊗]h and Ps i ≡s ⊗i∈Ngi imply ⊗(g)s. To prove (3), note that, if ⊗(g)s. i ≡s ⊗i∈Ngi gU [z] = ⊗(g)s → H2 ⊗(g)s, then ⊗i∈Nzihk i ∈ H2 ⊗(f )s with H2 ⊗(f )s⊥H2 ⊗(g)s → H2 U [z]Ps g. It is easy to see that, for each fixed i ∈ N, A(i) ∈ L[H(i)] commutes with any Ps g, Pw g or U [z], where g and z are arbitrary. (cid:3) Theorem 3.3. Every T ∈ L#[H2 where g and z are arbitrary. ⊗] commutes with all Ps g, Pw g and U [z], Proof. Let L be the set of all Ps g, Pw g or U [z], with g and z arbitrary. From the above observation, we see that all Ai ∈ L[H(i)], i ∈ N, commute with L and hence belong to its commutator L′. Since L′ is a closed algebra, this implies that L#[H2 ⊗] ⊆ L′ so that all T ∈ L#[H2 ⊗] commute with L. (cid:3) 3.2. Unbounded Operators on H2 ⊗. In this section, we consider a re- stricted class of unbounded operators and the notion of a strong convergence vector introduced by Reed [RE]. CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES 39 For each i ∈ N, let Ai be a closed densely defined linear operator on Hi, with domain D(Ai), and let Ai be its extension to H2 ⊗, with domain D(Ai) ⊃ D(Ai) = D(Ai) ⊗ (⊗k6=iHk). The next theorem follows directly from the definition of the tensor product of semigroups. Theorem 3.4. Let Ai, 1 6 i 6 n, be generators of a family of C0- 6 Mieωit. Then Sn(t) = ⊗i=1,n Si(t), semigroups Si(t) on Hi with kSi(t)kHi defined on ⊗i=1,nHi, has a unique extension (also denoted by Sn(t)) to all of H2 k=1 ⊗i∈Ngk l ∈ D(Al), 1 6 l 6 n, the infinitesimal generator for Sn(t) satisfies: ⊗, such that, for all vectors PK i with gk An" K Xk=1 ⊗i∈Ngk i# = n K Xl=1 Xk=1 Algk l (⊗i6=l i∈Ngk i ). Definition 3.5. Let {Ai}, i ∈ N, be a family of closed densely defined linear operators on Hi and let gi ∈ D(Ai) (respectively, fi ∈ D(Ai)), with kgikH = 1 (respectively, kfikH = 1), for all i ∈ N. (1) We say that g = ⊗i∈Ngi is a strong convergence sum (scs)-vector for i∈Ngi) exists. (2) We say that f = ⊗i∈Nfi is a strong convergence product (scp)-vector k=1 Akg =P∞k=1 Akgk(⊗i6=k the family {Ai} if s - lim n→∞ Pn n→∞ Qn for the family {Ai} if s - lim k=1 Akf = ⊗i∈NAifi exists. Let Dg be the linear span of {χ = ⊗i∈Nχi, χi ∈ D(Ai)}, with χi = gi (and let Df be the linear span of {η = ⊗i∈Nηi, ηi ∈ D(Ai)}, with ηi = fi) ⊗(g)s for all i > L, where L is arbitrary but finite. Clearly, Dg is dense in H2 (Dη is dense in H2 If there is a possible chance for confusion, we ⊗(f )s). 40 GILL, PANTSULAIA, AND ZACHARY let As, respectively Ap, denote the closure of P∞k=1 Ak on H2 tively Q∞k=1 Ak on H2 ⊗(g)s (respec- ⊗(f )s) are natural spaces for the study of infinite sums or products of unbounded ⊗(f )s). It follows that H2 ⊗(g)s (respectively H2 operators. (The notion of a strong convergence sum vector first appeared in Reed [RE].) Definition 3.6. We call H2 space ) for the family {Ai}. ⊗(g)s an RS-space (respectively, H2 ⊗(f )s an RP- Let {Uk(t)} be a set of unitary groups on {Hk}. It is easy to see that ⊗. However, we know from Theorem 3.2 (2), that it need not be reduced on any partial tensor product U (t) = ⊗∞k=1Uk(t) is a unitary group on H2 subspace. The following results are due to Streit [ST] and Reed [RE], as indicated. Theorem 3.7. (Streit) Suppose {Ak} is a set of selfadjoint linear operators ⊗(g)s, with corresponding unitary groups {Uk(t)}. If U (t) = on the space H2 ⊗∞k=1Uk(t), then Ps ⊗(g)s) and ⊗(g)s if and only if, for U (t) is a strongly continuous unitary group on H2 each c > 0, the following three conditions are satisfied: g (i.e., U (t) is reduced on H2 gU (t) = U (t)Ps (1) P∞k=1 hAkEk[−c, c]gk, gki < ∞, (2) P∞k=1(cid:12)(cid:12)(cid:10)A2 (3) P∞k=1 h(Ik − Ek[−c, c]gk, gki < ∞, kEk[−c, c]gk, gk(cid:11)(cid:12)(cid:12), CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES 41 where Ek[−c, c] are the spectral projectors of Ak and, in this case, U (t) = s − limn→∞ ⊗n k=1Uk(t). Corollary 3.8. Conditions 1-3 are satisfied if and only if there exists a strong convergence vector g = ⊗∞k=1gk for the family {Ak} such that gk ∈ D(Ak) and k=1 kAkgkk2 < ∞. X∞ k=1 hAkgk, gki < ∞, X∞ ⊗(g)s and U (t) is a strongly Theorem 3.9. (Reed) U (t) is reduced on H2 continuous unitary group on H2 ⊗(g)s if and only if g = ⊗∞k=1g is a strong convergence vector for the family {Ak} and P∞k=1 hAkgk, gki < ∞. If each case, A, the closure of P∞k=1 Ak, is the generator of U (t). Ak is positive, the statement is true without the above condition. In either The next result strengthens and extends Reed's theorem to contraction semigroups (i.e., the positivity requirement above can be dropped). Theorem 3.10. Let {Sk(t)} be a family of strongly continuous contraction semigroups with generators {Ak} defined on {Hk}, and let g = ⊗∞k=1gk be a strong convergence vector for the family {Ak}. Then S(t) = ⊗∞k=1Sk(t) is reduced on H2 ⊗(g)s and is a strongly continuous contraction semigroup. If S(t) = ⊗∞k=1Sk(t) is reduced on H2 ⊗(g)s and is a strongly continuous ⊗(g)s, then there exists a strong convergence contraction semigroup on H2 vector f = ⊗∞k=1fk ∈ H2 ⊗(g)s for the family {Ak}. 42 GILL, PANTSULAIA, AND ZACHARY Proof. Let g = ⊗∞k=1gk be a strong convergence vector for the family {Ak}. Without loss, we can assume that kgkk = 1. Let Sn(t) = ⊗n k=1Sk(t) ⊗(⊗∞k=n+1Ik) and observe that Sn(t) is a contraction semigroup ⊗(g)s for all finite n . Furthermore, its generator is the closure of on H2 An =Pn k=1 Ak, where Ak = Ak ⊗(⊗∞i6=kIi). If n and m are arbitrary, then [Sn(t) − Sm(t)] g =Z 1 = tZ 1 d dλ {Sn[λt]Sm[(1 − λ)t]} gdλ 0 Sn[λt]Sm[(1 − λ)t] [An − Am] gdλ, 0 where we have used the fact that, if two semigroups commute, then their corresponding generators also commute. It follows that: k[Sn(t) − Sm(t)] gk 6 tk[An − Am] gk . Since g = ⊗∞k=1g is a strong convergence vector for the family {Ak}, it fol- lows that s−limn→∞ Sn(t) = S(t) exists on a dense set in H2 ⊗(g)s and the convergence is uniform on bounded t intervals. It follows that S(t) extends to a bounded ⊗(g)s. To see that the closure of S(t) must be a con- linear operator on H2 traction, for any ε > 0, choose n so large that k[Sn(t) − S(t)] gk⊗ < εkgk⊗. It follows that kS(t)gk⊗ 6 kSn(t)gk⊗ + k[Sn(t) − S(t)] gk⊗ < kgk⊗ (1 + ε). Thus, S(t) is a contraction operator on H2 is a C0-semigroup. ⊗(g)s. It is easy to check that it CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES 43 Now suppose that S(t) = ⊗∞k=1Sk(t) is a strongly continuous contraction semigroup which is reduced on H2 ⊗(g)s. It follows that the generator A of S(t) is m-dissipative, and hence defined on a dense domain D(A) in H2 ⊗(g)s with S′(t)f = S(t)Af = AS(t)f for all f ∈ D(A). Since any such f is of the form f =P∞l=1 ⊗∞k=1f l k is in D(A). A simple computation shows that Af l =P∞k=1 Akf l, so that any f l is a strong convergence vector for the family {Ak}. k, each f l = ⊗∞k=1f l (cid:3) It is easy to see that, in the second part of the theorem, we cannot require that g = ⊗∞k=1gk itself be a strong convergence vector for the family {Ak} since it need not be in the domain of A. For example, g1 /∈ D(A1), while gk ∈ D(Ak), k 6= 1. 4. Function Spaces Let χIn be the indicator (or characteristic) function of In = ×∞k=n+1I. If we let L(Rn) represent the class of measurable functions on Rn, then for each measurable function fn ∈ L(Rn) we identify f ∈ L(Rn I ) by f = fn ⊗ χIn. Definition 4.1. A real-valued function f defined on the measure space (R∞I , B[R∞I ], λ∞) is said to be measurable if f−1(A) ∈ B[R∞I ] for every A ∈ B[R]. In this section we develop those aspects of function space theory that will be of use later. We note that all the standard theorems for Lebesgue measure apply. (The proofs are the same as for integration on Rn.) 44 GILL, PANTSULAIA, AND ZACHARY 4.1. L1-Theory. Let L1[Rn I ] be the class of integrable functions on Rn I . Since L1(Rn I ) ⊂ L1(Rn+1 I ), we define L1[R′∞I ] =S∞n=1 L1(Rn I ) and let L1[R∞I ] be the norm closure of L1[R′∞I ]. It follows that every function in L1[R∞I ] is the limit of a sequence of functions in L1[Rnk I ], for some sequence {nk} ⊂ N. Let Cc(Rn I ) be the class of continuous functions on Rn I which vanish out- side compact sets. We define Cc(R∞I ) to be the closure of S∞n=1 Cc(Rn I ) = Cc(R′∞I ) in the sup norm. Thus, for any f ∈ Cc(R∞I ), there always exists a sequence of functions {fnk} ∈ Cc(Rnk I ) such that fnk → f , for some sequence {nk} ⊂ N. We define C0(R∞I ), the functions that vanish at ∞, in the same manner. Lemma 4.2. If f ∈ Cc(R∞I ) or C0(R∞I ), then f is continuous. Proof. Let f (x) ∈ Cc(R∞I ) and let {xn n = 1, 2, . . . } be any sequence in Rn such that xn → x as n → ∞. If ε > 0 is given, choose K1 so that for k ≥ K1 and fk ∈ Cc(R∞I ), fk(xn) − f (xn) < ε 3 . Then choose K2 so that for k ≥ K2, fk(x) − f (x) < ε 3 . Choose N so that for n ≥ N, fk(xn) − fk(x) < ε 3 . If n ≥ N and k ≥ max{K1, K2}, we have: I f (xn) − f (x) ≤ fk(xn) − f (xn) + fk(x) − fk(xn) + fk(x) − f (x) < ε. The same proof applies to C0(R∞I ) (cid:3) Theorem 4.3. Cc(R∞I ) is dense in L1(R∞I ). Proof. We prove this result in the standard manner, by reducing the proof to positive simple functions and then to one characteristic function and finally CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES 45 using the approximation theorem to approximate a measurable set which contains a closed set and is contained in an open set. = 0 for all f ∈ L1 (by the DCT), we can prove the result for functions that vanish outside a compact First note that, since limk→∞(cid:13)(cid:13)f χBI (0,k) − f(cid:13)(cid:13)1 set. In this case, as f = f+ − f−, we need only consider positive f . How- ever, this function can be approximated by simple functions in S+. Since each simple function is a finite sum of characteristic functions (of bounded measurable sets) multiplied by finite constants, it follows that we need only show that we can approximate the characteristic function of a bounded mea- surable set by a continuous function which vanishes outside a compact set. Let ε > 0 be given and let g = χA, where A is any bounded measurable set. By the regularity of λ∞, there exists an open set O and a compact set H with H ⊂ A ⊂ O and λ∞(O \ H) < ε. Let {Vn} be the class of open intervals with rational end points. For each n ∈ N, let Fn ⊂ g−1[Vn] and Gn ⊂ (O \ g−1[Vn]) be compact sets, such that λ∞[(O \ Fn ∪ Gn)] < ε 2n . If H = ∩∞n=1[Fn ∪ Gn], then λ∞(O \ H) < ε. If x ∈ H, there is an n such that f (x) ∈ Vn and x ∈ Gc n ∩ H] ⊂ Vn. It follows that g restricted to H is continuous and λ∞(A \ H) ≤ λ∞(O \ H) < ε. n, so that g[Gc (cid:3) In a similar fashion we can define the Lp spaces, 1 < p < ∞. We should note that, each space is defined relative to the family of indicator functions 46 GILL, PANTSULAIA, AND ZACHARY for I. Thus, each space is the canonical one for that particular class of spaces. 5. Fourier Transform Theory In this section, we study the implications of Lebesgue measure on R∞ for the Fourier transform and discuss two different extensions of the Pontrjagin Duality theory for Banach spaces. Background. Let G be a locally compact abelian (LCA) group (c.f., Rn). The following is a restatement of Theorem 0.1 (see Rudin [RU1]). Theorem 5.1. If G is a LCA group and B(G) is the Borel σ-algebra of subsets of G, then there is a non-negative regular translation invariant mea- sure µ (i.e., µ(g + A) = µ(A), A ∈ B(G). The (Haar) measure µ is unique up to multiplication by a constant. Definition 5.2. A complex valued function α : G → C on a LCA group is called a character on G provided that α is a homomorphism and α(g) = 1 for all g ∈ G. The set of all continuous characters of G defines a new group G, called the dual group of G and (α1 + α2)(g) = α1(g) · α2(g). If we define a map G, by γg(α) = α(g), then the following theorem was proven by γ : G → Pontryagin: CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES 47 Theorem 5.3. (Pontryagin Duality Theorem) If G is a LCA-group, then the mapping γ : G → G is an isomorphism of topological groups. Thus, Pontrjagin Duality identifies those groups that are the character groups of their character groups. If the group is not locally compact The- orem 5.1 does not hold (e.g., there is no Haar measure). However Kaplan [KA1] has shown that the class of topological abelian groups for which the Pontrjagin Duality holds is closed under the operation of taking infinite products of groups. This result immediately implies that this class is larger than the class of locally compact abelian groups because the infinite product of locally compact groups (for example, R∞) may be non-locally compact (see also [KA2]). 5.1. Pontryagin Duality Theory I. In this section, we treat the Fourier transform as an operator. As will be seen, this approach has the advantage of being constructive. It also provides us with some insight into the problem that arises when we look at analysis on infinite dimensional spaces. We define F on L1[R, λ] by g(x) = F(g)(x) =ZR exp{−2πixy}g(y)dy. It is easy to check that F−1 is defined by g(y) = F−1(g)(y) =ZR exp{2πiyx}g(x)dx. 48 GILL, PANTSULAIA, AND ZACHARY This representation is more convenient for the infinite-dimensional case, be- cause we have no factors of √2π to worry about. It is possible to define F as a mapping on L1[Rn I , λ] to C0[Rn I , λ] for all n as one fixed linear operator. However, in the case of Hilbert spaces,Theorem 3.2(2) requires that we clearly specify our canonical domain and range space. The same is also true for L1[Rn I , λ](h) and C0[Rn I , λ](h) (see [GZ]). Since h = ⊗∞k=1χI (xk), an easy calculation shows that h = ⊗∞k=1 I ](h) by we can define F(fn)(x), mapping L1[Rn I ](h) into C0[Rn sin(πxk) πxk . Thus, (5.1) F(fn))(x) = ⊗n k=1Fk(f (n)) ⊗∞k=n+1 hk(xk). Theorem 5.4. The operator F extends to a bounded linear mapping of L1[R∞I ](h) into C0[R∞I ](h). Proof. Since L1[Rn I ](h) = lim n→∞ ∞ [n=1 L1[Rn I ](h) = L1[R′∞I ](h) and L1[R∞I ](h) is the closure of L1[R′∞I ](h) in the L1 - norm, it follows that F is a bounded linear mapping of L1[R′∞I ](h) into C0[R∞I ](h). Supposed that {fn} ⊂ L1[R′∞I ](h), converges to f ∈ L1[R∞I ](h). Thus, the sequence is Cauchy, so that kfn − fmk1 → 0 as m, n → ∞. It follows that F (fn(x) − fm(x)) 6ZR∞ I fn(y) − fm(y) dλ∞(y) = kfn − fmk1 , CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES 49 so that F (fn(x) − fm(x)) L1[R′∞I ](h) is dense in L1[R∞I ](h), it follows that F has a bounded extension, is a Cauchy sequence in C0[R∞I ](h). Since mapping L1[R∞I ](h) into C0[R∞I ](h). (cid:3) 5.2. L2-Theory. In the case of L2, the Fourier transform is an isometric isomorphism from L2[Rn] onto L2[Rn]. Theorem 5.5. The operator F is an isometric isomorphism of L2[R∞I ](h) onto L2[R∞I ](h). Proof. Let f ∈ L2[R∞I ](h). By construction, there exists a sequence of functions {fk ∈ L2[Rnk k→∞kf − fkk2 (h) = 0. Fur- thermore, since the sequence converges, it is a Cauchy sequence. Hence, I ], nk ∈ N} such that lim given ε > 0, there exists a N (ε) such that m, n ≥ N (ε) implies that kfm − fnk2 (h) < ε. Since F is an isometry, kF(fm) − F(fn)k2 (h) < ε, so that the sequence F(fk) is also a Cauchy sequence in L2[R∞I ](h). Thus, there is a f ∈ L2[R∞I ](h) with lim F(f ) = f . It is easy to see that f is unique. (h) = 0, and we can define f − F(fk)(cid:13)(cid:13)(cid:13)2 k→∞(cid:13)(cid:13)(cid:13) (cid:3) We can also prove a version of Theorems 5.4 and 5.5 for every separable Banach space (with a basis). Fix B and for each n, let Bn is clear that Bn , so that B is the norm closure of following have the same proofs as Theorems 5.4 and 5.5. I ⊂ Bn+1 I I . It I = B ∩ Rn n→∞Bn lim I . The Theorem 5.6. The operator F extends to a bounded linear mapping of L1[B](h) into C0[B](h). 50 GILL, PANTSULAIA, AND ZACHARY Theorem 5.7. The operator F is an isometric isomorphism from L2[B](h) onto L2[B](h). Theorems 5.4 - 5.7 show that ⊗∞i=1 hi is a strong (product) convergence vector for the Fourier transform operator F. In the L2-theory, we know that L2[R∞I ](h) and L2[R∞I ](h) are orthogonal subspaces of H2 ⊗. Thus, in this approach, the natural interpretation is that the Fourier transform induces a Pontryagin duality like theory that does not depend on the group structure of R∞I , but depends on the pairing of different function spaces. This approach is direct, constructive and applies to all separable Banach spaces (with a basis). Thus, the group structure of the underlying measure space plays no role. 5.3. Pontryagin Duality Theory II. In this section, we show that the standard form of Pontryagin duality theory is also possible, using the un- derlying measure space group structure. It is constructive but restrictive, in that, it does not apply to every separable Banach space with a basis. Let B be any uniformly convex separable Banach space UCB over the reals, so that B = B∗∗ (second dual). The next theorem follows from our theory of Lebesgue measure on Banach spaces. Theorem 5.8. If λB is our version of Lebesgue measure on B, then B and B∗ are also duals as character groups (i.e., B∗ = B). CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES 51 Proof. If we consider the restriction to L2[B, λB], we can define F directly by: (5.2) [F(f )](x∗) = f (x∗) =ZB exp{−2πihy, x∗i}f (y)dλB(y), where hy, x∗i is the pairing between B and B∗. From Plancherel's Theorem, we have: =(cid:16) f , f(cid:17)2 = (f, f )2 = kfk2 2 . It follows that B and B∗ are duals as character groups and 2 2 (cid:13)(cid:13)(cid:13) f(cid:13)(cid:13)(cid:13) f (y) =ZB∗ exp{2πihy, x∗i} f (x∗)dλB∗ (x∗). (cid:3) I , we can represent fn directly as a mapping from I = BI ∩ Rn If Bn I , λB] → L2[B∗,n L2[Bn , λB∗], by [F(fn)](x∗) = fn(x∗) =ZB I exp{−2πihy, x∗in}fn(y)dλB(y), where hy, x∗in is the restricted pairing of y and x∗ to Bn tively. It follows that equations 5.1 and 5.2 provide two distinct definitions I and B∗,n respec- I of the Fourier transform for B. Thus, in this approach the group structure of the underlying measure space changes. It is clear that representation for f (x∗) also applies if we use L1[B, λB], but in this case f (x∗) ∈ C0[B∗]. If we define y(·) mapping B → C, by y(x) = exp{−2πihy, x∗i}, then y(x) is a character of B. Furthermore, it is easy to see that (y1 + y2)(x) = 52 GILL, PANTSULAIA, AND ZACHARY y1(x)·y2(x). We now have the extension of the Pontryagin Duality Theorem to all UCB (with a basis). Theorem 5.9. If B is a UCB, then the mapping γx : B → γx(y) = y(x), is an isomorphism of topological groups. B, defined by In case B = H, is a Hilbert space, we can replace equation (5.2) by (5.3) f (x) = F[f ](x) =ZH exp{−2πihx, yiH}f (y)dλH(y), so that H is self-dual (as expected). From equation (5.3), we also get the expected result that: Fhexp{−π x2 H}i = exp{−π x2 H}. In closing, we observe that by Theorem 2.31 (see Example 2.32), if we use Gaussian measure on R∞, the dual character groups are R∞ and R∞ = R∞0 . From this we see that probability measures on (R∞, B[R∞]) induce a different character theory compared to that induced by λ∞ on (R∞I , B[R∞I ]). 5.4. Lp-Theory. We can obtain Lp[R∞I ] as in the construction of L1[R∞I ]. In this section we want to show the power of our approach to measure theory by establishing a version of Young's Theorem for every separable Banach space with a Schauder basis: Theorem 5.10. (Young) Let p, q, r ∈ [1,∞] with 1 r = 1 p + 1 q − 1. CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES 53 If f ∈ Lp[R∞I ] and g ∈ Lq[R∞I ], then the convolution of f and g, f ∗ g, exists (a.s.), belongs to Lr[R∞I ] and kf ∗ gkr 6 kfkp kgkq . Corollary 5.11. Let B be a separable Banach space with a Schauder basis and let p, q, r ∈ [1,∞] with 1 r = 1 p + 1 q − 1. If f ∈ Lp[B] and g ∈ Lq[B], then the convolution of f and g, f ∗ g, exists (a.s.), belongs to Lr[B] and kf ∗ gkr 6 kfkp kgkq . In order to prove Theorem 5.10, we first need the appropriate version of Fubini's Theorem. Since (R∞I , L∞I , λ∞) is a complete σ-finite measure space, a proof of the following may be found in Royden [RO] (see Theorems 19 and 20, pgs. 269-270): Theorem 5.12. (Fubini) If f ∈ L1[R∞I × R∞I ], then (1) for almost all x ∈ R∞I the function fx defined by fx(y) = f (x, y) ∈ L1[R∞I ](y): (2) for almost all y ∈ R∞I the function fy defined by fy(x) = f (x, y) ∈ L1[R∞I ](x): f (x, y)dλ∞(y) ∈ L[R∞I ](x); f (x, y)dλ∞(x) ∈ L[R∞I ](y); I (3) RR∞ (4) RR∞ I 54 GILL, PANTSULAIA, AND ZACHARY (5) f (x, y)d (λ∞ ⊗ λ∞) (x, y) I ×R∞ I ZR∞ =ZR∞ I "ZR∞ I f (x, y)dλ∞(y)# dλ∞(x) =ZR∞ I "ZR∞ I f (x, y)dλ∞(x)# dλ∞(y). Theorem 5.13. Let f, g ∈ L1[R∞I ], then (f ∗ g)(x) exists (a.s.); that is f (y)g(x − y) ∈ L1[R∞I ]. In addition, f ∗ g ∈ L1[R∞I ] and kf ∗ gk1 6 kfk1 kgk1 . Proof. First, it is easy to see that f (y)g(x − y) is a measurable function on R∞I . (There is no change from the case of Rn.) We can apply Fubini's theorem to get that: I ZR∞ =ZR∞ I (f ∗ g) (x)dλ∞(x) dλ∞(x)"ZR∞ =ZR∞ I I f (y)g(x − y)dλ∞(y)# =ZR∞ f (y)dλ∞(y) ·ZR∞ g(x)dλ∞(x). I I dλ∞(y)"ZR∞ I f (y)g(x − y)dλ∞(x)# It follows from the last equality that kf ∗ gk1 6 kfk1 kgk1. (cid:3) 5.4.1. Proof of Young's Theorem. CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES 55 Proof. First, assume that f and g are nonnegative and kfkp = kgkq = 1. Let 1 p . Now note that q and 1 q′ = 1 − 1 p′ = 1 − 1 1 1 q′ r 1 p′ + 1 r + (cid:16)1 − (cid:16)1 − 1 p = +(cid:18)1 − r(cid:17) q′ = p(cid:18) 1 p − r(cid:17) p′ = q(cid:18) 1 q − q(cid:19) +(cid:18)1 − r(cid:19) q′ = p(cid:18)1 − r(cid:19) p′ = q(cid:18)1 − 1 1 q 1 1 p(cid:19) = 1; q(cid:19) q′ = p; p(cid:19) p′ = q. 1 If we use Holder's inequality (for three functions), we can write (f ∗ g)(x) as: (f ∗ g) (x) =ZR∞ 6"ZR∞ I hf (y)p/rg(x − y)q/rihf (y)1−p/rg(x − y)1−q/ri dλ∞(y) "ZR∞ f (y)pg(x − y)qdλ∞(y)#1/r"ZR∞ dλ∞(y)#1/q′ f (y)(1−p/r)q′ I I I g(x − y)(1−q/r)p′ dλ∞(y)#1/p′ . This last inequality shows that I ⇒ f (y)pg(x − y)qdλ∞(y)#1/r (f ∗ g) (x) 6"ZR∞ f (y)pg(x − y)qdλ∞(y)# ⇒ (f ∗ g)r (x) 6 (f p ∗ gq) (x). (f ∗ g)r (x) 6"ZR∞ From Theorem 5.13, we have k(f ∗ g)rk1 6 kf pk1 kgqk1 = 1. In the general case, we know that f ∗ g exists (a.e.), so that f (y)g(x − y) ∈ L1[R∞I ]. But then, f (y)g(x − y) ∈ L1[R∞I ]. (cid:3) I In closing we note that, Beckner [BE] and Brascamp-Lieb [BL] have shown that on Rn we can write Young's inequality as kf ∗ gkr 6 (Cp,q,r;n)n kfkp kgkq, where Cp,q,r;n ≤ 1 is sharp. We conjecture that 1 is the sharp constant for R∞I . 56 GILL, PANTSULAIA, AND ZACHARY 6. Partial Differential Operators (Examples) In this section, we give examples of strong product and sum vectors for differential operators that have found interest in infinite dimensional analy- sis. Definition 6.1. For x ∈ R, 0 ≤ y < ∞ and 1 < a < ∞ define ¯g(x, y), ¯h(x) by: ¯g(x, y) = exp(cid:8)−yaeiax(cid:9) , ¯h(x) = ¯g(x, y)dy, x ∈ [− π otherwise . 2a , π 2a ], Z ∞ 0 0   The following properties of ¯g are easy to check: (1) (2) (3) so that ∂¯g(x, y) ∂x = −iayaeiax¯g(x, y), ∂¯g(x, y) ∂y = −aya−1eiax¯g(x, y), iy ∂¯g(x, y) ∂y = ∂¯g(x, y) ∂x . It is also easy to see that ¯h(x) is in L1[R] for x ∈ [− π 2a , π 2a ] and, (6.1) d¯h(x) dx =Z ∞ 0 ∂¯g(x, y) ∂x dy =Z ∞ 0 iy ∂¯g(x, y) ∂y dy. CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES 57 Integration by parts in the last expression of equation (6.1) shows that ¯h′(x) = −i¯h(x), so that ¯h(x) = ¯h(0)e−ix for x ∈ [− π R ∞0 exp{−ya}dy, an additional integration by parts shows that ¯h(0) = Γ( 1 2a ]. Since ¯h(0) = 2a , π a + 1). Let a = π 2a , π 1−ε , ¯h(x) = ¯hε(x), x ∈ [− π c expn ε2 2x2−ε2o , (6.2) fε(x) =  0, 2a ], where 0 < ε ≪ 1, and define x < ε/2, x > ε/2, where c is the standard normalizing constant. We now define ξ(x) = (¯h ∗ fε)(x), so that spt(ξ) = [− 1 2 ] = Iε. Thus, ξ(x) = 0, x /∈ Iε and otherwise, 2 , 1 ξ(x) =Z ∞ −∞ ¯h[x − z]fε(z)dz = e−ixZ ∞ −∞ eizfε(z)dz = αεe−ix. It follows from this that: α−1 ε ξ(ix) =  ex, x ∈ Iε 0, x /∈ Iε Define λε = λ and, I ε = ×∞k=1Iε, I ε n = ×∞k=n+1 and, λε ∞ = ⊗∞k=1λε. Example 6.2. In this example, we let hk(xk) = α−1 ε ξ(ixk), for each k ∈ N so that Dkhk = hk , for xk ∈ I. Let L2 D∞ = Q∞k=1 Dk and fn ∈ L2 D∞fn(x) = Dnfn(x) = Qn ∞]. I ] ∩ D(D∞), we can define D∞ on Rn l=1 Dlf(n)(x) ⊗(cid:0)⊗∞l=n+1hl(cid:1) , (a.s). This opera- tor is well-defined and has a closed densely defined extension to L2 I ] = L2[Rn I , λε ε[Rn ε[Rn I by If ε[R∞I ](h), where h = ⊗∞k=1hk. Thus, h is a strong product vector for D∞. 58 GILL, PANTSULAIA, AND ZACHARY The operator D∞ is required if we want to obtain the probability den- sity for a distribution function. (Note, by construction the density can be approximated from below by densities in a finite number of variables.) In the following example, we construct a general elliptic operator on L2 ε[R∞I ]. Example 6.3. If ∇ = (D1, D2,···) and σk : R∞I → R is a bounded analytic function for each k ∈ N, then let σ(x) = (σ1(x), σ2(x),···). We assume that ∞ ∞ (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) σj(x)σk(x) + Xj,k=1 Xk=1 We can now define ∆∞ by: bk(x)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)2 < ∞, where bk(x) =X∞ j=1 σj(x)Djσk(x). ∞ ∞ Xk=1 σj(x)σk(x)Dj Dk + For the same version of L2 ∆∞ = (σ(x) · ∇)2 = Xj,k=1 ε[R∞I ] as in the last example, if gn ∈ L2[Rn I ] ∩ D(∆∞) and cn(x) = P∞j=n+1 σj(x) [Djσk(x)], then ∆∞ is defined on Rn bk(x)Dk. I and ∆∞gn(x) = n Xj,k=1 σj(x)σk(x)Dj Dkgn(x) + bkDkψ(x) + cn(x)gn(x). n Xk=1 In order to obtain the same equation with cn(x) = 0, we use the version of L2 ε[R∞I ] defined with hk = ξI(0), k ≥ n + 1. In this case, for gn(x) ∈ ε[R∞I ] ∩ D(∆∞), we see that L2 ∆∞gn(x) = n Xj,k=1 σj(x)σk(x)Dj Dkgn(x) + bk(x)Dkgn(x). n Xk=1 CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES 59 In either case, ∆∞ is well-defined for each n and has a closed densely defined ε[R∞I ] and gn(x) → g(x) implies that limn→∞∆∞gn(x) = extension to L2 ∆∞g(x). It follows that different versions of L2[R∞I ] offer advantages for particular types of differential operators. (For other approaches, see [BK], [GZ] and [UM].) The following special cases have appeared in the literature (all can be obtained from the last example): (1) The natural infinite dimensional Laplacian: A = ∆∞ =X∞ i=1 i . ∂2(cid:14)∂x2 (2) The nonterminating diffusion generator in infinitely many variables (also known as the Ornstein-Uhlenbeck operator): A = 1 2 ∆∞ − Bx · ∇∞ = 1 2X∞ i=1 ∂2(cid:14)∂x2 i −X∞ i=1 bixi∂/∂xi. The infinite dimensional Laplacian of Umemura [UM]: A = ∞ Xi=1(cid:18) ∂2 i − ∂x2 xi c2 ∂ ∂xi(cid:19). Berezanskii and Kondratyev ([BK], pages 520-521) have also discussed op- erators analogous to (2) and (3). 60 GILL, PANTSULAIA, AND ZACHARY 6.1. Discussion. In a very interesting paper, Phillip Duncan Thompson [PDT] used the amplitudes of a set of orthogonal modes as the co-ordinates in an infinite-dimensional phase space. This allowed him to derive the prob- ability distribution for an ensemble of randomly forced two-dimensional vis- cous flows as the solution of the continuity equation for the phase flow. He obtained the following equation for the probability density: ∂ρ ∂t + ∞ Xk=1 Mk(x) ∂ρ ∂xk − ν ∞ Xk=1 (6.3) where ∂ ∂xk (cid:2)ρα2 kxk(cid:3) − ∂2ρ ∂x2 k ∞ Xk=1 = 0, Mk(x) = ∞ Xi=1 ∞ Xj=1 α2 j βijk αiαjαk (µiµjµk) 1 2 µk xixj. The coefficients βijk vanishes if any two indices are equal, is invariant under cyclic permutation of indices and reverses sign under non-cyclic permutation of indices, while the coefficients αi and µi are positive constants, determined by the problem. Thompson imposed the natural condition (6.4) ∞ Z−∞ ··· ∞ Z−∞ ρ(x, t) ∞ Yk=1 dxk = 1. At that time, he ran into the obvious mathematical criticism because equa- tion (6.4) was meaningless at the time. He also derived the equilibrium density (6.5) ρ0(x) = C exp(− 1 2 ν kx2 α2 k) . ∞ Xk=1 The results in section 2.5, see also equation (2.5), along with those in section 4.4, show that Thomson's paper was prescient. CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES 61 7. Conclusion In this paper we provided a reasonable version of Lebesgue measure on R∞, which together with the standard Gaussian measure on R∞, have al- lowed us to construct natural analogues of Lebesgue and Gaussian measure for every separable Banach space with a Schauder basis. We have extended the Fourier transform to L1[R∞, λ∞], L2[R∞, λ∞], defined sums and prod- ucts of unbounded operators, and presented a few constructive examples of partial differential operators in infinitely many variables. Acknowledgments. This work could not have been written without the generous help and critical remarks of Professor Frank Jones. We thank Pro- fessor Anatoly Vershik for appraising us of recent work by the Russian school and correcting some of our historical remarks. We would also like to thank an anonymous referee for corrections that have improve our presentation and for suggesting that we reconsider our approach to the Fourier transform and discuss its relationship to the Pontryagin duality theory, which led to a complete revision and extension of Section 5. References [BA] S. Banach Th´eorie des Op´erations lin´eaires, Monografj Matematyczn, Vol. 1, War- saw, (1932). [BA1] R. Baker "Lebesque measure" on R∞, Proc. Amer. Math. Soc. 113 (1991), 1023- 1029. 62 GILL, PANTSULAIA, AND ZACHARY [BA2] R. Baker "Lebesque measure" on R∞, II, Proc. Amer. Math. Soc. 132 (2004), 2577-2591. [BE] W. Beckner Inequalities in Fourier Analysis, Ann. of Math. 102 (1975), 159-182. [BK] Yu. M. Berezanskii and Yu. G. Kondratyev, Spectral methods in infinite dimensional analysis, Naukova Dumka, Kiev, (Russian) (1988). [BL] H. J. Brascamp and E. H. Lieb Best constants in Young's inequality, its converse, and its generalization to more than three functions, Ad. in Math. 20 (1976), 151-173. [BM] Y. Bakhtin and J. C. Mattingly, Malliavin calculus for infinite-dimensional systems with additive noise, arXiv:math.PR/0610754v1 (2006). [DI] J. Diestel , Sequences and Series in Banach Spaces, Grad. Texts in Math. Springer- Verlag, New York, (1984). [DP] G. DaPrato, An Introduction to Infinite-Dimensional Analysis, Springer Berlin, (2006). [DS] N. Dunford and J. T. Schwartz, Linear Operators Part I: General Theory, Wiley Classics edition, Wiley Interscience (1988). [EM] E. O. Elliott and A. P. Morse, General product measures, Trans. Amer. Math. Soc. 110, (1964) 245-283. [GU] A. Guichardet, Symmetric Hilbert Spaces and Related Topics, Lectures Notes in Mathematics, No. 261, Springer-Verlag, New York, (1969). [GZ] T. L. Gill and W. W. Zachary, Banach spaces of von Neumann type, Georgian International Journal of Science Technology 3 (2011), 1-35. [GZ1] T. L. Gill and W. W. Zachary, Feynman operator calculus: The constructive theory, Expositiones Mathematicae 29 (2011), 165-203. [HA] A. Haar, Der Massbegriff in der Theorie der kontinuierlichen Gruppe , Ann. Math., 34 (1933), 147-169. [HHK] H. H. Kuo , Gaussian Measures in Banach Spaces, Lecture Notes in Mathematics 463, Springer, New York (1975). CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES 63 [HI] D. Hill, σ-finite invariant measures on infinite product spaces, Trans. Amer. Math. Soc. 153, (1971) 347-370. [J] F. Jones , Lebesgue Integration on Euclidean Space, Revised Edition, Jones and Bartlett Publishers, Boston (2001). [KA] S. Kakutani, On equivalence of infinite product measures , Ann. Math., 49 (1948), 214-224. [KA1] S. Kaplan Extensions of Pontjagin duality I: Infinite products, Duke Math. J. 15 (1948), 649-659. [KA2] S. Kaplan Extensions of Pontjagin duality II: Direct and inverse sequences, Duke Math. J. 17 (1950), 419-435. [KH] A.B. Kharazishvili On invariant measures in the Hilbert space, Bull. Acad. Sci. Georgian SSR, 114 (1) (1984),4148 (in Russian). [KK] K. Kodaira and S. Kakutani, A nonseparable translation-invariant extension of Lebesgue measure space, Ann. Math., 52 (1950), 574-579. [KO] A. N. Kolmogorov, Grundbegriffe der Wahrscheinlichkeitsrechnung, Springer-Verlag, Vienna, (1933). [KP1] A. P. Kirtadze and G. R. Pantsulaia, Invariant measures in the space RN , (in Russian) Soobshch. Akad. Nauk Gruzii 141 (1991), 273-276. [KP2] A. P. Kirtadze and G. R. Pantsulaia, Lebesgue nonmeasurable sets and the unique- ness of invariant measures in infinite-dimensional vector spaces, Proc. A. Razmadze Math. Inst. 143 (2007), 95-101. [MO] C. C. Moore, Invariant measures on product spaces, Proc. 5th Berkeley Sym. Math. Stat. & Prob. (Berkeley, 1965) 2 Part 2, 447-459. [OX] J. C. Oxtoby, Invariant measures in groups which are not locally compact, Trans. Amer. Math. Soc. 60, (1946) 215237. [PA] G. Pantsulaia, Invariant and Quasiinvariant Measures in Infinite-Dimensional Topological Vector Spaces , Nova Science Publishers, New York, (2007). 64 GILL, PANTSULAIA, AND ZACHARY [PDT] P. D. Thompson, Some exact statistics of two-dimensional viscous flow with random forcing, Journal of Fluid Mechanics 55 (1972), 711-717. [RE] M. C. Reed, On self-adjointness in infinite tensor product spaces, Journal of Func- tional Analysis 5 (1970), 94-124. [RH] G. E. Ritter and E. Hewitt, Elliott-Morse measures and Kakutanis dichotomy the- orem, Math. Zeitschrift 211, (1992) 247263. [RO] H. L. Royden, Real Analysis, (2nd Ed.) Macmillan Press, New York, (1968). [RU] W. Rudin, Functional Analysis, McGraw-Hill Press, New York, (1973). [RU1] W. Rudin, Fourier Analysis on Groups, John Wiley & Sons, New York, (1990). [ST] L. Streit, Test function spaces for direct product representations, Commun. Math. Phys. 4 (1967), 22-31. [SU] V.N. Sudakov, Linear sets with quasi-invariant measure, Dokl.Akad.Nauk SSSR; 127 (1959), 524-525 (in Russian). [T] P. D. Thompson Some exact statistics of two-dimensional viscous flow with random forcing, J. Fluid Mech. 55 (1972), 711-717. [UM] Y. Umemura, On the infinite dimensional Laplacian operator, J. Math. Kyoto Univ. 4 (1964/1965), 477-492. [V] A. M. Vershik, Duality in the theory of measure in linear spaces, (English translation): Sov. Math. Dokl. 7, (1967) 1210-1214. [V1] A. M. Vershik, Does there exist the Lebesgue measure in the infinite-dimensional space?, Proceedings of the Steklov Institute of Mathematics, 259 (2007), 248-272. [V2] A. M. Vershik, The behavior of Laplace transform of the invariant measure on the hyperspace of high dimension, J. Fixed Point Theory Appl., 3 (2008), 317-329. [V3] D. L. Vandev, Invariant measures for the continual Cartan subgroup, J. Functional Analysis, 255 (2008), 2661-2682. CONSTRUCTIVE ANALYSIS IN INFINITELY MANY VARIABLES 65 [VA] A. M. Vershik, Action of groups on an infinite product measure, (in Russian) Pa- pers presented at the Fifth Balkan Mathematical Congress (Belgrade, 1974). Math. Balkanica 4 (1974), 643647. [VN1] J. von Neumann, The uniqueness of Haar's measure, Rec. Math. (Mat. Sbornik) N.S., 1 (1936), 721-734. [VN2] J. von Neumann, On infinite direct products, Compositio Mathematica, 6 (1938), 1-77. [WE] A. Weil, L'int´egration dans les groupes topologiques et ses applications, Actualit´es Scientifiques et Industrielles, no. 869, Paris, 1940. [WSRM] N. Wiener, A. Siegel, W. Rankin and W. T. Martin, Differential Space, Quantum Systems, and Prediction, M. I. T. Press, Cambridge, MA, (1966). [YA] Y. Yamasaki, Measures on infinite-dimensional spaces, World Scientific, Singapore, (1985). [YA1] Y. Yamasaki, Translationally invariant measure on the infinite-dimensional vector space, Publ. Res. Inst. Math. Sci. 16 (3) (1980), 693720. (Tepper L. Gill) Department of Mathematics, Physics and E&CE, Howard University, Washington DC 20059, USA, E-mail : [email protected] (Gogi R. Pantsulaia) Department of Mathematics, Georgian Technical Uni- versity, Tbilisi 0175, Georgia; I. Vekua Institute of Applied Mathematics, Tbilisi State University, Tbilisi 0143, Georgia E-mail : g:pantsulaia@gtu:ge (Woodford W. Zachary) Department of Mathematics and E&CE, Howard Uni- versity, Washington DC 20059, USA, E-mail : [email protected]
1503.06750	3	1503	2015-04-06T09:44:40	$C^{*}$ algebra and inverse chaos	[ "math.FA" ]	If an invertible linear dynamical systems is Li-York chaotic or other chaotic, what's about it's inverse dynamics? what's about it's adjoint dynamics? With this unresolved but basic problems, this paper will give a criterion for Lebesgue operator on separable Hilbert space. Also we give a criterion for the adjoint multiplier of Cowen-Douglas functions on $2$-th Hardy space. Last we give some chaos about scalars perturbation of operator and some examples of invertible bounded linear operator such that $T$ is chaotic but $T^{-1}$ is not.	math.FA	math	C ∗ algebra and inverse chaos Luo Lvlin∗ and Hou Bingzhe† January 24, 2015 Abstract: If an invertible linear dynamical systems is Li-York chaotic or other chaotic, what's about it's inverse dynamics? what's about it's adjoint dynamics? With this unresolved but basic problems, this paper will give a criterion for Lebesgue operator on separable Hilbert space. Also we give a criterion for the adjoint multiplier of Cowen-Douglas functions on 2-th Hardy space. Last we give some chaos about scalars perturbation of operator and some examples of invertible bounded linear operator such that T is chaotic but T −1 is not. Keywords: inverse, chaos, Hardy space, rooter function, Cowen-Douglas function, Spectrum, C∗ algebra, Lebesgue operator. 1.Introduction ∗E-mail:[email protected] and Adrress:Department of Mathematics,Jilin Univer- sity, ChangChun,130012,P. R. China. †E-mail:[email protected] and Adrress:Department of Mathematics,Jilin University, ChangChun,130012,P. R. China. 1 The ideas of chaos in connection with a map was introduced by Li T.Y.and his teacher Yorke,J.A.[22], after that there are various definitions of what it means for a map to be chaotic and there is a series of papers on Topological Dynamics and Ergodic Theory about chaos, such as [27][11][17][13][23]. Following Topological Dynamics,Linear Dynamics is also a rapidly evolv- ing branch of functional analysis,which was probably born in 1982 with the Toronto Ph.D.thesis of C.Kitai [3]. It has become rather popular be- cause of the efforts of many mathematicians, for the seminal paper [6] by G.Godefroy and J.H.Shapiro,the notes [12] by J.H.Shapiro,the authoritative survey [20] by K.-G.Grosse-Erdmann,,and finally for the book [4] by F.Bayart and E.Matheron,the book [19] by K.-G.Grosse-Erdmann and A.Peris. For finite-dimension linear space,the authors have made a topologically conjugate classification about Jordan blocks in [18][24]. So for the eigenvalues λ 6= 1, the operators of Jordan block are not Li-Yorke chaotic.With [4] we know that a Jordan block is not supercyclic when its eigenvalues λ = 1, and following a easy discussion it is not Li-Yorke chaotic too. So the definition of Li-Yorke chaos should be valid only on infinite- dimension Frechet space or Banach space such that in this paper the Hilbert space is infinite-dimensional. Because a finite-dimensional linear operator could be regard as a compact operator on some Banach spaces or some Hilbert spaces, we can get the same conclusion from [25]P12 or the Theorem 7 of [10]. For a Frechet space X,let L(X) denote the set of all bounded linear 2 operators on X. Let B denote a Banach space and let H denote a Hilbert space. If T ∈ L(B), then define σ(T ) = {λ ∈ C; T − λ is not invertible} and define rσ(T ) = sup{λ; λ ∈ σ(T )}. Definition 1. Let T ∈ L(B),if there exists x ∈ B satisfies: (1) lim n→∞ T n(x)k > 0; and (2) lim n→∞ kT n(x)k = 0. Then we say that T is Li-Yorke chaotic,and named x is a Li-Yorke chaotic point of T ,where x ∈ B, n ∈ N. Define a distributional function F n x (τ ) = 1 n ♯{0 ≤ i ≤ n : kT n(x)k < τ }, where T ∈ L(B), x ∈ B, n ∈ N. And define Fx(τ ) = lim inf n→∞ F n x (τ ); and F ∗x (τ ) = lim sup n→∞ F n x (τ ). Definition 2. Let T ∈ L(B),if there exists x ∈ B and (1) If Fx(τ ) = 0, ∃τ > 0, and F ∗x (ǫ) = 1, ∀ǫ > 0, then we say that T is distributional chaotic or I-distributionally chaotic . (2) If F ∗x (ǫ) > Fx(τ ), ∀τ > 0, and F ∗x (ǫ) = 1, ∀ǫ > 0, then we say that T is II-distributionally chaotic. (3) If F ∗x (ǫ) > Fx(τ ), ∀τ > 0,then we say that T is III-distributionally chaotic. Definition 3 ([25]). Let X is an arbitrary infinite-dimensional separable Frechet space, T ∈ L(X),If there exists a subset X0 of X satisfies: (1) For any x ∈ X0, {T nx}∞n=1 has a subsequence converging to 0; 3 (2) There is a bounded sequence {an}∞n=1 in span(X0) such that the se- quence {T nan}∞n=1 is unbounded. Then we say T satisfies the Li-Yorke Chaos Criterion. Theorem 1 ([25]). Let X is an arbitrary infinite-dimensional separable Frechet space, If T ∈ L(X),then the following assertions are equivalent. (i) T is Li-Yorke chaotic; (ii) T satisfies the Li-Yorke Chaos Criterion. Lemma 1 ([29]). There are no hypercyclic operators on a finite-dimensional space X 6= 0. Example 1 ([4]P8). Let φ ∈ H∞(D) and let Mφ : H2(D) → H2(D) be the associated multiplication operator. The adjoint multiplier M∗φ is hypercyclic if and only if φ is non-constant and φ(D)T T 6= ∅. Theorem 2 ([5]). Let X is an arbitrary separable Frechet space, T ∈ L(X).The following assertions are equivalent. (i) T is hypercyclic. (ii) T is topologically transitive; that is,for each pair of non-empty open sets U, V ⊆ X, there exists n ∈ N such that T n(U)T V 6= ∅. Theorem 3 ([4]). Let X is a topological vector space,T is a bounded linear 4 operator on X.Let   Λ1(T ) , span( Sλ=1,n∈N Λ+(T ) , span(Λ1(T )S Sλ>1,n∈N Λ−(T ) , span(Λ1(T )S Sλ<1,n∈N ker(T − λ)N T ran(T − λ)N ); ker(T − λ)N ); ker(T − λ)N ). If Λ+(T ) and Λ−(T ) are both dense in X,then T is mixing. 2.From Polar Decomposition to functional calculus The Polar Decomposition Theorem [15]P15 on Hilbert space is a useful theorem, especially for invertible bounded linear operator. We give some properties of C∗ algebra generated by normal operator. Let H be a separable Hilbert space over C and let X be a compact subset of C. Let C(X) denote the linear space of all continuous functions on the compact space X, let P(x) denote the set of all polynomials on X and let T be an invertible bounded linear operator on H. By the Polar Decomposition Theorem [15]P15 we get T = UT , where U is an unitary operator and T 2 = T ∗T . Let A(T ) denote the C∗ algebra generated by the positive operator T and 1. Lemma 2. Let 0 /∈ X be a compact subset of C. If P(x) is dense in C(X), then P( 1 x ) is also dense in C(X). Proof. By the property of polynomials we know that P( 1 x ) is a algebraic closed subalgebra of C(X) and we get: 5 (1) 1 ∈ P( 1 x ); (2) P( 1 x ) separate the points of X; (3) If p( 1 x) ∈ P(x),then ¯p( 1 x ) ∈ P(x). By the Stone-Weierstrass Theorem [14]P145 we get the conclusion. Lemma 3. Let X ⊆ R+. If P(x) is dense in C(X), then P(x2) is also dense in C(X). Proof. For X ⊆ R+, x 6= y ⇐⇒ x2 6= y2. By Lemma 2 we get the conclusion. By the GNS construction [14]P250 for the C∗ algebra A(T ), we get the following decomposition. Lemma 4. Let T be an invertible bounded linear operator on the separable Hilbert space H over C, A(T ) is the complex C∗ algebra generated by T and 1. There is a sequence of nonzero A(T )-invariant subspace. H1, H2, · · · such that: (1) H = H1 L H2 L · · · ; (2) For every Hi, there is a A(T )-cyclic vector ξi such that Hi = A(T )ξi and T Hi = Hi = T −1Hi. Proof. By [30]P54 we get (1),and T Hi ⊆ Hi,that is Hi ⊆ T −1Hi; by Lemma 2 we get T −1Hi ⊆ Hi. Hence we get T Hi = Hi = T −1Hi. 6 For ∀n ∈ N,T n is invertible when T is invertible. By the Polar Decom- position Theorem [15]P15 T n = UnT n, where Un is unitary operator and T n2 = T ∗nT n, we get the following conclusion. Lemma 5. Let T be an invertible bounded linear operator on the separable Hilbert space H over C, let A(T k) be the complex C∗ algebra generated by T k and 1 and let HT k invariant subspace, there is a decomposition H = Li Given a proper permutation of HT k and T −1HT k k be a sequence of non-zero A(T k)- k ∈ H, i, k ∈ N. i = HT (k+1) , we get T ∗HT k and HT (k+1) = A(T k)ξi i = HT (k+1) HT k , ξi . j i i i i i Proof. By Lemma 3,it is enough to prove the conclusion on P(T k2)ξi k. For any given ξi HT (k+1) . j k ∈ H = Lj HT (k+1) j , there is a unique j ∈ N such that ξi k ∈ (1) Because T is invertible, for any given ξi k, there is an unique ηi ∈ k. For ∀p ∈ P(x2), we get p(T (k+1)2)ηi = HT (k+1) s such that ηi = T −1ξi T ∗p(T k2)ξi k. Hence we get HT (k+1) s = P(T (k+1)2)ξs k+1 ⊇ T ∗P(T k2)ξi k = T ∗HT k i . (2) Similarly,for any given ξs k+1, there is an unique ηr ∈ HT k r such that ηr = T ξs k+1. For ∀p ∈ P(x2),we get p(T k2)ηr = p(T k2)T ξs k+1 = T ∗−1p(T (k+1)2)ξs k+1. Hence we get HT k r = P(T k2)ξr k ⊇ T ∗−1P(T (k+1)2)ξs k+1 = T ∗−1HT (k+1) s . 7 Let i = r,by (1)(2) we get T ∗−1HT (k+1) s ⊆ HT k i ⊆ T ∗−1HT (k+1) j . Fixed the order of HT k i = HT (k+1) . i i T ∗HT k , by a proper permutation of HT (k+1) j we get By Lemma 2 and T is invertible,we get (3) For any given ξi k, there is an unique ηi ∈ HT (k+1) i such that ηi = T ∗ξi k. For ∀p ∈ P(x−2), we get p(T (k+1)−2)ηi = T −1p(T k−2)ξi k. Hence we get HT (k+1) i = P(T (k+1)−2)ξi k+1 ⊇ T −1P(T k−2)ξi k = T −1HT k i . (4) For any given ξi k+1,there is an unique ηi ∈ HT k i such that ηi = T ∗−1ξi k+1. For ∀p ∈ P(x−2),we get T −1p(T k−2)ηi = T −1p(T k−2)T ∗−1ξi k+1 = p(T (k+1)−2)ξi k+1. Hence we get T −1HT k i = T −1P(T k−2)ξi k ⊇ P(T (k+1)−2)ξi k+1 = HT (k+1) i . By (3)(4) we get T −1HT k i ⊆ HT (k+1) i ⊆ T −1HT k i . that is,T −1HT k i = HT (k+1) i . Let ξ ∈ H is a A(T )-cyclic vector such that A(T )ξ is dense in H. Because of σT 6= ∅, on C(σ(T )) define the non-zero linear functional ρT:ρT(f ) =< f (T )ξ, ξ >, ∀f ∈ C(σ(T )). Then ρT is a positive lin- ear functional, by [30]P54 and the Riesz-Markov Theorem, on C(σ(T )) we 8 get that there exists an unique finite positive Borel measure µT such that f (z) dµT(z) =< f (T )ξ, ξ >, ∀f ∈ C(σ(T )). Rσ(T) Theorem 4. Let T be an invertible bounded linear operator on the separable Hilbert space H over C, there is ξ ∈ H such that A(T )ξ = H. For any given n ∈ N, let A(T n) be the complex C∗ algebra generated by T n and 1 and let ξn be a A(T n)-cyclic vector such that A(T n)ξn = H. Then: (1) For any given ξn,there is an unique positive linear functional Rσ(T n) f (z) dµT n(z) =< f (T n)ξn, ξn >, ∀f ∈ L2(σ(T n)). (2) For any given ξn,there is an unique finite positive complete Borel measure µT n such that L2(σ(T n)) is isomorphic to H. Proof. Because T is invertible, by Lemma 5 we get that if there is a A(T )- cyclic vector ξ, then there is a A(T n)-cyclic vector ξn. (1):For any given ξn, define the linear functional, ρT n(f ) =< f (T n)ξn, ξn >, by [30]P54 and the Riesz-Markov Theorem we get that on C(σ(T n)) there is an unique finite positive Borel measure µT n such that Rσ(T n) f (z) dµT n(z) =< f (T n)ξn, ξn >, ∀f ∈ C(σ(T n)). Moreover we can complete the Borel measure µT n on σ(T n), also us- ing µT n to denote the complete Borel measure, By [26] we know that the complete Borel measure is uniquely. 9 For ∀f ∈ L2(σ(T n)),because of ρT n(f 2) = ρT n( ¯f f ) =< f (T n)∗f (T n)ξn, ξn >= kf (T n)ξnk ≥ 0. we get that ρT n is a positive linear functional,hence (1) is right. (2) we know that C(σ(T n)) is a subspace of L2(σ(T n)) such that C(σ(T n)) is dense in L2(σ(T n)). For any f, g ∈ C(σ(T n)) we get < f (T n)ξn, g(T n)ξn >H=< g(T n)∗f (T n)ξn, ξn > = ρT n(¯gf ) = Rσ(T n) f (z)¯g(z) dµT n(z) =< f, g >L2(σ(T n)) . Therefor U0 : C(σ(T n)) → H, f (z) → f (T n)ξn is a surjection isometry from C(σ(T n)) to A(T n)ξn, also C(σ(T n)) and A(T n)ξn is a dense sub- space of L2(σ(T n)) and H,respectively. Because of U0 a closable operator,it closed extension U : L2(σ(T n)) → H, f (z) → f (T n)ξn is a unitary opera- tor. Hence for any given ξn, U is the unique unitary operator induced by the unique finite positive complete Borel measure µT n such that L2(σ(T n)) is isomorphic to H. By the Polar Decomposition Theorem [15]P15, we get U∗T ∗T U = T T ∗ and U∗T −2U = T −2 when T = UT . In fact,when T is invertible, we can choose an specially unitary operator such that T −1 and T −1 are unitary equivalent. We give the following unitary equivalent by Theorem 4. Theorem 5. Let T be an invertible bounded linear operator on the separable Hilbert space H over C and let A(T ) be the complex algebra generated by 10 T and 1. There is σ(T −1) = σ(T −1) and we get that T −1 and T −1 are unitary equivalent by the unitary operator F H xx∗, more over the unitary opera- xx∗ is induced by an almost everywhere non-zero function pφT,where tor F H pφT ∈ L∞(σ(T ), µT). That is,d µT −1 = φTd µT−1. Proof. By Lemma 4,lose no generally,let ξT is a A(T )-cyclic vector such that H = A(T )ξT. (1) Define the function Fx−1 : P(x) → P(x−1), Fx−1(f (x)) = f (x−1), it is easy to find that Fx−1 is linear.Because of f (z−1)dµT(z) =< f (T −1)ξ, ξ > = Rσ(T−1) Rσ(T) We get dµT−1(z) = z2dµT(z).Hence f (z)dµT−1(z). Fx−1(f (x)) ¯Fx−1(f (x))dµT−1(x) f (x−1) ¯f (x−1)dµT−1(x) kFx−1(f (x))kL2(σ(T−1),µT −1 ) = Rσ(T−1) = Rσ(T−1) = Rσ(T) ≤ sup σ(T )2 Rσ(T) ≤ sup σ(T )2kf (z)kL2(σ(T),µT ). z2f (z) ¯f (z)dµT(z) f (z) ¯f (z)dµT(x) So we get kFx−1k ≤ sup σ(T )2, by the Banach Inverse Mapping Theo- rem [14]P91 we get that Fx−1 is an invertible bounded linear operator from 11 L2(σ(T ), µT) to L2(σ(T −1), µT−1). Define the operator F H x−1 : A(T )ξ → A(T −1)ξ, F (f (T )ξ) = f (T −1)ξ. By Lemma 2 and [30]P55, we get that F H x−1 is a bounded linear operator on the Hilbert space A(T )ξ = H, kF H x−1k ≤ sup σ(T )2.Moreover we get H T −−−−−−−−→ H F H x−1 ↓ H −−−−−−−−−−→ T −1 ↓ F H x−1 H (2) Define the function Fxx∗ : P(xx∗) → P(x∗x), Fxx∗(f (x, x∗)) = f (x∗x). By [8] we get that Fxx∗ is a linear algebraic isomorphic. For any x ∈ σ(T −1), by Lemma 3 we get that P(x2) is dense in C(x), C(x) is dense in L2(σ(T −1), µT −1). Hence P(x2) is dense in L2(σ(T −1), µT −1). With a similarly discussion, for any y ∈ σ(T −1),we get that P(y2) is dense in L2(σ(T −1), µT−1). So Fxx∗ is an invertible bounded linear operator from L2(σ(T −1), µT−1) to L2(σ(T −1), µT −1). Therefor Fxx∗ ◦ Fx−1 is an invertible bounded linear operator from L2(σ(T ), µT) to L2(σ(T −1), µT −1). Because of λ ∈ σ(T ∗T ) ⇐⇒ 1 λ ∈ σ(T ∗−1T −1), we get that λ ∈ σ(T ) ⇐⇒ 1 λ ∈ σ(T −1), that is,σ(T −1) = σ(T −1). For any pn ∈ P(σ(T −1)) ⊆ A(σ(T −1)), because of T ∗−1pn(T −1) = 12 pn(T −1)T ∗−1, by [15]P60 we get that P(T −1) and P(T −1) are unitary equivalent. Hence there is an unitary operator U ∈ B(H) such that UP(T −1) = P(T −1)U, that is,UA(T −1) = A(T −1)U and UA(T −1)ξT = A(T −1)UξT. If let ξT −1 = UξT,then ξT −1 is a A(T −1)-cyclic vector and A(T −1)ξT −1 = H.Because of f ( 1 z )dµT(z). f (z)dµT−1(z) =< f (T −1)ξT, ξT > = Rσ(T) f (z)dµT −1(z) =< f (T −1)ξT −1, ξT −1 >. Rσ(T−1) Rσ(T −1) We get [dµT −1] = [dµT−1],that is, dµT −1 and dµT−1 are mutually ab- solutely continuous, by [14]IX.3.6Theorem and (1) we get that dµT −1 = φT( 1 z )dµT−1 = z2φT(z)dµT, where φT(z) 6= 0, a.e. and φT(z) ∈ L∞(σ(T ), µT). So we get Fxx∗ ◦ Fx−1(f (x))Fxx∗ ◦ Fx−1(f (x))dµT −1(x) f (x−1) ¯f (x−1)dµT −1(x) kFxx∗ ◦ Fx−1(f (x))kL2(σ(T −1),µT −1) = Rσ(T −1) = Rσ(T −1) = Rσ(T−1) = Rσ(T−1) = kqφT( 1 φT( 1 x )f (x−1) ¯f (x−1)dµT−1(x) Fx−1(pφT(x)f (x))Fx−1(pφT(x) ¯f (x))dµT−1(x) x)Fx−1(f (x))kL2(σ(T−1),µT −1 ) Hence Fxx∗ is an unitary operator that is induced by Uf ( 1 x) = qφT( 1 x )f ( 1 x). 13 Define F H xx∗:   A(T −2)ξT → A(T −2)ξT −1, xx∗(f (T −2)ξT) = f (T −2)ξT −1. F H Therefor F H By Lemma 4 and [8] we get A(T −1)ξ = H = A(T −1)ξ. That is,F H xx∗ is an unitary operator from A(T −1)ξT to A(T −1)ξT −1. xx∗ is an unitary operator and we get H T −1 −−−−−−−−−−→ H F H xx∗ ↓ H −−−−−−−−−−→ T −1 ↓ F H xx∗ H So T −1 and T −1 are unitary equivalent by F H xx∗ is induced by the function qφT( 1 F H Corollary 1. Let T be an invertible bounded linear operator on the separable xx∗, the unitary operator x ). Hilbert space H over C and let A(T ) be the complex algebra generated by T and 1. There is σ(T ) = σ(T ∗) and we get that T and T ∗ are unitary equivalent by the unitary operator F H xx∗ is induced by an almost everywhere non-zero function pφT,where pφT ∈ L∞(σ(T ), µT). That is,d µT ∗ = φTd µT. xx∗, more over the unitary operator F H 3.The chaos between T and T ∗−1 for Lebesgue operator For the example of singular integral in mathematical analysis, we know that is independent the convergence or the divergence of the weighted integral between x and x−1, however some times that indeed dependent for a special weighted function. For T is an invertible bounded operator on the separable Hilbert space H over C, we get 0 /∈ σ(T n) ⊆ R+. In the view of the 14 singular integral in mathematical analysis and by Theorem 4, we get that T and T ∗−1 should not be convergence or divergence at the same time for T is an invertible bounded operator. Anyway,they should be convergence or divergence at the same time for some special operators. Therefor we define the Lebesgue operator and prove that T and T ∗−1 are Li-Yorke chaotic at the same time for T is a Lebesgue operator. Then we give an example that T is a Lebesgue operator,but not is a normal operator. Let dx be the Lebesgue measure on L2(R+), by Theorem 4 we get that dµT n is the complete Borel measure and L2(σ(T n)) is a Hilbert space. If ∃N > 0, for ∀n ≥ N, n ∈ N,dµT n is absolutely continuity with respect to dx, by the Radon-Nikodym Theorem [14]P380 there is fn ∈ L1(R+) such that dµT n = fn(x) dx. Definition 4. Let T be an invertible bounded linear operator on the separable Hilbert space H over C, moreover if T satisfies the following assertions: (1) If ∃N > 0,for ∀n ≥ N, n ∈ N dµT n = fn(x) dx, x2fn(x) = fn(x−1), fn ∈ L1(R+). 0 < x ≤ 1.   (2) If ∃N > 0, for ∀n ≥ N, n ∈ N,there is a A(T n)-cyclic vector ξn. And for any given 0 6= x ∈ H and for any given 0 6= gn(t) ∈ L2(σ(T n)), there is an unique 0 6= y ∈ H such that y = gn(T n−1)ξn when x = gn(T n)ξn. Then we say that T is a Lebesgue operator, let LLeb(H) denote the set of all Lebesgue operators. 15 Theorem 6. Let T be a Lebesgue operator on the separable Hilbert space H over C, then T is Li-Yorke chaotic if and only if T ∗−1 is. Proof. We prove the conclusion by two parts. (1) Let H be A(T )-cyclic, that is,there is a vector ξ such that A(T )ξ = H. By Lemma 5 we get that if there is a A(T )-cyclic vector ξ, then there is also a A(T n)-cyclic vector ξn. Let x0 be a Li-Yorke chaotic point of T , by Theorem 4 and the Polar Decomposition Theorem [15]P15 and by the define of Lebesgue operator,we get that for enough large n ∈ N,there are gn(x) ∈ L2(σ(T n)), fn(x) ∈ L2(R+) and y0 ∈ H such that x0 = gn(T n)ξn,y0 = gn(T n−1)ξn and dµT n = fn(x) dx. kT nx0k =< T n∗T nx0, x0 > =< T n2gn(T n)ξn, gn(T n)ξn > =< gn(T n)∗T n2gn(T n)ξn, ξn > x2gn(x)¯g(x) dµT n(x) x2gn(x)2fn(x) dx 0 x2gn(x)2fn(x) dx + R +∞ 0 x2gn(x)2fn(x) dx + R 1 0 gn(x)2fn(x−1) dx + R 1 1 x−2gn(x−1)2fn(x) dx + R 1 x−2gn(x−1)2fn(x) dx 0 0 = Rσ(T n) = R +∞ = R 1 = R 1 , R 1 = R +∞ = R +∞ 1 x2gn(x)2fn(x) dx 0 x−4gn(x−1)2fn(x−1) dx 0 x−2gn(x−1)2fn(x) dx 0 x−2gn(x−1)2fn(x) dx 16 = Rσ(T n) x−2gn(x−1)¯gn(x−1) dµT n(x) =< gn(T n)∗T n−2gn(T n−1)ξn, ξn > =< T n−2gn(T n−1)ξn, gn(T n−1)ξn > =< T n−2y0, y0 > =< T −nT −n∗y0, y0 > = kT ∗−ny0k. Where , following the define of fn(x). (2) If H is not A(T )-cyclic, by Lemma 5 we get that for ∀n ∈ N, there k ∈ H, i, k ∈ N, where HT k is a decomposition H = Li is a sequence of A(T k)-invariant subspace, and do (1) for HT k HT k , ξi i i = A(T k)ξi k . i By (1)(2) we get that T is Li-Yorke chaotic if and only if T ∗−1 is Li-Yorke chaotic. Corollary 2. Let T be a Lebesgue operator on the separable Hilbert space H over C, then T is I-distributionally chaotic (or II-distributionally chaotic or III-distributionally chaotic) if and only if T ∗−1 is I-distributionally chaotic (or II-distributionally chaotic or III-distributionally chaotic). Theorem 7. There is an invertible bounded linear operator T on the sep- arable Hilbert space H over C, T is Lebesgue operator but not is a normal operator. Proof. Let 0 < a < b < +∞, then L2([a, b]) is a separable Hilbert space over R, because any separable Hilbert space over R can be expanded to a separable Hilbert space over C, it is enough to prove the conclusion on L2([a, b]). We 17 prove the conclusion by six parts: (1) Let 0 < a < 1 < b < +∞,M = {[a, b−a 2 ], [ b−a 2 , b]}. Construct measure preserving transformation on [a, b]. There is a Borel algebra ξ(M) generated by M, define Φ : [a, b] → [a, b], Φ([a, b−a 2 ]. Then Φ is an invertible measure 2 ]) = [ b−a 2 , b], Φ([ b−a 2 , b]) = [a, b−a preserving transformation on the Borel algebra ξ(M). By [27]P63,UΦ 6= 1 is a unitary operator induced by Φ, where UΦ is the composition UΦh = h ◦ Φ, ∀h ∈ L2([a, b]) on L2([a, b]). (2) Define Mxh = xh on L2([a, b]),then Mx is an invertible positive oper- ator. (3) For f (x) = ln x x , x > 0, define dµ = f (x)d x. then f (x) is continuous and f (x) > 0, a.e.x ∈ [a, b], hence d µ that is absolutely continuous with respect to d x is finite positive complete Borel measure, therefor L2([a, b], d µ) a separable Hilbert space over R. Moreover L2([a, b]) and L2([a, b], d µ) are unitary equivalent. (4) Let T = UΦMx,we get T ∗T = UΦT T ∗U∗Φ and UΦ 6= 1. Because of UΦMx 6= MxUΦ and UΦMx2 6= UΦMx2, we get that T is not a normal operator and σ(T ) = [a, b]. (5) The operator T = UΦMx on L2([a, b]) is corresponding to the operator T ′ on L2([a, b], d µ),T ′ is invertible bounded linear operator and is not a normal operator and σ(T ′) = [a, b]. 18 (6) From R b 1 n−1 n we get that fn(t) is continuous and almost everywhere positive, hence fn(t)d t a xnf (x)d x = R bn d t, let fn(t) = 1 n I[an,bn]f (t an tf (t 1 n−1 n nt t , 1 n ) 1 n ) is a finite positive complete Borel measure. For any E ⊆ R+ define IE = 1 when x ∈ E else IE = 0,so IE is the identity function on E. (i) fn(t−1) = 1 n I[an,bn]f (t− 1 n ) 1 n−1 n − t = 1 nI[an,bn]t ln t 1 n = 1 n I[an,bn] ln t − t 1 n 1 n−1 n − t − 1 n (i) t2fn(t) = 1 n I[an,bn]f (t 1 n ) t2 n−1 n t = 1 nI[an,bn]t ln t 1 n = 1 n I[an,bn] ln t 1 n t 1 n t2 n−1 n t By (i)(ii) we get x2fn(x) = fn(x−1). From σ(T ′n) = [an, bn] and R bn t2fn(t)dt, let d µT ′n = fn(t)d t,then d µT ′n is the finite positive complete Borel measure. d t = R +∞ an t2f (t 1 n−1 n nt 0 1 n ) For any given 0 6= h(x) ∈ L2([a, b]) we get 0 6= h(x−1) ∈ L2([a, b]). I[a,b] is a A(M n x )-cyclic vector of the multiplication M n x = Mxn, and I[an,bn] is a A(T ′n)-cyclic vector of T ′n, By Definition 4 we get that T ′ is Lebesgus operator but not is a normal operator. Corollary 3. There is an invertible bounded linear operator T on the sepa- rable Hilbert space H over C, T is a Lebesgue operator and also is a positive operator. 19 By the cyclic representation of C∗ algebra and the GNS construction, also by the functional calculus of invertible bounded linear operator, we could study the operator by the integral on R. This way neither change Li-Yorke chaotic nor the computing, but by the singular integral in mathematical analysis on the theoretical level we should find that there is a invertible bounded linear operator T that is Li-Yorke chaotic but T −1 is not Li-Yorke chaotic. 4. Cowen-Douglas function on Hardy space r<1 R π For D = {z ∈ C, z < 1}, if g is an complex analytic function on D and −π g(reiθ)2 dθ < +∞, then we denote g ∈ H2(D), hence H2(D) 2π , especially H2(D) is denoted as a Hardy space. By the completion theory of complex is a Hilbert space with the norm kgk2 −π g(reiθ)2 dθ there is sup H2 = sup r<1 R π analytic functions, the Hardy space H2(D) is a special Hilbert space that is relatively easy not only for theoretical but also for computing, and we should give some properties about adjoint multiplier operators on H2(D). Any given complex analytic function g has a Taylor expansion g(z) = anzn, so g ∈ H2(D) and +∞ a2 Pn=0 n < +∞ are naturally isomorphic. For T = ∂D, if L2(T) denoted the closed span of all Taylor expansions of functions +∞ Pn=0 in L2(T), then H2(T) is a closed subspace of L2(T). From the naturally isomorphic of H2(D) and H2(T) by the properties of analytic function, we denote H2(T) also as a Hardy space. By [14]P6 we get that any Cauchy sequence with the norm k • kH2 on 20 H2(D) is a uniformly Cauchy sequence on any closed disk in D, in particu- lar,we get that the point evaluations f → f (z) are continuous linear func- tional on H2(D), by the Riesz Representation Theorem[14]P13,for any g(s) ∈ H2(D), there is a unique fz(s) ∈ H2(D) such that g(z) =< g(s), fz(s) >, so define that fz is a reproducing kernel at z. Let H∞(D) denote the set of all bounded complex analytic function on D, for any given φ ∈ H∞(D), it is easy to get that kφk∞ = sup{φ(z); z < 1} is a norm on φ ∈ H∞(D). for any given g ∈ H2(D), the multiplication operator Mφ(g) = φg associated with φ on H2(D) is a bounded linear operator, and by the norm on H2(D) we get kMφ(g)k ≤ kφk∞kgkH2. If we denote H∞(T) as the closed span of all Taylor expansions of functions in L∞(T), then H2(T) is a closed subspace of L2(T) and H∞(D) and H∞(T) are naturally isomorphic by the properties of complex analytic functions [7]P55P97 and the Dirichlet Problem [7]P103. There are more properties about Hardy space in [4]P7, [16]P48, [21]P39, [28]P133 and [30]P106. Definition 5. For a connected open subset Ω of C,n ∈ N, let Bn(Ω) denotes the set of all bounded linear operator T on H that satisfies: (a) Ω ∈ σ(T ) = {ω ∈ C : T − ωnot invertible}; (b) ran(T − ω) = H for ω ∈ Ω; (c) W kerω∈Ω(T − ω) = H; (d) dim ker(T − ω) = n for ω ∈ Ω. 21 If T ∈ Bn(Ω), then say that T is a Cowen-Douglas operator. Theorem 8 ([2]). For a connected open subset Ω of C,T ∈ Bn(Ω),we get (1) If ΩT T 6= ∅,then T is Devaney chaotic. (2) If ΩT T 6= ∅,then T is distributionally chaotic. (3) If ΩT T 6= ∅,then T strong mixing. Definition 6 ([28]P141). Let P(z) be the set of all polynomials about z,where z ∈ T. Define a function h(z) ∈ H2(T) is an outer function if cl[h(z)P(z)] = H2(T). Lemma 6 ([28]P141). A function h(z) ∈ H∞(T) is invertible on the Banach algebra H∞(T), if and only if h(z) ∈ L∞(T) and h(z) is an outer function. Theorem 9 ([16]P81). Let P(z) be the set of all polynomials about z,where z ∈ D. then h(z) ∈ H2(D) is an outer function if and only if P(z)h(z) = {p(z)h(z); p ∈ P(z)} is dense in H2(D). Let φ is a non-constant complex analytic function on D, for any given z0 ∈ D, by [7]P29 we get that there exists δz0 > 0, exists kz0 ∈ N, when z − z0 < δz0, there is φ(z) − φ(z0) = (z − z0)kz0 hz0(z) where hz0(z) is complex analytic on a neighbourhood of z0 and hz0(z0) 6= 0. 22 Definition 7. Let φ is a non-constant complex analytic function on D, for any given z0 ∈ D, there exists δz0 > 0 such that φ(z) − φ(z0)z−z0<δz0 = pnz0 (z)hz0(z)z−z0<δz0 , hz0(z) is complex analytic on a neighbourhood of z0 and hz0(z0) 6= 0, (z) is a nz0-th polynomial and the nz0-th coefficient is equivalent 1. By pnz0 the Analytic Continuation Theorem [7]P28, we get that there is a unique complex analytic function hz0(z) on D such that φ(z)−φ(z0) = pnz0 then define hz0(z) is a rooter function of φ at the point z0. If for any given (z)hz0(z), z0 ∈ D, the rooter function hz0(z) has non-zero point but the roots of pnz0 are all in D and nz0 ∈ N is a constant on D that is equivalent m, then define (z) φ is a m-folder complex analytic function on D. Definition 8. Let φ(z) ∈ H∞(D), n ∈ N, Mφ is the multiplication by φ on H2(D). if the adjoint multiplier M∗φ ∈ Bn( ¯φ(D)), then define φ is a Cowen- Douglas function. By Definition 8 we get that any constant complex analytic function is not a Cowen-Douglas function Theorem 10. Let φ(z) ∈ H∞(D) be a m-folder complex analytic function, Mφ is the multiplication by φ on H2(D). If for any given z0 ∈ D, the rooter functions of φ at z0 is a outer function, then φ is a Cowen-Douglas function, that is,the adjoint multiplier M∗φ ∈ Bm( ¯φ(D)). Proof. By the definition of m-folder complex analytic function Definition 7 we get that φ is not a constant complex analytic function. For any given 23 z ∈ D, let fz ∈ H2(D) be the reproducing kernel at z. We confirm that M∗φ is valid the conditions of Definition 5 one by one. (1) For any given z ∈ D, fz is an eigenvector of M∗φ with associated eigenvalue λ = ¯φ(z).Because for any g ∈ H2(D) we get < g, M∗φ(fz) >H2=< φg, (fz) >H2= φ(z)f (z) =< g, ¯φ(z)fz >H2 By the Riesz Representation Theorem[14]P13 of bounded linear func- tional in the form of inner product on Hilbert space, we get M∗φ(fz) = ¯φ(z)fz = λfz, that is,fz is an eigenvector of M∗φ with associated eigenvalueλ = ¯φ(z). (2) For any given ¯λ ∈ φ(D), because of 0 6= φ ∈ H∞(D), we get that the multiplication operator Mφ − λ is injection by the properties of complex analysis, hence ker(Mφ − λ) = 0. Because of H2(D) = ker(Mφ − λ)⊥ = cl[ran(M∗φ − ¯λ)], we get that ran(M∗φ − ¯λ) is a second category space. By [14]P305 we get that M∗φ − ¯λ is a closed operator, also by [14]P93 or [31]P97 we get ran(M∗φ − ¯λ) = H2(D). (3) Suppose that span{fz; z ∈ φ(D)} = span{ 1 1−¯zs ; z ∈ φ(D)} is not dense in H2(D), By the definition of reproducing kernel fz and because of 0 6= φ ∈ H∞(D), we get that there exists 0 6= g ∈ H2(D),for any given z ∈ D, we have 0 =< g, ¯φ(z)fz >H2= φ(z)g(z) =< φ(z)g(z), fz >H2 . So we get g = 0 by the Analytic Continuation Theorem [7]P28,that is a 24 contradiction for g 6= 0. Therefor we get that span{fz; z ∈ φ(D)} is dense in H2(D), that is,W ker¯λ∈φ(D)(M∗φ − ¯λ) = H2(D). (4) By Definition 7 and the conditions of this theorem, for any given λ ∈ φ(D), there exists z0 ∈ D, exists m-th polynomial pm(z) and outer function h(z) such that φ(z) − λ = φ(z) − φ(z0) = pm(z)h(z), We give dim ker(M∗φ(z) − ¯λ) = m by the following (i)(ii)(iii) assertions. (i) Let the roots of pm(z) are z0, z1, · · · , zm−1, then there exists decompo- sition pm(z) = (z −z0)(z −z1) · · · (z −zm−1), and denote pm,z0z1···zm−1(z) is the decomposition of pm(z) by the permutation of z0, z1, · · · , zm−1, the following to get dim ker M∗pm,z0 z1···zm−1 = m. By the Taylor expansions of functions in H2(T), we get there is a naturally isomorphic Fs : H2(D) → H2(D − s), Fs(g(z)) → g(z + s), s ∈ C. It is easy to get that G = {Fs; s ∈ C} is a abelian group by the composite operation ◦, hence for 0 ≤ n ≤ m − 1,there is H2((D)) Fzn ↓ H2(D − zn) −−−−−−−−−−−−−−−−→ Mz−zn −−−−−−−−−−−−−−−−→ M ′ z H2((D)) ↓ Fzn H2(D − zn) Let T is the backward shift operator on the Hilbert space L2(N), that is, T (x1, x2, · · · ) = (x2, x3, · · · ). With the naturally isomorphic between H2(D− 25 zn) and H2(∂(D − zn)), M ′∗z H2(∂(D − zn)), that is,M ′∗z is a surjection and dim ker M∗z−zn = 1, where 0 ≤ n ≤ m − 1. is equivalent the backward shift operator T on is a surjection and dim ker M ′∗z = 1, hence M∗z−zn By the composition of Fzm−1 ◦ Fzm−2 ◦ · · · ◦ Fz0, M ′∗pm,z0z1···zm−1 is equivalent that is,M ′∗pm,z0 z1···zm−1 T m. hence M∗pm,z0z1···zm−1 is a surjection and dim ker M ′∗pm,z0 z1···zm−1 = m, is a surjection and dim ker M∗pm,z0 z1···zm−1 = m. (ii) Because H∞ is a abelian Banach algebra, Mpm is independent to the permutation of 1-th factors of pm(z), that is,M∗pm is independent to the 1-th factors multiplication of pm(z) = (z − z0)(z − z1) · · · (z − zm−1). Because G = {Fs; s ∈ C} is a abelian group by composition operation ◦, for 0 ≤ n ≤ m − 1, Fzm−1 ◦ Fzm−2 ◦ · · · ◦ Fz0 is independent to the permutation of composition. Hence M∗pm is a surjection and dim ker M∗pm = dim ker M∗pm,z0z1···zm−1 = m. (iii) By Definition 6 and Theorem 9, also by [14]P93 or [31]P97 and by [14]P305 we get that the multiplication operator Mh is surjection that associated with the outer function h. Hence we get ker M∗h(z) = (ranMh(z))⊥ = (H2(D))⊥ = 0. Because there exists decomposition M∗pm(z)h(z) = M∗h(z)M∗pm(z) on H2(D), we get dim ker(M∗φ − ¯λ) = dim ker M∗pm(z)h(z) = dim ker M∗pm(z) = m. 26 By (1)(2)(3)(4) we get the adjoint multiplier operator M∗φ ∈ Bm( ¯φ(D)). By Theorem 10 and Lemma 6, we get Corollary 4. Let φ ∈ H∞(D) is a m-folder complex analytic function, for any given z0 ∈ D, if the rooter function of φ at z0 is invertible in the Banach algebra H∞(D), then φ is a Cowen-Douglas function. Especially,for any given n ∈ D,if a and b are both non-zero complex, then a + bzn ∈ H∞(D) is a Cowen-Douglas function. The following gives some properties about the adjoint multiplier of Cowen- Douglas functions. Theorem 11. If φ ∈ H∞(D) is a Cowen-Douglas function, Mφ is the mul- tiplication by φ on H2(D), Then the following assertions are equivalent (1) M∗φ is Devaney chaotic; (2) M∗φ is distributionally chaotic; (3) M∗φ is strong mixing; (4) M∗φ is Li-Yorke chaotic; (5) M∗φ is hypercyclic; (6) φ(D)T T 6= ∅. 27 Proof. By Example 1 we get that M∗φ is hypercyclic if and only if φ is non- constant and φ(D)T T 6= ∅, hence (6) is equivalent to (5). First to get that (6) imply (1)(2)(3)(4). Because φ ∈ H2(D) is a Cowen-Douglas function, by Definition 8, M∗φ ∈ Bn( ¯φ(D)). By Theorem 8 we get that if φ(D)T T 6= ∅, then (1)(2)(3) is valid. On Banach spaces Devaney chaotic, distributionally chaotic and strong mixing imply Li-Yorke chaotic, respectively.Hence (4) is valid.Because ¯φ(D)T T 6= ∅ and φ(D)T T 6= ∅ are mutually equivalent, (6) imply (1)(2)(3)(4). Then to get that (1)(2)(3)(4) imply (6).By (1)(2)(3) imply (4),respectively, it is enough to get that (4) imply (6). If M∗φ is Li-Yorke chaotic, then we get that φ is non-constant and by φ k → ∞, hence kMφk = kM∗φk > 1, that [9]Theorem3.5 we get kM∗n sup n→+∞ is,sup z∈D that inf z∈D φ(z) > 1. Moreover,we also have inf z∈D φ ∈ H∞ and kM∗1 φ(z) ≥ 1 then 1 0 6= x ∈ H2(D) we get kM∗n φ xk ≥ contradiction to M∗φ is Li-Yorke chaotic. φ(z) < 1, Indeed,if we assume k = kM 1 φ 1 kM ∗−n φ k kxk ≥ 1 kM ∗ φ k ≤ 1. Hence for any φ kn kxk ≥ kxk. It is a 1 Therefor that M∗φ is Li-Yorke chaotic imply inf z∈D φ(z) < 1 < sup z∈D φ(z), By the properties of a simple connectedness argument of complex analytic functions we get φ(D)T T 6= ∅. Hence we get (1)(2)(3)(4) both imply (6). Corollary 5. If φ ∈ H∞(D) is a invertible Cowen-Douglas function in the Banach algebra H∞(D), and let Mφ be the multiplication by φ on H2(D). 28 Then M∗φ is Devaney chaotic or distributionally chaotic or strong mixing or Li-Yorke chaotic if and only if M∗−1 is. φ Proof. Because of T = (T −1)−1, it is enough to prove that M∗φ is Devaney chaotic or distributionally chaotic or strong mixing or Li-Yorke chaotic imply M∗−1 φ is. By Definition 8 we get M∗φ ∈ Bn( ¯φ(D)), with a simple computing we get φ ∈ Bn( 1 ¯φ (D)). M∗−1 If M∗φ is Devaney chaotic or distributionally chaotic or strong mixing or Li-Yorke chaotic, by Theoem 11 we get φ(D)T T 6= ∅, and by the properties of complex analytic functions we get 1 φ (D)T T 6= ∅. Because of M∗−1 φ ∈ B1( 1 ¯φ (D)) and by Theorem 11 we get M∗−1 φ is Devaney chaotic or distributionally chaotic or strong mixing or Li-Yorke chaotic. 5. The chaos of scalars perturbation of an operator We now study some properties about scalars perturbation of an operator inspired by [10] and [1] that research some properties about the compact perturbation of scalar operator. Definition 9. Let λ ∈ C,T ∈ L(H).Define (i) Let SLY (T ) denote the set such that λI + T is Li-Yorke chaotic for every λ ∈ SLY (T ). 29 (ii) Let SDC(T ) denote the set such that λI + T is distributionally chaotic for every λ ∈ SDC(T ). (iii) Let SDV (T ) denote the set such that λI + T is Devaney chaotic for every λ ∈ SDV (T ). (iv) Let SH (T ) denote the set such that λI + T is hypercyclic for every λ ∈ SH (T ). By Definition 9 we get SLY (λI + T ) = λ + SLY (T ), SDC(λI + T ) = λ + SDC(T ), SDV (λI + T ) = λ + SDV (T ) and SH(λI + T ) = λ + SH(T ). Lemma 7. Let T ∈ L(H) be a normal operator,then SLY (T ) = ∅. Proof. Because T is a normal operator,λI + T is a normal operator,too. by [10] we get SLY (T ) = ∅. Lemma 8. There is a quasinilpotent compact operator T ∈ L(H) such that SLY (T ) = T is closed and SLY (T ∗) = ∅,where T = ∂D, D = {z ∈ C, z < 1}. Proof. Let {en}n∈N be a orthonormal basis of L2(N) and let T be a weighted n }+∞n=1 such that Sω(e0) = backward shift operator with weight sequence {ωn = 1 0, Sω(en) = ωnen−1, where 0 < ωn < M < +∞,∀n > 0. By the Spectral Radius formula[14]P197 rσ(T ) = lim n→+∞ σ(T ) = {0}, hence σ(λI + T ) = λ. kT nk 1 n we get (1) If λ < 1, we can select ε > 0 such that λ + ε < q < 1. Then 30 ∀0 6= x ∈ H we get lim n→∞ k(λI + T )nxk ≤ lim n→∞ k(λI + T )nkkxk ≤ lim n→∞ (λ + ε)nkxk = 0. (2) If λ > 1,because of σ((λI + T )−1) = 1 λ,then we get lim n→∞ k(λI + T )nxk ≥ lim n→∞ 1 k(λI + T )−nk kxk ≥ lim n→∞ 1 k(λI + T )−1kn kxk ≥ kxk 6= 0. (3) By Theorem 3 we get that if λ = 1, then λI + T is mixing. Mixing imply Li-Yorke chaotic. (4) Because of σ(T ) = σ(T ∗), by (1)(2) we get that if λ 6= 1,then λ + T ∗ is not Li-Yorke chaotic. (5) If λ = 1,from the view of infinite matrix λI + T ∗ is lower triangu- lar matrix, then with a simple computing,for any 0 6= x ∈ L2(N), we get kλI + S∗ωxk > 0. Hence λI + T ∗ is not Li-Yorke chaotic. lim n→∞ By (1)(2)(3)(4)(5) we get that SLY (T ) = T is closed and SLY (T ∗) = ∅. Lemma 9. Let T be the backward shift operator on L2(N), T (x1, x2, · · · ) = (x2, x3, · · · ). Then SLY (T ) = SDC(T ) = SDV (T ) = SH(T ) = 2D \ {0}, SLY (2T ) = SDC(2T ) = SDV (2T ) = SH(2T ) = 3D, Hence SLY (T ) and SLY (2T ) are open sets. Proof. By [14]P209 we get σ(T ) = clD and σ(2T ) = cl2D, by Definition σ(T ) we get σ(λI + T ) = λ + clD. Because of the method to prove the conclusion is similarly for T and 2T , we only to prove the conclusion for T . 31 By the naturally isomorphic between H2(T) and H2(D). Let L2(N) = H2(T), by the definition of T we get (λI + T )∗ is the multiplication operator Mf by f (z) = ¯λ + z on the Hardy space H2(T). By the Dirichlet Problem [7]P103 we get that f (z) is associated with the complex analytic function φ(z) = ¯λ + z ∈ H∞(D) determined by the boundary condition φ(z)T = f (z). By Corollary 4 we get that φ is a Cowen-Douglas function. Therefor by the natural isomorphic between H2(T) and H2(D), λI + T is naturally equivalent to the operator M∗φ on H2(D). By Theorem 11 we get that M∗φ is hypercyclic or Devaney chaotic or distributionally chaotic or Li-Yorke chaotic if and only if φ(D)T T 6= ∅. Because of σ(λI +T ) = σ(¯λI +T ∗), we get σ(λI +T ) = σ(M∗φ) = σ(Mφ) ⊇ φ(D), hence SLY (T ) = SDC(T ) = SDV (T ) = SH(T ) = 2D \ {0} is an open set. Therefor we can get Corollary 6. Let T be the backward shift operator on L2(N), T (x1, x2, · · · ) = (x2, x3, · · · ). For λ 6= 0, a 6= 0, n ∈ N, if λ+aT n is a invertible bounded linear operator, then λ+aT n is strong mixing or Devaney chaotic or distributionally chaotic or Li-Yorke chaotic if and only if (λ + aT n)−1 is. Theorem 12. There is T ∈ L(H), SLY (T ) is neither open nor closed. Proof. Let T1, T2 ∈ L(H), because T1 or T2 is Li-Yorke chaotic if and only if T1 L T2 is, we get SLY (T1 L T2) = SLY (T1)S SLY (T2). 32 By Lemma 8,Lemma 9 and Definition 9 we get the conclusion. 6. Examples that T is chaotic but T −1 is not In the last we give some examples to confirm the theory giving by func- tional calculus on the begin that T is chaotic but T −1 is not. Example 2. Let {en}n∈N be a orthonormal basis of L2(N) and let Sω be a backward shift operator on L2(N) with weight sequence ω = {ωn}n≥1 such that Sω(e0) = 0, Sω(en) = ωnen−1,where 0 < ωn < M < +∞, ∀n > 0. (1) If λ = 1,then λI +Sω is Li-Yorke chaotic, but λI +S∗ω and (λI +S∗ω)−1 are not Li-Yorke chaotic. (2) Let (λI + Sω)n = Un(λI + Sω)n is the polar decomposition of (λI + Sω)n, {Un}∞n=1 is not a constant sequence. Proof. (1) By Theorem 3, we get that for λ = 1,λI +Sω is mixing and mixing imply Li-Yorke chaotic, hence λI + Sω is Li-Yorke chaotic. From the view of infinite matrix,λI + S∗ω and (λI + S∗ω)−1 are lower triangular matrix, with a simple computing,for any 0 6= x ∈ L2(N), we get lim kλI + S∗ωxk > 0,and n→∞ k(λI + S∗ω)−1xk > 0. Hence λI + S∗ω and (λI + S∗ω)−1 are not Li-Yorke lim n→∞ chaotic. (2) Let (λI + Sω)n = Un(λI + Sω)n is the polar decomposition of (λI + Sω)n. If {Un}∞n=1 is a constant sequence, then by Theorem 4 we get that (λI +S∗ω)−1 is Li-Yorke chaotic. A contradiction. Hence {Un}∞n=1 is not a constant se- quence. 33 Theorem 13 ([10]). For any ε > 0, there is a small compact operator Kε ∈ L(H) and kKεk < ε such that I + Kε is distributionally chaotic. In [10], I+Kε = (Ij+Kj) is distributionally chaotic. where H = +∞ Lj=1 +∞ Lj=1 Hj, nj = 2mj,Hj is the nj-dimension subspace of H. On Hj define: 0 2εj . . . Sj =   . . . . . . 2εj 0  nj×nj , Kj = −εj 2εj . . .   . . . . . . 2εj −εj  nj×nj . Ij + Kj = 1 − εj 2εj . . .   . . . . . . 2εj 1 − εj  nj×nj = (1 − εj)Ij + Sj. We can construct a invertible bounded linear operator I + Kε in the same way that is Li-Yorke chaotic, but (I + Kε)−1 is not. Example 3. There is a invertible bounded linear operator I + Kε on H = L2(N) such that I + Kε is Li-Yorke chaotic, but (I + Kε)−1,(I + Kε)∗−1 and (I + Kε)∗ are not Li-Yorke chaotic. Proof. Let {ei}∞i=1 is a orthonormal basis of H = L2(N) and Let H = +∞ Lj=1 Hj, j ∈ N, where Hj = span{ei}, 1 + j(j−1) 2 ≤ i ≤ j(j+1) 2 , Hj is j-dimension subspace of H. For any given positive sequence {εj}∞1 such that εj → 0 and (1 + εj)j → +∞, on Hj define: sup j→∞ 34 0 2εj . . . Sj =   . . . . . . 2εj 0  j×j , Kj =   −εj 2εj . . . . . . . . . 2εj −εj  j×j Ij + Kj = 1 − εj 2εj . . .   . . . . . . 2εj 1 − εj  j×j = (1 − εj)Ij + Sj First to prove that (I + Kε) is Li-Yorke chaotic. . . Let I+Kε = (I+Kj), fj = 1√j (11, · · · , 1j) and fj,n = 1√j (11, · · · , 1n, 0, · · · , 0) ∈ +∞ Lj=1 Hj. for any 1 ≤ n ≤ j we get k(I + Kε)n(fj)k = k(Ij + Kj)n(fj)k = k((1 − εj)Ij + Sj)n(fj)k = k ≥ k n n Pk=0 Pk=0 C k n(1 − εj)kSn−k j fjk C k n(1 − εj)k(2εj)n−k(11, · · · , 1n, 0, · · · , 0)k = (1 + εj)nkfj,nk. Hence we get 35 (a) lim j→∞ k(I + Kε)j(fj)k ≥ lim n→∞ kfjk(1 + εj)j = +∞. Because of rσ(I + Kε) < 1,we get (b) lim n→∞ k(I + Kε)n(fj)k = 0. By (a)(b) and by Definition 3 we get that λI + Kε satisfies the Li-Yorke Chaos Criterion, by Theorem 1 we get that λI + Kε is Li-Yorke chaotic. Then to prove that (I + Kε)−1 is not Li-Yorke chaotic. For convenience we define n?m by induction on m for any given n ∈ N. For any given j ∈ N,define: (1) j? = 1 + 2 + · · · + j, (2) If defined j?n , then define j?n+1 = 1?n + 2?n + · · · + j?n. Let A = +∞ Lj=1 Aj,where Aj = (Ij + Kj)−1.Because of Aj(Ij + Kj) = Aj(Ij + Kj) = Ij, we get A(I + Kε) = (I + Kε)A = +∞ Lj=1 Ij = I. By the Banach Inverse Mapping Theorem [14]P91 we get that A = (I + Kε)−1 is a bounded linear operator, that is,A = (I + Kε)−1.Hence we get Aj = 1 1 − εj (−2)j−1εj−1 j −2εj 1  j×j 1 −2εj . . .   . . . . . . 36 . A2 j = 1 (1 − εj)2   1 2 · (−2)εj j · (−2)j−1εj−1 j . . . . . . . . . 2 · (−2)εj 1  j×j A3 j = 1 (1 − εj)3   1 (1 + 2)(−2)εj j? · (−2)j−1εj−1 j . . . . . . . . . 2?(−2)εj 1  j×j For m ≥ 3, m ∈ N,If defined Am j = 1 (1 − εj)m   1 2?m−2(−2)εj j?m−2(−2)j−1εj−1 j . . . . . . . . . 2?m−2(−2)εj 1  j×j Then define . . . 37 A(m+1) j = AAm = 1 (1 − εj)(m+1)   1 2?m−1(−2)εj j?m−1(−2)j−1εj−1 j . . . . . . . . . 2?m−1(−2)εj 1  j×j . . For any given 0 6= x0 = (x1, x2, · · · , ) ∈ H, Let yj = (x(1+ j(j−1) y′j = (x 2 ); ), · · · , x j(j+1) , · · · , x( j(j+1) 2 (2+ j(j−1) ) 2 2 −1)); zj,m = x(1+ j(j−1) 2 ) + j Pk=2 k?(m−2)(−2)k−1εk j . Following a brilliant idea of Zermelo,we shall give the conclusion by in- duction. (1) If y1 6= 0,then we get lim n→∞ kAnx0k = lim n→∞ +∞ Pj=1 kAn j yjk ≥ lim n→∞ kAn 1 y1k = lim n→∞ x1 (1−ε1)n = +∞. (2) If y1 = 0,but y2 = (x2, x3) 6= 0. (i) If x3 6= 0, by (1) we get kAnx0k ≥ lim n→∞ lim n→∞ (ii) If x3 = 0 and ε2 > 1 x3 (1−ε1)n → +∞. 2?(n−2) ,because of y2 = (x2, x3) 6= 0,we get lim n→∞ kAnx0k = lim n→∞ +∞ Pj=1 kAn j yjk ≥ lim n→∞ kAn 2 y2k 38 = lim n→∞ 1 (1−ε1)np(x2 2 + (2?n−2(−2)ε2x3)2) ≥ lim n→∞ x2 (1−ε1)n = +∞. (3) Assume for k ≤ m−1, there is lim n→∞ Then for k = m and ym 6= 0 we get kAnx0k → +∞ for Ak and yk 6= 0. (i) If xm? 6= 0, by (1) we get lim n→∞ kAnx0k ≥ lim n→∞ xm? (1−ε1)n = +∞. (ii) If xm? = 0 and εm > 1 m?(n−2) ,because of ym 6= 0,we get lim n→∞ = lim n→∞ kAnx0k = lim n→∞ (1−ε1)n qz2 1 +∞ Pj=1 kAn j yjk ≥ lim n→∞ kAn mymk m,n + kAn m−1y′mk2. If y′m 6= 0,by the induction hypothesis we get kAn kAnx0k ≥ lim n→∞ lim n→∞ If y′m = 0,because of ym 6= 0 we get x(1+ m(m−1) m−1y′mk = +∞; 2 ) 6= 0. by (1) we get lim n→∞ kAnx0k ≥ lim n→∞ kzm,nk (1−ε1)n = lim n→∞ x ) m(m−1) (1+ 2 (1−ε1)n = +∞. Therefor for k = m and ym 6= 0, we get lim n→∞ induction we get that for any m ∈ N and ym 6= 0,there is lim n→∞ kAnx0k → +∞, by the kAnx0k → +∞. From (1)(2)(3) and H = +∞ Lj=0 Hj, we get that for any given 0 6= x0 = (x1, x2, · · · , ) ∈ H we can find m ∈ N such that ym 6= 0. Hence for any given 0 6= x0 = (x1, x2, · · · , ) ∈ H we get lim n→∞ is not Li-Yorke chaotic. From the view of infinite matrix, (I + Kε)∗ and kAnx0k = +∞. Therefor (I + Kε)−1 39 (I + Kε)∗−1 are lower triangular matrix, for any 0 6= x ∈ H,with a simple computing we get lim n→∞ (I + Kε)∗ and (I + Kε)∗−1 are not Li-Yorke chaotic. k(I + Kε)∗nxk > 0, k(I + Kε)∗−nxk > 0. Hence lim n→∞ Corollary 7. There is a invertible bounded linear operator I + Kε on H = L2(N) such that I + Kε is distributionally chaotic, but (I + Kε)−1,(I + Kε)∗−1 and (I + Kε)∗ are not distributionally chaotic. Proof. By the construction of Theorem 13, it is only to give the conclusion by induction on {ki}∞i=1 as Example 3. Theorem 14. There is T ∈ L(H), SLY (T ) = SDC(T ) = ω is an open arc of T = {λ = 1; λ ∈ C}, and for ∀λ ∈ ω,we get that (λ + T )∗,(λ + T )∗−1 and (λ + T )−1 are not Li-Yorke chaotic. Proof. As Example 3,give the same H = sequence {εj}∞1 such that εj → 0 and sup j→∞ Hj,Sj and Kj, give positive +∞ Lj=1 i + εjj → +∞,where i ∈ C. Let λI + Kε = +∞ Lj=1 (λI + Kj), so σ(λI + Kε) = {λ − εj; j ∈ N}. (i) If λ < 1,because of εj → 0, we get that ∃N > 0 when n > N, λ − εj < 1. With the introduction of this paper we get that Li-Yorke chaos is valid only on infinite Hilbert space. Loss no generally,for any j ∈ N,let λ − εj < 1, so rσ(λI+Kε) < 1. Hence for any 0 6= x ∈ H there is k(I + Kε)n(x)k = 0. lim n→∞ (ii) If λ > 1 or λ ∈ [ π 2 , 3π 2 ], because of εj > 0, for j ∈ N we get 40 λ − εj > 1, 1 λ−εj < 1, and σ(λI + Kε)−1 = { 1 λ−εj ; j ∈ N}. By Example 3, for any given x 6= 0 there is ym 6= 0, m ∈ N,hence we get k(λI + Kε)nx0k k(λIm + Km)nymk 1 kymk k(λIm+Km)−nk k(λIm+Km)−1kn kymk 1 lim n→∞ ≥ lim n→∞ ≥ lim n→∞ ≥ lim n→∞ ≥ kymk > 0. (iii) For ∀λ ∈ (− π 2 , π j > N, we get λ − εj < 1. Let H 2 ), because of εj → 0, there exists N > 0, when Hj , Hj, (λI + Kε)′ = (λI + Kε) L ′ = Lj>N j>N then for fj = 1√j (11, · · · , 1j) and fj,n = 1√j (11, · · · , 1n, 0, · · · , 0) ∈ H that when 1 ≤ n ≤ j,there is ′, we get k(λI + Kε)n(fj)k = k(λIj + Kj)n(fj)k = k((λ − εj)Ij + Sj)n(fj)k ≥ k n Pk=0 C k n(λ − εj)k(2εj)n−k(11, · · · , 1n, 0, · · · , 0)k = λ + εjnkfj,nk. By (i)(ii),if λ 6= 1 or λ ∈ [ π 2 , 3π 2 ], λI + Kε is not Li-Yorke chaotic. By (iii) and by the property of the triangle,if λ ∈ (− π 2 , π get λ + εj > i + εj and rσ((λI + Kε)′) < 1. Hence we get 2 ) and j > N, we k(λI + Kε)′n(fj)k = 0, and lim n→∞ lim j→∞ k(I + Kε)′j(fj)k ≥ lim j→∞ kfjkλ + εjj ≥ lim j→∞ i + εjj = +∞. 41 By Definition 3 we get that λI +Kε satisfies the Li-Yorke Chaos Criterion, by Theorem 1 we get that λI + Kε is Li-Yorke chaotic. Using the same proof of Example 3 we get that (λI + Kε)∗,(λI + Kε)∗−1 and (λI + Kε)−1 are not Li-Yorke chaotic. Using Corollary 7 and Theorem 13 we get that λI + Kε is distributionally chaotic,but (λI +Kε)∗,(λI +Kε)∗−1 and (λI +Kε)−1 are not Li-Yorke chaotic. Conjecture 1. For any given m ∈ N, there exists m-folder complex analytic function φ(z) ∈ H∞(D) such that φ is not a Cowen-Douglas function. Question 1. Gives the equivalent characterization of a m-folder complex analytic function; Gives the equivalent characterization of a rooter function; Gives the equivalent characterization of a Cowen-Douglas function. Gives the relations between them. Question 2. Let Mφ is the multiplication operator of the Cowen-Douglas function φ(z) ∈ H∞(D) on the Hardy space H2(D), then is M∗φ a Lebesgue operator? If not and if they have relations ,gives the relations between them. References [1] Bermdez, Bonilla, Martnez-Gimnez and Peiris. Li-Yorke and distributionally chaotic operators. J.Math.Anal.Appl.,(373)2011:1-83-93. [2] B.Hou, P.Cui and Y.Cao. Chaos for Cowen-Douglas operators. Pro.Amer.Math.Soc,138(2010),926-936. 42 [3] C.Kitai. Invariant closed sets for linear operators. Ph.D.thesis, University of Toronto, Toronto, 1982. [4] F.Bayart and E.Matheron. Dynamics of Linear Operators, Cambridge Uni- versity Press, 2009. [5] G.D.Birkhoff. Surface transformations and their dynamical applications. Acta Mathmatica, 1922:43-1-119. [6] G.Godefroy. Renorming of Banach spaces. In Handbook of the Geometry of Banach Spaces. volume 1,pp.781-835, North Holland, 2003. [7] Henri Cartan(Translated by Yu Jiarong). Theorie elementaire des fonctions analytuques d'une ou plusieurs variables complexes. Higer Education Press, Peking, 2008. [8] Hua Loo-kang. On the automorphisms of a sfield. Proc.Nat.Acad.Sci.U.S.A.,1949:35-386-389. [9] Hou B, Liao G, Cao Y. . Dynamics of shift operators. Houston Journal of Mathematics, 2012: 38(4)-1225-1239 [10] Hou Bingzhe,Tian Geng and Shi Luoyi. Some Dynamical Properties For Lin- ear Operators. Illinois Journal of Mathematics,Fall 2009. [11] Iwanik,A.. Independent sets of transitive points. Dynamical Systems and Er- godic Theory. vol.23,Banach Center publications,1989,pp277-282. [12] J.H.Shapiro. Notes on the dynamics of linear operators. Available at the au- thor'web page 2001. [13] John Milnor. Dynamics in one complex variable. Third Edition. Princeton University Press, 2006. 43 [14] John B.Conway. A Course in Functional Analysis. Second Edition,Springer- Verlag New York, 1990. [15] John B.Conway. A Course in Operator Theory. American Mathmatical Soci- ety, 2000. [16] John B.Garnett. Bounded Analytic Functions. Revised First Edition. Springer-Verlag, 2007. [17] J.R.Browm. Ergodic Theory and Topological Dynamics. Academic Press,New York,1976. [18] J.W.Robbin. Topological Conjungacy and Structural Stability for discrete Dynamical Systems. Bulletin of the American Mathematical Society, Vol- ume78, 1972. [19] K.-G.Grosse-Erdmann and A.Peris Manguillot. Linear Chaos. Springer, Lon- don, 2011. [20] K.-G.Grosse-Erdmann. Universal families and hypercyclic vectors. Bull. Amer. Math. Soc.,36(3):345-381,1999. [21] Kenneth Hoffman. Banach Spaces of Analytic Functions. Prentice-Hall, Inc., Englewood Cliffs, N.J., 1962. [22] Li T.Y. and Yorke J.A.. Period three implies chaos. Amer.Math.Monthly 82(1975), 985C992. [23] M.Shub. Global Stability of Dynamical Systems. Springer-Verlag New York, 1987. [24] N.H.Kuiper and J.W.Robbin. Topological Classification of Linear Endomor- phisms. Inventiones math.19,83-106, Springer-Verlag, 1973. 44 [25] Nilson C.Bernardes Jr., Antonio Bonilla, Vladimr Mller and A.peiris. Li- Yorke chaos in linear dynamics. Academy of Sciences Czech Republic, Preprint No.22-2012. [26] Paul R.Halmos. Measure Theory. Springer-Verlag New York, 1974. [27] Peter Walters. An Introduction to Ergodic Theory. Springer-Verlag New Yorke, 1982. [28] Ronald G.Douglas. Banach Algebra Techniques in Operator Theory. Second Edition. Springer-Verlag New York, 1998. [29] S.Rolewicz. On orbits of elements. Studia Math., 1969:32-17-22. [30] William Arveson. A Short Course on Spectral Theory. Springer Sci- ence+Businee Media, LLC, 2002. [31] Zhang Gongqing,Lin Yuanqu. Lecture notes of functional analysis(Volume 1). Peking University Press, Peking, 2006. 45
1812.11064	1	1812	2018-12-23T04:48:30	Extension of differentiable local mappings on linear topological spaces	[ "math.FA" ]	Usually, for extension of local maps, one uses multiplication by so called bump functions. However, majority of infinite-dimensional linear topological spaces do not have smooth bump functions. Therefore, in \cite{BR} we suggested a new approach for Banach spaces, based on the composition with locally identical maps. In the present work we discuss a possibility of generalization of this method for arbitrary spaces and applications of this theory.	math.FA	math	Extension of differentiable local mappings on linear topological spaces Genrich Belitskii and Victoria Rayskin Abstract. Usually, for extension of local maps, one uses multiplication by so called bump functions. However, majority of infinite-dimensional linear topological spaces do not have smooth bump functions. Therefore, in [5] we suggested a new approach for Banach spaces, based on the composition with locally identical maps. In the present work we discuss a possibility of gener- alization of this method for arbitrary spaces and applications of this theory. Bump functions and Blid maps and Topological spaces and Map extension and Linearization 1. Introduction Let X and Y be linear topological spaces. In this work, we will discuss the differentiable local maps f : X → Y and the possibility of differentiable extension of the maps. The map's extension is usually not unique and can be studied in the context of the equivalence class of f , i.e. a germ [f ]. Recall that a germ [f ] at x ∈ X is the equivalence class of local maps, such that any pair of the class members coincides on some neighborhood of x. Each element of the class is called a representative of a germ. Occasionally, we denote germ [f ] as f . In the future, without loss of generality, we will assume that x = 0. We are interested in the question of existence of a global representative of the germ. Recall that f : U ⊂ X → Y is differentiable at a point x ∈ U if f (x + h) = f (x) + A · h + r(h) with r(h) = o(h) and with the continuous operator A : X → Y . Each meaning of "smallness" o(h) defines a corresponding specific notion of r(h) and consequently a specific notion of differentiability. The value of the derivative is A (usually denoted by f ′(x)) and it is independent of the definition of r(h). Various definitions of differentiability are discussed in the works of F. and R. Nevanlinna ([8]), H. R. Fischer ([9]), H. H. Keller ([7]), E. inz ([2]), E. Binz, H. H. Keller ([3] ) and E. Binz, W. Meier - Solfrian ([4]). The main ideas of differentiation in abstract spaces were developed and clarified in the sequence of works by Averbuh, V. I. and Smoljanov, O. G ([1]), and later by M. Schechter in [12]. We remind all these definitions and connections between them in Section 2. Thus, the question investigated in this work is the following. Question 1.1. Let f be a germ at 0, differentiable in some a priori specified meaning. Does there exist a global differentiable (in the same sense) representative of the germ? The classical method of a smooth extension of a locally defined map f is mul- tiplication by the smooth bump function. Recall that a smooth (differentiable) 1 2 GENRICH BELITSKII AND VICTORIA RAYSKIN bump function is a real-valued function ranging between 0 and 1 with bounded support, which is equal to 1 in a neighborhood 0. Because it is known that very few spaces have smooth bump functions, we introduce1 here the new method based on the idea of considering the composition of f with a locally identical map. Be- cause we use composition in this construction, we will discuss only those notions of differentiability that satisfy the Chain Rule of differentiation. In Section 3 we show how locally defined maps with the help of our new method can be extended to the entire space. We present examples of the extension con- structed for the specific Banach spaces (Section 4) and for the specific non-Banach linear topological spaces (Section 5). In Section 6 we discuss applications of this theory to the problem of conjugation with linear map. We show that for the construction of differentiable conjugation, the assumption of the existence of smooth bump function is not necessary, and con- sequently the corresponding conjecture stated in the paper of W. Zhang, K. Lu and W. Zhang "Differentiability of the Conjugacy in the Hartman-Grobman Theorem" ([14]) is incorrect. 2. Background Definitions Let us recall three definitions of differentiation on linear topological spaces, which we use in this work. The reader can also find these definitions in [12] 1. Bounded differentiability Definition 2.1. The map f : X → Y is bounded-differentiable at x ∈ X, if for every bounded subset S ⊂ X and every h ∈ S and t ∈ R r(th)/t → 0 uniformly in h as t → 0. 2. Compact (Hadamard) differentiability Definition 2.2. The map f : X → Y is compact (Hadamard) differ- entiable at x ∈ X, if f (x + tnhn) − f (x) = tnAh + o(tn) as tn → 0, and hn → h. 3. If both X and Y are Banach spaces with the norms .1 and .2 respec- tively, then Fr´echet differentiation is well defined. Definition 2.3. The map f is Fr´echet differentiable at 0 if lim h→0 r(h)2/h1 = 0. Thus, as discussed in [1], bounded differentiability implies the compact one. The latter, in turn, is equivalent to Hadamard differentiability (see [12]). If both X and Y are Banach then bounded and Fr´echet differentiability coincide. For these types of differentiability many important rules, including the Chain Rule, hold. The compact (Hadamard) differentiability is the weakest for which the Chain Rule is satisfied (for instance, for the Gateaux derivative the Chain Rule does not hold). 1Earlier this idea was introduced in [5] for Banach spaces and for differentiability in Fr´echet sense. EXTENSION OF DIFFERENTIABLE LOCAL MAPPINGS ON LINEAR TOPOLOGICAL SPACES3 3. The Main Proposition As explained above, we need to define differentiability, which satisfies the Chain Rule. Thus, we assume that the Cain Rule holds for the next definition and for the later discussion. In our work [5] we defined the notion of the blid map for Banach spaces, which stands for Bounded Local Identity map. In this paper we generalize this idea for linear topological spaces. Definition 3.1. A space X satisfies blid-differentiable property if for every neighborhood U ⊂ X of 0 there is a differentiable map H defined on X, locally coinciding with the identity map, such that H(X) ⊂ U . Let us recall that a neighborhoods base of zero is a system B = {Vα} of neighborhoods of 0, such that for any neighborhood U ⊂ X of 0 there exists some Vβ ∈ B, Vβ ⊂ U . Therefore, if there is a neighborhoods base B such that for every Vα from B there exists local identity Hα, Hα(X) ⊂ Vα, then X satisfies the blid-property. Proposition 3.2 (The Main Proposition). If X satisfies blid-differentiable property, then every differentiable germ [f ] : X → Y has a global differentiable representative. Proof. Let f be a local representative of the germ defined on a neighborhood U ⊂ X of zero. Let H : X → X be a differentiable local identity map such that H(X) ⊂ U . Then the map is a global representative of the germ as we need. (cid:3) F (x) = f (H(x)), x ∈ X 4. Banach Spaces In this section we will consider a Banach space X and a general linear topolog- ical space Y . In [5] we introduced the following definition: Definition 4.1. A differentiable blid map for a space X is a global Bounded Local Identity differentiable map H : X → X. Lemma 4.2. If there exists differentiable blid map, then X satisfies differen- tiable blid-property. Proof. It is convenient to chose the balls Bc = {x ∈ X : x < c} to be the base of neighborhoods in the Banach space X. Let H be a blid map, such that H(x) < N . Then Hc(x) = c N H( N c x) is the blid map as well and its image is inside of Bc. (cid:3) The next statement immediately follows from the Lemma 4.2 and the Main Proposition 3.2. Corollary 4.3. If the space X admits differentiable blid map, then every differentiable germ at 0 ∈ X into Y has a global differentiable representative. Note, if Y is a Banach space as well, then this result is proved in [5] for differentiability in Fr´echet sense. Corollary 4.4. If the space X has a differentiable bump function h : X → R, then it satisfies differentiable blid-property, and every differentiable germ at 0 ∈ X has a global differentiable representative. 4 GENRICH BELITSKII AND VICTORIA RAYSKIN Proof. The map H(x) = h(x)x is a differentiable blid map for X. (cid:3) 4.1. The Spaces of Bounded Continuous Functions. In this section we will consider the space X = C(T ) of bounded continuous functions on a topological space T with the norm . = supT .. In the space X we will construct a blid map in the following way. H(x)(t) = h(x(t))x(t) where h is a smooth bump function on the real line. This blid map is Fr´echet- (consequently bounded-, consequently compact-) differentiable. Corollary 4.5. Any bounded- (consequently compact-) differentiable germ at 0 ∈ C(T ) has a global (in the corresponding sense) differentiable representative. 4.2. The Spaces of Smooth Functions. In this section we will consider X = Cq[0, 1], 0 < q < ∞, equipped with the norm xq = max0≤j≤q supt∈[0,1] x(j)(t). Recall that on the real line all types of differentiability coincide, and consequently can be viewed as Fr´echet differentiability. We will see that X has a differentiable blid map, and consequently satisfies differentiable blid property. Lemma 4.6. The space Cq[0, 1] admits Fr´echet-differentiable blid map. Proof. The map H(x)(t) = tj j! h(x(j)(0))x(j)(0) +Z t 0 dt1Z t1 0 dt2...Z tq−1 0 h(cid:16)x(q)(s)(cid:17) x(q)(s) ds, q−1 Xj=0 where h is a C∞ bump function on R, represents a bounded differentiable blid map. (cid:3) Corollary 4.7. Any bounded- (compact-) differentiable germ at 0 ∈ Cq[0, 1] has a global (in the corresponding sense) differentiable representative. 5. Metric Spaces Let X be a metric space with a metric d(x, y). Here we consider germs of maps from X into an arbitrary linear topological space Y . Although instead of Fr´echet differentiation (which is not defined for general metric spaces) we use bounded and compact (Hadamar) differentiation. The neighborhoods base B can be chosen as a collection {Bc}c = {x ∈ X : d(x, 0) < c}c. Then the space X satisfies differentiable-blid property if for every c there exists a differentiable, local identity map Hc : X → X such that d(Hc(x), 0) < c for all x, i.e., Hc(X) ⊂ Bc. In particular, if topology on X is defined by countable collection of norms xk, then the metric can be written as d(x, y) := 1 2k · ∞ Xk=0 x − yk x − yk + 1 . Lemma 5.1. Suppose for every k = 0, 1, ... there exists a global differentiable local identity map Hk such that Then X satisfies the differentiable blid property. Hk(x)k < ∞. sup x EXTENSION OF DIFFERENTIABLE LOCAL MAPPINGS ON LINEAR TOPOLOGICAL SPACES5 Proof. For a given c > 0 choose any (5.1) k > 1 − ln c/ ln 2 and let Hk be such that Set Hk(x)k < N, x ∈ X. Hc(x) = c 4N Hk(cid:18) 4N c x(cid:19) . Then inequality 5.1 and the fact that xj is monotonically increasing with j imply that d(Hc(x), 0) < c, i.e. Hc(X) ∈ Bc. (cid:3) 5.1. The Space of Smooth Functions on the Real Line. The space X = Cq(R) (0 ≤ q < ∞) of all smooth functions on R is endowed with the collection of norms xk = max t∈[−k,k] max l≤q x(l)(t). Lemma 5.2. The space X possesses the bounded- (consequently compact-) dif- ferentiable blid property. Proof. Let h(u) be a C∞-bump function on R, which equals to 1 in a neigh- borhood of 0 and such that a = supu∈R h(u)u < ∞. Then H(x)(t) =( h(x(t))x(t), q = 0 j=0 tj Pq−1 j! h(x(j)(0))x(j)(0) +R t 0 dt1R t1 0 dt2...R tq−1 0 is differentiable local identity map, and H(x)k < aek, k = 0, 1, ..., x ∈ X. ha(cid:0)x(q)(s)(cid:1) x(q)(s) ds, q ≥ 1 (cid:3) Corollary 5.3. Every bounded- (consequently compact-) differentiable germ at 0 ∈ C(R) has a global differentiable (in the corresponding sense) representative. 5.2. The Space of Infinitely Differentiable Functions on a Closed In- terval. The space X = C∞[0, 1] is endowed with the collection of norms xk = max j≤k max t∈[0,1] x(j)(t). Lemma 5.4. The space X possesses the bounded- (consequently compact-) dif- ferentiable property. Proof. Let h(u) be the same bump function on R as above. Then is differentiable local identity map, and H0(x)(t) = h(x(t))x(t) H0(x)0 < a. Further, let k > 0. Then Hk(x)(t) = k−1 Xj=0 tj j! h(x(j)(0))x(j)(0) +Z t 0 dt1Z t1 0 dt2...Z tk−1 0 h(cid:16)x(k)(s)(cid:17) x(k)(s) ds. is differentiable local identity map, and Hk(x)k < aek, k = 0, 1, ..., x ∈ X. (cid:3) 6 GENRICH BELITSKII AND VICTORIA RAYSKIN Corollary 5.5. Every bounded- (consequently compact-) differentiable germ at 0 ∈ C∞[0, 1] has a global representative. 5.3. The Space of Infinitely Differentiable Functions on the Real Line. The space X = C∞(R) is endowed with the collection of norms xk = max j≤k max t∈[−k,k] x(j)(t), k = 0, 1, 2, ... Lemma 5.6. The space X possesses the bounded- (consequently compact-) dif- ferentiable property. Proof. Let h(u) be the same bump function on R as above. Then is differentiable local identity map, and H0(x)(t) = h(x(t))x(t) H0(x)0 < a. Further, let k > 0. Then Hk(x)(t) = tp j! h(x(p)(0))x(p)(0) +Z t 0 dt1Z t1 0 dt2...Z tk−1 0 h(cid:16)x(k)(s)(cid:17) x(k)(s) ds. k−1 Xp=0 is differentiable local identity map, and Hk(x)k < aek, k = 0, 1, ..., x ∈ X. (cid:3) Corollary 5.7. Every bounded- (consequently compact-) differentiable germ at 0 ∈ C∞(R) has a global representative. In conclusion, we would like to pose the following Question 5.8. Which linear topological spaces have differentiable blid prop- erty? This question was not answered even for Banach spaces. 6. Applications Local linearization and normal forms are convenient simplification of complex dynamics. In this section we discuss differentiable linearization on Banach spaces. For a diffeomorphism F with a fixed point 0, we would like to find a smooth transformation Φ defined in a neighborhood of 0 such that Φ ◦ F ◦ Φ−1 has a simplified (polynomial) form called the normal form. If Φ ◦ F ◦ Φ−1 = DF = Λ is linear, the conjugation is called linearization. There are two major questions in this area of research: how to increase smoothness of the conjugation Φ, and whether it is sufficient to assume low smoothness of the diffeomorphism F . Hartman and Grobman independently showed that if Λ is hyperbolic, then for a diffeomorphism F there exists a local homeomorphism Φ such that Φ◦ F ◦ Φ−1 = Λ. Different proofs were given by Pugh [10]. A higher regularity of Φ has been an active area of research. The first attempt to answer the question of differentiability of Φ at the fixed point 0 under hyperbolicity assumption was made in [48], but an error was found and discussed in [11]. Later, in [6], Guysinsky, Hasselblatt and Rayskin presented correct proof. However, it was restricted to F ∈ C∞ (or more precisely, it was restricted to F ∈ Ck, where k is defined by complicated expression). It was con- jectured in the paper that the result is correct for F ∈ C2, as it was announced in [13]. EXTENSION OF DIFFERENTIABLE LOCAL MAPPINGS ON LINEAR TOPOLOGICAL SPACES7 Zhang, Lu and Zhang ([14], Theorem 7.1) showed that for a Banach space diffeomorphism F with a hyperbolic fixed point and α-Holder DF , the local con- jugating homeomorphism Φ is differentiable at the fixed point. Moreover, Φ(x) = x + O(x1+β ) and Φ−1(x) = x + O(x1+β ) as x → 0, for certain β ∈ (0, α]. There are two additional assumptions in this theorem. The first one is the spectral band width inequality. The authors explain that this inequality is sharp if the spectrum has at most one connected component inside of the unit circle in X, and at most one connected component outside of the unit circle in X. For the precise formulation of the spectral band width condition we refer the reader to the paper [14]. It is important (and it is pointed out in [14]) that this is not a non-resonance condition. The latter is required for generic linearization of higher smoothness. The second assumption is the assumption that the Banach space must possess smooth bump functions. It is conjectured in the paper that the second assumption is a necessary condition. In this section we explain that this conjecture is not correct (see Theorem 6.1). The bump function condition can be replaced with the less restrictive blid map condition. Blid maps allow to reformulate Theorem 7.1 in the following way: Theorem 6.1. Let X be a Banach space possessing a differentiable blid map with bounded derivative. Suppose F : X → X is a diffeomorphism with a hyperbolic fixed point, DF is α-Holder, and the spectral band width condition is satisfied. Then, there exists local linearizing homeomorphism Φ which is differentiable at the fixed point. Moreover, Φ(x) = x + O(x1+β ) and Φ−1(x) = x + O(x1+β ) as x → 0, for certain β ∈ (0, α]. In particular, we have the following Corollary 6.2. Let X = Cq[0, 1]. Suppose F : X → X is a diffeomorphism with a hyperbolic fixed point, DF is α-Holder, and the spectral band width condition is satisfied. Then, the local conjugating homeomorphism Φ is differentiable at the fixed point. Moreover, Φ(x) = x + O(x1+β ) and Φ−1(x) = x + O(x1+β ) as x → 0, for certain β ∈ (0, α]. Below we justify Theorem 6.1 Proof. Zhang, Lu and Zhang showed that for the conclusion of their Theorem 7.1 it is enough to satisfy the inequalities 1 and 2 (see 6.2 below), which are called condition (7.6) in their paper. In order to apply the blid maps instead of bump functions to the inequalities 6.2, it is sufficient to construct a bounded blid map, which has only first-order bounded derivative. I.e., let blid map H(x) : X → X be as follows: (6.1) 1. H(x) = x for x < 1 2. H ∈ C1 and H (j)(x) ≤ cj, j = 0, 1. The condition (7.6) of [14] is: (6.2) 1. 2. supx∈X DF (x) − Λ ≤ δη supx∈V \O {DF (x) − Λ/xα} = M < ∞ 8 GENRICH BELITSKII AND VICTORIA RAYSKIN Let DF − Λ = f . Define f (x) := f (δH(x/δ)) We will show that if f satisfies (7.6), then so does f . sup x∈X D f (x) ≤ sup x∈X Df (x) · sup x∈X DH(x) ≤ δη · c1. Thus, the first inequality of (7.6) holds for f . For the second inequality we have the following estimate: D f (x) xα ≤ Df (δH(x/δ)) δH(x/δ)α ·(cid:18) δH(x/δ) x (cid:19)α . The second multiple is bounded, because for small x (say, x/δ < ǫ for some ǫ > 0) we have while for x/δ ≥ ǫ δH(x/δ) x < c1 + o(1), δH(x/δ) x < c0/ǫ. I.e., δH(x/δ) x is less than some constant m. Then, D f (x) xα ≤ sup x∈V \O sup 0<x<δc0 {Df (x)/xα} · sup x∈X DH(x) · mα = sup 0<x<δc0 {Df (x)/xα} c1 · mα. This quantity is bounded by M c1mα if δ is sufficiently small. (cid:3) Other applications in the area of local analysis on Banach spaces possessing blid maps can be found in [5]. A generalization of Theorem 6.1 might be possible for the case of linear topo- logical spaces (e.g., space of smooth functions), which posses differentiable blid property. References [1] Averbuh, V. I.; Smoljanov, O. G, The theory of differentiation in linear topological spaces. Russian Mathematical Surveys. 1967. V. 22, #6, 201 -- 260. [2] E. inz, Ein Differenzierbarkeitsbegriff in limitierten Vektorraumen, Comment. Math. Helv. 41 (1966), 137 -- 156. [3] E. Binz, H. H. Keller, Funktionenraume in der Kategorie der Limesraume, Ann. Acad. Sci. Fennicae, ser. A, 383 (1966), 1 -- 21. [4] E. Binz, W. Meier - Solfrian, Zur Differentialrechnung in limitierten Vektorraumen, Comment. Math. Helv. 42, #4 (1967), 285 -- 296. [5] G. Belitskii, V. Rayskin, A New Method of Extension of Local Maps of Banach Spaces. Ap- plications and Examples [6] M. Guysinsky, B. Hasselblatt, and V. Rayskin, Differentiability of the Hartman-Grobman linearization, Discrete Contin. Dyn. Syst. 9 (2003), no. 4, 979984. [7] H. H. Keller, Raume stetiger multilinearen Abbildungen als Limesraume, Math. Ann. 159 (1965), 259 -- 270. [8] F. und R. Nevanlinna, Absolute Analysis, Springer Verl., 1959. [9] H. R. Fischer, Limesraume, Math. Ann. 137 (1959), 269 -- 303. [10] C. C. Pugh, On a theorem of P. Hartman, Amer. J. Math. 91 (1969), 363367. [11] V. Rayskin α-Holder linearization J. Differential Equations 147 (1998), no. 2, 271284. [12] M. Schechter, Differentiation in Abstract Spaces, Journal of Differentiable Equations, 55 (1984) 330 -- 345. [13] S. van Strien, Smooth linearization of hyperbolic fixed points without resonance conditions, J. Differential Equations 85 (1990), no. 1, 6690. EXTENSION OF DIFFERENTIABLE LOCAL MAPPINGS ON LINEAR TOPOLOGICAL SPACES9 [14] W. Zhang, K. Lu and W. Zhang Dfifferentiability of the Conjugacy in the Harman-Grobman Theorem, Transactions of the American Mathematical Society, 369, Number 7 (2017), 4995-5030. E-mail address: [email protected] E-mail address: [email protected]
1105.0467	1	1105	2011-05-03T02:01:15	A note on the boundedness of Riesz transform for some subelliptic operators	[ "math.FA", "math.AP", "math.DG" ]	Let $\M$ be a smooth connected non-compact manifold endowed with a smooth measure $\mu$ and a smooth locally subelliptic diffusion operator $L$ satisfying $L1=0$, and which is symmetric with respect to $\mu$. We show that if $L$ satisfies, with a non negative curvature parameter $\rho_1$, the generalized curvature inequality in \eqref{CD} below, then the Riesz transform is bounded in $L^p (\bM)$ for every $p>1$, that is \[\\| \sqrt{\Gamma((-L)^{-1/2}f)}\\|_p \le C_p \\| f \\|_p, \quad f \in C^\infty_0(\bM), \] where $\Gamma$ is the \textit{carr\'e du champ} associated to $L$. Our results apply in particular to all Sasakian manifolds whose horizontal Tanaka-Webster Ricci curvature is nonnegative, all Carnot groups with step two, and wide subclasses of principal bundles over Riemannian manifolds whose Ricci curvature is nonnegative.	math.FA	math	A NOTE ON THE BOUNDEDNESS OF RIESZ TRANSFORM FOR SOME SUBELLIPTIC OPERATORS FABRICE BAUDOIN AND NICOLA GAROFALO Abstract. Let M be a smooth connected non-compact manifold endowed with a smooth mea- sure µ and a smooth locally subelliptic diffusion operator L satisfying L1 = 0, and which is symmetric with respect to µ. We show that if L satisfies, with a non negative curvature param- eter ρ1, the generalized curvature inequality in (2.9) below, then the Riesz transform is bounded in Lp(M) for every p > 1, that is (cid:13)(cid:13)(cid:13)(cid:13) qΓ((−L)−1/2f )(cid:13)(cid:13)(cid:13)(cid:13)p ≤ Cpkf kp, f ∈ C∞ 0 (M), where Γ is the carr´e du champ associated to L. Our results apply in particular to all Sasakian manifolds whose horizontal Tanaka-Webster Ricci curvature is nonnegative, all Carnot groups with step two, and wide subclasses of principal bundles over Riemannian manifolds whose Ricci curvature is nonnegative. Contents 1. Introduction 2. Background 2.1. Assumptions 2.2. Some known results 3. Riesz transform 3.1. The case 1 < p ≤ 2 3.2. The case p > 2 4. Pointwise gradient estimates of the heat kernel References 1. Introduction 1 5 5 7 8 9 9 14 15 A central result in the analysis of Rn is the Lp continuity of singular integrals in the range 1 < p < ∞. One basic consequence of this result is the Lp boundedness of the Riesz transforms Rj = ∂ (−∆)−1/2, j = 1, ..., n, with their vector-valued counterpart ∂xj R = (R1, ...,Rn) = ∇(−∆)−1/2, see [S]. In [Str] Strichartz asked the question whether such Lp continuity of the Riesz transform could be extended to non-compact Riemannian manifolds under suitable assumptions on the latter. In this context the analogue of the vector-valued Riesz transform is the operator where we have denoted by ∆ the Laplacian on M in its realization as a positive self-adjoint operator on L2(M). Strichartz's question is important for the purpose of developing analysis on R = ∇∆−1/2, First author supported in part by NSF Grant DMS-0907326. Second author supported in part by NSF Grant DMS-1001317. 1 2 FABRICE BAUDOIN AND NICOLA GAROFALO manifolds. To explain this point let us indicate by (·,·) the inner product in L2(M), and with ∆1/2 the positive self-adjoint square root of ∆. Then one has the equality (∆f, f ) = (∆1/2f, ∆1/2f ). This immediately gives k∇fk2 = k∆1/2fk2, which in turn allows to identify the first-order Sobolev subspaces of L2(M) obtained by com- 0 (M) with respect to the seminorms k∇fk2 and k∆1/2fk2. Let us also notice in pletion of C ∞ passing that the latter equality can be reformulated in terms of R as follows kRfk2 = kfk2. However, when 1 < p < ∞ and p 6= 2, a similar identification of the two Sobolev spaces of order 0 (M) with respect to the seminorms k∇fkp and k∆1/2fkp is one obtained by completion of C ∞ no longer such a simple matter. It is a well-known fact that an estimate such as (1.1) Apk∆1/2fkp ≤ k∇fkp ≤ Bpk∆1/2fkp, f ∈ C ∞ 0 (M), 1 ′ p = 1. would suffice for such identification. It is also known that the validity of the right-hand inequality in (1.1) for a certain 1 < p < ∞ implies that of the left-hand inequality in Lp′ (M), where p + 1 Now the right-hand inequality in (1.1) is equivalent to the Lp continuity of the Riesz operator R. It is then clear that (1.1) is true for all 1 < p < ∞ if (1.2) Rfp ≤ Cpfp, f ∈ C ∞ 0 (M), for 1 < p < ∞, and this clarifies the relevance of the question raised by Strichartz. An interesting result due to Bakry [B] states that if the Ricci curvature of M is bounded from below by a non negative constant then (1.2), and therefore (1.1) hold for every 1 < p < ∞. The purpose of the present note is to extend this result to a sub-Riemannian framework by using the generalized curvature-dimension inequality recently introduced by the authors in [BG1]. This extension has been recently become possible thanks to a combination of the theory devel- oped in the two papers [BG1], [BBG], with the remarkable results in [CD], [ACDH]. The latter two works have established that (1.2) does hold in the range 1 < p < ∞ for complete, non- compact Riemannian manifolds satisfying suitable general assumptions which will be discussed below. In [CD] the authors have proved that (1.2) is true when 1 < p ≤ 2. In the paper [ACDH] the authors have established (1.2) in the remaining range 2 ≤ p < ∞. The essential new contri- bution of the present note is to verify that such general assumptions are verified (in a non-trivial manner) for a general class of locally subelliptic operators satisfying on a given smooth manifold M the generalized curvature-dimension inequality CD(ρ1, ρ2, κ, d) in (2.9) below, with curvature parameter ρ1 ≥ 0 (this in the Riemannian case corresponds to Ric≥ 0). Once this is done, the Lp continuity of an appropriately defined Riesz operator will follow by the general real variable methods developed in [CD], [ACDH]. To state the main result in this paper we assume that M be a C ∞ connected, non-compact manifold endowed with a smooth measure µ. Throughout the paper, the notation Lp(M), 1 ≤ p ≤ ∞, indicates the space of p-summable functions on M with respect to the measure µ. We assume that on M a second-order diffusion operator L with real coefficients is given. We also suppose that L be locally subelliptic, non-positive, and that it satisfy the assumptions listed in Section 2. There is a natural notion of (square of the length of the)"gradient" associated with RIESZ TRANSFORMS 3 L, namely Γ(f ) = 1 2{L(f 2) − 2f Lf}, and a canonical distance d, see (2.4) below. We assume throughout that the metric space (M, d) be complete. We also suppose that M be endowed with another bilinear differential form ΓZ , see (2.6) below, and that Γ and ΓZ satisfy all the hypothesis in Section 2 below. From our perspective, the most significant assumption is the so-called generalized curvature-dimension inequality CD(ρ1, ρ2, κ, d) in (2.9) below, which we now recall for the reader's convenience: There exist constants ρ1 ≥ 0, ρ2 > 0, κ ≥ 0, and d ≥ 2 such that the inequality (1.3) Γ2(f ) + νΓZ κ (Lf )2 +(cid:16)ρ1 − ν(cid:17) Γ(f ) + ρ2ΓZ(f ) 2 (f ) ≥ 1 d hold for every f ∈ C ∞(M) and every ν > 0, where Γ2 and ΓZ below. 2 are defined by (2.7) and (2.8) The assumption (1.3) constitutes a sub-Riemannian generalization of the classical curvature- diemension inequality CD(ρ, n) Γ2(f ) ≥ 1 n (∆f )2 + ρΓ(f ), which, as a consequence of the well-known Bochner's identity, is known to hold on any n- dimensional Riemannian manifold satisfying Ric≥ ρ. The parameter ρ1 in CD(ρ1, ρ2, κ, d) has the meaning of a lower bound on a sub-Riemannian Ricci tensor, see [BG1] for extensive details. Throughout the present paper the assumption ρ1 ≥ 0 will be in force. The semigroup Pt = etL is a strongly continuous semigroup of contraction operators on Lp(M) for 1 ≤ p ≤ ∞. We denote by p(x, y, t) = p(y, x, t) the positive heat kernel on M associated with the semigroup Pt. Given f ∈ C ∞ 0 (M), the function u(x, t) = Ptf (x) =ZM f (y)p(x, y, t)dµ(y), is a solution of the equation Lu − ut = 0 in M × (0,∞), corresponding to the initial datum u(x, 0) = f (x), x ∈ M. We now recall that, in the general framework described above, in the paper [BG1] we proved a generalized Li-Yau type inequality for solutions of the heat equation on M of the form u = Ptf . From such inequality, we were able to derive several basic facts, among which the following off-diagonal Gaussian upper bound: For any 0 < ε < 1 there exists a constant C(ρ2, κ, d, ε) > 0, which tends to ∞ as ε → 0+, such that for every x, y ∈ M and t > 0 one has (1.4) C(d, κ, ρ2, ε) d(x, y)2 p(x, y, t) ≤ V (x,√t) 1 2 V (y,√t) 1 2 exp(cid:18)− (4 + ε)t(cid:19) . Hereafter in this paper we adopt the notation V (x, r) = µ(B(x, r)), where for x ∈ M and r > 0 we have let B(x, r) = {y ∈ M d(y, x) < r}. In the paper [BBG] we further developed the program initiated in [BG1] and were able to obtain the following basic result: There exists a constant Cd > 0 depending only on ρ1, ρ2, κ, d, such that for every x ∈ M and r > 0 one has (1.5) V (x, 2r) ≤ CdV (x, r). 4 FABRICE BAUDOIN AND NICOLA GAROFALO For a purely analytical proof of (1.5) in the Riemannian setting we refer the reader to the paper [BG2]. Now in their work [CD] the authors proved that the two results (1.4) and (1.5) are enough to establish the weak-(1, 1) continuity of the Riesz transforms for the space of homogeneous type (M, d). Since from integration by parts, and from the identity (Lf, f ) = k(−L)1/2fk2 the strong L2 continuity of the Riesz transform kqΓ((−L)1/2)f )k2 = kfk2 trivially follows, by the Marcinckiewicz interpolation theorem we thus obtain the following result. Theorem 1.1. Let 1 < p ≤ 2. There is a constant Cp > 0 such that for every f ∈ C ∞ (1.6) 0 (M), qΓ((−L)−1/2f )(cid:13)(cid:13)(cid:13)(cid:13)Lp(M) ≤ CpkfkLp(M). (cid:13)(cid:13)(cid:13)(cid:13) Theorem 1.1 provides the Lp continuity of the absolute value of the Riesz operator T f =qΓ((−L)−1/2f ), within the range 1 < p ≤ 2. We emphasize is that T is a sublinear operator. For the remaining range 2 ≤ p < ∞ we appeal to the real variable theory developed in the work [ACDH]. We recall the salient ingredients of the general approach in that paper: 1) etL1 = 1 (stochastic completeness); 2) global doubling condition; 3) global Poincar´e inequality; 4) Caccioppoli type inequalities; 5) Gaffney type estimates; 6) bounds for √tpΓ(etL). As for 1) the stochastic completeness in our framework follows as a special case of Theorem 3.5 in [BG1], see also [Mu] for an extension of such result. Regarding 2) we have already discussed (1.5). As for 3), we mention that in [BBG] it was proved that there exists Cp > 0, depending only on ρ1, ρ2, κ, d, such that for every x ∈ M and r > 0 one has (1.7) Γ(f )dµ, ZB(x,r) f − fB2dµ ≤ Cpr2ZB(x,r) for every f ∈ C 1(B(x, r)). Thus 3. is available to us. We are thus missing ingredients 4), 5) and 6) In this note we establish these results, see Corollary 3.5, Lemmas 3.6, 3.7 and 3.10, and Theorem 4.1 below. This allows us to close the circle and, by using the work [ACDH], obtain the following result. Theorem 1.2. Let 2 ≤ p < ∞. There is a constant Cp > 0 such that for every f ∈ C ∞ (1.8) 0 (M), By combining Theorems 1.1 and 1.2 we obtain the following result. qΓ((−L)−1/2f )(cid:13)(cid:13)(cid:13)(cid:13)Lp(M) ≤ CpkfkLp(M). (cid:13)(cid:13)(cid:13)(cid:13) Theorem 1.3. Let 1 < p < ∞. There exist constants Ap, Bp > 0 such that f ∈ C ∞ (1.9) Apk(−L)1/2fkp ≤ kpΓ(f )kp ≤ Bpk(−L)1/2fkp, 0 (M), RIESZ TRANSFORMS 5 The results in this paper establish the continuity of the Riesz transform and the equivalence of the Sobolev spaces defined by completion of C ∞ 0 (M) with respect to the two seminorms in (3.2) for the various classes of sub-Riemannian manifolds which are encompassed by the general framework of [BG1]. While we refer the reader to that source for a detailed discussion of the examples, here we confine ourselves to mention the following basic result which is a corollary of our work. Theorem 1.4. Let (M, θ) be a CR manifold with real dimension 2n + 1 and vanishing Tanaka- Webster torsion, i.e., a Sasakian manifold. If there exists ρ1 ≥ 0 such that for every x ∈ M the Tanaka-Webster Ricci tensor satisfies the bound for every horizontal vector v ∈ Hx, then given any 1 < p < ∞ the Riesz transform associated with a sub-Laplacian on M is continuous on Lp(M). Ricx(v, v) ≥ ρ1v2, In connection with Theorem 1.4 we mention that it was proved in [BG1] that in the framework of Theorem 1.4 the generalized curvature-dimension inequality CD(ρ1, ρ2, κ, d) does hold with ρ1 ≥ 0. Thus these manifolds fall within the scope of the assumptions in Section 2. In closing, we mention some known partial results related to those in the present paper. In [LV] the boundedness of the Riesz transforms was proved on every stratified nilpotent Lie group. In [A] this result was generalized to Lie groups of polynomial growth. Acknowledgment: The second named author would like to thank Steve Hofmann for several helpful discussions. 2. Background 2.1. Assumptions. Hereafter in this paper, M will be a C ∞ connected and non-compact mani- fold endowed with a smooth measure µ. Throughout the paper, the notation Lp(M), 1 ≤ p ≤ ∞, indicates the space of p-summable functions on M with respect to the measure µ. We assume that on M a second-order diffusion operator L with real coefficients is given. We also suppose that L be locally subelliptic (for the relevant definition and properties of such operators see [FSC] and [JSC]), and that it satisfy: 1) L1 = 0; 2) RM f Lgdµ =RM gLf dµ; 3) RM f Lf dµ ≤ 0, for every f, g ∈ C ∞ There is a natural gradient (or rather, a natural square of the length of a gradient) canonically associated with L, and it is given by the quadratic functional Γ(f ) = Γ(f, f ), where 0 (M). (2.1) Γ(f, g) = 1 2 (L(f g) − f Lg − gLf ), f, g ∈ C ∞(M). The functional Γ(f ) is known as le carr´e du champ. Notice that Γ(1) = 0. Furthermore, using the results in [PS], locally in the neighborhood of every point x ∈ M we can write (2.2) X ∗ i Xi, L = − m Xi=1 where the vector fields Xi are Lipschitz continuous (such representation is not unique, but this fact is of no consequence for us. We note for further reference that the number m of vector fields entering in the local representation (2.2) is bounded above by the dimension of the manifold 6 FABRICE BAUDOIN AND NICOLA GAROFALO M). Therefore, for any x ∈ M there exists an open neighborhood Ux such that in Ux we have for any f ∈ C ∞(M) (2.3) Γ(f ) = (Xif )2. m Xi=1 This shows that Γ(f ) ≥ 0 and it actually only involves differentiation of order one. Furthermore, the value of Γ(f )(x) does not depend on the particular representation (2.2) of L. With the operator L we can also associate a canonical distance: d(x, y) = sup{f (x) − f (y) f ∈ C ∞(M),kΓ(f )k∞ ≤ 1} , (2.4) where for a function g on M we have let g∞ = ess sup A tangent vector v ∈ TxM is called subunit for L at x if v = Pm i ≤ 1, see [FP]. It turns out that the notion of subunit vector for L at x does not depend on the local representation (2.2) of L. A Lipschitz path γ : [0, T ] → M is called subunit for L if γ′(t) is subunit for L at γ(t) for a.e. t ∈ [0, T ]. We then define the subunit length of γ as ℓs(γ) = T . Given x, y ∈ M, we indicate with i=1 aiXi(x), with Pm x, y ∈ M, M g. i=1 a2 S(x, y) = {γ : [0, T ] → M γ is subunit for L, γ(0) = x, γ(T ) = y}. In this paper we assume that Under such assumption it is easy to verify that S(x, y) 6= ∅, for every x, y ∈ M. (2.5) ds(x, y) = inf{ℓs(γ) γ ∈ S(x, y)}, defines a true distance on M. Furthermore, thanks to Lemma 5.43 in [CKS] we know that hence we can work indifferently with either one of the distances d or ds. d(x, y) = ds(x, y), x, y ∈ M, In addition to the differential form (2.1), we assume that M be endowed with another smooth bilinear differential form, indicated with ΓZ , satisfying for f, g ∈ C ∞(M) (2.6) and ΓZ (f ) = ΓZ(f, f ) ≥ 0. Given the first-order bilinear forms Γ and ΓZ on M, we now introduce the following second-order differential forms: ΓZ(f g, h) = f ΓZ(g, h) + gΓZ (f, h), 1 (2.7) (2.8) Γ2(f, g) = 2(cid:2)LΓ(f, g) − Γ(f, Lg) − Γ(g, Lf )(cid:3), 2(cid:2)LΓZ(f, g) − ΓZ (f, Lg) − ΓZ(g, Lf )(cid:3). 2 ≡ 0 as well. As for Γ and ΓZ , we will use the notations 2 (f, g) = Observe that if ΓZ ≡ 0, then ΓZ Γ2(f ) = Γ2(f, f ), ΓZ 2 (f, f ). We make the following assumptions that will be in force throughout the paper: 2 (f ) = ΓZ ΓZ 1 (H.1) There exists an increasing sequence hk ∈ C ∞ 0 (M) such that hk ր 1 on M, and Γ(hk)∞ + ΓZ (hk)∞ → 0, as k → ∞. (H.2) For any f ∈ C ∞(M) one has Γ(f, ΓZ(f )) = ΓZ(f, Γ(f )). RIESZ TRANSFORMS 7 (H.3) The generalized curvature-dimension inequality CD (ρ1, ρ2, κ, d) be satisfied with ρ1 ≥ 0, that is: There exist constants ρ1 ≥ 0, ρ2 > 0, κ ≥ 0, and d ≥ 2 such that the inequality (2.9) Γ2(f ) + νΓZ 2 (f ) ≥ 1 d κ (Lf )2 +(cid:16)ρ1 − ν(cid:17) Γ(f ) + ρ2ΓZ(f ) hold for every f ∈ C ∞(M) and every ν > 0, where Γ2 and ΓZ 2 are defined by 2.7 and 2.8. For example, the assumptions (H.1)-(H.3) are satisfied in all Carnot groups of step two, and in all complete Sasakian manifolds whose horizontal Tanaka-Webster Ricci curvature is non negative. For further examples, including a wide class of bundles over Riemannian manifolds we refer the reader to [BG1]. In this framework: • L is essentially self-adjoint on C ∞ 0 (M), so that by using the spectral theorem for the Friedrichs extension of L in the Hilbert space L2(M), we may construct a strongly continuous contraction semigroup (Pt)t≥0 in L2(M) whose infinitesimal generator is L; • By hypoellipticity of L, (Pt)t≥0 admits a heat kernel, that is: There is a smooth function p(t, x, y), t ∈ (0,∞), x, y ∈ M, such that for every f ∈ L2(M) and x ∈ M , Ptf (x) =ZM p(t, x, y)f (y)dµ(y). Moreover, the heat kernel satisfies the two following conditions: (i) (Symmetry) p(t, x, y) = p(t, y, x); (ii) (Chapman-Kolmogorov relation) p(t + s, x, y) =RM p(t, x, z)p(s, z, y)dµ(z); • The semigroup (Pt)t≥0 is a sub-Markovian semigroup: If 0 ≤ f ≤ 1 is a function in L2(M), then 0 ≤ Ptf ≤ 1; • By the Riesz-Thorin interpolation theorem, (Pt)t≥0 defines a contraction semigroup on Lp(M), 1 ≤ p ≤ ∞. 2.2. Some known results. In this section we collect some results that, under the above listed assumptions, were proved in the works [BG1] and [BBG]. Such results constitute the backbone of the present paper. The first basic result is that the manifold M is stochastically complete with respect to L, i.e., for every t > 0 (2.10) Pt1 = etL1 = 1. We recall that this result is equivalent to the uniqueness of the bounded solution of the Cauchy problem (Lu − ut = 0, M × (0,∞), x ∈ M, u(x, 0) = ϕ(x), with bounded initial datum ϕ. The property (2.10) is a corollary of Theorem 3.5 in [BG1], see also [Mu] for an extension of such result. Another basic result is the following Gaussian upper bound that was proved in [BG1]. Theorem 2.1. For any 0 < ε < 1 there exists a constant C(ε) = C(d, κ, ρ2, ε) > 0, which tends to ∞ as ε → 0+, such that for every x, y ∈ M and t > 0 one has d(x, y)2 (4 + ε)t(cid:19) . exp(cid:18)− C(ε) V (x,√t) p(x, y, t) ≤ In the paper [BBG] it has been proved that the metric measure space (M, d, µ) satisfies the global volume doubling property and that, furthermore, the L2 Poincar´e inequality is satisfied on balls. More precisely, we have the following result. 8 FABRICE BAUDOIN AND NICOLA GAROFALO Theorem 2.2. There exist constants Cd, Cp > 0, depending only on ρ1, ρ2, κ, d, for which one has for every x ∈ M and every r > 0: (2.11) (2.12) V (x, 2r) ≤ Cd V (x, r); ZB(x,r) f − fB2dµ ≤ Cpr2ZB(x,r) Γ(f )dµ, for every f ∈ C 1(B(x, r)). We list for future use the following well-known consequence of the doubling condition (2.11). Corollary 2.3. With Cd as in (2.11) define Q = log2 Cd. Then, for every x ∈ M and any 0 < r < R < ∞ one has r(cid:19)Q V (x, R) ≤ Cd(cid:18) R (2.13) + 1(cid:19)Q In particular, if y, z ∈ M and t > 0 we have (2.14) V (y,√t) ≤ Cd(cid:18) d(y, z) √t V (x, r). V (z,√t), and also for any given α > 0, there exists a constant C > 0 depending on Cd and α, such that (2.15) ZM exp(cid:18)−α d(y, z)2 t (cid:19) dµ(z) ≤ CV (y,√t). Proof. The proof of (2.13) is standard and we omit it. As for (2.14) it is enough to apply (2.13) with R = d(y, z) + √t and r = √t, to obtain V (y,√t) ≤ V (z, d(y, z) + √t) ≤ Cd(cid:18) d(y, z) + √t √t (cid:19) Q V (z,√t). Finally, (2.15) easily follows by covering M with dyadic rings B(y, 2k+1√t) \ B(y, 2k√t), and (cid:3) then using (2.11). Our objective in this section is proving that the Riesz transform associated to L is bounded in Lp(M). 1 3. Riesz transform Theorem 3.1. Let for every 1 < p < ∞. There is a constant Cp > 0 such that for every f ∈ Lp(M), (3.1) As a consequence of Theorem 3.1, we obtain the following result. (cid:13)(cid:13)(cid:13)(cid:13) qΓ((−L)−1/2f )(cid:13)(cid:13)(cid:13)(cid:13)Lp(M) ≤ CpkfkLp(M). Apk(−L)1/2fkp ≤ kpΓ(f )kp ≤ Bpk(−L)1/2fkp, Theorem 3.2. Let 1 < p < ∞. There exist constants Ap, Bp > 0 such that f ∈ C ∞ (3.2) 1In the case where µ(M) < ∞ one has to consider the space Lp [ACDH]; this modification will be implicit in the text. 0 (M). 0(M) of functions in Lp(M) with mean 0, see RIESZ TRANSFORMS 9 3.1. The case 1 < p ≤ 2. Following our discussion in the introduction, the boundedness of the Riesz transform on Lp(M) for 1 < p ≤ 2 follows by combining Theorem 4.1 and (2.11) in Theorem 2.2 above with Theorem 1.1 in [CD]. We note explicitly that (2.11) implies that (M, d, µ) is a space of homogeneous type according to [CW], see also [C], and so all tools of real analysis are available. From these results the weak-(1, 1) continuity of the Riesz transform µ({x ∈ M qΓ((−L)−1/2f )(x)) > λ}) ≤ C λ kfkL1(M), λ > 0, can be established as in [CD]. Then, for the range 1 < p ≤ 2 the inequality (3.1) follows by applying Marcinckiewicz real interpolation theorem. The reader should notice that the operator T =qΓ((−L)−1/2f )) is a sublinear operator, i.e., T (f + g)(x) ≤ T f (x) + T g(x) for every f, g and a.e. x ∈ M. 3.2. The case p > 2. Following the general method in the proof of Theorem 3.1 in [ACDH], to establish the boundedness of the Riesz transform when p > 2, we need to the ingredients listed as 1)-6) in the introduction. As it was mentioned there the items which are at this point missing are Caccioppoli and Gaffney type estimates, as well as bounds for √tpΓ(etL). This section is devoted to filling this gap. We begin with establishing the former type of result. 3.2.1. Caccioppoli type estimates. In what follows we prove an a priori inequality of Caccioppoli type for the heat semigroup Pt. Proposition 3.3. Let f ∈ C ∞ 0 (M). For t ≥ 0 we have 1 + 2κ ρ2 Γ(Ptf ) + ρ2tΓZ (Ptf ) ≤ (Pt(f 2) − (Ptf )2). 2t Proof. Let us fix T > 0. Given a function f ∈ C0(M), for 0 ≤ t ≤ T we introduce the functionals φ1(x, t) = Γ(PT −tf )(x), φ2(x, t) = ΓZ (PT −tf )(x), which are defined on M × [0, T ]. It is is easy to check that, with Γ2 and ΓZ and (2.8), we have 2 defined as in (2.7) and Lφ1 + Lφ2 + ∂φ1 ∂t ∂φ2 ∂t = 2Γ2(PT −tf ). = 2ΓZ 2 (PT −tf ), see also [BG1]. Consider now the function φ(x, t) = a(t)φ1(x, t) + b(t)φ2(x, t) = a(t)Γ(PT −tf )(x) + b(t)ΓZ (PT −tf )(x), where a and b are two nonnegative functions that will be chosen later. At this point we observe that, since by hypothesis ρ1 ≥ 0, the generalized curvature-dimension CD(ρ1, ρ2, κ, d) in (2.9) trivially implies the generalized curvature-dimension inequality CD(0, ρ2, κ,∞). Applying the latter with the choice ν = b a we thus obtain Lφ + ∂φ ∂t = a′Γ(PT −tf ) + b′(PT −tf )ΓZ(PT −tf ) + 2aΓ2(PT −tf ) + 2b(PT −tf )ΓZ 2 (PT −tf ) ≥(cid:18)a′ − 2κ a2 b (cid:19) Γ(PT −tf ) + (b′ + 2ρ2a)ΓZ (PT −tf ). 10 FABRICE BAUDOIN AND NICOLA GAROFALO Let us now chose a(t) = 1 ρ2 (T − t), b(t) = (T − t)2. With this choice we have on [0, T ], and We find then Lφ + a2 b b′ + 2ρ2a ≡ 0, 1 ρ2 − 2(cid:19) Γ(PT −tf ). 1 ρ2 − a′ − 2κ ∂t ≥(cid:18)− = − 2κ ρ2 2 2κ ρ2 ∂φ . From a comparison theorem for parabolic partial differential equations (see for instance p.52 in [F] or Proposition 3.2 in [BG1]) we deduce (3.3) PT (φ(·, T ))(x) ≥ φ(x, 0) −(cid:18) 1 ρ2 + 2κ ρ2 2(cid:19)Z T 0 Pt(Γ(PT −tf ))dt. To conclude, we consider the functional A straightforward computation shows that Ψ(t) = 1 2 Pt(cid:0)(PT −tf )2)(cid:1) . This gives Ψ′(t) = = 1 2 1 2 ∂ ∂t (PT −tf )(cid:19) Pt(cid:0)L(PT −tf )2)(cid:1) + Pt(cid:18)PT −tf Pt(cid:18)L(PT −tf )2) − 2PT −tf L(PT −tf )(cid:19) = Pt (Γ(PT −tf )) . 2(cid:0)PT (f 2) − (PT f )2(cid:1) . Pt(Γ(PT −tf )dt = Ψ(T ) − Ψ(0) = 1 Z T 0 Replacing this information in (3.3), along with the identities φ(x, 0) = 1 ρ2 T Γ(PT f ) + T 2ΓZ (PT f ), φ(x, T ) = 0, we reach the desired conclusion. (cid:3) As a corollary of Proposition 3.3 and of the Lp continuity of Pt, for 1 ≤ p ≤ ∞, we obtain the following Caccioppoli type estimates. Corollary 3.4. For any f ∈ C ∞ 0 (M), 2 ≤ p ≤ ∞ we have kpΓ(Ptf )kp ≤s1 + 2κ 2t ρ2 kfkp. In what follows we will also need the following result. Corollary 3.5. Let n be the dimension of M. With C =q (2κ+ρ2)n 2ρ2 C √t (3.4) . This estimate implies that for any f ∈ L∞(M), kpΓ(Ptf )k∞ ≤ (3.5) sup x∈MZMpΓ(p(·, y, t))(x)dµ(y) ≤ C √tkfk∞. we have RIESZ TRANSFORMS 11 Proof. For x ∈ M fixed, let then Ux be a sufficiently small neighborhood of x in which we can write L as in (2.2), where the vector fields Xi are Lipschitz continuous. For f ∈ C ∞ 0 (M), we have (cid:12)(cid:12)(cid:12)(cid:12) ZM Xip(·, y, t)(x)f (y)dµ(y)(cid:12)(cid:12)(cid:12)(cid:12) Let ε > 0 and take now = XiPtf (x) ≤ kpΓ(Ptf )k∞ ≤s 1 + 2κ 2t ρ2 kfk∞ f (y) = hk(y) Xip(·, y, t)(x) Xip(·, y, t)(x) + ε , where 0 ≤ hk ≤ 1 is an increasing sequence in C ∞ Xip(·, y, t)(x)2 Xip(·, y, t)(x) + ε ZM hk(y) dµ(y) ≤s 1 + 2κ 2t ρ2 0 (M), converging to 1. We obtain By the monotone convergence theorem and by Fatou's theorem we deduce, first letting k → ∞ and then ε → 0, that ZM Xip(·, y, t)(x)dµ(y) ≤s 1 + 2κ 2t ρ2 . The estimate (3.4) follows immediately from the latter inequality. (cid:3) 3.2.2. Gaffney-type estimates. We now turn to the second main ingredient which are Gaffney type estimates. In what follows we indicate with E, F ⊂ M two closed subsets. We need the following results. Lemma 3.6. For every two closed sets E, F ⊂ M, and any f ∈ L∞(M) supported in E, one has PtfL2(F ) ≤ e− d(E,F )2 4t fL2(E). Proof. Let us suppose d(E, F ) > 0, otherwise the conclusion follows trivially from the L2 con- tinuity of Pt. As a first step we let ψ denote a function in Lip(M), such that Γ(ψ) ≤ 1 a.e. on M. With α > 0 and φ = e−αψ, we note that (3.6) Let f ∈ L∞(M), and consider Γ(φ) = α2φ2Γ(ψ) ≤ α2φ2. y(t) = φPtf2 L2(M). Denoting with < ·,· > the inner product in L2(M), we have y′(t) = 2 < L(Ptf ), φ2Ptf >= −2ZM Γ(Ptf, φ2Ptf )dµ = −2ZM ≤ −2ZM φ2Γ(Ptf )dµ − 4ZM εZM φ2Γ(Ptf )dµ + 2 φPtf Γ(Ptf, φ)dµ φ2Γ(Ptf )dµ + 2εα2ZM φ2(Ptf )2dµ, where to estimate the last term we have used (3.6). Choosing ε = 1 we conclude and, upon integrating this inequality, we obtain (3.7) φPtfL2(M) ≤ eα2tφfL2(M). y′(t) ≤ 2α2y(t), 12 FABRICE BAUDOIN AND NICOLA GAROFALO We now want to show that (3.7) yields the desired conclusion. Suppose that suppf ⊂ E, then we argue as follows. We take ψ(x) = d(x, F ), and for any α > 0 we let φ = e−αψ. Since φ ≡ 1 on F , from (3.7) we obtain PtfL2(F ) ≤ φPtfL2(M) ≤ eα2tφfL2(M) = eα2tφfL2(E). Now, for any x ∈ E we have ψ(x) ≥ d(E, F ), and thus φ ≤ e−αd(E,F ) on E. This gives PtfL2(F ) ≤ eα2t−αd(E,F )fL2(E). By choosing α = d(E,F ) 2t > 0 we reach the desired conclusion. (cid:3) For ω > 0 sufficiently small denote Sω = {z = reiθ ∈ C 0 < r < ∞, θ < π 2 + ω}. Since L generates an analytic semigroup ezL in a sector Sω, we have Pt = 1 2πiZΓδ etζ R(ζ; L)dζ, π where Γδ ⊂ Sω is the path composed of the two rays reiθ and re−iθ, with 0 < r < ∞ and 2 < θ < π 2 + δ, 0 < δ < ω, and R(ζ; L) is the resolvent of L. Using the same argument as in Lemma 3.6 it is easy to see that the Gaffney estimate (3.7) continues to be valid for Pzf with z ∈ Sδ for any fixed 0 < δ < ω. Similarly to what was done above this leads to an estimate of the type PzfL2(F ) ≤ eα2ℜz−αd(E,F )fL2(E), (3.8) for any f ∈ L∞(M) which is supported in E, and any α > 0. Lemma 3.7. There exists a constant C > 0 such that for every two closed sets E, F ⊂ M, and any f ∈ L∞(M) supported in E, one has z ∈ Sδ, tLPtfL2(F ) ≤ Ce− d(E,F )2 6t fL2(E). Proof. Consider the path Γt = {z = t(cid:18)1 + eiθ 2 (cid:19) ∈ C 0 ≤ θ ≤ 2π} ⊂ Sω. Using the analyticity of Pt in the sector Sω we can write t ∂Pt ∂t = t 2πiZΓt ezL (z − t)2 dz = Using (3.8), this gives 1 π Z 2π 0 t(cid:16)1+ eiθ 2 (cid:17)L e−iθe dθ. tLPtfL2(F ) ≤ t(cid:16)1+ eiθ 2 (cid:17)fL2(F ) )−αd(E,F )fL2(E) 0 1 P π Z 2π ≤ 2eα2t(1+ cos θ ≤ 2e 2 3 2 α2t−αd(E,F )fL2(E) Choosing α = d(E,F ) 3t yields the desired conclusion. (cid:3) RIESZ TRANSFORMS 13 Since the distance function y → d(x, y) is obviously a Lipschitz continuous function on M (with respect to d itself), in view of (2.2) and (2.3), and of a Rademacher type theorem for Lipschitz vector fields we can construct Lipschitz continuous cut-off functions on metric balls, see for instance Theorem 1.5 in [GN]. We collect this fact in the following lemma. Lemma 3.8. Let 0 < s < t < ∞. There exists a constant C > 0 such that for every B(x, s) ⊂ B(x, t) ⊂ M there exists a function ϕ ∈ Lip(M), with 0 ≤ ϕ ≤ 1, ϕ ≡ 1 on B(x, s), and supp ϕ ⊂ B(x, t), for which pΓ(ϕ) ≤ C t − s . Corollary 3.9. Given a closed set F ⊂ M, consider the open set Fε = {x ∈ M d(x, F ) < ε}. There exists a function ϕ ∈ Lip(M) such that 0 ≤ ϕ ≤ 1, ϕ ≡ 1 on Fε/2, and ϕ ≡ 0 in M \ Fε, and for which Proof. It follows from Lemma 3.8 by a standard partition of unity argument. C ε . pΓ(ϕ) ≤ (cid:3) We are now in a position to establish the third Gaffney type estimate which we will need. Lemma 3.10. There exist constants C ≥ 0 and α > 0 such that for any two closed sets E, F ⊂ M, and every function f ∈ L∞(M) supported in E, one has kfkL2(E). Proof. We adapt the argument on p. 930 in [ACDH]. If d(E, F ) ≤ √t there is nothing to prove. We can thus assume that d(E, F ) > √t. With ε = d(E,F ) function ϕ ∈ Lip(M) supported in Fε as in Corollary 3.9. We have tZM √tkpΓ(Ptf )kL2(F ) ≤ Ce−α d(E,F )2 , consider the set Fε, and pick a ϕ2Γ(Ptf )dµ = t 3 t ϕ2(cid:2)L(cid:0)(Ptf )2(cid:1) − 2Ptf L(Ptf )(cid:3) dµ ϕ2Ptf LPtf dµ − 2tZM 2ZM = −tZM ≤ tLPtfL2(Fε)PtfL2(M) + 2(cid:18)tZM ϕPtf Γ(ϕ, Ptf )dµ We now use Lemma 3.7 and the L2 continuity of Pt to obtain for some α > 0 f2 tLPtfL2(Fε)PtfL2(M) ≤ Ce− d(E,Fε)2 f2 6t Recalling that √t < d(E, F ), and using the support property of Γ(ϕ), and the estimate Γ(ϕ) ≤ Cd(E, F )−2 from Corollary 3.9, we conclude L2(E). Γ(ϕ)(Ptf )2dµ(cid:19)1/2 . ϕ2Γ(Ptf )dµ(cid:19)1/2(cid:18)tZM L2(E) ≤ Ce−α d(E,F )2 t (cid:18)√tZM Γ(ϕ)(Ptf )2dµ(cid:19)1/2 (Ptf )2dµ(cid:19)1/2 ≤ C(cid:18)ZFε ≤ Ce− d(E,Fε)2 4t fL2(E) ≤ Ce−α d(E,F )2 t for an appropriate α > 0. We conclude tZM ϕ2Γ(Ptf )dµ ≤ Ce−α d(E,F )2 f2 A trivial estimate allows to conclude that t L2(E) + Ce−α d(E,F )2 t ϕ2Γ(Ptf )dµ ≤ Ce−α d(E,F )2 t tZM fL2(E)(cid:18)tZM f2 L2(E). fL2(E), ϕ2Γ(Ptf )dµ(cid:19)1/2 . 14 Since FABRICE BAUDOIN AND NICOLA GAROFALO we have reached the desired conclusion. tkpΓ(Ptf )k2 L2(F ) ≤ tZM ϕ2Γ(Ptf )dµ, (cid:3) 3.2.3. The completion of the proof Theorem 3.1 in the range 2 ≤ p < ∞. We are now in a position to prove Theorem 3.1 in the range 2 ≤ p < ∞. As mentioned above, all we need to do at this point is combine Theorem 2.2, Corollary 3.5 and Lemmas 3.6, 3.7 and 3.10 with the real variable results in [ACDH]. 4. Pointwise gradient estimates of the heat kernel In this section we investigate the validity of pointwise gradient estimates of the heat kernel. Besides being interesting in their own right, such estimates are also connected with Theorem 3.1 in the range 2 ≤ p < ∞. For a detailed discussion of this aspect we refer the reader to [ACDH]. Theorem 4.1. There exists a constant C = C(d, κ, ρ2) > 0 such that for every x, y ∈ M and t > 0 one has pΓ (p(·, y, t))(x) ≤ C √tV (y,√t) . Proof. We fix a point x ∈ M and s > 0 and begin with the observation that if we consider the function fx,s(z) = p(x, z, s), then thanks to Theorem 4.1 there exists a constant C > 0 (independent of x ∈ M and s > 0) such that fx,s ∈ L∞(M) and kfx,skL∞(M) ≤ V (x,√s) . C We next observe that, given points x, y ∈ M and t > 0, then we can write p(x, y, t) = P t 2 (fy, t 2 )(x). For x ∈ M fixed, let then Ux be a sufficiently small neighborhood of x in which we can write L as in (2.2). We thus have 2 2 (fy, t Xip(·, y, t)(x) = Xi(cid:16)P t ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) rΓ(cid:16)P t √tV (y,pt/2) (fy, t ≤ C 2 2 )(cid:17) (x) ≤rΓ(cid:16)P t )(cid:17)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)L∞(M) √t(cid:13)(cid:13)(cid:13) ≤ C , 2 (fy, t 2 fy, t )(cid:17)(x) 2(cid:13)(cid:13)(cid:13)L∞(M) where in the last inequality we have used the L∞ Caccioppoli inequality in Corollary 3.5. This estimate finally gives pΓ (p(·, y, t))(x) =vuut d Xi=1 (Xip(·, y, t)(x))2 ≤ C √tV (y,pt/2) The proof is completed by an application of (2.11) in Theorem 2.2. . (cid:3) RIESZ TRANSFORMS 15 Theorem 4.2. For every ε > 0 there exists a constant C(ε) = C(d, κ, ρ2, ε) > 0 such that for every x, y ∈ M and t > 0 one has pΓ (p(·, y, t))(x) ≤ C(ε) √tV (y,√t) exp(cid:18)− d(x, y)2 4(1 + ε)t(cid:19) . Proof. It suffices to appeal to Theorem 4.11 in [CS]. In such result the authors, by a beautiful use of Phragm´en-Lindelof theory, prove that the combination of (2.13) in Corollary 2.3, and of the estimates and p(x, x, t) ≤ pΓ (p(·, y, t))(x) ≤ √tV (y,√t)(cid:18)1 + C pΓ (p(·, y, t))(x) ≤ C V (x,√t) , C √tV (y,√t) , allows to improve the estimate in Theorem 4.1 into the following one d(x, y)2 4t (cid:19)1+3Q exp(cid:18)− d(x, y)2 4t (cid:19) , where Q is as (2.13). From the latter estimate the desired conclusion follows. (cid:3) References [ACDH] P. Auscher, T. Coulhon, X.T. Duong, S. Hofmann, Riesz transform on manifolds and heat kernel regu- larity Ann. Sci. ´Ecole Norm. Sup. (4) 37 (2004), no. 6, 911-957. [A] G. Alexopoulos, An application of homogenization theory to harmonic analysis: Harnack inequalities and Riesz transforms on Lie groups of polynomial growth. Canad. J. Math. 44 (1992), no. 4, 691 -- 727. [B] D. Bakry, ´Etude des transformations de Riesz dans les vari´et´es riemanniennes `a courbure de Ricci minor´ee. (French) [A study of Riesz transforms in Riemannian manifolds with minorized Ricci curvature] S´eminaire de Probabilit´es, XXI, 137 -- 172, Lecture Notes in Math., 1247, Springer, Berlin, 1987. [BG1] F. Baudoin & N. Garofalo, Curvature-dimension inequalities and Ricci lower bounds for sub-Riemannian manifolds with transverse symmetries, Arxiv preprint 1101.3590, submitted paper (2009). [BG2] , Perelman's entropy and doubling property on Riemannian manifolds, J. Geom. Anal., to appear. [BBG] F. Baudoin. M. Bonnefont & N. Garofalo, A sub-Riemannian curvature-dimension inequality, volume doubling property and the Poincar´e inequality, arXiv:1007.1600, submitted, (2010). [CKS] E. Carlen, S. Kusuoka & D. Stroock, Upper bounds for symmetric Markov transition functions, Ann. Inst. H. Poincar´e Probab. Statist. 23 (1987), no. 2, suppl., 245 -- 287. [C] M. Christ, A T (b) theorem with remarks on analytic capacity and the Cauchy integral, Colloq. Math. 60/61 (1990), no. 2, 601628. [CW] R. R. Coifman & G. Weiss, Analyse harmonique non-commutative sur certains espaces homog`enes. (French) ´Etude de certaines int´egrales singuli`eres, Lecture Notes in Mathematics, Vol. 242. Springer-Verlag, Berlin- New York, 1971. v+160 pp. [CD] T. Coulhon, X. T. Duong, Riesz transforms for 1 ≤ p ≤ 2, Trans. Amer. Math. Soc., 351 3 (1999), 1151-1169. [CS] T. Coulhon, A. Sikora, Gaussian heat kernel upper bounds via the Phragmn-Lindelf theorem. Proc. Lond. Math. Soc. (3) 96 (2008), no. 2, 507 -- 544. [FP] C. Fefferman & D. H. Phong, Subelliptic eigenvalue problems, Conference on harmonic analysis in honor of Antoni Zygmund, Vol. I, II (Chicago, Ill., 1981), 590 -- 606, Wadsworth Math. Ser., Wadsworth, Belmont, CA, 1983. [FSC] C. L. Fefferman & A. S´anchez-Calle, Fundamental solutions for second order subelliptic operators, Ann. of Math. (2) 124 (1986), no. 2, 247 -- 272. [F] A. Friedman, Partial differential equations of parabolic type, Dover, 2008. [GN] N. Garofalo & D.-M. Nhieu, Lipschitz continuity, global smooth approximations and extension theorems for Sobolev functions in Carnot-Carathodory spaces, J. Anal. Math. 74 (1998), 67-97. [JSC] D. Jerison & A. S´anchez-Calle, Subelliptic second order differential operators, Lecture. Notes in Math., 1277 (1987), pp. 46-77. 16 FABRICE BAUDOIN AND NICOLA GAROFALO [LV] Lohou´e, N., Varopoulos, N.: Remarques sur les transform´ees de Riesz sur les groupes nilpotents. C.R.A.S. Paris, 301, 11 (1985) 559-560. [Mu] I. Munive, Generalized curvature-dimension inequalities, stochastic completeness and volume growth, preprint, 2011. [PS] R. S. Phillips & L. Sarason, Elliptic-parabolic equations of the second order, J. Math. Mech. 17 1967/1968, 891-917. [S] E. M. Stein, Singular integrals and differentiability properties of functions, Princeton Univ. Press, 1970. [Str] R. Strichartz, Analysis of the Laplacian on the complete Riemannian manifold, J. Funct. Anal. 52 (1983), 48-79. Department of Mathematics, Purdue University, West Lafayette, IN 47907 E-mail address, Fabrice Baudoin: [email protected] Department of Mathematics, Purdue University, West Lafayette, IN 47907 E-mail address, Nicola Garofalo: [email protected]
1708.09549	1	1708	2017-08-31T03:29:09	Necessary and sufficient conditions for boundedness of commutators of bilinear Hardy-Littlewood maximal function	[ "math.FA" ]	Let $\mathcal{M}$ be the bilinear Hardy-Littlewood maximal function and $\vec{b}=(b,b)$ be a collection of locally integrable functions. In this paper, the authors establish characterizations of the weighted {\rm BMO} space in terms of several different commutators of bilinear Hardy-Littlewood maximal function, respectively; these commutators include the maximal iterated commutator $\mathcal{M}_{\Pi \vec{b}}$, the maximal linear commutator $\mathcal{M}_{\Sigma\vec{b}}$, the iterated commutator $[\Pi \vec{b},\mathcal{M}]$ and the linear commutator $[\Sigma \vec{b},\mathcal{M}]$.	math.FA	math	Appl. Math. J. Chinese Univ. 2013, 28(): -* Necessary and sufficient conditions for boundedness of commutators of bilinear Hardy-Littlewood maximal function Wang Ding-huai and Zhou Jiang∗ Abstract. Let M be the bilinear Hardy-Littlewood maximal function and ~b = (b, b) be a collection of locally integrable functions. In this paper, the authors establish characterizations of the weighted BMO space in terms of several different commutators of bilinear Hardy-Littlewood maximal function, respectively; these commutators include the maximal iterated commutator Σ~b, the iterated commutator [Π~b, M] and the linear Π~b, the maximal linear commutator M M commutator [Σ~b, M]. §1 Introduction A locally integrable function f is said to belong to BMO space if there exists a constant C > 0 such that for any cube Q ⊂ Rn, 1 Q ZQ f (x) − fQdx ≤ C, where fQ = 1 Q RQ f (x)dx and the minimal constant C is defined by kf k∗. There are a number of classical results that demonstrate BMO functions are the right collections to do harmonic analysis on the boundedness of commutators. A well known result of Coifman, Rochberg and Weiss [7] states that the commutator [b, T ](f ) = bT (f ) − T (bf ) is bounded on some Lp, 1 < p < ∞, if and only if b ∈ BMO, where T be the classical Calder´on- Zygmund operator. Chanillo [5] proved that if b ∈ BMO, the commutator is bounded from Lp to Lq with 1 < p < n/α and 1/q = 1/p − α/n, where Iα be a fractional [b, Iα]f (x) = b(x)Iαf (x) − Iα(bf )(x) Received: 2016--. MR Subject Classification: 42B20, 42B25, 42B35. Keywords: BMO function, Characterization, Commutator, Hardy-Littlewood maximal function. Digital Object Identifier(DOI): 10.1007/s11766-013-***-. Supported by the National Natural Science Foundation of China (11661075) ∗Corresponding author 2 Appl. Math. J. Chinese Univ. Vol. 28, No. * integral operator. Moreover, if n − α is even, the reverse is also valid. A complete characteriza- tion of BMO via the commutator [b, Iα] was shown by Ding [8]. During the past thirty years, the theory was then extended and generalized to several directions. For instance, Bloom [3] in- vestigated the characterization of BMO spaces in the weighted setting. In 1991, Garc´ıa-Cuerva, Harboure, Segovia and Torrea [11] showed that the maximal commutator Mb(f )(x) = sup Q∋x 1 Q ZQ b(x) − b(y)f (y)dy is bounded on Lp, 1 < p < ∞, if and only if b ∈ BMO. In 2000, Bastero, Milman and Ruiz [1] studied the necessary and sufficient conditions for the boundedness of [b, M ] on Lp spaces when 1 < p < ∞. They showed that the commutator of Hardy-Littlewood maximal operator [b, M ](f )(x) = b(x)M (f )(x) − M (bf )(x) is bounded on Lp, 1 < p < ∞, if and only if b ∈ BMO with b− ∈ L∞, where b−(x) = − min{b(x), 0}. In 2014, Zhang [24] considered the characterization of BMO via the commutator of the fractional maximal function on variable exponent Lebesgue spaces. In the multilinear setting, the boundedness of commutators has been extensively studied already, as in P´erez and Torres' [16], Tangs [19], Lerner, Ombrosi, P´erez, Torres, and Trujillo- Gonz´alezs [12] and Chen and Xues [6], and P´erez, Pradolini, Torres, and Trujillo-Gonz´alezs [15]. Specially, Chaffee and Torres [4], Wang, Pan and Jiang [20] and Zhang [23] contributed the theory of characterization of BMO spaces by considering the linear commutator of Multilinear operators, respectively. In this paper, we will extend Zhang's result to weighted case and we replace the linear commutators by iterated commutators. Our main results as follows. Theorem 1.1. Let 1 < p1, p2 < ∞,~b = (b, b), 1/p = 1/p1 + 1/p2 and ω ∈ A1. Then the following are equivalent, (A1) b ∈ BMO(ω); (A2) MΣ~b is bounded from Lp1 (ω) × Lp2 (ω) to Lp(ω1−p); (A3) MΠ~b is bounded from Lp1(ω) × Lp2(ω) to Lp(ω1−2p). Theorem 1.2. Let 1 < p1, p2 < ∞,~b = (b, b), 1/p = 1/p1 + 1/p2 and ω ∈ A1. Then the following are equivalent, (B1) b ∈ BMO(ω) and b−/ω ∈ L∞; (B2) [Σ~b, M] is bounded from Lp1 (ω) × Lp2 (ω) to Lp(ω1−p); (B3) [Π~b, M] is bounded from Lp1(ω) × Lp2(ω) to Lp(ω1−2p). Wang Ding-huai and Zhou Jiang. Commutators of bilinear Hardy-Littlewood maximal function 3 §2 Some preliminaries and notations In 2009, Lerner, Ombrosi, P´erez, Torres and Trujillo-Gonz´alez [12] introduced the following multilinear maximal function that adapts to the multilinear Calder´on-Zygmund theory. In this paper, we only consider the bilinear case. A similar argument also works for the multilinear cases. Definition 2.1. For a collection of locally integrable functions ~f = (f1, f2), the bilinear maxi- mal function M is defined by M( ~f )(x) = sup Q∋x 2 Yi=1 1 Q ZQ fi(yi)dyi. We now give the definitions of the maximal commutators and the commutators related to the bilinear maximal function M. Definition 2.2. For two collections of locally integrable functions ~f = (f1, f2) and ~b = (b1, b2), the maximal linear commutator MΣ~b is defined by Xi=1 MΣ~b( ~f )(x) = ( ~f )(x), 2 M(i) bi where M(i) bi ( ~f )(x) = sup Q∋x 1 Q2 ZQZQ bi(x) − bi(yi) 2 Yj=1 fj(yj)dy1dy2. The maximal iterated commutator MΠ~b is defined by MΠ~b( ~f )(x) = sup Q∋x 1 Q2 ZQZQ 2 Yi=1 The linear commutator of M is defined by bi(x) − bi(yi)fi(yi)dy1dy2. [Σ~b, M]( ~f )(x) = [b1, M](1)( ~f )(x) + [b2, M](2)( ~f )(x), [b1, M](1)( ~f )(x) = b1(x)M( ~f )(x) − M(b1f1, f2)(x) where and [b2, M](2)( ~f )(x) = b2(x)M( ~f )(x) − M(f1, b2f2)(x). The iterated commutator of M is defined by [Π~b, M]( ~f )(x) = b1(x)b2(x)M( ~f )(x) − b1(x)M(f1, b2f2)(x) −b2(x)M(b1f1, f2)(x) + M(b1f1, b2f2)(x). We now recall the definition of Ap weight introduced by Muckenhoupt [13]. Definition 2.3. For 1 < p < ∞ and a nonnegative locally integrable function ω on Rn, ω is in the Muckenhoupt Ap class if it satisfies the condition sup Q (cid:18) 1 Q ZQ ω(x)dx(cid:19)(cid:18) 1 Q ZQ ω(x)− 1 p−1 dx(cid:19)p−1 < ∞. 4 Appl. Math. J. Chinese Univ. Vol. 28, No. * And a weight function ω belongs to the class A1 if there exists C > 0 such that for every cube Q, 1 Q ZQ ω(x)dx ≤ C ess inf x∈Q ω(x). We write A∞ = S1≤p<∞ Ap. Definition 2.4. Let 1 ≤ p < ∞. Given a a nonnegative locally integrable function ω, the weighted BMO space BMOp(ω) is defined be the set of all functions f ∈ L1 loc(Rn) such that kf kBMOp(w) := sup Q (cid:18) 1 w(Q) ZQ f (y) − fQpω(y)1−pdy(cid:19)1/p < ∞, BMO1(ω) = BMO(ω) simple. where the supremum is taken over all cubes Q ⊂ Rn and ω(Q) = RQ ω(x)dx. We write Remark For 1 ≤ p < ∞ and ω ∈ A1, Garc´ıa-Cuerva [10] proved that BMO(ω) = BMOp(ω) with equivalence of the corresponding norms. Standard real analysis tools as the weighted maximal function Mω(f ), the sharp maximal function M ♯(f ) carries over to this context, namely, Mω(f )(x) = sup Q∋x 1 ω(Q) ZQ f (y)ω(y)dy; M ♯(f )(x) = sup Q∋x inf c 1 Q ZQ f (y) − cdy ≈ sup Q∋x 1 Q ZQ f (y) − fQdy. A variant of weighted maximal function and sharp maximal operator Mω,s(f )(x) = (cid:0)Mω(f s)(cid:1)1/s , which will become the main tool in our scheme. and M ♯ δ (f )(x) = (cid:0)M ♯(f δ)(x)(cid:1)1/δ §3 Main lemmas To prove Theorem 1.1 and Theorem 1.2, we need the following results. Lemma 3.1. Let ω ∈ A1, ~b = (b, b) and b ∈ BMO(ω). Then M ♯ 1 3(cid:0)MΠ~b( ~f )(cid:1)(x) . kbk2 BMO(ω)ω(x)2M (M( ~f )(x)) +kbk2 BMO(ω)ω(x)2 2 Yi=1 Mω,s(fi)(x) + 2 Xi=1 kbkBMO(ω)ω(x)M 1 2 (M(i) b ( ~f ))(x), for any 1 < s < ∞ and bounded compact supported functions f1, f2. Proof. First of all, we give the definition of the following auxiliary maximal function, which has been studied in [17] and [18] for the linear case. Let ϕ(x) ≥ 0 be a smooth function such that ϕǫ(t) = ǫ−2nϕ( t ǫ ), ϕ′(t) . t−1 and χ[0,1](t) ≤ ϕ(t) ≤ χ[0,2](t). Let Φ(f1, f2)(x) = sup ǫ>0ZRn ZRn ϕǫ(x − y1 + x − y2) 2 Yi=1 fi(yi)dy1dy2, Wang Ding-huai and Zhou Jiang. Commutators of bilinear Hardy-Littlewood maximal function 5 and ΦΠ~b(f1, f2)(x) = sup ǫ>0 ZRn ZRn ϕǫ(x − y1 + x − y2) 2 Yi=1 b(x) − b(yi)fi(yi)dy1dy2. We first show that Φ(f1, f2)(x) ≈ M(f1, f2)(x). In fact, let Bǫ = {y ∈ Rn : x − y ≤ ǫ}. It is easy to see that The bounded compact supported condition of ϕ gives B ǫ 2 × B ǫ 2 ⊂ (cid:8)(y1, y2) : x − y1 + x − y2 ≤ ǫ(cid:9) ⊂ Bǫ × Bǫ. ǫ>0ZRn ZRn ϕǫ(x − y1 + x − y2)f1(y1)f2(y2)dy1dy2 Φ(f1, f2)(x) = sup ≤ sup ǫ>0 1 ǫ2n ZBǫ ZBǫ ϕ(cid:16) x − y1 + x − y2 ǫ (cid:17)f1(y1)f2(y2)dy1dy2 . M(f1, f2)(x) and Φ(f1, f2)(x) ≥ sup ǫ>0 1 ǫ2n ZB ǫ 2 ZB ǫ 2 ϕ(cid:16) x − y1 + x − y2 ǫ (cid:17)f1(y1)f2(y2)dy1dy2 & M(f1, f2)(x). We can also obtain that ΦΠ~b(f1, f2)(x) ≈ MΠ~b(f1, f2)(x). Now, we shall estimate the sharp maximal function of the auxiliary maximal function. Let Q be a cube and x ∈ Q. Then, for any z ∈ Q we have (cid:12)(cid:12)ΦΠ~b(f1, f2)(z) − cQ(cid:12)(cid:12) ≤ (cid:12)(cid:12)b(z) − bQ2Φ(f1, f2)(z) +(cid:12)(cid:12)b(z) − bQΦ(f1, (b − bQ)f2)(z)(cid:12)(cid:12) +(cid:12)(cid:12)b(z) − bQΦ((b − bQ)f1, f2)(z)(cid:12)(cid:12) +(cid:12)(cid:12)Φ((b − bQ)f1, (b − bQ)f2)(z) − cQ(cid:12)(cid:12) 1 (z) + AQ 2 (z) + AQ 3 (z) + AQ i will be defined later. 4 (z), ΦΠ~b(f1, f2)(z) − cQδ(cid:12)(cid:12)(cid:12) (cid:19)1/δ . (cid:18) 1 Q ZQ(cid:12)(cid:12)(cid:12) Xj=1 Aj, 4 . 1 , (b − bQ)f ∞ Therefore, where cQ = (cid:0)Φ((b − bQ)f ∞ (cid:18) 1 Q ZQ(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)ΦΠ~b(f1, f2)(z)(cid:12)(cid:12) δ − cQδ(cid:12)(cid:12)(cid:12) =: AQ 2 )(cid:1)Q and f ∞ dz(cid:19)1/δ where Aj = (cid:0) 1 Q RQ(cid:0)AQ Let us consider first the term A1. By averaging AQ , j = 1, 2, 3, 4 and taking δ = 1/3. j (z)(cid:1)δ dz(cid:1)1/δ A1 = (cid:18) 1 Q ZQ(cid:16)(cid:12)(cid:12)b(z) − bQ2Φ(f1, f2)(z)(cid:17)1/3 1 over Q, we get dz(cid:19)3 . kbk2 BMO(ω) ω(Q)2 Q2 · 1 Q ZQ M(f1, f2)(z)dz . kbk2 BMO(ω)ω(x)2M (M(f1, f2))(x). 6 Appl. Math. J. Chinese Univ. Vol. 28, No. * Let us consider next the term A2. We write AQ 2 (z) = (cid:12)(cid:12)b(z) − bQΦ(f1, b − bQf2)(z) ≤ (cid:12)(cid:12)b(z) − bQΦ(f1,(cid:0)b(z) − bQ + b(z) − b(cid:1)f2)(z) ≤ (cid:12)(cid:12)b(z) − bQ2M(f1, f2)(z) +(cid:12)(cid:12)b(z) − bQM(2) 21(z) + AQ =: AQ 22(z). b (f1, f2)(z) For AQ 21(z), the fact that Φ(f1, f2)(z) . M(f1, f2)(z) gives A21 Q ZQ(cid:0)AQ 21(z)(cid:1)δ := (cid:18) 1 . (cid:16) ω(Q) . kbk2 kbkBMO(ω)(cid:17)2 1 Q BMO(ω)ω(x)2M (M(f1, f2))(x). M(f1, f2)(z)dz dz(cid:19)1/δ Q ZQ For AQ 22(z), A22 := (cid:18) 1 Q ZQ(cid:0)AQ . ω(x)kbkBMO(ω) dz(cid:19)1/δ 22(z)(cid:1)δ Q2(cid:18)ZQ(cid:12)(cid:12)M(2) 1 . ω(x)kbkBMO(ω)M1/2(M(2) b (f1, f2))(x). 1/2(cid:19)2 b (f1, f2)(z)(cid:12)(cid:12) The same process also follows that A3 . kbk2 BMO(ω)ω(x)2M (M(f1, f2))(x) + ω(x)kbkBMO(ω)M1/2(M(1) b (f1, f2))(x). To estimate A4, we split fj to fj = f 0 j + f ∞ j = fjχ2Q. We write j with f 0 1 , (b − bQ)f 0 AQ 4 ≤ (cid:12)(cid:12)Φ((b − bQ)f 0 +(cid:12)(cid:12)Φ(((b − bQ)f 0 +(cid:12)(cid:12)Φ((b − bQ)f ∞ +(cid:12)(cid:12)Φ((b − bQ)f ∞ 41(z) + AQ =: AQ 1 , (b − bQ)f ∞ 1 , (b − bQ)f 0 1 , (b − bQ)f ∞ 2 )(z)(cid:12)(cid:12) 2 )(z)(cid:12)(cid:12) 2 )(z)(cid:12)(cid:12) 2 )(z) − cQ(cid:12)(cid:12) 44(z). 42(z) + AQ 43(z) + AQ Then A4 ≤ (cid:18) 1 4 AQ Q ZQ(cid:0) Xj=1 (cid:18) 1 Q ZQ(cid:0)AQ Xj=1 4j (z)(cid:1)δ 4j(z)(cid:1)δ 4 dz(cid:19)1/δ dz(cid:19)1/δ . . 4 Xj=1 A4j . Wang Ding-huai and Zhou Jiang. Commutators of bilinear Hardy-Littlewood maximal function 7 By Kolmogorov inequality and the fact that M is bounded from L1 × L1 to L1/2,∞, we have 1 , (b − bQ)f 0 2 )kL1/2,∞ 1 , (b − bQ)f 0 2 )kL1/2,∞ b(yi) − bQfi(yi)dyi C Q2 kΦ((b − bQ)f 0 C Q2 kM((b − bQ)f 0 C Q2 2 Z2Q (cid:18)Z2Q Yi=1 Yi=1 2 C Q2 A41 ≤ ≤ ≤ ≤ . 2 Yi=1 kbkBMOs′ (ω)ω(x)Mω,s(fi)(x). b(yi) − bQs′ ω1−s′ (yi)dyi(cid:19)1/s′ (cid:18)Z2Q fi(yi)sω(yi)dyi(cid:19)1/s For A42, it is easy to see that ϕǫ(z − y1 + z − y2) . then A42 . 1 Q ZQZ2QZRn\2Q 1 (cid:0)z − y1 + z − y2(cid:1)2n , b(y1) − bQf1(y1)b(y2) − bQf2(y2) (cid:0)z − y1 + z − y2(cid:1)2n b(y1) − bQf1(y1)b(y2) − bQf2(y2) dy1dy2dz dy1dy2dz . ZQZ2QZRn\2Q . 1 Q Z2Q (cid:0)z − y1 + z − y2(cid:1)2n b(y1) − bQf1(y1)dy1ZQZRn\2Q 2kQ Z2kQ 2−kn ∞ . kbkBMOs′ (ω)ω(x)Mω,s(f1)(x) Xk=1 Xk=1 ∞ 2−kn 2kQ . kbkBMOs′ (ω)ω(x)Mω,s(f1)(x) b(y2) − bQf2(y2)dy2 b(y2) − bQf2(y2) z − y22n dy2dz ×(cid:20)Z2kQ b(y1) − m2kQ(b)f2(y2)dy2 +Z2kQ m2kQ(b) − bQf2(y2)dy2(cid:21) . kbkBMOs′ (ω)ω(x)Mω,s(f1)(x) ∞ Xk=1 2−kn 2kQ ×(cid:20)kbkBMOs′ (ω)ω(x)Mω,s(f2)(x) + kkbkBMO(ω)ω(x)M (f2)(x)(cid:21) . 2 Yi=1 kbkBMOs′ (ω)ω(x)Mω,s(fi)(x). Similarly, for A43, we have A43 . 2 Yi=1 kbkBMOs′ (ω)ω(x)Mω,s(fi)(x). 8 Appl. Math. J. Chinese Univ. Vol. 28, No. * For z − z′ ≤ 1 2 max{z − y1, z − y2}, (cid:12)(cid:12)ϕǫ(z − y1 + z − y2) − ϕǫ(z′ − y1 + z′ − y2)(cid:12)(cid:12) . Therefore, 1 , (b(z) − b)f ∞ z − z′ (cid:0)z − y1 + z − y2(cid:1)2n+1 . 2 )(z′)(cid:12)(cid:12) ϕǫ(z − y1 + z − y2) − ϕǫ(z′ − y1 + z′ − y2)(cid:12)(cid:12)(cid:12) 2 )(z) − Φ((b(z) − b)f ∞ 1 , (b(z) − b)f ∞ b(yi) − bQfi(yi)dyi b(yi) − bQfi(yi)dyi b(yi) − bQfi(yi)dy1dy2 2 2 × . sup (cid:12)(cid:12)Φ((b(z) − b)f ∞ ǫ>0 ZRn\2QZRn\2Q(cid:12)(cid:12)(cid:12) Yi=1 ZRn\2Q Yi=1 Yi=1 Yi=1 2kQ Z2kQ Xk=1 −2knǫi . . 2 ∞ . 2 z − z′ǫi z − yin+ǫi kbkBMOs′ (ω)ω(x)Mω,s(fi)(x), where ǫ1, ǫ2 > 0 with ǫ1 + ǫ2 = 1. Collecting our estimates, we have shown that M ♯ 1 3(cid:0)MΠ~b( ~f )(cid:1)(x) . kbk2 BMO(ω)ω(x)2M (M( ~f )(x)) +kbk2 BMO(ω)ω(x)2 2 Yi=1 Mω,s(fi)(x) for any 1 < s < ∞ and bounded compact supported functions f1, f2. + 2 Xi=1 kbkBMO(ω)ω(x)M 1 2 (M(i) b ( ~f ))(x), Lemma 3.2. Let ω ∈ A1, ~b = (b, b) and b ∈ BMO(ω). Then there exist a constant C such that M ♯ 1 2(cid:0)M(1) b ( ~f )(cid:1)(x) . kbkBMO(ω)ω(x)M (M( ~f )(x)) +kbkBMO(ω)ω(x)Mω,s(f1)(x)M (f2)(x), for any 1 < s < ∞ and bounded compact supported functions f1, f2. Proof. Let Q be a cube and x ∈ Q. Then, for z ∈ Q we have (cid:12)(cid:12)Φ(1) b (f1, f2)(z) − cQ(cid:12)(cid:12) ≤ (cid:12)(cid:12)b(z) − bQΦ(f1, f2)(z) +(cid:12)(cid:12)Φ((b − bQ)f1, f2)(z) − cQ(cid:12)(cid:12) 1 (z) + BQ 2 (z). =: BQ Wang Ding-huai and Zhou Jiang. Commutators of bilinear Hardy-Littlewood maximal function 9 Therefore, where Bj = (cid:0) 1 Q RQ(cid:0)BQ j (z)(cid:1)δ Let us consider first the term B1. By averaging BQ 1 over Q, we get dz(cid:19)2 2 . 1/2 1/2 Bj, b (f1, f2)(z)(cid:12)(cid:12) − cQ1/2(cid:12)(cid:12)(cid:12) dz(cid:19)2 (cid:18) 1 Q ZQ(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Φ(1) . (cid:18) 1 Q ZQ(cid:12)(cid:12)ΦΠ~b(f1, f2)(z) − cQ(cid:12)(cid:12) Xj=1 dz(cid:1)1/δ Q ZQ(cid:16)(cid:12)(cid:12)b(z) − bQΦ(f1, f2)(z)(cid:17)1/2 , j = 1, 2. . kbkBMO(ω) M(f1, f2)(z)dz ω(Q) Q · 1 Q ZQ . kbkBMO(ω)ω(x)M (M(f1, f2))(x). dz(cid:19)2 B1 = (cid:18) 1 Let us consider next the term B2. We split fj to fj = f 0 j with f 0 j = fjχ2Q. We write BQ 1 , f 0 2 ≤ (cid:12)(cid:12)Φ((b − bQ)f 0 +(cid:12)(cid:12)Φ((b − bQ)f ∞ 21(z) + BQ =: BQ 2 )(z)(cid:12)(cid:12) +(cid:12)(cid:12)Φ(((b − bQ)f 0 2 )(z)(cid:12)(cid:12) +(cid:12)(cid:12)Φ((b − bQ)f ∞ 23(z) + BQ 24(z). 22(z) + BQ 1 , f 0 1 , f ∞ 2 )(z)(cid:12)(cid:12) 2 )(z) − cQ(cid:12)(cid:12) j + f ∞ 1 , f ∞ By Kolmogorov inequality and the fact that M is bounded from L1 × L1 to L1/2,∞, we have B21 ≤ . . C Q2 kΦ((b − bQ)f 0 1 Q2 kM((b − bQ)f 0 1 Q2 Z2Q 1 , f 0 2 )kL1/2,∞ 1 , f 0 2 )kL1/2,∞ b(y1) − bQf1(y1)dy1Z2Q f2(y2)dy2 . kbkBMOs′ (ω)ω(x)Mω,s(f1)(x)M (f2)(x). For B22, B22 . . . 1 1 b(y1) − bQf1(y1)f2(y2) b(y1) − bQf1(y1)f2(y2) Q ZQZ2QZRn\2Q Q ZQZ2QZRn\2Q Q Z2Q (cid:0)z − y1 + z − y2(cid:1)2n (cid:0)z − y1 + z − y2(cid:1)2n b(y1) − bQf1(y1)dy1ZQZRn\2Q 2kQ Z2kQ 2−kn ∞ 1 Xk=1 . kbkBMOs′ (ω)ω(x)Mω,s(f1)(x) . kbkBMOs′ (ω)ω(x)Mω,s(f1)(x)M (f2)(x). dy1dy2dz dy1dy2dz f2(y2) z − y22n dy2dz f2(y2)dy2 10 Appl. Math. J. Chinese Univ. Vol. 28, No. * For B23, we have B23 . . . 1 1 Q ZQZRn\2QZ2Q Q ZQZ2QZRn\2Q Q ZQZRn\2Q 1 b(y1) − bQf1(y1)f2(y2) b(y1) − bQf1(y1)f2(y2) (cid:0)z − y1 + z − y2(cid:1)2n (cid:0)z − y1 + z − y2(cid:1)2n dy1dzZ2Q z − y12n b(y1) − bQf1(y1) dy1dy2dz dy1dy2dz f2(y2)dy2 . kbkBMOs′ (ω)ω(x)Mω,s(f1)(x)M (f2)(x). Concerning the last estimate for B24. For any z′ ∈ Q and y1, y2 ∈ Rn\2Q, we have 1 , f ∞ 2 )(z′)(cid:12)(cid:12) ϕǫ(z − y1 + z − y2) − ϕǫ(z′ − y1 + z′ − y2)(cid:12)(cid:12)(cid:12) . sup 1 , f ∞ 2 )(z) − Φ((b − bQ)f ∞ ×b(y1) − bQf1(y1)f2(y2)dy1dy2 (cid:12)(cid:12)Φ((b − bQ)f ∞ ǫ>0ZRn\2QZRn\2Q(cid:12)(cid:12)(cid:12) . ZRn\2Q Xk=2 . kbkBMOs′ (ω)ω(x)Mω,s(f1)(x)M (f2)(x). 2kQ Z2kQ dy1ZRn\2Q b(y1) − bQf1(y1)dy1 b(y1) − bQf1(y1) z − y12n+ǫ1 2−knǫ1 . ∞ f2(y2) z − y2ǫ2 dy2 ∞ Xi=2 2−knǫ2 2kQ Z2iQ f2(y2)dy2 where ǫ1, ǫ2 > 0 with ǫ1 + ǫ2 = 1. Taking the mean over Q for z and z′ respectively, we obtain B24 . . 1 1 , f ∞ Q ZQ(cid:12)(cid:12)Φ((b − bQ)f ∞ Q ZQ Q ZQ(cid:12)(cid:12)Φ((b − bQ)f ∞ 1 1 2 )(z) − cQ(cid:12)(cid:12)dz . kbkBMOs′ (ω)ω(x)Mω,s(f1)(x)M (f2)(x). 1 , f ∞ 2 )(z) − Φ((b − bQ)f ∞ 1 , f ∞ 2 )(z′)(cid:12)(cid:12)dzdz′ Collecting our estimates, we have shown that M ♯ 1 2(cid:0)M(1) b ( ~f )(cid:1)(x) . kbkBMO(ω)ω(x)M (M( ~f )(x)) +kbkBMO(ω)ω(x)Mω,s(f1)(x)M (f2)(x), for any 1 < s < ∞ and bounded compact supported functions f1, f2. Similarly, we have Lemma 3.3. Let ω ∈ A1, ~b = (b, b) and b ∈ BMO(ω). Then there exist a constant C such that M ♯ 1 2(cid:0)M(2) b ( ~f )(cid:1)(x) . kbkBMO(ω)ω(x)M (M( ~f )(x)) +kbkBMO(ω)ω(x)Mω,s(f2)(x)M (f1)(x), for any 1 < s < ∞ and bounded compact supported functions f1, f2. Lemma 3.4. Let ω ∈ A1 and 0 < p < ∞. Then ω1−p ∈ A∞. Proof. If 0 < p ≤ 1, then 1 − p ∈ [0, 1). It is easy to see that ω1−p ∈ A1 ⊂ A∞. If 1 < p < ∞, if follows from ω ∈ A1 ⊂ Ap that ω1−p ∈ Ap′ ⊂ A∞. Wang Ding-huai and Zhou Jiang. Commutators of bilinear Hardy-Littlewood maximal function 11 Lemma 3.5. Let ω ∈ A1, 1 < s < p1, p2 < ∞ and 1/p = 1/p1 + 1/p2. Then both M( ~f ) and Q2 i=1 Mω,s(fi) are bounded from Lp1(ω) × Lp2(ω) to Lp(ω). 1 < s < p1, p2 < ∞, it is easy to obtain that both M( ~f ) and Q2 Proof. From the fact that M (f )(x) . Mω,s(f )(x) and Mω,s(f )(x) is bounded on Lp(ω) for i=1 Mω,s(fi) are bounded from Lp1 (ω) × Lp2 (ω) to Lp(ω). The following relationships between Mδ and M ♯ to be used is a version of the classical ones due to Fefferman and Stein [9]. Lemma 3.6. Let 0 < p, δ < ∞ and ω ∈ A∞. There exist a positive C such that ZRn (Mδf (x))pω(x)dx ≤ CZRn (M ♯ δ f (x))pω(x)dx, for any smooth function f for which the left-hand side is finite. Lemma 3.7. Let Q0 be any fixed cube and b be a locally integral function. Then, for any x ∈ Q0, we get (1) (2) (3) M(χQ0 , χQ0)(x) ≡ 1; M(bχQ0, χQ0 )(x) = M(χQ0 , bχQ0)(x) = MQ0 (b)(x); where MQ0(b)(x) = supQ0⊃Q∋x 1 M(bχQ0, bχQ0)(x) = M2 Q RQ b(y)dy. Q0 (b)(x), Proof. We only give the proof of (3) and the proof of (1),(2) are similar. For any x ∈ Q0, we have M 2 Q0(b)(x) = (cid:16) sup Q0⊃Q∋x 1 sup 1 Q ZQ Q ZQ = Q0⊃Q∋x b(y)dy(cid:17)2 b(y1)χQ0 (y1)dy1 · 1 Q ZQ b(y2)χQ0 (y2)dy2 ≤ M(bχQ0, bχQ0 )(x). On the other hand, for any cube Q ⊂ Rn, we can construct a cube Q1 such that and Q1 ≤ Q. Therefore, Q0 ⊃ Q1 ⊃ Q0 ∩ Q ∋ x 1 Q ZQ∩Q0 b(y)dy ≤ 1 Q1 ZQ1 b(y)dy ≤ MQ0(b)(x). Thus, then (3) is proved. M(bχQ0 , bχQ0)(x) = sup Q∋x(cid:16) 1 Q ZQ b(y)χQ0 (y)dy(cid:17)2 ≤ M 2 Q0(b)(x), §4 Proofs of Theorem 1.1 and Theorem 1.2 Proof of Theorem 1.1. (A1) ⇒ (A2): It is enough to prove Theorem 1.1 for f1, f2 being bounded functions with compact support. We observe that to use the Fefferman-Stein inequal- ity, one needs to verify that certain terms in the left-hand side of the inequalities are finite. 12 Appl. Math. J. Chinese Univ. Vol. 28, No. * Applying a similar argument as in [12, pp.32-33], the boundedness properties of M and Fatou's lemma, one gets the desired result. Since Lemma 3.4 and ω ∈ A1, then ω1−p ∈ A∞. By Lemma 3.2 and Lemma 3.3 with 1 < s < min{p1, p2}, from a standard argument that we can obtain kMΣ~b( ~f )kLp(ω1−p) . kM 1 2(cid:0)MΣ~b( ~f )(cid:1)kLp(ω1−p) . kM ♯ . kbkBMO(ω)(cid:18)(cid:13)(cid:13)M(cid:0)M( ~f )(cid:1)(cid:13)(cid:13)Lp(ω) +(cid:13)(cid:13) 2(cid:0)MΣ~b( ~f )(cid:1)kLp(ω1−p) Mω,s(fi)(cid:13)(cid:13)Lp(ω)(cid:19) Yi=1 2 1 2 . kbkBMO(ω) kfikLpi (ω). Yi=1 (A2) ⇒ (A1): Let Q be any fixed cube. Suppose that MΣ~b is bounded from Lp1 (ω)×Lp2(ω) into Lp(ω1−p), then kMΣ~b(χQ, χQ)kLp(ω1−p) . kχQkLp1 (ω)kχQkLp2 (ω) . ω(Q) 1 p , which implies that b(x) − bQdx 2 ω(Q) ZQ ω(Q) ZQ 1 ≤ Q−2ZQZQ b(x) − b(y1)χQ(y1)χQ(y2)dy1dy2dx + 1 ω(Q) ZQ Q−2ZQZQ b(x) − b(y2)χQ(y1)χQ(y2)dy1dy2dx . . 1 MΣ~b(χQ, χQ)(x)dx ω(Q) ZQ ω(Q)(cid:16)ZQ(cid:12)(cid:12)M~b(χQ, χQ)(x)(cid:12)(cid:12) 1 1 . ω(Q)1/p kMΣ~b(χQ, χQ)kLp(ω1−p) . kMΣ~bkLp1 (ω)×Lp2 (ω)→Lp(ω1−p). p ω(x)1−pdx(cid:17)1/p(cid:16)ZQ ω(x)dx(cid:17)1/p′ Thus showing that b ∈ BMO(ω). (A1) ⇒ (A3): Since ω ∈ A1, Lemma 3.4 implies that ω1−2p ∈ A∞. From Lemma 3.1, Lemma 3.2 and Lemma 3.3 with 1 < s < min{p1, p2}, we get kMΠ~b( ~f )kLp(ω1−2p) . kM 1 . kbk2 3(cid:0)MΠ~b( ~f )(cid:1)kLp(ω1−2p) . kM ♯ BMO(ω)(cid:18)(cid:13)(cid:13)M(cid:0)M( ~f )(cid:1)(cid:13)(cid:13)Lp(ω) +(cid:13)(cid:13) Xi=1 kbkBMO(ω)(cid:13)(cid:13)M 1 (M(i) 2 2 2 + . kbk2 BMO(ω) kfikLpi (ω). Yi=1 1 2 3(cid:0)MΠ~b( ~f )(cid:1)kLp(ω1−2p) Mω,s(fi)(cid:13)(cid:13)Lp(ω)(cid:19) Yi=1 b ( ~f ))(x)(cid:13)(cid:13)Lp(ω1−p) Wang Ding-huai and Zhou Jiang. Commutators of bilinear Hardy-Littlewood maximal function 13 (A3) ⇒ (A1): By MΠ~b is bounded from Lp1 (ω) × Lp2 (ω) into Lp(ω1−2p), we get 1 1 b(x) − bQ2ω(x)−1dx ω(x)−1Q−2ZQZQ MΠ~b(χQ, χQ)(x)ω(x)−1dx ω(Q) ZQ ω(Q) ZQ ω(Q) ZQ ω(Q)(cid:16)ZQ(cid:12)(cid:12)MΠ~b(χQ, χQ)(x)(cid:12)(cid:12) 1 1 1 . . . . ω(Q)1/p kMΠ~b(χQ, χQ)kLp(ω1−2p) . kMΠ~bkLp1 (ω)×Lp2 (ω)→Lp(ω1−2p). b(x) − b(y1)b(x) − b(y2)dy1dy2dx p ω(x)1−2pdx(cid:17)1/p(cid:16)ZQ ω(x)dx(cid:17)1/p′ Thus we complete the proof of Theorem 1.1. Proof of Theorem 1.2. (B1) ⇒ B2): By the definition of M( ~f ), we have M (bf1, f2)(x) = M (bf1, f2)(x), M (f1, bf2)(x) = M (f1, bf2)(x). Then (cid:12)(cid:12)[b, M](1)( ~f )(x) − [b, M](1)( ~f )(x)(cid:12)(cid:12) . (cid:12)(cid:12)(cid:12) b(x)M( ~f )(x) − M(bf1, f2)(x) − b(x)M( ~f )(x) + M(bf1, f2)(x)(cid:12)(cid:12)(cid:12) . b−(x)M( ~f )(x). Similarly, we also have (cid:12)(cid:12)[b, M](2)( ~f )(x) − [b, M](2)( ~f )(x)(cid:12)(cid:12) . b−(x)M( ~f )(x). Since (cid:12)(cid:12)a − c(cid:12)(cid:12) ≤ a − c for any real numbers a and c, there holds for i = 1, 2. This shows that (cid:12)(cid:12)[b, M](i)(f1, f2)(x)(cid:12)(cid:12) ≤ M(i) b (f1, f2)(x), Applying (4) and Theorem 1.1 we have (cid:12)(cid:12)Σ~b, M]( ~f )(x)(cid:12)(cid:12) . MΣ~b( ~f )(x) + b−(x)M( ~f )(x). (4) (cid:13)(cid:13)[Σ~b, M]( ~f )(x)(cid:13)(cid:13)Lp(ω1−p) . (cid:13)(cid:13)MΣ~b( ~f )(cid:13)(cid:13)Lp(ω1−p) +(cid:13)(cid:13)b−M( ~f )(cid:13)(cid:13)Lp(ω1−p) . (cid:0)kb−/ωkL∞ + kbkBMO(ω)(cid:1)kf1kLp1 (ω)kf2kLp2 (ω). Therefore, b ∈ BMO(ω) with b−/ω ∈ L∞ implies that [Σ~b, M] is bounded from Lp1(ω)× Lp2(ω) to Lp(ω1−p). (B2) ⇒ (B1): Let Q0 be any fixed cube. By Lemma 3.7, for any x ∈ Q0, b(x) = b(x)M(χQ0 , χQ0)(x), MQ0 (b)(x) = M(bχQ0, χQ0 )(x) = M(χQ0 , bχQ0)(x), 14 Then, Appl. Math. J. Chinese Univ. Vol. 28, No. * = 2 2 ω(Q0) ZQ0 ω(Q0) ZQ0 ω(Q0) ZQ0 1 = b(x) − MQ0(b)(x)dx b(x)M(χQ0 , χQ0)(x) − M(bχQ0 , χQ0)(x)dx b(x)M(χQ0 , χQ0)(x) − M(bχQ0 , χQ0)(x) +b(x)M(χQ0 , χQ0 )(x) − M(χQ0, bχQ0 )(x)dx 1 . . . 1 ω(Q0) ZQ0 (cid:12)(cid:12)[Σ~b, M](χQ0, χQ0 )(x)(cid:12)(cid:12)dx ω(Q0)(cid:18)ZQ0 (cid:12)(cid:12)(cid:12) [Σ~b, M](χQ0, χQ0 )(x)(cid:12)(cid:12)(cid:12) ω(Q0)1/p(cid:13)(cid:13)[Σ~b, M](χQ0 , χQ0)(cid:13)(cid:13)Lp(ω1−p) 1 p . k[Σ~b, M]kLp1 (ω)×Lp2 (ω)→Lp(ω1−p). ω(x)1−pdx(cid:19)1/p ·(cid:18)ZQ0 ω(x)dx(cid:19)1/p′ Now, we have all the ingredients to prove b ∈ BMO(ω) and b−/ω ∈ L∞. 1 ω(Q0) ZQ0 b(x) − bQ0dx . 1 ω(Q0) ZQ0 b(x) − MQ0(b)(x)dx . k[Σ~b, M]kLp1 (ω)×Lp2 (ω)→Lp(ω1−p). which implies that b ∈ BMO(ω). In order to show show that b−/ω ∈ L∞, observe that for any x ∈ Q0, MQ0 (b)(x) ≥ b(x). Therefore, which gives 0 ≤ b−(x) . MQ0 (b)(x) − b+(x) + b−(x) = MQ0 (b)(x) − b(x), 1 Q0 ZQ0 b−(x) ω(x) dx . . 1 Q0 ZQ0 Q0 ZQ0 1 b−(x)dx · 1 inf x∈Q0 ω(x) b(x) − MQ0 (b)(x)dx · Q0 ω(Q0) . k[Σ~b, M]kLp1 (ω)×Lp2 (ω)→Lp(ω1−p), this yields that Thus, the boundedness of b−/ω follows from Lebesgue's differentiation theorem. (b−/ω)Q0 . k[Σ~b, M]kLp1 (ω)×Lp2 (ω)→Lp(ω1−p). (B1) ⇒ (B3): Let ~B = (b, b) and ~B = (b, b). Then b(x)b(x)M( ~f )(x) − b(x)M(f1, bf2)(x) (cid:12)(cid:12)(cid:12) [Π~b, M](f1, f2)(x) − [Π ~B, M](f1, f2)(x)(cid:12)(cid:12)(cid:12) . (cid:12)(cid:12)(cid:12) −b(x)b(x)M( ~f )(x) + b(x)M(f1, bf2)(x)(cid:12)(cid:12)(cid:12) . b−(x)(cid:12)(cid:12)[b, M](2)(f1, f2)(x)(cid:12)(cid:12). Wang Ding-huai and Zhou Jiang. Commutators of bilinear Hardy-Littlewood maximal function 15 Similarly, we also have Noting that which yields that b(x)b(x)M( ~f )(x) − b(x)M(bf1, f2)(x) (cid:12)(cid:12)(cid:12) [Π~B, M](f1, f2)(x) − [Π ~B, M](f1, f2)(x)(cid:12)(cid:12)(cid:12) . (cid:12)(cid:12)(cid:12) −b(x)b(x)M( ~f )(x) + b(x)M(bf1, f2)(x)(cid:12)(cid:12)(cid:12) . b−(x)(cid:12)(cid:12)[b, M](1)(f1, f2)(x)(cid:12)(cid:12). (cid:12)(cid:12)[Π~B, M]( ~f )(x)(cid:12)(cid:12) ≤ MΠ~b( ~f )(x), It follows from Theorem 1.1 and b−/ω ∈ L∞ that (cid:12)(cid:12)[Π~b, M]( ~f )(x)(cid:12)(cid:12) . MΠ~b( ~f )(x) + b−(x)MΣ~b( ~f )(x) + (b−(x))2M( ~f )(x). (cid:13)(cid:13)[Π~b, M]( ~f )(x)(cid:13)(cid:13)Lp(ω1−2p) . (cid:13)(cid:13)MΠ~b( ~f )(cid:13)(cid:13)Lp(ω1−2p) +(cid:13)(cid:13)b−MΣ~b( ~f )(cid:13)(cid:13)Lp(ω1−2p) +(cid:13)(cid:13)(b−)2M( ~f )(cid:13)(cid:13)Lp(ω1−2p) BMO(ω)kf1kLp1 (ω)kf2kLp2 (ω) + kb−/ωkL∞(cid:13)(cid:13)MΣ~b( ~f )(cid:13)(cid:13)Lp(ω1−p) . kbk2 +kb−/ωk2 L∞kM( ~f )kLp(ω) . (cid:0)kb−/ωkL∞ + kbkBMO(ω)(cid:1)2 this leads to our results. kf1kLp1 (ω)kf2kLp2 (ω), (B3) ⇒ (B1): Let Q0 be any fixed cube. By Lemma 3.5, for any x ∈ Q0, b(x)2 = b(x)2M(χQ0 , χQ0 )(x), b(x)MQ0 (b)(x) = b(x)M(bχQ0 , χQ0 )(x) = b(x)M(χQ0 , bχQ0)(x), M 2 Q0 (b)(x) = M(bχQ0 , bχQ0)(x). Then, 1 1 1 1 = b(x) − MQ0(b)(x)2ω(x)−1dx ω(Q0) ZQ0 ω(Q0) ZQ0 (cid:16)b(x)2 − 2b(x)MQ0 (b)(x) + M 2 ω(Q0) ZQ0 [Π~b, M](χQ0, χQ0 )(x)ω(x)−1dx ω(Q0)(cid:18)ZQ0 (cid:12)(cid:12)(cid:12) [Π~b, M](χQ0 , χQ0 )(x)(cid:12)(cid:12)(cid:12) ω(Q0)1/p(cid:13)(cid:13)[Π~b, M](χQ0 , χQ0 )(cid:13)(cid:13)Lp(ω1−2p) . k[Π~b, M]kLp1 (ω)×Lp2 (ω)→Lp(ω1−2p). = 1 = . Q0(b)(x)(cid:17)ω(x)−1dx p ω(x)1−2pdx(cid:19)1/p ·(cid:18)ZQ0 ω(x)dx(cid:19)1/p′ 16 Appl. Math. J. Chinese Univ. Vol. 28, No. * Now, we have all the ingredients to prove b ∈ BMO(ω) and b−/ω ∈ L∞. 1 ω(Q0) ZQ0 ω(Q0) ZQ0 1 . b(x) − bQ0 2ω(x)−1dx b(x) − MQ0(b)(x)2ω(x)−1dx + 1 ω(Q0) ZQ0 bQ0 − MQ0(b)(x)2ω(x)−1dx 1 . b(x) − MQ0(b)(x)2ω(x)−1dx ω(Q0) ZQ0 . k[Π~b, M]kLp1 (ω)×Lp2 (ω)→Lp(ω1−2p). which implies that b ∈ BMO2(ω); that is, b ∈ BMO(ω). For any x ∈ Q0, we have b−(x) ω(x) Q0 ZQ0 1 dx . . 1 1 Q0 ZQ0 Q0 ZQ0 ω(Q0) ZQ0 b−(x)dx · 1 inf x∈Q0 ω(x) b(x) − MQ0(b)(x)dx · Q0 ω(Q0) . (cid:18) 1 . k[Π~b, M]kLp1 (ω)×Lp2 (ω)→Lp(ω1−2p), b(x) − MQ0 (b)(x)2ω(x)−1dx(cid:19)1/2 which yields Thus, the boundedness of b−/ω follows from Lebesgue's differentiation theorem. (b−/ω)Q0 . k[Π~b, M]kLp1 (ω)×Lp2 (ω)→Lp(ω1−2p). The proof of Theorem 1.2 is complete. §5 Acknowledgments The research was supported by National Natural Science Foundation of China (Grant NO.11661075). References [1] J Bastero, M Milman, F J Ruiz. Commutators for the maximal and sharp functions, Proc. Amer. Math. Soc., 2000(128): 3329-3334. [2] ´A B´enyi, W Dami´an, K Moen and R H Torres. Compactness properties of commutators of bilinear fractional integrals, Math. Z. (2015), doi: 10.1007/s00209-015-1437-4. [3] S Bloom. A commutator theorem and weighted BMO, Trans. Amer. Math. Soc., 1985(292): 103-122. [4] L Chaffee, R H Torres. Characterization of Compactness of the Commutators of Bilinear Fractional Integral Operators, Potential Anal., 2015(43): 481-494. Wang Ding-huai and Zhou Jiang. Commutators of bilinear Hardy-Littlewood maximal function 17 [5] S Chanillo. A note on commutators, Indiana Univ. Math. J., 1982(31): 7-16. [6] X Chen, Q Xue. Weighted estimates for a class of multilinear fractional type operators, J. Math. Anal. Appl. (2010(362): 355-373. [7] R Coifman, R Rochberg, G Weiss. Factorization theorems for Hardy spaces in several variables, Ann. of Math, 1976(103): 611-635. [8] Y Ding. A characterization of BMO via commutators for some operators, Northeast. Math. J., 1997(13): 422-432. [9] C Fefferman, E M Stein. H p spaces of several variables Acta Math., 1972(129): 137-193. [10] J Garc´ıa-Cuerva. Weighted H p spaces, Dissertationes Math. 1979(162). [11] J Garc´ıa-Cuerva, E Harboure, C Segovia, J L Torrea. Weighted norm inequalities for commutators of strongly singular integrals, Indiana Univ. Math. J., 1991(40): 1397-1420. [12] A K Lerner, S Ombrosi, C P´erez, R H Torres, R Trujillo-Gonz´alez. New maximal functions and multiple weights for the multilinear Calder´on-Zygmund theory, Adv. Math., 2009(220) 1222-1264. [13] B Muckenhoupt. Weighted norm inequalities for the Hardy maximal function, Trans. Amer. Math. Soc., 1972(165): 207-226. [14] C P´erez. Endpoint estimates for commutators of singular integral operators, J. Funct. Anal., (1995(128): 163-185. [15] C P´erez, G Pradolini, R H Torres, R Trujillo-Gonz´alez. End-points estimates for iterated commutators of multilinear singular integrals, Bull. London Math. Soc., 2014(46): 26-42. [16] C. P´erez and R.H. Torres, Sharp maximal function estimates for multilinear singular integrals, Contemp. Math. 320 (2003) 323-331. 1045.42011 MR1979948 [17] C Segovia, J L Torrea. Weighted inequalities for commutators of fractional and singular integrals, Publicacions Matem`atuques, (1991(35): 209-235. [18] C Segovia, J L Torrea. Higher order commutators for vector-valued Calder´on-Zygmund operators, Trans. Amer. Math. Soc., 1993(336): 537-556. [19] L Tang. Weighted estimates for vector-valued commutators of multilinear operators, Proc. Roy. Soc. Edinburgh Sect. A, (2008(138): 897-922. [20] S B Wang, J B Pan, Y S Jiang. Necessary and sufficient conditions for boundedness of commutators of multilinear fractional integral operators, Acta Math. Sin., 2015(35): 1106- 1114. (in Chinese) 18 Appl. Math. J. Chinese Univ. Vol. 28, No. * [21] M Wilson. Weighted Littlewood-Paley Theory and Exponential-Square Integrability, Lec- ture Notes in Math vol. 1924. Springer, Berlin (2008). [22] M M Rao, Z D Ren. Theory of Orlicz Spaces, Marcel Dekker, New York (1991). [23] P Zhang, Multiple Weighted Estimates for Commutators of Multilinear Maximal Function, Acta Math. Sin. (Engl. Ser.), 2015(31): 973-994. [24] P Zhang, Commutators of the fractional maximal function on variable exponent Lebesgue spaces, Czech. Math. J., 2014(64): 183-197. College of Mathematics and System Sciences, Xinjiang University, Urumqi 830046, Republic of China Email:[email protected]; [email protected].
1003.5066	4	1003	2012-06-28T06:46:31	A Bernstein-type inequality for rational functions in weighted Bergman spaces	[ "math.FA" ]	Given $n\geq1$ and $r\in[0, 1),$ we consider the set $\mathcal{R}_{n, r}$ of rational functions having at most $n$ poles all outside of $\frac{1}{r}\mathbb{D},$ were $\mathbb{D}$ is the unit disc of the complex plane. We give an asymptotically sharp Bernstein-type inequality for functions in $\mathcal{R}_{n, r}\:$ (as n tends to infinity and r tends to 1-) in weighted Bergman spaces with "polynomially" decreasing weights. We also prove that this result can not be extended to weighted Bergman spaces with "super-polynomially" decreasing weights.	math.FA	math	A BERNSTEIN-TYPE INEQUALITY FOR RATIONAL FUNCTIONS IN WEIGHTED BERGMAN SPACES ANTON BARANOV AND RACHID ZAROUF Abstract. Given n ≥ 1 and r ∈ [0, 1), we consider the set Rn, r of rational functions having at most n poles all outside of 1 D, were D is the unit disc of the complex plane. We give an asymptotically sharp Bernstein-type inequality for functions in Rn, r in weighted Bergman spaces with "polynomially" decreasing weights. We also prove that this result can not be extended to weighted Bergman spaces with "super-polynomially" decreasing weights. r 1. Introduction Estimates of the norms of derivatives for polynomials and rational functions (in different functional spaces) is a classical topic of complex analysis (see surveys by A.A. Gonchar [10], V.N. Rusak [16], and P. Borwein and T. Erd´elyi [3, Chapter 7]). Such inequalities have applications in many domains of analysis; to mention just some of them: 1) matrix analysis and in operator theory (see "Kreiss Matrix Theorem" [12, 17] or [19, 18] for resolvent estimates of power bounded matrices), 2) inverse theorems of rational approximation (see [4, 15, 14]), 3) effective Nevanlinna -- Pick interpolation problems (see [23, 22]). Rn, r =(cid:26) p q Here, we present Bernstein-type inequalities for rational functions f of degree n with poles in {z : z > 1}, involving Hardy norms and weighted Bergman norms. Let Pn be the complex space of polynomials of degree less or equal to n ≥ 1. Let D = {z ∈ C : z < 1} be the unit disc of the complex plane and D = {z ∈ C : z ≤ 1} its closure. Given r ∈ [0, 1), we define : p, q ∈ Pn, d◦p < d◦q, q(ζ) 6= 0 ζ < r(cid:27) , (where d◦p denotes the degree of p ∈ Pn), the set of all rational functions in D of degree less or equal than n ≥ 1, having at most n poles all outside of 1 D. Notice that for r = 0, we get Rn, 0 = Pn−1. 1.1. Definitions of Hardy spaces and radial weighted Bergman spaces. We denote by Hol (D) the space of all holomorphic functions on D. From now on, if f ∈ Hol (D) then for every ρ ∈ (0, 1) we define ξ ∈ fρ : ξ 7→ f (ρξ) , 1 ρ D. 1 r We consider the two following scales of Banach spaces X ⊂ Hol (D) : 2000 Mathematics Subject Classification. Primary 32A36, 26A33; Secondary 26C15, 41A10. Key words and phrases. Rational function, Bernstein-type inequality, weighted Bergman norm. The first author is supported by the Chebyshev Laboratory (St.Petersburg State University) under RF Government grant 11.G34.31.0026 and by RFBR grant 12-01-00434. 1 2 ANTON BARANOV AND RACHID ZAROUF a. The Hardy spaces H p = H p(D), 1 ≤ p ≤ ∞ : H p =(cid:26)f ∈ Hol (D) : kfkp Hp = sup 0≤ρ<1ZT fρ(ξ)p dm(ξ) < ∞(cid:27) , where m stands for the normalized Lebesgue measure on T = {z ∈ C : z = 1}. As usual,we denote by H ∞ the space of all bounded analytic functions in D. a (w), 1 ≤ p < ∞ (where "a" means b. The radial weighted Bergman spaces Lp analytic), 0 Lp Lp a (w) =(cid:26)f ∈ Hol (D) : kfkp a(w) =Z 1 where the weight w satisfies w ≥ 0 and R 1 weights w(ρ) = wβ(ρ) = (1 − ρ)β, β > −1, we have Lp A being the normalized area measure on D. ρw (ρ)ZT fρ (ζ)p dm(ζ)dρ < ∞(cid:27) , a(cid:0)(1−z)βdA(z)(cid:1), 0 w(ρ)dρ < ∞. For the classical power For general properties of these spaces we refer to [11, 24]. From now on, for two positive functions a and b, we say that a is dominated by b, denoted by a . b, if there is a constant c > 0 such that a ≤ cb; and we say that a and b are comparable, denoted by a ≍ b, if both a . b and b . a. 1.2. Statement of the problem and known results. By Bernstein-type in- equalities for rational functions one usually understands the inequalities of the form a (wβ) = Lp (1.1) kf ′kX ≤ φX, Y (n)kfkY , f ∈ Rn, where Rn is the set of all proper rational functions of degree at most n with the poles in {z > 1}, X and Y are some normed spaces of functions analytic in the unit disc, and φ is some increasing (often polynomially growing) function. Thus, for a given pair of the function spaces X and Y , the question is to determine the dependence on n for the norm of the differentiation operator (Rn,k · kX ) to Y . Bernstein-type inequalities of E.P. Dolzhenko [5] and A.A. Pekarskii [14] are of this form; e.g., it is shown in [5] that kfkH 1 1 ≤ c1nkfk∞, kfkB 1/2 2,2 ≤ c2n1/2kfk∞, f ∈ Rn, 1 is the Hardy -- Sobolev space, and B1/2 where H 1 2,2 is the Besov (or Dirichlet) space, see the definition in Section 3. Let us also mention that this problem is a part of a more general one given by G. Lorentz in a letter sent to T. Erd´elyi in 1988 (see [9]). Looking at (1.1), we notice that for some choices of X and Y , we have φX, Y (n) = +∞ for every n = 1, 2, . . . . Indeed, it may happen for instance when the poles of our function f are allowed to be arbitrary close to the torus T : we can observe this phenomenon for example in the special case X = Y = H p, 1 ≤ p ≤ +∞ but also when X = Y = Lp a(w), 1 ≤ p ≤ +∞. This observation leads us to come back on the problem in (1.1) and to state it more generally : that is replacing Rn by Rn, r (for any fixed r ∈ [0, 1)) and φX, Y (n) by φX, Y (n, r) so that to focus on this phenomenon of "natural dependence on the parameter r". For most of the classical cases already studied by others (for instance E. P. Dolzhenko [5], A. A. Pekarskii [14], V.V. Peller [15]) the spaces X and Y are such that supr∈(0, 1) φX, Y (n, r) < +∞: in this case we can set φX, Y (n) = supr∈(0, 1) φX, Y (n, r). As a consequence, A BERNSTEIN-TYPE INEQUALITY FOR RATIONAL FUNCTIONS 3 if supr∈(0, 1) φX, Y (n, r) = +∞, it may be of interest to search (as a continuation of the investigations of the second author [20, 21]) for the "best possible" φX, Y (n, r) in an asymptotically sense, that is to say as n → ∞ and r → 1−. This question has already been answered for the case X = Y = H p, 1 ≤ p ≤ +∞ by K. M. Dyakonov [6] see (1.2) below. In this paper, we answer the same question for the case X = Y = Lp a(w), 1 ≤ p ≤ +∞. Let us give a general formulation of our problem for the special case X = Y for which we set Cn, r(X) = φX, Y (n, r) : given a Banach space X of holomorphic functions in D, we are searching for the best possible constant Cn, r(X) such that kf ′kX ≤ Cn, r(X)kfkX , f ∈ Rn, r. For the case where X = H p is a Hardy space, an estimate which gives a correct order of growth for Cn, r(X) was obtained by K.M. Dyakonov [6] (as a very special case of more general results): for any p ∈ [1,∞] there exist positive constants Ap and Bp such that (1.2) Ap n 1 − r ≤ Cn, r(H p) ≤ Bp n 1 − r for all n ≥ 1 and r ∈ [0, 1). More precisely, the upper estimate for p ∈ (1, +∞) is treated in [6, Theorem 1], the case p = 1, in [6, Corollary 1], and the case p = +∞ (known much earlier) is given in [3, Theorem 7.1.7]. The below estimate follows trivially when applying the differentiation operator to the test function f (z) = (1 − rz)−n. any r ∈ (0, 1) there exists the limit For the case p = 2 an asymptotically sharp result was obtained later in [20]: for lim n→∞ Cn, r(cid:0)H 2(cid:1) n = 1 + r 1 − r . Related results about Bernstein-type inequalities in a more general setting of the so-called model or star invariant subspaces may be found in [8, Theorems 10,11], [7, Theorem 1], and [1, 2]. 1.3. Main results. We obtain estimates for the derivatives of rational functions with respect to weighted Bergman norms. It turns out that there is an essen- tial difference between slowly (polynomially) decreasing weights and fast (super- polynomially) decreasing weights. In the first case we have a two-sided estimate analogous to (1.2), while in the second case only the above estimate remains true. Let us give the precise definitions. Recall that w is always an integrable nonnegative function on (0, 1). Definition 1.1. (Polynomially decreasing weights) The weight w is said to be γ-polynomially decreasing if there exists γ > 0 such that is increasing on [r0, 1) for some 0 ≤ r0 < 1. We say that w is polynomially decreasing if it is γ-polynomially decreasing for some γ > 0. ρ 7→ (1 − ρ)−γw(ρ), Definition 1.2. (Super-polynomially decreasing weights) The weight w is said to be super-polynomially decreasing if for any γ > 0 there exists r(γ) ∈ (0, 1) such that the function decreases on the interval [r(γ), 1). ρ 7→ (1 − ρ)−γw(ρ), 4 ANTON BARANOV AND RACHID ZAROUF Typical example of the weights from the first class are given by w(r) = (1 − r)β, β > −1, or w(r) = (1 − r)β ( log(1 − r) + 1)γ, β > −1, γ ∈ R. The weights Our first result may be considered as an analogue of Dyakonov's theorem for the w(r) = exp(cid:0) − c(1 − r)−γ(cid:1), c > 0, γ > 0 are super-polynomially decreasing. radial weighted Bergman spaces. Theorem 1.3. Let 1 ≤ p < ∞ and let w be an integrable nonnegative function on [0, 1). Then there exists a positive constant K depending only on p (but not on the weight w) such that (1.3) Cn, r (Lp a (w)) ≤ K n 1 − r for all r ∈ [0, 1) and n ≥ 1. Moreover, if we fix r ∈ (0, 1) and let n tend to infinity, then we have (1.4) eKr 1 − r ≤ lim inf n→∞ a(w)) Cn,r(Lp n ≤ lim sup n→∞ a(w)) Cn,r(Lp n ≤ K 1 − r , where eK is, as K, a positive constant depending only on p. The next theorem shows that for the polynomially decreasing weights the quan- a(w)) admits a below estimate of the same form. tity Cn, r(Lp Theorem 1.4. If w is γ-polynomially decreasing, then there exists a positive con- stant K ′ depending only on w and p such that (1.5) K ′ n 1 − r ≤ Cn, r (Lp a (w)) ≤ K n 1 − r , where K is defined in (1.3) and where the left-hand side inequality of (1.5) holds for all r ∈ [0, 1) and n ≥ γ+3 p + 1. In particular, (1.5) holds for the classical weights w(ρ) = wβ (ρ) = (1 − ρ)β ρ, β > −1. The polynomial decrease is essential and provides a sharp bound for the validity of the uniform estimate (1.5) for all possible values of n and r. Namely, if the weight is super-polynomially decreasing, then (1.5) will fail along some sequence of radii. Theorem 1.5. Suppose that w is super-polynomially decreasing. Then there exists a sequence rn → 1− such that for any p, Crn,n(Lp a(w)) n = o(cid:18) 1 1 − rn(cid:19), n → ∞. Acknowledgements. The authors are deeply grateful to Nikolai Nikolski, Alexander Borichev, and Evgueny Doubtsov for many helpful discussions, con- structive comments and precious remarks which definitively helped to improve the manuscript. 2. Proofs of Theorems 1.3 and 1.4 Proof of Theorem 1.3. First, we notice that for any 0 ≤ α < 1, (2.1) kfkp Lp a(w) ≍ZZu∈Cα ρf (ρξ)p w(ρ)dm(ζ)dρ A BERNSTEIN-TYPE INEQUALITY FOR RATIONAL FUNCTIONS 5 for all f ∈ Lp a(w), where Cα = {z : α < z < 1}. Let f ∈ Rn, r with r ∈ [0, 1) and n ≥ 1. Using (2.1) with α = 1 2 we get 1 Lp ρ w(ρ) kf ′kp a(w) ≍ ZZρξ∈C1/2 f ′(ρξ)p w (ρ) dm(ζ)dρ = Z 1 ρp(cid:16)(cid:13)(cid:13)(cid:13)(fρ) ′(cid:13)(cid:13)(cid:13) Hp(cid:17) dρ. Hp(cid:17) dρ ≤ (2Cn, r (H p))pZ 1 ρp(cid:16)(cid:13)(cid:13)(cid:13)(fρ) ′(cid:13)(cid:13)(cid:13) ≍ (Cn, r (H p))p kfkp a(w) . Lp 1 1 2 1 2 p p ρw (ρ)(cid:0)kfρkHp(cid:1)p dρ Now using the fact that fρ ∈ Rn, ρr ⊂ Rn, r for every ρ ∈ (0, 1), we get Z 1 1 2 ρw(ρ) In particular, using the right-hand side inequality of (1.2), we get Cn, r (Lp a (w)) ≤ Kp n 1 − r , for all p ∈ [1, ∞), and β ∈ (−1, ∞), where Kp is a constant depending on p only. Now, let us prove (1.4). Let 1 Lp fn(z) = kfnkp and, analogously, (1 − rz)n ∈ Rn,r, and D = {z ∈ D : 1 − rz ≤ 21 − r}. We claim that a(w) ∼ZD fn(z)pw(z)dA(z), a(w) ∼ZD f ′ nkp ZD\D fn(z)pw(z)dA(z) ≤ Indeed, by a very rough estimate n(z)pw(z)dA(z), kf ′ Lp C1 2pn(1 − r)pn , n → ∞, n → ∞. where C1 > 0 depends only on w. On the other hand, if we put D = {z ∈ D : 1 − rz ≤ 3 21 − r}, then Since r (thus D and D) are fixed we see that w(z)dA(z). 1 ZD fn(z)pw(z)dA(z) ≥ 2pn(1 − r)pn = o(cid:18) a(w) ∼ZD f ′ nkp kf ′ kfnkp a(w) Lp Lp 1 (3/2)pn(1 − r)pnZ D w(z)dA(z)(cid:19), 1 (3/2)pn(1 − r)pnZ D n(z)pw(z)dA(z).ZD fn(z)pw(z)dA(z). n → ∞. Thus, 6 ANTON BARANOV AND RACHID ZAROUF Obviously, ZD f ′ n(z)pw(z)dA(z) =ZD nprp 1 nprp 1 − rzpn+p w(z)dA(z) 2p(1 − r)pZD 1 − rzpn w(z)dA(z) 2p(1 − r)pZD fn(z)pw(z)dA(z). nprp ≥ = Thus, lim inf n→∞ kf ′ nkLp nkfnkLp a(w) a(w) ≥ r 2(1 − r) . (cid:3) For the proof of Theorem 1.4 we will need two lemmas. Lemma 2.1. Let r ∈ [0, 1) and t ≥ 0. We set I(t, r) =ZT 1 − rξ−t dm(ξ) and ϕr(t) =ZT 1 + rξt dm(ξ). Then, I(t, r) = 1 (1 − r2)t−1 ϕr(t − 2) for every t ≥ 2, and t 7→ ϕr(t) is an increasing function on [0, +∞) for every r ∈ [0, 1). Moreover, both r 7→ ϕr(t − 2) and r 7→ I(t, r), are increasing on [0, 1), for all t ≥ 0. Proof. Indeed, supposing that t ≥ 2, we can write I(t, r) = 1 1 − r2ZT b′ 1 r(ξ) 1 − rξt−2 dm(ξ), where br(z) = r−z in the above integral we get 1−rz ). Using the fact that br ◦ br(z) = z and changing the variable I(t, r) = = = 1 1 − rbr ◦ br(ξ)t−2 dm(ξ) r(ξ) 1 − rbr(ξ)t−2 dm(ξ) 1 1 1 1 1 − r2ZT b′ 1 − r2ZT (1 − r2)t−1 ϕr(t − 2), exp(cid:18) t 1−rz . Now, = 1−r2 2 ln(cid:0)1 + r2 − 2r cos s(cid:1)(cid:19) ds, 0 1−rz since 1 − rbr(z) = 1−rz−r(r−z) ϕr(t) =Z 2π ln(cid:0)1 + r2 + 2r cos s(cid:1) exp(cid:18) t 4Z 2π exp(cid:18) t 4Z 2π (cid:2)ln(cid:0)1 + r2 − 2r cos s(cid:1)(cid:3)2 r (t) = r(t) = and ϕ′′ ϕ′ 1 1 2 0 0 2 ln(cid:0)1 + r2 + 2r cos s(cid:1)(cid:19) ds, ln(cid:0)1 + r2 − 2r cos s(cid:1)(cid:19) ds ≥ 0, A BERNSTEIN-TYPE INEQUALITY FOR RATIONAL FUNCTIONS 7 for every t ≥ 0, r ∈ [0, 1). Thus, ϕr is a convex fonction on [0, ∞) and ϕ′ increasing on [0, ∞) for all r ∈ [0, 1). Moreover, r is Thus, ϕ′ 1 r(0) = 4Z 2π ϕ′ r(t) ≥ ϕ′ 0 ln(cid:0)1 + r2 − 2r cos s(cid:1) ds = 0. r(0) = 0, ∀t ∈ [0, ∞), r ∈ [0, 1), and so ϕr is increasing on [0, ∞). The fact that r 7→ I(t, r), is increasing on [0, 1) for all t ≥ 0 is obvious since 2 I(t, r) =(cid:13)(cid:13)(cid:13)(cid:13) 1 (1 − rz)t/2(cid:13)(cid:13)(cid:13)(cid:13) H 2 =Xk≥0 a2 k(t)r2k, where ak(t) is the kth Taylor coefficient of (1 − z)−t/2. The same reasoning gives that r 7→ ϕr(t) is increasing on [0, 1). (cid:3) Lemma 2.2. If for some r0 ∈ [0, 1) and γ > 0 the function w(ρ) on [r0, 1), then (1−ρ2)γ is increasing Z 1 r ρw(ρ)I(t, rρ)dρ ≍Z 1 r0 ρw(ρ)I(t, rρ)dρ, for all t such that t ≥ γ + 3 and for all r ≥ r0, with constants independent on t. Proof. Clearly, Moreover, r r0 ρw(ρ)I(t, rρ)dρ ≥Z 1 Z 1 ρw(ρ)I(t, rρ)dρ =Z 1 Z 1 ρw(ρ)I(t, rρ)dρ = Z r Z r r0 r0 r and applying Lemma 2.1, ρw(ρ)I(t, rρ)dρ, r ∈ [r0, 1). ≤ ≤ r0 r0 ρw(ρ)I(t, rρ)dρ, ρw(ρ) (1 − ρ2)γ (1 − r2)γ Z r w(r) (1 − r2)γ ϕr2(t)Z 1 ρw(ρ)I(t, rρ)dρ +Z r (cid:0)1 − ρ2(cid:1)γ (1 − (rρ)2)t−1 ϕrρ(t)dρ ρ(cid:0)1 − ρ2(cid:1)γ (1 − (rρ)2)t−1 ϕrρ(t)dρ ρ(cid:0)1 − ρ2(cid:1)γ (1 − (rρ)2)t−1 dρ, ρ(cid:0)1 − ρ2(cid:1)γ (1 − (rρ)2)t−1 ϕrρ(t)dρ ρ(cid:0)1 − ρ2(cid:1)γ (1 − (rρ)2)t−1 dρ. w(ρ) (1 − ρ2)γ (1 − r2)γ ϕr2 (t)Z 1 w(r) w(r) ≥ r0 r0 r because u 7→ ϕu(t) is increasing for all t > 0. For the same reason, Z 1 r ρw(ρ) 1 (1 − (rρ)2)t−1 ϕrρ(t)dρ = Z 1 r 8 ANTON BARANOV AND RACHID ZAROUF Now note that Z r r0 ρ(cid:0)1 − ρ2(cid:1)γ (1 − (rρ)2)t−1 dρ .Z 1 r ρ(cid:0)1 − ρ2(cid:1)γ (1 − (rρ)2)t−1 dρ, r ∈ [r0, 1), with constants independent on t ≥ γ + 3. Indeed, this estimate holds for t = γ + 3, and, hence, by monotonicity of the function ρ 7→ (1 − (ρr)2)−1, for all t ≥ γ + 3. Thus, using Lemma 2.1 and the fact that the function (1−ρ)−γw(ρ) is increasing on [r0, 1), we obtain Z r r0 ρw(ρ)I(t, rρ)dρ ≤ r0 w(r) w(r) (1 − r2)γ ϕr2 (t − 2)Z r (1 − r2)γ ϕr2 (t − 2)Z 1 ρ(cid:0)1 − ρ2(cid:1)γ (1 − (rρ)2)t−1 dρ ρ(cid:0)1 − ρ2(cid:1)γ (1 − (rρ)2)t−1 dρ (1 − (rρ)2)t−1 ϕrρ(t)dρ, ρw(ρ) 1 r ≤ κ1 ≤ κ2Z 1 r (where κ1, κ2 are positive constants which do not depend on t), which completes the proof. (cid:3) Proof of Theorem 1.4. We need to prove only the lower bound, the upper bound is already proved in Theorem 1.3. Let us prove the minoration with the test function f (z) = (1−rz)n . Using (2.1) with α = r0, we need to show that 1 Lp a(w) nprp =Z 1 kf ′kp r0 ρw(ρ)I(pn + p, rρ)dρ C (1 − r)pZ 1 r0 ≥ ρw(ρ)I(pn, rρ)dρ = C (1 − r)p kfkp a(w). Lp Since r ∈ [r0, 1) and n ≥ γ+3 this means that p , by Lemma 2.2 applied with t = pn + p and t = pn Z 1 r ρw(ρ)I(pn + p, rρ)dρ ≥ C (1 − r)pZ 1 r ρw(ρ)I(pn, rρ)dρ. By Lemma 2.1, this is equivalent to the estimate Z 1 r ρw(ρ) ϕrρ(pn + p − 2) (1 − (rρ)2)pn+p−1 dρ ≥ The last statement is obvious since C (1 − r)pZ 1 r ρw(ρ) ϕrρ(pn − 2) (1 − (rρ)2)pn−1 dρ. where bλ(z) = λ−z 1−λz , λ ∈ D. Define the model space KBσ by bλ, Bσ = Yλ∈σ KBσ =(cid:0)BσH 2(cid:1)⊥ and ek(z) =(cid:18)k−1Yj=1 = H 2 ⊖ BσH 2. bλj (z)(cid:19) (1 − λk)1/2 1 − λkz , k ∈ [2, n], Consider the family (ek)1≤k≤n in KBσ (known as Malmquist basis, see [13, p. 117]), A BERNSTEIN-TYPE INEQUALITY FOR RATIONAL FUNCTIONS 9 Z 1 r ρw(ρ) ϕrρ(pn + p − 2) (1 − (rρ)2)pn+p−1 dρ (1 − r2)pZ 1 ≥ 1 r ρw(ρ) ϕrρ(pn + p − 2) (1 − (rρ)2)pn−1 dρ (1 − r2)pZ 1 1 r ϕrρ(pn − 2) (1 − (rρ)2)pn−1 dρ, where the last inequality is due to the fact that t 7→ ϕu(t) is increasing for all 0 ≤ u < 1. (cid:3) ρw(ρ) ≥ 3. The case of super-polynomially decreasing weights. Proof of Theorem 1.5: For the proof of Theorem 1.5 we will need a definition from the theory of model subspaces of the Hardy space. For a finite subset σ of D with card σ = n, consider the finite Blaschke product e1(z) = (1 − λ1)1/2 1 − λ1z The family (ek)1≤k≤n associated with σ is an orthonormal basis of the n-dimensional space KBσ . In what follows we denote by Lp a(w, sD) and by H p(sD), s > 0, the weighted Bergman space and the Hardy space in the disc sD = {z : z < s}, respectively. If w ≡ 1, we write simply Lp Lemma 3.1. Let n ≥ 1, r, s ∈ [0, 1) and p ∈ [1, +∞]. We set a(sD) and we write Lp a if s = 1. Mp, s(n, r) = supnf (ξ) : ξ ∈ D, f ∈ Rn, r, kfkLp a(sD) ≤ 1o. Then (3.1) Mp, 2 3 (n, r) ≤ d cn (1 − r)n+b , where d > 0, b > 0, c > 1 are some absolute positive constants (may be, depending on p). 3 , but for every s ∈ (0, 1), with Remark 3.2. Lemma 3.1 is valid not only for s = 2 constants d > 0, b > 0, c > 1 depending both on s and p. Proof. For every f ∈ Rn, r and ξ ∈ D, we have f 2 (cid:12)(cid:12)(cid:12)(cid:12)f(cid:18) 1 2 ξ(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12)(cid:12)f 2 3(cid:18) 3 4 ξ(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12)(cid:12)ZD (u)(cid:16)k 3 4 ξ(u)(cid:17)2 3 dA(u)(cid:12)(cid:12)(cid:12)(cid:12) , 10 ANTON BARANOV AND RACHID ZAROUF where kλ(z) = 1 is the normalized area measure on D. Applying Holder's inequality we obtain is the standard Cauchy kernel associated with λ ∈ D, and A 1−λz , ξ ∈ T, where p′ is such that 1 (cid:12)(cid:12)(cid:12)(cid:12)f(cid:18) 1 2 3 a a( 2 p + 1 kfkLp p′ = 1. Now, note that D)(cid:13)(cid:13)(cid:13)(cid:13)(cid:16)k 3 4 ξ(cid:17)2(cid:13)(cid:13)(cid:13)(cid:13)Lp′ a(cid:13)(cid:13)(cid:13)(cid:13)(cid:16)k 3 ξ(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) ≤(cid:13)(cid:13)(cid:13)f 2 =(cid:18) 3 2(cid:19)1/p 3(cid:13)(cid:13)(cid:13)Lp (cid:13)(cid:13)(cid:13)(cid:13)(cid:16)k 3 4 ξ(cid:17)2(cid:13)(cid:13)(cid:13)(cid:13)H∞ 4 ξ(cid:17)2(cid:13)(cid:13)(cid:13)(cid:13)Lp′ ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:16)k 3 4(cid:19)2 =(cid:18) 1 1 − 3 2(cid:19)1/p D) ≤ 16(cid:18) 3 D) ≤ 1, we obtain D) ≤ kfkH∞( 1 kfkLp a( 2 a( 1 a 2 2 3 ≤ 24, 4 ξ(cid:17)2(cid:13)(cid:13)(cid:13)(cid:13)Lp′ a = 16. Finally, supposing kfkLp which gives (3.2) Mp, 2 3 (n, r) ≤ 24M2, 1 2 (n, r). It remains to obtain a suitable upper bound for M2, 1 2 (n, r). Let us prove that (3.3) M2, 1 2 (n, r) ≤ 2√n(cid:18) 2 1 − r(cid:19)n+ 1 2 . 2 ∈ Rn, 1 2 r ⊂ Rn, r. If {1/λ1, . . . , 1/λn} is the set of For every f ∈ Rn, r, we have f 1 the poles of f (thus, λj ≤ r, j = 1, . . . , n), then f ∈ KBσ with σ = {λ1, . . . , λn} ⊂ rD), whereas the set {2/λ1, . . . , 2/λn} is the set of the poles of the function f 1 and D. Hence, there exist α1, . . . , αn ∈ C f 1 such that 2 ∈ KBσ′ with σ′ = (cid:8) 1 2 λ1, . . . , 1 2 2 2 λn(cid:9) ⊂ r nXk=1 f 1 = 2 αkek, (3.4) andPn in C. Thus, on D, where (ek)n k=1 is the Malmquist basis associated with the set σ′. Since both f 1 k=1 αkek are meromorphic in C the equality (3.4) is in fact valid everywhere 2 λj f (ξ) = 2 − 2ξ nXk=1 and by the Cauchy -- Schwarz inequality, αk 1 − λj ξ k−1Yj=1 αk2(cid:17)1/2 nXk=1(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:18)k−1Yj=1 f (ξ) ≤(cid:16) nXk=1 4 − ξ(cid:1) 2(cid:0) λ 1 − λ 4 ξ for any ξ ∈ D. Now, if λ ∈ rD and ξ ∈ D, 1 − λ 4 ξ 1 − λξ 2 − 2ξ 1 − λξ (3.5) = λ (cid:16)1 − 1 4 λk2(cid:17)1/2 1 − λjξ , ξ ∈ D, λj 2 − 2ξ 1 − λjξ 1 − λjξ(cid:19)(cid:0)1 − 1 (ξ)(cid:18)1 + 4λk2(cid:1)1/2 4(1 − λξ)(cid:19), 3λ 4 = 2b λ 2!1/2 . (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) A BERNSTEIN-TYPE INEQUALITY FOR RATIONAL FUNCTIONS 11 22(k−1)(cid:18) 1 1 − r(cid:19)2(k−1) 2 1 − r . which gives 2 λ 1 1 λj ≤ (3.6) 3r 4 We get 2 − 2ξ 2 − 2ξ 1 − λjξ (cid:18)k−1Yj=1 nXk=1(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 4λk2(cid:1)1/2 (cid:12)(cid:12)(cid:12)(cid:12) 1 − λξ(cid:12)(cid:12)(cid:12)(cid:12) ≤ 2(cid:18)1 + 1 − λj ξ(cid:19)(cid:0)1 − 1 1 − r(cid:19) = 4 − r 2(1 − r) ≤ (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) nXk=1 4(cid:18) 2 1 − r(cid:19)2n+1 2(cid:13)(cid:13)(cid:13)H 2 =(cid:13)(cid:13)(cid:13)f 1 αk2(cid:17)1/2 For any function ϕ(z) =Pk≥0bϕ(k)zk in H 2, one has √k + 1 bϕ(k) ≤ kϕkL2 (1 − r)2 1 Now we first notice that ≤ . . a kϕkB 1/2 2, 2 , H 2 =Xk≥0 kϕk2 kϕk2 1/2 2, 2 B We now use the upper bound of [21, Theorem A, (4)]: for ϕ ∈ Rn, ρ one has √k + 1 (cid:16) nXk=1 bϕ(k) = kϕ′k2 ≤ (2 + r) ≤ 4n L2 a H 2 + kϕk2 n 1 − rkϕk2 H 2 + kϕk2 H 2 which gives In particular, with ϕ = f 1 (3.7) that is, 2 a . H 2 , we get ϕ ∈ Rn, 1 1 − ρ kϕkL2 2 r and 1 − rkϕk2 kϕkH 2 ≤ 2r n 2(cid:13)(cid:13)(cid:13)H 2 ≤ 2r n a ≤ 2√2n(cid:13)(cid:13)(cid:13)f 1 (cid:13)(cid:13)(cid:13)f 1 2(cid:13)(cid:13)(cid:13)L2 2(cid:13)(cid:13)(cid:13)L2 1 − r/2(cid:13)(cid:13)(cid:13)f 1 2(cid:13)(cid:13)(cid:13)H 2(cid:18) 1 2(cid:18) 2 1 − r(cid:19)n+ 1 (1 − r)2n+1(cid:19) 1 2√2n(cid:13)(cid:13)(cid:13)f 1 2(cid:13)(cid:13)(cid:13)L2 f (ξ) ≤(cid:13)(cid:13)(cid:13)f 1 √2n(cid:18) 2 1 − r(cid:19)n+ 1 f (ξ) ≤ kfkL2 22n+1 ≤ 4 1 . a 2 2 2 , a We conclude from (3.5), (3.6) and (3.7) that for any ξ ∈ D, a( 1 D) , Combining (3.2) and (3.3) and choosing d = 48, b = 1 ξ ∈ D. Taking the supremum over ξ ∈ D and f ∈ Rn, r we obtain (3.3). 2n√n ≤ cn for any n ≥ 1, we complete the proof and obtain (3.1). Proof of Theorem 1.5. Take r ∈ (0, 1) and R ∈ (0, r) and let us represent the norm kf ′kp a(w) of a function f ∈ Rn, r as I1 + I2, Lp 2 and c > 2 such that (cid:3) 2 I1 =Z R 0 k(fρ)′kp pw(ρ)dρ, I2 =Z 1 R k(fρ)′kp pw(ρ)dρ. where the last inequality follows from Lemma 3.1. Note that p, 2/3(n, r)w(ρ)dρ R 3 3 Lp a( 2 M p ∞w(ρ)dρ I2 ≤ C3(cid:16) n ≤ C3kfkp ≤ C3kfkp 1 − r(cid:17)pZ 1 R kfρkp 1 − r(cid:17)pZ 1 D)(cid:16) n 1 − r(cid:17)p D)(cid:16) n kfkp D) ≤ (w(2/3))−1kfkp I2 ≤ C4(cid:16) n 1 − r(cid:17)p cpn cpn a( 2 Lp a( 2 Lp 3 (1 − r)pn+pb w(R), a(w). Lp (1 − r)pn+pb w(R)kfkp a(w). Lp Hence, 12 ANTON BARANOV AND RACHID ZAROUF Here and everywhere below in this proof, Ci, i = 1, . . . , 5, are positive constants, depending, may be, only on p and w (but not on n and r). By (1.2), we have for the first integral I1 ≤ C1(cid:16) n 1 − R(cid:17)pZ R 0 kfρkp pw(ρ)dρ ≤ C2(cid:16) n 1 − R(cid:17)p a(w). Lp kfkp (n, r)kfρkLp D). Ap- Note that fρ ∈ Rn, ρr ⊂ Rn, r, and, thus, kfρk∞ ≤ Mp, 2 plying (1.2) once again together with an obvious inequality kfρkp ≤ kfρk∞, we get a( 2 3 3 Now, choose a positive increasing sequence (γn)n∈N such that n = o (γn), as n → +∞. For any n we fix r◦ n, 1). Now for a fixed n take r, R so that r◦ 1 − R = (1 − r)1/2, n such that the function w(r)(1− r)−γn decreases on [r◦ n = (1 − r)1/4. n < R < r < 1 and 1 − r◦ Hence, using the fact that w is bounded on [r◦ 1, 1), we obtain We have w(R) ≤ w(r◦ n) I2 ≤ C4(cid:16) n 1 − r(cid:17)p kfkp Let us show that for sufficiently large n, cpn (1 − r)γn/4 (1 − R)γn n)γn (1 − r◦ = w(r◦ n)(1 − r)γn/4. Lp a(w) · cpn (1 − r)γn/4 (1 − r)pn+pb . (1 − r)pn+pb → 0, r → 1 − . Indeed, choosing r so that c < (1 − r)−1, we get cpn (1 − r)γn/4 (1 − r)pn+pb ≤ (1 − r) = o(cid:18) 1 I 1/p 2 nkfkLp a(w) 1 − rn(cid:19), n → ∞. since n = o(γn), n → ∞. Hence, there exists a sequence (rn), rn → 1−, such that γn 4 −2pn−pb → 0, r → 1−, The corresponding estimate for I1 is obvious since 1 − Rn = (1 − rn)1/2. (cid:3) A BERNSTEIN-TYPE INEQUALITY FOR RATIONAL FUNCTIONS 13 References 1. A.D. Baranov, Bernstein-type inequalities for shift-coinvariant subspaces and their applications to Carleson embeddings, J. Funct. Anal. 223 (2005), no. 1, 116 -- 146. 2. A.D. Baranov, Embeddings of model subspaces of the Hardy space: compactness and Schatten -- von Neumann ideals, Izvestia RAN Ser. Matem. 73 (2009), no. 6, 3 -- 28; English transl.: Izv. Math. 73 (2009), no. 6, 1077 -- 1100. 3. P. Borwein and T. Erd´elyi, Polynomials and Polynomial Inequalities, Springer, New York, 1995. 4. V.I. Danchenko, An integral estimate for the derivative of a rational function, Izv. Akad. Nauk SSSR Ser. Mat. 43 (1979), no. 2, 277 -- 293; English transl. Math. USSR Izv., 14 (1980), no. 2, 257 -- 273. 5. E.P. Dolzhenko, Rational approximations and boundary properties of analytic func- tions, Mat. Sb. (N.S.), 69(111):4 (1966), 497-524. 6. K.M. Dyakonov, Kernels of Toeplitz operators, smooth functions, and Bernstein-type inequalities, Zap. Nauchn. Semin. S. Peterburg. Otdel. Mat. Inst. Steklov. (POMI) 201 (992), 5 -- 21; English transl.: J. Math. Sci. 78 (1996), 131 -- 141. 7. K.M. Dyakonov, Differentiation in star-invariant subspaces I. Boundedness and com- pactness, J. Funct. Anal. 192 (2002), 364 -- 386. 8. K.M. Dyakonov, Smooth functions in the range of a Hankel operator, Indiana Univ. Math. J. 43 (1994), 805 -- 838. 9. T. Erd´elyi, George Lorentz and inequalities in approximation, Algebra i Analiz 21 (2009), no. 3, 1 -- 57. 10. A.A. Gonchar, Degree of approximation by rational fractions and properties of func- tions, Proc. Internat. Congr. Math. (Moscow, 1966), Mir, Moscow, 1968, 329 -- 356; English transl.: Amer. Math. Soc. Transl. (2) 91 (1970). 11. H. Hedenmalm, B. Korenblum, and K. Zhu, Theory of Bergman spaces, Graduate Texts in Mathematics, 199, Springer-Verlag, New York, 2000. 12. R.J. Leveque, L.N. Trefethen, On the resolvent condition in the Kreiss matrix Theo- rem, BIT 24 (1984), 584 -- 591. 13. N. Nikolski, Treatise on the Shift Operator, Springer-Verlag, Berlin, 1986. 14. A.A. Pekarskii, Inequalities of Bernstein type for derivatives of rational functions, and inverse theorems of rational approximation, Mat. Sb. 124(166) (1984), 571-588; English transl.: Math. USSR-Sb. 52 (1985), no. 2, 557 -- 574. 15. V.V. Peller, Hankel operators of class Sp and their applications (rational approxima- tions, Gaussian processes, the problem of majorizing operators), Mat. Sb. 113(155) (1980), no. 4, 538 -- 581; English transl.: Math. USSR Sb. 41 (1982), no. 4, 443 -- 479. 16. V.N. Rusak, Rational Functions as Approximation Apparatus, Izdat. Beloruss. Gos. Univ., Minsk, 1979 (Russian). 17. M.N. Spijker, On a conjecture by LeVeque and Trefethen related to the Kreiss matrix theorem, BIT 31 (1991), 551 -- 555. 18. R. Zarouf, Sharpening a result by E.B. Davies and B. Simon, C. R. Acad. Sci. Paris 347 (2009), no. 16, 939 -- 942. 19. R. Zarouf, Analogs of the Kreiss resolvent condition for power bounded matrices, Actes des journ´ees du GDR "Analyse Fonctionnelle Harmonique et Applications", Metz, 2010. 20. R. Zarouf, Asymptotic sharpness of a Bernstein-type inequality for rational functions in H 2, Algebra i Analiz 23 (2011), no. 2, 147 -- 161. 21. R. Zarouf, Application of a Bernstein type inequality to rational interpolation in the Dirichlet space, Zap. Nauchn. Semin. S. Peterburg. Otdel. Mat. Inst. Steklov. (POMI) 39:101 -- 112, (2011). 22. R. Zarouf, Effective H∞ interpolation constrained by Hardy and Bergman weighted norms, Ann. Funct. Anal., 2(2):59 -- 74 (2011). 23. R. Zarouf, Effective H∞ interpolation, to appear in Houston J. Math. 24. K. Zhu, Operator theory in function spaces, Monographs and Textbooks in Pure and Applied Mathematics, 139. Marcel Dekker, Inc., New York, 1990. 14 ANTON BARANOV AND RACHID ZAROUF Department of Mathematics and Mechanics, Saint Petersburg State University, 28, Universitetski pr., St. Petersburg, 198504, Russia E-mail address: [email protected] CMI-LATP, UMR 6632, Universit´e de Provence, 39, rue F.-Joliot-Curie, 13453 Mar- seille cedex 13, France E-mail address: [email protected]
1003.2224	1	1003	2010-03-10T23:09:52	A characterization of separable conjugate spaces	[ "math.FA" ]	Necessary and sufficient conditions for a separable Banach space to be a dual space are proved. Some applications are discussed	math.FA	math	A CHARACTERIZATION OF SEPARABLE CONJUGATE BANACH SPACES By STEFANO ROSSI Abstract The following elegant characterization of dual Banach spaces will be proved: a separable Banach space X is a dual space if and only exists a norm closed subspace M ⊂ X∗ with (Dixmier) characteristic equal to 1, whose functionals are norm-attaining. Contents 1 Introduction 2 Notations and preliminaries 3 A general result 4 The separable case 5 Applications 1 Introduction 1 2 2 4 7 Among all Banach spaces, dual spaces have some additional properties, which assure a more far-reaching treatment of their structure. Probably the most useful property is represented by weak-compactness of the unit ball of such spaces. Giving necessary and sufficient conditions for a Banach space X to be a dual space is a long standing problem; it goes back essentially to J. Dixmier [3]. Many theorems concerning this intriguing area are known; these results usually require the compactness property for X1, the unit ball of X, with re- spect to suitable locally convex topology. Elegant as theoretical results, these theorems, however, do not provide any practical tool to solve the question in the applications, because compactness property is very difficult to be checked in infinite-dimensional contexts. In spite of this, reflexive Banach spaces (which are indeed dual spaces with unique predual) are well understood thanks to a celebrate thorem by R. C. James [4], which states that a Banach space X is reflexive if and only if every continuous linear functional attains its norm on X1. The result quoted above can be regarded as a straightforward application of the following powerful characterization of weak compactness, which is due as well to James: 1 Theorem 1 (James, [6]). Let Y be a complete locally convex space and C ⊂ Y a bounded, σ(Y, Y∗) closed subset The following are equivalent: 1. C is a weakly compact subset. 2. Every ϕ ∈ Y∗ attains its sup on C,i.e. for every ϕ ∈ Y∗ there exists c ∈ C such that ϕ(c) = supy∈C ϕ(y). The proof of the theorem announced in the abstract essentially depends on James ’ theorem 1, where completeness assumption unfortunately cannot be removed, as James himself has shown in [5]. 2 Notations and preliminaries If X is a Banach space, X1 will stand for its unit ball, that is X1 = {x ∈ X : kxk ≤ 1} The dual (conjugate) space of X will be denoted by X∗. A Banach space X is a dual space if there exists a Banach space Y such that Y∗ ∼= X (isometric isomorphism); in this case Y is said to be a predual. Accordig to Dixmier, we define the characteristic of a subspace M ⊂ X∗ as the the real number given by c(M) . = inf x6=0(cid:26) sup ϕ∈M1 ϕ(x) kxk (cid:27) When c(M) = 1, we say that M is a determinant subspace of the conjugate space; a standard application of the bipolar theorem shows that M is determi- nant iff M1 is weakly dense in X∗ 1. Finally, we recall that a continuous linear functional ϕ ∈ X∗ 1 is said to be norm- attaining if there exist x ∈ X1 such that ϕ(x) = kϕk. A subspace M ⊂ X∗ is norm-attainig if each ϕ ∈ M is a norm-attaining linear functional. 3 A general result This section is devoted to present a general result for the existence of preduals. In what follows, we will need some tools from the general theory of locally convex space, in particular some basic facts belonging to the theory of dual pair. For an extensive treatment see, for istance, [8]. For completeness’ sake, here we recall that, given a dual pair (X, Y), the Mackey topology on X is the locally convex topology described by the base of seminorms pK given by pK(x) = supy∈Khx, yi where K is a σ(Y, X)-compact, (circled) convex subset of Y. The Mackey topology is often denoted by τ (X, Y). Analogously, one can define the Mackey topology τ (y, X) on Y. 2 We begin our analysis by observing that any conjugate space X∗ is a complete locally convex space with the Mackey topology τ (X∗, X). This can be proved, for instance, with the aid Grothendieck completeness theorem or as an easy application of the lemma 2 from section 4 of the present paper. A second observation needed is that any predual of a Banach space M should be sought as a suitable subspace of the conjugate space X∗. More precisely, one has the following: Lemma 1. Let X and M be Banach spaces. If the isometric isomorphism M∗ ∼= X holds, then there is an isometric injection i : M → X∗ such that i(M) is a determinant subspace, whose functionals are norm-attaining. Proof. Let Φ : X → M∗ be an isometric isomorphism. Taking the (Banach) adjoint of Φ, we get an isometric isomorphism Φ∗ : M∗∗ → X∗. Let i : M → X∗ the linear isometry given by the composition Φ∗ ◦ j, where j is the canonical injection of M into its second conjugate space M∗∗. It only remains to prove that i(M) has the required properties. Let x ∈ M, then we have kxk = kΦ(x)k = sup m∈M1 hΦ(x), mi = sup m∈M1 hi(m), xi The last equality says that i(M) is determinant. In order to prove that i(M) is norm attaining, observe that, given any m ∈ M, one has ki(m)k = kmk = ϕ(m), for some ϕ ∈ M∗ 1 by virtue of the Hanh- Banach theorem. Now, if x be is the unique element of X such that Φ(x) = m, we have ki(m)k = hi(m), xi. This concludes the proof, since kxk = kΦ(x)k = kmk ≤ 1. Under suitable additional requests the previous lemma can be reverted to get existence of a predual, as indicated in the following result (due to Dixmier [3]), which can be regarded as a partial convers of Alaoglu theorem. For the completeness’ sake, we provide a different prove based on the Krein-Smulian theorem. Theorem 2. Let M ⊂ X∗ be a norm closed determinant subspace, such that X1 is σ(X, M)- compact. Then M∗ ∼= X, through the isometric isomorphism Φ : X → M∗ given by hΦ(x), ϕi = ϕ(x) for all ϕ ∈ M∗ and for all x ∈ X. Proof. The linear map Φ : X → M∗ described in the statement is surely isomet- ric, since M is determinant. Let N ⊂ M∗ be its range. N is weakly* dense in M∗, since it is a total subspace, i.e. N⊥ = 0 (in the duality (M∗, M)). To get the conclusion, it is enough to observe that N is weakly* closed by virtue of the Krein-Smulian theorem (here is the point werw completeness of M is needed). In fact N1 = Φ(X1) is weakly* compact as the continuous image of a compact set (Φ is apparently a σ(X, M) − σ(M∗, M) continuous map). Actually the theorem remains true under the milder assumption that M is just a total subspace, see [7], anyway Theorem 2 is enough to state and prove the following general characterization of dual Banach space: Theorem 3. Let X be a Banach space. The following conditions are equivalent: 3 1. X is a dual space. 2. There exists a norm closed subspace M ⊂ X∗ which is determinant, norm- attaining and such that X is complete with respect to the Mackey topology τ (X, M). Moreover, each subspace as in 2. is canonically1 a predual of X . Proof. The implication 1. ⇒ 2. has been already proved. Conversely, if 2. holds, we can equip X with the Mackey topology τ (X, M). By the Mackey- Arens theorem, the Mackey dual of X is M. The unit ball X1 ⊂ X is clearly Mackey-bounded since it is even norm bounded; moreover X1 is σ(X, M)-closed. In fact, thanking to the fact that M is determinant, we can rewrite X1 as {x ∈ X : ϕ(x) ≤ 1 for all ϕ ∈ M1} so it is a σ(X, M)- closed subset as intersection of σ(X, M)-closed sets. By the assumption of completeness, James’theorem applies, so we get σ(X, M)- compactness of X1. Hence M is canonically a predual of X. Remark 1. Unfortunately completeness hypothesi is not omissible, as the fol- lowing counterexample shows. Let us consider the Banach space l1[0, 1] of (real) function f defined on [0, 1] such that Px∈[0,1] f (x) < ∞, endowed with the cor- responding k · k1- norm. Its conjugate space is clearly represented by B[0, 1], the space of all bounded function over [0, 1] with the sup norm, acting by hg, f i = Px∈[0,1] g(x)f (x) for all g ∈ B[0, 1], f ∈ l1[0, 1]. The norm closed subspace of continuous C[0, 1] ⊂ B[0, 1] is determinant and norm-attaining. However it is not a predual, since C[0, 1]∗ ∼= M([0, 1]) (the space of all finite Borel measures on [0,1]), while the inclusion2 l1[0, 1] ⊂ M([0, 1]) is proper. Ac- tually the trouble is that l1[0, 1] is not complete under the Mackey topology τ (l1[0, 1], C[0, 1]), as pointed out in [1]. Although elegant, the previous theorem is not completely satisfactory, be- cause to handle completeness notions in the general context of locally convex spaces is rather difficult; however the result can be considerably strengthened in the separable case, as it is shown in the next section. 4 The separable case In this section we devote our analysis to separable Banach spaces X, the rea- son being that separability assumption allows us to prove the following useful completeness result: Theorem 4. Let X be a separable Banach space and M ⊂ X∗ a determinant subspace. X is a complete locally convex space with respect to the Mackey topology τ (X, M). 1 M∗ ∼= X through the isometric isomorphism Φ : X → M∗ given by the restriction on M of the canonical injection j : X → X∗∗, i.e. hΦ(x), ϕi = ϕ(x) for each ϕ ∈ M. 2One think f ∈ l1[0, 1] as the Borel measure Px∈[0,1] f (x)δx. 4 The proof of theorem 4 relies on the following two lemmas: Lemma 2. Let X be topological vector space. If {fα : α ∈ I} ⊂ X ∗ is a net of continuous linear functionals converging to f uniformly over the compact subset of X, then f is sequentially-continuous. Proof. It is a standard ε 3 -argument. Let {xn : n ∈ N} ⊂ X be a sequence . convergin to x ∈ X. The set K = {xn : n ∈ N} ∪ {x} is compact, thus fα converges to f uniformly on K. This means that, given ε > 0, there exists α0 ∈ I such that fα0 (y) − f (y) < sup y∈K By the triangle inequality we get f (xn) − f (x) ≤ ε ε 3 for all n ≥ Nε, where Nε is any natural number such that fα0(xn)−fα0(x) ≤ ε 3 for all n ≥ Nε. The second lemma deals with the Mazur property of M with respect the σ(M, X)-topology. More precisely, we have the following statement: Lemma 3. Let X be a separable Banach space. If M ⊂ X ∗ is a determinant subspace, then M equipped with the σ(M, X)-topology is a Mazur space, that is every sequantially-continuous linear form f : M → C is continuous with respect to the σ(M, X)-topology. Proof. Since X is separable, the weak* topology on X1 is metrizable as well its restriction to M1. Let f : M → C be a σ(M, X) sequentially-continuous linear form. The restriction f ↾M1 is thus a σ(M, X)- (uniformly) continuous function. Since M is determinant, M1 is weakly* dense in X∗ 1, so f ↾M1 extends to a weakly* continuous function g : X∗ 1 → C, which is apparently the restricton to 1 of a linear form G defined on the whole X∗. By the Krein-Smulian theorem X∗ (see, for instance, [2]), this linear functional is weakly*-continuous, that is there is a unique x ∈ X such that G(ϕ) = ϕ(x) for all ϕ ∈ X∗. This concludes the proof. Finally we can prove theorem 4. Proof. Let {xα : α ∈ I} ⊂ X be a τ (X, M)- Cauchy net. This means that, given any3 σ(M, X)-compact subset K ⊂ M and any ε > 0, we have ϕ(xα − xβ) < ε sup ϕ∈K (1) for all α, β ∈ I such that α, β (cid:23) α0 for some α0 ∈ I. In particular, fixed ϕ ∈ M, {ϕ(xα) : α ∈ I} is a numerical Cauchy net, so we can define a linear form on . = limα ϕ(xα). Thanks to inequality 1, the limit is uniform M, say G, by G(ϕ) 3One should note that if K ⊂ M is σ(M, X)-compact set, then conv(K) σ(M,X) is yet σ(M, X) compact thank to Alaoglu theorem. 5 over the σ(M, X)-compact subset of M. By lemma 2, G is sequentially σ(M, X)- continuous, so it continuous in this topology thanks to the lemma 3. This means that there is a unique x ∈ X such that G(ϕ) = ϕ(x) for all ϕ ∈ M. Taking the limit (with respect to β ∈ I) of the inequality 1, one easily get xα → x in the τ (M, X)-topology. This concludes our proof. Finally, we can state our main theorem, as announced in the abstract: Theorem 5. Let X be a separable Banach space. Every determinant norm closes subspace M ⊂ X∗, whose functional are norm-attaining, is canonically a predual of X. Proof. Theorem 3 of the previous secrion applies, since X is complete with respect the Mackey topology τ (X, M), by virtue of completeness theorem 4. Observation 1. The assumption that M is norm closed cannot be dropped. Let us consider, for instance, the separable Banach space C[0, 1] of continuous function on the compact interval [0, 1]. Let us consider the subspace N ⊂ C[0, 1]∗, algebraically spanned by the set {δx : x ∈ [0, 1]}, where δx is the Dirac measure concentrated on x ∈ [0, 1]. It is not difficult to check that N is a determinant subspace, whose functionals are norm-attaining. Nevertheless N cannot be a predual of C[0, 1], since the last space has not preduals at all, as it is well known. Observe that N is not norm-attaining. The following corollaries are almost immediate. Corollary 1. Let X be a separable Banach space. If M ⊂ X∗ is a norm closed subspace as in theorem 5, then it is a maximal subspace with respect the property to be norm closed and norm-attaining. Proof. If N ⊇ M is norm closed and norm-attaining, then it is (canonically) a predual of X as well as M, hence N = M, since no proper inclusion relationship among preduals are allowed, thanks to the theorems of Alaoglu-Banach and Krein-Smulian. In particular we get the following non-trivial result: Corollary 2. If X is a (separable) Banach space, such that X∗ is separable, then j(X) ⊂ X∗∗ is a closed maximal norm-attaining subspace. As concrete applications of the corollaries stated above, we get, for instance, that c0 ⊂ l∞ is a closed subspace, which is maximal with respect to the property to be norm-attaining (on l1), since c0 is a (not unique) predual of l1. The same is true for K(H) ⊂ B(H), where K(H) is the norm closed ideal of all compact operators on the separable Hilbert space H, thought as a predual of S1, the nuclear (trace-class) operators on H. 6 5 Applications To appreciate theorem 5, we illustrate two nice applications of it, simplifying the proofs of two classical results. The first is the Riesz-Frechet representation theorem for the dual of a (separa- ble) Hilbert space, the second one is the Riesz representation theorem for the dual spaces of the Lebesgue spaces Lp (p > 1) in the separable case. Theorem 6 (Riesz-Frechet). Let H be a separable Hilbert space, with inner product (·, ·). If ϕ ∈ H∗, then there is a unique y ∈ H such that ϕ(x) = (x, y) for each x ∈ H. Proof. There is no loss of generality in the assumption that H is a real Hilber space; in this case the map Φ : H → H∗, given by hΦ(y), xi = (x, y), is lin- ear and isometric. Note that Φ(H) ⊂ H∗ is norm closed, determinant and norm-attaining. This means that Φ(H) is canonically a predual of H, i.e. Φ is surjective. The application to duality theory for Lp spaces is more interesting, since the usual proofs requires rather involved argument invoking Clarkson inequality or Radon-Nikodym theorem. We will consider a measure space (X, S, µ), with the property that S is countably generated: this assures that the corresponding Lebesgue spaces Lp(X, µ) are separable for all p ≥ 1 (for the exception of the value p = ∞ ). Given p > 1, we denote by q > 1 its conjugate exponent, that is the real number q such that 1 p + 1 q = 1. With these notations, the classical Riesz theorem can be stated as follows Theorem 7. If p > 1 is a real number, then the map Φ : Lq(X, µ) → Lp(X, µ)∗ given by hΦ(f ), gi = RX f gdµ for all f ∈ Lq(X, µ), g ∈ Lp(X, µ) , is an isometric Proof. Let Ψ : Lp(X, µ) → Lq(X, µ)∗ be the linear map give by isomorphism . hΨ(f ), gi . = ZX f gdµ for all f ∈ Lp(X, µ), g ∈ Lq(Xµ) By Holder inequality one has kψ(f )k ≤ kf kp for all f ∈ Lp(X, µ). Given f ∈ Lp(X, µ), let us consider the measurable function g given by g(x) = (cid:26) f (x)p−2f (x) 0 if f (x) = 0 p p, it follows that hΨ(f ), hi = kf kp, where h It is easy to check that g ∈ Lq(X, µ) and that kgkq = kf kp . Since one has hΨ(f ), gi = kf kp ∈ Lq(X, µ)1. By virtue of the last equality, we conclude that kΨ(f )k = kf kp and the norm of the functional is attained on the function h. This means that Ψ(Lp(X, µ)) is a closed subspace of Lq(X, µ)∗, whose functionals are norm- attaining. A similar argument as above also shows that Ψ(Lp(X, µ)) is deter- minant. By theorem 5, we argue that Ψ(Lp(X, µ)) is canonically a predual of Lq(X, µ),i.e. the isometric isomorphism Lp(X, µ)∗ ∼= Lq(X, µ) holds as in the statement. p q = kf kp−1 . = g kgkq 7 Remark 2. Our proof of Riesz representation theorem only depend on Holder inequality. I wish to thank S. Doplicher for some useful discussions during the peparation of the present paper. References [1] J. Bonet, B. Cascales, Non complete Mackey topology on Banach spaces, Bull. Australian Matth. Soc [2] , J. B. Conway, A Course in Functional Analysis, Springer, GraduatTexts in Mathematics, 1989 (2nd edition). [3] J. Dixmier, Sur un theoreme de Banach, Duke Math. J. 15, 1057-1071, 1948. [4] R. C. James, Characterization of reflexivity, Studia Math. 23, 205-216, 1964. [5] R. C. James, A counterexample for a sup theorem in normed space, Isreael J. Math 9, 511-512, 1971. [6] R. C. James, Reflexivity and the sup of linear functionals, Israel J. Math 13, 289-300, 1972. [7] S. Kaijser, A note on dual Banach spaces, Math. Scand. 41, 325-330, 1977. [8] A. Wilansky, Modern Methods in Topological Vector Spaces, McGraw-Hill International Book Company, 1978. DIP. MAT. CASTELNUOVO, UNIV. DI ROMA LA SAPIENZA, ROME, ITALY E-mail address: [email protected] 8
1710.00702	1	1710	2017-10-02T14:52:30	p-Riesz bases in quasi shift invariant spaces	[ "math.FA" ]	Let $ 1\leq p< \infty$ and let $\psi\in L^{p}(\R^d)$. We study $p-$Riesz bases of quasi shift invariant spaces $V^p(\psi;Y)$.	math.FA	math	p-Riesz bases in quasi shift invariant spaces Laura De Carli, Pierluigi Vellucci Abstract. Let 1 ≤ p < ∞ and let ψ ∈ Lp(Rd). We study p−Riesz bases of quasi shift invariant spaces V p(ψ; Y ). 1. Introduction Let 1 ≤ p < ∞ and let ψ ∈ Lp(Rd). We consider the shift invariant space V p(ψ) = Span{τkψ}k∈Zd , where τsf (x) = f (x + s) is the translation and "bar" denotes the closure in Lp(Rd). Shift-invariant spaces appear nat- urally in signal theory and in other branches of applied sciences. In [2] [13] and in the recent preprint [14] quasi-shift invariant spaces of functions are considered. Given X = {xj}j∈Zd, a countable and discrete1 subset of Rd and a function ψ ∈ Lp(Rd), we let (1.1) V p(ψ; X) = Span{τxj ψ}. Thus, V p(ψ) = V p(ψ; Zd). Quasi-shift invariant spaces are also called Spline-type spaces in [9], [10] [11], [23]. Following [2], [7], we say that the translates {τxj ψ}j∈Zd form a p-Riesz basis in V p(ψ; X) if there exist constants A, B > 0 such that, for every finite set of coefficients ~d = {dj} ⊂ C, (1.2) Ak ~dkℓp ≤ kXj dj τxj ψkp ≤ Bk ~dkℓp. Here and throughout the paper, we have let kf kp = (cid:0)RRd f (x)pdx(cid:1) 1 k~ckℓp = (Pj∈Zd cjp) p and p . If x = (x1, ..., xd), y = (y1, ..., yd) ∈ Rd, we will 1 2 . We will also let often let x · y = x1y1 + ... + xdyd and x2 = (x · x) x∞ = sup1≤j≤d xj. 1 1 A countable set X ⊂ Rd is discrete if for every xj ∈ X there exists δj > 0 such that xj − xk2 > δj for every k 6= j. 1 2 LAURA DE CARLI, PIERLUIGI VELLUCCI If (1.2) holds, then it is possible to prove that (1.3) 2V p(ψ; X) = {f = Xk∈Zd dkτxk ψ(x), ~d ∈ ℓp } and the sequence {dk}k∈Zd is uniquely determined by f . p−Riesz bases allow a stable reconstruction of functions in V p(ψ; X); when X = Zd and B = {τj ψ}j∈Zd is a p−Riesz basis of V p(ψ), the coefficient dj in (1.3) can be expressed in an unique way in terms of the functions in the dual basis of B. See [30], [5] and also [1] for explicit reconstruction formulas. for every y ∈ [− 1 2 , 1 When ψ has compact support, it is known (see e.g. [1, Prop. 1.1], [16], [24]) that B is a p−Riesz basis in V p(ψ) if and only ifPm∈Zd ψ(y +m)2 6= 0 We have denoted with ψ(y) =RRd e2πix·yf (x)dx the Fourier transform of 2 )d and every m ∈ Zd. See also Lemma 5 in Section 2. ψ. The proof of the aforementioned result relies on the lattice structure of Zd and on standard Fourier analysis technique and does not easily generalize to other sets of translations. Let ψ ∈ Lp(Rd), 1 ≤ p < ∞, and let X = {xj}j∈Zd be a discrete set of Rd. It is natural to consider the following problem: Problem 1. Let BX = {τxj ψ}j∈Zd be a p−Riesz basis for V p(ψ; X); can we find δ > 0 so that, for every Y = {yj}j∈Zd ⊂ Rd with supj yj − xj2 < δ, the set BY = {τyj ψ}j∈Zd is a p−Riesz basis for V p(ψ; X)? This problem cannot be solved if ψ has compact support. For example, let ψ(x) = rect(x) be the characteristic function of the interval [− 1 2 ) and let X = Z; let Y = {yn}n∈Z be such that y0 = δ > 0 and yn = n when n 6= 0. All functions in V p(rect; Y ) vanish in the interval [− 1 2 + δ] and so V p(rect; Y ) 6= V p(rect). 2 , − 1 2 , 1 We prove in Section 3 that Problem 1 can be solved when p = 2 and ψ is band-limited, i.e., when the Fourier transform of ψ has compact support. See also Section 5 for more remarks and comments on problem 1 We are concerned with the following problem: Problem 2. With the notation of Problem 1: let BX be a p−Riesz basis for V p(ψ; X) and let Y = {yn}n∈Zd that satisfies supn yn − xn2 < δ; is BY a p−Riesz basis for V p(ψ; Y ) whenever δ is sufficiently small? It is proved in [10] that Problem 2 has always solution when X is rel- atively separated, i.e., when X = X1 ∪ ... ∪ Xk, with Xj = {xj,n}n∈Zd and inf n6=m xj,n − xj,m2 > 0. In Section 2 we prove the following theorem. Theorem 1. Suppose that that ψ is in the Sobolev space W 1,p(Rd), with 1 < p < ∞, and that {τxj ψ}j∈Zd is a p−Riesz basis of V p(ψ; X). For 2 A proof of this ientity was kindly provided to us by K. Hamm P-RIESZ BASES IN QUASI SHIFT INVARIANT SPACES 3 every j ∈ Zd there exists δj > 0 such that {τyj ψ}j∈Zd is a p−Riesz basis of V p(ψ; Y ) whenever xj − yj2 < δj. We recall that W 1,p(D) is the space of Lp(D) functions whose partial (D) is the closure distributional derivatives are also in Lp(D) and that W 1,p of C ∞ 0 (D) in W 1,p(D). When X is not relatively separated the δj's in Theorem 1 may not have a positive lower bound, but we can still solve Problem 2 in the cases considered in Theorems 2 and 3 below. 0 Theorem 2. Assume that ψ ∈ L1(Rd) ∩ L2(Rd) satisfies (1.4) 0 < c = Xk∈Zd inf x∈[0,1)d ψ(x + k)2 < Xk∈Zd sup ψ(x + k)2 = C < ∞. x∈[0,1)d Then, Problem 2 can be solved when p = 2 and {e2πixn·x}n∈Zd is a Riesz basis in L2([0, 1)d). We recall that measurable functions f (cid:16)Pn∈Zd supx∈[0,1)d f (x + n)q 2(cid:17) 1 be defined in a similar manner. the amalgam space W (L∞, ℓq) set of : Rd → C for which f W (L∞, ℓq) = the is q < ∞. The amalgam space W (Lr, ℓq) can The assumption (1.4) implies that ψ is in the amalgam space W (L∞, ℓ2). From Theorem 2 follows that if ψ satisfies (1.4), then {τyk ψ}k∈Zd is a 4 . See the remark after 2−Riesz basis of V 2(ψ, Y ) whenever k − yk∞ < 1 the proof of Theorem 2 in Section 4. Exponential Riesz bases in L2(0, 1) are completely understood and clas- sified [20]. To the best of our knowledge, no such characterization exists for exponential bases on L2((0, 1)d) when d > 1. For our next theorem we consider ψ in the Sobolev space W 1,p denote with ∂j ψ = ∂ψ ∂xj we let ∇ψ = (∂1ψ, ..., ∂dψ) be the gradient of ψ. Let Y = {yk}k∈Zd and (Rd); we the partial derivative (in distribution sense) of ψ and 0 L = sup k∈Zd yk − k2 < ∞. We prove the following Theorem 3. Let D = (a1, b1) × ... × (ad, bd) be a bounded rectangle in (D), with 1 ≤ p < ∞, and let {τkψ}k∈Zd be a p−Riesz Rd. Let ψ ∈ W 1,p basis of V p(ψ) with frame constants 0 < A ≤ B < ∞. If 0 (1.5) C = L dXj=1 (1 + [bj − aj + L])p−1k∂j ψkp p < A, the set {τyk ψ}k∈Zd is a p−Riesz basis of V p(ψ; Y ) with constants B + C and A − C. 4 LAURA DE CARLI, PIERLUIGI VELLUCCI The proofs of Theorems 2 and 3 are in Section 3. Our Theorem 3 can be compared to [10, Theorem 3.5]. In this theorem it is assumed that ∇(ψ) is in the amalgam space W (L∞, ℓ1), and that inf x∈[0,1)dPk∈Zd ∇ψ(x + k)2 > 0. In the aforementioned theorem is proved that {τyk ψ}k∈Zd is a Riesz basis of V 2(ψ; Y ) if C ′ = L2(1 + 2L)2d∇ψ2 W (L∞, ℓ1) < A. Generalizations to functions for which ∇(ψ) is in the amalgam space W (Lq, ℓ1), with q > d are also possible (see Remark 3.2 in [10]). Our Theorem 3 reduces to [10, Theorem 3.5] when p > d and ψ has compact support. For example, when ψ has support in [0, 1)d, the norm in W (Lp, ℓ1) reduces to ∇ψp. The constant C in Theorem 3 may be smaller than C ′ defined above when the support of ψ is small. Theorem 3 does not apply when ψ = rect or when ψ is a step function; For J ≥ 1, we let SJ = ns(t) =Pj≤J sj rect(t − j), sj ∈ Ro . We let p−1 be the dual exponent of p. The following theorem is proved in p′ = p Section 4.2. Theorem 4. Assume that g ∈ SJ and that {τkg}k∈Z is a p−Riesz basis for V p(g), with frame constants A and B. If the sequence {τyk g}k∈Z is a Riesz basis for V p(g; Y ). 2pLJ kgkp p′ < A , Acknowledgement. We are grateful to the anonymous referee of this paper for her/his thorough reading of our manuscript and for providing suggestions that have improved the quality of our work. We also wish to thank K. Hamm for providing a proof of the identity (1.3) for p 6= 2. 2. Preliminaries 2.1. Notation. We denote with hf, gi = RRd f (x)¯g(x)dx and kf k2 = phf, f i the standard inner product and norm in L2(Rd). For a given p ∈ Rd and δ > 0, we let B(p, δ) = {x ∈ Rd : x − p2 < δ}. 2 , 1 We let rect(x) = χ[− 1 2 )(x) be the characteristic function of the interval 2 , 1 [− 1 2 ) and βs = rect(s+1)(x) = rect ∗... ∗ rect(x) be the s + 1−times iterated convolution of rect. The function βs(x), a piecewise polynomial function of degree s, is a B-spline of order s. See [27], where the B-splines were first introduced, and [21], [32] and the references cited there. 2.2. p− Riesz bases. Recall that a Schauder basis in a separable Banach space V is a linearly independent set B = {vj}j∈Z such that: span(B) = V , and there exists a sequence of bounded linear functions fj : X → C (the functional coefficients of the basis) such that x =Pj fj(x)vj for every x ∈ V . P-RIESZ BASES IN QUASI SHIFT INVARIANT SPACES 5 Following [6], [7] and [2], we say that B is a p−Riesz basis of V , with 1 ≤ p < ∞, if Span(B) = V , if every series Pn anvn converges in V when ~a = (an)n∈Z ∈ ℓp and if there exist constants A, B > 0 such that, for every finite sequence of coefficients {dj }j∈Z ⊂ C, we have Ak ~dkℓp ≤ kXj dj vjkp ≤ Bk ~dkℓp. Every p−Riesz basis is a Schauder basis. As mentioned in the introduction, when V = V p(ψ) and ψ has compact support, the functional coefficients of the basis {τkψ}k∈Z of V p(ψ) can be written in terms of the dual functions of the basis. The following results are well known (see e.g. [1, Prop. 1.1], [16], [24]). Lemma 5. a) Let ψ ∈ Lp 0(Rd). The set B = {τkψ}k∈Zd is a p−Riesz basis in V p(ψ) if and only if (2.1) ψ(y + m)2 6= 0 for every y ∈ [− 1 2 , 1 2 )d. Xm∈Zd b) If ψ ∈ W (L∞, ℓ1) is continuous and if B is Riesz basis in V 2(ψ), then B is a p-Riesz basis in V p(ψ) for every 1 ≤ p < ∞. Proof. For the convenience of the reader we prove that if ψ ∈ L2 0(Rd), B is a Riesz basis of V 2(ψ) with constants 0 < A ≤ B < ∞ if and only if the following inequality holds for every y ∈ Q = [− 1 2 , 1 2 )d. (2.2) A ≤ Xm∈Zd ψ(y + m)2 ≤ B. We can verify (using e.g. the Poisson summation formula) that the function in (2.1) is continuous in Q, and so (2.2) is equivalent to (2.1). Let {ck}k∈Zd ⊂ C be a finite set of coefficients such that Pk ck2 = 1. The Fourier transform of f =Pk∈Zd ckτkψ is cke2πiy·k = ψ(y)M (y). f (y) = ψ(y) Xk∈Zd and by Plancherel's theorem kf k2 2 = k f k2 2 = Xm∈Zd M (y)2 Xm∈Zd Zm+[− 1 , 1 2 )d ψ(y + m)2dy. 2 Xm∈Zd RQ M (y)2g(y)dy andRQ M (y)2dy =Pk ck2 = 1 follows that A ≤ f 2 2 ≤ BPk ck2 and the above consid- ψ(y + m)2M (y)2dy =ZQ ZQ Let g = Pm∈Zd ψ(y + m)2; Conversely, from APk ck2 ≤ kf k2 erations, follows that 2 = 2 ≤ from f 2 (2.2) holds, B. if f (y)2dy = (2.3) AkM 2kL1(Q) ≤ZQ M (y)2g(y)dy ≤ BkM 2kL1(Q). 6 LAURA DE CARLI, PIERLUIGI VELLUCCI Every non-negative h ∈ L1(Q) can be written as h = M 2, with M ∈ L2(Q). The dual of L1(Q) is L∞(Q) and so kgkL∞(Q) = supkhkL1(Q)=1RQ f (y)g(y)dy. From (2.3) follows that A ≤ kgkL∞(Q) ≤ B as required. We will use the following Paley-Wiener type result. (cid:3) Lemma 6. Let X, Y ⊂ Rd be countable and discrete. Suppose that {τxj ψ}j∈Zd is a p−Riesz basis of V p(ψ; X) with constants A ≤ B. If the inequality Xj (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) aj(τxj ψ − τyj ψ)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) p p ≤ CXn anp holds for all finite sequences {an}n∈Zd ∈ C with a constant C < A, the sequence {τyj ψ}j∈Zd is a p−Riesz basis of V p(ψ; Y ) with constants B + C and A − C. Xj Proof. Assume that Pn anp = 1; we have: (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) aj τyj ψ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)p aj(τxj ψ − τyj ψ)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)p +(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) aj τxj ψ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)p Xj Xj aj(τxj ψ − τyj ψ)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)p ajτxj ψ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)p ajτyj ψ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)p −(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xj Xj ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≥(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xj and ≤ C + B ≥ A − C. (cid:3) p′ = p Proof of Theorem 1 . Assume p ∈ (1, ∞) and Pj ajp = 1. Let constants such that Pj δj p′ p−1 is the dual exponent of p and let {δj}j∈Zd be a sequence of positive < ∞, We recall that, when 1 < p < ∞, a function f ∈ Lp(Rd) is in the Sobolev space W 1,p(Rd) if and only if there is a constant c > 0 that depends on f but not on δ, such that (2.4) ωp(δ, f ) = sup t<δ τtf − f p ≤ cδ for every δ > 0. Furthermore, one can choose c = k∇f kp. See e.g. Prop. 9.3 in [4]. By (2.4) and Holder' s inequality, (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) aj(τxj ψ − τyj ψ)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xj ≤Xj p  ajp ≤ c Xj Xj  1 aj kτxj ψ − τyj ψkp ≤Xj δj p′ δjp′   = c Xj 1 p′ 1 p′ . aj δj P-RIESZ BASES IN QUASI SHIFT INVARIANT SPACES 7 We can chose the δj so small that c(cid:16)Pj δjp′(cid:17) 1 complete the proof. p′ < A and use Lemma 6 to (cid:3) 3. Problem 1 (p = 2) In this section we prove that Problem 1 can be solved when p = 2 and ψ has compact support. Theorem 7. Let ψ ∈ L2(Rd). Assume that ψ has compact support and that there exist constants c, C > 0 such that c ≤ inf x∈Rd ψ(x) ≤ sup x∈Rd ψ(x) ≤ C. Let {τxj ψ}j∈Zd be a Riesz basis in V 2(ψ, X). There exists δ > 0 such that if Y = {yj}j∈Zd ⊂ Rd satisfies supj xj − yj < δ, then also {τxj ψ}j∈Zd is a Riesz basis of V 2(ψ). Proof. Let D = supp( ψ). When p = 2, Plancherel theorem implies that the set {τxj ψ}j∈Zd is a Riesz basis in V 2(ψ) if and only if the set {e2πixj ·x}j∈Zd is a Riesz basis on L2(Rd, ψ dx). Our assumptions on ψ imply that the norm on L2(Rd, ψ dx) is equivalent to the norm on L2(D) and that {e2πixj ·x}j∈Zd is an exponential Riesz basis on L2(D). Exponential Riesz bases on bounded domains of Rd are stable under small perturbations (see [22] and also Section 2.3 in [18]); we can find δ > 0 such that, if Y = {yj}j∈Zd ⊂ Rd satisfies supj xj − yj < δ, then also {e2πiyj ·x}j∈Zd is a Riesz basis on L2(D) and hence also in L2(Rd, ψ dx). (cid:3) Example. Let d = 1 and let ψ = sinc(x) = sin(πx) ; the Fourier transform of τkψ(x) = sinc(x − k) is e2πikx rect(x) = e2πikxχ[− 1 2 )(x), and so V 2(ψ) 2 , 1 is isometrically isomorphic to Span{e2πijx rect(x)}j∈Zd = L2(− 1 2 ). By Kadec's theorem ([17], [33]) if Y = {yn}n∈Zd ⊂ R is such that supn yn − n ≤ δ < 1 2 ) and therefore, the set {sinc(x − yn)}n∈Zd is a Riesz basis for V 2(sinc). Thus, V 2(sinc; Y ) = V 2(sinc). 4 , the set {e2πiynx}n∈Zd is still a Riesz basis of L2(− 1 2 , 1 πx 2 , 1 Things are not so clear when p 6= 2. For example, the trigonometric system B = {e2πinx}n∈Zd is a Schauder basis in Lp(− 1 2 ) when 1 < p < ∞, but it is not a p−Riesz basis and the previous example cannot be generalized 2 , 1 in an obvious way. Stability results for the Schauder basis B in Lp(− 1 2 ) are proved in [25] and in [26]. 2 , 1 4. Problem 2 In this section we prove Theorems 2 and 3. Let X = {xn}n∈Zd and B = {e2πix·xn}n∈Zd. We first prove the following 8 LAURA DE CARLI, PIERLUIGI VELLUCCI Lemma 8. Let ψ ∈ L2(Rd) ∩ L1(Rd) be as in (1.4); if B is a Riesz basis in L2([0, 1)d) with constants A1 and B1 then {τxn ψ} is a Riesz basis of V 2(ψ, X)} with constants A = A1c and B = B1C. Proof. For k ∈ Zd, we let ck = inf x∈(0,1]d ψ(x + k)2 and Ck = supx∈(0,1]d ψ(x + k)2. Let {dj} be a finite set of complex coefficient such that Pj dj2 = 1. Since B is a Riesz basis in L2((0, 1]d), for every given k ∈ Zd we have that ≤ B1. 2 L2((0,1]d) From this inequality follows at once that A1 ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ckA1 ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xn A1c ≤ Xk∈Zd dne−2πixn·k e2πixn·y(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xn dne−2πixn·ke2πixn·y ψ(. − k)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) With c =Pk∈Zd ck and C =Pk∈Zd Ck = ψ2 dne2πixn·(.−k) ψ(. − k)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xn In view of Pk∈Zd kg(. − k)kL2((0,1]d) = g2, we obtain dne2πixn·y ψ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) A1c ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xn ≤ B1C. 2 2 2 2 ≤ CkB1. L2((0,1]d) W (L∞, ℓ2), we have ≤ B1C. L2((0,1]d) By Plancherel's theorem, the latter is equivalent to A1c ≤ kPn dnτxnψk2 ≤ B1C and so {τxk ψ}k∈Zd is a Riesz basis of V 2(ψ, X), as required. (cid:3) Proof of Theorem 2. Let B = {e2πix·xn}n∈Zd be a Riesz basis in L2([0, 1)d); it is proved in [22] (see also Section 2.3 in [18]) that we can find δ > 0 such that, if Y = {yj}j∈Zd ⊂ Rd satisfies supj xj − yj2 < δ, then also {e2πiyj ·x}j∈Zd is a Riesz basis in L2([0, 1)d). By Lemma 8, {τynψ} is a Riesz basis of V 2(ψ, Y ). (cid:3) Remark. When Y = {yk}k∈Zd is such that supk∈Zd k − yk∞ < 1 4 , by the multi-dimensional generalization of Kadec's theorem proved in [31] we have that {e2πiyj ·x}j∈Zd is a Riesz basis in L2([0, 1)d) and by Lemma 8, {τyn ψ}n∈Zd is a Riesz basis of V 2(ψ, Y )}. 4.1. Proof of Theorem 3. In order to prove Theorem 3 we need some preliminary result: first, we prove the following P-RIESZ BASES IN QUASI SHIFT INVARIANT SPACES 9 Lemma 9. Let (a, b) ⊂ R, with a < b < ∞, and let 1 ≤ p < ∞. Let ψ ∈ Lp 0(a, b). For every finite set of coefficients {αj } ⊂ C, we have that Xk (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) αkτkψ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) p p ≤ kψkp p([b − a] + 1)p−1Xk αkp where [ ] denotes the integer part. Proof. For simplicity we let a = 0. When b ≤ 1 the supports of the τkψ's are disjoint and so kf kp b > 1 the supports of the τkψ overlap, and there are at most [b] + 1 of such supports that intersect at each point. By the elementary inequality (x1 + · · · + xm)p ≤ mp−1 (xp m) (which is valid when the xj are non-negative) we have that 1 + · · · + xp pPk αkp. When p = kPk αkτkψkp p = kψkp f (t)p = Xk akτkψ(t)p ≤ ([b] + 1)p−1Xk pPk αkp as required. akpτkψ(t)p and so kf kp p ≤ ([b] + 1)p−1kψkp (cid:3) Let Y = {yk}k∈Zd be a discrete subset of Rd. Assume that L = ≤ L(1 + [b − a + L])p−1kψ′kp p where ψ′(t) denotes the distributional derivative of ψ. Assume first that k ψ′(t + x)dx and that finite set of coefficients {αj} ⊂ C such that Pk αkp = 1, we have that p 0 (D). Then, for every (1 + [bj − aj + L])p−1k∂j ψkp p. supk∈Zd yk − k2 < ∞. We prove the following p p p ≤ L (4.2) (4.1) dXj=1 Proof. When d = 1 and D = (a, b), we prove that j=1[aj , bj] and let ψ ∈ W 1,p Lemma 10. Let D =Qd (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) αk(τkψ − τyk ψ)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xk αk(τkψ − τykψ)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xk (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) αk(τkψ − τyk ψ)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xk ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) αkZ t+k+L Xk where we have let g(t) = R t+L yk > k. Observing that ψ(t + yk) − ψ(t + k) = R yk αkZ t+yk =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xk =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xk ψ′(x)dx(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) k − yk ≤ L, we have that ψ′(x)dx. t+k t+k p p p p p p ψ′(x)dx(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) αkτkg(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) p p It is easy to verify that g(t) is supported in the interval [a − L, b]. Indeed, ψ′ is supported in [a, b] and so g(t) ≡ 0 whenever t + [0, L] ∩ [a, b] = ∅. Thus, g(t) ≡ 0 when t + L < a or t > b, or: g(t) ≡ 0 when t ∈ R − [a − L, b], as required. t 10 LAURA DE CARLI, PIERLUIGI VELLUCCI We apply a change of variables and Minkowsky's integral inequality; we gather ≤ (1 + [b − a + L])p−1kgkp p . By Lemma 9 (4.3) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xk αk(τkψ − τyk ψ)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) kgkp =(cid:13)(cid:13)(cid:13)(cid:13) Z t+L t p p p p ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) αkτkg(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xk ψ′(x)dx(cid:13)(cid:13)(cid:13)(cid:13)p =(cid:13)(cid:13)(cid:13)(cid:13) Z L ≤ Lkψ′kp 0 ψ′(x + t)dx(cid:13)(cid:13)(cid:13)(cid:13)p t [a, b + L]. When d = 2 we can let yk = (yk,1, yk,2) and k = (k1, k2) and write which together with the inequality (4.3) concludes the proof of (4.2). When yk < k the proof if similar, but the function g(t) defined above should ψ′(x)dx, a function supported in the interval be replaced by g(t) = R t−L (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) αk(τkψ − τyk ψ)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)p Xk +(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) αk(τ(k1,k2)ψ − τ(yk,1,k2)ψ)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)p αk(τ(yk,1,k2)ψ − τ(yk,1,yk,2)ψ)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)p ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xk Xk αk(τk2 ψ2 − τyk,2ψ2)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)p +(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) αk(τk1ψ1 − τyk,1ψ1)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)p =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xk Xk where we have let ψ1 = τ(0,k2)ψ and ψ2 = τ(yk1 applied to ψ1 and ψ2, yields (4.1). The case d > 2 is similar. ,0)ψ. The inequality (4.2), Proof of Theorem 3. Follows from Lemmas 6 and 10. (cid:3) (cid:3) 4.2. rect and step functions. Since Sobolev spaces W 1,p(R) do not contain discontinuous functions, we cannot apply Theorem 3 when ψ is a step function. Let ψ = rect; it is immediate to verify that, for every 1 ≤ p < ∞, the set {τj rect}j∈Z is a p−Riesz basis of V p(rect) with frame constants A = B = 1. Throughout this section we let Y = {yk}k∈Z ⊂ R, with L = supk∈Zd yk − k and we assume 1 ≤ p < ∞. Lemma 11 below is an easy generalization of Lemma 10 in [8]. Lemma 11. Assume 0 ≤ L < 1. For every finite set of coefficients {αk}n∈Zd ⊂ C we have that (4.4) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xk αk(rect(t − k) − rect(t − yk))(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) p p < 2pLXk αkp. P-RIESZ BASES IN QUASI SHIFT INVARIANT SPACES 11 Proof. Assume Pk αkp = 1. Let (4.5) f (t) =Xk αk (rect(t − k) − rect(t − yk)) =Xk αkχIk, where Ij denotes the support of rect(t − j) − rect(t − yj). When yj 6= j, Ij is union of two intervals that we denote with I + j . When yj > j, we let j and I − I − j = (j − 1 2 , yj − 1 2 ), I + j = (j + 1 2 , yj + 1 2 ). We use (improperly) the same notation to denote I + yj < j. j and I − j also when h there is at most another interval I ± Since we have assumed yh − h ≤ L < 1, for every given interval J = I ± k that overlap with J; thus, for every t ∈ R, the sum in (4.5) has at most 2 terms. By the ele- mentary inequality (x1 + · · · + xm)p ≤ mp−1 (xp m) we have that p ≤ 2p−1 supk Ik = 2p−1(2L) = 2p L (cid:3) f (t)p ≤ 2p−1Pk αkpχIk (t), and kf kp and the proof of the Lemma is concluded. 1 + · · · + xp Lemma 11 and Lemma 6 yield the following Theorem 12. With the notation of Lemma 11, the set {τyk rect}k∈Z is a p−Riesz basis in V p(rect; Y ) if 2pL < 1. Corollary 13. Let ψ0 ∈ L1(R) and let ψ = rect ∗ψ0. Suppose that {τkψ}k∈Z is a p−Riesz basis for V p(ψ). For every finite set of coefficients {αk}n∈Z ⊂ C with Pk αkp = 1, we have that αk(ψ(t − k) − ψ(t − yk))(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xk p p < 2pLkψ0kp 1 and the set {ψ(t − yk)}k∈Z is a p−Riesz basis for for V p(ψ; Y ) whenever (4.6) 2pLkψ0kp 1 < A. Remark. If ψ0(x) 6= 0 for every x ∈ R, then the set {τkψ}k∈Z is a p−Riesz basis for V p(ψ). 0 whenever x ∈ [− 1 Indeed, Pm∈Z drect(y + m)2 = Pm∈Z sinc(x + k)2 6= 2 ) and so also Pm∈Z ψ(x + k)2 = Pm∈Z ψ0(x + k)drect(x + k)2 6= 0; by Lemma 5 the set {τkψ}k∈Z is a p−Riesz basis for V p(ψ). 2 , 1 proof of Corollary 13. Let αk (ψ(t − k) − ψ(t − yk)) , F (t) =Xk f (y) =Xk∈Z αk (rect(y − k)−rect(y − xk) 12 LAURA DE CARLI, PIERLUIGI VELLUCCI and we show that kF kp p ≤ 2pLkψ0kp 1. We gather F (t) =Z ∞ −∞ ψ0(t − y)Xk∈Z αk (rect(y − k)−rect(y − xk)) dy = ψ0 ∗ f (t). Thus, by Young's inequality and Lemma 11, kF kp p ≤ kψ0kp 1kf kp p ≤ 2pLkψ0kp 1 and the proof of the corollary is concluded. (cid:3) Let βm(x) = rect(m+1) be the B-spline of order m > 1. We recall that βm is supported in the interval [− m+1 2 ] and βm(x) ∈ W 1,p(R) whenever m ≥ 1. It is easy to verify by induction on m that kβmkp ≤ 1 and mkp ≤ 2. It is known that {τkβm}k∈Z is a Riesz basis of V 2(βm) whose kβ′ Riesz constants A(m) and B(m) are explicitly evaluated in [19]. See also [28]. By the observations after Lemma 5, {τkβm}k∈Z is a p−Riesz basis of V p(βm) with constants Ap(m) > 0 and Bp(m) < ∞. 2 , m+1 We prove the following Corollary 14. Assume that L < 2−pAp(m). Then, the set {τyk βm}k∈Z is a p−Riesz basis of V p(ψ, Y ). Proof. We apply Corollary 13 with ψ0 = βm−1. (cid:3) Remark. We could have used Theorem 3 to prove Corollary 14, but we would have obtained a lower upper bound for L (namely, L < Ap(m) 2(2+m)p−1 ). proof of Theorem 4. Let g(t) = Pj≤J sj rect(t − j). {αk}n∈Zd ⊂ C be a finite set of coefficients such that Pk αkp = 1. Let Let f (t) =Xk αk (g (t − k) − g (t − xk)) . As in previous theorems, we find conditions on L for which kf kp p < A. We have f (t) = Xj≤J = Xj≤J sjXk αk (rect(t − j − k) − rect(t − j − xk)) sjfj(t). p dt ZR αk (rect(t − j − k) − rect(t − j − xk))(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) fj(t)pdt =ZR Xk (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) αk(cid:0)rect(t′ − k) − rect(t′ − xk)(cid:1)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) =ZR Xk αk(rect(t − k) − rect(t − yk)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xk dt′ p p p . P-RIESZ BASES IN QUASI SHIFT INVARIANT SPACES 13 By Minkowski and Holder inequalities, and noting that Pj≤J sjq = kgkq it follows that q, kf kp ≤ Xj≤J (4.7) sjkfjkp ≤ sjp′ Xj≤J  = kgkp′  Xj≤J 1 p′  Xj≤J p  1 p kfjkp , 1 p kfjkp p  With the change of variables t − j = t′ in the integral below, we obtain From Lemma 11, follows that the integral above is ≤ 2pL. We gather: kf kp p ≤ 2pLJ kgkp p′ < A, and by Lemma 6 Theorem 4 follows. (cid:3) p′ . By assumption 2pLJ kgkp 5. Remarks and open problems We have discussed Problem 1 when p = 2 and the Fourier transform of the window function ψ has compact support. When ψ is not band-limited, Plancherel's theorem implies that the set {τxj ψ}j∈Zd is a Riesz basis in V 2(ψ, X) if and only if the set V = {e2πixj ·x ψ}j∈Zd is a Riesz sequence in L2(Rd), and hence a Riesz basis in V = Span(V)). By a theorem of [33, Theorem 11]) for every j ∈ Zd there Krein-Milman-Rutman (see e.g. exists ǫj > 0 such that every set of functions {gj }j∈Zd ⊂ V is a Riesz basis of V whenever gj − eπixj ·x ψ2 < ǫj. We can find δj > 0 such that (eπixj ·x − eπiyj·x) ψ2 < ǫj whenever xj − yj2 < δj, but we do not know whether the δj's have a lower bound or not. For functions ψ in Lp(Rd) for every p ∈ [1, ∞) it would be interesting to prove conditions that would ensure that a q-Riesz basis in V q(ψ, X) for some q is automatically a p-Riesz basis in V p(ψ, X) for all p. Lemma 5 (b) shows that, for certain ψ, if the set {τkψ}k∈Zd is a 2-Riesz basis of V 2(ψ), it is also a p-Riesz in V p(ψ) but the method of proof of this result does not generalize well to other sets of translations. Results in [3] and [29] may help generalize Lemma 5. 14 LAURA DE CARLI, PIERLUIGI VELLUCCI It would also be interesting to define and investigate p-Riesz bases in quasi-shift invariant spaces V p(ψ, X) when 0 < p < 1. Wavelet in Lp with 0 < p < 1 have been considered in [12]. We feel that the results contained in Section 3 of [12] may help the understanding of V p(ψ, X) when 0 < p < 1. References [1] A. Aldroubi, Q. Sun, Connection between p−frames and p−Riesz bases in locally finite SIS of Lp(R) Proceedings of SPIE - The International Society for Optical Engineering, February 1970 [2] A. Aldroubi, Q. Sun, W. Tang, p-Frames and shift invariant subspaces ofLp , J. Fourier Anal. Appl. 7 (2001) 1-21. [3] A. Aldroubi, A. Baskakov, I. Krishtal, Slanted matrices, Banach frames, and sampling, J. Funct. Anal. 255 (2008) 1667 -- 1691. [4] H. Brezis, Functional Analysis, Sobolev Spaces and Partial Differential Equations, Springer Verlag 2011. [5] A. Ben-Artzi, A. Ron, On the Integer Translates of a Compactly Supported Function: Dual Bases and Linear Projectors, SIAM J. Math. Anal., 21(6), 15501562. [6] P. Casazza, O. Christensen, D.T. Stoeva, Frame expansions in separable Banach spaces, J. Math. Anal. Appl. 307 (2005) 710 -- 723. [7] O. Christensen, D.T. Stoeva, p-Frames in separable Banach spaces, Adv. Comput. Math. 18 (2003) 117-126. [8] L. De Carli, P. Vellucci, Stability theorems for the n-order hold model, arXiv:1605.01706 (2016) (submitted) [9] H. Feichtinger, Spline-type spaces in Gabor analysis. Wavelet analysis (Hong Kong, 2001), 100122, Ser. Anal., 1, World Sci. Publ., River Edge, NJ, 2002. [10] H. Feichtinger, U. Molter, J.L.Romero, Perturbation techniques in irregular spline- type spaces, World Scientific Publishing Co. Inc., Int. J. Wavelets Multiresolut. Inf. Process, 6 (2) (2008) 249277. [11] H. Feichtinger, H.; D.M. Onchi, Constructive reconstruction from irregular sampling in multi-window spline-type spaces. Progress in analysis and its applications, 257265, World Sci. Publ., Hackensack, NJ, 2010. [12] G. Garrigos; R. Hochmuth; A. Tabacco, Wavelet characterizations for anisotropic Besov spaces with 0 < p < 1 Proc. Edinburgh Math. Soc. 47 (2004) 573-59. [13] K. Grochenig, J. Stockler, Gabor Frames and Totally Positive Functions, Duke Math. J. 162(6) (2013), 1003 -- 1031. [14] K. Hamm, J. Ledford, On the structure and interpolation properties of quasi-invariant shift spaces, arxiv:1703.01533 (2017) [15] C. Heil, A basis theory primer, Appl. Num. Harm. Analysis, Birkhauser 2011. [16] R.Q. Jia, C. A. Micchelli, Using the refinement equation for the construction of pre- wavelets II: power of two, In "Curve and Surface" (P. J. Laurent, A. Le Mehaute and L. L. Schumaker eds.), Academic Press, New York 1991, pp. 209 -- 246. [17] M.I. Kadec, The exact value of the Paley-Wiener constant, Soviet Math. Dokl., 5 (1964), 559 -- 561. [18] G. Kozma, S. Nitzan, Combining Riesz bases, Inv. Math. 199 (1) (2015), , pp 267-285 [19] E.V. Mischenko, Determination of Riesz bounds for the spline basis with the help of trigonometric polynomials. Sib. Math. J. 51(4), 660-666 (2010) [20] B. Pavlov, Basicity of an exponential system and Muckenhoupt's condition, Soviet Math. Dokl. 20 (1979) 655 -- 659. [21] Prautzsch, H., Boehm, W., Paluszny, M., Bezier and B-Spline Techniques, Springer Science and Business Media (2002). [22] R. Paley and N. Wiener, Fourier transforms in the complex domain. Amer. Math. Soc. Colloquium Publications, vol. 19; Amer. Math. Soc., New York, 1934. P-RIESZ BASES IN QUASI SHIFT INVARIANT SPACES 15 [23] J.L. Romero, Explicit localization estimates for spline-type spaces. Sampl. Theory Signal Image Process. 8 (2009), no. 3, 249-259. [24] A. Ron, A necessary and sufficient condition for the linear independence of the integer translates of a compactly supported distribution, Constructive Approximation 5 (1), 297-308. [25] A. M. Sedletskii, Izv. Vyssh. Uchebn. Zaved., Mat., No. 7, 85-91 (1973). [26] A. M. Sedletskii, Equivalence of the trigonometric system and its perturbations in Lp(π, π), Doklady Mathematics. Vol. 94. No. 1. Pleiades Publishing, 2016. [27] I. J. Schonberg, Cardinal interpolation and spline functions. J. Approx. teory 2 (2) (1969), pp. 167 -- 206. [28] A. Antony Selvan, R. Radha, Sampling and Reconstruction in Shift Invariant Spaces of B-Spline Functions Acta Appl. Math. DOI 10.1007/s10440-016-0053-6 (2016) [29] C. E. Shin, Q. Sun, Stability of localized operators, J. Funct. Anal. 256 (2009) 2417 -- 2439. [30] Q. Sun, Stability of the Shifts of Global Supported Distributions J.Math. Analysis and Appl. 261 (1)(2001) 113 -- 125. [31] W. Sun, X. Zhou, On the stability of multivariate trigonometric systems. J. Math Anal. Appl. 235 (1999), 159 -- 167. [32] M. Unser, A. Aldroubi, M. Eden, B-spline signal processing. I. Theory, Signal Pro- cessing, IEEE Transactions on 41 (2), 821 -- 833. [33] R. M. Young, An introduction to nonharmonic Fourier series, Academic Press, 2001 Laura De Carli: Florida International Univ., Univ. Park, Miami (FL) E-mail address: [email protected] Pierluigi Vellucci: Dept. of Economics, Roma Tre University, via Silvio D'Amico 77, 00145 Rome, Italy. E-mail address: [email protected]
1002.3583	2	1002	2011-07-25T14:11:34	Zone diagrams in compact subsets of uniformly convex normed spaces	[ "math.FA", "cs.CG", "math.GN" ]	A zone diagram is a relatively new concept which has emerged in computational geometry and is related to Voronoi diagrams. Formally, it is a fixed point of a certain mapping, and neither its uniqueness nor its existence are obvious in advance. It has been studied by several authors, starting with T. Asano, J. Matousek and T. Tokuyama, who considered the Euclidean plane with singleton sites, and proved the existence and uniqueness of zone diagrams there. In the present paper we prove the existence of zone diagrams with respect to finitely many pairwise disjoint compact sites contained in a compact and convex subset of a uniformly convex normed space, provided that either the sites or the convex subset satisfy a certain mild condition. The proof is based on the Schauder fixed point theorem, the Curtis-Schori theorem regarding the Hilbert cube, and on recent results concerning the characterization of Voronoi cells as a collection of line segments and their geometric stability with respect to small changes of the corresponding sites. Along the way we obtain the continuity of the Dom mapping as well as interesting and apparently new properties of Voronoi cells.	math.FA	math	ZONE DIAGRAMS IN COMPACT SUBSETS OF UNIFORMLY CONVEX NORMED SPACES EVA KOPECK ´A, DANIEL REEM, AND SIMEON REICH Abstract. A zone diagram is a relatively new concept which has emerged in computational geometry and is related to Voronoi diagrams. Formally, it is a fixed point of a certain mapping, and neither its uniqueness nor its existence are obvious in advance. It has been studied by several authors, starting with T. Asano, J. Matousek and T. Tokuyama, who considered the Euclidean plane with singleton sites, and proved the existence and uniqueness of zone diagrams there. In the present paper we prove the existence of zone diagrams with respect to finitely many pairwise disjoint compact sites contained in a compact and convex subset of a uniformly convex normed space, provided that either the sites or the convex subset satisfy a certain mild condition. The proof is based on the Schauder fixed point theorem, the Curtis-Schori theorem regarding the Hilbert cube, and on recent results concerning the characterization of Voronoi cells as a collection of line segments and their geometric stability with respect to small changes of the corresponding sites. Along the way we obtain the continuity of the Dom mapping as well as interesting and apparently new properties of Voronoi cells. 1. Introduction A zone diagram is a relatively new concept related to geometry and fixed point theory. In order to understand it better, consider first the more familiar concept of a Voronoi diagram. In a Voronoi diagram we start with a set X, a distance function d, and a collection of subsets (Pk)k∈K in X (called the sites or the gen- erators), and with each site Pk we associate the Voronoi cell Rk, that is, the set of all x ∈ X the distance of which to Pk is not greater than its distance to the union of the other sites Pj, j 6= k. On the other hand, in a zone diagram we associate with each site Pk the set Rk of all x ∈ X the distance of which to Pk is not greater than its distance to the union of the other sets Rj, j 6= k. Figures 1 and 2 show the Voronoi and zone diagrams, respectively, corresponding to the same ten singleton sites in the Euclidean plane. At first sight, it seems that the definition of a zone diagram is circular, because the definition of each Rk depends on Rk itself via the definition of the other cells Rj, j 6= k. On second thought, we see that, in fact, a zone diagram is defined to be a fixed point of a certain mapping (called the Dom mapping), that is, a solution of a certain equation. While the Voronoi diagram is explicitly defined, so its existence (and uniqueness) are obvious, neither the existence nor the uniqueness of a zone diagram are obvious in advance. As a result, in addition to the problem of finding algorithms for computing zone diagrams, we are faced with the more 1 2 EVA KOPECK ´A, DANIEL REEM, AND SIMEON REICH Figure 1. A Voronoi diagram of 10 point sites in a square in (R2, ℓ2). Figure 2. A zone diagram of the same 10 points as in Figure 1. fundamental problem of establishing their existence (and uniqueness) in various settings, and with the problem of reaching a better understanding of this concept. The concept of a zone diagram was first defined and studied by T. Asano, J. Matousek and T. Tokuyama [2, 3] (see also [4]), in the case where (X, d) was the Euclidean plane, each site Pk was a single point, and all these (finitely many) points were different. They proved the existence and uniqueness of a zone diagram in this case, and also suggested a natural iterative algorithm for approximating it. Their proofs rely heavily on the above setting. Several other papers related to zone diagrams in the plane have been published, e.g., those of T. Asano and D. Kirkpatrick [1], of J. Chun, Y. Okada and T. Tokuyama [6], and recently of S. C. de Biasi, B. Kalantari and I. Kalantari [9]. Shortly after [3], the authors of [22] considered general sites in abstract spaces, called m-spaces, in which X is an arbitrary nonempty set and the "distance" function should only satisfy the condition d(x, x) ≤ d(x, y) ∀x, y ∈ X and can take any value in the interval [−∞, ∞]. They introduced the concept of a double zone diagram, and using it and the Knaster-Tarski fixed point theorem, proved the existence of a zone diagram with respect to any two sites in X. They also showed that in general the zone diagram is not unique. In a recent work by K. Imai, A. Kawamura, J. Matousek, Y. Muramatsu and T. Tokuyama [14], the existence and uniqueness of the zone diagram with respect to any number of general positively separated sites in the n-dimensional Euclidean space Rn was announced. The proof is based on results from [22] and on an elegant geometric argument specific to Euclidean spaces. Very recently some of these authors have generalized this result to finite dimensional normed spaces which are both strictly convex and smooth [15]. In the present paper we prove the existence of zone diagrams with respect to finitely many pairwise disjoint compact sites contained in a compact and convex subset of a (possibly infinite dimensional) uniformly convex space, provided that ZONE DIAGRAMS 3 either the sites or the convex subset satisfy a certain mild condition. This mild condition holds, for instance, if either the convex subset X has a strictly convex boundary or the sites are contained in the interior of X relative to the affine hull spanned by X (and X is arbitrary). The proof is based on the Schauder fixed point theorem, the Curtis-Schori theorem regarding the Hilbert cube, and on re- cent results concerning the characterization of Voronoi cells as a collection of line segments and their geometric stability with respect to small changes of the corre- sponding sites (see Sections 3 and 4). Along the way we obtain the continuity of the Dom mapping (Proposition 6.1) in a general setting and interesting properties of Voronoi cells, namely Lemma 5.1 and Lemma 5.2. Although Voronoi diagrams have been the subject of extensive research during the last decades [5, 18, 12], this research has been mainly focused on Euclidean finite dimensional spaces (in many cases just R2 or R3), and it seems that these lemmata are new even for R2 with a non-Euclidean norm. It may be of interest to compare our main existence result with the recent existence result described in [15]. On the one hand, our result is weaker than that result, since we only prove the existence of a zone diagram in a compact and convex set, while in [15] uniqueness is also proved and the setting is the whole space Rn. In addition, it seems that some of the arguments in [15], although only formulated for finitely many compact sites, can be extended to infinitely many, positively separated closed sites. On the other hand, our result is stronger in the sense that we allow infinite dimensional spaces and we do not require the smoothness of the norm. As a matter of fact, the counterexamples mentioned in [15] show that uniqueness does not necessarily hold if the norm is not smooth. In any case, the strategies used for proving these two results are completely different: in [15] the authors use the existence of double zone diagrams (based on the Knaster- Tarski fixed point theorem) and several geometric arguments, and here we use the Schauder fixed point theorem, the Curtis-Schori theorem regarding the Hilbert cube, and several general results about Voronoi cells in uniformly convex normed spaces and elsewhere. The structure of the paper is as follows: in Section 2 we present the basic definitions and notation. In Section 3 we provide the outline of the proof of the main result. In Section 4 we formulate several claims which are needed in the proof. The proof itself is given in Sections 5, 6 and 7. We conclude the paper with Section 8, which contains open problems and two pictures of zone diagrams. We note that although the setting in the main result is a compact and convex subset of a uniformly convex normed space, many of the auxiliary results actually hold in a more general setting with essentially the same proofs, and therefore we formulate and prove them there. 2. Definitions and Notation In this section we present our notation and basic definitions. We consider a closed and convex set X 6= ∅ in some normed space (eX, · ), real or complex, 4 EVA KOPECK ´A, DANIEL REEM, AND SIMEON REICH finite or infinite dimensional. The induced metric is d(x, y) = x − y. We assume that X is not a singleton, for otherwise everything is trivial. The notation θ = 1 will always mean a unit vector θ. Given p ∈ X we set Θp = {θ : θ = 1, p + tθ ∈ X for some t > 0}. This is the set of all directions such that rays emanating from p in these directions intersect X not only at p. We denote by [p, x] the closed line segment connecting p and x, i.e., the set {p + t(x − p) : t ∈ [0, 1]}. We denote by B(x, r) the open ball of center x and radius r. The notation Y ≈ Z means that the topological spaces Y and Z are homeomorphic. Definition 2.1. Given two nonempty sets P, A ⊆ X, the dominance region dom(P, A) of P with respect to A is the set of all x ∈ X the distance of which to P is not greater than its distance to A, that is, dom(P, A) = {x ∈ X : d(x, P ) ≤ d(x, A)}. Here d(x, A) = inf{d(x, a) : a ∈ A}. Definition 2.2. Let K be a set of at least 2 elements (indices), possibly infinite. Given a tuple (Pk)k∈K of nonempty subsets Pk ⊆ X, called the generators or the sites, the Voronoi diagram induced by this tuple is the tuple (Rk)k∈K of nonempty subsets Rk ⊆ X, such that for all k ∈ K, Rk = dom(Pk,[j6=k Pj) = {x ∈ X : d(x, Pk) ≤ d(x, Pj) ∀j 6= k, j ∈ K}. In other words, each Rk, called a Voronoi cell, is the set of all x ∈ X the distance of which to Pk is not greater than its distance to the union of the other Pj, j 6= k. Definition 2.3. Let (X, d) be a metric space and let K be a set of at least 2 elements (indices), possibly infinite. Given a tuple (Pk)k∈K of nonempty subsets Pk ⊆ X, a zone diagram with respect to that tuple is a tuple R = (Rk)k∈K of nonempty subsets Rk ⊆ X such that Rk = dom(Pk, Sj6=k Rj) ∀k ∈ K. In other words, if we define Xk = {C : Pk ⊆ C ⊆ X}, then a zone diagram is a fixed point of the mapping Dom : Qk∈K Xk → Qk∈K Dom(R) = (dom(Pk, Sj6=k (1) Xk, defined by Rj))k∈K. For example, let P1 = {p1} and P2 = {p2} be two different sets in the Euclidean plane, or more generally, in a Hilbert space X. In this case the corresponding Voronoi diagram consists of two half-spaces: dom(P1, P2) is the half-space con- taining P1 and determined by the hyperplane passing through the middle of the line segment [p1, p2] and perpendicular to it, and dom(P2, P1) is the other half- space. The zone diagram of (P1, P2) exists by the results of [3] (in the case of the Euclidean plane), or [22] (in the general case), but it is not clear how to describe the zone diagram explicitly. ZONE DIAGRAMS 5 We now recall the definition of strictly and uniformly convex spaces. Definition 2.4. A normed space (eX, · ) is said to be strictly convex if for all x, y ∈ eX satisfying x = y = 1 and x 6= y, the inequality (x + y)/2 < 1 holds. (eX, · ) is said to be uniformly convex if for any ǫ ∈ (0, 2], there exists δ ∈ (0, 1] such that for all x, y ∈ eX, if x = y = 1 and x − y ≥ ǫ, then (x + y)/2 ≤ 1 − δ. Roughly speaking, if the space is uniformly convex, then for any ǫ > 0, there exists a uniform positive lower bound on how deep the midpoint between any two unit vectors must penetrate the unit ball, assuming the distance between them is at least ǫ. In general normed spaces the penetration is not necessarily positive, since the unit ball may contain line segments. The plane R2 endowed with the max norm · ∞ is a typical example of this. A uniformly convex space is always strictly convex, and if it is also finite dimensional, then the converse is true too. The n-dimensional Euclidean space Rn, or more generally, inner product spaces, the sequence spaces ℓp, the Lebesgue spaces Lp(Ω), p ∈ (1, ∞), and a uniformly convex product of a finite number of uniformly convex spaces, are all examples of uniformly convex spaces. See [7] and, for instance, [16] and [11] for more information regarding uniformly convex spaces. We now recall three definitions of a topological character. Definition 2.5. The Hilbert cube I ∞ is the set I ∞ =Q∞ n=1[0, 1/n] as a topological space the topology of which is induced by the ℓ2 norm, or, equivalently, by the product topology. Definition 2.6. A topological space X is said to be locally (path) connected if for any x ∈ X and any open set U containing x, there exists a (path) connected open set V ⊆ U such that x ∈ V . X is said to be weakly locally (path) connected, or (path) connected im kleinen, if for any x ∈ X and any open set U containing x, there exists a (path) connected set C ⊆ U and an open set V ⊆ C such that x ∈ V . Definition 2.7. Let (X, d) be a metric space. Given two nonempty sets A1, A2 ⊆ X, the Hausdorff distance between them is defined by D(A1, A2) = max{ sup a1∈A1 d(a1, A2), sup a2∈A2 d(a2, A1)}. Recall that the Hausdorff distance is different from the usual distance between two sets which is defined by d(A1, A2) = inf{d(a1, a2) : a1 ∈ A1, a2 ∈ A2}. Recall also that if (X, d) is compact and we consider the set of all its nonempty closed subsets, then this space is a compact metric space with the Hausdorff distance as its metric [13]. We end this section with a definition which is pertinent to the formulation of It is followed by a brief our main result and some of our auxiliary assertions. discussion. 6 EVA KOPECK ´A, DANIEL REEM, AND SIMEON REICH Definition 2.8. Let X be a closed and convex subset of a normed space. Let p ∈ X. Let θ ∈ Θp. Let L(θ) ∈ (0, ∞] be the length of the line segment generated from the intersection of X and the ray emanating from p in the direction of θ. The point p is said to have the emanation property (or to satisfy the emanation condition) in the direction of θ if for each ǫ > 0 there exists β > 0 such that for any φ ∈ Θp, if φ − θ < β, then the intersection of X and the ray emanating from p in the direction of φ is a line segment of length at least L(θ) − ǫ. In other words, L(φ) ≥ L(θ) − ǫ. The point p is said to have the emanation property if it has the emanation property in the direction of every θ ∈ Θp. A subset C of X is said to have the emanation property if each p ∈ C has the emanation property. The following examples illustrate the emanation property. In the first four the emanation property holds, and in the last one it does not hold. Example 2.9. X is any bounded closed convex set and p ∈ X is an arbitrary point in the interior of X relative to the affine hull spanned by X. Example 2.10. The boundary of the bounded closed and convex X is strictly convex (if a 6= b are two points in the boundary, then the open line segment (a, b) is contained in the interior of X relative to the affine hull spanned by X) and p ∈ X is arbitrary. Any ball in a strictly convex space has a strictly convex boundary. Example 2.11. X is a cube (of any finite dimension) and p ∈ X is arbitrary. Example 2.12. X is a closed linear subspace and p ∈ X is arbitrary. Example 2.13. This example shows that the emanation condition does not hold in general. Consider the Hilbert space ℓ2. Let (en)∞ n=1 be the standard basis. Let y1 = e1 and for each n > 1 let yn = e1/2 + en/n. Let A = {−e1}. Let X be the closed convex hull generated by AS{yn : n = 1, 2, . . .}. Let p = 0 and let Indeed, limn→∞ θn = θ1 but L(θn) = p0.25 + 1/n2 < 0.99 = θn = yn/yn for each n. Then p does not satisfy the emanation condition in the direction of θ1. L(θ1) − 0.01 for each n > 1. The subset X is in fact compact since the sequence (yn)∞ n=1 converges. For more details about the emanation property, see [21]. 3. Outline of the proof of the main result In this section we outline the proof of our main result, Theorem 7.2. It states that there exists a zone diagram with respect to finitely many pairwise disjoint compact sites in a compact and convex subset of a uniformly convex normed space, provided that either the sites or the convex subset satisfy a certain mild condition. The idea of the proof is to find a certain space Y homeomorphic to the Hilbert cube I ∞ (Y ≈ I ∞ for short) such that Dom(Y ) ⊆ Y and Dom is continuous on Y . Now, if h : I ∞ → Y is a homeomorphism, then f = h−1 ◦ Dom ◦ h : I ∞ → I ∞ is a continuous mapping which maps a compact and convex subset of ℓ2 into itself, so the Schauder fixed point theorem [23] (see also [10, p. 119] and Theorem 3.1 ZONE DIAGRAMS 7 below) ensures that f has a fixed point q ∈ I ∞. By taking R = h(q), we see that R is a fixed point of Dom, that is, R is a zone diagram. In order to apply this idea, one has to find the set Y , to prove that it is homeomorphic to I ∞, and to prove the continuity of Dom on Y . It has turned out that even in the case of singleton sites in a square in the Euclidean plane the proof is not obvious (the main difficulty is to prove the continuity of Dom), and, in fact, such a proof has never been published. The above strategy was suggested by the first author, and was briefly mentioned in [3, p. 1188]. The space Y was taken to beQk∈K Yk, where K was finite, Yk was {C : Pk ⊆ C ⊆ Qk and C is closed} and Qk was the intersection of the Voronoi cell of Pk with X (the square). Since each site Pk is taken to be a singleton, it follows that each Qk is actually convex, so, in particular, it is a connected and locally connected compact metric space. Since, in addition, Pk 6= Qk, it follows from the theorem of D. Curtis and R. Schori [8, Theorem 5.2] stated below (see also [13, p. 91]) that Yk, as a metric space endowed with the Hausdorff metric, is homeomorphic to I ∞. The topology on Y is the product topology, induced by the k) : k ∈ K}, so Y , as a finite product of spaces homeomorphic to I ∞, is also homeomorphic to I ∞. uniform Hausdorff metric eD((Sk)k∈K, (S′ k)k∈K) = max{D(Sk, S′ Theorem 3.1. (Schauder) Let X be a nonempty convex and compact subset of a normed space. If f : X → X is continuous, then it has a fixed point. Theorem 3.2. (Curtis-Schori) Let X be a Peano continuum, that is, a con- nected and locally connected compact metric space, and let P ⊆ X, P 6= X be closed and nonempty. Let 2X P = {C : P ⊆ C ⊆ X, C is closed}, endowed with the Hausdorff metric. Then 2X P ≈ I ∞. In the general case, the application of the above strategy, and, in particular, the verification of the hypotheses of Theorem 3.2, are not a simple task, and they require several additional tools related to dominance regions, such as their characterization as unions of line segments, and their stability with respect to small perturbations of the relevant sets. These results will be stated in the next section. They have recently been established in [19, 21], and their proofs can be found there. Using these results, we first prove the existence of a zone diagram with respect to finite sites, and then, approximating compact sets by finite subsets of them and applying a continuity argument, we extend this existence result to any compact sites with the emanation property. 4. Several technical tools In this section we either prove or cite several technical claims needed for estab- lishing the main result; see [21] and [19] for those proofs not included here. The following theorem is a new representation theorem for dominance regions. Theorem 4.1. Let X be a closed and convex subset of a normed space. Let P, A ⊆ X be nonempty. Suppose that the distance between x and P is attained 8 EVA KOPECK ´A, DANIEL REEM, AND SIMEON REICH for all x ∈ X. Then dom(P, A) is a union of line segments starting at the points of P . More precisely, given p ∈ P and θ = 1, let (2) T (θ, p) = sup{t ∈ [0, ∞) : p + tθ ∈ X and d(p + tθ, p) ≤ d(p + tθ, A)}. Then dom(P, A) = [p∈P [θ=1 [p, p + T (θ, p)θ]. When T (θ, p) = ∞, the notation [p, p+T (θ, p)θ] means the ray {p+tθ : t ∈ [0, ∞)}. The proof of Theorem 4.1 is based on the following simple observation, which will also be needed for a different purpose later (see the proof of Lemma 5.2). Lemma 4.2. Let (eX, · ) be a normed space, and let ∅ 6= A ⊆ eX. Suppose that y, p ∈ eX satisfy d(y, p) ≤ d(y, A). Then d(x, p) ≤ d(x, A) for any x ∈ [p, y]. The next two theorems describe a continuity property of dominance regions and the mapping T defined in (2) in uniformly convex normed spaces. We note that condition (3) below expresses the fact that the set A is "well distributed in X". It obviously holds when X is bounded, but it may also hold even when X is not bounded, as in the case where X = Rn and A is the lattice of points with integer coordinates. Theorem 4.3. Let X be a closed and convex subset of a uniformly convex normed space. Then, under certain conditions, the mapping dom(·, ·) has a uniform con- tinuity property with respect to the Hausdorff distance. More precisely, assume that P and A are nonempty with d(P, A) > 0. Suppose that ∃ρ ∈ (0, ∞) such that ∀x ∈ X the open ball B(x, ρ) intersects A. (3) Then for each ǫ ∈ (0, d(P, A)/6) there exists ∆ > 0 such that if D(A, A′) < ∆, D(P, P ′) < ∆, and the distances between any point x ∈ X and both P and P ′ are attained, then the inequality D(dom(P, A), dom(P ′, A′)) < ǫ holds. Theorem 4.4. Let X be a closed and convex subset of a uniformly convex normed space. Let p ∈ X and suppose that p has the emanation property. Then the mapping T (·, p) has a certain continuity property. More precisely, let A ⊂ X be nonempty such that (3) holds and that d(p, A) > 0. Then for each ǫ ∈ (0, d(p, A)/6) and each θ ∈ Θp there exists ∆ > 0 such that for each φ ∈ Θp, if θ − φ < ∆, then T (θ, p) − T (φ, p) ≤ ǫ. In addition, the range of the mapping T is bounded by ρ. We remark in passing that in the proofs of Theorems 4.3 and 4.4, the uni- form convexity of the space enters, inter alia, through Clarkson's strong triangle inequality [7, Theorem 3]. The following lemma will be needed for proving the local connectedness of certain dominance regions (Lemma 5.1 and Lemma 5.2). Note that the set Bp is not necessarily the unit ball (or a ball of some closed affine hull), as in the case where X is the Hilbert cube I ∞ in ℓ2, or more generally, a compact and convex subset of an infinite dimensional normed space. ZONE DIAGRAMS 9 Lemma 4.5. Let X be a convex set in a normed space and let p ∈ X. Define the set Bp = {rθ : r ∈ [0, 1], θ ∈ Θp}. Then Bp is convex. Proof. Let b1, b2 ∈ Bp. Then bi = riθi for some ri ∈ [0, 1] and θi = 1. Let b = λ1b1 + λ2b2 for some λ1, λ2 ∈ [0, 1] with λ1 + λ2 = 1. Suppose that b 6= 0, for otherwise b ∈ Bp. Then b = rθ with r = λ1b1 + λ2b2 and θ = (λ1b1 + λ2b2)/r. Clearly, 0 ≤ r ≤ r1λ1 + r2λ2 ≤ 1, and θ = 1, so it remains to prove that p + tθ ∈ X for some t > 0. By the definition of θi, there is ti > 0 such that yi := p + tiθi ∈ X. We can assume that λi 6= 0 and ri 6= 0 for i ∈ {1, 2}, for otherwise θ = θj for j 6= i and p + tjθ ∈ X. Let β1 = λ1r1t2/(λ1r1t2 + λ2r2t1), β2 = λ2r2t1/(λ2r2t1 + λ1r1t2) and t = (β1t1 + β2t2)r/(λ1r1 + λ2r2). These values were obtained by equating the coefficients of the θi in the equation p + tθ = β1y1 + β2y2. It is easy to check that βi ∈ (0, 1), β1 + β2 = 1, t > 0 and p + tθ = β1y1 + β2y2. Thus p + tθ ∈ X because X is convex. (cid:3) The next two lemmata are probably known, at least in a version close in its spirit to their formulation (see, e.g., [17, p. 162, exercise 6] and [13, Proposition 10.7, pp. 82-83]), but we include their proof for the sake of completeness. Lemma 4.6. If Z is a topological space which is weakly locally (path) connected, then it is locally (path) connected. Proof. The proof is based on the fact that Z is locally (path) connected if and only if for every open set U of Z each (path) connected component of U is open in Z [17, p. 161]. Given an open set U, let x ∈ U. Let Cx be the (path) connected component of x in U. Since Z is weakly locally (path) connected, there exist a (path) connected set C ⊆ U and an open set V ⊆ C such that x ∈ V . By the maximality of Cx, we have C ⊆ Cx, and hence V ⊆ Cx. Thus x belongs to the interior of Cx, and in the same way all the other points of Cx are in its interior. Hence Cx is open in X and the same is true for all other (path) components of U. (cid:3) Lemma 4.7. Let (Z, d) be a metric space and suppose that Z =Sm i=1 Zi for some locally (path) connected closed subsets Zi of Z. Then Z is locally (path) connected. Proof. It suffices to prove the assertion for m = 2; the general case follows by induction. Let x ∈ Z. Then x ∈ Zi for some i. Let U be an open neighborhood of Z with x ∈ U. Assume first that x /∈ Zj for j 6= i. Then the intersection U ∩ B(x, d(x, Zj)/2) ⊆ Zi contains an open (path) connected subset V of Zi with x ∈ V , since Zi is locally (path) connected. Note that V is also (path) connected with respect to the topology of Z. Since d(V, Zj) > 0, it follows that V is actually open in Z, so x has an open (path) connected neighborhood contained in U in this case. Assume now that x ∈ Z1 ∩ Z2. Then U ∩ Zi is an open neighborhood of x in Zi, so for each i ∈ {1, 2}, there is an open (path) connected subset Vi ⊆ U ∩ Zi of Zi with x ∈ Vi, because Zi is locally (path) connected. The union V1 ∪ V2 is a (path) connected subset of Z which is contained in U, but it is not clear whether 10 EVA KOPECK ´A, DANIEL REEM, AND SIMEON REICH it is open in Z. However, by definition, Vi = V ′ and i ∩ Zi for some open set V ′ i of Z, x ∈ V ′ 2 ∩ Z2) ⊆ V1 ∪ V2 ⊆ U. 1 ∩ V ′ 2 = (V ′ 1 ∩ V ′ 2 ∩ Z1) ∪ (V ′ 1 ∩ V ′ Since x is arbitrary, this proves that Z is weakly locally (path) connected, so by Lemma 4.6 it is locally (path) connected. (cid:3) The following lemma will be useful for proving the main result of Section 5. Lemma 4.8. Let X be a convex set in a normed space, and suppose that P, A ⊂ X satisfy r := d(P, A) > 0. Let B(P, r/2) := {x ∈ X : d(x, P ) < r/2}. Then B(P, r/2) ⊆ {x ∈ X : d(x, P ) < d(x, A)} and d(x, P ) ≥ r/2 for any x ∈ X satisfying x ∈ dom(A, P ). Proof. The set B(P, r/2) is contained in {x ∈ X : d(x, P ) < d(x, A)} because if x ∈ B(P, r/2) and a ∈ A, then r ≤ d(a, P ) ≤ d(a, x) + d(x, P ) < d(a, x) + r/2, so d(x, P ) < r/2 ≤ d(x, A). Now let x ∈ dom(A, P ). Since d(x, A) ≤ d(x, P ), this point does not belong to B(P, r/2) by the above paragraph. Let p ∈ P be arbitrary, and consider the line segment [x, p]. It intersects the boundary of B(P, r/2) (otherwise the connected space [x, p] would have a decomposition as a union of two disjoint open sets) at some point y, and since B(P, r/2) is open, it follows that d(y, P ) ≥ r/2. Hence d(x, p) = d(x, y) + d(y, p) ≥ r/2, i.e., d(x, P ) ≥ r/2, as claimed. (cid:3) We finish with the following lemma, the proof of which is a simple consequence of the definition. Lemma 4.9. The equality dom(Sn i=1 Pi, A) = Sn nonempty subsets P1, . . . , Pn and A of the metric space X. i=1 dom(Pi, A) holds for any 5. The space homeomorphic to the Hilbert cube The main result of this section is Proposition 5.4 below. It is based on Lemma 5.2 which shows that dom(p, A) is locally path connected whenever d(p, A) > 0, and on Lemma 5.3 which generalizes this result to dom(P, A) for any finite set P . As Lemma 4.2 shows, dom(p, A) is star-shaped. However, this property by itself is not sufficient for concluding that it is locally (path) connected, since there are simple examples of star-shaped sets in R2 which are not locally connected. In fact, a result of T. Zamfirescu [24, Theorem 4] shows that a large class (in the sense of Baire category) of star-shaped sets in Rn are not locally connected. As a result, the local path connectedness of dom(p, A) must be proved. It has turned out that the proof, given in Lemma 5.2, is somewhat technical, and in a special but important case, namely Lemma 5.1 below, a simpler proof can be presented. This special case is of interest in itself, and its proof also casts some light on the strategy for proving Lemma 5.2. The condition on p and s described in Lemma 5.1 holds, for instance, when p is in the interior of X relative to the closed affine hull spanned by it. See also the short discussion before Lemma 4.5 regarding the set Bp. ZONE DIAGRAMS 11 Figure 3. An example related to Lemma 5.1 in a square in (R2, ℓ7). Here p = pk, A is the collection of the other singleton sites pj, j 6= k, and dom(p, A) is the Voronoi cell of pk. The set Bp is simply the unit ball. 4. An the conclusion example Figure of where Lemma 5.1 fails if the uni- form convexity assumption is removed. Here X is a square in (R2, ℓ∞), p = (0, 0), and A = {(2, 0), (−2, 0), (0, −2)}. The green region is dom(p, A). Lemma 5.1. Let X be a closed and convex subset of a uniformly convex normed space. Let p ∈ X. Suppose that there exists some s > 0 such that p + sθ ∈ X for each θ ∈ Θp. Suppose also that p has the emanation property. Let ∅ 6= A ⊂ X be such that d(p, A) > 0. Suppose also that condition (3) holds. Then dom(p, A) is homeomorphic to the set Bp = {rθ : r ∈ [0, 1], θ ∈ Θp}, and, in particular, dom(p, A) is path connected and locally path connected. Proof. Let f : Bp → dom(p, A) be defined by f (rθ) = p + rT (θ, p)θ, i.e., f (0) = p and f (x) = p + T (x/x, p)x for x 6= 0, where T (θ, p) is defined in (2). Let ǫ = min{d(p, A)/2, s/2}. Let θ ∈ Θp and t ∈ [0, ǫ] be arbitrary. Then for x = p + tθ we have x ∈ X because t ≤ s and X is convex. In addition, 2ǫ ≤ d(p, A) ≤ d(p, x) + d(x, A) ≤ ǫ + d(x, A), so d(x, p) ≤ ǫ ≤ d(x, A) and hence x ∈ dom(p, A). Thus T (θ, p) ≥ ǫ by the definition of T . From Theorem 4.4 it follows that T is bounded, so f is well defined. By Theorem 4.1, f is onto, and it is one-to-one by a direct calculation. If y = p + rT (θ, p)θ, then θ = (y − p)/y − p and r = y − p/T (θ, p), so the inverse function is defined by f −1(y) =(cid:26) y−p T ((y−p)/y−p,p) y 6= p y = p. 0 It now follows from Theorem 4.4 that both f and f −1 are continuous. Since Bp is convex by Lemma 4.5, it is path connected and locally path connected, so dom(p, A) is both path connected and locally path connected, as claimed. (cid:3) 12 EVA KOPECK ´A, DANIEL REEM, AND SIMEON REICH An example related to Lemma 5.1 is given in Figure 3. This may be somewhat surprising, but in general the conclusion of Lemma 5.1 is not true, and Figure 4 presents a counterexample in the non-uniformly convex space (R2, ℓ∞) with two simple sets. An examination of the above proof suggests that problems can appear if no positive number s satisfies the condition in the formulation of Lemma 5.1, because in this case the function f −1 may not be continuous. Such a phenomenon occurs in infinite dimensional spaces, for instance, when X is the Hilbert cube I ∞ in ℓ2, but cannot happen if p is in the interior of X relative to the closed affine hull spanned by it. In order to replace Lemma 5.1 one has to show directly by other arguments that dom(p, A) is connected and locally connected. This will be done in Lemma 5.2 below. Lemma 5.2. Let X be a closed and convex set in a uniformly convex normed space. Let p ∈ X and ∅ 6= A ⊂ X and suppose that d(p, A) > 0. Suppose also that condition (3) holds and that p has the emanation property. Then dom(p, A) is path connected and locally path connected. Proof. By Lemma 4.2, any two points x1, x2 ∈ dom(p, A) can be connected via p, so dom(p, A) is path connected. In order to prove that dom(p, A) is locally path connected it suffices by Lemma 4.6 to show that it is weakly locally path connected. Let x0 ∈ dom(p, A). If x0 = p, then let V = B(x0, r) ∩ X for r = d(p, A)/3. Given y ∈ V , the triangle inequality shows that 3r = d(p, A) ≤ d(p, y) + d(y, A) ≤ r + d(y, A). Hence d(y, p) ≤ r < 2r ≤ d(y, A) and y ∈ dom(p, A), so V ⊆ dom(p, A). Since V is convex, x0 = p has a path connected open neighborhood. Now suppose that x0 6= p, and for any ǫ > 0 let U = B(x0, ǫ) ∩ dom(p, A) be any neighborhood of x0. Let θ0 = (x0 − p)/x0 − p. Then θ0 = 1, and since x0 = p+x0 −pθ0 ∈ dom(p, A), the definition of T (θ0, p) (in (2)) implies that d(x0, p) ≤ T (θ0, p). Hence x′ 0 := (x0 −p)/T ((x0 −p)/x0 −p, p) = (x0 −p/T (θ0, p))θ0 is well defined, different from 0 and belongs to Bp = {rθ : r ∈ [0, 1], θ = 1, p + tθ ∈ X for some t > 0}. In addition, x′ > 0 for each x′ ∈ B(x′ 0/2) ∩ Bp. 0, x′ Since T (θ0, p) > 0, the continuity of T (·, p) implies that there exists δ2 ∈ (0, 1) such that if θ ∈ Θp and θ − θ0 < δ2, then T (θ, p) > 0. Let δ3 = δ2x0 − p/2. 0, x′ Define f : B(x′ 0 < δ1, then f (x′) − f (x′ T (·, p) is continuous (Theorem 4.4) and since f (x′ continuous, so for the given ǫ > 0, there exists some δ1 ∈ (0, x′ x′ − x′ Bp. Since U ′ continuity of f implies that f (U ′ containing x0 = f (x′ dom(p, A), it suffices to show that f (U ′ 0/2) ∩ Bp → dom(p, A) by f (x′) = p + T (x′/x′, p)x′. Since 0) = x0, it follows that f is 0/2) such that if 0, δ1) ∩ 0, the 1) is a path connected set contained in U and 0). In order to establish the weak (path) connectedness of 1) contains an open neighborhood of x0. 1 = B(x′ 1 is a path connected (in fact, convex) open set containing x′ 0) < ǫ. Hence f (U ′ 1) ⊂ U for U ′ Given x ∈ B(x0, δ3) ∩ dom(p, A), we have ZONE DIAGRAMS 13 (4) x − p x − p − x0 − p(x − p) − x − p(x0 − p) x − px0 − p = (cid:12)(cid:12)(cid:12)(cid:12) x0 − p x0 − p(cid:12)(cid:12)(cid:12)(cid:12) = (x0 − p − x − p)(x − p) + x − p((x − p) − (x0 − p)) x − px0 − p ≤ 2x0 − x x0 − p < δ2. Thus T ((x − p)/x − p, p) > 0 for each x ∈ B(x0, δ3) ∩ dom(p, A). For each such x, let g(x) = (x − p)/T ((x − p)/x − p, p). Since g is continuous, for δ1 from the definition of U ′ 1 there exists δ4 ∈ (0, δ3) such that if x − x0 < δ4, then g(x) − g(x0) < δ1. But g(x0) = x′ 1 for each x in the set V := B(x0, δ4) ∩ dom(p, A). In addition, by the definition of f and g we have x = f (g(x)) for each x ∈ V (just let θ = (x − p)/x − p and x′ = g(x), and note that x′/x′ = θ). Hence V is an open neighborhood of x0 which is contained in f (U ′ 1), so dom(p, A) is weakly locally path connected and hence locally path connected by Lemma 4.6. (cid:3) 0 and g(x) ∈ U ′ Lemma 5.3. Let X be a closed and convex subset of a uniformly convex normed space. Let P = {p1, . . . , pm} ⊂ X and assume that ∅ 6= A ⊂ X satisfies d(P, A) > 0. Suppose also that condition (3) holds and that the subset P has the emanation property. Then dom(P, A) is locally path connected. Sm Proof. This assertion is a simple consequence of the facts that dom(P, A) = i=1 dom(pi, A)) (by Lemma 4.9), that dom(pi, A) is locally path connected (by Lemma 5.2) and closed for each i, and the fact that the metric space dom(P, A) is a finite union of locally path connected closed sets and hence locally path con- nected by Lemma 4.7. Note that the topology of dom(pi, A), which is induced by the norm of the space, coincides with its topology as a subspace of dom(P, A), and hence we can apply Lemma 4.7. (cid:3) Proposition 5.4. Let X be a compact and convex subset of a uniformly convex normed space, and let (Pk)k∈K be a finite tuple of finite sets in X which are pairwise disjoint. Suppose that for each k ∈ K the site Pk has the emanation {C : C is closed and Pk ⊆ C ⊆ Qk}, endowed with the Hausdorff metric. Let property. For each k ∈ K, let Ak = Sj6=k Pj and Qk = dom(Pk, Ak). Let Yk = Y =Qk∈K Yk, endowed with the uniform Hausdorff metric. Then Y ≈ I ∞. Proof. Given k ∈ K and p ∈ Pk, let Cp be the path connected component of p in Qk, and let Ep = {q ∈ Pk : q ∈ Cp}. Since d(Pk, Ak) > 0 for each k by assumption, we have d(q, Ak) > 0 for any q ∈ Ep, so dom(q, Ak) is path connected by Lemma 5.2. Thus dom(q, Ak) ⊆ Cp for each q ∈ Ep, and hence dom(Ep, Ak) ⊆ Cp by Lemma 4.9. On the other hand, if x ∈ Cp, then x ∈ dom(p′, Ak) for some p′ ∈ Pk, dom(p, Ak) by Lemma 4.9. Lemma 5.2 implies that there is a path between p′ and x, and since x ∈ Cp, there is a path between x and p. Thus p′ ∈ Ep, so x ∈ dom(p′, Ak) ⊆ dom(Ep, Ak) and hence dom(Ep, Ak) = Cp. because Qk =Sp∈Pk Since for each p1, p2 ∈ Pk, either Ep1 = Ep2 or Ep1 ∩ Ep2 = ∅, it follows that {Ep : p ∈ Pk} = {Ek,1, . . . , Ek,mk} for some disjoint sets Ek,l, l = 1, . . . , mk, 14 EVA KOPECK ´A, DANIEL REEM, AND SIMEON REICH the union of which is Pk. Hence the path connected components of Qk are dom(Ek,1, Ak), . . . , dom(Ek,mk, Ak), and since they are closed and disjoint in the compact set X, they are, in fact, compact and positively separated. Since Ek,l ⊆ Pk, we have d(Ek,l, Ak) ≥ d(Pk, Ak) > 0 for each k and each l. Hence, by Lemma 5.3, the sets dom(Ek,l, Ak) are also locally path connected. By Lemma 4.8, we have Ek,l & dom(Ek,l, Ak) for each l = 1, . . . , mk and each k, because there are points x with d(x, Ek,l) = r/4 for r = d(Ek,l, Ak) and each such point is in dom(Ek,l, Ak), but not in Ek,l. Consequently, Theo- rem 3.2 implies that 2dom(Ek,l,Ak) ≈ I ∞. Using the fact that the sets dom(Ek,1, Pk), dom(Ek,2, Pk),. . . , dom(Ek,mk, Pk) are positively separated (and also the fact that l=1(C ∩dom(Ek,l, Ak)) as a disjoint union of closed sets contained in the components dom(Ek,l, Ak)), we can easily verify that Yk with the Hausdorff metric is homeomorphic to the fi- , endowed with the uniform Hausdorff metric. (cid:3) any closed set C in Yk has the unique decomposition C =Smk nite product spaceQmk Therefore Yk ≈ (I ∞)mk ≈ I ∞, and hence Y ≈ I ∞. l=1 2dom(Ek,l,Ak) Ek,l Ek,l 6. Continuity of the Dom mapping Proposition 6.1. Let X be a convex subset of a uniformly convex normed space, and let (Pk)k∈K be a tuple of nonempty and positively separated sets in X, that is, inf{d(Pk, Pj) : j, k ∈ K, j 6= k} > 0. Assume that condition (3) holds (with the that the distance between each x ∈ X and each Pk, k ∈ K, is attained. For each same ρ) for each k ∈ K with Pk and Ak =Sj6=k Pj instead of P and A. Suppose k ∈ K let Qk = dom(Pk,Sj6=k Pj) and Yk = {C : C is closed and Pk ⊆ C ⊆ Qk}. Let Y = Qk∈K Yk, endowed with the uniform Hausdorff metric eD defined by eD((Sk)k∈K, (S′ k) : k ∈ K}. Then Dom maps Y into itself and is uniformly continuous there. k)k∈K) = sup{D(Sk, S′ Proof. First, note that for any nonempty subsets P, A, B of X, if A ⊆ B, then dom(P, B) ⊆ dom(P, A). Now let S = (Sk)k∈K ∈ Y be given. By the defini- tion of Yk, we have Pk ⊆ Sk for each k ∈ K, so Wk := dom(Pk,Sj6=k Sj) ⊆ dom(Pk,Sj6=k Pj) = Qk by the definition of Qk. Since the k-th component of W := Dom(S) is the closed subset Wk, we conclude that (Pk)k∈K ⊆ Dom(S) ⊆ (Qk)k∈K, and hence Dom maps Y into itself. j) ≤ sup{D(Sj, S′ correspond to ǫ in Theorem 4.3. Let S = (Sk)k∈K, S′ = (S′ Now let ǫ > 0 satisfy 12ǫ ≤ inf{d(Pk, Pj) : j, k ∈ K, j 6= k}, and let ∆ k)k∈K be any two tuples k) < ∆ for each j) ≤ ∆ for each k ∈ K, because in Y satisfying eD(S, S′) < ∆. As a result, we also have D(Sk, S′ k ∈ K. This implies that D(Sj6=k Sj,Sj6=k S′ D(Sj6=k Sj,Sj6=k S′ If x ∈ Sj6=k Sj, then x ∈ Sj for some j 6= k, so x ∈ Qj = dom(Pj,Si6=j Pi) ⊆ dom(Pj, Pk). Hence d(x, Pk) ≥ 6ǫ by Lemma 4.8. Thus d(Pk,Sj6=k Sj) ≥ 6ǫ for each k ∈ K. Hence all the conditions of Theorem 4.3 are satisfied (here P = Pk = P ′, A = Sj6=k Sj, A′ = Sj6=k S′ (3) of Theorem 4.3 for A follows from condition (3) for Ak = Sj6=k Pj because j and condition j) : j 6= k}. Fix k ∈ K. ZONE DIAGRAMS 15 Pj ⊆ Sj). Thus D(dom(Pk,Sj6=k Sj), dom(Pk,Sj6=k S′ the definition of Dom and eD, we have eD(Dom(S), Dom(S′)) ≤ ǫ, so Dom is indeed k)) < ǫ for each k ∈ K. By uniformly continuous on Y . (cid:3) 7. The main result Proposition 7.1. Let X be a compact and convex subset of a uniformly convex normed space, and let (Pk)k∈K be a finite tuple of finite sets in X which are pairwise disjoint. Suppose that for each k ∈ K, the site Pk has the emanation property. Then there exists a zone diagram with respect to these sites. Proof. Given the finite sites (Pk)k∈K, let Y be as in Proposition 6.1. Since Dom maps Y into itself and is continuous there by Propostion 6.1, and since Y is homeomorphic to the Hilbert cube by Proposition 5.4, we deduce the existence of a zone diagram from the Schauder fixed point theorem, as explained in the beginning of Section 3. (cid:3) Theorem 7.2. Let X be a compact and convex subset of a uniformly convex normed space, and let (Pk)k∈K be a finite tuple of compact sets in X which are pairwise disjoint. Suppose that for each k ∈ K, the site Pk has the emanation property. Then there exists a zone diagram with respect to these sites. Proof. Let (Pk)k∈K be the given finite tuple of pairwise disjoint compact sites. Let r = min{d(Pk, Pj) : j 6= k}. By the compactness of the sites, for each positive integer m and for each k ∈ K, there exists a finite subset Pm,k of Pk such that D(Pm,k, Pk) < 1/m. Note that d(Pm,k, Pm,j) ≥ r whenever j 6= k. For each m, let Rm = (Rm,k)k∈K be a zone diagram corresponding to the tuple (Pm,k)k∈K, the existence of which is guaranteed by Proposition 7.1. In other words, (5) Rm,k = dom(Pm,k,[j6=k Rm,j) ∀k ∈ K. Since the space of all nonempty compact subsets of X (endowed with the Haus- dorff metric) is compact [13], and since K is finite, we can find a convergent subsequence of the finite sites and the components of the corresponding zone di- agrams. Therefore for some subsequence (ml)∞ l=1 of positive integers the sequence (Rml,k)∞ l=1 converges to l=1 converges to some set Rk, and the sequence (Pml,k)∞ Pk by the definition of Pm,k. Hence Sj6=k Rml,j converges to Sj6=k Rj, because D(Sj6=k Rml,j,Sj6=k Rj) ≤ max{D(Rml,j, Rj) : j 6= k}. and x ∈Sj6=k Rml,j, then x ∈ Rml,j for some j 6= k. Therefore Let ǫ′ = r/12. Then 12ǫ′ ≤ min{d(Pml,k, Pml,j) : j 6= k}. Let k ∈ K. If j 6= k x ∈ dom(Pml,j,[i6=j Rml,i) ⊆ dom(Pml,j,[i6=j Pml,i) ⊆ dom(Pml,j, Pml,k). Hence d(x, Pml,k) ≥ 6ǫ′ by Lemma 4.8, and, as a result, d(Pml,k,Sj6=k Rml,j) ≥ 6ǫ′ for each k ∈ K and l. Thus d(Pk,Sj6=k Rj) ≥ 6ǫ′ and hence the positive distance condition in Theorem 4.3 holds. From (5) and the continuity of dom(·, ·) 16 EVA KOPECK ´A, DANIEL REEM, AND SIMEON REICH (Theorem 4.3) we conclude that R = (Rk)k∈K is a zone diagram with respect to (Pk)k∈K. (cid:3) The hypotheses of Theorem 7.2 are satisfied, in particular, when all the compact sites are contained in the interior of a closed ball (or another compact and convex set) in Rn. 8. Concluding remarks and open problems In this short section we describe several interesting questions and directions for further investigation. An interesting problem is whether the dominance region dom(P, A) is locally connected for general positively separated sets P and A. If so, then this will suggest an alternative method for proving the existence of a zone diagram with respect to compact sites. Second, it would be interesting to extend the existence result further, say to all normed spaces, without any compactness requirement on the sites and the subset X. The question of uniqueness is interesting too. It seems that the uniform convexity assumption on the norm is not sufficient even in the plane with two singleton sites, as shown in [15], but perhaps, following [15], uniform convexity combined with uniform smoothness of the norm will imply uniqueness in the infinite dimensional case too. It would also be of interest to establish our main theorem (Theorem 7.2) and auxiliary results (e.g., Lemmata 5.1 and 5.2) without imposing the condition of the emanation property on the points. Another interesting and natural problem is how to approximate a zone diagram in the setting described in this paper. It turns out that there is a way to do it, and a key point in carrying out this task is to use the algorithm for computing Voronoi diagrams of general sites in general spaces described in [19, 20]. For the sake of completeness, we include two pictures of zone diagrams in the plane with two different norms. The proof of Theorem 7.2 shows that one can approximate the given compact sites by finite subsets of them and then the corresponding zone diagram approximates the real one. Unfortunately, no error estimates are obtained in the proof, and it would indeed be of interest to find such estimates. Acknowledgements We thank Akitoshi Kawamura, Jir´ı Matousek, and Takeshi Tokuyama for pro- viding a copy of their recent paper [15] before posting it on the arXiv. The first author was supported by Grant FWF-P19643-N18. The third author was par- tially supported by the Israel Science Foundation (Grant 647/07), the Fund for the Promotion of Research at the Technion (Grant 2001893), and by the Technion President's Research Fund (Grant 2007842). All the authors thank Orr Shalit and the referee for their helpful suggestions and corrections. We are very grateful to Eva Goldman for kindly redrawing the figures so expertly. ZONE DIAGRAMS 17 Figure 5. A zone diagram of 5 sites, each with 4 points, in a square in (R2, ℓ6). Figure 6. A zone diagram of 3 sites in a square in (R2, ℓp), p = 3.14159, each with 3 points. References 1. T. Asano and D. Kirkpatrick, Distance trisector curves in regular convex distance met- rics, Proceedings of the 3rd International Symposium on Voronoi Diagrams in Science and Engineering (ISVD 2006), pp. 8 -- 17. 2. T. Asano, J. Matousek, and T. Tokuyama, The distance trisector curve, Adv. Math. 212 (2007), 338 -- 360, a preliminary version in STOC 2006, pp. 336-343. 3. , Zone diagrams: Existence, uniqueness, and algorithmic challenge, SIAM J. Com- put. 37 (2007), no. 4, 1182 -- 1198, a preliminary version in SODA 2007, pp. 756-765. 4. T. Asano and T. Tokuyama, Drawing equally-spaced curves between two points, Abstracts of the 14th Annual Workshop on Computational Geometry with a Focus on Open Problems, 2004, pp. 24 -- 25. 5. F. Aurenhammer, Voronoi diagrams - a survey of a fundamental geometric data structure, ACM Computing Surveys, vol. 3, 1991, pp. 345 -- 405. 6. J. Chun, Y. Okada, and T. Tokuyama, Distance trisector of segments and zone diagram of segments in a plane, Proceedings of the 4th International Symposium on Voronoi Diagrams in Science and Engineering (ISVD 2007), pp. 66 -- 73. 7. J. A. Clarkson, Uniformly convex spaces, Trans. Amer. Math. Soc. 40 (1936), 396 -- 414. 8. D. W. Curtis and R. M. Schori, Hyperspaces of Peano continua are Hilbert cubes, Fund. Math. 101 (1978), no. 1, 19 -- 38. 9. S. C. de Biasi, B. Kalantari, and I. Kalantari, Maximal zone diagrams and their computa- tion, Proceedings of the 7th international symposium on Voronoi Diagrams in Science and Engineering (ISVD 2010), pp. 171 -- 180. 10. J. Dugundji and A. Granas, Fixed point theory, Springer Monographs in Mathematics, Springer-Verlag, New York, 2003. 11. K. Goebel and S. Reich, Uniform convexity, hyperbolic geometry, and nonexpansive map- pings, Monographs and Textbooks in Pure and Applied Mathematics, vol. 83, Marcel Dekker Inc., New York, 1984. 12. C. Gold, The Voronoi Web Site, 2008, http://www.voronoi.com/wiki/index.php?title=Main Page. 18 EVA KOPECK ´A, DANIEL REEM, AND SIMEON REICH 13. A. Illanes and S. B. Nadler Jr, Hyperspaces. Fundamentals and recent advances, Monographs and Textbooks in Pure and Applied Mathematics, vol. 216, Marcel Dekker Inc., New York, 1999. 14. K. Imai, A. Kawamura, J. Matousek, Y. Muramatsu, and T. Tokuyama, Distance k- sectors and zone diagrams, Extended abstract in EuroCG 2009, pp. 191 -- 194. Available at http://www.cs.toronto.edu/∼kawamura/publ/090317. 15. A. Kawamura, J. Matousek, and T. Tokuyama, Zone diagrams in Euclidean spaces and in other normed spaces, Mathematische Annalen, to appear. Preliminary versions in SoCG 2010, pp. 216-221, arXiv 0912.3016 (2009). 16. J. Lindenstrauss and L. Tzafriri, Classical Banach spaces, II: Function spaces, Springer, Berlin, 1979. 17. J. Munkres, Topology, second ed., Prentice Hall, Upper Saddle River, NJ, 2000. 18. A. Okabe, B. Boots, K. Sugihara, and S. N. Chiu, Spatial tessellations: concepts and ap- plications of Voronoi diagrams, second ed., Wiley Series in Probability and Statistics, John Wiley & Sons Ltd., Chichester, 2000, with a foreword by D. G. Kendall. 19. D. Reem, An algorithm for computing Voronoi diagrams of general generators in general normed spaces, Proceedings of the sixth international symposium on Voronoi diagrams in science and engineering (ISVD 2009), pp. 144 -- 152. 20. 21. , Voronoi and zone diagrams, Ph.D. thesis, The Technion, Haifa, 2010. , The geometric stability of Voronoi diagrams with respect to small changes of the sites, (2011), Complete version in arXiv 1103.4125, Extended abstract in SoCG 2011, pp. 254-263. 22. D. Reem and S. Reich, Zone and double zone diagrams in abstract spaces, Colloquium Mathematicum 115 (2009), 129 -- 145, arXiv 0708.2668 (2007). 23. J. Schauder, Der Fixpunktsatz in Funktionalraumen, Studia Math. 2 (1930), 171 -- 180. 24. T. Zamfirescu, Typical starshaped sets, Aequ. Math. 36 (1988), 188 -- 200. Institute of Mathematics, Czech Academy of Sciences, Zitn´a 25, CZ-11567 Prague, Czech Republic, and Institut fur Analysis, Johannes Kepler Univer- sitat, Altenbergerstrasse 69, A-4040 Linz, Austria. E-mail address: [email protected] The Technion - Israel Institute of Technology, 32000 Haifa, Israel. Current address (July 2011): Department of Mathematics, University of Haifa, Mount Carmel, 31905 Haifa, Israel. E-mail address: [email protected] The Technion - Israel Institute of Technology, 32000 Haifa, Israel E-mail address: [email protected]
1406.6712	2	1406	2015-06-19T12:57:30	Stability of low-rank matrix recovery and its connections to Banach space geometry	[ "math.FA", "cs.IT", "cs.IT" ]	There are well-known relationships between compressed sensing and the geometry of the finite-dimensional $\ell_p$ spaces. A result of Kashin and Temlyakov can be described as a characterization of the stability of the recovery of sparse vectors via $\ell_1$-minimization in terms of the Gelfand widths of certain identity mappings between finite-dimensional $\ell_1$ and $\ell_2$ spaces, whereas a more recent result of Foucart, Pajor, Rauhut and Ullrich proves an analogous relationship even for $\ell_p$ spaces with $p < 1$. In this paper we prove what we call matrix or noncommutative versions of these results: we characterize the stability of low-rank matrix recovery via Schatten $p$-(quasi-)norm minimization in terms of the Gelfand widths of certain identity mappings between finite-dimensional Schatten $p$-spaces.	math.FA	math	STABILITY OF LOW-RANK MATRIX RECOVERY AND ITS CONNECTIONS TO BANACH SPACE GEOMETRY JAVIER ALEJANDRO CH ´AVEZ-DOM´INGUEZ AND DENKA KUTZAROVA Abstract. There are well-known relationships between compressed sensing and the ge- ometry of the finite-dimensional ℓp spaces. A result of Kashin and Temlyakov [20] can be described as a characterization of the stability of the recovery of sparse vectors via ℓ1- minimization in terms of the Gelfand widths of certain identity mappings between finite- dimensional ℓ1 and ℓ2 spaces, whereas a more recent result of Foucart, Pajor, Rauhut and Ullrich [16] proves an analogous relationship even for ℓp spaces with p < 1. In this paper we prove what we call matrix or noncommutative versions of these results: we charac- terize the stability of low-rank matrix recovery via Schatten p-(quasi-)norm minimization in terms of the Gelfand widths of certain identity mappings between finite-dimensional Schatten p-spaces. 1. Introduction A mathematical problem that appears often in real-world situations is the following: we wish to recover a high-dimensional vector x ∈ RN from a measurement Ax where A : RN → Rm is a linear map and m is smaller than N . As stated the problem of course cannot be solved, but that changes if we have the additional condition that the unknown vector x is sparse, i.e. it has a small number of non-zero coordinates. This is the subject matter of compressed sensing, a very active area of research with numerous applications; the book [17] is a recent comprehensive reference. Formally, this sparse recovery problem can be stated as (1.1) minkxk0 subject to Ax = y, where k·k0 represents the number of nonzero coordinates of a vector. This is an NP-hard [27] and non-convex problem, so we are interested in conditions (especially on the map A) that would allow us to solve an easier problem and still arrive to the right solution. In that spirit, a basic technique in compressed sensing is that of ℓ1-minimization: if the vector x is sparse enough, then minimizing kx′kℓ1 over all vectors x′ so that Ax′ = Ax actually allows us to recover x. Formally, instead of problem (1.1) we are considering its convex relaxation (1.2) minkxkℓ1 subject to Ax = y. Aditionally, one can consider the analogous problem of ℓp-minimization. In practice the unknown vectors are not necessarily sparse, but are close to sparse ones. Thus for any method of recovery it is of utmost importance to investigate its stability, that is, having a control on the distance between the original vector and its reconstruction in terms of the distance from the original vector to the sparse vectors. It turns out that the 1 2 JAVIER ALEJANDRO CH ´AVEZ-DOM´INGUEZ AND DENKA KUTZAROVA stability of sparse vector recovery through ℓp-minimization has connections to the Banach- space geometry of finite-dimensional ℓp-spaces. More generally, it is known that there are connections between recovery -- in particular the compressed sensing model -- and geometric quantities called Gelfand widths, see e.g. [28, 10, 9, 20]. In many practical situations, there is extra structure in the space of unknown vectors. A good example is the famous matrix completion problem (also known as the Netflix problem), where the unknown is a matrix and the measurement map gives us a subset of its entries. In this case sparsity gets replaced by the more natural condition of having low rank, and the last few years have witnessed an explosion of work in this area. In what follows, MN will denote the space of N × N real-valued matrices. We now consider a linear operator A : MN → Rm, and a fixed vector y ∈ Rm. The low-rank recovery problem can thus be stated as the problem of finding the solution to (1.3) min rank(X) subject to AX = y. This is again an NP-hard problem, so once again we would like to replace it by another one which is simpler to solve but has the same solution. In noncommutative functional analysis the Schatten p-spaces are usually considered to be the counterparts of the classical ℓp spaces (recall that the Schatten p-norm of a matrix X is the ℓp-norm of its vector of singular values), so from that point of view it is natural to wonder whether the Schatten p-norm minimization approach can work in the matrix context. We would like to consider operators A for which the previous problem is equivalent to (1.4) minkXkS1 subject to AX = y. Where kXkSp denotes the Schatten p-norm of the matrix X ∈ MN . This has already been studied in several situations of interest, with the idea going back to the Ph.D. thesis of M. Fazel [13]. Schatten 1-norm (also known as nuclear norm) minimization in the particular case of the matrix completion problem was studied by Cand`es and Recht [4] (and later on Cand`es and Tao [7] gave optimality results quantifying the minimum number of entries needed to recover a matrix of low rank exactly by any method whatsoever, and showed that nuclear norm minimization is nearly optimal). Plenty of concepts from the classical theory of compressed sensing have found matrix counterparts: Cand`es and Recht [4] use the idea of coherence; Recht, Fazel and Parrilo [32] used the matrix version of the restricted isometry property [32]; whereas both Recht, Xu and Hassibi [33] and Fornasier, Rauhut and Ward [14] consider null-space conditions; the spherical section property was used by Dvijotham and Fazel [11] and Oymak, Mohan, Fazel and Hassibi [29]. One thing that does not appear to have been explicitly studied in the matrix context is the aforementioned relationship to Gelfand widths. Recall that the Gelfand k-width of a subset K of a normed space E is defined as dk(K, E) := inf(cid:26) sup x∈K∩LkxkE : L subspace of E with codim(L) ≤ k(cid:27) . A closely related concept that is more commonly used in Banach space theory is that of a Gelfand number : if T : X → Y is a linear operator between normed spaces, its k-Gelfand STABILITY OF LOW-RANK MATRIX RECOVERY 3 number is defined by ck(T ) := inf( sup x∈L,kxk≤1 kT xk : L subspace of X with codim(L) < k) . The speed of convergence to zero of the sequence of Gelfand numbers (ck(T ))∞ k=1 is a measure of the compactness of the operator T , and is an example of a sequence of s-numbers; see [30, 22] for more details. In the cases under consideration in this paper the concepts of Gelfand numbers and Gelfand widths actually coincide (up to a small shift in the index), so we will freely use them both depending on the particular context. It should be mentioned that there is a general concept of Gelfand width for a linear map that is not always the same as the corresponding Gelfand number (see [31, Sec. 6.2.6] for the details), but both concepts do coincide in nice situations (see [12]). The work of Kashin and Temlyakov [20] made more precise the already-known connection between compressed sensing and the Kashin-Garnaev-Gluskin [19, 18] result that calculates the m-Gelfand numbers of the identity map from ℓN 1 to ℓN 2 , namely cm+1(id : ℓN 1 → ℓN 2 ) ≤ Cr1 + log(N/m) m . In a nutshell, the main result of Kashin and Temlyakov shows that the stability of sparse recovery via ℓ1-minimization is equivalent to the kernel of the measurement map being a "good" subspace where the Gelfand number of a certain order is achieved. This idea was taken further by Foucart, Pajor, Rauhut and Ullrich [16], who used compressed sensing ideas to calculate the Gelfand numbers of identity maps from ℓN for 0 < p ≤ 1, p < q ≤ 2. In this paper, we prove matrix versions of the aforementioned theorems, relating the stability of low-rank matrix recovery to the Gelfand numbers of identity maps between finite-dimensional Schatten p-spaces. As far as we know the only part of our results that is already written down in the literature is the following analogue of the Kashin-Garnaev- Gluskin result due to Carl and Defant [8, p. 252], namely the calculation of the m-Gelfand numbers of the identity map from SN p to ℓN q 1 to SN cm(id : SN 1 → SN 2 : for 1 ≤ m ≤ N 2, m(cid:27)1/2 2 ) ≍ min(cid:26)1, N . Here and in the rest of the paper, the symbol ≍ means that the quantities on the left and the right are equivalent up to universal constants. If we want to emphasize the dependance of the constants on some parameters, those will appear as subindices of the equivalence symbol (≍p,q, for example). The rest of this paper is organized as follows. In section 2 we introduce our notation and state several known results that will be needed in the sequel. In section 3 we show the first relationships between the stability of low-rank matrix recovery and the geometry of Banach spaces, by proving a matrix version of the Kashin-Temlyakov theorem. Section 4 contains a technical result, a matrix version of the main theorem from [15] that gives conditions on the measurement map A that guarantee the stability of the Schatten p-minimization scheme. A very similar theorem was recently obtained independently by Liu, Huang and Chen [25], though our proof is different and we require a weaker hypothesis. In the final 4 JAVIER ALEJANDRO CH ´AVEZ-DOM´INGUEZ AND DENKA KUTZAROVA section, the technical result from Section 4 is used to calculate the Gelfand numbers of the identity maps from SN for 0 < p ≤ 1, p < q ≤ 2 in the spirit of the work of Foucart, Pajor, Rauhut and Ullrich. p to SN q 2. Notation and preliminaries to rectangular ones. For p > 0 we will denote by SN Schatten p-quasi-norm, given by In this paper we will only consider square matrices, but all the results can be adapted p the space of N × N matrices with the σip(cid:17)1/p kXkSp =(cid:16) N , Xi=1 i=1 is the vector of singular values of the matrix X. Similarly, SN where (σi)N the space of N × N matrices with the weak-Schatten-p-quasi-norm given by p,∞ will denote k1/pσ∗ k i=1 is the non-increasing rearrangement of (σi)N i=1. For any quasi-normed space kXkSp,∞ = max 1≤k≤N where (σ∗ X, BX will denote its unit ball. i )N We will need to consider the best s-rank approximation error in the Schatten p-quasi- norm, It is well known that the infimum is actually attained at the s-spectral truncation Y = X[s] (that is, keeping only the s largest singular values in the singular value decomposition). ρs(X)Sp := inf(cid:8)kX − Y kSp : rank(Y ) ≤ s(cid:9). Given a linear map A : MN → Rm and a vector y ∈ Rm, for 0 < p ≤ 1 we will denote by ∆p(y) a solution to minimize kZkSp subject to AZ = y. That is, ∆p is the Schatten p-quasi-norm minimization reconstruction map. The map ∆p of course depends on the measuring map A, but for simplicity we do not make this dependence explicit in the notation. 2.1. The Restricted Isometry Property. The Restricted Isometry Property (RIP) for a linear map A : RN → Rm was introduced by Cand`es and Tao [5], and quickly became a key concept in the analysis of sparse recovery via ℓp-norm minimization. The s-order restricted isometry constant of such a map is the smallest δ > 0 such that for every vector x ∈ RN of sparsity at most s, (1 − δ)kxk2 ℓ2 ≤ kAxk2 ℓ2 ≤ (1 + δ)kxk2 ℓ2 . The importance of the RIP stems from the fact that small restricted isometry constants imply exact recovery via ℓp-quasi-norm minimization for 0 < p ≤ 1, and it should be noted that it is well known that random choices of the matrix A give small RIP constants of order s, as long as m is at least of the order of s ln(eN/s) [6, 1, 26]. The version of the RIP for matrix recovery was introduced by Recht, Fazel and Parrilo [32], and is as follows: a linear map A : MN → Rm is said to have the Restricted Isometry Property of rank s with constant δ > 0 if for every matrix Z ∈ MN of rank at most s, (1 − δ)kZk2 S2 ≤ kAZk2 ℓ2 ≤ (1 + δ)kZk2 S2 . STABILITY OF LOW-RANK MATRIX RECOVERY 5 The best such constant is denoted by δs(A). result follows from [3, Thm. 2.3], and will be very important for us in the sequel. Just as in the vector case, random constructions give small RIP constants. The next Theorem 2.1. Given a prescribed δ ∈ (0, 1), there is a constant Cδ such that if the entries of the map A (seen a matrix with respect to the canonical bases in MN and Rm) are indepen- dent gaussians with mean zero and variance 1/m, then with positive (even overwhelming) probability δs(A) ≤ δ holds provided that (2.1) m ≥ CδsN. 3. A noncommutative Kashin-Temlyakov theorem We will prove a matrix version of the Kashin-Temlyakov characterization of the stability of sparse recovery via ℓ1-norm minimization in terms of widths. To this end, we define three properties modeled after the ones studied in [20]. Definition 3.1. Let N 2 > m and A : MN → Rm a linear operator. We say that A has a: (a) Matrix Strong Compressed Sensing Property (MSCSP) if for any X ∈ MN we have kX − ∆1(AX)kS2 ≤ Cs−1/2ρs(X)S1 for s ≍ m/N . (b) Matrix Weak Compressed Sensing Property (MWCSP) if for any X ∈ MN we have kX − ∆1(AX)kS2 ≤ Cs−1/2 kXkS1 for s ≍ m/N . (c) Matrix Width Property (MWP) if for any X ∈ ker(A), kXkS2 ≤ C(N/m)−1/2 kXkS1 . (3.1) Notice that the MSCSP is a weakening of condition (i) in [29, Lemma 8], since we are only considering X ′ = ∆1(AX). Also, the name of the MWP comes from its clear relationship to the definition of the Gelfand numbers/widths. The following theorem is a matrix version of the Kashin-Temlyakov theorem [20, Thm. 2.2]: Theorem 3.2. For a linear operator A : MN → Rm, the MSCSP, MWCSP and MWP are equivalent (up to a change in the constants). Proof. The MSCSP trivially implies the MWCSP, since ρs(X)S1 ≤ kX − 0kS1 = kXk. Assume that A has the MWSCSP. Given X ∈ ker(A), note that AX = 0 = A0, so clearly 0 = ∆1(0) = ∆1(AX) and thus from the MWSCSP we have kXkS2 ≤ Cs−1/2 kXkS1, giving the MWP. Assume now that we have the MWP, that is, that equation (3.1) holds. If s < 1 4 C −2N/m, from [29, Thm. 2] (which is a matrix version of [20, Thm. 2.1]) we obtain kX − ∆1(AX)kS1 ≤ C ′ρs(X)S1 for C ′ = 2(1− 2pC 2sm/N )−1. Since X − ∆1(AX) ∈ ker A, the previous equation together with (3.1) imply the MSCSP. (cid:3) 6 JAVIER ALEJANDRO CH ´AVEZ-DOM´INGUEZ AND DENKA KUTZAROVA The aforementioned Kashin-Temlyakov theorem says, in a nutshell, that the stability of sparse-vector recovery via ℓ1-minimization has limits imposed by the geometry of Banach spaces encoded in the appropriate Gelfand widths. In the previous proposition, we showed a similar relationship relating the stability of low-rank recovery via nuclear norm minimization with some other Gelfand widths. As in the vector case, following [17, Cor. 10.6], there is a relationship between the geometry of SN 1 and the stability of compressed sensing by any method. See Theorem 5.5 below for the precise statement. 4. Stability of low-rank matrix recovery through Schatten p quasi-norm minimization In this technical section we prove a general result (a matrix version of the main theorem in [15]) that gives RIP-style conditions on the measuring map A that guarantee the stability of the Schatten p-norm minimization scheme. For that we will need some notation: Let αs, βs ≥ 0 be the best constants in the inequalities αs kZkS2 ≤ kAZkℓ2 ≤ βs kZkS2 , rank(Z) ≤ s. The results will be stated in terms of a quantity invariant under the change A ← cA, namely Note that this constant is related to the RIP constant, in fact γ2s := β2 2s 2s ≥ 1. α2 γ2s = 1 + δ2s 1 − δ2s . Unlike in the rest of the paper, we will consider the more general situation of approximate recovery when measurements are moderately flawed, namely the problem kAZ − ykℓ2 ≤ β2s · θ. minimize kZkSp subject to (Pp,θ) For simplicity, we will write (Pp) instead of (Pp,0). Note that by a compactness argument, a solution of (Pp,θ) exists for any 0 < p ≤ 1 and any θ ≥ 0. The following theorem is a matrix version of [15, Thm. 3.1]. It gives conditions (in the spirit of the RIP) that guarantee not only the stability but also the robustness (that is, resistance to errors in the measurements) of the Schatten p-quasi-norm-minimization for low-rank matrix recovery. Theorem 4.1. Given 0 < p ≤ 1, if for some integer t ≥ s s(cid:19)1/p−1/2 (4.1) γ2t − 1 < 4(√2 − 1)(cid:18) t then a solution X ∗ of (Pp,θ) approximates the original matrix X with errors (4.2) (4.3) kX − X ∗kSp ≤ C1ρs(X)Sp + D1 · s1/p−1/2 · θ, kX − X ∗kS2 ≤ C2 ρs(X)Sp t1/p−1/2 + D2 · θ, where the constants C1, C2, D1 and D2 depend only on p, γ2t and the ratio s/t. STABILITY OF LOW-RANK MATRIX RECOVERY 7 Proof. We will need to recall some properties of the Sp-quasi-norm. Namely, for any ma- trices U and V , (4.4) kUkS1 ≤ kUkSp , kUkSp ≤ N 1/p−1/2 kUkS2 , STEP 1: Consequence of the assumption on γ2t. kU + V kp Sp ≤ kUkp Sp + kV kp Sp . We will consider certain matrix decompositions similar to the ones in [21]. Consider the singular value decomposition of X, given by X = U diag(λi(X))V T where U , V are unitary matrices and λ(X) = (λ1(X), . . . , λN (X)) are the singular values of X arranged in decreasing order. For any matrix Z ∈ MN , we will consider a block decomposition of Z with respect to X as follows: let U T ZV have the block form U T ZV =(cid:18)Z11 Z12 Z21 Z22(cid:19) where Z11, Z12, Z21, Z22 are of sizes s × s, s × (N − s), (N − s) × s, (N − s) × (N − s), respectively. We then decompose Z as Z = Z(s) + Z c (s) where Z21 Z(s) = U(cid:18)Z11 Z12 0 (cid:19) V T (s) = U(cid:18)0 0 Z22(cid:19) V T Z c 0 Furthermore, we now consider the singular value decomposition of Z22 given by Z22 = P diag(λ(Z22))QT with P and Q being (N − s) × (N − s) unitary matrices, and λ(Z22) is the vector of the N − s singular values of Z22 arranged in decreasing order. We decompose λ(Z22) as a sum of vectors ZTi, each of sparsity at most t, where T1 corresponds to the locations of the t largest entries of λ(Z22), T2 to the locations of the next t largest entries, and so on. For i ≥ 1 we now define ZTi = U(cid:18)0 0 P diag(λTi(Z22))QT(cid:19) V T , 0 and denote ZT0 := Z(s). We first observe that kZT0k2 S2 + kZT1k2 S2 = kZT0 + ZT1k2 S2 ≤ 1 2t kA(ZT0 + ZT1)k2 α2 ℓ2 1 α2 1 α2 = (4.5) 2t(cid:10)A(Z − ZT2 − ZT3 − ··· ),A(ZT0 + ZT1)(cid:11) 2t(cid:10)AZ,A(ZT0 + ZT1)(cid:11) + Let us renormalize the vectors −ZTk and ZT0 so that their S2-norms equal one by setting and Y0 := ZT0/kZT0kS2. We then obtain, using the polarization 2t Xk≥2(cid:2)(cid:10)A(−ZTk ),AZT0(cid:11) +(cid:10)A(−ZTk ),AZT1(cid:11)(cid:3) Yk := −ZTk /kZTkkS2 1 α2 = 8 JAVIER ALEJANDRO CH ´AVEZ-DOM´INGUEZ AND DENKA KUTZAROVA identity (cid:10)A(−ZTk ),AZT0(cid:11) kZTkkS2 kZT0kS2 1 1 = hAYk,AY0i = 4(cid:2)β2 2t kYk + Y0k2 ≤ 4(cid:2)kA(Yk + Y0)k2 2t kYk − Y0k2 S2 − α2 ℓ2 − kA(Yk − Y0)k2 ℓ2(cid:3) 2t − α2 [β2 S2(cid:3) = 2t]. 1 2 An analogous argument with T1 in place of T0 allows us to conclude β2 2t − α2 2t 2 (4.6) On the other hand, we have (cid:10)A(−ZTk ),AZT0(cid:11) +(cid:10)A(−ZTk ),AZT1(cid:11) ≤ kZTkkS2(cid:2)kZT0kS2 + kZT1kS2(cid:3). (4.7) (cid:10)AZ,A(ZT0 +ZT1)(cid:11) ≤ kAZkℓ2·kA(ZT0 + ZT1)kℓ2 ≤ kAZkℓ2·β2t(cid:2)kZT0kS2 +kZT1kS2(cid:3). (cid:2)kZT0kS2 + kZT1kS2(cid:3). Substituting the inequalities (4.6) and (4.7) into (4.5) we have γ2t − 1 2 Xk≥2 S2 ≤  γ2t β2t kAZkℓ2 + kZTkkS2 S2 + kZT1k2 kZT0k2 inequality is If we set c := kAZkℓ2 · γ2t/β2t, d := (γ2t − 1)/2 and Σ = Pk≥2 kZTkkS2 S2 − (c + dΣ)kZT1kS2 ≤ 0, S2 − (c + dΣ)kZT0kS2 + kZT1k2 kZT0k2 , the previous or equivalently, c + dΣ (cid:20)kZT0kS2 − (cid:21)2 by getting rid of the second squared term, this easily implies 1 + √2 +(cid:20)kZT1kS2 − (cid:21)2 c + dΣ c + dΣ 2 2 (4.8) (c + dΣ)2 2 . ≤ c + dΣ √2 2 = + kZT0kS2 ≤ kZT0kSp ≤ s1/p−1/2 kZT0kS2 ≤ s1/p−1/2 1 + √2 2 (c + dΣ). By Holder's inequality (see (4.4)) we get 2 p Sp (c + dΣ). . Therefore, (4.9) We now proceed to bound Σ. For k ≥ 2, let η, η′ be singular values of ZTk , ZTk−1, respectively. By definition, we must have η ≤ η′. Raising to the p-th power and averaging , and hence η2 ≤ t−2/p(cid:13)(cid:13)ZTk−1(cid:13)(cid:13) over all singular values of ZTk−1, ηp ≤ t−1(cid:13)(cid:13)ZTk−1(cid:13)(cid:13) . Adding over all singular values of ZTk and taking the square root, this yields kZTkkS2 ≤ t1/2−1/p(cid:13)(cid:13)ZTk−1(cid:13)(cid:13)Sp kZTkkSp ≤ t1/2−1/p"Xk≥1 Σ =Xk≥2 2β2t · kAZkℓ2 · s1/p−1/2 + µ ·(cid:13)(cid:13)(cid:13) kZTkkS2 ≤ t1/2−1/pXk≥1 Combining the above inequality with (4.9), we obtain = t1/2−1/p(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)Z(s)(cid:13)(cid:13)Sp ≤ (s)(cid:13)(cid:13)(cid:13)Sp (s)(cid:13)(cid:13)(cid:13)Sp Sp#1/p kZTkkp (4.10) 2 Sp Z c Z c λ . . STABILITY OF LOW-RANK MATRIX RECOVERY 9 where the constants λ and µ are given by λ := (1 + √2)γ2t and µ := 1 4 (1 + √2)(γ2t − 1)(cid:16) s t(cid:17)1/p−1/2 . Note that the assumption on γ2t translates into the inequality µ < 1. STEP 2: From now on let Z := X − X ∗. Because X ∗ is a minimizer of (Pp,θ), we have (4.11) kX ∗kp Sp ≤ kXkp Sp . From [21, Lemma 2.2], whenever B, C ∈ MN satisfy BT C = 0 and BC T = 0 one has kB + Ckp (4.12) . In particular, note that X c Z c Sp Sp p p p . +(cid:13)(cid:13)(cid:13) Sp (s)(cid:13)(cid:13)(cid:13) and (cid:13)(cid:13)(cid:13) X(s) − Z c Sp (s)(cid:13)(cid:13)(cid:13) p Sp =(cid:13)(cid:13)X(s)(cid:13)(cid:13) +(cid:13)(cid:13)(cid:13) Sp (s)(cid:13)(cid:13)(cid:13) From the p-triangle inequality (see (4.4)), since = kBkp kXkp Sp p Sp Sp + kCkp =(cid:13)(cid:13)X(s)(cid:13)(cid:13) X(s) − Z c (s) = X − Z − X c (s) + Z(s) = X ∗ − X c (s) + Z(s), Together with (4.11) and both equalities in (4.12), this yields we get p p Sp Z c p Sp (cid:13)(cid:13)(cid:13) X(s) − Z c Sp ≤ kX ∗kp +(cid:13)(cid:13)(cid:13) (s)(cid:13)(cid:13)(cid:13) (s)(cid:13)(cid:13)(cid:13) +(cid:13)(cid:13)(cid:13) Sp ≤(cid:13)(cid:13)X(s)(cid:13)(cid:13) After a cancellation and noticing that (cid:13)(cid:13)(cid:13) (s)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13) (s)(cid:13)(cid:13)(cid:13) STEP 3: Error estimates. Sp ≤ 2ρs(X)p (cid:13)(cid:13)X(s)(cid:13)(cid:13) +(cid:13)(cid:13)(cid:13) (4.13) X c p Sp Z c Sp Sp p p We first note the bound p Sp . p X c Sp (s)(cid:13)(cid:13)(cid:13) (s)(cid:13)(cid:13)(cid:13) Sp(cid:13)(cid:13)(cid:13) p +(cid:13)(cid:13)Z(s)(cid:13)(cid:13) (s)(cid:13)(cid:13)(cid:13) Sp p X c X c = ρs(X)p Sp , we obtain p Sp . +(cid:13)(cid:13)Z(s)(cid:13)(cid:13) p Sp . +(cid:13)(cid:13)Z(s)(cid:13)(cid:13) kAZkℓ2 = kAX − AX ∗kℓ2 ≤ kAX − ykℓ2 + ky − AX ∗kℓ2 ≤ 2β2s · θ. For the Sp-error, we combine the estimates in (4.10) and (4.13) to obtain Sp ≤ 2ρs(X)p As a consequence of µ < 1, we have Z c Sp (cid:13)(cid:13)(cid:13) p Sp . Z c + λp · s1−p/2 · θp + µp ·(cid:13)(cid:13)(cid:13) (s)(cid:13)(cid:13)(cid:13) 1 − µp · s1−p/2 · θp. λp + Sp p Sp ≤ 2 1 − µp ρs(X)p p (s)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13) (s)(cid:13)(cid:13)(cid:13) Z c 10 JAVIER ALEJANDRO CH ´AVEZ-DOM´INGUEZ AND DENKA KUTZAROVA Using the estimate (4.10) once again, we can derive that p p 2 Z c p Sp +(cid:13)(cid:13)(cid:13) (s)(cid:13)(cid:13)(cid:13) kZkp Sp ≤(cid:13)(cid:13)Z(s)(cid:13)(cid:13) 1 − µp (1 + µp)ρs(X)p ≤ ≤ 21−1/p" (1 − µp)1/p 21/p Sp ≤ (1 + µq) ·(cid:13)(cid:13)(cid:13) 2λp + Sp Z c (s)(cid:13)(cid:13)(cid:13) 1 − µp · s1−p/2 · θp (1 + µp)1/pρs(X)Sp + + λp · s1−p/2 · θp Sp 21/pλ (1 − µp)1/p · s1/p−1/2 · θ#p where we have used the inequality (ap + bp)1/p ≤ 21/p−1(a + b) for a, b ≥ 0. The desired estimate (4.2) follows with C1 := 22/p−1(1 + µp)1/p (1 − µp)1/p , D1 := 22/p−1λ (1 − µp)1/p . 1/2 S2 kZTkk2  For the S2-error, let us observe that the bound in (4.8) also holds if we replace kZT0kS2 by kZT1kS2, and hence kZkS2 = Xk≥0 where ν = (λ + 1 − √2)/2. We also have that Σ ≤ t1/2−1/p(cid:13)(cid:13)(cid:13) kZTkkS2 ≤ (1 + √2) · (c + dΣ) + Σ ≤ ν · Σ + 2λ · θ, 1 − µp · s1−p/2 · θp(cid:21)1/p ≤Xk≥0 (s)(cid:13)(cid:13)(cid:13)Sp ≤ t1/2−1/p(cid:20) ≤ t1/2−1/p21/p−1" 1 − µp ρs(X)p (1 − µp)1/p 21/p Z c λp + Sp 2 and hence we conclude that kZkS2 ≤ νt1/2−1/p21/p−1" 21/p (1 − µp)1/p This gives the estimate (4.3) with ρs(X)Sp + C2 = 22/p−2(λ + 1 − √2) (1 − µp)1/p , D2 = λ ρs(X)Sp + (1 − µp)1/p · s1/p−1/2 · θ# , (1 − µp)1/p · s1/p−1/2 · θ# + 2λθ. 21/p−2λ(λ + 1 − √2) + 2λ. λ (1 − µp)1/p (cid:3) As consequences of Theorem 4.1, we obtain two corollaries that are matrix versions of the ones in [15]. The first one corresponds to the case of exact recovery. Corollary 4.2. Given 0 < p ≤ 1, if γ2t − 1 < 4(√2 − 1)(cid:18) t s(cid:19)1/p−1/2 for some integer t ≥ s, then every rank s matrix is exactly and stably recovered by solving (Pp). The second one deals with the special case of nuclear norm minimization. STABILITY OF LOW-RANK MATRIX RECOVERY 11 Corollary 4.3. Under the assumption that γ2s < 4√2 − 3 ≈ 2.6569, every rank s matrix is exactly and stably recovered by solving (P1). This last Corollary is clearly related to existing results on the RIP, it corresponds to the condition δ2s < 2(3 − √2)/7 ≈ 0.4531. Note that for the specific case of p = 1 this is in fact δ2s < 1/√2 ≈ 0.7071. Another recent result similar to our Theorem 4.1 is [25, condition is not the best possible: a very recent result of Cai and Zhang [2] shows that the optimal condition to have exact recovery of rank s matrices via nuclear norm minimization Thm. 6] (which in turn generalizes results of Lee and Bresler [24, 23]), where they get a conclusion of the same form as (4.3) but requiring a stronger hypothesis. Finally, note also that Theorem 2.1 guarantees the existence of maps A satisfying the hypothesis of Theorem 4.1. 5. The Gelfand widths of Sp-balls for 0 < p ≤ 1 In this section we calculate the Gelfand numbers p → SN q ) cm(id : SN for 0 < p ≤ 1, p < q ≤ 2. This can be considered as a noncommutative version of the results from [16], where they use compressed sensing ideas to calculate the corresponding Gelfand numbers Inspired by their approach, our proof is based on low-rank matrix recovery ideas. cm(id : ℓN p → ℓN q ). Our main result is the following (compare to [16, Thm. 1.1]). Theorem 5.1. For 0 < p ≤ 1 and p < q ≤ 2, if 1 ≤ m < N 2, then m(cid:27)1/p−1/q m(cid:27)1/p−1/q q ) ≍p,q min(cid:26)1, q ) ≍p,q min(cid:26)1, and, if p < 1, Sp,∞ , SN dm(BN dm(BN Sp, SN N N . . Before the proof, let us go through some preliminaries. Recall that it is classical to show that, for q > p, (5.1) (5.2) ρs(X)Sq ≤ ρs(X)Sq ≤ 1 s1/p−1/q kXkSp , Dp,q s1/p−1/q kXkSp,∞ , Dp,q := (q/p − 1)−1/q. 5.1. Lower bounds. In this section we prove a result that will easily imply the desired lower bounds in Theorem 5.1. It is a matrix version of [16, Thm. 2.1] and, just like in their result, we note that the restriction q ≤ 2 is not imposed here. Proposition 5.2. For 0 < p ≤ 1 and p < q ≤ ∞, there exists a constant cp,q > 0 such that dm(BN Sp, SN q ) ≥ cp,q min(cid:26)1, N m(cid:27)1/p−1/q 12 JAVIER ALEJANDRO CH ´AVEZ-DOM´INGUEZ AND DENKA KUTZAROVA Proof. With c = (1/2)2/p−1/q and µ = min{1, N/2m}, we are going to prove that We proceed by contradiction, assuming that dm(BN existence of a linear map A : MN → Rm such that for all V ∈ ker(A) \ {0}, q ) < cµ1/p−1/q. This implies the dm(BN Sp, SN q ) ≥ cµ1/p−1/q. Sp, SN kV kSq < cµ1/p−1/q kV kSp . For a fixed V ∈ ker(A) \ {0}, in view of the inequalities kV kSp ≤ N 1/p−1/q kV kSq and c ≤ (1/2)1/p−1/q , we derive 1 < (µN/2)1/p−1/q , so 1 ≤ 1/µ < N/2. We then define s := ⌊1/µ⌋ ≥ 1, so 2s < N and 1 2µ < s ≤ 1 µ . Now for V ∈ ker(A) \ {0}, (cid:13)(cid:13)V[2s](cid:13)(cid:13)Sp ≤ (2s)1/p−1/q(cid:13)(cid:13)V[2s](cid:13)(cid:13)Sq ≤ (2s)1/p−1/q kV kSq < c(2sµ)1/p−1/q kV kSp ≤ and therefore, using that kV kp , we conclude p Sp p Sp Sp 1 21/p kV kSp p =(cid:13)(cid:13)V[2s](cid:13)(cid:13) (cid:13)(cid:13)V[2s](cid:13)(cid:13) +(cid:13)(cid:13)V − V[2s](cid:13)(cid:13) Sp ≤(cid:13)(cid:13)V − V[2s](cid:13)(cid:13) p Sp . This means that A satisfies the sufficient conditions in [29, Thm. 3], which implies that Schatten p-quasinorm minimization gives exact recovery of rank s matrices. By well-known arguments (see, for example, the discussion after the statement of theorem 2.3 in [3]), this gives m ≥ N s > N 1 2µ ≥ N 2 1 µ ≥ N 2 2m N = m, a blatant contradiction. (cid:3) 5.2. Upper bounds. In this subsection we establish a result from which the desired upper bounds in Theorem 5.1 will follow easily. The proof relies on low-rank matrix recovery methods, and the reader will notice similarities with the proof of Theorem 4.1. It should be mentioned that the bound for the case p ≥ 1 follows easily from the result of Carl and Defant [8] mentioned in the introduction together with an interpolation argument, but the bound for the case p < 1 is new (as far as the authors know). Our result is a matrix version of [16, Thm. 3.2], but the essence of the argument can be traced back to Donoho [10, Thm. 9]. As in the case of the result of [16], it is interesting to note that even when p < 1, an optimal reconstruction map ∆ for the realization of the number Em(BN q ) can be chosen to be the S1-minimization mapping, at least when q ≥ 1. Theorem 5.3. For 0 < p < 1 and p < q ≤ 2, there exists a linear map A : MN → Rm such that, with r = min{1, q}, , SN Sp,∞ sup X∈BN p,∞ kX − ∆r(AX)kSq ≤ Cp,q minn1, where Cp,q > 0 is a constant that depends only on p and q. N mo1/p−1/q , STABILITY OF LOW-RANK MATRIX RECOVERY 13 Proof. Let C be the constant in (2.1) relative to the RIP associated with δ = 1/3, say. Case 1: m ≥ CN . We define s ≥ 1 as the largest integer smaller than m m . (5.3) CN Let t = 2s. It is then possible to find a linear map A : MN → Rm with δt(A) ≤ δ. In particular, we have δs(A) ≤ δ. Now, given Z := X − ∆r(AX) ∈ kerA, we decompose Z into matrices ZT1, ZT2, ZT3, . . . of rank at most s by taking the s largest singular values of Z for ZT1, then the next s largest ones for ZT2 and so on. CN , so that < s ≤ 2CN m This easily implies (cid:0)kZTkk2 S2 (5.4) kZTkkS2 ≤ Using the r-triangle inequality, we have , 1 r Sr /s(cid:1)1/r, i.e., k ≥ 2. /s(cid:1)1/2 ≤(cid:0)(cid:13)(cid:13)ZTk−1(cid:13)(cid:13) s1/r−1/2 (cid:13)(cid:13)ZTk−1(cid:13)(cid:13)Sr Sq ≤Xk≥1(cid:0)s1/q−1/2 kZTkkS2(cid:1)r ≤Xk≥1(cid:16) s1/q−1/2 √1 − δ kAZTkkℓ2(cid:17)r . r Sq kZTkkr ≤Xk≥1 Sq =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) kZkr Xk≥1 The fact that Z ∈ ker A implies AZT1 = −Pk≥2 AZTk . It follows that Xk≥2 s1/q−1/2(cid:17)r ZTk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Sq ≤ (cid:16) s1/q−1/2 √1 − δ (cid:17)r(cid:16)Xk≥2 ≤ 2(cid:16) s1/q−1/2 √1 − δ (cid:17)r Xk≥2 +(cid:16) s1/q−1/2 √1 − δ (cid:17)r ℓ2 ≤ 2(cid:16)r1 + δ 1 − δ kAZTkkℓ2(cid:17)r kAZTkkr kZkr We then derive, using the inequality (5.4), kZkr Sq ≤ 2(cid:16)r 1 + δ 1 − δ 1 s1/r−1/q(cid:17)r Xk≥1 kZTkkr Sr . In view of the choice δ = 1/3 and of (5.3), we deduce kAZTkkr ℓ2 kZTkkr S2 . Xk≥2 (5.5) kX − ∆r(AX)kSq ≤ 21/r√2(cid:16) 2CN m (cid:17)1/r−1/q kX − ∆r(AX)kSr . Moreover, in view of δ2s ≤ 1/3 and of Theorems 4.1 and 2.1, there exists a constant C1 > 0 such that (5.6) kX − ∆r(AX)kSr ≤ (C1)1/rρs(x)Sr . Finally, using (5.2) and (5.3), we have (5.7) m (cid:17)1/p−1/r Putting (5.5), (5.6), and (5.7) together, we obtain, for any x ∈ BN s1/p−1/r ≤ Dp,r(cid:16) 2CN ρs(X)Sr ≤ Dp,r . p,∞, kX − ∆r(AX)kSq ≤ 21/r√2C 1/r 1 Dp,r(cid:16) 2CN m (cid:17)1/p−1/q . 14 JAVIER ALEJANDRO CH ´AVEZ-DOM´INGUEZ AND DENKA KUTZAROVA Case 2: m ≤ CN . We simply choose the map A as the zero map. Then, for any X ∈ BN kX − ∆r(AX)kSq = kXkSq ≤ Dp,q kXkp,∞ ≤ Dp,q, Sp,∞ for some constant Dp,q > 0. This completes the proof. , we have (cid:3) Remark 5.4. When p = 1, the same proof but using inequality (5.1) instead of (5.2) gives the following: for 1 < q ≤ 2, there exists a linear map A : MN → Rm such that, sup X∈BN 1 kX − ∆1(AX)kSq ≤ Cq minn1, where Cq > 0 is a constant that depends only on q. N mo1−1/q , 5.3. Proof of theorem 5.1. Proof. First, an observation. As in the vector case, the simple inclusion BN implies Sp ⊆ BN Sp,∞ dm(BN Sp, SN q ) ≤ dm(BN Sp, SN Sp,∞, SN q ) and upper bounds for dm(BN q ), hence it suffices to show lower bounds for dm(BN q ). The lower bounds follow immediately from Proposition 5.2. When 0 < p < 1, the upper bounds follow from Theorem 5.3. For p = 1, the upper bound when 1 ≤ m ≤ N follows from the trivial inequality kXkSq ≤ kXkS1, whereas when N ≤ m ≤ N 2 it follows from Remark 5.4. , SN Sp,∞ (cid:3) 5.4. Relation to compressive widths. As promised after the proof of our matrix version of the Kashin-Temlyakov theorem, the relationship between the Banach space geometry of the finite-dimensional Schatten p-classes and matrix recovery goes beyond the norm minimization scheme. Below we use the notation from [17, Sec. 10.1]: the quantities Em and Em ada measure the worst-case reconstruction errors of optimal measurement/reconstruction schemes in the nonadaptive and adaptive settings, respectively. Theorem 5.5. For 0 < p ≤ 1 and p < q ≤ 2, if 1 ≤ m < N 2 then the adaptive and nonadaptive compressive widths satisfy Em ada(BSN p , SN Proof. Since −BSN dm(BSN p p p = BSN q ) ≤ Em , SN But now, since dm(BSN the proof. p , SN q ) ≍p,q min(cid:26)1, N m(cid:27)1/p−1/q . , [17, Thm. 10.4] implies q ) ≤ 21/pdm(BSN , SN q ) , SN p q ) ≍p,q Em(BSN + BSN and BSN p p ada(BSN p , SN p p ⊆ 21/pBSN q ) ≤ Em(BSN p → SN p , SN q ) = cm+1(id : SN q ), an appeal to Theorem 5.1 finishes (cid:3) In the ℓp case the lower estimate is of particular importance in compressed sensing, since it allows one to prove lower bounds for the number of measurements required to stably recover s-sparse vectors in RN . In the matrix case, that is no longer the case. Trying it only gives that (under certain conditions), the minimum number of measurements m STABILITY OF LOW-RANK MATRIX RECOVERY 15 required to stably recover rank s matrices in MN is ≥ CN s, which is not an improvement over the information-theoretical limit. The reason behind this is that, unlike in the ℓp case, there are compressed sensing algorithms (including norm minimization) that give stability with a number of measurements of that order [3]. Acknowledgements We would like to thank Rachel Ward for suggesting the reference [15], and also thank the Workshop in Analysis in Probability at Texas A&M University. The first author was partially supported by NSF grant DMS-1400588. References [1] Richard Baraniuk, Mark Davenport, Ronald DeVore, and Michael Wakin. A simple proof of the re- stricted isometry property for random matrices. Constr. Approx., 28(3):253 -- 263, 2008. [2] T. Tony Cai and Anru Zhang. Sparse representation of a polytope and recovery of sparse signals and low-rank matrices. IEEE Trans. Inform. Theory, 60(1):122 -- 132, 2014. [3] Emmanuel J. Cand`es and Yaniv Plan. Tight oracle inequalities for low-rank matrix recovery from a minimal number of noisy random measurements. IEEE Trans. Inform. Theory, 57(4):2342 -- 2359, 2011. [4] Emmanuel J. Cand`es and Benjamin Recht. Exact matrix completion via convex optimization. Found. Comput. Math., 9(6):717 -- 772, 2009. [5] Emmanuel J. Cand`es and Terence Tao. Decoding by linear programming. IEEE Trans. Inform. Theory, 51(12):4203 -- 4215, 2005. [6] Emmanuel J. Candes and Terence Tao. Near-optimal signal recovery from random projections: universal encoding strategies? IEEE Trans. Inform. Theory, 52(12):5406 -- 5425, 2006. [7] Emmanuel J. Cand`es and Terence Tao. The power of convex relaxation: near-optimal matrix comple- tion. IEEE Trans. Inform. Theory, 56(5):2053 -- 2080, 2010. [8] Bernd Carl and Andreas Defant. Asymptotic estimates for approximation quantities of tensor product identities. J. Approx. Theory, 88(2):228 -- 256, 1997. [9] Albert Cohen, Wolfgang Dahmen, and Ronald DeVore. Compressed sensing and best k-term approxi- mation. J. Amer. Math. Soc., 22(1):211 -- 231, 2009. [10] David L. Donoho. Compressed sensing. IEEE Trans. Inform. Theory, 52(4):1289 -- 1306, 2006. [11] K. Dvijotham and Maryam Fazel. A nullspace analysis of the nuclear norm heuristic for rank minimiza- tion. In Proc. of ICASSP 2010, Dallas, TX, March 2010. [12] David E. Edmunds and Jan Lang. Gelfand numbers and widths. J. Approx. Theory, 166:78 -- 84, 2013. [13] Maryam Fazel. Matrix rank minimization with applications. PhD thesis, Stanford University, 2002. [14] Massimo Fornasier, Holger Rauhut, and Rachel Ward. Low-rank matrix recovery via iteratively reweighted least squares minimization. SIAM J. Optim., 21(4):1614 -- 1640, 2011. [15] Simon Foucart and Ming-Jun Lai. Sparsest solutions of underdetermined linear systems via lq- minimization for 0 < q ≤ 1. Appl. Comput. Harmon. Anal., 26(3):395 -- 407, 2009. [16] Simon Foucart, Alain Pajor, Holger Rauhut, and Tino Ullrich. The Gelfand widths of ℓp-balls for 0 < p ≤ 1. J. Complexity, 26(6):629 -- 640, 2010. [17] Simon Foucart and Holger Rauhut. A mathematical introduction to compressive sensing. Applied and Numerical Harmonic Analysis. Birkhauser/Springer, New York, 2013. [18] A. Yu. Garnaev and E. D. Gluskin. The widths of a Euclidean ball. Dokl. Akad. Nauk SSSR, 277(5):1048 -- 1052, 1984. [19] B. S. Kashin. The widths of certain finite-dimensional sets and classes of smooth functions. Izv. Akad. Nauk SSSR Ser. Mat., 41(2):334 -- 351, 478, 1977. [20] B. S. Kashin and V. N. Temlyakov. A remark on the problem of compressed sensing. Mat. Zametki, 82(6):829 -- 837, 2007. [21] Lingchen Kong and Naihua Xiu. Exact low-rank matrix recovery via nonconvex schatten p- minimization. Asia-Pacific Journal of Operational Research, pages 1340010 -- 1 -- 1340010 -- 13, 2013. 16 JAVIER ALEJANDRO CH ´AVEZ-DOM´INGUEZ AND DENKA KUTZAROVA [22] Hermann Konig. Eigenvalues of operators and applications. In Handbook of the geometry of Banach spaces, Vol. I, pages 941 -- 974. North-Holland, Amsterdam, 2001. [23] Kiryung Lee and Yoram Bresler. Guaranteed minimum rank approximation from linear observations by nuclear norm minimization with an ellipsoidal constraint. ArXiv preprint, page arXiv:0903.4742. [24] Kiryung Lee and Yoram Bresler. ADMiRA: atomic decomposition for minimum rank approximation. IEEE Trans. Inform. Theory, 56(9):4402 -- 4416, 2010. [25] Lu Liu, Wei Huang, and Di-Rong Chen. Exact minimum rank approximation via Schatten p-norm minimization. J. Comput. Appl. Math., 267:218 -- 227, 2014. [26] Shahar Mendelson, Alain Pajor, and Nicole Tomczak-Jaegermann. Uniform uncertainty principle for Bernoulli and subgaussian ensembles. Constr. Approx., 28(3):277 -- 289, 2008. [27] B. K. Natarajan. Sparse approximate solutions to linear systems. SIAM J. Comput., 24(2):227 -- 234, 1995. [28] Erich Novak. Optimal recovery and n-widths for convex classes of functions. J. Approx. Theory, 80(3):390 -- 408, 1995. [29] Samet Oymak, Karthik Mohan, Maryam Fazel, and Babak Hassibi. A simplified approach to recov- ery conditions for low rank matrices. In 2011 IEEE International Symposium on Information Theory Proceedings, pages 2318 -- 2322, Piscataway, NJ, 2011. IEEE. [30] Albrecht Pietsch. Eigenvalues and s-numbers, volume 13 of Cambridge Studies in Advanced Mathemat- ics. Cambridge University Press, Cambridge, 1987. [31] Albrecht Pietsch. History of Banach spaces and linear operators. Birkhauser Boston, Inc., Boston, MA, 2007. [32] Benjamin Recht, Maryam Fazel, and Pablo A. Parrilo. Guaranteed minimum-rank solutions of linear matrix equations via nuclear norm minimization. SIAM Rev., 52(3):471 -- 501, 2010. [33] Benjamin Recht, Weiyu Xu, and Babak Hassibi. Null space conditions and thresholds for rank mini- mization. Math. Program., 127(1, Ser. B):175 -- 202, 2011. Department of Mathematics, University of Texas at Austin, 2515 Speedway Stop C1200, Austin, TX 78712-1202. Current address: Instituto de Ciencias Matem´aticas, CSIC-UAM-UC3M-UCM, C/Nicol´as Cabrera, n◦ 13-15, Campus de Cantoblanco, UAM, 28049, Madrid, Spain. E-mail address: [email protected] Institute of Mathematics, Bulgarian Academy of Sciences, Sofia, Bulgaria. Current address: Department of Mathematics, University of Illinois at Urbana-Champaign, 1409 W. Green Street Urbana, IL 61801. E-mail address: [email protected]
1107.0275	1	1107	2011-07-01T17:12:39	Multiresolution analysis for Markov Interval Maps	[ "math.FA", "math.OA" ]	We set up a multiresolution analysis on fractal sets derived from limit sets of Markov Interval Maps. For this we consider the $\mathbb{Z}$-convolution of a non-atomic measure supported on the limit set of such systems and give a thorough investigation of the space of square integrable functions with respect to this measure. We define an abstract multiresolution analysis, prove the existence of mother wavelets, and then apply these abstract results to Markov Interval Maps. Even though, in our setting the corresponding scaling operators are in general not unitary we are able to give a complete description of the multiresolution analysis in terms of multiwavelets.	math.FA	math	MULTIRESOLUTION ANALYSIS FOR MARKOV INTERVAL MAPS JANA BOHNSTENGEL AND MARC KESSEBÖHMER Abstract. We set up a multiresolution analysis on fractal sets derived from limit sets of Markov Interval Maps. For this we consider the Z-convolution of a non-atomic measure supported on the limit set of such systems and give a thorough investigation of the space of square integrable functions with respect to this measure. We define an abstract multiresolution analysis, prove the existence of mother wavelets, and then apply these abstract results to Markov Interval Maps. Even though, in our setting the corresponding scaling operators are in general not unitary we are able to give a complete description of the multiresolution analysis in terms of multiwavelets. Contents Introduction and main results 1. 2. Markov Interval Maps 3. Abstract Multiresolution analysis 3.1. Proof of Theorem 1.6 3.2. Abstract multiplicative Multiresolution analysis 3.3. Translation completeness 4. Applications to Markov Interval Maps 4.1. Multiresolution Analysis for MIM 4.2. Mother wavelets for MIM 4.3. MRA for a Markov measures 4.4. Examples 5. Operator algebra References 1 9 11 13 16 18 20 20 25 25 29 29 31 1. Introduction and main results The main aim of this paper is to construct a wavelet basis on limit sets of Markov Interval Maps (MIM) in the unit interval translated by Z. In this way we extend the results in [DJ06, BK10], where wavelet bases with respect to singular measures were provided. For this we first prove that a MIM gives rise to a multiresolution analysis (MRA) and study the particular case where the underlying measure is Markovian. This MRA is then reformulated in an abstract way allowing us to prove the existence of a wavelet basis in this abstract setting. In the case of a fractal given by an iterated function system (IFS) on [0, 1] there are several approaches to construct a wavelet basis on the L2-space with respect to a suitable singular measure which is supported on a so-called enlarged fractal. The enlarged fractal is derived from the original fractal by first mapping scaled copies of it to each gap interval and then by taking the union of translats by Z Date: July 1, 2011. Key words and phrases. multiwavelets, multiresolution analysis, Markov interval maps. 1 2 JANA BOHNSTENGEL AND MARC KESSEBÖHMER defining a dense set in R. In [DJ06] the authors construct a wavelets basis for fractals on self-similar Cantor sets, i.e. sets that are given by affine IFS with the same scaling factor 1/N, N ≥ 2, for all p ≤ N branches. They consider the L2-space with respect to µ, the δ-dimensional Hausdorff measure restricted to the enlarged fractal, where δ denotes the dimension of the Cantor set. In this situation the analysis depends on the two unitary operators U and T , where U denotes the pf (N·) and T denotes the translation operator scaling operator given by U f := given by T f := f (· − 1) for f ∈ L2(µ). Furthermore, a natural choice for a father wavelet ϕ is the characteristic function on the original fractal. The authors show that for a family of closed subspaces (Vj)j∈Z of L2(µ) the following six conditions are satisfied. √ • cl(cid:83) • (cid:84) • ··· ⊂ V−2 ⊂ V−1 ⊂ V0 ⊂ V1 ⊂ V2 ⊂ ··· , j∈Z Vj = L2(µ), j∈Z Vj = {0}, • Vj+1 = U Vj, j ∈ Z, • {T nϕ : n ∈ Z} is an orthonormal basis in V0, • U−1T U = T N. These observations allow the authors to construct a wavelets basis for L2(µ) explic- itly in terms of certain filter functions. In [BK10] we generalize this approach by allowing conformal IFS satisfying the open set condition on [0, 1]. We choose the measure of maximal entropy supported on the fractal and this measure is extended to a measure µ supported on the enlarged fractal in R. Then similarly as in [DJ06] we construct the wavelet basis via MRA in terms of the unitary scaling operator U and the unitary translation operator T . Again via filter functions the mother wavelets ψi, i ∈ {1, . . . , N − 1} are defined such that(cid:8)U nT kψi : n, k ∈ Z, i ∈ {1, . . . , N − 1}(cid:9) provides an orthonormal basis of L2(µ). (cid:83) by X :=(cid:84)∞ n=0 F −nI, where I :=(cid:83) Here, our aim is to extend the construction of wavelet bases with respect to fractal measures to the construction of wavelet bases on the by Z translated limit set of a Markov Interval Map (MIM). A Markov Interval Map consists of a family (Bi)N−1 i=0 of closed subintervals in [0, 1] with disjoint interior, and a function F : i∈N Bi → [0, 1], such that FBi is expanding and C 1, i ∈ N := {0, . . . , N − 1} and such that F (Bi) ∩ Bj (cid:54)= ∅ implies Bj ⊂ F (Bi). Its (fractal) limit set is given i∈N Bi. By considering its inverse branches τi := −1, i ∈ N, we obtain a Graph Directed Markov System (see [MU03]) with (FBi) incidence matrix A = (Aij)i,j∈N, where Aij = 1 if F (Bi) ⊃ Bj and 0 otherwise. For the precise definition see Definition 2.1 and for an explicit example of an MIM see Example 1.2 where we consider the β-transformation. The limit set X is -- up to a countable set where it is finite-to-one -- homeomorphic to the topological Markov chain ΣA := {ω = (ω0, ω1, . . . ) ∈ N : Aωiωi+1 = 1∀i ≥ 0}. For the definition of Given a Markov measure(cid:101)ν on the shift space ΣA with a probability vector (pi)i∈N the canonical coding map π from ΣA to X see (2.1). and stochastic matrix (πij)i,j∈N, we consider the probability measure ν :=(cid:101)ν ◦ π−1, to which we also refer as ν a Markov measure. The Z-convolution (by translations) of ν is given by N (cid:88) k∈Z νZ := ν(· − k). Similar to the construction in [BK10] we introduce the scaling operator (1.1) U f (x) := · 1[ji](x − k) · f (τ−1 j (x − k) + j + N k) (cid:88) (cid:88) (cid:88) (cid:114) pi k∈Z j∈N i∈N pjπji MULTIRESOLUTION ANALYSIS FOR MARKOV INTERVAL MAPS 3 Figure 1.1. The graph of the β-transform. and the translation operator T f (x) := f (x − 1) (1.2) for f ∈ L2(νZ) and x ∈ R, where [ji] ⊂ R, i, j ∈ N, denotes a cylinder set (see Section 2). It is important to note that in contrast to the construction of the scaling operator for IFS the operator U is in general not unitary. Nevertheless, we have the following properties. Proposition 1.1. Let (ϕi)i∈N denote a family of father wavelets given by ϕi := 1[i], i ∈ N. The translation operator T and the scaling operator U satisfy −1 (cid:112)ν([i]) (2) ϕi = U(cid:80) the following properties. √ πijT iϕj, i ∈ N, (1) T U = U T N, j∈N (3) (cid:104)T kϕiT lϕj(cid:105) = δ(k,i),(l,j), k, l ∈ Z, i, j ∈ N, (4) U U∗ = id, (5) U∗U = id if and only if Aij = 1 for all i, j ∈ N . For an explicit formula of U∗ see (4.2). As an example for this setting we consider the β-transformation. √ 5 denote the golden mean. Then Example 1.2 (β-Transformation). Let β := 1+ the β-transformation is given by F : [0, 1] → [0, 1], x (cid:55)→ βx mod 1 (see Figure 1.1 2 for the graph of F ). This map can be considered as a MIM as follows. In this β , x ∈ [0, 1], and case we have X := [0, 1] and the inverse branches are τ0(x) := x β , x ∈ [0, β − 1]. We may choose the two intervals B0 := [0, β − 1] (cid:19) τ1(x) := x+1 and B1 := [β − 1, 1] and the corresponding transition matrix is then given by A := (cid:18) 1 1 . 1 0 From [Rén57, Par60] we know that there exists an invariant measure ν absolutely continuous to the Lebesgue measure restricted to [0, 1] with density h given by (cid:40) 5+3 √ √ 10 5 5 5+ 10 h(x) := √ for 0 ≤ x < for 5−1 2 2 ≤ x < 1. 5−1 √ , 4 JANA BOHNSTENGEL AND MARC KESSEBÖHMER Figure 1.2. The graph of U(cid:0)id[0,1] (cid:1). The measure ν can be represented on ΣA by a stationary Markov measure with the stochastic matrix and probability vector p := is then given for x ∈ R by . The scaling operator U acting on L2 (νZ) (cid:19) 2 − β 0 (cid:18) β − 1 (cid:17) 1 5 5 k∈Z Π := , β−1√ (cid:88) (cid:16) β√ (cid:16)(cid:112)β1[0,β−2)(x − k) + 1[β−2,β−1)(x − k) (cid:17) · f (β(x − k) + 2k) . 5/β(cid:1)1/2 + β · 1[β−1,1)(x − k) 1[0,β−1) and ϕ1 =(cid:0)√ (cid:17) (cid:16)(cid:112)β1[0,β−2)(x) + 1[β−2,β−1)(x) + β · 1[β−1,1)(x) βx. U f (x) = U(cid:0)id[0,1] (cid:1) x = For the father wavelets we may choose ϕ0 =(cid:0)√ 5β(cid:1)1/2 1[β−1,1). We illustrate the action of U in Figure 1.2 where U is applied to the identity map id[0,1] : x (cid:55)→ x, restricted to [0, 1], that is for x ∈ [0, 1] we have consider a family of scaling operators (cid:0)U (n)(cid:1) We further generalize our construction by considering non-atomic probability measures ν on X which we do not assume to be Markovian. In this case it is natural to consider more than just one scaling operator U. More precisely, we n∈Z which allow us to construct an orthonormal wavelet basis. For this we define U (0) := id and for f ∈ L2(νZ) and n ∈ N we let (cid:88) (cid:88) (cid:88) k∈Z ω∈Σn A j∈N ·f (cid:115) (cid:16) 1[ωj](x − k) νZ([j]) νZ([ωj]) τ−1 ω (x − k) +(cid:80)n−1 i=0 ωn−1−iN i + N nk (cid:17) U (n)f (x) := (1.3) 5 (cid:33) (cid:33) (cid:33) ωn−1−iN i − N nk ωn−1−iN i − N nk + k . It is straight forward to verify that if the measure ν is Markovian, then we have U (n) = U n for n ∈ N0 and U (−n) = (U∗)n, n ∈ N. More details are provided n∈Z and T satisfy the following relations. Proposition 1.3. Let (ϕj)j∈N denote the family of father wavelets given by ϕi = 1[i], i ∈ N. The translation operator T and the family of scaling (νZ([i])) −1/2 A i=0 i=0 ·f τω := 1[j] (cid:115) k∈Z a j∈N ω∈Σn (1.4) (cid:32) (cid:88) (cid:88) (cid:88) (cid:88) νZ([ωj]) νZ([j]) and U (−n)f (x) (cid:32) x − n−1(cid:88) (cid:32) x − n−1(cid:88) in Section 4.3. Furthermore, the operators (cid:0)U (n)(cid:1) operators(cid:0)U (n)(cid:1) (3) ϕi = U (1)T i(cid:80) The properties of (cid:0)U (n)(cid:1) (1) Let(cid:0)U (n)(cid:1) (1) T U (n) = U (n)T N n, n ∈ N, (2) U (−n)T ϕj = T N n j∈N (cid:113) νZ([ij]) νZ([i]) ϕj, i ∈ N, U (−n)ϕj, n ∈ N, j ∈ N, n∈Z satisfy the following. MULTIRESOLUTION ANALYSIS FOR MARKOV INTERVAL MAPS (4) (cid:104)U (n)T kϕiU (n)T lϕj(cid:105) = δ(k,i),(l,j), n, k, l ∈ Z, i, j ∈ N, (5) U (n)U (−n) = id, n ∈ N, (6) if U (n)T kϕj (cid:54)= 0, then U (−n)U (n)T kϕj = T kϕj, n ∈ N, k ∈ Z, j ∈ N. n∈Z and T lead us to the following abstract definition of a multiresolution analysis which involves more than one father wavelet. In the literature these functions are sometimes called multiwavelets (cf. [Alp93]). Definition 1.4 (Abstract MRA). Let µ be a non-atomic measure on (R,B). (cid:16) µ,(cid:0)U (n)(cid:1) n∈Z ,T(cid:17) n∈Z and T be bounded, linear operators on L2(µ) such that T is unitary and U (0) = id. We say allows a two-sided mul- tiresolution analysis (two-sided MRA) if there exists a family {Vj : j ∈ Z} of closed subspaces of L2 (µ) and a family of functions (called father wavelets) ϕj ∈ L2 (µ), j ∈ N, N ∈ N, with compact support, such that the following conditions are satisfied. (a) ··· ⊂ V−2 ⊂ V−1 ⊂ V0 ⊂ V1 ⊂ V2 ⊂ . . . , (b) cl(cid:83) (c) (cid:84) (d) (cid:0)U (n)(cid:8)T kϕj : k ∈ Z, j ∈ N(cid:9)(cid:1)\{0}, n ∈ Z, is an orthonormal basis of (e) U (n)(cid:8)T kϕi : k ∈ N n, i ∈ N(cid:9) ⊂ spanU (n+1)(cid:8)T kϕi : k ∈ N n+1, i ∈ N(cid:9), n ∈ N0 and U (−n) {ϕi : i ∈ N} ⊂ spanU (−n+1)(cid:8)T kϕi : i ∈ N , k ∈ N(cid:9), j∈Z Vj = L2(µ), j∈Z Vj = {0}, Vn, n ∈ N, (2) Let (cid:0)U (n)(cid:1) (cid:16) n∈N0 ,T(cid:17) µ,(cid:0)U (n)(cid:1) (f) T U (n)V0 = U (n)T N nV0 and U (−n)T V0 = T N nU (−n)V0, n ∈ N. and T be linear operators on L2(µ), T unitary, and let U (0) = id. We say allows a one-sided multiresolution analysis ( one-sided MRA) if there exists a family (Vj : j ∈ N0) of closed subspaces of L2 (µ) and a family of functions (called father wavelets) ϕj ∈ L2 (µ), j ∈ N, N ∈ N, with compact support, such that the following conditions are satisfied. (a) V0 ⊂ V1 ⊂ V2 ⊂ . . . , Vj = L2(µ), (b) cl(cid:83) n∈N0 j∈N0 introduced above. Theorem 1.5. Let(cid:0)U (n)(cid:1) (cid:16) νZ,(cid:0)U (n)(cid:1) (cid:17) 6 JANA BOHNSTENGEL AND MARC KESSEBÖHMER (c) (cid:0)U (n)(cid:8)T kϕj : k ∈ Z, j ∈ N(cid:9)(cid:1)\{0}, n ∈ N0, is an orthonormal basis (d) U (n)(cid:8)T kϕi : k ∈ N n, i ∈ N(cid:9) ⊂ spanU (n+1)(cid:8)T kϕi : k ∈ N n+1, i ∈ N(cid:9), of Vn, n ∈ N, (e) T U (n)V0 = U (n)T N nV0, n ∈ N. Our next theorem shows that the abstract MRA holds in particular for MIM as n∈N0 lows a one-sided MRA, where the father wavelets are set to be ϕi := (νZ([i])) i ∈ N. n∈N0 be given as in (1.3). Then , T −1/2 al- 1[i], (cid:17) basis. For the abstract MRA we show that there always exists an orthonormal wavelet bounded, linear operators on L2(µ) and T be a unitary operator on L2(µ). Theorem 1.6. Let µ be a non-atomic measure on R, (cid:0)U (n)(cid:1) (cid:16) µ,(cid:0)U (n)(cid:1) with dn ≥ qn, d−n ≥ q−n, and two families of mother wavelets(cid:0)ψn,l : l ∈ dn − qn (cid:16) n∈Z be a family of If allows a two-sided MRA with father wavelets ϕj, j ∈ N, then there exist for every n ∈ N0 numbers dn ∈ N n+2, d−n ∈ N 2, qn ∈ N n+1, q−n ∈ N, , n ∈ N0, such that the following set of functions defines an n∈Z ,T(cid:17) ψ−n,l : l ∈ d−n − q−n orthonormal basis for L2(µ) (cid:8)T kψn,l : n ∈ N0, l ∈ dn − qn, k ∈ Z(cid:9) ∪(cid:110)T N nkψ−n,l : n ∈ N, l ∈ d−n − q−n, k ∈ Z(cid:111) (cid:1) and show (cid:9), and respectively for n < 0 Remark 1.7. We give a precise construction for the family of mother wavelets ψn,l in Section 3.1. More precisely, for each n ∈ Z we consider the linear subspaces Wn := Vn+1 (cid:9) Vn, where the closed subspaces Vn of L2(µ) are given in Defi- nition 1.4 (1d), and the finite family of functions (cid:0)ψn,l : l ∈ dn − qn that that for n ≥ 0 (cid:8)T kψn,l : k ∈ Z, l ∈ dn − qn (cid:110)T Nnkψn,l : k ∈ Z, l ∈ d−n − q−n , defines an orthonormal basis of Wn. (cid:1), (cid:111) . Note that for IFS the mother wavelets are typically constructed in terms of so- called filter functions. We will see in Section 5 that an analog construction is still possible if the measure ν is Markovian. sponding result for the one-sided MRA. An immediate consequence of the proof of Theorem 1.6 is the following corre- Corollary 1.8. Let µ be a non-atomic measure on R, (cid:0)U (n)(cid:1) wavelets(cid:0)ψn,l : l ∈ dn − qn µ,(cid:0)U (n)(cid:1) (cid:1), n ∈ N0, such that the following set of functions defines bounded, linear operators on L2(µ) and T a unitary operator on L2(µ). If allows a one-sided MRA with the father wavelets ϕj, j ∈ N, then there exists for every n ∈ N0 numbers dn ∈ N n+2, qn ∈ N n+1 with dn ≥ qn and a family of mother (cid:8)T kψn,l : n ∈ N0, l ∈ dn − qn, k ∈ Z(cid:9) ∪(cid:8)T kϕi : k ∈ Z, n ∈ N(cid:9) . an orthonormal basis for L2(µ) a family of n∈N0 (cid:16) The construction for a MIM with an underlying Markov measure ν belongs to a specific class. In this class the scaling operators U (n) can be represented multiplica- tively. In our general framework we say that is multiplicative if there exists a linear, bounded operator U on L2 (µ) such that for n ∈ N0 U (n) = U n µ,(cid:0)U (n)(cid:1) n∈N0 ,T(cid:17) (cid:16) ,T(cid:17) n∈N0 (cid:16) µ,(cid:0)U (n)(cid:1) ,T(cid:17) MULTIRESOLUTION ANALYSIS FOR MARKOV INTERVAL MAPS 7 and U (−n) = (U∗)n. The results concerning the mother wavelets simplify in this case as a consequence of the following lemma. Lemma 1.9. Let us assume that allows a two-sided MRA with the closed subspaces Vn of L2(µ) from Definition 1.4 (1d) and set Wn := Vn+1(cid:9) Vn, n ∈ Z. n∈N0 • If U (n) = U n for n ∈ N then Wn = U nW0, n ∈ N. • If U (−n) = (U∗)n for n ∈ N0 then W−n = (U∗)n−1 W−1. Thus, we only have to find appropriate mother wavelets for W0 and W−1 and ob- tain a wavelet basis then by applying repeatedly U. More precisely, this observation allows us to derive the following corollary from the Theorem 1.6. is multiplicative, then there exists an or- (cid:41) (cid:111) , Corollary 1.10. If n∈N0 thonormal basis of L2(µ) of the form (cid:40) (cid:16) µ,(cid:0)U (n)(cid:1) ,T(cid:17) n−1(cid:88) i=0 A, k = ∪(cid:110) U nT kψl : n ∈ N0, ω ∈ Σn ωiN i + N nm, m ∈ Z, l ∈ d1 − N (U∗)n T kψ−,l : n ∈ N0, k ∈ Z, l ∈ d−1 − N where the functions ψl, l ∈ d1 − N, and ψ−,l, l ∈ d−1 − N, are given explicitly in Remark 3.6. The above corollary applied to Example 1.2 leads to the following construction. Example (Example 1.2 (continued)). The mother wavelet is (cid:16)√ ψ = 5(2 − β) (cid:17)1/2 1[0,(β−1)2) −(cid:16)√ 5 (cid:17)1/2 1[(β−1)2,β−1) and so a basis is given by (cid:8)T kϕ1 : k ∈ 2Z + 1(cid:9) ∪(cid:8)U nT kψ : k ∈ Dn, n ∈ N(cid:9) ∪(cid:8)(U∗)n T kψ : k ∈ Z, n ∈ N(cid:9) ,  . kj2j + 2nl : (kj)j∈n ∈ {0, 1}n , kj · kj−1 = 0, j ∈ n − 1, l ∈ Z n−1(cid:88) where Dn := j=0 The proof that this indeed defines a orthonormal basis will be postponed to Section 4.4. −1/2 (cid:16) (cid:17) n∈Z , T νZ,(cid:0)U (n)(cid:1) In the case of a MRA for a MIM with Markov measure ν we have in particular that U (n) = U n and U (−n) = (U∗)nand we even obtain a stronger correspondence between Markov measures for MIM and a two-sided MRA. Theorem 1.11. We have that the father wavelets ϕi := (νZ([i])) Markovian. allows a two-sided MRA with 1[i], i ∈ N, if and only if the measure ν is appart from being multiplicative, that is we have ϕj ∈ span U(cid:8)T jϕi : i ∈ N(cid:9) for In the case of ν being a Markov measure we even have a stronger property each j ∈ N. We call a MRA with this property translation complete. We further investigate multiplicative MRA which are translation complete in Section 3.2. In this situation we derive a 0-1-valued transition matrix A given by Aij = 0 if and only if UT iϕj = 0 and show that for MIM the matix coincides with the incidence matrix. This observation is used to construct the mother wavelets in a simpler way by considering for each father wavelet a unitary matrix to obtain coefficients for the 8 JANA BOHNSTENGEL AND MARC KESSEBÖHMER corresponding mother wavelets. We will use this approach to construct the mother wavelets for MIM. We would like to point out some interesting connections to C∗-algebras of Cuntz- Krieger type, [KSS07]. We start by further considering the scaling operator U for the MRA in the setting of a MIM with the incidence matrix A and Markov measure ν. We can also write the operator U in a different way using the representation of a Cuntz-Krieger-algebra. For this we consider the partial isometries Si given for i ∈ N, f ∈ L2(ν), x ∈ supp(ν) by Sif (x) = (ν([i])) −1/2 1[i](x)f (τ−1 i (x)). It has been shown in [KSS07] that this gives a representation of the Cuntz-Krieger- algebra OA by bounded operators acting on L2(ν), that is the Si, i ∈ N, are partial isometries and satisfy (cid:88) (cid:88) j∈N i∈N S∗ i Si = 1 = AijSjS∗ j , SiS∗ i . The scaling operator U acting on L2 (νZ) can then alternatively be written in terms of the partial isometries as (cid:88) (cid:88) (cid:88) (cid:114) pi k∈Z j∈N i∈N πji U = T kSj 1[i]T −(j+N k), where we notice that Sj 1[i], j, i ∈ N, acts on L2 (νZ). We can also write U∗ in terms of the partial isometries Si, i ∈ N. In this way we obtain j T −k. T j+N k1[i]S∗ (cid:114) πji (cid:88) (cid:88) (cid:88) U∗ = k∈Z j∈N i∈N pi The spaces Vn, n ∈ N0, can also be written in terms of the isometries Si, i ∈ N, that is for n ∈ N a basis of Vn is given by (cid:26)(cid:114) pi πωn−1i T lSωϕi : l ∈ Z, ω ∈ Σn A, i ∈ N . (cid:27) Let us finish this section by commenting on some known results in the literature connected to the results in here. Up to our knowledge there are at least two further approaches to construct a wavelet basis on the limit sets of MIM, namely [MP09, KS10] and there is one approach for the specific case of a β-transformation given in [GP96]. In [MP09] Marcolli and Paolucci consider the limit set X of a MIM N for i ∈ N inside the unit interval consisting of the inverse branches τi(x) = x+i with some transition rule encoded in a matrix A. This limit set can be associated with a Cantor set inside the unit interval. The Cantor set is then equipped with the Hausdorff measure of the appropriate dimension δ. If all transitions were allowed, the limit set would coincide with a usual Cantor set given by an affine iterated function system. They then use the representation of the Cuntz-Krieger-algebra OA, where A is the transition matrix, for the construction of the orthonormal (cid:1) and not a multiresolution analysis. Their proofs system of wavelets on L2(cid:0)H δX mainly rely on results in [Bod07, Jon98]. Finally, Marcolli and Paolucci give a possible application where they adapt the construction of a wavelet basis to graph wavelets for finite graphs with no sinks, which can be associated to Cuntz-Krieger- algebras. These graph wavelets are a useful tool for spatial network traffic analysis, compare [MP09, CK03]. MULTIRESOLUTION ANALYSIS FOR MARKOV INTERVAL MAPS 9 In [KS10] the authors construct a Haar basis analogous to the wavelet basis construction in [DJ06] for the middle third Cantor set for a one-sided topologically exact subshift of finite type and with respect to a Gibbs measure µφ for a Hölder continuous potential φ. The construction is then used to obtain a spectral triple in the framework of non-commutative geometry. The construction of wavelet basis in different spaces than L2(λ), where λ is the Lebesgue measure on R, may lead to a further understanding of non-commutative geometry in the sense that we can obtain a "Fourier" or wavelet basis for quasi lattices or quasi crystals. As an essential non-linear example for the construction of a wavelet basis on limit sets of MIM one can take the limit set of a Kleinian group together with the measure of maximal entropy or the Patterson-Sullivan measure, compare Example 2.3. 5 √ As an example we apply the construction to a β-transformation, where β denotes , compare Example 1.2. In this way we obtain a the golden mean, i.e. β = 1+ 2 wavelet basis for L2 (νZ) , where ν is the invariant measure for this transformation, compare [Rén57, Par60]. This measure is absolutely continuous with respect to the Lebesgue measure. In [GP96], Gazeau and Patera construct a similar basis to ours for the β-transformation with respect to the Lebesgue measure on R. They use instead of a translation by the group Z a translation by so called β-integers which consider the β-adic expansion and are obtained by a so-called greedy algorithm. There are some common features between our construction and the one in [GP96] like both give characteristic functions on intervals depending on powers of β. But since we consider different measures we have different coefficients. n∈Z and give a proof of Theorem 1.6. operators (cid:0)U (n)(cid:1) In Section 4 we start with a family of operators (cid:0)U (n)(cid:1) The paper is organized as follows. In Section 2 we provide some basic definitions and introduce MIM. In Section 3 we elaborate the abstract MRA for families of In Section 3.2 we then consider the special case of multiplicative systems. In Section 3.3 we prove how the condition of translation completeness simplifies the construction of the mother wavelets. The rest of this paper is devoted to the special case of a MRA for MIM. n∈Z acting on L2 (νZ) for an arbitrary non-atomic probability measure ν on the limit set of an MIM in the unit interval translated by Z and show that a one-sided MRA is always satisfied. If on the other hand a two-sided MRA holds, we then prove that the measure ν is necessarily Markovian. The construction of the mother wavelets will be given explicitly. In Section 4.3 we give an explicit construct of the wavelet basis if the measure ν is Markovian. Finally, in Section 5 we show how low-pass filters and high-pass filters can be em- ployed to construct mother wavelets for multiwavelets for MIM with an underlying Markov measure. 2. Markov Interval Maps In this section we give some basic definitions and notations. We consider fractals given as limit sets of one-dimensional Markov Interval Maps. fine I :=(cid:83) a Markov Interval Map and its limit set is defined as X :=(cid:84)∞ (cid:17) Definition 2.1. Let (Bi)i∈N be closed intervals in [0, 1] with disjoint interior. De- i∈N Bi and F : I → [0, 1] exanding and C 1 on each Bi, i ∈ N, such that (Bi)i∈N , F if F (Bi)∩ Bj (cid:54)= ∅ then Bj ⊂ F (Bi) for i, j ∈ N. We call the system (cid:16) n=0 F −nI. Remark 2.2. 10 JANA BOHNSTENGEL AND MARC KESSEBÖHMER (a) A fundamental do- main of the action of (cid:104)g, h(cid:105) on the Poincaré disc model. (b) The correspond- ing Bowen-Series map. Figure 2.1. Example of a Fuchsian group. function system (IFS). (1) If F (Bi) = [0, 1] for each i ∈ N, then (X, F ) corresponds to an iterated −1, i ∈ N. The family (τi)i∈N (2) We define the inverse branches τi := (FBi) is called a one-dimensional graph directed Markov system (GDMS) with the incidence matrix A = (Aij)N−1 (cid:40) and it follows that F (Bi) =(cid:83) Aij := 1, 0, i,j=0 which is obtained by if Bj ⊂ F (Bi) else, j∈N : Aij =1 Bj. Example 2.3. An example is a convex, co-compact Kleinian group, as an example consider Figure 2.1a. The limit set can be considered as the limit set of the Bowen- Series map, which gives rise to a Markov Interval Map, compare Figure 2.1b. The limit set is the set that is obtained by successive application of these four maps, where the composition of gi and g−1 i are forbidden. A typical measure to be studied would be the measure of maximal entropy or the conformal measure (of maximal dimension). Next we consider the corresponding shift space. Consider the alphabet N = {0, . . . , N − 1}. The limit set X is then homeomorphic (mod ν) to the set of all admissible words ΣA := {ω = (ω0, ω1, . . . ) ∈ N N The homeomorphism is given, for x ∈ X, by : Aωiωi+1 = 1∀i ≥ 0}. π : ΣA → X ω (cid:55)→ limn→∞ τω0 ◦ ··· ◦ τωn (x), (2.1) which is independent of the particular choice of x ∈ X. Remark 2.4. Furthermore, we define the cylinder sets for ω0, . . . , ωk ∈ N, k ∈ N0, as [ω0 . . . ωk] := {(ω(cid:48) 0, ω(cid:48) 1, . . . ) ∈ ΣA : ωi = ω(cid:48) i, i ∈ {0, . . . , k}} . MULTIRESOLUTION ANALYSIS FOR MARKOV INTERVAL MAPS 11 If for some i ∈ {0, . . . , k − 1} Aωiωi+1 = 0 then [ω0 . . . ωk] = ∅. Then the sets Bi and F (Bi) for i ∈ N are homeomorphic (mod ν) to the following sets in the shift space: π−1 (Bi) = [i] π−1 (F (Bi)) = {ω = (ω0, ω1, . . . ) ∈ ΣA : Aiω0 = 1}. and The dynamic of F is conjugated to the shift dynamic σ : ΣA → ΣA, σ (ω0, ω1, . . . ) = (ω1, ω2, . . . ) and consequently, the functions τi correspond to the inverse branches of the shift function, i.e. τi ◦ π (ω0, ω1, . . . ) = π (i, ω0, ω1, . . . ), for ω ∈ π−1 (F (Bi)), i ∈ N. A := (cid:8)ω = (ω0, . . . , ωn−1) ∈ N n : Aωiωi+1 = 1 for all i ∈ {0, . . . , n − 1}(cid:9) Furthermore, let us fix the following notation. • Σn • Σ∗ • For ω ∈ Σn • For ω ∈ Σn defines the set of admissible words of length n ∈ N. A stands for all finite words, i.e. Σ∗ A. n≥1 Σn A we define τω := τω0 ◦ τω1 ◦ ··· ◦ τωn−1. A, τ ∈ Σm A we define their concatenation ωτ := (ω0, . . . , ωn−1, τ0, . . . , τm−1) A =(cid:83) which is an element of Σn+m whenever Aωn−1τ0 = 1. A As a measure on X we could consider for instance the pull-back under π of Gibbs measures on ΣA (for definitions see e.g. [KS10]). Now we define the appropriate space for which we want to construct a wavelet Definition 2.5. Let(cid:101)ν be a probability measure on (ΣA,B) and ν =(cid:101)ν◦π−1. Define basis. the enlarged fractal by (cid:91) (cid:88) k∈Z k∈Z R = X + k νZ(B) = ν(B − k), and define the Z-convolution νZ of the measure ν for a Borel set B by which clearly is an invariant measure under Z-translation. Remark 2.6. One example is the space L2 (ΣA, µφ), where ΣA denotes a one-sided topologically exact subshift of finite type. An important class of measures on ΣA are given by invariant Gibbs measure with respect to a Hölder continuous potential φ ∈ C (ΣA, R), denoted by µφ, compare [KS10]. µφ corresponds to the measure (cid:101)ν in Definition 2.5. In the following we use the convention 0−1 · 1∅ = 0. For simplicity we let [ω0, . . . , ωn−1] also denote the sets τω0 ◦ ··· ◦ τωn−1 (X) using the identification by to a measure(cid:101)ν on ΣA by ν =(cid:101)ν ◦ π−1 and νZ denotes the measure obtained from ν π. Furthermore, in Section 4 the measure ν supported on [0, 1] always corresponds by Z-convolution. 3. Abstract Multiresolution analysis In this section we give a proof of Theorem 1.6. To do so we first construct mother wavelets explicitly and in the next step we prove that these give indeed which allows a an orthonormal basis. two-sided MRA. In this section we fix µ,(cid:0)U (n)(cid:1) n∈Z ,T(cid:17) (cid:16) For the construction of an ONB we cannot define the mother wavelets in terms of filter functions due to the fact that we have more than one father wavelet. 12 JANA BOHNSTENGEL AND MARC KESSEBÖHMER Remark 3.1. If we have the usual setting from the literature, compare e.g. [Dau92], then we have a multiplicative MRA with a unitary operator U and the operator T given in (1.2). In this case there is only one father wavelet ϕ and there exists a k∈Z akzk, ak ∈ C, such that ϕ = Um0(T )ϕ. For the construction of the mother wavelets we look for N − 1 k ∈ C, j ∈ N\{0}, where N ∈ N is connected to the scaling since it indicates on which interval [0, N ] the unit interval is mapped when the operator U is applied to 1[0,1]. The high-pass filters are chosen, such that the matrix 1√ N so-called low-pass filter m0 : T → T of the form m0(z) =(cid:80) high-pass filters mj : T → T of the form mj : z (cid:55)→(cid:80) (cid:0)mj(ρlz)(cid:1) kzk, bj M (z) := k∈Z bj j,l∈N , where ρ = e2πi/N, is unitary for almost all z ∈ T. In terms of these high-pass filters the mother wavelets are defined as ψj = Umj(T )ϕ, j ∈ N\{0}. We first notice that for n ∈ N (cid:26)(cid:18)(cid:106) k (cid:107) (cid:19) (cid:27) {(l, j) ∈ N n × N} = (3.1) where (m)N := m mod N and (cid:98)x(cid:99) = maxk∈Z,k≤x(k) is the largest integer not exceeding x. , (k)N N , Clearly from the definition of the MRA, Definition 1.4 (1e), we have the following: (1) If for n ∈ N0, k ∈ N n+1, U (n)T (cid:98) k N (cid:99)ϕ(k)N (cid:54)= 0 , there exists uniquely : k ∈ N n+1 m∈N n+2 ∈ CN n+2 such that m and U (n)T (cid:98) k (cid:1) N (cid:99)ϕ(k)N = U (n+1)(cid:80) determined(cid:0)an,k (cid:0)an,k U (−n)ϕi = U (−n+1)(cid:80) (cid:0)bn,i m∈N 2 ∈ CN 2 such that m = 0, m ∈ N n+2, m = 0, m ∈ N 2, (cid:0)bn,i and (cid:1) m (3.2) (3.3) m∈N n+2 an,k if U (n+1)T (cid:98) m m T (cid:98) m N (cid:99)ϕ(m)N N (cid:99)ϕ(m)N = 0(cid:1) . m∈N 2 bn,i if U (−n+1)T (cid:98) m m T (cid:98) m N (cid:99)ϕ(m)N N (cid:99)ϕ(m)N = 0(cid:1) . (2) If U (−n)ϕi (cid:54)= 0, n ∈ N, i ∈ N, there exists uniquely determined coefficients m m (cid:1) m∈N 2, i ∈ N, n ∈ N. (1) For fixed n ∈ N0, define Remark 3.2. We only consider U−(n)ϕi, since U (−n)T kϕi = T N nkU (−n)ϕi by (1f) of Definition 1.4. Lemma 3.3. The following holds for the coefficients (cid:0)an,k (cid:1) n ∈ N0, and(cid:0)bn,i m∈N n+2, k ∈ N n+1, (cid:111) N (cid:99)ϕ(m)N (cid:54)= 0 (cid:111) m∈N n+2, k ∈ Qn, are orthonormal. (2) For fixed n ∈ N, define m ∈ N n+1 : U (n)T (cid:98) m Then the vectors vk =(cid:0)an,k (cid:110) (cid:1) Then the vectors wi =(cid:0)bn,i m∈N 2, i ∈ Q−n, are orthonormal. m ∈ N : U (−n)ϕm (cid:54)= 0 Q−n := Qn := (3.5) (3.4) (cid:110) (cid:1) m m . . MULTIRESOLUTION ANALYSIS FOR MARKOV INTERVAL MAPS 13 Proof. ad (1): For fixed n ∈ N0, let k, l ∈ Qn, then N (cid:99)ϕ(k)NU (n)T (cid:98) l m T (cid:98) m an,k N (cid:99)ϕ(l)N(cid:105) N (cid:99)ϕ(m)N (cid:12)(cid:12)U (n+1) (cid:88) m∈N n+2 δ(k,l) = (cid:104)U (n)T (cid:98) k =(cid:10)U (n+1) (cid:88) (cid:88) m∈N n+2 m an,l an,k m . = m∈N n+2 ad (2): Follows analogously to (1). (cid:11) m T (cid:98) m an,l N (cid:99)ϕ(m)N (cid:3) 3.1. Proof of Theorem 1.6. The aim is to prove the existence of a basis as given in Theorem 1.6. For this we divide the proof in two parts. First we construct coefficients such that the functions ψn,k given in (3.6) and (3.7) give an orthonormal basis. In the second part we verify that these functions give indeed an orthonormal basis. We prove these parts first for n ∈ N0 and then for n ∈ Z, n < 0. We define the mother wavelets for each scale n ∈ Z such that we obtain with their translates a basis for Wn = Vn+1 (cid:9) Vn, where Vn is given in Definition 1.4 (1d). Define for n ∈ N0 m (cid:54)= 0 for some k ∈ Qn m (cid:54)= 0 for some k ∈ Q−n (cid:9) , (cid:9) , (3.6) (3.7) and dn := card Dn, d−n := card D−n. • The mother wavelets for the subspaces Wn, n ∈ N0, of the MRA shall have the form for k ∈ dn − qn with qn := card Qn m T (cid:98) m cn,k N (cid:99)ϕ(m)N , m∈N n+2 m ∈ C are given in (3.8). where the coefficients cn,k • For the negative index subspaces W−n, n ∈ N, of L2(µ) we define the mother wavelets in terms of the coefficients of the matrix in (3.9) for n ∈ N and k ∈ d−n − q−n, where q−n := card Q−n, as m T (cid:98) m c−n,k ψ−n,k := U (−n+1) (cid:88) N (cid:99)ϕ(m)N . m∈N 2 (cid:1) matrix. m ∈ C, c−n,k The coefficients cn,k Schmidt process. m ∈ C are determined in the following via the Gram- For the definition of the basis we fix n ∈ N0 and we construct an orthonormal . This is a (qn × dn)- Now we consider dn − qn vectors ei, i ∈ dn − qn, of length dn which are linearly , k ∈ Qn. Via the Gram-Schmidt process (cid:1) , i ∈ dn − qn, of length dn which to m∈Dn m = 0 if m ∈ N n+2\Dn and we define a matrix Cn := k∈Qn,m∈N n+2 of basis for Cdn in the following way. Consider(cid:0)an,k independent of the vectors (cid:0)an,k (cid:1) we obtain dn − qn orthonormal vectors(cid:0)cn,i are orthonormal to (cid:0)an,k (cid:0)cn,k (cid:1) , k ∈ Qn. We extend the vectors (cid:0)cn,i k∈dn−qn,m∈N n+2 of size (dn − qn) × N n+2 and An := (cid:0)an,k some of length N n+2 by cn,i k∈Qn,m∈Dn m m∈Dn m m∈Dn m m∈Dn (cid:1) (cid:1) (cid:1) m size qn × N n+2. So we obtain a matrix of size dn × N n+2 by m m m Dn :=(cid:8)m ∈ N n+2 : an,k D−n :=(cid:8)m ∈ N 2 : bn,k ψn,k := U (n+1) (cid:88) (3.8) Mn := Now we turn to the construction of the coefficients for ψ−n,k, n ∈ N, in (3.7). For each −n, n ∈ N, we define an orthonormal basis of Cd−n in the following way. (cid:18) AnCn (cid:19) . m 14 (cid:1) k∈Q−n,m∈D−n JANA BOHNSTENGEL AND MARC KESSEBÖHMER Consider(cid:0)bn,k (cid:1) q−n vectors which are linearly independent of(cid:0)bn,k process we obtain d−n − q−n orthonormal vectors(cid:0)c−n,j length d−n which are orthonormal to(cid:0)bn,i (cid:1) vectors(cid:0)c−n,i Now we define Dn := (cid:0)c−n,i . This is a (q−n × d−n)-matrix. Now we consider d−n− . Via the Gram-Schmidt m∈D−n , j ∈ d−n − q−n, of m∈D−n m . In the last step we extend the m = 0 if m ∈ N 2\D−n. k∈Q−n,m∈N 2 such to some of length N 2 by defining c−n,i i∈d−n−q−n,m∈N 2 and Bn := (cid:0)bn,k m∈D−n m∈D−n (cid:1) (cid:1) (cid:1) (cid:1) m m m m m that (3.9) (cid:102)Mn := (cid:19) (cid:18) BnDn is a matrix of size d−n × N 2. In the next step we show that we obtain indeed an orthonormal basis with these mother wavelets given in (3.6) and (3.7). First we prove this for n ∈ N0. Recall n∈N0 k=0 Wk ⊕ V0. Now we show (cid:1) = L2(µ) since that Wn = Vn+1 (cid:9) Vn for n ∈ N0. Consequently, cl(cid:0)(cid:83) for every n ∈ N0 it follows iteratively that Vn+1 =(cid:76)n that for fixed n ∈ N we have that(cid:8)T lψn,k : k ∈ dn − qn, l ∈ Z(cid:9) is an orthonormal Wn ∪ V0 basis of Wn. First we show the orthonormality. To show the orthonormality of T rψn,k and T sψn,l, r, s ∈ Z, k, l ∈ dn − qn, it is sufficient to consider T rψn,k and ψn,l since the operator T is unitary. The orthonormality follows then from (cid:69) m U (n+1)T (cid:98) m cn,l N (cid:99)ϕ(m)N (cid:69) N (cid:99)+N n+1rϕ(m)N N (cid:99)+N n+1rϕ(m)N (cid:12)(cid:12)(cid:12) (cid:88) (cid:68)T N n+1r+(cid:98) m m∈N n+2 N (cid:99)ϕ(m)N (cid:12)(cid:12)(cid:12) (cid:88) (cid:12)(cid:12)(cid:12)T (cid:98) s m∈N n+2 m T (cid:98) m cn,l N (cid:99)ϕ(m)N (cid:69) N (cid:99)ϕ(s)N · δ(N n+1r+(cid:98) m N (cid:99),(cid:98) s N (cid:99)),((m)N ,(s)N ) (cid:104)T rψn,kψn,l(cid:105) = = m∈N n+2 (cid:68) (cid:88) m U (n+1)T (cid:98) m (cid:68) (cid:88) cn,k m T (cid:98) m (cid:88) (cid:88) cn,k (cid:88) (cid:88) = δr,0 · (cid:88) cn,k m cn,l s cn,k m cn,l s m∈N n+2 m∈N n+2 m∈N n+2 s∈N n+2 s∈N n+2 cn,k m cn,l m = = m∈N n+2 = δr,0 · δk,l. In the next step we show that Vn+1 = Vn ⊕ Wn. We consider a basis element N (cid:99)ϕ(k)N , k ∈ N n+2, and show that it is a linear N (cid:99)ϕ(l)N and ψn,m, l ∈ N n+1, m ∈ dn − qn. It is N (cid:99)ϕ(k)N = 0 N (cid:99)ϕ(k)N (cid:54)= 0, k ∈ N n+2, it can be written as the of Vn+1 of the form U (n+1)T (cid:98) k combination of functions U (n)T (cid:98) l sufficient to consider only k ∈ N n+2 by Definition 1.4 (1e). If U (n+1)T (cid:98) k it is obvious satisfied. If U (n+1)T (cid:98) k MULTIRESOLUTION ANALYSIS FOR MARKOV INTERVAL MAPS 15 =U (n+1) following linear combination: U (n+1)T (cid:98) k  N (cid:99)ϕ(k)N (cid:88) (cid:88) m∈N n+2 an,l k l∈Qn (cid:88) (cid:124) (cid:88) l∈Qn =U (n+1) (cid:88) (cid:88) l∈Qn = = k U (n)T (cid:98) l an,l m∈N n+2 N (cid:99)ϕ(l)N + k U (n)T (cid:98) l an,l N (cid:99)ϕ(l)N + an,l k an,l m m + l∈dn−qn cn,l k cn,l (cid:88)  (cid:125) (cid:88) k U (n+1) (cid:88) (cid:123)(cid:122) m T (cid:98) m N (cid:99)ϕ(m)N + (cid:88) an,l (cid:88) l∈dn−qn cn,l =δk,m l∈dn−qn m∈N n+2 cn,l k ψn,l.  T (cid:98) m N (cid:99)ϕ(m)N (cid:88) cn,l k m∈N n+2 m T (cid:98) m cn,l N (cid:99)ϕ(m)N  m T (cid:98) m cn,l N (cid:99)ϕ(m)N l∈dn−qn l∈Qn If we consider T lψn,k and T rψm,s for n, m ∈ N, n (cid:54)= m, l, r ∈ Z, k ∈ dn − qn, s ∈ dm − qm, the orthonormality follows from T lψn,k ∈ Wn, T rψm,s ∈ Wm and by the definition of Wn, Wm. ogous results. We show that for fixed n ∈ N(cid:110)T N nkψ−n,l : l ∈ d−n − q−n, k ∈ Z(cid:111) Now we consider the closed subspaces Vn of L2(µ) with n < 0 and show the anal- is an orthonormal basis of W−n = V−n+1 (cid:9) V−n. First we show that any function U (−n+1)ϕj can be written as a linear combination of functions U (−n)ϕi and ψ−n,l, i ∈ N, l ∈ d−n − q−n. This linear combination is precisely U (−n+1)ϕj (cid:68) (cid:88) (cid:88) m∈N 2 = m∈N 2 = δl,k. c−n,l m c−n,k m =U (−n+1) n,i j bn,i m + b −n,l c j c−n,l m N (cid:99)ϕ(m)N  (cid:124) (cid:88)  (cid:88)  (cid:88) m∈N 2 i∈Q−n =U (−n+1) (cid:88) (cid:88) i∈Q−n i∈Q−n = = n,i j b i∈Q−n j U (−n)ϕi + n,i b j U (−n)ϕi + n,i b (cid:88) =δj,m l∈d−n−q−n (cid:123)(cid:122) N (cid:99)ϕ(m)N + (cid:88) U (−n+1) (cid:88) m∈N 2 −n,l j m∈N 2 −n,l j c ψ−n,l. m∈N 2 (cid:88) m T (cid:98) m bn,i (cid:88) (cid:88) l∈d−n−q−n c l∈d−n−q−n  T (cid:98) m (cid:125) (cid:88)  −n,l c j m T (cid:98) m c−n,l N (cid:99)ϕ(m)N l∈d−n−q−n m T (cid:98) m c−n,l N (cid:99)ϕ(m)N We have to show the orthonormality only for ψ−n,l and ψ−n,k since T is a unitary operator and U (−n+1)T kϕj are mapped to orthonormal functions. For ψ−n,l and ψ−n,k, l, k ∈ d−n − q−n, the orthonormality follows from (cid:104)ψ−n,lψ−n,k(cid:105) = m U (−n+1)T (cid:98) m c−n,k m U (−n+1)T (cid:98) m c−n,l N (cid:99)ϕ(m)N N (cid:99)ϕ(m)N (cid:69) (cid:12)(cid:12)(cid:12) (cid:88) m∈N 2 16 JANA BOHNSTENGEL AND MARC KESSEBÖHMER Furthermore, it follows that L2(µ) =(cid:76) cl(cid:0)(cid:83) (cid:1) = L2(µ). Consequently, we have that (cid:8)T lψn,k : n ∈ Z, k ∈ dn − qn, l ∈ Z(cid:9) Wn ∪ V0 n∈N0 k∈Z Wk, since we have shown before that is an ONB of L2(µ). Remark 3.4. For the proof of Corollary 1.8 we have to consider the first part of the proof of Theorem 1.6 and show the orthonormality between ψn,k and ϕi in addition, which follows from the construction of the mother wavelets. two-sided multiplicative MRA. 3.2. Abstract multiplicative Multiresolution analysis. In this section we want to consider how the general results simplify if we impose the extra condi- tion of a multiplicative MRA. Recall from the introduction that in the case of Definition 1.4, we say that we have a multiplicative MRA if there exists an operator U such that U (n) = U n for all n ∈ N and U (−n) = (U∗)n, n ∈ N. We then say(cid:0)µ, ((U)n , (U∗)n)n∈N0 ,T(cid:1) allows a Proof of Lemma 1.9. Recall that (cid:8)T lψ0,k : k ∈ d0 − N , l ∈ Z(cid:9) is an orthonormal basis of W0. We have ψ0,k =(cid:80) N (cid:99)ϕ(m)N and we show that for fixed n ∈ N, U nW0 = Wn. First it follows that U nT mψ0,k ∈ Wn ⊂ Vn+1, n ∈ N, m ∈ Z, since The key observation is contained in Lemma 1.9 which we prove first. m UT (cid:98) m m∈N 2 c0,k (cid:88) l∈N 2 l U n+1T (cid:98) l c0,k N (cid:99)+N mϕ(l)N U nT mψ0,k = and (cid:104)U nT mψ0,kU nT rϕi(cid:105) l UT (cid:98) l c0,k =(cid:104)U nT m (cid:88) =(cid:104)U n+1T N m (cid:88) =δm,r · (cid:88) l∈N 2 c0,k l a0,i l∈N 2 l = 0. l T (cid:98) l c0,k l∈N 2 N (cid:99)ϕ(l)NU nT r (cid:88) N (cid:99)ϕ(l)NU n+1T N r (cid:88) j∈N 2 j∈N 2 j UT (cid:98) j a0,i N (cid:99)ϕ(j)N(cid:105) j T (cid:98) j a0,i N (cid:99)ϕ(j)N(cid:105) Consequently, U nW0 ⊂ Wn. Now consider U n+1T mϕj ∈ Vn+1, m ∈ Z, j ∈ N, and we show that this can be written as a linear combination of functions U nT lϕi and U nT rψ0,k, l, r ∈ Z, i ∈ N, k ∈ d0 − N, by considering the scalar product. First we recall that UT (cid:98) k k ψ0,l for k ∈ N 2 from the proof of Theorem 1.6, and hence for k ∈ N 2 N (cid:99)ϕ(k)N =(cid:80) k ϕi +(cid:80) l∈d0−N c0,l i∈N a0,i 1 = (cid:104)UT (cid:98) k a0,i k ϕi + k ψ0,l(cid:105) c0,l N (cid:99)ϕ(k)N(cid:88) (cid:88) i∈N (cid:88) i∈N = a0,i k a0,i k + c0,l k c0,l k . l∈d0−N (cid:88) l∈d0−N MULTIRESOLUTION ANALYSIS FOR MARKOV INTERVAL MAPS 17 It follows that for m ∈ N n+2 with m = k + N 2k1, k ∈ N 2, k1 ∈ N n, we have N (cid:99) = (cid:98) k (cid:98) m k ϕi + U nT k1 (cid:88) (cid:88) l∈d0−N N (cid:99)ϕ(l)N(cid:105) k ψ0,l(cid:105) c0,l i UT (cid:98) i c0,l N (cid:99)ϕ(i)N(cid:105) l UT (cid:98) l (cid:88) a0,i i∈N 2 N (cid:99)ϕ(l)N(cid:105) =(cid:104)U n+1T (cid:98) k N (cid:99) + N k1 and (m)N = (k)N, and so (cid:104)U n+1T (cid:98) k a0,i N (cid:99)+N k1 ϕ(k)NU nT k1(cid:88) N (cid:99)+N k1 ϕ(k)NU nT k1(cid:88) N (cid:99)+N k1ϕ(k)NU nT k1 (cid:88) (cid:88) l∈d0−N l T (cid:98) l (cid:88) a0,i N (cid:99)+N k1ϕ(k)NT N k1(cid:88) N (cid:99)+N k1 ϕ(k)NT N k1 (cid:88) + (cid:104)U n+1T (cid:98) k l∈N 2 c0,l k l∈N 2 c0,l k + (cid:104)T (cid:98) k =(cid:104)T (cid:98) k a0,i k a0,i k i∈N i∈N i∈N l∈d0−N i∈N 2 = a0,i k a0,i k + c0,l k c0,l k (cid:88) l∈d0−N (cid:88) i∈N i T (cid:98) i c0,l N (cid:99)ϕ(i)N(cid:105) =1. Now we notice that we can write any element k ∈ Z as k = k0 + N n+2l for some k0 ∈ N n+2 and l ∈ Z. Consequently, with UT NV0 = T UV0 we obtain the general result for U n+1T kϕj, k ∈ Z, j ∈ N. To obtain W−n = (U∗)n W−1, W−1 = V0 (cid:9) V−1, n ∈ N, we can proceed as above. N (cid:99)ϕ(l)N −1,l j ψ−1,l. First we have from the proof of Theorem 1.6 that ψ−1,k =(cid:80) and that ϕj, j ∈ N, can be represented as ϕj =(cid:80) 0,j U∗ϕi+(cid:80) −1,k T (cid:98) l l∈N 2 c l l∈d−1−N c i∈N b −1,i With these observations we obtain as above that (cid:104)(U∗)n T mϕj (U∗)n−1 T rψ−1,k(cid:105) = 0 and (cid:104)(U∗)n−1 T kϕj (U∗)n(cid:88) i∈N b j ϕi + (U∗)n−1 (cid:88) −1,i l∈d−1−N −1,l j ψ−1,l(cid:105) = 1. c (cid:3) Remark 3.5. If we have UU∗ = id, then W0 = U (W−1). Notice that U is not necessarily injective on W−1. Now we turn to the mother wavelets. Remark 3.6. If U (n) = U n, U (−n) = (U∗)n, then we only consider the mother wavelets for k ∈ d0 − N. So ψ0,k = U (cid:88) l∈N 2 l T (cid:98) l c0,k N (cid:99)ϕ(l)N , where the coefficients are from (3.2) and we define ψk := ψ0,k. For the negative indexed part of the construction we write for k ∈ d−1 − N (cid:88) l∈N 2 ψ−,k = −1,k c l T (cid:98) l N (cid:99)ϕ(l)N . and under condition (3.10) this simplifies to U nT kϕj = N (cid:99)ϕ(l)N (cid:88) l U n+1T N k+(cid:98) l a0,j (cid:88) l∈N 2 i∈N := a0,j l U nT kϕj = i U n+1T N k+jϕi. a0,j 18 JANA BOHNSTENGEL AND MARC KESSEBÖHMER 3.3. Translation completeness. In the following we assume a stronger condition than (1e) of Definition 1.4, namely that the father wavelets are translation complete, i.e. for j ∈ N (3.10) ϕj ∈ spanU(cid:8)T jϕi : i ∈ N(cid:9) . This condition implies that there exist complex numbers a0,j , i ∈ N, such that i (cid:88) i∈N ϕj = i UT jϕi. a0,j We would like to point out that this condition is also satisfied for the particular case of MIM where the father wavelets ϕj, j ∈ N , are chosen to be the scaled characteristic functions on the cylinder sets [j] (see Section 4). We have that (3.2) takes the following form for k ∈ Z, j ∈ N, n ∈ N, j j . We now show that condition (3.10) To simplify the notation we set aj l allows us to simplify the construction of the mother wavelets. Lemma 3.7. Under condition (3.10) one possible choice of the matrix M0 in (3.8) has a block structure consisting of N blocks. (cid:0)ak (cid:1) Proof. Define Qk := {j ∈ N : UT kϕj (cid:54)= 0} and qk := card Qk for each k ∈ N. Then (cid:1) (cid:0)ak j∈Qk is a vector of length qk and we choose qk−1 linearly independent vectors to l ∈ qk\{0}, orthonormal to(cid:0)ak , j∈Qk of length qk. Via the Gram-Schmidt process we obtain vectors j∈Qk j = 0 if j ∈ N\Qk we extend (cid:16) (cid:17)  is a The matrix (cid:99)M0 = (hij)i∈q1,j∈N 2 given with the blocks Mk, k ∈ N, by for k = 0 the vectors j∈Qk matrix of size qk × N. j∈Qk. By setting ck,l . Then Mk := (cid:0)ak (cid:17) l∈qk\{0},j∈N  (cid:1) (cid:17) (cid:16) (cid:16) (cid:17) (cid:16) ck,l j ck,l j j j∈N ck,l j to ck,l j j∈N (cid:1) j (hij)i∈q0,j∈N = M0, for k ∈ N\{0} (hij)i∈(cid:80)k l=0 ql\(cid:80)k−1 l=0 ql,j∈(k+1)N\kN = Mk and otherwise zero satisfies the conditions imposed on M0 in (3.8), i.e. restrict the columns to those in D1, it is unitary, and (cid:100)M0 is of size q1 × N 2 since (cid:80) k∈N qk = q1. We notice that (cid:99)M0 is ordered in a different way than M0, since the rows(cid:0)ak i=0 qi\(cid:80)k−1 (cid:1) simpler form for k = 0, l ∈ q0\{0} and for k ∈ N\{0}, l ∈(cid:80)k (1) If U (n) = U n, U (−n) = (U∗)n and (3.10), the mother wavelets take the i=0 qi, j∈Qk are not grouped in M0. Remark 3.8. if we (cid:3) j as (cid:88) j∈N ψl = j UT kϕj, ck,l MULTIRESOLUTION ANALYSIS FOR MARKOV INTERVAL MAPS 19 where the coefficients are as constructed in Lemma 3.7. For negative in- dexed part we define for k ∈ d−1 − N −1,k c l N (cid:99)ϕ(l)N . ψ−,k = T (cid:98) l (2) In the case of (3.10), or the slightly weaker statement (3.11) U (n)T (cid:98) k N (cid:99)ϕ(k)N = i U (n+1)T N(cid:98) k an,k N (cid:99)+(j)N ϕi (cid:88) (cid:88) l∈N 2 i∈N we can obtain the coefficients for the mother wavelets by constructing for each k ∈ N n with U (n)T kϕj (cid:54) (cid:54)= 0 for at least one j ∈ N a matrix of size qn,k × qn,k, where qn,k = card{j ∈ N : U (n)T kϕj (cid:54) (cid:54)= 0} instead of one unitary matrix of size dn × dn. In this way we need at most N n matrices on the scale n ∈ N. i Now we turn to a correspondence to the construction of a wavelet basis for MIM. The next proposition shows how the incidence matrix of MIM plays a role in the MRA. holds that a0,j k ∈ Z, U nT kϕj (cid:54)= 0 if and only if for all i = 0, . . . , n − 2, UT ki+1ϕki (cid:54) (cid:54)= 0 if and only if UT jϕi (cid:54)= 0, i, j ∈ N, then we have for n ∈ N, (cid:54)= 0 and Proposition 3.9. In the case of U (n) =(cid:0)U (1)(cid:1)n, n ∈ N0, (3.10) and if it further UT k0 ϕj (cid:54)= 0, where k =(cid:80)n−1 iteratively. Notice that U 2T k0+N k1ϕj = UT k1(cid:0)UT k0ϕj UT k1ϕk0 = UT k1(cid:88) i UT k0ϕi = U 2T N k1+k0(cid:88) Proof. We prove this for k = k0 + N k1, k0, k1 ∈ N. The general result follows U 2T k0+N k1 ϕj (cid:54)= 0 it follows that UT k0ϕj (cid:54)= 0. Besides we have that (cid:1). Consequently, from i=0 kiN i + lN n, ki ∈ N , i ∈ n, and l ∈ Z. i ϕi (cid:54)= 0 ak0 ak0 i∈N i∈N if U 2T k0+N k1 ϕj (cid:54)= 0. If we assume that UT k1ϕk0 (cid:54)= 0 and UT k0ϕj (cid:54)= 0 then U 2T N k1+k0 ϕj = UT k1UT k0 ϕj (cid:16) (cid:16) ak0 j ak0 j = = ϕk0 − (cid:88) (cid:17)−1 UT k1 UT k1 ϕk0 − (cid:88) (cid:17)−1 i∈N\{j} i∈N\{j}  i UT k0ϕi ak0 i U 2T N k1+k0ϕi ak0  (cid:54)= 0, since UT k1ϕk0 −(cid:80) c(cid:0)U (1)(cid:1)n T kϕj for all n ∈ N, k ∈ N n and j ∈ N and (3.10), where c may depend (cid:3) Remark 3.10. We can show the same result if for some c ∈ R, we have U (n)T kϕj = i U 2T N k1+k0 ϕi = ak0 j UT N k1+k0ϕj (cid:54)= 0. i∈N\{j} ak0 on n, k, j. Under the conditions of Proposition 3.9 we can give a N × N matrix A, which coincides with the incidence matrix in the case of MIM given by A = (Aij)i,j∈N with (cid:40) Aij := if UT iϕj = 0, else. 0, 1, 20 JANA BOHNSTENGEL AND MARC KESSEBÖHMER 4. Applications to Markov Interval Maps 4.1. Multiresolution Analysis for MIM. Now we apply the results of Section 3 to Markov Interval Maps. More precisely, we construct a wavelet basis on the L2-space of a limit set of a Markov Interval Map translated by Z with respect to a measure. First we consider the case where we do not have any relation between the measures of νZ ([ij]) and νZ ([i]), νZ ([j]). In this case we cannot define only one operator U, but on each scale n ∈ Z we consider a different operator U (n). n∈Z and T are defined in (1.3), (1.4) and (1.2) respectively. (1) Notice that in general we have U (1)U (1) (cid:54)= U (2) since the multiplicative νZ([kij]) for U (2) on the cylinder νZ([k) νZ([kl]) for U (1)U (1) and νZ([ij]) Remark 4.1. Consequently, we obtain a family of operators(cid:0)U (n)(cid:1) The operators(cid:0)U (n)(cid:1) (cid:113) νZ([j]) (cid:113) νZ([j]) (3) The operators(cid:0)U (n)(cid:1) constant sets may differ. (2) The operator T is unitary. n∈Z. U (n)f ∈ L2(νZ). Define the N father wavelets as ϕi := (µ([i])) Remark 4.2. Notice that for ω ∈ Σn N nl, l ∈ Z, we have (cid:40) n∈Z are well defined, namely for f ∈ L2(νZ) we have −1/2 A, j ∈ N and k ∈ Z with k =(cid:80)n−1 1[i] for i ∈ N. i=0 ωn−1−iN i + −1/2(cid:88) j∈N ϕi(x) = j∈N = (νZ([i])) (cid:115) (cid:88) (cid:115) (cid:88) =U (1)T i(cid:88) j∈N = j∈N νZ([ij]) νZ([i]) (cid:115) 1[ij](x) (cid:115) (cid:88) k∈N νZ([ij]) νZ([i]) ϕj(x). νZ([ij]) νZ([i]) νZ([j]) νZ([ij]) · (µ([j])) −1/2 1[ij](x) νZ([k]) νZ([ik]) 1[ik](x) · ϕj (cid:0)τ−1 (x)(cid:1) i (4.1) U (n)T kϕj = Now we turn to the proof of the properties of(cid:0)U (n)(cid:1) −1/2 T l1[ωj], 0, (νZ([ωj])) if Aωn−1j = 0, else. n∈Z and T stated in Propo- sition 1.3. Proof of Proposition 1.3. ad (1): Let n ∈ N, f ∈ L2(νZ), x ∈ R, then T U (n)f (x) νZ([j]) νZ([ωj]) 1[ωj](x − 1 − k) · f ω (x − 1 − k) + τ−1 ωn−1−iN i + N nk (cid:32) (cid:32) n−1(cid:88) n−1(cid:88) i=0 i=0 ωn−1−iN i + N nl − N n (cid:33) (cid:33) νZ([j]) νZ([ωj]) 1[ωj](x − l) · f ω (x − l) + τ−1 (cid:88) (cid:88) k∈Z (cid:88) (cid:88) ω∈Σn A = = (cid:115) (cid:115) (cid:88) (cid:88) j∈N j∈N l∈Z ω∈Σn =U (n)T N n A f (x). ad (3): Let i ∈ N, x ∈ R, then MULTIRESOLUTION ANALYSIS FOR MARKOV INTERVAL MAPS 21 (cid:33) ωn−1−iN i (cid:115) νZ([ωj]) νZ([l]) ϕj (cid:32) x − n−1(cid:88) (cid:32) x − n−1(cid:88) i=0 i=0 νZ([ωj]) νZ([l]) ϕj ωn−1−iN i − N nk (cid:33) . ad (2): Notice that for n ∈ N, l ∈ N, k ∈ Z, j∈N and A:ω0=l ω∈Σn (cid:88) U (−n)ϕl(x) = (cid:88) (cid:115) ad (4): Let n ∈ N and k = (cid:80)n−1 l =(cid:80)n−1 i=0 ωn−1−iN i + N nl1, ω ∈ Σn U (−n)T kϕl(x) = (cid:88) (cid:88) ω∈Σn A:ω0=l j∈N i, j ∈ N then Consequently, T N nkU (−n)ϕj = U (−n)T kϕj for all k ∈ Z, n ∈ N, j ∈ N. i=0 ωn−1−iN i + N nk1, ω ∈ Σn A, k1 ∈ Z and A, l1 ∈ Z and Aωn−1i = 1 and Aωn−1j = 1 for (cid:104)U (n)T kϕiU (n)T lϕj(cid:105) = (cid:104)(νZ([ωi])) −1/2 T k1 1[ωi] (νZ([ωj])) −1/2 T l1 1[ωj](cid:105) = δk1,l1δ(ω,i),(ω,j). Otherwise, we have U (n)T kϕi = 0 or U (n)T lϕj = 0. Furthermore for n ∈ N, k, j ∈ Z, i, m ∈ N, we have = ω∈Σn (cid:104)U (−n)T kϕiU (−n)T lϕm(cid:105) νZ([ωj]) νZ([i]) (cid:115) (cid:68) (cid:88) (cid:88) (cid:115) (cid:88) (cid:88) =δk,l · δi,m · (cid:88) (cid:88) A:ω0=m ω∈Σn A:ω0=i j∈N j∈N νZ([ωj]) νZ([m]) T ω∈Σn A:ω0=i j∈N =δ(k,i),(l,m), νZ([ωj]) νZ([i]) (cid:80)n−1 i=0 ωn−1−iN i+N nkϕj (cid:80)n−1 i=0 ωn−1−iN i+N nlϕj T (cid:69) where we used in the second equality that (cid:104)T kϕjT lϕi(cid:105) = δ(k,j),(l,i). (cid:115) ad (5): Let n ∈ N, f ∈ L2(νZ), x ∈ R, then U (n)U (−n)f (x) (cid:88) l∈Z f (cid:88) A r∈N νZ([r]) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) νZ([(cid:101)ωr]) 1[(cid:101)ωr](x − l) (cid:32) (cid:101)ω∈Σn n−1(cid:88) (cid:101)ωn−1−iN i + N nl − n−1(cid:88) τ−1(cid:101)ω (x − l) + (cid:32) (cid:32) (cid:101)ωn−1−iN i + N nl − n−1(cid:88) n−1(cid:88) τ−1(cid:101)ω (x − l) + (cid:88) (cid:88) ω∈Σn j∈N k∈Z τω 1[ωj](x − k) · f (x) i=0 i=0 i=0 i=0 A = = 1[j] k∈Z ω∈Σn A j∈N (cid:115) νZ([ωj]) νZ([j]) (cid:33) (cid:33) (cid:33) + k ωn−1−iN i − N nk ωn−1−iN i − N nk =f (x), where we used in the third equality that i = r, ω =(cid:101)ω and k = l since otherwise it is zero. (cid:17) (cid:33) (cid:3) JANA BOHNSTENGEL AND MARC KESSEBÖHMER 22 l ∈ Z, with k =(cid:80)n−1 ad (6): For n ∈ N, k ∈ Z, j ∈ N, x ∈ R, with U (n)T kϕj (cid:54)= 0, there is ω ∈ Σn A, i=0 ωn−1−iN i + N nl and so U (−n)U (n)T kϕj(x) = U (−n)(cid:16) = T N nlU (−n)(cid:16) (cid:80)n−1 (νZ([ωj])) (cid:17) −1/2 T l1[ωj](x) (νZ([ωj])) −1/2 1[ωj](x) −1/2 i=0 ωn−1−iN i (νZ([j])) 1[j](x) = T N nlT = T kϕj(x). (cid:88) (cid:88) k∈Z ωj∈Σn+1 A 1[j] (cid:32) x − n−1(cid:88) i=0 Remark 4.3. We further notice that for n ∈ N, x ∈ R, we have f ∈ L2(νZ), U (−n)U (n)f (x) = ωn−1−iN i − N nk · f (x), and consequently, in general we do not have U (−n)U (n) = id. Now we can turn to the proof of Theorem 1.5. (cid:110) (cid:111) . Proof of Theorem 1.5. We show the properties (2a) to (2e) of Definition 1.4 with the father wavelets ϕi = (ν([i]) 1[i], i ∈ N. −1/2 We define the closed subspaces of L2(νZ) for j ∈ N as V0 := cl span(cid:8)T kϕi : k ∈ Z, i ∈ N(cid:9) , Vj := cl span U (n)T kϕj = i=0 ωn−1−iN i+ ad (2d): We notice that for ω ∈ Σn ad (2c): By the definition of Vj we obviously have that(cid:8)U (j)T kϕi : k ∈ Z, i ∈ N(cid:9) U (j)T kϕi : k ∈ Z, i ∈ N spans Vj, j ∈ Z. The orthonormality follows from Proposition 1.3 (4). N nl, l ∈ Z, we have A, j ∈ N and k ∈ Z with k =(cid:80)n−1 (cid:115) (cid:88) A so that k = (cid:80)n−1 i=0 ωn−1−iN i + N nl, l ∈ Z, then ad (2a): Notice that for n ∈ N, k ∈ Z and i ∈ N we obtain with Proposition 1.3 (cid:115) U (n)T kϕi = U (n)T kU (1)T i(cid:88) (cid:115) = U (n)U (1)T N k+i(cid:88) If there is not such an ω ∈ Σn U (n)T kϕj = 0. νZ([ωji]) νZ([ωj]) U (n+1)T N k+jϕi. (2) and (3) that νZ([i]) νZ([ij]) j∈N i∈N ϕj νZ([i]) νZ([ij]) ϕj. j∈N So it follows that Vn ⊂ Vn+1 by Remark 4.1 (1). ad (2b): First we notice that X is either totally disconnected or we can consider X as an interval in [0, 1]. Furthermore, every characteristic function on a cylinder [ω] ⊂ ΣA can be obtained by U (n)T kϕj, n ∈ N0, k ∈ Z, j ∈ N. Thus, we are left to show that(cid:8)T k1[ω] : k ∈ Z, ω ∈ Σ∗ If X is totally disconnected, it follows by the Stone-Weierstrass Theorem that (cid:9) is dense in L2(νZ). (cid:9) (cid:8)T k1[ω] : k ∈ Z, ω ∈ Σ∗ A A MULTIRESOLUTION ANALYSIS FOR MARKOV INTERVAL MAPS in L2(νZ) and so cl span(cid:8)T k1[ω] : k ∈ Z, ω ∈ Σ∗ is dense in C(R, C), see e.g. [KSS07]. Besides it is well known that C(R, C) is dense If X = [a, b], notice that every interval I ⊂ [0, 1] can be approximated by τω(X), ω ∈ Σ∗ A, generates B, thus every element A ∈ B can be approximated by elements of {τω(X) : ω ∈ Σ∗ A}. Consequently, every elementary function can be approximated by functions 1τω(X) and so all functions in L2(νZ) can be approximated by elements of A. Hence τω(X), ω ∈ Σ∗ (cid:9) = L2(νZ). 23 A (cid:8)T k1[ω] : ω ∈ Σ∗ A, k ∈ Z(cid:9) = (cid:110) Consequently, cl(cid:83) ad (2e): This follows from Proposition 1.3 (1) and (2). k∈N Vn = L2(νZ) . (cid:111) . (cid:3) U (n)T lϕi : n ∈ N0, l ∈ Z, i ∈ N Next we prove the forward direction of Theorem 1.11. The backward direction will be shown in Section 4.3. Proof of Theorem 1.11 "=⇒". We assume that sided MRA with the father wavelets ϕi = (νZ([i]) holds by (2d) of Definition 1.4 that for n ∈ N (cid:17) U (−n) {ϕi : i ∈ N} ⊂ span U (−n+1)(cid:8)T kϕi : i ∈ N , k ∈ N(cid:9) . νZ,(cid:0)U (n)(cid:1) n∈Z , T (cid:16) −1/2 allows a two- 1[i]. Then in particular, it We further notice that for n ∈ N, k, i ∈ N, U (−n)ϕk = ω∈Σn A:ω0=k (cid:88) (cid:88) (cid:115) (cid:88) (cid:115) j∈N (cid:88) ω∈Σn−1 A :ω0=i j∈N and U (−n+1)T kϕi = (cid:80)n−1 l=0 ωn−1−lN l ϕj νZ([ωj]) νZ([k]) T (cid:80)n−2 l=0 ωn−2−lN l+N n−1kϕj. νZ([ωj]) νZ([i]) T From the precise from of U (−n)ϕk and U (−n+1)T mϕi, n ∈ N, k, m, i ∈ N, it follows that (cid:104)U (−n)ϕkU (−n+1)T mϕi(cid:105) (cid:54)= 0 only if m = k since (cid:115) (cid:88) (cid:104)U (−n)ϕkU (−n+1)T mϕi(cid:105) νZ([ωj]) νZ([k]) (cid:88) j∈N A:ω0=k ω∈Σn =(cid:104) (cid:88) (cid:88) (cid:88) (cid:80)n−1 (cid:88) ω∈Σn A:ω0=k (cid:104)T ω∈Σn−1 = =δm,k A :ω0=i νZ([ωj]) νZ([i]) T j∈N (cid:115) (cid:88) (cid:101)ω∈Σn−1 A :(cid:101)ω0=i (cid:115) ϕj1T (cid:88) (cid:88) j1∈N l=0 ωn−1−lN l (cid:80)n−1 l=0 ωn−1−lN l ϕj T (cid:80)n−2 l=0 ωn−2−lN l+N n−1mϕj(cid:105) (cid:115) (cid:88) νZ([(cid:101)ωj2]) (cid:80)n−2 l=0 (cid:101)ωn−2−lN l+N n−1mϕj2(cid:105) νZ([ωj1]) νZ([k]) (cid:115) νZ([i]) j2∈N (cid:115) j∈N ω∈Σn−1 A :ω0=i νZ([k]) νZ([kωj]) νZ([ωj]) νZ([i]) , where we used in the third equality the property of Proposition 1.3 (4), namely (cid:104)T kϕj1T lϕj2(cid:105) = δ(k,j1),(l,j2) for any k, l ∈ Z and j1, j2 ∈ N. T l=0 ωn−1−lN l (cid:80)n−1 (cid:80)n−2 l=0 ωn−2−lN l+N n−1kϕj ϕj (cid:80)n−2 l=0 ωn−2−lN l+N n−1kϕj. T νZ([kωj]) νZ([k]) (cid:115) νZ([ωj]) νZ([k]) νZ([kωj]) νZ([k]) T (cid:115) (cid:88) j∈N j∈N A:ω0=k ω∈Σn (cid:88) (cid:88) (cid:115) (cid:88) (cid:88) (cid:88) (cid:88) (cid:113) νZ([kωj]) ω∈Σn−1 ω∈Σn−1 j∈N i∈N A = = A :ω0=i νZ([kωj]) = νZ([ωj]) 24 JANA BOHNSTENGEL AND MARC KESSEBÖHMER As a consequence of (2c), (2d) of Definition 1.4 and the observation above it ∈ CN such that follows that for every n ∈ N, k ∈ N there exist unique αn,k (cid:16) (cid:17) i i∈N i U (−n+1)T kϕi αn,k (cid:88) (cid:88) αn,k i ω∈Σn−1 A :ω0=i j∈N (cid:115) (cid:80)n−2 l=0 ωn−2−lN l+N n−1kϕj. νZ([ωj]) νZ([i]) T On the other hand, from the precise form of U (−n)ϕk it follows that U (−n)ϕk = = (cid:88) (cid:88) i∈N i∈N U (−n)ϕk = (cid:113) νZ([ωj]) By comparing the coefficients it follows that for every ω ∈ Σn−1 αn,k νZ([k]) . Consequently, αn,k νZ([i]) = i A , ω0 = i, we have (cid:16) (cid:17)2 νZ([k]) i ∈ R+ and αn,k i . νZ([i]) Now it remains to be shown that cn,k with ω0 = i and k ∈ N it follows that i νZ([kω]) = νZ([kωj]) = (cid:88) j∈N are independent of n ∈ N. For n ∈ N, ω ∈ Σn A (cid:88) (cid:16) (cid:17)2 αn,k i νZ([k]) (cid:16) αn,k i j∈N νZ([i]) νZ([ωj]) (cid:17)2 A with ω0 = i as ω =(cid:101)ωωn−1 for a suitable νZ([k]) . νZ([i]) = νZ([ω]) On the other hand we can write ω ∈ Σn (cid:101)ω ∈ Σn−1 A , (cid:101)ω0 = i, and so (cid:16) (cid:17)2 αn−1,k i νZ([k]) νZ([i]) νZ([kω]) = νZ([k(cid:101)ωωn−1]) = νZ([(cid:101)ωωn−1]) (cid:16) (cid:17)2 αn−1,k i νZ([k]) . νZ([i]) = νZ([ω]) i i i = αm,k = αn,k i for αn,k := (cid:0)αk and so αn,k . for all n, m ∈ N, k, i ∈ N. In the following Thus, αn−1,k (cid:1)2 we write αk νZ([k])/νZ([i]) for k, i ∈ N, then we have νZ([kωj]) = κk,ω0νZ([ωj]) for all ω ∈ Σ∗ A, j, k ∈ N. From this property we conclude the Markov (cid:88) relation since for any k, i ∈ N Define κk,i (cid:88) i i i ν([ki]) = ν([kij]) = κk,iν([ij]) = κk,iν([i]) j∈N j∈N MULTIRESOLUTION ANALYSIS FOR MARKOV INTERVAL MAPS 25 and so Define πki := κk,iν([i])/ν([k]) =(cid:0)αk κk,iν([i]) (cid:1)2, then πki is a incidence probability. Conse- ν([k]) . ν([ki]) = ν([k]) quently, we have that if a two-sided MRA holds then the measure ν is Markovian. (cid:3) The reversed implication will be shown in Section 4.3. i 4.2. Mother wavelets for MIM. In this section we are in the case of Remark 3.8 (2) and so we consider for each father wavelet ϕi, i ∈ N, a matrix of coefficients; more precisely on each scale we have to consider for each element of the alphabet for N a matrix of coefficients. We slightly change the notation from cn,k,l ω ∈ Σn A, since the information about n and k are coded; n is given by the length of A we need a matrix of size qωn × qωn, where qωn = card{j ∈ N : ∈ C, j ∈ N, k ∈ qωn\{0}, such that the a word and k =(cid:80)n−1 For ω ∈ Σn+1 i=0 ωn−1−iN i. Aωnj = 1}. First we determine cω,k (qωn × qωn)-matrix to cω,l j j where Dωn = {j ∈ N : Aωnj = 1} is unitary. This is done as explained above via the Gram-Schmidt process. i=0 ωn−iN i the basis functions as: for l ∈ qωn\{0} We define for ω ∈ Σn+1  , j∈Dωn k∈qωn\{0},j∈Dωn Mω := j j (cid:17) (cid:16)(cid:112)νZ([ωj]) (cid:17) Aωnjcω,k  (cid:16) A , k =(cid:80)n ψω,l = U (n)T k(cid:88) (cid:88) j∈N Aωnjcω,k j ψω,k = j∈N These functions can be written differently for ω ∈ Σn+1 A , k ∈ qωn\{0}, as Aωnjcω,l j ϕj. · (νZ([ωj])) −1/2 · 1[ωj]. From Theorem 1.6 and Theorem 1.5 the following corollary follows. Corollary 4.4. An orthonormal basis for L2(νZ) is given by (cid:8)T lψω,k : l ∈ Z, ω ∈ Σ∗ cl span(cid:8)T lψω,k : l ∈ Z, ω ∈ Σn A, k ∈ {1, . . . , qωω−1 − 1}(cid:9) ∪(cid:8)T lϕj : l ∈ Z, j ∈ N(cid:9) . A, k ∈ {1, . . . , qωn−1 − 1}(cid:9) = Vn (cid:9) Vn−1. Remark 4.5. In fact, the proofs of Theorem 1.6 and Theorem 1.5 show that we have for n ∈ N 4.3. MRA for a Markov measures. In this section we construct a wavelet basis on the limit set translated by Z where the underlying measure ν is Markovian. For this fix a probability vector p = (p0, p1, . . . , pN−1) and a (N × N ) stochastic matrix Π = (πjk)j,k∈N such that for ω ∈ Σn A we have n−2(cid:89) ν([ω]) = pω0 πωiωi+1. i=0 Furthermore, we have that πjk = 0 if Ajk = 0. This is a special case of the one in the last section. Therefore, we omit some proofs here and mainly state the results, so that the differences become clear. In this construction we only have to define one operator U since we obtain U (n) by U n, i.e. by iteration of U. Another main difference is that we do not need one matrix for every ω ∈ Σ∗ A to obtain the mother wavelets, but we only need matrices JANA BOHNSTENGEL AND MARC KESSEBÖHMER 26 for ω ∈ Σ1 1.9. A = N. So we need not more than N 2 matrices. This follows from Lemma The setting is as defined in Section 2. Set U := U (1)and so it takes the form in (1.1). By the Markov property we have νZ([i]) and hence one easily verifies that U (n) = U n. Also notice that U is not unitary unless we have that Aij = 1 for all i, j ∈ N. νZ([ji]) = pi pj πji Now we turn to the form of U∗. Lemma 4.6. U∗ has the form (4.2) U∗f (x) = (cid:88) (cid:88) (cid:88) k∈Z j∈N i∈N (cid:114) pjπji pi · 1[i](x − j − N k) · f (τj(x − j − N k) + k). Remark 4.7. Notice that U∗ = U (−1) and (U∗)n = U (−n). Proof. To prove that U∗ has the form above we use the Z-translation invariance of the measure νZ and the fact that dνZ◦τj on [i]. We obtain this Radon- Nikodym derivative since for a cylinder set [ω], ω ∈ Σ∗ A, we have dνZ = pj πji pi νZ (τj([ω])) = pjπjω0 πωiωi+1 n(cid:89) i=0 and νZ([ω]) = pω0 i=0 πωiωi+1. Consequently, we obtain that for f, g ∈ L2(νZ) (cid:104)U fg(cid:105) j∈N i∈N (cid:81)n (cid:114) pi (cid:88) (cid:88) (cid:114) pi (cid:88) (cid:88) (cid:114) pi (cid:88) (cid:88) (cid:114) pi (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) i∈N i∈N i∈N j∈N j∈N j∈N pjπji pjπji pjπji k∈Z (cid:88) (cid:88) (cid:88) (cid:88) k∈Z k∈Z k∈Z = = = = = f (x) k∈Z j∈N i∈N =(cid:104)fU∗g(cid:105), with U∗g as in (4.2). · 1[ji](x − k) · f (τ−1 j (x − k) + j + N k)g(x)dνZ(x) · 1[ji](x) · f (τ−1 j (x) + j + N k)g(x + k)dνZ(x) · 1[ji](τj(x)) · f ((x) + j + N k)g(τj(x) + k)dνZ(τj(x)) pjπji (cid:114) pjπji pi · 1[ji](τj(x)) · f ((x) + j + N k)g(τj(x) + k) · pjπji pi · dνZ(x) · 1[i](x − j − N k) · g(τj(x − j − N k) + k)dνZ(x) (cid:3) Now we turn to the definition of the father wavelets which we use in the MRA. Define the N father wavelets as ϕi = (νZ([i])) Remark 4.8. Notice that the family of father wavelets (ϕi)i∈N is orthonormal by definition. 1[i] for i ∈ N. −1/2 Now we turn to the properties of the operators U and T given in Proposition 1.1. Proof of Proposition 1.1. We have that (1), (2), (3) and (4) follow directly from Proposition 1.3 since it is a special case of U (n) in the section above. MULTIRESOLUTION ANALYSIS FOR MARKOV INTERVAL MAPS 27 (cid:88) (cid:88) ad (5): This proof is analogous to the one of (4) or Proposition 1.3 (5). We obtain for f ∈ L2(νZ) k∈Z(cid:80) (cid:80) j∈N all i, j ∈ N. U∗U f (x) = Aji1[i](x − j − N k) · f (x). (cid:80) i∈N Aji1[i](x − j − N k) = 1 for all x ∈ R if and only if Aji = 1 for (cid:3) j∈N k∈Z Now we turn to the proof of the backward direction of Theorem 1.11. Some of the properties follow directly from the proof of Theorem 1.5. Proof of Theorem 1.11 "⇐=". We show the properties (1a) to (1f) of Definition 1.4. The property (1b) follows from Theorem 1.5. ad (1e): For n ∈ N0 it follows directly from Theorem 1.5. For n ∈ Z, n < 0, x ∈ R, k ∈ N, it follows by n i∈N ϕk(x) (cid:88) (cid:118)(cid:117)(cid:117)(cid:116)n−2(cid:89) (cid:88) (cid:32) (cid:88) x − n−2(cid:88) n A :ω0=k √ n−1 A ω∈Σ πk,j l=1 (U∗) = = = ω∈Σ (cid:88) j∈N ϕi (cid:88) j∈N ωn−2−lN l − kNn−1 √ l=0 πk,j (U∗) n−1 T kϕj(x). x − n−1(cid:88) l=0 ωn−1−lN l  πωl,ωl+1 · πk,ω1πωn−1,iϕi (cid:88) (cid:118)(cid:117)(cid:117)(cid:116)n−3(cid:89) i∈N :ω0=j l=1 πωl,ωl+1 · πj,ω1πωn−2,i (cid:33) (cid:88) j∈N ad (1a): For n ∈ N0 it follows directly from Theorem 1.5. For n ∈ Z, n < 0, k ∈ N, it follows from (U∗) n ϕk = √ πk,j (U∗) n−1 T kϕj. ad (1c): We have that(cid:84) grows in n ∈ N. More precisely, for j ∈ N νZ (supp ((U∗)n ϕj)) = Consequently, {0} =(cid:84) n∈Z Vn = {0} , because the support of (U∗)n ϕj, j ∈ N, (cid:88) : ω0 = j, ωn = i(cid:9)(cid:1) . νZ ([i])(cid:0)card(cid:8)ω ∈ Σn+1 j∈Z Vj since any function f ∈(cid:84) j∈Z Vj must be constant for i∈N A every n ∈ N on supp ((U∗)n ϕj), for j ∈ N. ad (1d): This property follows directly from the definition of the spaces Vj and Proposition 1.3 (4) with the observation that U (n) = U n and U (−n) = (U∗)n, n ∈ N0. (cid:3) ad (1f): This property follows from Proposition 1.1 (4) and (5). Remark 4.9. Now we give some remarks concerning the father wavelets. (1) The relation for the functions ϕi, i ∈ N, can also be written as (cid:88) l∈N (cid:0)U T lϕj (cid:1) (ϕj)t j∈N = Ml j∈N , (cid:40)√ 28 JANA BOHNSTENGEL AND MARC KESSEBÖHMER (cid:40) where the Ml are (N × N )-matrices with (Ml)n,k = n, k ∈ N. l ∈ Z, i.e. k is in the N-adic expansion. Then we obtain (2) Notice that for k ∈ Z we can write k = a0 + N l, where a0 ∈ N and some 0, else, πlk, n = l, for 0, (pa0 · πa0j) U T kϕj = (3) Notice that in (cid:8)U nT kϕi : n ∈ N, k ∈ Z, i ∈ N(cid:9) some functions are con- in the N-adic expansion, k =(cid:80)n−1 stantly zero. These functions are precisely those where for k ∈ Z written j=0 kn−1−jN i + lN n, kj ∈ N, l ∈ Z, either Akj kj+1 = 0 for some j ∈ {0, . . . , n − 2} or Akn−1i = 0. −1/2 T l1[a0j], if Aa0j = 0, else. 4.3.1. Mother wavelets for Markov measures. The construction of the mother wavelets simplifies in this setting because we only have to consider mother wavelets for one scale and obtain the other by iterative application of the operators U and T by Lemma 1.9. The mother wavelets are constructed via N matrices as given in Lemma 3.7 and so the mother wavelets are defined for k ∈ N and l ∈ qk\{0}, by ψk,l = U T k(cid:88) j∈N Akjck,l j ϕj (3) Alternatively we can define the mother wavelets as the elements of the k∈N qk ≤ N 2. In the case of N 2 mother wavelets we are back in the case of fractals given by an IFS. for coefficients ck,l Remark 4.10. j ∈ C as in Lemma 3.7. √ l=1 AkiAkjck,l (1) The number of mother wavelets we obtain is(cid:80) (2) Notice that(cid:80)qk−1 vector(cid:0)ψk,l(cid:1)t (cid:88) (cid:18)(cid:16) (cid:118)(cid:117)(cid:117)(cid:116)pω0 (cid:17) n−2(cid:89) (νZ([ωj]))1/2 = l∈{1,...,qk−1} = l∈qk\{0},j∈N (cid:88) Akjck,l j i ck,l πi(i+1) Akjck,i j j + √ πki πkj = δi,j. (cid:19)(cid:0)U T kϕj (cid:1)t j∈N . √ Akjck,i j πωn−1j = 0, (4) Here we can see that we only need mother wavelets for W0 since i=1 j∈N j∈N which was the crucial condition in the case of the last section. Corollary 4.11.(cid:8)U nT mψk,l : n ∈ N0, m ∈ Dn,k, k ∈ N , l ∈ qk\{0}(cid:9) ∪(cid:8)(U∗)n T mψk,l : n ∈ N, m ∈ Z, k ∈ N , l ∈ qk\{0}(cid:9) ∪(cid:8)(U∗)n T kϕj : n ∈ N, k ∈ NZ + l, j, l ∈ N , Ajl = 0(cid:9) gives an ONB for L2(νZ), where (cid:40) n−1(cid:88) m ∈ Z : m = (cid:41) and Aω0k = 1, l ∈ Z i=0 . Dn,k = Remark 4.12. ωn−1−iN i + N nl, ωi ∈ N , (ω0, . . . , ωn−1) ∈ Σn A MULTIRESOLUTION ANALYSIS FOR MARKOV INTERVAL MAPS 29 (1) Because of U W−1 = W0 we only have to add those functions T kϕj, k ∈ Z, j ∈ N, with U T kϕj = 0 to the basis of U∗ (W0) to obtain a basis of W−1. (2) Notice that ψk,l = U T k(cid:88) (cid:88) i∈N Akick,l i = Akick,l i ϕi · (pk · πki) −1/2 · 1[ki]. (cid:1) i∈N matrix are considered, i.e. the IFS has the form(cid:0)τi(x) = x+i 4.4. Examples. In the construction of [MP09] only Cantor sets with incidence i∈N, and there exists a incidence matrix A. The limit set has then the Hausdorff dimension δ = dimH (X) = log N , where r(A) is the spectral radius of A. So we consider the δ-dimensional log r(A) Hausdorff measure µ restricted to the by Z translated set X. It follows that pj = µ([j]) and πij = N−2δpj . Consequently, in this case we can rewrite our conditions for obtaining the coefficients of the mother wavelets in a simpler way. pi N More precisely, for k ∈ N instead of(cid:88) (cid:88) we obtain the condition j∈N √ πkj = 0 Akjck,i j √ pj = 0. Akjck,i j j∈N Although the basis in [MP09] is only given in terms of the representation of a Cuntz-Krieger algebra we can now give a scaling operator U in the sense of (1.1) for this case. More precisely, we obtain 1[j](x − k) · f (τ−1 j (x − k) + j + N k). U f (x) = N δ(cid:88) (cid:88) k∈Z j∈N Proof of Example 1.2: We clearly have that the β-transformation belongs to the class of Markov measures. Consequently, we have that (µ, U, T ) allows a MRA. We can construct the mother wavelets along the lines of Section 4.3. Since we have that in this case d0 = 2 and d1 = 1 we only have to construct coefficients for ϕ0 to obtain the mother wavelets. These coefficients are given in the following matrix which is unitary: Thus, the mother wavelet is ψ = U(cid:0)√ √ (cid:18) √ (cid:19) √ 2 − β β − 1 2 − β −√ β − 1 2 − βϕ0 − √ . β − 1ϕ1 (cid:1). To obtain the basis we further notice that U T ϕ1 = 0 and so we have to keep T kϕ1, k ∈ 2Z + 1 in the basis. 5. Operator algebra In the case of one father wavelet we obtain a so-called low-pass filter function and high-pass filter functions, in terms of which the mother wavelets are given. Via these filter functions we obtain a representation of the Cuntz algebra ON, where N is the number of filter functions. In the case of multiwavelets we can obtain weaker relations. Here we restrict to the case of MIM with underlying Markov measure as treated in Section 4.3. These results are in correspondence to results in [BFMP10]. The relations for the father and the mother wavelets can be written in the fol- For the following we introduce for z ∈ T := {ω ∈ C : ω = 1} the low-pass filter lowing way: H(z) =(cid:0)√ πklzk(cid:1) l,k∈N 30 and for each k ∈ N and z ∈ T the high-pass filter JANA BOHNSTENGEL AND MARC KESSEBÖHMER (cid:16) l zk(cid:17) With these definitions we obtain the following immediate lemma. Gk(z) = Aklck,j j∈qk\{0},l∈N . j∈N, then φ = U H(T )φ and let ψk =(cid:0)ψk,j(cid:1)t Lemma 5.1. Let φ = (ϕj)t j∈qk\{0} for k ∈ N, then ψk = U Gk(T )φ, where the operators U and T are applied to evey entry in the vector. Remark 5.2. It follows that for z ∈ T (cid:88) j∈N √ πkjπljzl−k  k,l∈N H(z)H t(z) = and for k ∈ N, z ∈ T, These filter functions lead us to the definitions of certain "isometries". Definition 5.3. For z ∈ T and f = (f0, . . . , fN−1), fj ∈ L2(T, λ), define and for k ∈ N, z ∈ T, SH f (z) = Gk(z)Gt k(z) = I. √ N H t(z)f(cid:0)zN(cid:1) k(z)f(cid:0)zN(cid:1) . SGk f (z) = Gt For these "isometries" we have the following properties. Proposition 5.4. The following relations hold: (1) S∗ (2) S∗ (3) S∗ (4) S∗ H SH = I, SGk = I, k ∈ N, Gk H SGk = 0 and S∗ SGj = 0, i, j ∈ N, i (cid:54)= j . Gk Gi SH = 0, k ∈ N, Remark 5.5. Realize that for z ∈ T, f = (f0, . . . , fN−1), fj ∈ L2(T, λ), and for k ∈ N, z ∈ T, f = (f0, . . . , fN−1), fj ∈ L2(T, λ), S∗ H f (z) = 1√ N H(ω)f (ω) S∗ Gk f (z) = 1 N Gk(ω)f (ω). (cid:88) (cid:88) ωN =z ωN =z Proof. ad (1): Let z ∈ T, f = (f0, . . . , fN−1), fj ∈ L2(T), then S∗ H SH f (z) = = H(ω)H t(ω)f (ωN ) H(ω)H t(ω)f (z) = f (z) ad (2): Let k ∈ N, z ∈ T, f = (f0, . . . , fN−1), fj ∈ L2(T), then S∗ SGk f (z) = k(ω)f (z) = f (z) ad (3): Let k ∈ N, z ∈ T, f = (f0, . . . , fN−1), fj ∈ L2(T), then GK Gk(ω)Gt S∗ H SGk f (z) = H(ω)Gt k(ω)f (z) = 0, ωN =z ωN =z (cid:88) (cid:88) (cid:88) (cid:88) ωN =z 1 N 1 N ωN =z since(cid:80) MULTIRESOLUTION ANALYSIS FOR MARKOV INTERVAL MAPS 31 ωN =z H(ω)Gt Gk k(ω) = 0 by summing up the roots of unity. For S∗ ad (4): Let i, j ∈ N, i (cid:54)= j, z ∈ T, f = (f0, . . . , fN−1), fj ∈ L2(T), then SH we use that Gk(ω)H t(ω) = 0 by the choice of the coefficients ck,l j . S∗ Gi SGj f (z) = 1 N by summing up the roots of unity. Gi(ω)Gt j(ω)f (z) = 0, ωN =z (cid:3) (cid:88) Here we have seen that in contrast to the filter functions for a usual MRA with one father wavelet and a unitary scaling operator U, we do not obtain a representation of a Cuntz algebra since we do not neccessarily have that SH S∗ H + = I. So we only obtain weaker relations between these filter (cid:80) k∈N\{0} SGk S∗ GK functions. References [Alp93] Bradley K. Alpert, A class of bases in L2 for the sparse representation of integral operators, SIAM J. Math. Anal. 24 (1993), no. 1, 246 -- 262. MR 1199538 (93k:65104) [BFMP10] Lawrence W. Baggett, Veronika Furst, Kathy D. Merrill, and Judith A. Packer, Clas- sification of generalized multiresolution analyses, J. Funct. Anal. 258 (2010), no. 12, 4210 -- 4228. MR 2609543 Jana Bohnstengel and Marc Kesseböhmer, Wavelets for iterated function systems, J. Funct. Anal. 259 (2010), no. 3, 583 -- 601. MR 2644098 [BK10] [Bod07] Mats Bodin, Wavelets and Besov spaces on Mauldin-Williams fractals, Real Anal. [CK03] [Dau92] [DJ06] [GP96] [Jon98] [KS10] Exchange 32 (2006/07), no. 1, 119 -- 143. MR 2329226 (2008h:42063) Mark Crovella and Eric Kolaczyk, Graph wavelets for spatial traffic analysis, Proceed- ings of IEEE Infocom, April 2003. Ingrid Daubechies, Ten lectures on wavelets, CBMS-NSF Regional Conference Series in Applied Mathematics, vol. 61, Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 1992. MR 1162107 (93e:42045) Dorin E. Dutkay and Palle E. T. Jorgensen, Wavelets on fractals, Rev. Mat. Iberoam. 22 (2006), no. 1, 131 -- 180. MR 2268116 (2008h:42071) Jean-Pierre Gazeau and Jiri Patera, Tau-wavelets of Haar, J. Phys. A 29 (1996), no. 15, 4549 -- 4559. MR 1413218 (97f:42054) Alf Jonsson, Wavelets on fractals and Besov spaces, Journal of Fourier Analysis and Applications 4 (1998), 329 -- 340, 10.1007/BF02476031. Marc Kesseböhmer and Tony Samuel, Spectral metric spaces for Gibbs measures, arXiv:1012.5152 (2010). [MP09] [KSS07] Marc Kesseböhmer, Manuel Stadlbauer, and Bernd Stratmann, Lyapunov spectra for KMS states on Cuntz - Krieger algebras, Mathematische Zeitschrift 256 (2007), 871 -- 893, 10.1007/s00209-007-0110-y. Matilde Marcolli and Anna Paolucci, Cuntz - Krieger algebras and wavelets on fractals, Complex Analysis and Operator Theory (2009), 1 -- 41, 10.1007/s11785-009-0044-y. Daniel Mauldin and Mariusz Urbański, Graph directed Markov systems, Cambridge Tracts in Mathematics, vol. 148, Cambridge University Press, Cambridge, 2003, Ge- ometry and dynamics of limit sets. MR 2003772 (2006e:37036) [MU03] [Par60] William Parry, On the β-expansions of real numbers, Acta Math. Acad. Sci. Hungar. [Rén57] 11 (1960), 401 -- 416. MR 0142719 (26 #288) Alfréd Rényi, Representations for real numbers and their ergodic properties, Acta Math. Acad. Sci. Hungar 8 (1957), 477 -- 493. MR 0097374 (20 #3843) Fachbereich 3 - Mathematik und Informatik, Universität Bremen, Bibliothekstrasse 1, 28359 Bremen, Germany E-mail address: [email protected], [email protected]
1502.05214	3	1502	2016-01-28T03:56:13	Weak amenability of Fourier algebras and local synthesis of the anti-diagonal	[ "math.FA", "math.OA" ]	We show that for a connected Lie group $G$, its Fourier algebra $A(G)$ is weakly amenable only if $G$ is abelian. Our main new idea is to show that weak amenability of $A(G)$ implies that the anti-diagonal, $\check{\Delta}_G=\{(g,g^{-1}):g\in G\}$, is a set of local synthesis for $A(G\times G)$. We then show that this cannot happen if $G$ is non-abelian. We conclude for a locally compact group $G$, that $A(G)$ can be weakly amenable only if it contains no closed connected non-abelian Lie subgroups. In particular, for a Lie group $G$, $A(G)$ is weakly amenable if and only if its connected component of the identity $G_e$ is abelian.	math.FA	math	WEAK AMENABILITY OF FOURIER ALGEBRAS AND LOCAL SYNTHESIS OF THE ANTI-DIAGONAL HUN HEE LEE, JEAN LUDWIG, EBRAHIM SAMEI AND NICO SPRONK Abstract. We show that for a connected Lie group G, its Fourier algebra A(G) is weakly amenable only if G is abelian. Our main new idea is to show that weak amenability of A(G) implies that the anti-diagonal, ∆G = {(g, g−1) : g ∈ G}, is a set of local synthesis for A(G × G). We then show that this cannot happen if G is non-abelian. We conclude for a locally compact group G, that A(G) can be weakly amenable only if it contains no closed connected non-abelian Lie subgroups. In particular, for a Lie group G, A(G) is weakly amenable if and only if its connected component of the identity Ge is abelian. 0.1. Background. Questions on the nature of bounded derivations on (commuta- tive) Banach algebras A have been around for a long time, in particular vanishing of bounded Hochschild cohomologies H 1(A,M) for certain Banach A-modules M. See, for example, [SiWe, Kam]. Johnson systematized many of these questions in [Joh1]. In particular, he showed that for a locally compact group G, its group algebra is amenable (i.e. H 1(L1(G),M∗) = {0} for each dual module M∗) if and only if G is an amenable group. He also started the problem of determining when H 1(L1(G), L1(G)∗) = {0}. For a commutative Banach algebra A, Bade, Curtis and Dales ([BCD]) intro- duced the concept of weak amenability, which is defined as having H 1(A,M) = {0} for all symmetric Banach modules. They observed that this is equivalent to having H 1(A,A∗) = {0}. There is an interesting universal module also exhibited by Runde ([Run]). The above observation of [BCD], leads us to refer to any Banach algebra B as weakly amenable if H 1(B,B∗) = {0}. Weak amenability was established for all L1(G) by Johnson ([Joh2]). The Fourier algebras, A(G), as defined by Eymard ([Eym]), are dual objects to the group algebras L1(G) in a sense which generalizes Pontryagin duality. It was long expected that the amenability properties enjoyed by group algebras would also extend to Fourier algebras. Hence it was a surprise when Johnson ([Joh3]) showed that A(G) fails to be weakly amenable for any compact simple Lie group, in particular for G = SO(3). This motivated Ruan ([Rua]) to consider the operator Date: July 24, 2018. 2000 Mathematics Subject Classification. Primary 43A30; Secondary 43A45, 43A80, 22E15, 22D35, 46H25, 46J40. Key words and phrases. Fourier algebra, weak amenability, local synthesis. The first named author was supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF), grant NRF-2015R1A2A2A01006882. The second named author was supported by Institut ´Elie Cartan de Lorraine. The third named author was supported by NSERC Grant 366066-2014, and the Professeur Invit´e program at Institut ´Elie Cartan de Lorraine. The fourth named was supported by NSERC Grant 312515-2010, and the Korean Brain Pool Program. 1 2 HUN HEE LEE, JEAN LUDWIG, EBRAHIM SAMEI AND NICO SPRONK space structure A(G) inherits by virtue of being the predual of a von Neuman algebra. He proved that A(G) is operator amenable if and only if G is amenable. Operator weak amenability for general A(G) was determined by Spronk ([Spr]) and, independently, by Samei ([Sam1]). The question of amenability of A(G) was settled by Forrest and Runde ([FoRu]): it happens exactly when G is virtually abelian. They also showed that if the connected component Ge is abelian, then A(G) is weakly amenable. The following is suggested. Question 0.1. If A(G) is weakly amenable, then must Ge be abelian? Much progress has been made in answering this question. Building on work of Plymen ([Ply]) -- which was written to answer a question in [Joh3] -- Forrest, Samei and Spronk ([FSS1]) showed that A(G) is not weakly amenable whenever G contains a non-abelian connected compact subgroup. Exciting recent progress was made by Choi and Ghandehari. In [ChGh1] they show for the affine motion group, and hence any simply connected semisimple Lie group, and also for the reduced Heisenberg group, that the Fourier algebra is not weakly amenable. In [ChGh2] they used completely different techniques to show the same for Heisenberg groups. Our main theorem generalizes all of these results. Let us briefly review the history of ideas around spectral synthesis. All concepts below will be defined in Section 1.1. The study of sets of spectral synthesis, or, for us, simply "synthesis", for A(G), especially for abelian G, has a long history. See, for example, the historical notes in [HeRoII, §42]. Herz ([Her]) appears to have been the first author to consider local synthesis for general G. There, he proved this property is enjoyed by closed subgroups. Herz's work has inspired, in part, the injection theorem of Lohou´e ([Loh]) and has motived aspects of the work of Ludwig and Turowska ([LuTu1]). Thanks to the existence of a bounded approximate identity consisting of compactly supported functions ([Lep]), sets of local synthesis are sets of synthesis when G is amenable. Weak synthesis has its origins in work of Varopoulos ([Var]) on spheres in Rn (n ≥ 3), was used in Kirsch and Muller ([KiMu]), and was formalized by Warner ([War]). The first explicit mention of what we call smooth synthesis is by Muller ([Mul]), which was applied to certain manifolds in Rn. The first use of this concept for non-abelian G is due to Ludwig and Turowska ([LuTu2]). Following their work, Park and Samei ([PaSa]) showed that for a connected Lie group G, the anti diagonal ∆G is a set of smooth synthesis, and also of weak synthesis for A(G × G). For compact G, building on work of Grønbaek ([Grø]), the article [Joh3] used the failure of a weak form of spectral synthesis of the diagonal ∆G = {(g, g) : g ∈ G} for the projective tensor product algebra A(G) ⊗A(G), to obtain the failure of weak amenability. The local synthesis of ∆G for general locally compact G was used by Samei ([Sam2]) to study a property which implies weak amenability of A(G). Returning to compact groups, the ideas of [Joh3] were formalized and capitalized upon in [FSS2]. These were used in [FSS1] to show that for compact G, A(G) is weakly amenable exactly when the anti-diagonal ∆G = {(g, g−1) : g ∈ G} is a set of synthesis for A(G × G). For groups containing open subgroups products of abelian groups and compact groups, i.e. G ⊇ H ∼= A × K, this last result was extended to local synthesis. We recall that A(G) ⊗A(G) 6= A(G × G), generally ([Los]). Hence we do not expect obvious connections between sets of local synthesis for these two algebras. WEAK AMENABILITY OF A(G) 3 0.2. Structure. The starting point for the present investigation lies in the afore- mentioned results of [PaSa]. Let G be a connected Lie group. We use the fact that ∆G is simultaneously of weak and smooth synthesis for A(G × G), along with the characterization of weak amenability of [Run], to show that for a connected Lie group, weak amenability of A(G) implies that ∆G is a set of local synthesis for A(G × G). The techniques rely intrinsically on the Lie theory, especially having finite dimension for G. We see no way, at present, to extend them to arbitrary connected locally compact groups. In Section 1, we show for any connected Lie G that that weak amenability of A(G) implies local synthesis of ∆G for A(G× G) -- a property we shall hereafter call "local synthesis for G× G". Also in that section we discuss our two main functorial properties satisfied by local synthesis of ∆G for G × G for locally comapct G: the restriction to a closed connected Lie subgroup and an injection theorem with quotients by discrete normal subgroups. In section 2 we give a criterion for testing local synthesis of the anti-diagonal, and we show for five (classes of) low-dimensional Lie groups that this criterion is satisfied. In Section 3 we tie the investigation together by noting that any non-abelian connected Lie group contains one of the five aforementioned groups or its simply connected covering group, and use the functorial properties to draw our conclusion for connected G. We can thus answer Question 0.1 affirmatively for all Lie groups. 0.3. Basic notation. Let G be a locally compact group. The following spaces will be used in this note: the space Cc(G) of compactly supported continuous functions; and the Lp-spaces with respect the left Haar measure, Lp(G), p = 1, 2,∞. We follow Eymard ([Eym]) for all definitions and concepts around the Fourier al- gebra A(G). We recall that A(G) consists of all matrix coefficients u(g) = hλ(g)ξηi where λ : G → U(L2(G)) is the left regular representation, and ξ, η ∈ L2(G). Fur- thermore, the norm is given by kukA = inf{kξk2 kηk2 : u = hλ(·)ξηi as above}. The bounded linear dual is given by the group von Neuman algebra V N (G) = λ(G)′′ ⊂ B(L2(G)). If u = hλ(·)ξηi ∈ A(G) and T ∈ V N (G) we write T (u) = hT ξηi. We shall denote the operator norm of an element S of V N (G) by kSkV N . We recall that the maps u 7→ u (u(g) = u(g−1)) and maps of left and right transla- tion are all isometric automorphisms of A(G). We always let A(G) ⊗A(G) denote the projective tensor product. We shall have no need for completely bounded maps, and will make no use of the operator projective tensor product, except, of course, implicitly, when we discuss A(G × G). We shall define more specialized notions in situ as their necessities arise. 1. Weak amenability and local synthesis for the anti-diagonal 1.1. Weak amenability implies local synthesis for the anti-diagonal. Let G be a locally compact group. We let Ac(G) denote the subalgebra of elements u of A(G) for which supp u = {g ∈ G : u(g) 6= 0} is compact. It is well known that the Tauberian condition holds: Ac(G) is dense in A(G); and that A(G) is regular on G: given compact K and a neighbourhood U of K, there is u in Ac(G) for which uK = 1 and supp u ⊂ U . See [Eym, (3.38) and (3.2)]. 4 HUN HEE LEE, JEAN LUDWIG, EBRAHIM SAMEI AND NICO SPRONK Given a closed subset E of G we let IG(E) = {u ∈ A(G) : uE = 0} I 0 G(E) = {u ∈ Ac(G) : supp u ∩ E = ∅}, and JG(E) = IG(E) ∩ Ac(G). G(E) ⊆ JG(E) ⊆ IG(E). We say that E is a set of It is evident that I 0 • spectral synthesis for G if I 0 • local synthesis for G if I 0 G(E) = IG(E); and G(E) = JG(E). For a linear subspace S ⊂ A(G) we let S(d) = span{ud : u ∈ S}. We say that E is a set of • weak synthesis for G if there is d in N for which IG(E)(d) ⊆ I 0 • local weak synthesis for G if there is d in N for which JG(E)(d) ⊆ I 0 We now let G be a connected Lie group and let D(G) denote the space of com- If K is a compact subset of G we let pactly supported smooth functions on G. DK(G) = {u ∈ D(G) : supp u ⊂ K}, which is a Fr´echet space. For example, we fix a basis β = (X1, . . . , Xd) for the Lie algebra g of G and for u, v in DK(G) set G(E); and G(E). (1.1) ∂X u(g) = , for X in g ρK β,n(u) = X1≤i1,...,in≤d(cid:13)(cid:13)∂Xi1 . . . ∂Xin u(cid:13)(cid:13)∞ and ρK β (u, v) = d dt u(g exp(tX))(cid:12)(cid:12)(cid:12)(cid:12)t=0 ∞Xn=0 2n + ρK ρK β,n(u − v) . β,n(u − v) Then given a compact neighbourhood K of the identity, D(G) = S∞ n=1 DK n (G) is an inductive limit of Fr´echet spaces, a so called LF-space. See, for example, [Tre]. It follows from [Eym, (3.26)] or [LuTu2, Lemma 3.3 and Remark 4.2], that D(G) ⊂ Ac(G), and each inclusion DK(G) ֒→ A(G) is continuous. Furthermore, since D(G) is dense in L2(G), D(G) ⊇ D(G) ∗ D(G) and {u : u ∈ D(G)} = D(G), D(G) is dense in A(G). For a closed subset E of G we let J D G (E) = D(G) ∩ IG(E) where closure is in A(G). Since D(G) is dense in A(G), and D(G)J D we have that J D have inclusions I 0 G (E), G (E) is an ideal in A(G). Since A(G) is regular and Tauberian we G(E) ⊆ J D G (E) ⊆ J D G (E) ⊆ JG(E) ⊆ IG(E). We say that E is a set of • smooth synthesis for G if J D • local smooth synthesis for G if J D G (E) = IG(E); and G (E) = JG(E). The projective tensor product E ⊗F of two locally convex spaces E and F is the completion of the algebraic tensor product E ⊗F in the final topology with respect to the embedding E ×F ֒→ E ⊗F . The standard proof of the following, for G = Rd, uses techniques of Fourier analysis on Schwarz class functions. Hence we need to take a little care to see that it holds in our setting, though the proof is standard. Lemma 1.1. Let G be a connected Lie group and K and M be compact subsets of G. Then DK(G) ⊗DM (G) ∼= DK×M (G × G) linearly and homeomorphically. WEAK AMENABILITY OF A(G) 5 Proof. Let us first assume that there is a neighbourhood U of K for which there is a diffeomorphism ϕ : U → U ′ ⊂ Rd. Then the map u 7→ u ◦ ϕ : DK (G) → Dϕ(K)(Rd) is a linear homeomorphism. The same fact holds for M , with a neighbourhood V of M and a diffeomorphism ψ : V → V ′ ⊂ Rn. Thus we get linear homeomorphisms DK(G) ⊗DM (G) ∼= Dϕ(K)(Rd) ⊗Dψ(M)(Rd) ∼= Dϕ(K)×ψ(M)(R2d) ∼= DK×M (G × G) where the middle identification is provided by [Tre, Theorem 51.6], and the last one by the map w 7→ w ◦ (ϕ−1 × ψ−1). i=1 of K and {Vj}n j=1 of M , such that i=1,{vj}n each member is diffeomorphic to an open subset of Rn. Let {ui}m j=1 in D(G) be smooth partitions of unity, subordinate to the respective covers. Then we obtain, (G), Generally, there are finite open covers {Ui}m whose inverse is mere inclusion. Then we obtain linear homeomorphisms for example, a linear homeomorphism u 7→Pn (G) ∼= mMi=1 DM ∩Vj DK∩Ui as above. Hence we obtain linear homeomorphisms i=1 uui : DK(G) →Pn mMi=1 DK∩Ui×M ∩Vj nMj=1 i=1 DK∩Ui (G × G) (G)! ⊗ nMj=1 nXj=1 mXi=1 nXj=1 mXi=1 ∼= DK(G) ⊗DM (G) ∼= DK∩Ui (G) ⊗DM ∩Vj (G) DK∩Ui×M ∩Vj (G × G) ∼= DK×M (G × G) by using both injectivity and projectivity of tensor product of these nuclear spaces. (cid:3) We let ∆G = {(g, g−1) : g ∈ G} ⊂ G × G. Lemma 1.2. Let G be a connected Lie group. Then (Ac(G) ⊗ Ac(G)) ∩ JG×G( ∆G) is dense in JG×G( ∆G). Proof. For simplicity, we let A = A(G × G). Let u ∈ JG×G( ∆G). It was shown in [PaSa, Theorem 7] that ∆G is a set of local smooth synthesis, so we may assume that u ∈ D(G × G). Hence we can find a compact subset M of G for which u ∈ DM ×M (G × G). We can further assume that M is symmetric: M −1 = M . Lemma 1.1 provides that u ∈ DM (G) ⊗DM (G). Fix ε > 0. Let ρ be any invariant metric on DM ×M (G × G) which comes form the Fr´echet structure. There is a δ1 > 0 for which for w ∈ DM ×M (G × G) ⊂ A(G × G). Furthermore, it is straightforward to check that the map ρ(w, 0) < δ1 ⇒ kwkA < ε Λ : DM ×M (G × G) → DM (G) given by Λw(g) = w(g, g−1) is continuous since ∆G, qua submanifold of G × G, is diffeomorphic to G. If we fix vM in Ac(G) for which vMM = 1, there is δ2 > 0 for which ε ρ(w, 0) < δ2 ⇒ kΛ(w)kA(G) < . kvMkA(G) 6 HUN HEE LEE, JEAN LUDWIG, EBRAHIM SAMEI AND NICO SPRONK Thus, if δ = min{δ1, δ2}, then given u as in the paragraph above, there is a v in the algebraic tensor product DM (G)⊗DM (G) for which ρ(u, v) < δ, and hence our choice of δ entails that ku − vkA < ε and kΛ(u − v)kA(G) < ε kvMkA(G) . Now let w = v− Λ(v)⊗ vM which is an element of Ac(G)⊗ Ac(G). Since M = M −1 G×G( ∆G), it is easy to see that w ∆G = 0. Moreover, notice that Λ(u) = 0, as u ∈ J D so ku − wkA ≤ ku − vkA + kΛ(u − v) ⊗ vMkA < 2ε since kΛ(u − v) ⊗ vMkA ≤ kΛ(u − v)kA(G) kvMkA(G). (cid:3) We let A(G)♯ denote the unitization of A(G) and m♯ : A(G)♯ ⊗A(G)♯ → A(G)♯ and m : A(G) ⊗A(G) → A(G) be the continuous linearization of the respective multiplication maps. Since A(G) is Tauberian and regular, A(G)2 = A(G). Hence It follows from [Grø, Theorem 3.2] that A(G) is weakly amenable if and only if (1.2) (ker m)2 = A(G) ⊗A(G) · ker m♯. Above, and in what follows, we use for a subspace S of an algebra A, the notation Sd = span{u1 . . . ud : u1, . . . , ud ∈ S}. Theorem 1.3. Let G be a connected Lie group. If A(G) is weakly amenable, then ∆G is a set of local synthesis for A(G × G). Proof. We let m : A(G) ⊗A(G) → A(G) be given on elementary tensors by m(u ⊗ v) = uv, and likewise define m♯. Since v 7→ v is an isometry on A(G), if A(G) is weakly amenable then (1.2) implies that (ker m)2 = A(G) ⊗A(G) · ker m♯. (1.3) For simplicity, we write J = JG×G( ∆G), below. Let u ∈ J. We wish to approximate G×G( ∆G). We let ι : A(G) ⊗A(G) → A(G × G) be the linear u by elements from I 0 contraction which embeds A(G) ⊗ A(G) into A(G × G). We have that [Ac(G) ⊗ Ac(G)] ∩ J ⊂ ι(Ac(G) ⊗ Ac(G) · ker m♯) ⊆ ι(ker m)2 ∩ Ac(G × G) ⊆ J 2 where the first inclusion follows from regularity of A(G), the second inclusion is provided by (1.3), and the third inclusion follows from regularity of A(G × G). By Lemma 1.2, [Ac(G)⊗ Ac(G)]∩ J is dense in J. We thus conclude that J = J 2, and induction shows that J = J m for any m in N. The identity 4uv = (u + v)2 − (u − v)2 shows that J 2 = J (2), where the latter notation was used in the definition of weak synthesis, above. By induction we see that J 2m = J (2m) for any m in N. We conclude that J = J (2m) for any m in N. It is shown in [PaSa, Theorem 7], that ∆G is a set of local weak synthesis, G×G( ∆G). Hence we see that for some m, i.e. if n ≥ dim(G)/2, then J (n) ⊂ I 0 J = J (2m) = I 0 G×G( ∆G), and local synthesis is established. (cid:3) WEAK AMENABILITY OF A(G) 7 1.2. Functorial properties for local synthesis of the anti-diagonal. Our ultimate goal is to show that a non-abelian connected Lie group does not allow local synthesis of the anti-diagonal. The following will allow us to reduce our calculations to certain computable cases. We shall make use of the well-known result that for any locally compact group G, and closed subgroup H, the restriction map RH : A(G) → A(H) is a quotient map. See [Her], [McM, (4.21)], [Ars, (3.23)] or [DeDe]. We even have RH (Ac(G)) = Ac(H). Theorem 1.4. Let G be a locally compact group and H is a closed connected Lie subgroup. If ∆G is of local synthesis for G × G, then ∆H is of local synthesis for H × H. Proof. We first claim that RH×H (JG×G( ∆G)) = JH×H ( ∆H ). i, v′ i in Ac(G) for which RH u′ It is clear that RH×H (JG×G( ∆G)) ⊆ JH×H ( ∆H ). To see the converse inclusion, let u ∈ JH×H ( ∆H ). Since H is a connected Lie group, Lemma 1.2 allows us to assume that u ∈ Ac(H) ⊗ Ac(H), hence u =Pn i=1 ui ⊗ vi. The restriction theorem assures that there are elements u′ i = vi for each i. Hence if w in Ac(G) satisfies that wS = 1 where S =Sn and, sincePn i) ∈ IG×G( ∆G) ∩ Ac(G × G) iv′ i − w ⊗ u′ i=1 ui(h)vi(h) = u(h−1, h) = 0 for h in H, we have i=1 supp(ui), then i = ui, RH v′ nXi=1 (u′ i ⊗ v′ u′ = RH×H (u′) = (ui ⊗ vi − RH (w) ⊗ uivi) = u. nXi=1 It then follows that u ∈ RH×H (JG×G( ∆G)). It is evident that RH×H (I 0 G×G( ∆G)) ⊆ I 0 H×H ( ∆H ), Our assumption that JG×G( ∆G) = I 0 paragraph, shows that I 0 H×H ( ∆H ) = JH×H ( ∆H ). so RH×H(cid:16)I 0 G×G( ∆G)(cid:17) ⊆ I 0 H×H ( ∆H ). G×G( ∆G), coupled with the result of the prior (cid:3) The proof of the next functorial property is more general, though its proof is a little more involved. We shall use the following fact about local synthesis. This is related to well-known facts about spectral synthesis in regular function algebras; see, for example, expositions in [ReSt, Kan]. If G is a locally compact group and E is a closed subset of G, we have for u in Ac(G) that G(E) ⇔ u is "locally in I 0 bourhood Ug of g, and a ug in I 0 G(E)", i.e. for every g in E there is a neigh- u ∈ I 0 G(E), for which uUg = ugUg . Indeed, we need to only prove sufficiency. In this case, take such a collection U1 = Ug1 , . . . , Un = Ugn of open sets covering supp(u)∩E, and associated functions k=1 Uk with Un+1∩ E = ∅. Any partition of unity v1, . . . , vn+1 of supp(u) subordinate to U1, . . . , Un+1, satisfies u1, . . . , un. Let Un+1 be a neighbourhood of supp(u)\Sn uvk = ukvk, for each k = 1, . . . , n, so u =Pn k=1 ukvk + uvn+1 ∈ I 0 G(E). 8 HUN HEE LEE, JEAN LUDWIG, EBRAHIM SAMEI AND NICO SPRONK The following holds for any regular Banach function algebra. To avoid introduc- ing new notation, we state it only for a Fourier algebra. Proposition 1.5. Let G be a locally compact group and E be a closed subset which admits a partition, E = Fi∈I Ei, for which each Ei is closed in G and relatively open in E. Then E is of local synthesis for G if and only if each Ei is of local synthesis. Proof. (⇒) Let F be a subset of E which is closed in G and relatively open in E; in particular we may consider F = Ei for a fixed i. Let u ∈ JG(F )∩ Ac(G). For every g in F , our assumptions on F allows us to choose a compact neighbourhood U of g for which U ∩ E = U ∩ F . We find v in Ac(G) for which supp(v)∩ E = supp(v)∩ F and vU = 1. We have that u = uv on U and uv ∈ JG(E) = I 0 G(E) ⊆ I 0 G(F ). Thus u is locally in I 0 G(F ), whence in JG(F ). (⇐) Let u ∈ JG(E) ∩ Ac(G). Since each Ei is of local synthesis, we have u ∈ JG(E) ⊆ JG(Ei) = I 0 G(Ei) G(E), thus u ∈ I 0 G(E). Hence u is locally in I 0 i.e. u = limn→∞ un, where each un ∈ I 0 G(Ei). Fix g in Ei and, again, our as- sumption of relative openness of Ei in E allows us to take a compact neighbour- hood U of g for which U ∩ E = U ∩ Ei. If v is any element of Ac(G) for which supp(v)∩ E = supp(v)∩ Ei and vU = 1, then u = uv on U , and uv = limn→∞ unv, where each unv ∈ I 0 Theorem 1.6. Let G be a locally compact group and Γ a discrete normal subgroup. Then ∆G is of local synthesis for G × G if and only if ∆G/Γ is of local synthesis for G/Γ × G/Γ. Proof. The injection theorem of [Loh] tells us that ∆G/Γ is of local synthesis for G/Γ× G/Γ if and only if q−1( ∆G/Γ), where q : G× G → G/Γ× G/Γ is the quotient map, is a set of local synthesis for G × G. Hence we wish to verify that ∆G is of local synthesis ⇔ q−1( ∆G/Γ) is of local synthesis (1.4) each for G × G. Hence we wish to examine the structure of the latter set. We first observe that since gΓ = Γg and (Γg)−1 = g−1Γ in G/Γ, we have G(E). (cid:3) q−1( ∆G/Γ) = {(γg, g−1γ′−1) : g ∈ G, γ, γ′ ∈ Γ}. Hence if we define an action of Γ × Γ on G × G by (γ, γ′) · (g, g′) = (γg, g′γ′−1), then q−1( ∆G/Γ) = (Γ × Γ) · ∆G, the orbit of the set ∆G under this action. It is easy to check that i.e. if (γg, g−1γ′−1) = (h, h−1) then γg = (g−1γ′−1)−1; and hence (γ, γ′) · ∆G = ∆G ⇔ γ′ = γ Let us see that (γ, γ′) · ∆G = (λ, λ′) · ∆G ⇔ λ−1γ = λ′−1γ′ ⇔ λ′λ−1 = γ′γ−1. Thus q−1( ∆G/Γ) = F(γ,γ ′)∆Γ∈(Γ×Γ)/∆Γ(γ, γ′) · ∆G, where ∆Γ = {(γ, γ) : γ ∈ Γ}. the individual fibres (λ, λ′) · ∆G are relatively open in q−1( ∆G/Γ). (1.5) Suppose that (λh, h−1λ′−1) is approached by a net of elements (γngn, g−1 n γ′ n Then the net of elements (λ−1γngnh−1, hg−1 −1). −1λ′) would approach (e, e), which n γ′ n WEAK AMENABILITY OF A(G) 9 implies that λ−1γnγ′ n thus γnγ′ n −1λ′ approaches e, and hence is ultimately e, by discreteness; −1 is ultimately λλ′−1, i.e. the net is ultimately in the fibre (λ, λ′) · ∆G. (⇒) If ∆G is of local synthesis for G × G, then so too is each fibre (γ, γ′)· ∆G = (γ, e) ∆G(e, γ′−1). We then appeal to Proposition 1.5 and (1.5) to obtain necessity in (1.4). (⇐) The anti-diagonal ∆G is an open fibre of q−1( ∆G/Γ), thanks to (1.5). Then Proposition 1.5 delivers sufficiency in (1.4). (cid:3) In practice we will use this last result with a central discrete subgroup. Centrality seems to offer no meaningful simplification to the proof, however. 2. Failure of local synthesis for the anti-diagonal 2.1. The basic strategy. We wish to show that for certain given 2 or 3-dimensional non-abelian connected Lie groups G, that its anti-diagonal fails to be a set of local synthesis for G×G . We shall exploit the density of the space of test functions D(G) in A(G), as outlined in Section 1.1. This implies that any T in V N (G) ∼= A(G)∗ may be understood as a distribution, i.e. T ∈ D(G)∗. The following informs all of the choices made through the rest of this section. Lemma 2.1. Let G be a connected Lie group with Lie algebra g and X ∈ g with X 6= 0, hence (X, 0) is an element of the Lie algebra g × g of G × G. Let 0 6= v ∈ L1(G). If there exists S = SX,v in V N (G × G) for which ∂(X,0)u(g, g−1)v(g) dg S(u) =ZG Let us see that S ∈ I 0 G×G( ∆G)⊥. Let u ∈ I 0 G×G( ∆G) ( JG×G( ∆G). whenever u ∈ D(G × G), then ∆G is not a set of local synthesis for G × G. Proof. We will verify that S ∈ I 0 annihilator of the subspace K. This will show that I 0 G×G( ∆G)⊥ \ JG×G( ∆G)⊥, where K ⊥ denotes the G×G( ∆G), and ε > 0. By the regularity of the algebra D(G × G), there is w in D(G × G) such that wsupp(u) = 1 and supp(w) ∩ ∆G = ∅. Since I 0 G×G( ∆G), there is u′ in D(G × G) ∩ IG×G( ∆G) such that ku − u′kA < ε/ kwkA. But then, since u = wu, we have that G×G( ∆G), it is obvious that ∂(X,0)(wu′) ∆G = 0, so ku − wu′kA < ε. As wu′ ∈ I 0 S(wu′) = 0. Hence S(u) = S(u − wu′) ≤ kSkV N ε. As ε > 0 may be chosen arbitrarily, S(u) = 0. Let us now see that S 6∈ JG×G( ∆G)⊥. We consider general x, y, z in D(G), and let w = xy ⊗ z − x ⊗ yz, which is an element of D(G × G) ∩ JG×G( ∆G). But then S(w) =ZG x(g)∂X y(g)z(g−1)v(g) dg. [∂X (xy)(g)−∂X x(g)y(g)]z(g−1)v(g) dg = −ZG G×G( ∆G) ⊆ J D We may choose x, y, z for which S(w) 6= 0. (cid:3) We shall require disintegration of the left regular representation of G into irre- ducible components. For this purpose and to introduce notation, we summarize the Plancherel theorem of [Tat]. Our presentation is influenced by [Fol, (7.50)]. We purposely restrict the description, to fit our needs. We let bG denote the space of (equivalence classes of) irreducible representations of G, and accept the standard 10 HUN HEE LEE, JEAN LUDWIG, EBRAHIM SAMEI AND NICO SPRONK abuse of notation where we conflate an equivalence class with one of its represen- tatives. If u ∈ Cc(G) we let its Fourier transform be given at π in bG by u(g)π(g) dg ∈ B(Hπ), i.e. hu(π)ξηi =ZG u(π) =ZG u(g)hπ(g)ξηi dg where ξ, η ∈ Hπ, and Hπ is the space on which π acts. Proposition 2.2. Suppose G is a connected Lie group for which the kernel of the modular function, K = ker ∆, is type I, and G acts on bK regularly in the sense that there is a Borel cross-section for the space of orbits of G on bK. Then bG contains • a dense Borel subset S(bG) of elements, each of which is the induced repre- • a Borel parametrization b 7→ πb : B → S(bG) and a Borel measure µ on B; • for each b in B, a positive operator δb on Hπb, which satisfies sentation from some closed subgroup of ker ∆; and πb(g)δbπb(g−1) = 1 δb ∆(g) for which the choice of the triple ((δb)b∈B, (πb)b∈B, µ) is unique, up to measure equivalence of ((πb)b∈B, µ), and such that the Plancherel transform on Cc(G), given by extends to a unitary identifying b )b∈B u 7→ (bu(πb)δ1/2 B Hπb ⊗2 Hπb dµ(b). L2(G) ∼=Z ⊕ Each Hilbertian tensor product Hπb ⊗2 Hπb is identified with the space of Hilbert- Schmidt operators on Hπb , above. Hence the left regular representation λ of G admits disintegration up to unitary equivalence, and quasi-equivalence respectively, as (2.1) λ ∼=Z ⊕ B πb ⊗ IHπb dµ(b) and λ ≃Z ⊕ B πb dµ(b). In (2.1) we need to concern ourselves with only the equivalence class of µ in the relation of mutual absolute continuity of measures on B. Furthermore, the left regular representation of G × G may now be represented by the quasi-equivalence (2.2) πb × πb′ d(µ × µ)(b, b′) λ × λ ≃Z ⊕ B×B where we use Kroenecker products of representations. Hence if v ∈ L1(G), then the the operator field operator on A(G× G) given by Ev(u) =RG u(g, g−1)v(g) dg may be represented by (2.3) Ev(b, b′) =ZG v(g)πb(g) ⊗ πb′ (g−1) dg for b, b′ in B. If G is unimodular, we set each δb = IHπb . When G is not unimodular, and πb is induced from a character χ of abelian subgroup H of ker ∆, then Hπb may be identified with a completion of Fπb = {f ∈ C(G) : f (gh) = χ(h)f (g)}, and δb with multiplication on the latter space by ∆(·)−1; compare with the description in [Fol, (7.49)]. WEAK AMENABILITY OF A(G) 11 We also have for any g in G, the Fourier inversion formula of [Tat, Corollary 2]: (2.5) (2.6) Tr(πb(g)bw(πb)δb) dµ(b). By density of span λ(G) in V N (G), and of each span πb(G) in B(Hπb) (Schur's lemma), we also have for any T in V N (G) the duality formula for w = hλ(·)uvi, where u, v ∈ Cc(G), we have (2.4) w(g) =ZB Tr(πb(g−1) w(πb)δb) dµ(b) =ZB T (w) =ZB Tr(T (b)bw(πb)δb) dµ(b) where T ≃ (T (b))b∈B ∈ L∞(B, µ;B(Hπb)). Let G be a connected Lie group and π ∈ bG. It is well known that for any X in the Lie algebra g of G, and any π in bG, there is a dense subspace HX π of vectors ξ for which dπ(X)ξ = lim h→0 [π(exp(hX)) − I]ξ 1 h exists. For example HX π = span{u(π)ξ : u ∈ D(G), ξ ∈ Hπ}. Thus dπ(X) is an (unbounded) operator on Hπ. Given the parameterization b 7→ πb above, it can be checked that, as the limit of a measurable field of operators, (dπb(X))b∈B is also measurable. π ⊇ HD Lemma 2.3. Suppose that G is a connected Lie group with Lie algebra g, and X in g is such that, for each b in B, there is a subspace Fb of HX πb which is dense in Hπb , and for which δ−1 πb . If T in V N (G) satisfies that (S(b))b∈B = (T (b)dπb(X))b∈B is a bounded field of operators, then for u in D(G) we have that b Fb ⊆ HX S(u) = T (∂X u). ∂X u(g−1)πb(g)δbξ dg b Fb then we have for u in D(G) that Proof. Recall that as in (1.1), the symbol ∂X denotes a derivative on the right. First we fix π. If ξ ∈ δ−1 \(∂X u)∨(πb)δbξ =ZG h(cid:20)ZG u(g−1 exp(hX))πb(g)δbξ dg −ZG h(cid:20)ZG u(g−1)πb(exp(hX)g)δbξ dg −ZG [πb(exp(hX) − I](cid:19)ZG u(g−1)πb(g)δbξ dg(cid:21) u(g−1)πb(g)δbξ dg(cid:21) u(g−1)πb(g)δb dgξ = lim h→∞ = lim h→∞ 1 1 1 h h→∞ =(cid:18) lim = dπb(X)bu(πb)δbξ where in the second through fourth lines, the limit is understood in the weak sense, and we may use dominated convergence theorem on associated scalar integrals. Since δ−1 b Fb is dense in Hπb , the computation above, coupled with (2.5), tells us that S(u) =ZB =ZB Tr(T (b)dπb(X)bu(πb)δb) dµ(b) Tr(T (b) \(∂X u)∨(πb)δb) dµ(b) = T (∂Xu). (cid:3) 12 HUN HEE LEE, JEAN LUDWIG, EBRAHIM SAMEI AND NICO SPRONK We now will embark on using Lemma 2.3 on the group G× G and operator fields with fibres d(πb × πb′ )(X, 0) = dπb(X) ⊗ IHb′ (X ∈ g) to verify the conditions of Lemma 2.1. As a first illustration, let us apply these methods to the special unitary group SU(2). We note that ∆SU(2) is a set of non-synthesis is known (see [FSS1]). We recall the well known fact that SU(2) admits as its Lie algebra su(2), which may be identified with 2× 2-complex matrices which are skew-Hermitian and trace zero. each πn acts on a space of dimension n + 1, and if g ∼= diag(z, ¯z) (∼= is similarity by conjugation in SU(2)) then πn(g) ∼= diag(zn, zn−2, . . . , z−n). Proposition 2.4. Let X be any non-zero element of su(2). Then there is an S = SX,1 in V N (SU(2)) for which The representation theory of SU(2) is well-known: cSU(2) = {πn : n = 0, 1, 2, . . .}, S(u) =ZG ∂(X,0)u(g, g−1) dg for u in D(SU(2) × SU(2)). Proof. Since X ∗ = −X and TrX = 0, there is x in R for which we have equivalence X ∼= diag(ix,−ix), hence exp(hX) ∼= diag(eihx, e−ihx) in SU(2). It is immediate that dπn(X) ∼= diag(inx, i(n − 2)x, . . . ,−inx). The Schur orthogonality relations immediately give, in the notation of (2.3) with v = 1, that E1(n, n′) =ZSU(2) πn(g) ⊗ πn′ (g)∗ dg =( 1 n+1 Tn 0 if n = n′ otherwise where Tn is an (n + 1)2 × (n + 1)2 permutation matrix with respect to some basis. It is then obvious that for each n, n′. We appeal to Lemma 2.3. kE1(n, n′)(dπn(X) ⊗ I)k ≤ x (cid:3) 2.2. Two unimodular groups. We let E = C⋊T be the 3-dimensional Euclidean motion group with multiplication and inversion given by (x, z)(x′, z′) = (x + zx′, zz′) and (x, z)−1 = (−¯zx, ¯z). This group is unimodular with Haar integral the same as that on C × T. The following data are well known; see, for example [Sug, IV]. The additive group C = R2 admits a real inner product (x, x′) 7→ x· x′ = Re x Re x′ + Im x Im x′. For each a in C\{0} we obtain an irreducible unitary representation πa by inducing from the character χa (χa(x) = e−ix·a). Then πa ∼= πb if and only if a = b. Hence we parameterize this family by r in (0,∞). Each representation is given πr(x, z)ξ(w) = e−ix·(rw)ξ(¯zw). Then the disintegration formula (2.1) takes the form πr : E → U(L2(T)), λ ≃Z ⊕ (0,∞) πr r dr ≃Z ⊕ (0,∞) πr dr where the middle formula is with respect to the group's Plancherel measure which is mutually absolutely equivalent to Lebesgue measure on (0,∞). WEAK AMENABILITY OF A(G) 13 We recall that E has Lie algebra e = hT, X1, X2 : [T, X1] = X2, [T, X2] = −X1, [X1, X2] = 0i where exp(hX1) = (h, 1), exp(hX2) = (ih, 1) and exp(hT ) = (0, eih). We also consider the Sobolev-type space H 2,1(R2) = {v : R2 → C v, ∂2 xv ∈ L1(R2), for all x ∈ R2} dt v(y + tx)(cid:12)(cid:12)t=0, and the second order derivatives may be consid- where ∂xv(y) = d ered in the distributional sense. Theorem 2.5. Let v ∈ H 2,1(C) = H 2,1(R2), and v in L1(E) be given by v(x, z) = v(x). Let X ∈ span{X1, X2}. Then there is an S = SX,v in V N (E) for which S(u) =ZE ∂(X,0)u(g, g−1)v(g) dg for u in D(E × E). Proof. First, for ξ in L2(T), (2.3) provides that Ev(r, r′)ξ(w, w′) =ZTZC =ZT v(x)e−ix·(rw−r′zw′)ξ(¯zw, zw′) dx dz 2πbv(rw − r′zw′)ξ(¯zw, zw′) dz U ∗Ev(r, r′)U ξ(w, w′) = Ev(r, r′)U ξ(w, ¯ww′) L2(T2) given by U ξ(w, w′) = ξ(w, ww′), which has adjoint given by U ∗ξ(w, w′) = ξ(w, ¯ww′). We have where bv is the Fourier transform of v on C = R2. Consider the unitary U on = 2πZTbv(rw − r′z ¯ww′)ξ(¯zw, w′) dz = 2πZTbv(rw − r′zw′)ξ(¯z, w′) dz The fact that we choose v from H 2,1(C) allows that there is a constant Cv, such that for any y in C we have Cv 1 + y2 . bv(y) ≤ In fact we could choose Cv = kbvk∞ +kcLvk∞, where L = ∂2 x1 + ∂2 x2 is the Laplacian on R2 ∼= C. Hence we estimate kU ∗Ev(r, r′)U ξ(w, w′)k2 ≤ (2πCv)2ZTZT(cid:20)ZT ≤ (2πCv)2(cid:20)ZT dz 1 + r − r′z2(cid:21)2 kξk2 2 2 ≤ZT2(cid:20)2πCvZT 1 + r − r′zw′ ¯w2(cid:21)(cid:20)ZT dz ξ(¯z, w′) 1 + rw − r′zw′2 dz(cid:21)2 d(w, w′) ξ(¯z, w′)2 dz 1 + r − r′w′z ¯w2(cid:21) dw dw′ where we have used the Cauchy-Schwarz inequality in the second line, and Tonelli's theorem, in the third. Hence (2.7) kEv(r, r′)k = kU ∗Ev(r, r′)Uk ≤ZT 2πCv dz 1 + r − r′z2 for r, r′ > 0. 14 HUN HEE LEE, JEAN LUDWIG, EBRAHIM SAMEI AND NICO SPRONK Now we consider for each r > 0 the operator dπr(X). (x, 1) = exp X we have If ξ ∈ L2(T) then for dπr(X)ξ(w) = lim h→0 1 h [e−i(hx)·(rw) − 1]ξ(w) = −irx · wξ(w) where convergence is uniform in w. Hence by (2.6) dπr(X) is the multiplication operator by w 7→ −irx· w. In particular dπr(X) is bounded with kdπr(X)k ≤ rx. Combining with (2.7) we see that kEv(r, r′)(dπr(X) ⊗ I)k ≤ZT 2πCvr dz 1 + r − r′z2 =Z 2π 0 Cvr dt 1 + r2 + r′2 − 2rr′ cos t . However, using either methods of complex analysis, or the table of integrals [GrRy, 2.553-3], we obtain that the latter integral is equal to the first expression in the elementary estimate Cvr Cvr 2q(1 + r2 + r′2)2 − 4(rr′)2 ≤ 2q1 + 2(r2 + r′2) which is clearly uniformly bounded in r and r′. Hence, by Lemma 2.3, we are done. (cid:3) , x + x′) Now we consider the reduced Heisenberg group Hr = (R × T) ⋊ R with multipli- cation and inversion given by (y, z, x)(y′, z′, x′) = (y + y′, zz′eixy ′ and (y, z, x)−1 = (−y, ¯zeixy,−x). We identify the centre of Hr with T. The group is unimodular, and its Haar integral is the same as that on the product group R × T × R. All of the infinite- dimensional irreducible representations are known to be obtained by inducing from the characters χ0,n on the normal subgroup R× T, χ0,n(y, z) = ¯zn, for n ∈ Z\ {0}. This follows, for example, from [KaTa, 4.38] and the fact that Hr is a quotient of the usual Heisenberg group. For each n in Z \ {0}, the representation is given by πn : Hr → L2(R), The left regular representation admits a decomposition πn(y, z, x)ξ(t) = ¯zne−intyξ(t − x). λ ≃Mn∈Z πn (2.8) where π0 : Hr → L2(R2) is, effectively, the left regular representation of Hr/T. Indeed, if we let Hn = {ξ ∈ L2(Hr) : ξ(g(0, z, 0)) = ¯znξ(g) for a.e. g in Hr}, then λn = λ(·)Hn has λn(0, z, 0) = ¯znI, and is hence quasi-equivalent to πn by the Stone-von Neumann theorem (see, for example, [Fol, (6.49)]). We note that the Lie algebra of Hr is given by h = hX, Y, Z : [X, Y ] = Z, [Z, X] = 0 = [Y, Z]i where exp(hX) = (0, 1, h), exp(hY ) = (h, 1, 0) and exp(hZ) = (0, eih, 0). Theorem 2.6. Let v1, v2 in L1(R) be so that v1 is (essentially) bounded and v2 ∈ L1(R), and set v(y, z, x) = v1(x)v2(y). Then there is S = SZ,v in V N (Hr) for which S(u) =ZHr ∂Z u(g, g−1)v(g) dg for u ∈ D(Hr × Hr). WEAK AMENABILITY OF A(G) 15 Proof. We allow (2.8) to substitute for (2.1) and we obtain a likewise decomposition for λ × λ. Thus, using the appropriate analogue of (2.3) we compute for n, n′ in Z \ {0} that for ξ in L2(R2) ∼= L2(R) ⊗2 L2(R) that Ev(n, n′)ξ(t, t′) =ZRZTZR v1(x)v2(y)¯zne−intyzeixy = δn,n′ZRZR = √2πδn,n′ZR v1(x)v2(y)e−in(x+t−t′)yξ(t − x, t′ + x) dy dx v1(x)v2(n(x + t − t′))ξ(t − x, t′ + x) dx ein′t′yξ(t − x, t′ + x) dy dz dx where δn,n′ is the Kroenecker delta symbol. Likewise Ev(n, 0) = 0 for n in Z \ {0}. We consider the unitary on L2(R) given by U ξ(t, t′) = ξ(t, t′ + t). We have for n in Z \ {0} that n′ U ∗Ev(n, n)U ξ(t, t′) = Ev(n, n)U ξ(t, t′ − t) Hence we compute kU ∗Ev(n, n)U ξk2 v1(x)v2(n(x + 2t − t′))ξ(t − x, t′) dx v1(t − x)v2(n(−x + 3t − t′))ξ(x, t′) dx = √2πZR = √2πZR 2π(cid:20)ZR v1(t − x)v2(n(−x + 3t − t′))ξ(x, t′) dx(cid:21)2 ×(cid:20)ZR v1(t − x)v2(n(−x + 3t − t′))ξ(x, t′)2 dx(cid:21) dt dt′ ∞ kv2k2 2 ≤ZRZR ≤ 2πZRZR(cid:20)ZR v1(t − x)v2(n(−x + 3t − t′)) dx(cid:21) dt dt′ ≤ 2π kv1k2 1 3n2 kξk2 2 where we have used the Cauchy-Schwarz inequality for the second inequality and Tonelli's theorem, a Holder inequality and a change of variables for the third. In summary (2.9) kEv(n, n′)k ≤ δn,n′ It is trivial to see that √2π √3n kv1k∞ kv2k1 , for (n, n′) in (Z \ {0}) × Z. Combining with (2.9) we see that dπn(Z) = −inI for n in Z. sup (n,n′)∈Z2 kE(n, n′)(dπn(Z) ⊗ I)k < ∞ and we may hence appeal to Lemma 2.3. (cid:3) 2.3. Two non-unimodular groups. We shall consider a class of groups we call Gr´elaud's groups. Fix a parameter θ > 0 and for s in R let τ (s) = exp s(cid:20)1 −θ 1 (cid:21) = es(s), where (s) =(cid:20)cos sθ − sin sθ cos sθ (cid:21) . sin sθ θ 16 HUN HEE LEE, JEAN LUDWIG, EBRAHIM SAMEI AND NICO SPRONK We now let Gθ = R2 ⋊τ R with multiplication given by (x, s)(x′, s′) = (x + τ (s)x′, s + s′), hence (x, s)−1 = (−τ (−s)x,−s). Notice that we have det τ (s) = e2s from which we get left Haar integral, for u in Cc(Gθ), and modular function u(x, s)e−2s ds dx and ∆(x, s) = e2s. ZG u(x, s) d(x, s) =ZR2ZR πy : Gθ → U(L2(R)), If y ∈ R2, with associated character χy(x) = e−ix·y on R2, we get induced repre- sentation πy(x, s) = e−i(τ (−t)x)·yξ(t − s) = e−ie−tx·((t)y)ξ(t − s). Notice that if y′ = (t′)y, then πy ∼= πy ′ via the intertwiner ρ(t′), where ρ is the left regular representation on L2(R). Hence we may parameterize these representations by the unit sphere S1. Furthermore, by [KaTa, 7.35], the family {πy}y∈S1 is a dense, To learn the disintegration of the left regular representation (2.1), we will have to obtain the Plancherel formula for this group. Since we do not know of a reference for this, we compute it ourselves. We use the notation of Proposition 2.2. compact subset of bGθ. Proposition 2.7. The Plancherel decomposition of L2(Gθ) is given by L2(Gθ) ∼=Z ⊕ S1 L2(R)y ⊗2 L2(R)y dν(y) where each L2(R)y is a copy of L2(R), ν is the unique rotationally invariant prob- ability measure on S1 and we have for ξ in L2(R)y, δyξ(t) = ∆(t)−1ξ(t) = e−2tξ(t) for a.e. t. Hence we obtain λ ≃Z ⊕ S1 πy dν(y) Proof. We will simply verify that the choices above satisfy Proposition 2.2. Let for y in S1 and u in Cc(Gθ) and ξ in L2(R) u(πy)δ1/2 y u(g)πy(g)(δ1/2 y ξ)(t) dg ξ(t) =ZG =ZRZR2 =ZR u(x, s)e−ie−tx·((t)y)es−tξ(t − s) dx ds 2πu1(e−t(t)y, t − s)e−sξ(s) ds where u1 is the partial Fourier transform in the R2-variable. We note the well- known fact that the Hilbert-Schmidt norm ku(πy)k2 is given in terms of the kernel function as(cid:0)RRRR 2πu1(e−t(t)y, s + t)2 e−2sds dt(cid:1)1/2 ZS1 ku(πy)δyk2 2 dν(y) =ZS1ZRZR 2πu1(e−t(t)y, t − s)2 e−2sds dt dν(y) =ZS1ZRZR 2πu1(e−ty, s)2 e2(t−s)ds dt dν(y) =ZR2ZR 2πu1(z, s)2 e−2sds dz =ZRZR2 u(x, s)2 dx e−2sds = kuk2 . We then see that 2 . WEAK AMENABILITY OF A(G) 17 where we have used the invariance of the chosen measures on S1 and R, then an obvious change of variables in R2, and finally the Plancherel formula in R2. (cid:3) We note that Gθ admits the Lie algebra gθ = hT, X1, X2 : [T, X1] = X1 − θX2, [T, X2] = θX1 + X2, [X1, X2] = 0i where exp(hXj) = (hej, 0) for j = 1, 2 ((e1, e2) is the standard basis for R2) and exp(hT ) = (0, h). Theorem 2.8. Let v1 in L1(R2) be so that v1 ∈ Ac(R2), let v2 ∈ Cc(R), and then let v(x, s) = v1(x)v2(s). Let X ∈ span{X1, X2}. Then there is an S = SX,v in V N (Gθ) for which S(u) =ZGθ ∂(X,0)u(g, g−1)v(g) dg for u in D(Gθ × Gθ). Proof. Our assumptions on v1 and v2 ensure that v is integrable. We use (2.3) to compute for ξ in L2(R2) ∼= L2(R) ⊗2 L2(R) that Ev(y, y′)ξ(t, t′) v1(x)v2(s)e−i(τ (−t)x)·y+(−τ (−t′−s)x)·y ′ ξ(t − s, t′ + s) dx e−2sds 2πv1(e−t(t)y − e−t′−s(t′ + s)y′)v2(s)e−2sξ(t − s, t′ + s) ds. =ZRZR2 =ZR = 2πZR = 2πZR Letting, now, U ξ(t, t′) = ξ(t, t′ + t) we compute U ∗Ev(y, y′)U ξ(t, t′) = Ev(y, y′)U ξ(t, t′ − t) (2.10) (2.11) v1(e−t(t)y − e−t′+t−s(t′ − t + s)y′)v2(s)e−2sξ(t − s, t′) ds v1(e−t(t)y − e−t′+s(t′ + s)y′)v2(t − s)e2(s−t)ξ(s, t′) ds Let us now compute dπy(X). Let (x, 0) = exp X, and for ξ ∈ Cc(R) we have [e−ie−t(hx)·((t)y) − 1]ξ(t) = −ie−tx · ((t)y)ξ(t) dπy(X)ξ(t) = lim h→0 1 h where convergence is uniform on compact sets. Hence dπy(X) admits Fy = Cc(R) in its domain, and by (2.6), and is given by pointwise multiplication by t 7→ −ie−tx · ((t)y). Notice that δ−1 y Fy ⊂ Fy. Hence we may appeal to Lemma 2.3 and it suffices to see that the field of operators (Ev(y, y′)(dπy(X) ⊗ I))y,y ′∈S1 is bounded (2.12) to gain our conclusion. It is clear that U ∗(dπy(X) ⊗ I)U = dπy(X) ⊗ I, and it suffices to work with the field (U ∗Ev(y, y′)U (dπy(X) ⊗ I))y,y ′∈S1 . We combine (2.10) and (2.11) to see that for ξ in L2(R2) we have U ∗Ev(y, y′)U (dπy(X) ⊗ I)ξ(t, t′) = 2πZR = −2πiZR v1(e−t(t)y − e−t′+s(t′ + s)y′)v2(t − s)e2(s−t)(dπy(X) ⊗ I)ξ(s, t′) ds v1(e−t(t)y − e−t′+s(t′ + s)y′)v2(t − s)es−2t(x · ((s)y))ξ(s, t′) ds. 18 HUN HEE LEE, JEAN LUDWIG, EBRAHIM SAMEI AND NICO SPRONK Our assumptions on v1 allow us to find a non-increasing ϕ in Cc([0,∞)) for which Let us observe that e−t(t)y − e−t′+s(t′ + s)y′ ≥ e−t − e−t′+s. Now with an application of Cauchy-Schwarz inequality we obtain 2πv1(y) ≤ ϕ(y). 2 kU ∗Ev(y, y′)U (dπy(X) ⊗ I)ξk2 ≤ x2ZR2(cid:20)ZR ≤ x2ZR2(cid:20)ZR ×(cid:20)ZR ϕ(e−t − e−t′+s)v2(t − s)es−2tξ(s, t′) ds(cid:21)2 ϕ(e−t − e−t′+s)e−tv2(t − s)es−t ds(cid:21) ϕ(e−t − e−t′+s)e−tv2(t − s)es−tξ(s, t′)2 ds(cid:21) d(t, t′). d(t, t′) Since e−t ≤ e−t − e−t′+s + e−t′+s we have ZR ϕ(e−t − e−t′+s)e−tv2(t − s)es−t ds ≤ZR ϕ(e−t − e−t′+s)e−t − e−t′+sv2(t − s)es−t ds +ZR ϕ(e−t − e−t′+s)e−t′+sv2(t − s)es−t ds ≤ kϕιk∞ kδ−1/2v2k1 + kϕk1 kδ−1/2v2k∞ where ι(r) = r for r ≥ 0 and δv2(s) = e2sv2(s). It then follows that kU ∗Ev(y, y′)U (dπy(X) ⊗ I)ξk2 2 ≤ x2 (kϕιk∞ kδv2k1 + kϕk1 kδv2k∞) ×ZRZR(cid:20)ZR ϕ(e−t − e−t′+s)e−tv2(t − s)es−t dt(cid:21) ξ(s, t′)2 ds dt′ ≤ x2(cid:16)kϕιk∞ kδ−1/2v2k1 + kϕk1 kδ−1/2v2k∞(cid:17)kϕk1 kδ−1/2v2k∞ kξk2 2 . Hence we have verified (2.12). (cid:3) We now consider the real affine motion group, also known as the "ax + b group", F = R ⋊ R with product and inverse given by The group has left Haar integral and modular function given by (b, a)(b′, a′) = (b + eab′, a + a′) and (b, a)−1 = (−e−ab,−a). u(b, a)d(b, a) =ZRZR ZF and ∆(b, a) = ea. u(b, a) e−ada db It is extremely well known, see, for example [Kha] or [Fol, 7.6-2], that the only inequivalent infinite-dimensional irreducible representations are given by π±, which are gained by inducing characters of positive/negative index from the normal sub- group B = {(b, 0); B ∈ R} ∼= R. Explicitly, for ξ in L2(R) we have π±(b, a)ξ(t) = e∓ie−tbξ(t − a). WEAK AMENABILITY OF A(G) 19 Furthermore, we get a quasi-equivalence (2.13) λ ≃ π+ ⊕ π−. We recall that the Lie algebra for F is given by where exp(hX) = (0, h) and exp(hY ) = (h, 0). f = hX, Y : [X, Y ] = Y i v1(b)v2(a). Then there is S = SY,v in V N (F ) such that Theorem 2.9. Let v1 in L1(R) be sobv1 ∈ Ac(R), let v2 ∈ Cc(R), and set v(b, a) = S(u) =ZF ∂(Y,0)u(g, g−1)v(g) dg for u in D(F × F ). Proof. The details of this proof are similar to those in the proof of Theorem 2.8. Indeed, we obtain for σ, σ′ in {±} and ξ in L2(R2) that v1(σe−t − σ′e−a−t′ )v2(a)e−aξ(t − a, t′ + a) da and if U ξ(t, t′) = ξ(t, t′ + t) we have Ev(σ, σ′)ξ(t, t′) = √2πZR U ∗E(σ, σ′)U ξ(t, t′) = √2πZR v1(σe−t − σ′ea−t′ )v2(t − a)ea−tξ(a, t′) da. We compute for ξ in F± = Cc(R) the derivative dπ±(Y )ξ(t) = lim g→0 1 h [e∓ie−th − I]ξ(t) = ∓ie−tξ(t). The operators from the Plancherel formula (Proposition 2.2) are given by δ±ξ(t) = e−tξ(t), and we see that δ±F± ⊆ F±. Hence we may appeal to Lemma 2.3, and it suffices to see that each of the operators Ev(σ, σ′)(dπσ ⊗ I) is bounded for each σ, σ′ in {±}. (2.14) Since U ∗(dπσ ⊗ I)U = dπσ ⊗ I it suffices compute the norms of the operators U ∗Ev(σ, σ′)U (dπσ ⊗ I). We first observe that U ∗Ev(σ, σ′)U (dπσ ⊗ I)ξ(t, t′) = √2πiσZR We then note that e−t ≤ σe−t − σ′ea−t′ v1(σe−t − σ′ea−t′ + ea−t′ and observe that )v2(t − a)e−tξ(a, t′) da. ZR v1(σe−t−σ′ea−t′ )v2(t − a)e−t da ≤ZR v1(σe−t − σ′ea−t′ +ZR v1(σe−t − σ′ea−t′ ≤ kv1αk∞ kv2k1 + kv1k1 kv2k∞ . )σe−t − σ′ea−t′ v2(t − a) da )ea−t′ v2(t − a) da 20 HUN HEE LEE, JEAN LUDWIG, EBRAHIM SAMEI AND NICO SPRONK where α(t) = t for t in R. Hence by our usual technique we see that kU ∗Ev(σ, σ′)U (dπσ ⊗ I)ξk2 2 ≤ 2π (kv1αk∞ kv2k1 + kv1k1 kv2k∞) ×ZRZRZR v1(σe−t − σ′ea−t′ ≤ 2π (kv1αk∞ kv2k1 + kv1k1 kv2k∞)kv1k1 kv2k∞ kξk2 2 . )e−tv2(t − a)ξ(a, t′)2 dt da dt′ Hence (2.14) is verified. (cid:3) 2.4. Remarks. We note that the failure of weak amenability of A(G), for G either F or Hr, is shown in [ChGh1]; but this does not automatically imply failure of local synthesis. The one aspect in common in the strategies employed for each of the five (classes of) groups G above is that the Lie derivative is always taken from a direction which is trivial in any abelian quotient. We suspect that to do otherwise would entail that for an abelian quotient, G/N , we would find that ∆G/N would be a set of non-synthesis, which would contradict Theorem 1.3 as A(G/N ) ∼= L1([G/N ) is amenable. Our choice of Lie derivatives X, in forming operators SX,v, shares a property in common with the choices made in [Joh3] for SO(3), and [ChGh1, ChGh2] for F , Hr and the Heisenberg group H. Of course, there is an enormous gulf between exploiting these for a failure of spectral synthesis calculation and showing that these derivatives may be used to build non-trivial elements of H 1(A(G), V N (G)). It is plausible that for each of our four semi-direct product groups G = N ⋊ A, above, with the associated Lie algebra n ⋊ a, that the space of bounded derivations from A(G) to V N (G) is isomorphic to the Lie algebra n. 3. Weak amenability of Fourier algebras The following is well-known. We include a proof for convenience of non-specialists. Proposition 3.1. (i) Let g be a non-abelian real Lie algebra. Then g contains either su(2) or a non-abelian solvable algebra m. (ii) Let m be a non-abelian solvable real Lie algebra. Then m contains one of the following Lie algebras: f = hX, Y : [X, Y ] = Y i (affine motion) e = hT, X1, X2 : [T, X1] = X2, [T, X2] = −X1, [X1, X2] = 0i (Euclidean motion) gθ = hT, X1, X2 : [T, X1] = X1 − θX2, [T, X2] = θX1 + X2, [X1, X2] = 0i, θ > 0, (Gr´elaud), or h = hX, Y, Z : [X, Y ] = Z, [X, Z] = 0 = [Y, Z]i (Heisenberg). Proof. Every concept and fact from Lie theory which is used in this proof is well- known and may be found in [HiNe], for example. (i) Let r denote the solvable radical ideal of g. If r is non-abelian, we set m = r. If r is abelian, we use the Levi-decomposition g = s + r, where s is semisimple. For s we have the Iwasawa decomposition s = k + a + n. If k is non-abelian, then k contains su(2) and we are done. Otherwise m = a + n is non-abelian and solvable; WEAK AMENABILITY OF A(G) 21 indeed, see the construction in [HiNe, §13.3]. [In [ChGh1, Proposition 5.3] a further refinement shows that m ⊇ f, in this case.] (ii) We let m ) m′ ) ··· ) m(d−1) ) {0} be the derived series, so m(d−2) is non abelian. For simplicity we assume m = m(d−2); thus we now have derived series m ) m′ ) m′′ = {0}. Suppose, first there is an element S of m for which ad S : m → m admits a non-zero eigenvalue a + ib, a, b ∈ R. If b = 0, then there is an eigenvector Y for ad S, i.e. [S, Y ] = aY . Put X = 1 a S and we get f = span{X, Y }. Otherwise, we appeal to the real Jordan decomposition of ad S, which allows us to find elements Y1, Y2 of m for which [S, Y1] = aY1 − bY2 and [S, Y2] = bY1 + aY2. a S and Xj = Yj (j = 1, 2) if ab > 0; and let T = 1 If a = 0, we let T = 1 b S and we notice immediately that Y1 = X1, Y2 = X2 ∈ m′, so [Y1, Y2] = 0. Thus we get e = span{T, Y1, Y2}. Finally, if ab 6= 0, we let θ = b/a. Let T = 1 b S, X1 = Y2, X2 = Y1 if ab < 0. It is straightforward to check that X1, X2 ∈ m′ so [X1, X2] = 0. Hence, gθ = span{T, X1, X2}, in this case. If no S in m has that ad S admits a non-zero eigenvalue, then each ad S is nilpotent, and hence by Engel's theorem m is nilpotent. We consider the central series m ) C(m) ) ··· ) Cn(m) ) {0}. Any Z in Cn(m) is hence central and we further consider such Z for which there are X, Y in m (one of which is in Cn−1(m)) for which Z = [X, Y ]. Then h = span{X, Y, Z}. Theorem 3.2. Let G be a connected Lie group. Then the following are equivalent: (cid:3) (i) A(G) is weakly amenable, (ii) the anti-diagonal ∆G is of local synthesis for G × G, and (iii) G is abelian. we may restrict ourselves to the case of simply connected groups. Proof. The result (i) ⇒ (ii) is Theorem 1.3, while (iii) ⇒ (i) is well-known, i.e. A(G) ∼= L1(bG), in this case. Hence it remains to prove that (ii) ⇒ (iii). We will assume that G is non-abelian, and show that ∆G is not of local synthesis for G× G. All Lie theoretic concepts and nomenclature, used in this proof, may be found in [HiNe]. We will show that when G is non-abelian, then ∆G fails to be of local synthesis for G × G. Thanks to Theorem 1.6, and the fact that any G ∼= eG/Γ for a simply connected connected Lie group eG and a discrete central subgroup Γ of eG, Let G be simply connected with Lie algebra g. The Levi decomposition g = s + r begets respective integral subgroups S and R of G, with S semisimple and R normal, and hence, essentially by the smooth splitting theorem, each is closed and simply connected. Given a Lie subalgebra m of g, in particular one of the algebras arising from Proposition 3.1, above, we let M be the integral subgroup generated by m. If m = su(2), so m ⊂ s, then M ∼= SU(2)/C, where C is a subgroup of the centre {±1}, and hence M is closed. If m ⊂ s and is non-abelian and solvable, then M ⊆ AN (Iwasawa decomposition of S); or if m ⊆ r, then M ⊆ R. Any integral subgroup of a simply connected solvable group is closed and simply connected. Thus, corresponding to each of the solvable Lie algebras f, e, gθ or h, as obtained in Proposition 3.1, above, we see that M is closed and isomorphic to one of the affine motion group F , the simply connected cover eE of the Euclidean motion group E, G´elaud's group Gθ, or the Heisenberg group H. 22 HUN HEE LEE, JEAN LUDWIG, EBRAHIM SAMEI AND NICO SPRONK Proposition 2.4, Theorem 2.5, Theorem 2.6, Theorem 2.8, and Theorem 2.9, and then Lemma 2.1, tells us that ∆H fails to be a set of (local) synthesis for H × H, for each of H = SU(2), E, Hr, Gθ and F . Then, Theorem 1.6 tells us the same for H = SU(2)/C, eE and H. Hence applying Theorem 1.4 to the subgroup M of G, above, we obtain (ii). Let G be a locally compact group. The following is an immediate consequence of the restriction theorem for closed subgroups H, i.e. RH (A(G)) = A(H) (see the comments before Theorem 1.4); and the fact that any quotient of a commutative weakly amenable Banach algebra is again weakly amenable. (cid:3) Corollary 3.3. Let G be a locally compact group. If A(G) is weakly amenable, then any connected Lie subgroup of G is abelian. Example 3.4. Let us consider an example of a non-abelian solvable connected group which contains no non-abelian closed connected Lie subgroups. We consider motion group. We let Rap denote the almost periodic compactification of the real line, and for each r in R we let χr : Rap → T be the unique continuous character extending and is isomorphic to the discretized reals, Rd. If K is a closed connected Lie a group related to the simply connected covering group, eE = C⋊R, of the Euclidean t 7→ eitr : R → T. We observe that the dual group is given by dRap = {χr : r ∈ R} subgroup of Rap, then K ∼= Tn, where n = 0, 1, 2, . . . , so bK ∼= Zn. On the other hand bK is a quotient of Rd, and hence is either trivial or divisible. Thus K = {1}, and we conclude that Rap contains no non-trivial closed Lie subgroups. Now let E = C ⋊ Rap with multiplication given by (x, ζ)(x′, ζ ′) = (x + χ1(ζ)x′, ζζ ′). We may think of this group as a partial compactification of eE along its quotient subgroup. We note that q(x, ζ) = (x, χ1(ζ)) defines a quotient homomomorphism from E onto E. For notation convenience we identify Rap ∼= {0} ⋊ Rap in E, and T ∼= {0} ⋊ T in E, and we identify C with C ⋊ {1} in either E, or in E, as should be clear by context. Notice that qC is a homeomorphism. We now show that E admits no non-abelian closed connected Lie subgroups. We first note that the proper Lie subalgebras of e are either one-dimensional, the two- dimensional ideal n = RX1 + RX2, or {0}. Since every closed connected subgroup of E corresponds to a Lie subalgebra of e, we find that the one-dimensional closed connected subgroups are of the form MS+X = exp(R(S + X)) = {((z − 1)x, z) : z ∈ T} and LX = exp(RX) = {(tx, 1) : t ∈ R} where X ∈ n with exp X = (x, 1), while the only two-dimensional closed connected subgroup is C. Taking q−1(M ) for each subgroup M of E, listed above, we get M S+X = {((χ1(ζ) − 1)x, ζ) : ζ ∈ Rap}, LX ⋊ ker χ1 and C ⋊ ker χ1. Let H be a proper closed connected subgroup of E and H0 = H ∩ C. Notice that in E, q(H0) = q(H) ∩ C ∼= H0. If H0 = C then for (x, ζ) in H we have (0, ζ) = (−x, 1)(x, ζ) ∈ H ∩ Rap, so the image of H under the quotient map (x, ζ) 7→ ζ is closed, and hence either the trivial group {1} or is non-Lie. If H0 ( C, then WEAK AMENABILITY OF A(G) 23 q(H) ∩ C ∼= H0 ( C, and it follows that H is a subgroup of one of the abelian groups M S+X , LX ⋊ ker χ1 or Rap, and hence is abelian. At present, we are aware of no method for determining if A(E) is weakly amenable. Acknowledgement. We are grateful to Søren Knudby for pointing out a flaw in an earlier version of the proof of Theorem 3.2, and for guidance in correcting it. We also thank Yemon Choi for pointing out an error in a proposition which purported to reduce the problem to the case of pro-solvable Ge. We realized there was a serious gap in the proof and have deleted the proposition from the final version of the paper. References [Ars] G. Arsac. Sur l'espace de Banach engendr´e par les coefficients d'une repr´esentation unitaire. Publ. D´ep. Math. (Lyon) 13 (1976), no. 2, 1 -- 101. [BCD] W.G. Bade, P. Curtis and H.G. Dales. Amenability and weak amenability for Beurling and Lipschitz algebras. Proc. London Math. Soc. (3) 55 (1987), no. 2, 359 -- 377. [ChGh1] Y. Choi and M. Ghandehari. Weak and cyclic amenability for Fourier algebras of con- nected Lie groups. J. Funct. Anal. 266 (2014), no. 11, 6501 -- 6530. [ChGh2] Y. Choi and M. Ghandehari. Weak amenability for Fourier algebras of 1-connected nilpo- tent Lie groups, pre-print, 20 pages. See arXiv:1405.6403. [DeDe] J. Delaporte and A. Dehrighetti. On Herz' extension theorem. Boll. Un. Mat. Ital. A (7) 6 (1992), no. 2, 245 -- 247. [Eym] P. Eymard. L'alg`ebre de Fourier d'un groupe localement compact. Bull. Soc. Math. France 92 (1964), 181 -- 236. [Fol] G.B. Folland. A course in abstract harmonic analysis. CRC Press, Boca Raton, FL, 1995. [FoRu] B.E. Forrest and V. Runde. Amenability and weak amenability of the Fourier algebra. Math. Z. 250 (2005), no. 4, 731 -- 744. [FSS1] B.E. Forrest, E. Samei and N. Spronk. Weak amenability of Fourier algebras on compact groups. Indiana Univ. Math. J. 58 (2009), no. 3, 1379 -- 1393. [FSS2] B.E. Forrest, E. Samei and N. Spronk. Convolutions on compact groups and Fourier alge- bras of coset spaces. Studia Math. 196 (2010), no. 3, 223 -- 249. [GrRy] I.S. Gradshteyn and I.M. Ryzhik. Table of integrals, series, and products. Translated from the Russian. Translation edited and with a preface by Alan Jeffrey and Daniel Zwillinger. Seventh edition. Elsevier/Academic Press, Amsterdam, 2007. [Grø] N. Groenbaek. A characterization of weakly amenable Banach algebras. Studia Math. 94 (1989), no. 2, 149 -- 162. [Her] C. Herz. Harmonic synthesis for subgroups. Ann. Inst. Fourier (Grenoble) 23 (1973), no. 3, 91 -- 123. [HeRoII] E. Hewitt and K.A. Ross. Abstract harmonic analysis. Vol. II: Structure and analysis for compact groups. Analysis on locally compact Abelian groups. Springer-Verlag, New York- Berlin, 1970. [HiNe] J. Hilgert and K.H. Neeb. Structure and geometry of Lie groups. Springer, New York, 2012. [Joh1] B.E. Johnson. Cohomology in Banach algebras. Memoirs of the American Mathematical Society, No. 127. American Mathematical Society, Providence, R.I., 1972. [Joh2] B.E. Johnson. Weak amenability of group algebras. Bull. London Math. Soc. 23 (1991), no. 3, 281 -- 284. [Joh3] B.E. Johnson. Non-amenability of the Fourier algebra of a compact group. J. London Math. Soc. (2) 50 (1994), no. 2, 361 -- 374. [Kam] H. Kamowitz. Cohomology groups of commutative Banach algebras. Trans. Amer. Math. Soc. 102 (1962), 352 -- 372. [Kan] E. Kaniuth. A course in commutative Banach algebras. Springer, New York, 2009. [KaTa] E. Kaniuth and K.F. Taylor. Induced representations of locally compact groups. Cam- bridge University Press, Cambridge, 2013. [Kha] I. Khalil. Sur l'analyse harmonique du groupe affine de la droite. Studia Math. 51 (1974), 139 -- 167. 24 HUN HEE LEE, JEAN LUDWIG, EBRAHIM SAMEI AND NICO SPRONK [KiMu] W. Kirsch and D. Muller. On the synthesis problem for orbits of Lie groups in Rn. Ark. Mat. 18 (1980), no. 2, 145 -- 155. [Lep] H. Leptin. Sur l'alg`ebre de Fourier d'un groupe localement compact. C. R. Acad. Sci. Paris S´er. A-B 266 (1968), A1180 -- A1182. [Loh] N. Lohoue. Remarques sur les ensembles de synth`ese des alg'ebres de groupe localement compact. J. Funct. Anal. 13 (1973), 185 -- 194. [Los] V. Losert. On tensor products of Fourier algebras. Arch. Math. (Basel) 43 (1984), no. 4, 370 -- 372. [LuTu1] J. Ludwig and L. Turowska. On the connection between sets of operator synthesis and sets of spectral synthesis for locally compact groups. J. Funct. Anal. 233 (2006), no. 1, 206 -- 227. [LuTu2] J. Ludwig and L. Turowska. Growth and smooth spectral synthesis in the Fourier algebras of Lie groups. Studia Math. 176 (2006), no. 2, 139 -- 158. [McM] J.R. McMullen. Extensions of positive-definite functions. Memoirs of the American Math- ematical Society, No. 117. American Mathematical Society, Providence, R. I., 1972. [Mul] D. Muller. On the spectral synthesis problem for hypersurfaces of RN . J. Funct. Anal. 47 (1982), no. 2, 247 -- 280. [PaSa] B.D. Park and E. Samei. Smooth and weak synthesis of the anti-diagonal in Fourier alge- bras of Lie groups. J. Lie Theory 19 (2009), no. 2, 275 -- 290. [Ply] R.J. Plymen. Fourier algebra of a compact Lie group, unpublished. See ArXiv math/0104018. [ReSt] H. Reiter and J. Stegeman. Classical harmonic analysis and locally compact groups. Second edition. The Clarendon Press, Oxford University Press, New York, 2000. [Rua] Z.-J. Ruan. The operator amenability of A(G). Amer. J. Math. 117 (1995), no. 6, 1449 -- 1474. [Run] V. Runde. A functorial approach to weak amenability for commutative Banach algebras. Glasgow Math. J. 34 (1992), no. 2, 241 -- 251. [Sam1] E. Samei. Bounded and completely bounded local derivations from certain commutative semisimple Banach algebras. Proc. Amer. Math. Soc. 133 (2005), no. 1, 229 -- 238 [Sam2] E. Samei. Hyper-Tauberian algebras and weak amenability of Fig`a-Talamanca -- Herz alge- bras. J. Funct. Anal. 231 (2006), no. 1, 195 -- 220. [SiWe] I.M. Singer and J. Wermer. Derivations on commutative normed algebras. Math. Ann. 129, (1955). [Spr] N. Spronk. Operator weak amenability of the Fourier algebra. Proc. Amer. Math. Soc. 130 (2002), no. 12, 3609 -- 3617. [Sug] M. Sugiura. Unitary representations and harmonic analysis. An introduction. Second edi- tion. North-Holland Publishing Co., Amsterdam; Kodansha, Ltd., Tokyo, 1990. [Tat] N. Tatsuuma. Plancherel formula for non-unimodular locally compact groups. J. Math. Kyoto Univ. 12 (1972), 179 -- 261. [Tre] F. Treves. Topological vector spaces, distributions and kernels. Academic Press, New York- London, 1967. [Var] N.Th. Varopoulos. Spectral synthesis on spheres. Proc. Cambridge Philos. Soc. 62 (1966), 379?387. [War] C.R. Warner. Weak spectral synthesis. Proc. Amer. Math. Soc. 99 (1987), no. 2, 244?248. Addresses: Department of Mathematical Sciences and Research Institute of Math- ematics, Seoul National University, Gwanak-ro 1, Gwanak-gu, Seoul 08826, Republic of Korea Institut ´Elie Cartan de Lorraine, Universit´e de Lorraine -- Metz, Batiment A, Ile du Saulcy, F-57045 Metz, France Department of Mathematics and Statistics, University of Saskatchewan, Room 142 McLean Hall, 106 Wiggins Road Saskatoon, SK, S7N 5E6, Canada Department of Pure Mathematics, University of Waterloo, Waterloo, ON, N2L 3G1, Canada WEAK AMENABILITY OF A(G) 25 Email-adresses: [email protected] [email protected] [email protected] [email protected]
1107.1700	1	1107	2011-07-08T19:04:59	Wavelet analysis on adeles and pseudo-differential operators	[ "math.FA" ]	This paper is devoted to wavelet analysis on adele ring $\bA$ and the theory of pseudo-differential operators. We develop the technique which gives the possibility to generalize finite-dimensional results of wavelet analysis to the case of adeles $\bA$ by using infinite tensor products of Hilbert spaces. The adele ring is roughly speaking a subring of the direct product of all possible ($p$-adic and Archimedean) completions $\bQ_p$ of the field of rational numbers $\bQ$ with some conditions at infinity. Using our technique, we prove that $L^2(\bA)=\otimes_{e,p\in\{\infty,2,3,5,...}}L^2({\bQ}_{p})$ is the infinite tensor product of the spaces $L^2({\bQ}_{p})$ with a stabilization $e=(e_p)_p$, where $e_p(x)=\Omega(\|x\|_p)\in L^2({\bQ}_{p})$, and $\Omega$ is a characteristic function of the unit interval $[0,\,1]$, $\bQ_p$ is the field of $p$-adic numbers, $p=2,3,5,...$; $\bQ_\infty=\bR$. This description allows us to construct an infinite family of Haar wavelet bases on $L^2(\bA)$ which can be obtained by shifts and multi-delations. The adelic multiresolution analysis (MRA) in $L^2(\bA)$ is also constructed. In the framework of this MRA another infinite family of Haar wavelet bases is constructed. We introduce the adelic Lizorkin spaces of test functions and distributions and give the characterization of these spaces in terms of wavelet functions. One class of pseudo-differential operators (including the fractional operator) is studied on the Lizorkin spaces. A criterion for an adelic wavelet function to be an eigenfunction for a pseudo-differential operator is derived. We prove that any wavelet function is an eigenfunction of the fractional operator. These results allow one to create the necessary prerequisites for intensive using of adelic wavelet bases and pseudo-differential operators in applications.	math.FA	math	WAVELET ANALYSIS ON ADELES AND PSEUDO-DIFFERENTIAL OPERATORS A. YU. KHRENNIKOV, A. V. KOSYAK, AND V. M. SHELKOVICH Abstract. This paper is devoted to wavelet analysis on adele ring A and the theory of pseudo-differential operators. We develop the technique which gives the possibility to generalize finite-dimensional results of wavelet analysis to the case of adeles A by using infinite tensor products of Hilbert spaces. The adele ring is roughly speaking a subring of the direct product of all possible (p-adic and Archimedean) completions Qp of the field of rational numbers Q with some conditions at infinity. Using our technique, we prove that L2(A) = ⊗e,p∈{∞,2,3,5,... }L2(Qp) is the infinite tensor product of the spaces L2(Qp) with a stabilization e = (ep)p, where ep(x) = Ω(xp) ∈ L2(Qp), and Ω is a characteristic function of the unit interval [0, 1], Qp is the field of p-adic numbers, p = 2, 3, 5, . . . ; Q∞ = R. This description allows us to construct an infinite family of Haar wavelet bases on L2(A) which can be obtained by shifts and multi-delations. The adelic multiresolution analysis (MRA) in L2(A) is also constructed. In the framework of this MRA another infinite family of Haar wavelet bases is constructed. We introduce the adelic Lizorkin spaces of test functions and distributions and give the characterization of these spaces in terms of wavelet functions. One class of pseudo-differential operators (including the fractional operator) is studied on the Lizorkin spaces. A criterion for an adelic wavelet function to be an eigenfunction for a pseudo-differential operator is derived. We prove that any wavelet function is an eigenfunction of the fractional operator. These results allow one to create the necessary prerequisites for intensive using of adelic wavelet bases and pseudo-differential operators in applications. 1. Introduction 1.1. p-Adic and adelic analysis. During a few hundred years theoretical physics has been developed on the basis of real and, later, complex numbers. The p-adic numbers were described by K. Hensel in 1897 to transfer the ideas and techniques of power series methods to number theory. According to the well-known Ostrovsky theorem, any nontrivial valuation on the field Q of rational numbers is equivalent either to the real valuation · or to one of the p-adic valuations · p. The corresponding completions of Q give the fields R or Qp. The theory of p-adic numbers has already penetrated intensively into several areas of mathematics and its applications. In the last 20 years the field of p-adic numbers Qp (as well as its algebraic extensions, including the field of complex p-adic numbers Cp) has been intensively used in theoretical and mathematical physics, p-adic string theory, gravity and cosmology, the theory of stochastic differential equations over the field of p-adic numbers, Feynman path integration over p-adics, the Date: 2000 Mathematics Subject Classification. 11F85, 42C40, 47G30; Secondary 26A33, 46F10. Key words and phrases. Adeles, wavelets, multiresolution analysis, pseudo-differential operators, fractional operator. The first and third authors (A. K. and V. S.) were supported by the grant of the Profile Mathematical Modeling and System Collaboration of Vaxjo University (Sweden). The third author (V. S.) was supported in part by Grant 09-01-00162 of Russian Foundation for Basic Research. The second (A. K.) and third (V. S.) were supported by the DFG Projects. 1 2 A. YU. KHRENNIKOV, A. V. KOSYAK, AND V. M. SHELKOVICH theory of p-adic valued probabilities and dynamical systems, in theory of disordered systems (spin glasses) (see [8], [26], [27], [30], [38], [40], [64] and the references therein). Applications were, however, not only restricted to physics. p-Adic models were also proposed in psychology, cognitive and social sciences, and, e.g., in biology, image analysis (see [27], [28]). These applications induced and stimulated a development of new branches of p-adic analysis, in particular, the theory of p-adic wavelets. Recall that nowadays wavelets are applied in a lot of branches of modern mathematics and engineering area. The first real wavelet basis (3.5), (3.6) was introduced by Haar in 1910. However, for almost a century nobody could find another wavelet function (a function whose shifts and delations form an orthogonal basis). Only in the early nineties a method for a more general construc- tion of the wavelet functions appeared [46], [47], [50], [51]. This method is based on the notion of multiresolution analysis (MRA in the sequel). In the p-adic setting the situation was the following. In 2002 S. V. Kozyrev [39] constructed a compactly sup- ported p-adic wavelet basis (3.21) in L2(Qp), which is an analog of the real Haar basis (3.5), (3.6). J. J. Benedetto and R. L. Benedetto [10], R. L. Benedetto [11] suggested a method for constructing wavelet bases on locally compact abelian groups with compact open subgroups. This method is applicable for the p-adic setting. It is based on a theory of wavelet sets and only allows the construction of wavelet functions whose Fourier trans- forms are the characteristic functions of some sets (see [10, Proposition 5.1.]). Moreover, these authors doubted that the development of the MRA approach is possible. In spite of the above opinions and arguments [10], [11], in [57], the p-adic MRA theory in L2(Qp) was developed and new p-adic wavelet bases were constructed. Some important results in p-adic wavelet theory were obtained in [1], [2], [34], [35]. It turned out that the the- ory of p-adic wavelets plays an important role in the study of p-adic pseudo-differential operators and equations [3], [4], [5], [33], [39], [40], [57]. This theory gives a powerful technique to deal with p-adic pseudo-differential equations. Recall that on complex- valued functions defined on Qp, the operation of differentiation is not defined . As a result, a large number of p-adic models use pseudo-differential equations instead of dif- ferential equations. The p-adic multidimensional fractional operator Dα was introduced by M. Taibleson [58] (see also [59]) in the space of distributions D′(Qn p ). The spectral theory of this fractional operator was developed by V.S. Vladimirov in [62], in particular, explicit formulas for the eigenfunctions of this operator were constructed (see also [64]). In [63] (see also [64]) V.S. Vladimirov constructed the spectral theory of the Schrodinger- type operator Dα + V (x), which was further developed by A.N. Kochubei [36], [37] (see also [38]). The adele ring A is some subring of the direct product of all possible (p-adic and Archimedean) completions Qp of the field of rational numbers Q. The group of ideles was introduced by Chevalley in 1936 [16] as a part of his program to formulate class field theory so that it worked for infinite-degree extensions. The adeles were introduced by Weil in the late 1930s as an additive analogue of ideles. The ring A is often used in advanced parts of number theory (for example, see [13], [15], [17], [48], [49], [65]). For example, Tate's proof (see in [15, J. T. Tate, pp.305 -- 347]) of the functional equa- tion Z(s) = Z(1 − s), where Z(s) = π−s/2Γ(s/2)ζ(s) for the Riemann zeta func- ns , ℜ(s) > 1, is based on the Fourier analysis on adeles A and ideles (see also [17, 4.7]). Recently the theory of adeles has been successfully ap- plied in various parts of contemporary mathematical and theoretical physics. Namely, tion ζ(s) = Pn∈N 1 WAVELET ANALYSIS ON ADELES AND PSEUDO-DIFFERENTIAL OPERATORS 3 there are adelic constructions, e.g., in statistical mechanics, stochastics, string the- ory, quantum cosmology, and quantum mechanics (see [9], [14], [19] -- [21], [25], [38, 5.8] [26], [27], [60], [61], [64], [66], and the references therein). However it should be noted that in contrast to the real and p-adic analysis, the adelic analysis practically (in particular, the adelic wavelet theory and the theory of adelic pseudo-differential operators and equations) has not been developed so far. In particular, the necessity of the development the wavelet theory on adeles was mentioned in [10]. 1.2. Contents of the paper. In this paper some problems of the adelic harmonic analysis are studied. The points to be considered are, first, the theory of adelic Haar wavelets; and second, the theory of simplest adelic pseudo-differential operators in con- nection with wavelets. In Sec. 2, some facts from the p-adic and adelic analysis are given. In Sec. 3, we recall definitions of the real MRA and present some results on p-adic MRA and wavelet bases from the papers [35], [57]. In Sec. 4, the theory of infinite tensor products of Hilbert spaces [52], [12] is used to generalize finite-dimensional results to the case of adeles. We recall the constructions of infinite tensor product of Hilbert spaces He = ⊗e,n∈NHn, the complete von Neumann product of infinitely many Hilbert spaces. The space He can be obtained as the closure of the union of some subspaces. We observe certain stability of the space He with respect to the variation of the corresponding subspaces (see Lemma 4.2). In Subsec. 4.4, using Lemma 4.2 for the special stabilizing sequence e = (ep)p, ep(xp) = φp(xp) = Ω(xpp) (here Ω(t) is the characteristic function of the segment [0, 1] ⊂ R) we show that L2(A) coincides with the infinite tensor product of the Hilbert spaces L2(Qp) over all possible completions of the field Q: L2(A) = ⊗e,pL2(Qp) (see Lemma 4.3). In Sec. 5, we apply the above scheme to construct adelic wavelet bases (5.5) on L2(A, dx) generated by the tensor product of one-dimensional Haar wavelet bases (3.5), (3.22), (3.25). We would like to stress that to construct adelic wavelet bases, we need the p-adic wavelet bases that contain functions φp(xp) = Ω(xpp). Thus, instead of the Haar wavelet basis (3.20) we will use modified Haar basis (3.22). According to [1], there are no orthogonal p-adis MRA based wavelet bases except for those described in Theorem 3.1 and Corollary 3.1. Thus, formula (5.5) gives all adelic Haar wavelet bases generated by the tensor product of one-dimensional Haar wavelet bases. Here infinite tensor product depends on the special stabilization e = (ep)p, where ep(xp) = Ω(xpp). Recall that in [20, Sec.4, formula (4.1)], a basis in the space L2(A) associated with eigenfunctions of harmonic oscillators was constructed (see Remark 5.1). To construct adelic wavelet bases, we use the standard Haar measures on Qp and on the adele ring A. For measures on Qp different from Haar measure, see [3, Appendix D], based on [42]. In Sec. 6, using the idea of constructing separable multidimensional MRA by means of the tensor product of one-dimensional MRAs suggested by Y. Meyer [50] and S. Mal- lat [46], [47], we construct adelic wavelet bases in L2(A). In a general situation we show how using some system of closed subspaces V (k) , j ∈ Z, in a Hilbert spaces Hn, n ∈ N, with properties (a) -- (c) of the MRA (see Definitions 3.1, 3.2) one can construct various systems of subspaces with the same properties in the infinite tensor product of the spaces Hn. In Theorem 6.2 adelic separable MRA is constructed. The refinable function of this MRA given by (6.14) is an infinite product refinable functions of all one-dimensional j 4 A. YU. KHRENNIKOV, A. V. KOSYAK, AND V. M. SHELKOVICH MRAs. In the framework of this MRA an infinite family of adelic Haar wavelet bases (6.25) are constructed. In Sec. 7, we introduce the adelic Lizorkin spaces of test functions Φ(A) and distri- butions Φ′(A). These spaces are constructed by using the original real Lizorkin spaces introduced in [43] -- [45] and the p-adic Lizorkin spaces introduced in [4] (see also [3, Ch. 7]). A basic motivation for using Lizorkin spaces rather than the Schwartz and Bruhat -- Schwartz spaces of distributions S ′(R) and D′(Qp) is due to the fact that the latter spaces are not invariant under the fractional operators. Next, in Sec. 8, the char- acterization of the adelic Lizorkin spaces in terms of wavelets is given. Namely, it is proved that any test function from Φ(A) can be represented in the form of a finite combination of adelic wavelet functions (5.6), and any distribution from Φ′(A) can be represented as an infinite linear combination of adelic wavelet functions (5.6) (in [7], assertions of these types were stated for ultrametric Lizorkin spaces). In the framework of our constructions the following three facts seem to have the same deep reason: (1) functions φ(p)(xp) = Ω(xpp) are stabilization functions in the adelic Bruhat -- Schwartz space (see Definition 2.3), (2) we use the stabilization sequence e = (ep)p where ep(xp) = Ω(xpp) in proving the fact that L2(A) = ⊗e,pL2(Qp), (3) under the projection of the space L2(Qp) onto L2(Zp) some elements of wavelet basis (3.20) and Kozyrev's wavelet basis (3.21) are transformed into functions which are proportional to the same function Ω(xpp) (see Propositions 3.1 -- 3.3). In Sec. 9, by Definition (9.17), (9.18) a class of pseudo-differential operators on the adelic Lizorkin spaces is introduced. The fractional operators Dbγ, bγ ∈ C∞ (see Def- inition (9.9), (9.10), (9.13)), and Dγ, γ ∈ C (see Definition (9.14), (9.15)) belong to this class. We prove that the Lizorkin spaces of test functions Φ(A) and distributions Φ′(A) are invariant under the above-mentioned pseudo-differential operators. More- over, a family of fractional operators on the space of distributions Φ′(A) forms an abelian group. Thus the Lizorkin spaces constitute "natural" domains of definition for this class of pseudo-differential operators. Note, that in [31] the fractional operator was considered in L2(A). In Subsec. 9.4, the spectral theory of adelic pseudo-differential operators is developed. By Theorem 9.1, we derive a criterion for an adelic wavelet function to be an eigenfunction for a pseudo-differential operator. It is proved that any wavelet function is an eigenfunction of a fractional operator. Thus the adelic wavelet analysis is closely connected with the spectral analysis of pseudo-differential operators. Using results of Sec. 9, similarly to the p-adic case, one can develop the "variable separation method" (an analog of the classical Fourier method) to reduce solving adelic pseudo-differential equations to solving ordinary differential equations with respect to the real variable t (for details, see [3, Ch. 10], [6]). 2. Preliminary results 2.1. p-Adic numbers. We shall systematically use the notation and results from the book [64]. Let N, Z, Q, R, C be the sets of positive integers, integers, rational, real, and complex numbers, respectively. According to the well-known Ostrovsky theorem, any nontrivial valuation on the field of rational numbers Q is equivalent either to the real valuation · or to one of the p- adic valuations · p. This p-adic norm · p is defined as follows: if an arbitrary rational number x 6= 0 is represented as x = pγ m n , where γ = γ(x) ∈ Z and the integers m, n are not divisible by p, then xp = p−γ, 0p = 0. (2.1) x 6= 0, WAVELET ANALYSIS ON ADELES AND PSEUDO-DIFFERENTIAL OPERATORS 5 The norm · p is non-Archimedean and satisfies the strong triangle inequality x + yp ≤ max(xp,yp). The completion of Q with respect to the usual absolute value · gives the field of real numbers R. The field Qp of p-adic numbers is defined as the completion of the field of rational numbers Q with respect to the norm · p. Next we will denote x∞ = x, Q∞ = R and Z∞ = Z. By (2.2) VQ = {∞, 2, 3, 5, . . .} we denote the set of indices for all valuations on the field Q. Any p-adic number x ∈ Qp, x 6= 0, is represented in the canonical form (2.3) x = where γ = γ(x) ∈ Z, xk ∈ Fp = {0, 1, . . . , p − 1}, xγ 6= 0, γ ≤ k < ∞. The series is convergent in the p-adic norm · p, and one has xp = p−γ. The fractional part of the number x ∈ Qp (given by (2.3)) is defined as follows (2.4) if γ(x) ≥ 0 or x = 0, xγpγ + · · · + x−1p−1, if γ(x) < 0. {x}p =(cid:26) 0, The set Q× p = Qp \ {0} is the multiplicative group of the field Qp. p-Adic numbers Zp = {x ∈ Qp : xp ≤ 1} are called integer p-adic numbers. In view of (2.3), Zp consists of p-adic numbers xkpk, ∞Xk=γ xkpk. ∞Xk=0 (2.5) x = Zp is a subring of the ring Qp. The multiplicative group of the ring Zp is the set Z× p =(cid:8)x ∈ Zp : xp = 1(cid:9) =nx ∈ Zp : x = ∞Xk=0 xkpk, x0 6= 0o. Denote by Bγ(a) = {x ∈ Qp : x − ap ≤ pγ} the ball of radius pγ with the center at the point a ∈ Qp and by Sγ(a) = {x ∈ Qp : x − ap = pγ} = Bγ(a) \ Bγ−1(a) the corresponding sphere, γ ∈ Z. For a = 0 we set Bγ = Bγ(0) and Sγ = Sγ(0). 2.2. p-Adic distributions. A complex-valued function f defined on Qp is called locally- constant if for any x ∈ Qp there exists an integer l(x) ∈ Z such that f (x + y) = f (x), y ∈ Bl(x). Let E(Qp) and D(Qp) be the linear spaces of locally-constant C-valued functions on Qp and locally-constant C-valued compactly supported functions (so-called Bruhat -- Schwartz test functions), respectively [64, VI.1.,2.]. If ϕ ∈ D(Qp), then according to Lemma 1 from [64, VI.1.], there exists l ∈ Z, such that ϕ(x + y) = ϕ(x), y ∈ Bl, x ∈ Qp. The largest of such numbers l = l(ϕ) is called the parameter of constancy of the function ϕ. Let us denote by Dl N (Qp) from D(Qp) having supports in the ball BN and with parameters of constancy ≥ l. The following embedding holds: N (Qp) ⊂ Dl′ Dl N (Qp) the finite-dimensional space of test functions Dl N ′(Qp), N ≤ N ′, l ≥ l′. 6 A. YU. KHRENNIKOV, A. V. KOSYAK, AND V. M. SHELKOVICH N→∞ indDN (Qp), where DN (Qp) = lim We have D(Qp) = lim N (Qp) (see [64, VI.2]). These representations give us the inductive limit topology on the corresponding spaces. Denote by D′(Qp) the set of all linear functionals (Bruhat -- Schwartz distributions) on The Fourier transform of ϕ ∈ D(Qp) is defined by the formula D(Qp) [64, VI.3.]. l→−∞ indDl bϕ(ξ) = F [ϕ](ξ) =ZQp χp(ξx)ϕ(x) dx, ξ ∈ Qp, where dx is the Haar measure on Qp such thatRxp≤1 dx = 1, and χp(x) = e2πi{x}p (2.6) is the additive character on Qp (see [64, III.1.]), {x}p is the fractional part (2.4) of the number x ∈ Qp. The Fourier transform is a linear isomorphism D(Qp) onto D(Qp) [59, III,(3.2)], [64, VII.2.]. Moreover, ϕ ∈ Dl iff F [ϕ] ∈ D−N −l (Qp). N (Qp) (2.7) The Fourier transform of a distribution f ∈ D′(Qp) is the distribution bf = F [f ] defined by the relation hF [f ], ϕi = hf, F [ϕ]i, ∀ ϕ ∈ D(Qp). Here and in the sequel hf, ϕi denotes the action of a distribution f on a test function ϕ. If f ∈ D′(Qp), a ∈ Q× p , b ∈ Qp, then [64, VII,(3.3)]: (2.8) F [f (ax + b)](ξ) = a−1 According to [64, IV,(3.1)], p χp(cid:16) − b a a(cid:17), ξ(cid:17)F [f (x)](cid:16) ξ x, ξ ∈ Qp. F [Ω(p−k · p)](x) = pkΩ(pkxp), (2.9) where Ω(t) is the characteristic function of the segment [0, 1] ⊂ R. In particular, (2.10) x ∈ Qp, k ∈ Z, F [Ω(ξp)](x) = Ω(xp). 2.3. Adeles. We use the notation and results from [24, Ch. III,§1,2] and [17]. Definition 2.1. The adeles of Q are and the ideles of Q are A = AQ =(cid:8)(xp)p∈VQ ∈ Yp∈VQ J = JQ =(cid:8)(xp)p∈VQ ∈ Yp∈VQ Qp : xp ∈ Zp for almost all p 6= ∞(cid:9), p for almost all p 6= ∞(cid:9), p : xp ∈ Z× Q× where VQ = {∞, 2, 3, 5, . . .} is the set of indices. The adele ring A is the restricted direct product of R and Qp for all p = 2, 3, . . . with respect to the integer rings Zp (see [17]). If componentwise operations of addition and multiplication are introduced, A is a ring of adeles and J is a multiplicative group. The additive group of the ring A is called the adelic group. There is a natural imbedding Q 7→ A given by Q ∋ r 7→ (r, r, . . . , r, . . . ) ∈ A. WAVELET ANALYSIS ON ADELES AND PSEUDO-DIFFERENTIAL OPERATORS 7 Indeed, any constant sequence (r, r, . . . , r, . . . ) is an adele since r ∈ Zp for any p not dividing the denominator of r. The adeles (ideles) of the form (r, r, . . . , r, . . . ), where r ∈ Q, are called principal adeles (respectively, ideles). To define a topology on A, we show that A is a union of locally compact groups. Note Qp is not a locally compact group. that the direct productQp∈VQ Definition 2.2. Let S be a finite subset of VQ such that ∞ ∈ S ⊂ VQ. Define the S-adeles of Q as AS = AQ,S =Yp∈S Qp ×Yp6∈S Zp. For an arbitrary S, the space AS of S-adeles is locally compact (in the Tikhonov product topology) as the product of a finite product of locally compact spaces Qp, p ∈ S by an infinite product of compact spaces Zp, p 6∈ S. Since we have A =SS AS [17, Theorem 2.15], the adele group A is locally compact (i.e. any neighborhood of a point contains a compact neighborhood of this point). A fundamental system of open neighborhoods of zero in A is the following set (see [17, Theorem 2.17]) (2.11) Yp∈S Up ×Yp6∈S Zp, where S is a finite subset of VQ that contains ∞ and Up is an open set in Qp containing 0 ∈ Qp for p ∈ S. infinite sequences The non-Archimedean part eA of the adele ring A is defined (see [31]) as the set of x′ = (x2, . . . xp, . . . ), where xp ∈ Qp, p = 2, 3, . . . , and there exists a prime number P = P (x′) such that xp ∈ Zp for p ≥ P . p → xp, n → ∞, for any p ∈ VQ, where VQ is defined by (2.2); The sequence of adeles {x(n), n ∈ N} converges to the adele x (x(n) → x, n → ∞) if (1) x(n) (2) there exists N such that x(n) Since the adele group A is a locally compact commutative group, it possesses the Haar measure which will be denoted by dx, where x = (x∞, x2, . . . , xp, . . . ). The Haar measure dx can be expressed in terms of the measures dxp on the groups Qp as follows: p − xp ∈ Zp for all n ≥ N. (2.12) where dx = dx∞ dx2 · · · dxp · · · , Z 1 0 ZZp dx∞ = 1, dxp = 1, p ∈ VQ. Here formula (2.12) is understood in the following sense: if f (x) = f∞(x∞)f2(x2)· · · fp(xp) is a cylindrical function, then ZA f (x) dx =ZR f∞(x∞) dx∞ZQ2 f2(x2) dx2 · · ·ZQp fp(xp) dxp. Any additive character χ(x) on the adelic ring has the form χ(x) = χ0(ax), x = (x∞, x2, . . . xp, . . . ) ∈ A, 8 A. YU. KHRENNIKOV, A. V. KOSYAK, AND V. M. SHELKOVICH for some a = (a∞, a2, . . . ap, . . . ) ∈ A (see [24, Ch. III,§1.5,(3)]). Here (2.13) χp(ap), a ∈ A, χ0(a) = Yp∈VQ where χ∞(a∞) = e2πia∞, and χp(ap) is defined by (2.6), p = 2, 3, . . . . It is clear that for any a ∈ A there exists a prime number P = P (a) such that χp(ap) = 1 for p ≥ P , i.e., in fact, the product (2.13) is finite. In other words, χ0(a) = exp(cid:0)2πiσ(a)(cid:1), where σ(a) =Pp∈VQ Definition 2.3. (see [24, Ch. III,§2.1]) Let S(R) be the real Schwartz space of tempered test functions. The space of Bruhat -- Schwartz adelic test functions S(A) consists of finite linear combinations of elementary functions of the form ap (mod 1). (2.14) ϕ(x) = Yp∈VQ ϕp(xp), x ∈ A, where the factors ϕp(xp) are such that: (i) ϕ∞(x∞) ∈ S(R); (ii) ϕp(xp) ∈ D(Qp), p = 2, 3, . . . ; (iii) there exists P = P (ϕ) such that ϕp(xp) = Ω(xpp) for all p > P (the number P (ϕ) is called the parameter of finiteness of an elementary function ϕ). In view of condition (iii) the space of test functions S(A) admits a natural represen- tation (2.15) S(A) = lim m∈VQ\∞ indS [m](A), where S [m](A) is the subspace of all test functions with the parameter of finiteness m, m ∈ VQ \ ∞. The representation (2.15) equips the space S(A) with the inductive limit topology. The Bruhat -- Schwartz adelic test functions are continuous on A. The Bruhat -- Schwartz space of adelic distributions S ′(A) was studied in [31], [22]. The spaces of test functions and distributions connected with the non-Archimedean part of adeles eA we denote by S(eA) and S ′(eA), correspondingly (see [31]). The Fourier transform of ϕ ∈ S(A) is defined by the formula χ0(ξx)ϕ(x) dx, ξ ∈ A, where χ0 is defined by (2.13). It is clear that F [S(A)] = S(A). If f ∈ S ′(A), then def = ZA bϕ(ξ) = F [ϕ](ξ) hF [f ], ϕi def = hf, F [ϕ]i, ∀ ϕ ∈ S(A), and F [f ] ∈ S ′(A). For f, g ∈ L2(A) we have (2.16) where (f, g) = (F [f ], F [g]), f2 = F [f ]2, (f, g) =ZA f (x)g(x) dx is the scalar product of the functions f and g in L2(A). It is clear that the space S(A) is dense in L2(A). WAVELET ANALYSIS ON ADELES AND PSEUDO-DIFFERENTIAL OPERATORS 9 3. Real and p-adic MRAs and wavelet bases 3.1. Real MRA.. Now we recall the definitions of the real and p-adic multiresolution analysis. Definition 3.1. (for example, see [53, 1.2]) A collection of closed spaces Vj ⊂ L2(R), j ∈ Z, is called a multiresolution analysis (MRA) in L2(R) if the following axioms hold (a) Vj ⊂ Vj+1 for all j ∈ Z; Vj is dense in L2(R); Vj = {0}; (b) Sj∈Z (c) Tj∈Z (3.4) (3.5) where (3.6) (d) f (·) ∈ Vj ⇐⇒ f (2·) ∈ Vj+1 for all j ∈ Z; (e) there exists a function φ ∈ V0 such that the system {φ(· − n), n ∈ Z} is an orthonormal basis for V0. The function φ from axiom (e) is called refinable or scaling. One also says that a MRA is generated by its scaling function. The function φ is a solution of a special kind of functional equations which are called refinement equations. Their solutions are called refinable functions. According to the standard MRA-scheme (see, e.g., [53, §1.3]) for each j, we define (wavelet spaces) (3.1) It is easy to see that Wj = Vj+1 ⊖ Vj, j ∈ Z. f ∈ Wj ⇐⇒ f (2·) ∈ Wj+1, for all j ∈ Z and Wj ⊥ Wk, j 6= k. Taking into account axioms (b) and (c), we obtain (3.2) L2(R) =Mj∈Z Wj = V0 ⊕(cid:16)Mj∈Z+ Wj(cid:17), where Z+ = {0} ∪ N. It is well known that the Haar refinement equation has the following form (3.3) φH(t) = φH(2t) + φH(2t − 1), t ∈ R. Its solution (the characteristic function χ[0,1](t) of the unit interval [0, 1]) φH(t) = χ[0,1](t) =(cid:26) 1, 0, t ∈ [0, 1], t /∈ [0, 1], , t ∈ R, generates the Haar MRA. In the framework of the Haar MRA one can construct the well-known Haar wavelet basis in L2(R) ψH jn(t) = 2j/2ψH(cid:0)2jt − n(cid:1), ψH(t) = 0 ≤ t < 1 2, 2 ≤ t < 1, t 6∈ [0, 1), 1, −1, 0, 1 t ∈ R, j ∈ Z, n ∈ Z, = χ[0, 1 2 )(t) − χ[ 1 2 ,1)(t), t ∈ R, is called the Haar wavelet function (whose dyadic shifts and delations form the Haar basis (3.5)). 10 A. YU. KHRENNIKOV, A. V. KOSYAK, AND V. M. SHELKOVICH Using the second decomposition (3.2), instead of the Haar wavelet basis (3.5) in L2(R), (3.7) one can consider the following wavelet basis φH(t − n) ∈ V0, jn(t) = 2j/2ψH(cid:0)2jt − n(cid:1) ∈ Wj, ψH n ∈ Z, j ∈ Z+, n ∈ Z, t ∈ R, where the wavelet function ψH is given by (3.6). Basis (3.7) will be called modified Haar basis. 3.2. p-adic MRA.. In the p-adic case a "natural" set of shifts for Qp is the following (see [39] and [57]): (3.8) Ip = {a ∈ Qp : {a}p = a} = {a = a−γ pγ + · · · + a−1 p : aj ∈ Fp, −γ ≤ j ≤ −1}. In [57], similarly to Definition 3.1, the following definition was introduced. Definition 3.2. ([57]) A collection of closed spaces Vj ⊂ L2(Qp), j ∈ Z, is called a multiresolution analysis (MRA) in L2(Qp) if the following axioms hold (a) Vj ⊂ Vj+1 for all j ∈ Z; Vj is dense in L2(Qp); Vj = {0}; (b) Sj∈Z (c) Tj∈Z (d) f (·) ∈ Vj ⇐⇒ f (p−1·) ∈ Vj+1 for all j ∈ Z; (e) there exists a function φ ∈ V0 such that the system {φ(· − a), a ∈ Ip} is an orthonormal basis for V0. It follows immediately from axioms (d) and (e) that the functions pj/2φ(p−j · −a), a ∈ Ip, form an orthonormal basis for Vj, j ∈ Z. According to the standard scheme (see, e.g., [53, §1.3]) for the construction of MRA- based wavelets, for each j, we define a space Wj (wavelet space) as the orthogonal complement of Vj in Vj+1, i.e., (3.9) It is not difficult to see that Vj+1 = Vj ⊕ Wj, j ∈ Z. f ∈ Wj ⇐⇒ f (p−1·) ∈ Wj+1, (3.10) and Wj ⊥ Wk, j 6= k. Taking into account axioms (b) and (c), we obtain (3.11) j ∈ Z for all In view of (3.9) and axiom (a), we have (3.12) L2(Qp) =Mj∈Z Vj = V0 ⊕(cid:16) M0≤k≤j−1 Wj = V0 ⊕(cid:16)Mj∈Z+ Wj(cid:17). Wk(cid:17) j ∈ N. If we now find a finite number of functions ψν ∈ W0, ν ∈ A, such that the system {ψν(x − a) : a ∈ Ip, ν ∈ A} forms an orthonormal basis for W0, then, due to (3.10), (3.11), the system {pj/2ψν(p−j · −a) : a ∈ Ip, j ∈ Z, ν ∈ A}, is an orthonormal basis for L2(Qp). Such functions ψν, ν ∈ A, are called wavelet functions and the corresponding basis is called a wavelet basis. WAVELET ANALYSIS ON ADELES AND PSEUDO-DIFFERENTIAL OPERATORS 11 In [32], the following conjecture was proposed: to construct a p-adic analog of the real Haar MRA, we can use the following refinement equation (3.13) φ(x) = p−1Xr=0 φ(cid:16)1 p x − r p(cid:17), x ∈ Qp, whose solution (a refinable function) φ is the characteristic function Ω(cid:0)xp(cid:1) of the unit disc. The above refinement equation is natural and reflects the fact that the unit disc B0 = {x : xp ≤ 1} is the union of p mutually disjoint discs of radius p−1: B0 = p−1[r=0 B−1(r), where B−1(r) =nx :(cid:12)(cid:12)x − r(cid:12)(cid:12)p ≤ 1 po, r ∈ Fp. This geometric fact is the result of the ultrametric structure of the p-adic field Qp. The term p-adic Haar MRA is connected with the fact that for p = 2 the equation (3.13) is a 2-adic refinement equation φ(x) = φ(cid:16)1 2 x(cid:17) + φ(cid:16)1 2 x − 1 2(cid:17), x ∈ Q2, which is a direct analog of the refinement equation (3.3) generating the Haar MRA and the Haar wavelet basis (3.5), (3.6) in the real case. In [57], using the above relation (3.13) as a refinement equation, the Haar MRA was constructed for p = 2 (for the case p 6= 2, see [3, 8.4]). Here (3.14) Vj = span(cid:8)pj/2φ(cid:0)p−j · −a(cid:1) : a ∈ Ip(cid:9), where φ(x) = Ω(cid:0)xp(cid:1). In contrast to the Haar MRA in L2(R) (which generates only one wavelet basis (3.5), (3.6)), in the p-adic setting there exist infinity many different Haar orthogonal bases for L2(Qp) generated by the same MRA. Explicit formulas for generating p-adic wavelet functions were obtained in [57] for p = 2 and later in [35] for p 6= 2. Let us recall these results. Theorem 3.1. ( [35]) The set of all compactly supported wavelet functions is given by j ∈ Z, (3.15) ψµ(x) = (here supp ψµ ⊂ Bs(0), s ≥ 0), where (3.16) ψ(0) αµ ν;kψ(0) ps−1Xk=0 p−1Xν=1 ν (x) = χp(cid:16) ν p ν (cid:16)x − x(cid:17)Ω(cid:0)xp(cid:1), are Kozyrev's wavelet functions, s = 0, 1, 2, . . . , and k ps(cid:17), µ ∈ F× p = {1, 2, . . . , p − 1}, ν ∈ F× p , x ∈ Qp, αµ ν;k = (3.17) = m=0 e−2πi −p−sPps−1 −p−2sPps−1 m=0Pps−1 − ν p +m ps kσµmuµµ, n=0 e−2πi − ν p +m ps k 1−e 2πi 1−e 2πi µ−ν p µ−ν p +m−n ps σνmuνµ, µ = ν, µ 6= ν, σµm = 1, uµν are entries of an arbitrary unitary (p− 1)× (p− 1) matrix U; µ, ν ∈ F× p ; k = 0, 1, . . . , ps − 1. 12 A. YU. KHRENNIKOV, A. V. KOSYAK, AND V. M. SHELKOVICH Corollary 3.1. ( [57]) Let p = 2. For every s = 0, 1, 2, . . . the function (3.18) ψ(x) = 2s−1Xk=0 αkψ(0)(cid:16)x − k 2s(cid:17), is a compactly supported wavelet function (supp ψ ⊂ Bs(0)) for the Haar MRA if and only if (3.19) αk = 2−s γre−iπ 2r−1 2s k, 2s−1Xr=0 where the wavelet function ψ(0) := ψ(0) 1 such that γr = 1; k, r = 0, 1, . . . , 2s − 1. is given by (3.16), γr ∈ C is an arbitrary constant According to the general wavelet theory, all dilations and shifts of wavelet functions (3.18), (3.19) or (3.15), (3.17) form a p-adic orthonormal Haar wavelet basis in L2(Qp): (3.20) ψk;ja(x) = pj/2ψk(p−jx − a), k ∈ F× p , j ∈ Z, a ∈ Ip. In particular, Kozyrev's wavelet basis in L2(Qp) generated by wavelet functions (3.16) is the following ψ(0) k;ja(x) = pj/2ψ(0) k (p−jx − a) The wavelet functions (3.16) can be expressed in terms of the refinable function φ(x) = (3.21) Ω(cid:0)xp(cid:1): = pj/2χp(cid:16) k p ψ(0) k (x) = (p−jx − a)(cid:17)Ω(cid:0)p−jx − ap(cid:1), p−1Xr=0 p φ(cid:16)1 p(cid:17), e2πi kr x − p r k ∈ F× p , x ∈ Qp. k ∈ F× p , j ∈ Z, a ∈ Ip. It was proved in [1] that there are no orthogonal MRA based wavelet bases except for those described in Theorem 3.1 and Corollary 3.1. To construct adelic wavelet bases, we will need the p-adic wavelet bases that contain the function φ(x). Therefore, taking into account the second decomposition (3.11), instead of the p-adic wavelet basis (3.20) in L2(Qp), we consider the following wavelet basis (3.22) φ(x − a) ∈ V0, ψk;ja(x) = pj/2ψk(p−jx − a) ∈ Wj, a ∈ Ip, k ∈ F× p , j ∈ Z+, a ∈ Ip, p , are given by (3.15) -- (3.17). In particular, we where the wavelet functions ψk, k ∈ F× have the following wavelet basis (3.23) ψ(0) k;ja(x) = pj/2χp(cid:16)k p φ(x − a) ∈ V0, (p−jx − a)(cid:17)Ω(cid:0)p−jx − ap(cid:1) ∈ Wj, k ∈ F× The fact that bases (3.22) and (3.23) contain the refinable function φ(x) = Ω(cid:0)xp(cid:1) plays a crucial role in our construction (see below Sec. 5). Bases (3.22), (3.23) will be called modified Haar type bases. Taking into account the second decomposition (3.11), we obtain the following state- ment. a ∈ Ip, p , j ∈ Z+, a ∈ Ip. WAVELET ANALYSIS ON ADELES AND PSEUDO-DIFFERENTIAL OPERATORS 13 p): (3.24) and I j p where dx0 Proposition 3.1. The restriction of basis (3.23) on Zp constitutes an orthonormal basis in L2(Zp, dx0 , k ∈ F× p , a ∈ I j p, j ∈ Z+, p = {a ∈ Ip : pja ∈ Zp}. Proof. It is easy to see that the restriction of Kozyrev's wavelets (3.21) is the following φ(x) = Ω(cid:0)xp(cid:1), k;ja(x) = pj/2ψk(p−jx − a)Ω(cid:0)p−jx − ap(cid:1)(cid:12)(cid:12)Zp eψ(0) p = dxp(cid:12)(cid:12)Zp (p−jx − a)(cid:17)Ω(cid:0)p−jx − ap(cid:1)(cid:12)(cid:12)(cid:12)Zp k;ja(x) = pj/2χp(cid:16)k eψ(0) p (p−jx − a)(cid:1)Ω(cid:0)p−jx − ap(cid:1), pj/2χp(cid:0) k = k;ja(x)(cid:12)(cid:12)Zp 6= 0 only if p−jx−ap ≤ 1. Let a = a−γp−γ +· · ·+a−1p−1 For j ∈ Z+, we have eψ(0) and x =P∞ p−jx − a =(cid:0)x0p−j + · · · + xj−1p−1(cid:1) −(cid:0)a−γp−γ + · · · + a−1p−1(cid:1) + k ∈ Fp, j ∈ Z+, a ∈ I j p, j ≤ −1, a = 0, j ≤ −1, a 6= 0. k=0 xkpk. Then we have Since p−jx − a ∈ Zp, the latter relation implies that x0p−j + · · · + xj−1p−1 = a−γp−γ + · · · + a−1p−1. Hence pja ∈ Zp. pj/2φ(x), 0, ∞Xk=j xkpk−j. Let us calculate the scalar product p k′ p (p−j ′ χp(cid:16)k (p−jx − a)(cid:17)χp(cid:16) − x − a′p(cid:1) dx, k′;j ′a′(cid:1) (cid:0)eψ(0) k;ja,eψ(0) = p(j+j ′)/2ZZp x − a)(cid:17)× ×Ω(cid:0)p−jx − ap(cid:1)Ω(cid:0)p−j ′ k, k′ ∈ Fp, j, j′ ∈ Z+, a, a′ ∈ Ip. Here (cid:0)eψ(0) k′;j ′a′(cid:1) 6= 0 only if p−jx − ap ≤ 1 and p−j ′x − a′p ≤ 1. Just as above, k;ja,eψ(0) p , we have supp Ω(cid:0)p−jx − ap(cid:1), supp Ω(cid:0)p−j ′x − a′p(cid:1) ∈ Zp. Now, taking into k′;j ′a′(cid:1) = δkk′δjj ′δaa′, where δss′ is the Kronecker symbol. (cid:0)eψ(0) k;ja,eψ(0) we conclude that a, a′ ∈ Ip are such that pja ∈ Zp, pj ′a′ ∈ Zp. Thus for a ∈ I j p, a′ ∈ I j ′ account orthonormality of wavelet functions ψ(0) Using Proposition 3.1 and Theorem 3.1, one can prove the following statement. k′;j ′a′ in Qp, one can conclude that (cid:3) k;ja, ψ(0) Proposition 3.2. The restriction of the basis (3.22) to Zp constitutes in L2(Zp, dx0 p) the orthonormal basis (3.25) φ(x) = Ω(cid:0)xp(cid:1), eψk;ja(x) = pj/2ψk(p−jx − a)(cid:12)(cid:12)Zp p , are given by (3.15) -- (3.17), and the set I (j) where the wavelet functions ψk, k ∈ F× defined in Remark 3.1. Remark 3.1. We note that the set I (j) a ∈ I (j) p are nonzero, is finite and its description is similar to the description of the set I j p, though more complicated. For other wavelet bases, we will have a similar situation. p , for which the restricted functions eψk;ja(x), is p , k ∈ F× p , a ∈ I (j) p , j ∈ Z+, 14 A. YU. KHRENNIKOV, A. V. KOSYAK, AND V. M. SHELKOVICH Consider the orthogonal projection P [0] : L2(Qp, dxp) 7→ L2(Zp, dx0 p), L2(Qp, dxp) ∋ f (x) 7→ (P [0]f )(x) = f (x)Ω(cid:0)xp(cid:1) ∈ L2(Zp, dx0 In our case, the projections of modified Haar bases (3.22) and (3.23) to the subspace L2(Zp, dx0 p is the restriction of the Haar measure dxp on Qp to Zp. We note that the projections of some basis elements become zeros. p) are also bases in L2(Zp, dx0 p), where dx0 p). Proposition 3.3. The restriction of the p-adic Haar MRA (given by formula (3.14)) to the space L2(Zp) consists of the spaces (3.28) where Vj(cid:12)(cid:12)Zp eVj = Vj(cid:12)(cid:12)Zp = span(cid:8)pj/2φ(cid:0)p−j · −a(cid:1), x ∈ Zp : a ∈ Ip(cid:9), = eV0 = span(cid:8)φ(·)(cid:9) for all j ≤ −1, φ(x) = Ω(cid:0)xp(cid:1), and j ∈ Z+, Here according to (3.9) -- (3.11), (3.26) given by (3.27) (3.29) where eV0 ⊂ eV1 ⊂ eV2 ⊂ · · · . L2(Zp) = eV0M(cid:16)Mj∈Z+fWj(cid:17), p , a ∈ Ip(cid:9), : k ∈ F× fWj = eVj+1 ⊖eVj = span(cid:8)pj/2ψk(p−jx − a)(cid:12)(cid:12)Zp and the wavelet functions ψk, k ∈ F× use the wavelet functions (3.23). p , are given by (3.15) -- (3.17). In particular, we can j ∈ Z+, 4. Basis in the space L2(A, dx) 4.1. Infinite tensor product of Hilbert spaces. We recall [52] (see also [12, Ch.1, Hn of Hilbert spaces Hn, n ∈ N. §2.3]) the definition of infinite tensor product He = Ne,n∈N Fix the sequence (4.1) e = (e(n))∞ n=1, e(n) ∈ Hn, ke(n)kHn = 1, called a stabilizing sequence. Denote by E the set of all stabilizing sequences. Fix an orthonormal basis (o.n.b) (e(n) 1 , n ∈ N. Let Λ be the set of multi-indices α = (αn)n∈N, αn ∈ N, n ∈ N such that αn = 1 for sufficiently big n depending on α. By definition, the o.n.b. of the space ⊗e,n∈NHn with the stabilizing sequence e consists k )k∈N in the Hilbert space Hn such that e(n) = e(n) of all vectors (eα)α∈Λ of the following form (4.2) eα = e(1) α1 ⊗ e(2) α2 ⊗ · · · ⊗ e(n) αn ⊗ e(n+1) ⊗ · · · , α ∈ Λ where αn = 1 for sufficiently big n depending on α. An arbitrary element f of the space He has the following form: (4.3) fαeα, with kfk2 f2 α < ∞ f =Xα∈Λ He =Xα∈Λ WAVELET ANALYSIS ON ADELES AND PSEUDO-DIFFERENTIAL OPERATORS 15 and the scalar product (f, g)He of two vectors f, g ∈ He has the following form: (4.4) fαgα. (f, g)He =Xα∈Λ It will often be convenient for us (see [12, Ch.1,§ 2.3]) to represent the set Λ as the union of disjoint sets, each consisting of "finite" sequences. Namely, for an α ∈ Λ let ν(α) denote the minimal m = 1, 2, . . . such that αm+1 = αm+2 = · · · = 1. Let (4.5) Λn = {α ∈ Λ : ν(α) = n}, n ∈ N. Obviously, ΛnT Λm = ∅ (n 6= m) and Λ =S∞ n=1 Λn. Example 4.1. ([12, Ch.1, § 2.3, example 1]) Let L2(Xk, µk), k ∈ N be the space of square integrable complex functions on the measurable space Xk with a probability measure µk. Choose the stabilizing sequence e = (e(k))∞ k=1, where e(k)(x) ≡ 1, x ∈ Xk, k ∈ N. In this case we have Theorem 4.1. The following two spaces are isomorphic: (4.6) Oe, k∈N L2(Xk, µk) ∼= L2(cid:16)Yk∈N Xk,⊗k∈Nµk(cid:17). Example 4.2. If in the previous example the first m measures µk are not necessarily probability, i.e., µk(Xk) = ∞, then we get following statement. Theorem 4.2. The following two spaces are isomorphic: (4.7) Xk,⊗m L2(cid:16) mYk=1 k=1µk(cid:17) ⊗(cid:16) ∞Oe, k=m+1 L2(Xk, µk)(cid:17) ∼= L2(cid:16)Yk∈N Xk,⊗k∈Nµk(cid:17). Remark 4.1. Let us consider in the infinite tensor product space He = Ne,n∈N Hn the infinite tensor product A = ⊗n∈NAn of bounded operators An acting in the space Hn, n ∈ N. By definition, an operator A acts on the total set of well-defined vectors f = ⊗n∈Nfn in the space He (see Lemma 4.1) in the following way: Af := ⊗n∈NAnfn = A1f1 ⊗ · · · ⊗ Amfm ⊗ · · · , (4.8) if the latter expression is well-defined in He. In the sequel we shall use the projections P [m] : He → H (m), where the subspaces H [m] ⊂ He are defined on the total family of vectors f = ⊗n∈Nfn as follows: He ∋ f = ⊗n∈Nfn 7→ P [m]f := ⊗m n=1fn ⊗ e(m+1) ⊗ e(m+2) ⊗ · · · ∈ H [m], and H [m] = Hn ⊗ e(m+1) ⊗ e(m+2) ⊗ · · · mOn=1 We shall denote by A(m) the projections P [m]AP [m] of the operator A in He onto the space H [m], i.e. define A(m) := P [m]AP [m]. It is clear that A(m) = mOn=1 An ⊗On>m Idn, where Idn is the identity operator in Hn. We have the strong convergence A(m) → A as m → ∞ on a suitable set of vectors f = ⊗n∈Nfn. It is sufficient to estimate k(A − A(m))fkHe. For details, see [12, Ch. I, §2.7.]. 16 A. YU. KHRENNIKOV, A. V. KOSYAK, AND V. M. SHELKOVICH As usual, we denote by AX the restriction of an operator A acting in a Hilbert space H to the invariant subspace X ⊂ H. 4.2. Complete von Neumann product of infinitely many Hilbert spaces. The complete von Neumann tensor product H = ⊗k∈NHk of Hilbert spaces Hk, k ∈ N is by definition the orthogonal sum of the spaces He ([52], see also [12, Ch.1,§ 2.10]) (4.9) H = Me∈E/∼ To be more precise, fix the space He = Ne,n∈N He over all possible equivalence classes E/∼ of stabilizing sequences e. Hn. We define the vector f = ⊗n∈Nf (n), where f (n) ∈ Hn, as the week limit (if it exists) in He of the vectors f [m] = f (1) ⊗ · · · ⊗ f (m) ⊗ e(m+1) ⊗ e(m+2) ⊗ · · · (4.10) as m → ∞. Since the set span {eα : α ∈ Λ} is dense in He, the week limit of the vectors f [m] exists if and only if: 1) the norms kf [m]kHe are uniformly bounded with respect to m = 1, 2, . . . , and 2) limm→∞(f [m], eα)He exists for each α ∈ Λ. The following statements are proved in [12, Ch.I,§ 2.10]. Lemma 4.1. The strong limit of vectors (4.10) exists in He, as m → ∞, if and only n=q(f (n), e(n))Hn (q = 1, 2, . . . ) converge to finite if the product Q∞ numbers, and we haveQ∞ limm→∞ f [m] exists if and only if for some q = 1, 2, . . . the product Q∞ Corollary 4.1. If f (n) in (4.10) are taken to be unit vectors, then the strong limit n=q(f (n), e(n))Hn converges to a finite nonzero number. Definition 4.1. (see [12, Ch. I,§ 2.10, Theorem 2.9]) Consider the set E of all stabi- lizing sequences e = (e(n))∞ n=1 of the form (4.1). A stabilizing sequence l ∈ E is said to be equivalent to a stabilizing sequence e ∈ E (l ∼ e), if each strong limit n=1 kf (n)kHn = 0 whenQ∞ n=1 kf (n)kHn and Q∞ n=q(f (n), e(n))Hn = 0 for each q. ⊗ l(2)′ ⊗ · · · = lim ⊗ e(m+2) ⊗ e(m+2) ⊗ · · · , exists in He, where (l(n)′)∞ k=1, "diluted" by the vectors e(n), i.e. each l(n)′ is equal either to l(n) or to e(n). The relation ∼ is an equivalence relation and we denote by E/∼ the set of all equivalent classes of E. ⊗ · · · ⊗ l(m)′ k=1 is the sequence (l(n))∞ l(1)′ l(1)′ m→∞ Theorem 4.3. Two infinite tensor products He = Ne,n∈N Hn and Hl = Nl,n∈N responding to two equivalent stabilizing sequences e = (e(n))∞ isomorphic. For e 6∼ l, the spaces He and Hl are orthogonal. 4.3. On some subspaces of the infinite tensor products. Let Xn be some sub- spaces in the Hilbert spaces Hn, n ∈ N and let (e(n) k )k∈Z be an orthonormal basis in Hn such that the orthonormal basis in Xn is (e(n) k )k∈N, and let (e(n))n∈N be the stabilizing sequence e(n) = e(n) n=1 and l = (l(n))∞ n=1 are Hn cor- (4.11) 1 , n ∈ N. Note that e(n) = e(n) Consider two spaces He and Hl 1 ∈ Xn ⊂ Hn, n ∈ N. e(X), l ∈ N, where Hn = H1 ⊗ H2 ⊗ · · · ⊗ Hl ⊗ Hl+1 ⊗ Hl+2 ⊗ · · · , He = Oe,n∈N WAVELET ANALYSIS ON ADELES AND PSEUDO-DIFFERENTIAL OPERATORS 17 (4.12) Hl e(X) = lOk=1 Hk ⊗ ∞Oe,k=l+1 Xk = H1 ⊗ H2 ⊗ · · · ⊗ Hl ⊗ Xl+1 ⊗ · · · . In the particular case Xn = Ce(n) we have Hl e(X) = H1 ⊗ H2 ⊗ · · · ⊗ Hl ⊗ Ce(l+1) ⊗ Ce(l+2) ⊗ · · · , l ∈ N. Lemma 4.2. For arbitrary subspaces Xn in Hn and a stabilizing sequence (e(n))n∈N with properties (4.11), we have Proof. The basis in the space He is the following (see (4.2)): (4.14) eα = e(1) α1 ⊗ e(2) α2 ⊗ · · · ⊗ e(n) where Λn is defined by (4.5), and the basis in the space Hl el α = e(1) α1 ⊗ e(2) α2 ⊗ · · · ⊗ e(l) αl ⊗ e(l+1) Hl e(X). He =[l≥1 αn ⊗ e(n+1) ⊗ · · · , α ∈ Λ = [n∈N αl+k ⊗ e(l+k+1) ⊗ · · · , αl+1 ⊗ · · · ⊗ e(l+k) α ∈ Λl = [k∈N∪{0} e(X) is Λn, Λl,k, (4.13) (4.15) where (4.16) Λl,k = {α ∈ Λ : αi ∈ Z, 1 ≤ i ≤ l, αi ∈ N, l + 1 ≤ i ≤ l + k, αl+k+i = 1, i > 1}. Obviously He ⊇ Hl e(X) for all l ∈ N, hence He ⊇Sl≥1 Hl We show that He ⊆ Sl≥1 Hl form (4.14) are contained in the family of vectors el take in (4.15) l = n and k = 0, we obtain all vectors eα of the form (4.14). e(X). It is sufficient to show that all vectors eα of the α of the form (4.15). Indeed, if we (cid:3) e(X). 4.4. Infinite tensor product of Hilbert spaces L2(Qp, dxp). Using the results of the previous section we can define the infinite tensor product He(Q) = Oe, p∈VQ L2(Qp, dxp) of Hilbert spaces L2(Qp, dxp) with an arbitrary stabilizing sequence e = (e(p))p∈VQ e(p) ∈ L2(Qp, dxp). Let {e(∞) αp }αp∈Ip be arbitrary orthonormal bases in L2(Q∞, dx∞) and L2(Qp, dxp), respectively, where I∞ and Ip are the corresponding multi-indices. Then the orthonormal basis in the space He(Q) is the following: (4.17) α∞ }α∞∈I∞ and {e(p) α∞ ⊗Op≤m eα = e(∞) e(p), α = (α∞, α2, α3, . . . ) ∈ Λ, αp ⊗Om<p e(p) where (e(p))p∈VQ, e(p) ∈ L2(Qp, dxp), is some stabilizing sequence and Λ is defined below. Definition 4.2. Define Λ as the set of multi-indices α = (αp)p∈VQ, αp ∈ Ip such that e(p) αp = e(p) for sufficiently big p depending on α. 18 A. YU. KHRENNIKOV, A. V. KOSYAK, AND V. M. SHELKOVICH Fix the stabilizing sequence of the form (4.18) e = (e(p))p∈VQ Using Lemma 4.2, for the stabilizing sequence (4.18), we obtain the following descrip- e(p)(xp) = φp(xp) = Ω(cid:0)xpp(cid:1) ∈ L2(Zp, dx0 p), p ∈ VQ. tion of the space L2(A, dx). Lemma 4.3. We have (4.19) L2(A, dx) = He(Q) = Oe, p∈VQ L2(Qp, dxp). L2(Ap, µp) =Nq≤p L2(Qq) ⊗Nq>p Proof. It is sufficient to use Lemma 4.2 and set Xp = L2(Zp), Hp = L2(Qp), Hp e = p ∈ VQ \ ∞ (see (4.27)), µp is the restriction on Ap of the Haar measure dx on A. 4.5. Some remarks about the stabilizations. Consider the space (4.16) L2(Zq), p ∈ VQ, where Ap = Q∞ ×Qp′≤p Qp′ ×Qp′>p Zp′, He(Q) = Oe, p∈VQ L2(Qp, dxp), (cid:3) (4.20) where the stabilizing sequence e = (e(p))p∈VQ has the special form e(p)(xp) = φp(xp) ∈ L2(Qp, dxp). We show that the sequence e and M j e for j ∈ Z are not equivalent (see VQ Definition 4.1), where (4.21) M j VQ e := (M j p φ(p))p∈VQ and M j p φ(p)(xp) := p−j/2φ(p)(pjxp). Indeed, we have (φ(p), M j p φ(p))L2(Qp) =ZQp Set c∞ = (φ(∞), M j ∞φ(∞))L2(Qp∞), then we get ∞ (e, M j c−1 VQ (φ(p), M j e) = Yp∈VQ\∞ pj/2, φ(p)(xp)p−j/2φ(p)(pjxp)dx =(cid:26) p−j/2, p φ(p))L2(Qp) = (cid:16)Qp∈VQ\∞ p(cid:17)−j/2 (cid:16)Qp∈VQ\∞ p(cid:17)j/2 j ≥ 0, j ≤ −1. , , j ≥ 0, j ≤ −1, e) = 0 hence the sequences e Thus in both cases the product is divergent, i.e., (e, M j VQ and M j VQ e are not equivalent! αp in the space L2(Qp), i.e., e = (ψ(p) A similar argument holds if we take as the stabilizing sequence some elements of αp )p∈VQ. In this case we immediately αp )p∈VQ and Kozurev's basis ψ(p) get (ψ(p) M j VQ The above considerations force us to consider the following space αp )L2(Qp) = 0 for all j ∈ Z. Hence the sequences e = (ψ(p) αp )p∈VQ are not equivalent. αp , M j e = (M j p ψ(p) p ψ(p) to be sure that the operator M j (4.22) (4.23) (4.24) He,Z(Q) = ⊕j∈ZHM j VQ is well-defined. Further we set e(Q) VQ e(Q) VQ Wk = HM k k=−∞Wk = ⊕j Vj = ⊕j k=−∞HM k VQ e(Q). WAVELET ANALYSIS ON ADELES AND PSEUDO-DIFFERENTIAL OPERATORS 19 Theorem 4.4. The collection of closed spaces Vj ⊂ He,Z(Q), j ∈ Z, defined by (4.24) is a multiresolution analysis in He,Z(Q), i.e., the following properties hold: (a) Vj ⊂ Vj+1 for all j ∈ Z; Vj is dense in He,Z(Q); Vj = {0}; (b) Sj∈Z (c) Tj∈Z (d) f (·) ∈ Vj ⇐⇒ f (M j (e) there exists a basis (ei)i∈I in the space W0 such that (M k VQ·) ∈ Vj+1 for all j ∈ Z; VQei)i∈I is a basis in the space Wk, k ∈ Z. 4.6. Complete von Neumann product of Hilbert spaces L2(Qp, dxp). It is nat- ural to consider also the complete von Neumann product (see (4.25)) of Hilbert spaces L2(Qp, dxp), p ∈ VQ (4.25) H(Q) = Me∈E/∼ He(Q), where He(Q) is defined by (4.20). Perhaps this space could be useful in further devel- opment of analysis on the adele space A. The space He,Z(Q) defined by (4.22) certainly, contains the space L2(A) and is roughly speaking an infinite direct sum of non isomor- phic copies of the space similar to L2(A). 4.7. Basis on L2(A, dx). We can construct a basis in L2(A, dx) using description (4.19) of the space L2(A, dx) in Lemma 4.3. Without using this description we can proceed as follows. Ap (see (4.27)), then we realize L2(Ap, µp) as the infinite tensor product of Hilbert spaces (see (4.30)) and use the following considerations. We present A as the union of some subgroups Ap: A =Sp∈VQ Xn, X =Sn∈N Xn. Set µn = µXn, νn = µXn\Xn−1, n ≥ 2, L2(X, µ) = ⊕n∈NL2(Xn \ Xn−1, νn) = [n∈N Let (X, µ) be some measurable space and let X be the union of the measurable sets ν1 = µ1 and X0 = ∅. In this case we have For any prime p we denote by p− and p+ the previous and the following primes. L2(Xn, µn). (4.26) Define the subgroup Ap of the group A as follows: p be the restriction of the Haar measure dxp on Qp to the subgroup Zp, let dµp be the restriction to the subgroup Ap of the measure dx on A, and let dνp be the restriction of the measure dx to the subset Ap \ Ap−. Then we have (4.28) dµp = dx∞ ⊗(cid:16) ⊗p′≤p dxp′(cid:17) ⊗(cid:16) ⊗p′>p dx0 p′(cid:17). −, νp′)(cid:17) = [p∈VQ L2(A, dx) = L2(Ap, µp) ⊕(cid:16)Mp′>p \ Ap′ L2(Ap′ Lemma 4.4. For an arbitrary prime p we have the following description: (4.29) L2(Ap, µp). and using (4.26) we get. Ap = Q∞ ×Yp′≤p Qp′ ×Yp′>p Zp′, p ∈ VQ. (4.27) We have A =Sp∈VQ Let dx0 Ap. 20 A. YU. KHRENNIKOV, A. V. KOSYAK, AND V. M. SHELKOVICH To construct a basis in the space L2(A, dx) using the latter description, it is sufficient to construct a basis in the space L2(Ap, µp). The latter space L2(Ap, µp) is the infinite tensor product of the spaces L2(Qp, dxp) and L2(Zp, dx0 p): (4.30) L2(Ap, µp) = L2(Q∞, dx∞)⊗(cid:16)Op′≤p L2(Qp′, dxp′)(cid:17)⊗(cid:16) Oe,p′>p L2(Zp′, dx0 p′)(cid:17). Remark 4.2. Since the measures dx0 p on Zp are probability measures, decomposition (4.30) allows us to use an explicit description of the basis in the infinite tensor product ⊗k∈NL2(Xk, µk) of Hilbert spaces L2(Xk, µk) with probability measures µk, k ∈ N (for example, see [12]) in order to construct a basis in the spaces L2(Ap, µp) and L2(A, dx). Suppose that {e(∞) α∞ }α∞∈I∞, and {e(p) αp}αp∈Ip are arbitrary orthonormal bases in the spaces L2(Q∞) and L2(Qp), respectively, I∞ and Ip are the corresponding indices. Fix the stabilizing sequence (e(p))p∈VQ, where e(p) is some element of the basis {e(p) αp}αp∈Ip for all p ∈ VQ such that e(p)(x) ∈ L2(Zp, dx0 p). We can construct a basis in the space L2(Ap, µp) for all p ∈ VQ using decomposition (4.30). In such a way we can construct a basis in L2(A, dx) using (4.29). As it was mentioned before, using Lemma 4.3 and the basis in the infinite tensor product we obtain. Theorem 4.5. The vectors eα = e(∞) (4.31) α∞ ⊗Oq≤m e(q) αq ⊗Om<q e(q), α ∈ Λ = [m∈VQ Λm, form the orthonormal basis in the space L2(A, dx), where Λ is the set of multi-indices α = (αp)p∈VQ, such that e(p) αp = e(p) for sufficiently big p depending on α (see (4.2)), Λp = {α ∈ Λ : νa(α) = p}, p ∈ VQ, and νa(α) denote the minimal p ∈ VQ such that e(q) αq = e(q) for q > p. Corollary 4.2. The vectors (4.32) eeα =Oq≤m e(q) αq ⊗Om<q e(q), α ∈eΛ, form an orthonormal basis in the space L2(eA, dx), where eΛ is the set of multi-indices αp = e(p) for sufficiently big p depending on α, i.e., α = (αp)p∈VQ\{∞} such that ψ(p) eΛ = [m∈VQ\∞eΛm and eΛm = {α ∈eΛ : νa(α) = m}, m ∈ VQ \ ∞. (4.33) 5. Adelic wavelet bases generated by tensor product of one-dimensional wavelet bases To construct an adelic wavelet basis, we apply the above scheme from Subsec. 4.7 to the one-dimensional bases (3.5), (3.22), (3.25). In particular, one can use the Haar wavelet bases (3.5), (3.23), (3.24). Let j∞a∞(x∞) = ψH j∞a∞(x∞), α∞ = (j∞, a∞) ∈ I∞ := (Z, Z); α∞ (x∞) = ψ(∞) ψ(∞) (5.1) be the real Haar wavelet basis (3.5) in L2(R), (5.2) αp (xp) = ψ(p) ψ(p) αp (xp) = ψ(p) ψ(p) kp;jpap(xp), αp = (kp, jp, ap) ∈ I+ p , Z+, Ip), ap (xp) = φ(p)(xp − ap), αp = (0, 0, ap) ≡ ap ∈ Ip; p := (F× WAVELET ANALYSIS ON ADELES AND PSEUDO-DIFFERENTIAL OPERATORS 21 be the p-adic modified Haar wavelet basis (3.22) in L2(Qp), where Ip is defined by (3.8), and let (5.3) αp (xp) = ψ(p) ψ(p) αp (xp) = ψ(p) ψ(p) kp; jpap(xp), αp = (kp, jp, ap) ∈ I+ 0 (xp) = φ(p)(xp), αp = (0, 0, 0) = 0; p = (F× p , Z+, Ip), be the p-adic Haar wavelet basis (3.25) in L2(Zp), which is the restriction of the basis (5.2) to Zp. We recall that the restrictions of some basis elements (5.2) are equal to zero (see Remark 3.1 and Proposition 3.1). Here and in what follows, for a stabilizing sequence we take (4.18). Hence we have the following bases in L2(Qp) and in L2(Zp), respectively, p [ Ip; p [{0} ⊂ Ip, p (0) p (5.4) (5.5) := I+ αp }αp∈I ψ(p) αp , αp ∈ Ip = I+ ψ(p) αp , αp ∈ I(0) where {ψ(p) αp }αp∈Ip on L2(Zp). Now, using Lemma 4.3, we obtain the following orthonormal wavelet basis in the space L2(A, dx) (see (4.31)): Ψα = Ψ(bk,bj,ba)(x) = ψ(∞) is the projection of {ψ(p) α∞ ⊗Oq≤m eΨα = eΨ(bk,bj,ba)(x) = O2≤q≤m φ(q), α ∈ Λ = [m∈VQ and the orthonormal wavelet basis in the space L2(eA, dx) (see (4.32)) φ(q), α ∈eΛ, with the stabilization φ = (φ(p))p∈VQ\∞. Herebk = (kp)p,bj = (jp)p,ba = (ap)p (see (5.2)). Remark 5.1. We note that in [20, Sec.4, formula (4.1)] the basis in the space L2(A, dx) is constructed as follows αq ⊗Om<q αq ⊗Om<q (5.6) ψ(q) ψ(q) Λm. ψα,β = ψ∞ n0(x∞) Yp=2,3,... ψαp,βp(xp), where ψ∞ n0(x∞) and ψαp,βp(xp) are orthonormal eigenfunctions (of harmonic oscillators) in the real and p-adic cases, respectively, where ψαp,βp(xp) = Ω(cid:0)xpp(cid:1) for almost all p . Let us introduce the adelic set of shifts IA and dilations ZA: (5.7) IA =(cid:8)ba = (a∞, a2, . . . , ap, . . . ) : a∞ ∈ I∞ = Z, ap ∈ Ip, ZA =(cid:8)bj = (j∞, j2, . . . , jp, . . . ) : j∞ ∈ Z, jp ∈ Z, Set m(ba) = min{p : aq = 0 for all q > p} and m(bj ) = min{p : jq = 0 for all q > p}. On the space L2(A) we define the operators of shifts Tba,ba ∈ ZA, and multi-dilation bj,bj ∈ ZA which are defined as the infinite tensor product of one-dimensional operators there exists n depending onba such that ap = 0 for all p > n(cid:9). there exists n depending onbj such that jp = 0 for all p > n(cid:9). M (4.21): (5.8) (5.9) M bj = M −j∞ ∞ ⊗ O2≤q≤m M jq q ⊗Oq>m Idq, 22 A. YU. KHRENNIKOV, A. V. KOSYAK, AND V. M. SHELKOVICH where m = m(bj ) and Idq is the identity operator on L2(Qq). We suppose that the p in the spaces L2(Q∞) and L2(Qp) (on the functions f (∞) ∈ operators M j L2(Q∞) and f (p) ∈ L2(Qp)) act as follows ∞ and M j (M j ∞f (∞))(x∞) = 2−j/2f (∞)(2−jx∞), (M j p f (p))(x) = p−j/2f (p)(pjxp). Let f (x) = ⊗q∈VQfq(xq) ∈ L2(A), where f (∞) ∈ L2(Q∞), f (q) ∈ L2(Qq), q ≥ 2, and f (q) = φ(q) for almost all q > m. Then (cid:0)Tbaf(cid:1)(x) bjf(cid:1)(x) (cid:0)M (5.10) (5.11) (cid:0)M bjTbaf(cid:1)(x) def = f (x −ba) = f∞(x∞ − a∞) ⊗O2≤q fq(xq − aq), def = 2−j∞/2f (∞)(2−j∞x∞)⊗ ⊗(cid:16) O2≤q≤m q−jq/2f (q)(qjqxq)(cid:17) ⊗(cid:16)Om<q f (q)(xq)(cid:17), def = 2−j∞/2f (∞)(2−j∞x∞ − a∞)⊗ f (q)(xq)(cid:17). Now one can obtain the adelic wavelet basis functions Ψα given in (5.5) by all adelic (5.12) ⊗(cid:16) O2≤q≤m shifts and dilations of wavelet functions q−jq/2f (q)(qjqxq − aq)(cid:17) ⊗(cid:16)Om<q In the latter relation, we assume that m = m(bj ) = m(ba ). (kq, 0 0)(xq) ⊗Om<q bjTbaΨ(bk,0,0)(cid:1)(x), (0,0)(x∞) ⊗Oq≤m Ψ(bk,0,0) := ψ(∞) (5.13) Namely, ψ(q) φ(q)(xq). Ψα(x) =: Ψ(bk,bj,ba)(x) =(cid:0)M where bk = (0, k2, . . . , km, 0, 0, . . . ), kq ∈ F× q ; ba = (a∞, a2, . . . , am, 0, 0, . . . ), a∞ ∈ Z, aq ∈ Iq;bj = (j∞, j2, . . . , jm, 0, 0, . . . ), j∞ ∈ Z, jq ∈ Z+; m ∈ VQ \∞. We stress that here m(ba) = m(bj) = m(bk) = m. The wavelet systems (5.5) described above is of considerable interest and can be useful in various situations. Nevertheless, they do not possess all the advantages of one-dimensional wavelet bases, in particular, the localization property, which is of great value for applications. In the one-dimensional case, the real Haar basis functions with large indices j∞ ∈ Z and the p-adic Haar basis functions with large indices jp ∈ Z+, p = 2, 3, . . . , have small supports. Thus the support of the multidimensional basis function may be large along one or several directions and small along some other directions. To avoid these drawbacks, we shall use a different approach. Its main idea is using the tensor product of the MRAs generating these bases instead of the tensor product of available wavelet bases. WAVELET ANALYSIS ON ADELES AND PSEUDO-DIFFERENTIAL OPERATORS 23 6. Separable adelic MRA generated by tensor product of one-dimensional MRAs Using the idea of constructing separable multidimensional MRA by means of the ten- sor product of one-dimensional MRAs (suggested by Y. Meyer [50] and S. Mallat [46], [47] (see, e.g., [53, §2.1]), we construct wavelet bases for the space L2(A, dx). We start with some general facts. Let we have a collection of closed subspaces V (k) , j ∈ Z, in a Hilbert space Hk, k = 1, 2, . . . , m, having the properties (a), (b) and (c) of the Definition 3.1: (a) V (k) , j+1 for all j ∈ Z, k = 1, 2, . . . , m; (b) Hk =Sj∈Z V (k) V (k) j = {0}. We use the following notations for the finite tensor product: j ⊂ V (k) j j (6.1) H [m] = Hk, V (k) j+1 = V (k) , j ∈ Z, k = 1, 2, . . . , m, mOk=1 V [m] j+1 = j j ⊕ W (k) mOk=1 V (k) j+1 = mOk=1 (V (k) j ⊕ W (k) j ) = V [m] j ⊕ W [m] j , V [m] j = mOk=1 V (k) j , W [m] j = (i1,i2,...,im;j)∈{1,2}m\{1,1,...,1} M i2,j ⊗ · · · ⊗ Z (m) im,j, W(i1,i2,...,im;j) = Z (1) i1,j ⊗ Z (2) W(i1,i2,...,im;j), Z (k) 1,j = V (k) j , Z (k) 2,j = W (k) j , k = 1, 2, . . . , m. (c) Tj∈Z where (6.2) and (6.3) where (6.4) (6.5) For the infinite tensor product we keep the similar notations: H = He = Oe,n∈N V [m] j+1 = mOk=1 Hn, H [m] = ⊗m k=1HkO e(m+1) ⊗ e(m+2) ⊗ · · · , m ∈ N. V (k) j+1 ⊗ e(m+1) ⊗ e(m+2) ⊗ · · · = mOk=1 j ⊕ W (k) (V (k) j ) ⊗ e(m+1) ⊗ e(m+2) ⊗ · · · = V [m] j ⊕ W [m] j , mOk=1 V [m] j = W [m] j = V (k) j ⊗ e(m+1) ⊗ e(m+2) ⊗ · · · , M i2,j ⊗ · · · ⊗ Z (m) W(i1,i2,...,im;j) (i1,i2,...,im)∈{1,2}m\{1,1,...,1} W(i1,i2,...,im;j) = Z (1) i1,j ⊗ Z (2) im,j ⊗ e(m+1) ⊗ e(m+2) ⊗ · · · . Define the space Vj as an appropriate limit of the spaces V [m] j (6.6) Vj = span(V [m] j : m ∈ N). 24 A. YU. KHRENNIKOV, A. V. KOSYAK, AND V. M. SHELKOVICH j }j∈Z of the space Hn, n ∈ N satisfy the following Then the subspaces Vj, j ∈ Z of the space He = ⊗e,n∈NHn defined by (6.6) satisfy the (aH) Vj ⊂ Vj+1 for all j ∈ Z; j following properties: j = {0} in Hn. j+1 for all j ∈ Z and n ∈ N; is dense in Hn; Theorem 6.1. Let subspaces {V (n) properties: j ⊂ V (n) (a) V (n) (b)Sj∈Z V (n) (c)Tj∈Z V (n) (bH)Sj∈Z Vj is dense in He = ⊗e,n∈NHn; (cH)Tj∈Z Vj = {0}. j+1 for all n ∈ N, we conclude that V [m] : m ∈ N) ⊂ span(V [m] j+1 : m ∈ N) = Vj+1. (bH) It is clear that for any fixed m ∈ N the spaceSj∈Z V [m] mOk=1 Hk ⊗ e(m+1) ⊗ e(m+2) ⊗ · · · Proof. (aH) Since V (n) j ⊂ V (n) (6.4)), hence Vj = span(V [m] H [m] = j j j ⊂ V [m] j+1, m ∈ N (see is dense in the space hence all vectors eα, α ∈ Λm of the basis (4.14) are contained in ∪j∈ZV [m] . To finish the proof, we note that all elements of the basis eα, α ∈ Λ in He = ⊗e,n∈NHn are given by (4.14), where Λ = ∪m∈NΛm. j (cH) 1. Let us suppose that for any m ∈ N, we have (6.7) 2. In this case we get ∩j∈Z V [m] j = {0}. (6.8) ∩j∈Z Vj = ∩j∈Zspan(V [m] j 1. To prove (6.7), we shall use (c). We set W (n) : m ∈ N) ⊂ span(∩j∈ZV [m] j+1 ⊖ V (n) j = V (n) j j : m ∈ N) = {0}. for j ∈ Z and n ∈ N. Then we get (6.9) Hn = ⊕k∈ZW (n) k , and V (n) j = ⊕j−1 k=−∞W (n) k . k Let (e(n) αn )αn∈I (n) be an orthonormal basis in Wk, then by construction, (e(n) αn )αn∈I (n) is an orthonormal basis in Hn, where I (n) =Sk∈Z I (n) k . Define Λ(I) (see Definition (4.2)) as the set of multi-indices α = (αn)n∈N, αn ∈ I (n) such that e(n) αn = e(n) for sufficiently big n depending on α, here e = (e(n))n∈N is a stabilizing sequence. Using (6.9) and (6.4), we get V [m] j =(cid:16) mOk=1 V (k) j (cid:17) ⊗ e(m+1) ⊗ e(m+2) ⊗ · · · so the basis in the space V [m] j = mOk=1(cid:16) j−1Mn=−∞ W (k) n (cid:17) ⊗ e(m+1) ⊗ e(m+2) ⊗ · · · , can be chosen as eα, α ∈ Λm,j(I), where α1 ⊗ e(2) αm ⊗ e(m+1) ⊗ · · · , α2 ⊗ · · · ⊗ e(m) eα = e(1) WAVELET ANALYSIS ON ADELES AND PSEUDO-DIFFERENTIAL OPERATORS 25 Λm,j(I) =(cid:16)α ∈ Λ(I) : α = (αn)n∈N, Note thatTj∈Z Λm,j(I) = limj→−∞(cid:0)Sj−1 (compare also with Lemma 6.1 below, where Λ1,j(I) = (−∞, j − 1] ∩ Z). Indeed, for f ∈ V [m] )(cid:17). (cid:1) = ∅, this imply (6.7) (α1, . . . , αm) ∈ ( j−1[k=−∞ j−1[k=−∞ j we have I (1) k , . . . , I (m) k f = Xα∈Λm,j (I) then kfk2 = lim If f ∈Tj∈Z V [m] j 2. To prove (6.8), using (6.4) we get cα2 < ∞. k k=−∞ I (1) k=−∞ I (m) k , . . . ,Sj−1 cαeα with kfk2 = Xα∈Λm,j (I) j→−∞ Xα∈Λm,j (I) j ⊗ e(m+1) ⊗ e(m+2) ⊗ · · · , m ∈ N, cα2 =Xα∈∅ cα2 = 0. V [m] j = ⊗m V [1] j = V (1) V [2] j = V (1) V [m] j = V (1) k=1V (k) j ⊗ e(2) ⊗ · · · ⊗ e(m) ⊗ e(m+1) ⊗ · · · , j ⊗ V (2) j ⊗ · · · ⊗ e(m) ⊗ e(m+1) ⊗ · · · , j ⊗ · · · ⊗ V (m) j ⊗ V (2) j ⊗ e(m+1) ⊗ · · · . (e) = span(e(k), V (k) ) for k − 1 ∈ N, then we get Denote by V (k) j j j span(V [1] , V [3] , V [2] , V [2] j ) = V (1) j ) = V (1) , k = 2, . . . , m) = V (1) j j span(V [1] span(V [k] j Since V (n) j Vj = span(V [k] So j j (e) = span(e(n), V (n) : k ∈ N) = V (1) Vj =\j∈Z ⊂\j∈Z(cid:0)V (1) span(V [k] j ⊗ Oe,n≥2 j \j∈Z j j j j j ⊗ V (2) j ⊗ V (2) j ⊗ V (2) ) ⊂ V (n) j ⊗ Oe,n≥2 (e) ⊗ e(3) ⊗ · · · ⊗ e(m) ⊗ e(m+1) ⊗ · · · , (e) ⊗ V (3) (e) ⊗ · · · ⊗ e(m) ⊗ e(m+1) ⊗ · · · , (e) ⊗ · · · ⊗ V (m) (e) ⊗ V (3) for j ≤ 0, we conclude that j ⊗ Oe,n≥2 V (n) V (n) j j j Oe,n≥2 (e) ⊂ V (1) (e) = V (1) 0 j j (e) ⊗ e(m+1) ⊗ · · · . V (n) 0 : k ∈ N) ⊂ V (n) 0 (cid:1) =\j∈Z(cid:0)V (1) j (cid:1) ⊗ Oe,n≥2 V (n) 0 = {0}. . (cid:3) Remark 6.1. We can give another definition of the spaces Vj, j ∈ Z, namely, for fixed n ∈ N denote by Vj(n), j ∈ Z the following family of spaces : m ∈ N, m ≥ n). (6.10) Vj(n) = span(V [m] For this system of spaces, an analog of Theorem 6.1 holds. Moreover, in the proof of (cH) we have the following formula (see the end of the proof of the Theorem 6.1): j \j∈Z Vj(n) ⊂\j∈Z(cid:16) nOk=1 V (k) j (cid:17) ⊗ Oe,k>n V (k) 0 = {0}. 26 A. YU. KHRENNIKOV, A. V. KOSYAK, AND V. M. SHELKOVICH We would like to present the following elementary lemma. Lemma 6.1. Let we have the decomposition of the Hilbert space H = ⊕k∈ZWk and let k=−∞Wk, thenTj∈Z Vj = (eα)α∈Ik be an orthonormal basis in the space Wk. Set Vj := ⊕j−1 {0}. Proof. By construction, (eα)α∈J is an orthonormal basis in H, where J =Sk∈Z Jk. We have for any f ∈ H (6.11) f =Xα∈J cαeα =Xk∈Z Xα∈Jk If f ∈ Vn = ⊕n−1 k=−∞Wk for all n ∈ Z, using (6.11), we conclude that α! < ∞. c2 cαeα! and kfk2 =Xk∈Z Xα∈Jk n−1Xk=−∞ Xα∈Jk cα2! = 0. (cid:3) kfk2 = lim n→−∞ Let {V (ν) j }j∈Z be one-dimensional MRA (see Definitions 3.1, 3.2), and let φ(ν) be its refinable function, ν ∈ VQ. We introduce the spaces (see (6.4)) (6.12) V [m] V (ν) j ⊗ φ(m+) ⊗ · · · ⊂ L2(A, dx), m ∈ VQ, j = Oν∈{∞,2,3,...,m} j ∈ Z. Now we introduce in L2(A, dx) the delation operator M j, which is defined by its projection M j (m) on any subspace V [m] j (for details, see Remark 4.1): (6.13) M j (m) = M −j ∞ ⊗ O2≤q≤m M j q ⊗Oq>m Idq, m ∈ VQ \ ∞, (m)f and (m)Tba)f similarly to (5.11) and (5.12), respectively, where IA is given by (5.7). As where Idq is the identity operator in L2(Qq). For f ∈ L2(A) we define M j (M −j before, we assume that m = m(ba ). Theorem 6.2. Let {V (ν) and let φ(ν) be its refinable function, ν ∈ VQ. Let (6.14) j }j∈Z be the one-dimensional MRA (see Definitions 3.1, 3.2), Φ(x) = φ(∞)(x∞) ⊗ φ(2)(x2) ⊗ · · · ⊗ φ(p)(xp) ⊗ · · · , x ∈ A, and (6.15) Then (6.16) Vj = span(V [m] j : m ∈ VQ) = span{(cid:0)M −j (m)TbaΦ(cid:1)(·) : ba ∈ IA, m(ba) = m ∈ VQ}, Vj = Oe;ν∈VQ V (ν) j ⊂ L2(A, dx), j ∈ Z+, j ∈ Z. (where e = (φ(ν))ν∈VQ\∞ is the stabilization sequence (4.18) and the subspaces (6.15) satisfy the following properties: (a) Vj ⊂ Vj+1 for all j ∈ Z; (b) ∪j∈ZVj is dense in L2(A, dx); (c) ∩j∈ZVj = {0}; WAVELET ANALYSIS ON ADELES AND PSEUDO-DIFFERENTIAL OPERATORS 27 0 j ⇐⇒(cid:0)M −1 (m)f(cid:1)(·) ∈ V [m] j+1 for all j ∈ Z, for any m ∈ VQ; basis for V0. Proof. 1. For each ν ∈ VQ, the system of functions {φ(ν)(·− aν)}aν ∈Iν is an orthonormal basis of the space V (ν) (see axioms (e) in Definitions 3.1, 3.2). Using the construction of (d) f (·) ∈ V [m] (e) the function Φ ∈ V0 is such that the system {Φ(x−ba) : ba ∈ IA} is an orthonormal Sec. 4 and taking into account that any shiftba = (a∞, a2, . . . , ap, . . . ) ∈ IA is finite and φ(ν) ∈ V (ν) that the system {Φ(x −ba) : ba ∈ IA} is an orthonormal basis in V0 and V0 = Oe; ν∈VQ Using definition (5.12) we get for m(ba) = m. (cid:0)M −j (m)TbaΦ(cid:1)(x) = 2j/2φ(∞)(2jx∞ − a∞)⊗ 0 ⊂ L2(Qν, dxν) is a stabilization sequence (4.18) in L2(A, dx), we conclude V (ν) 0 ⊂ L2(A, dx). (6.17) νj/2φ(ν)(ν−jxν − aν)(cid:17) ⊗ φ(m+)(xm+) ⊗ · · · ∈ V [m] ⊗(cid:16) mOν=2 for anyba ∈ IA, m ∈ VQ. Since {νj/2φ(ν)(ν−jxν − aν) : aν ∈ Iν} is the basis in V (ν) j ∈ N, using (6.15) and (6.17), we conclude that j j ⊂ Vj and φ(ν) ∈ V (ν) 0 ⊂ V (ν) j for Vj = Oe; ν∈VQ (6.15), (6.13) imply that f ∈ V [m] 0 Taking into account axioms (d) in Definitions 3.1, 3.2, one can see that Definitions if and only if M −j (m)f (·) ∈ V [m] j , j ∈ N, m ∈ VQ. Thus representation (6.16) and properties (e) and (d) hold. 2. Due to [12, Ch. I, §2.2, Theorem 2.2], we immediately obtain that property (a) holds for j ∈ N. For any j ∈ Z the property (a) holds by Theorem 6.1. Here we give an independent proof. According to (6.15), V0 = span{Φ(cid:0)x −ba(cid:1) : ba ∈ IA}, where Φ(cid:0)x −ba(cid:1) = φ(∞)(x∞ − a∞) ⊗ φ(2)(x2 − a2) ⊗ · · · ⊗ φ(p)(xp − ap) ⊗ φ(p+)(xp+) ⊗ · · · . Applying the refinement equation (3.13) to any factor of the above product and taking into account that the numbers aν ν ∈ Iν for all aν ∈ Iν and r = 1, 2, . . . , ν − 1, ν = 2, 3, . . . , we conclude that V0 ⊂ V1. By Definition (6.15) of the spaces Vj and property (d), this yields property (a). 3. The property (b) holds by Theorem 6.1. Here we give an independent proof. It is clear that any element g ∈ L2(A, dx) can be approximated by a finite linear combination of the basis elements Ψα(x) = Ψ(bk;bj ba)(x) of L2(A, dx) given by (5.5). At the same time any element Ψα(x) = Ψ(bk;bj ba)(x) is a finite product of the one-dimensional basis elements (5.1) -- (5.4): ν + r ν , aν V (ν) j ⊂ L2(A, dx), j ∈ N. α∞ (x∞) = ψ(∞) ψ(∞) αq (xq) = ψ(q) ψ(q) αq (xq) = ψ(q) ψ(q) j∞a∞(x∞) ∈ L2(Q∞), (kq;jqaq)(xq) ∈ L2(Qq), (kq;jqaq)(xq) ∈ L2(Zq), αp ∈ Ip = (F× p = (F× αp ∈ I(0) α∞ ∈ (Z, Z), p , Z+, Ip)[ Ip, p , Z+, Ip)[{0}. In turn, it follows from the completeness property for each MRA {V (ν) j }j∈Z (see axioms (b) in Definitions 3.1, 3.2) that any element f∞ ∈ L2(Q∞) and fν ∈ L2(Qν) can be 28 A. YU. KHRENNIKOV, A. V. KOSYAK, AND V. M. SHELKOVICH approximated by a finite linear combination of the functions 2j/2φ(∞)(2j · −a∞), a∞ ∈ I∞, and νj/2φ(ν)(ν−j · −aν), aν ∈ Iν, respectively, for sufficiently large j (j ∈ Z+), ν ∈ {2, 3, . . . , p}. Hence any basis element Ψα(x) = Ψ(bk;bj ba)(x) ∈ L2(A, dx) can be approximated by an element of the space Vj for some sufficiently large j (j ∈ Z+). This implies that the element g can be also approximated by an element of the space Vj for some j, i.e., the collection of all the spaces Vj, j ∈ Z, is dense in L2(A, dx). 4. The property (c) holds by Theorem 6.1. Here we give an independent proof. Let us prove that the intersection of all Vj does not contain a nonzero element. Assume that there exists an element g ∈ L2(A, dx) such that g ∈ Vj for all j ∈ Z and kgk 6= 0. According to Sec. 5, the wavelet functions Ψα = Ψ(bk;bj ba) ∈ L2(A, dx) given by (5.5) form an orthonormal basis in L2(A, dx), then g can be represented as g =Pα∈Λ cαΨα ∈ Vj for all j ∈ Z, where the coefficients cα are independent of j. Here α = (αp)p∈VQ = (bk;bj ba), Consider a basis element Ψα = Ψ(bk;bj ba),bj = (j∞, j2, . . . , jm, 0, 0, . . . ). Let us take the index jα = min(0, j∞, j2, j3, . . . , jm). For any wavelet function given by the first formula in (5.2), we have ψ(p) , p = 2, 3, . . . . In view of (3.1), (3.2) we have ψ(∞) for j < j∞. Consequently, the basis element Ψα = Ψ(bk;bj ba) ⊥ Vj, j < jα, i.e., g ⊥ Ψα. In such a way we prove that g ⊥ Ψα for all indices α ∈ Λ. The fact that the system {Ψα : α ∈ Λ} is an orthonormal basis in L2(A, dx) contradicts the assumption kgk 6= 0. p , Z+, Ip)S Ip, p = 2, 3, . . . . kp;jpap ⊥ V (p) α∞ = ψ(∞) α∞ ∈ (Z, Z), αp ∈ Ip = (F× , jp ∈ Z+, and V (p) j∞ ⊥ V (∞) j ⊂ V (p) αp = ψ(p) The collection of the spaces Vj, j ∈ Z+, satisfying conditions (a) -- (e) of Theorem 6.2 will be called the adelic separable MRA in the space L2(A). The function (6.14) is a refinable function of this MRA. (cid:3) 0 0 j Next, following the standard finite-dimensional scheme (see [18, Ch. 10.1], [53, §2.1]), we define the wavelet space Wj as the orthogonal complement of Vj in Vj+1: (6.18) Wj = Vj+1 ⊖ Vj, j ∈ Z+. It follows from properties (a) and (b) of Theorem 6.2 that (cf. (3.2) and (3.11)) Since according to (3.1), (3.2), (3.9) -- (3.11), we have V (ν) Z (ν) 1,j = V (ν) obtain , setting , using (6.19), (6.20), and taking into account (6.17), we j ⊕ W (ν) j+1 = V (ν) 2,j = W (ν) and Z (ν) j j j Vj+1 = Oν∈VQ V (ν) j+1 = Oν∈VQ(cid:0)V (ν) = Vj ⊕(cid:16)Mp∈VQ j (cid:1) j ⊕ W (ν) M (i∞,i2,...,ip)∈{1,2}♯(p)+1\{1,1,...,1} W(i∞,i2,...,ip),j(cid:17), j ∈ Z+, L2(A) = V0 ⊕Mj∈Z+ Wj. Let us present some evident formulas in ⊗n k=1Hk. Namely, we have (6.19) and similarly, (6.20) (a11 ⊕ a12) ⊗ · · · ⊗ (an1 ⊕ an2) = M(i1,...,in)∈{1,2}n k=1ank) = M(i1,...,in)∈{1,...,m}n (⊕m k=1a1k) ⊗ · · · ⊗ (⊕m a1i1 ⊗ · · · ⊗ anin, a1,i1 ⊗ · · · ⊗ an,in, WAVELET ANALYSIS ON ADELES AND PSEUDO-DIFFERENTIAL OPERATORS 29 where we set ♯(p) = ♯{2, 3, . . . , p} (♯(A) is the number of elements of the set A), and (6.21) = W [p] (i∞,i2,...,ip),j = Z (∞) i∞,j ⊗ Z (2) i2,j ⊗ · · · ⊗ Z (p) ip,j ⊗ φ(p+) ⊗ · · · , W [p] bip,j bip = (i∞, i2, . . . , ip) ∈ {1, 2}♯(p)+1 \ {1, 1, . . . , 1}. Thus the space Wj given by (6.18) is a direct sum of infinite number of subspaces (6.21). Thus we have It is clear that an orthonormal basis for the space bip=(i∞,i2,...,ip)∈{1,2}♯(p)+1\{1,1,...,1} Mbip∈{1,2}♯(p)+1\{1,1,...,1} W [p] bip,j(cid:17). W [p] (i∞,i2,...,ip),0 (6.22) where L2(A) = V0 ⊕Mj∈Z+ W0 =Mp∈VQ Wj = Mj∈Z+(cid:16)Mp∈VQ M i∞,0 ⊗ Z (2) W [p] bip,0 = W [p] (i∞,i2,...,ip),0 = Z (∞) i2,0 ⊗ · · · ⊗ Z (p) is formed by the shifts (with respect toba ∈ IA) of the functions i2 ⊗ · · · ⊗ ϑ(p) := Ψbk,(i∞,i2,...,ip) = ϑ(∞) i∞ ⊗ ϑ(2) (6.23) Ψbk,bip ip ⊗ φ(p+) ⊗ · · · , ip,0 ⊗ φ(p+) ⊗ · · · , kp− kp− 1 = ψ(ν) 1 = ψH , ϑ(∞) 2 = φH (see (3.5), (3.4)); ϑ(ν) According to (6.22), an orthonormal Haar wavelet basis in L2(A) is formed by all where all factors in this product starting from a certain prime p constitute a stabilization sequence (φ(ν))ν>p; ϑ(∞) kν , kν ∈ F× ν , ϑ(ν) (i∞, i2, . . . , ip) ∈ {1, 2}♯(p)+1 \ {1, 1, . . . , 1}. 2 = φ(ν), (see (3.22), (3.25) or (3.23), (3.24)), ν = 2, 3, . . . , p; p = 2, 3, . . . ; bip = shiftsba ∈ IA of the refinable function (6.14) φ(∞)(x∞)φ(2)(x2)· · · φ(p−)(xp−)φ(p)(xp)φ(p+)(xp+)· · · , and all shiftsba ∈ IA and dilations j ∈ Z+ of wavelet functions (6.23): ψ(∞)(x∞)ψ(2) k2 (x2)· · · ψ(p−) kp (xp)φ(p+)(xp+)· · · , ψ(∞)(x∞)ψ(2) k2 (x2)· · · ψ(p−) (xp+)· · · , . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . φ(∞)(x∞)ψ(2) k2 (x2)· · · ψ(p−) (xp+)· · · , . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ψ(∞)(x∞)φ(2)(x2)· · · φ(p−)(xp−)φ(p)(xp)φ(p+)(xp+)· · · , . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . φ(∞)(x∞)φ(2)(x2)· · · φ(p−)(xp−)ψ(p) kp (xp)φ(p+)(xp+)· · · , (xp−)ψ(p) (xp−)φ(p)(xp)φ(p+,0) kp (xp)ψ(p+,0) (xp−)ψ(p) (6.24) kp+ kp+ kp− (6.25) Thus, this wavelet basis has the form where p = 2, 3, . . . ; the real Haar wavelet function ψ(∞)(x∞) and refinable function φ(∞)(x∞) = Ω(x∞) are given by (3.6) and (3.4), x∞ ∈ Q∞; the r-adic Haar wavelet functions ψ(r) kr (xr), kr ∈ F×, are given by formulas (3.15) -- (3.17), the refinable function φ(r)(xr) = Ω(cid:0)xrr(cid:1), xr ∈ Qr; r = 2, 3, . . . . wherebip = (i∞, i2, . . . , ip) ∈ {1, 2}♯(p)+1 \ {1, 1, . . . , 1}, bk = (k2, . . . , kp), ks ∈ F× 2, 3, . . . , p; j ∈ Z+;ba = (a∞, a2, . . . , ap) ∈ IA, p ∈ VQ. We stress that here p(ba) = p(j) = p(bk) = p. TbaΦ, Ψbk,bip;j ba := (2 · 2 · 3 · · · p)j/2(cid:0)M −j (p) Tba(cid:1) Ψbk,bip s , s = , 30 A. YU. KHRENNIKOV, A. V. KOSYAK, AND V. M. SHELKOVICH Due to Theorem 3.1 and [1], by formula (6.25) all Haar bases generated by adelic Haar MRA are described. Corollary 6.1. Let (see (6.23)) (6.26) i2 ⊗ · · · ⊗ ϑ(m) im ⊗ φ(m+) ⊗ · · · . p eΨbk,bi′ (x′) := eΨbk,(i2,...,im) = ϑ(2) The vectors (see (6.25)) Tba′Φ, (6.27) p;j ba′ := (2 · 3 · · · p)j/2(cid:0)M −j eΨbk,bi′ form the orthonormal basis in the space L2(eA), wherebi′ {1, . . . , 1}, bk = (k2, . . . , kp), ks ∈ F× s , s = 2, 3, . . . , m; j ∈ Z+; ba′ = (a2, . . . , ap) ∈ IA, p ∈ VQ \ ∞. We stress that here p(ba′) = p(j) = p(bk′) = p. p = (i2, . . . , ip) ∈ {1, 2}♯(p−1)+1 \ (p) Tba(cid:1) Ψbk,bi′, 7. Lizorkin spaces on adeles 7.1. Real and p-adic Lizorkin spaces. Recall some facts from [43] -- [45], [54, 2.], [55, §25.1]. Consider the following space Ψ(R) = {ψ(ξ) ∈ S(R) : ψ(j)(0) = 0, j = 0, 1, 2, . . .}, . where S(R) is the real Schwartz space of tempered test functions. The space of functions Φ(R) = {φ : φ = F [ψ], ψ ∈ Ψ(R)}. is called the real Lizorkin space of test functions. The Lizorkin space can be equipped with the topology of the space S(R), which makes Φ a complete space [54, 2.2.], [55, §25.1.]. Since the Fourier transform is a linear isomorphism S(R) onto S(R), this space admits the following characterization: φ ∈ Φ(R) if and only if φ ∈ S(R) is orthogonal to polynomials, i.e., (7.1) ZR xnφ(x) dx = 0, n = 0, 1, 2, . . . . The space Φ′(R) is called the real Lizorkin space of distributions. According to [4], [3, Ch. 7] the p-adic Lizorkin space of test functions is defined as Φ(Qp) = {φ : φ = F [ψ], ψ ∈ Ψ(Qp)}, where Ψ(Qp) = {ψ(ξ) ∈ D(Qp) : ψ(0) = 0}. The space Φ(Qp) equipped with the topology of the space D(Qp) is a complete space. Lemma 7.1. ( [4]) We have (a) φ ∈ Φ(Qp) iff φ ∈ D(Qp) and (7.2) φ(x) dx = 0; ZQp N (Qp) ∩ Φ(Qp) (b) φ ∈ Dl The topological dual of the space Φ(Qp) is the space of p-adic Lizorkin distributions Φ′(Qp). The space Φ′(Qp) can be obtained from D′(Qp) by "sifting out" constants (see [4]). iff ψ = F −1[φ] ∈ D−N −l (Qp) ∩ Ψ(Qp). WAVELET ANALYSIS ON ADELES AND PSEUDO-DIFFERENTIAL OPERATORS 31 7.2. Adelic Lizorkin spaces. Consider the subspace Ψ(A) of the space of test func- tions S(A) (see Subsec. 2.3) consisting of finite linear combinations of elementary func- ηp(ξp) (where VQ is defined by (2.2)) with properties (i)− (iii) of Definition 2.3, where ηp(ξp) = Ω(ξpp) for all p > P ) and such that tions (2.14) η(ξ) =Qp∈VQ = 0, s = 0, 1, 2, . . . , (iv) ds dξs ∞ η(ξ∞, ξ2, ξ3, . . . , ξr−, ξr, ξr+, . . . , ξP , ξP+, . . . )(cid:12)(cid:12)(cid:12)ξr=0 for all r ∈ {∞, 2, 3, . . . , P}, where P = P (η) is the parameter of finiteness of an elemen- tary function η. ηp(ξp) = Ω(ξpp) for all p > P ) and such that Let Ψ(eA) be the space of test functions S(eA) connected with the non-Archimedean part of adeles eA (see Subsec. 2.3) consisting of finite linear combinations of elemen- tary functions η(ξ′) =Qp∈VQ\∞ ηp(ξp) with properties (i) − (iii) of Definition 2.3 (here (iv∗) η(ξ2, ξ3, . . . , ξr−, ξr, ξr+, . . . , ξP , ξP+, . . . )(cid:12)(cid:12)ξr=0 = 0, Φ(A) =(cid:8)ζ : ζ = F [η], η ∈ Ψ(A)(cid:9) and Φ(eA) =(cid:8)ζ : ζ = F [η], η ∈ Ψ(eA)(cid:9) where P = P (η) is the parameter of finiteness of an elementary function η. are called the adelic Lizorkin spaces of test functions. r ∈ {2, 3, . . . , P}, Definition 7.1. The spaces Φ(A) can be equipped with the topology of the space S(A) (see (2.15)). Since the Fourier transform is a linear isomorphism S(A) onto S(A) (see [24, Ch. III,§2.2]), we have Φ(A) ⊂ S(A). The Lizorkin spaces Φ(A) and Φ(eA) admit the following characterizations. Lemma 7.2. (1) Any elementary function ζ(x) = ζ∞(x∞) Yp∈VQ\∞ ζp(xp) ∈ Φ(A) (here ζp(xp) = Ω(xp) for all p > P ) if and only if ζ ∈ S(A) and (7.3) xs rζ(x∞, x2, x3, . . . , xr−, xr, xr+, . . . , xP , xP+, . . . ) dxr = 0, ZQr where if r = ∞, then s = 0, 1, 2, . . . , and if r = 2, 3, . . . , P , then s = 0. (2) Any elementary function eζ(x′) = Yp∈VQ\∞ ζp(xp) ∈ Φ(eA) (7.4) if and only if eζ ∈ S(eA) and Remark 7.1. In view of condition (iii), similarly to (2.15), the spaces Φ(A) and Φ(eA) ZQreζ(x2, x3, . . . , xr−, xr, xr+, . . . , xP , xP+, . . . ) dxr = 0, Φ(eA) = lim ind Φ[m](eA), admit a representation in the form r = 2, 3, . . . , P. ind Φ[m](A), m∈VQ\∞ Φ(A) = lim m∈VQ\∞ (7.5) 32 A. YU. KHRENNIKOV, A. V. KOSYAK, AND V. M. SHELKOVICH functions with the parameter of finiteness m, m ∈ VQ. The representation (7.5) equip where Φ[m](A) ⊂ Φ(A) and Φ[m](eA) ⊂ Φ(eA) are subspaces of the corresponding test the spaces Φ(A) and Φ(eA) with the inductive limit topology. The spaces Φ′(A) and Φ′(eA) are called the adelic Lizorkin spaces of distributions. We define the Fourier transform of distributions f ∈ Φ′(A) and g ∈ Ψ′(A) by the relations hF [f ], ηi = hf, F [η]i, hF [g], ζi = hg, F [ζ]i, ∀ η ∈ Ψ(A), ∀ ζ ∈ Φ(A). (7.6) (7.7) By definition, F [Φ(A)] = Ψ(A), F [Ψ(A)] = Φ(A), i.e., (7.6) give well defined objects. Moreover, we have F [Φ′(A)] = Ψ′(A), F [Ψ′(A)] = Φ′(A). 7.3. Adelic Lizorkin spaces and wavelets. It is well known that p-adic Haar wavelet functions (3.20) from L2(Qp) satisfy the condition (see [3, Ch. 8]) ZQp ψk;ja(x) dx = 0, k ∈ F× p , j ∈ Z, a ∈ Ip, where ψk;ja(x) = pj/2ψk(p−jx − a) and wavelet functions ψk, k ∈ F× Theorem 3.1. p , are given by In view of the formula (see [3, Ch. 8] or [39]) p-adic Haar wavelet functions (3.24) from L2(Zp) satisfy the following condition: a = 0, j ≤ −1, otherwise, k ∈ F× p , j ∈ Z+, a ∈ I j p, 0, (cid:0)Ω(xp), ψk;ja(x)(cid:1) =(cid:26) pj/2, ZZp eψ(0) k;ja(x) dx =(cid:0)Ω(xp), ψk;ja(x)(cid:1) = 0, ZZp eψk;ja(x) dx = 0, is described in Remark 3.1. (7.8) (7.9) where I (j) p where I j functions (3.25) from L2(Zp) we also have p is defined in Proposition 3.1. It is easy to see that for the p-adic Haar wavelet k ∈ F× p , j ∈ Z+, a ∈ I (j) p , φ(x) dx = 1, and Lemma 7.2, we conclude that the adelic wavelet functions (5.5) and (6.25) belong to the Lizorkin space Thus, taking into account relations (7.7) -- (7.9),RQp Φ(eA) only if products (5.6) and (6.27) do not contain the functions φ(p)(xp − ap) as factors for p ≤ m. 8. Characterization of the adelic Lizorkin spaces in terms of wavelets To construct adelic wavelet bases in L2(A), we used the real Haar basis (3.5) in L2(R), jn(t) dt = 0 does not hold for n ∈ N, it is clear that ψH jn /∈ Φ(R). Therefore, in contrast to the p-adic case (see [3, 8.14]), the characterization of the adelic Lizorkin spaces in terms of wavelets is possible only for jn(t) dt = 0. SinceRR tnψH For simplicity here and in what follows we will suppose that the adelic wavelets (5.6) are constructed by using the one-dimensional wavelet basis (3.23) and its restriction (3.24) to Zp. for which we haveRR ψH the space Φ(eA). WAVELET ANALYSIS ON ADELES AND PSEUDO-DIFFERENTIAL OPERATORS 33 (8.1) Lemma 8.1. Any test functioneζ ∈ Φ(eA) can be represented in the form of a finite sum where eΛ is the set of indices (4.33); cα are constants; eΨα(x′) are wavelet functions (5.6) eζ(x′) =Xα∈ eΛ cαeΨα(x′), eΨα = eΨ(bk;bj ba) = O2≤q≤m x′ = (x2, x3, . . . ) ∈eA, αq ⊗Om<q φ(q), α ∈eΛ, ψ(q) αq (xq) = φ(q)(xq − aq), ap ∈ Ip, (see (5.2), (5.3)) for which do not contain the factors ψ(q) q ≤ m. Proof. By definition of the space Φ(eA) (see Subsec. 7.2), it is sufficient to prove this lemma for the case of an elementary function eζ(x′) =Qp∈VQ\∞ ζp(xp), where ζp(xp) = φ(p)(xp) = Ω(xpp) for all p > P , and P = P (eζ) is the parameter of finiteness of an elementary functioneζ. It is clear that eζ ∈ L2(eA). Then (8.1) holds. We will prove that According to Subsec. 7.3, eΨα(x′) ∈ Φ(eA) only if the product (5.6) does not contain only the finite number of cα 6= 0. functions φ(p)(xp − ap) as factors for p ≤ m. Now we need to calculate the coefficients (8.2) cα = c(bk;bj ba) =(cid:0)eζ(x′),eΨα(x′)(cid:1) = Yp∈VQ\∞(cid:0)ζq(xq), ψ(q) p = q , Z+, Iq), are given by (3.23), (3.24), (5.2), (5.3), and among them there are no kq;jqaq (xq), αq = (kq, jq, aq) ∈ I+ where one-dimensional wavelet functions ψ(q) (F× wavelet functions φ(q)(xq − aq). αq = ψ(q) αq (xq)(cid:1), 1. Suppose that m > P . Then we have i.e., cα = 0. 2. Let m < P . Consider the part of product (8.2) Ym<q≤P(cid:0)ζq(xq), φ(q)(xq)(cid:1) = Ym<q≤P(cid:0)ζ 0 q (xq), φ(q)(xq)(cid:1), where ζ 0 support in Zq (see Sec. 7.1). Using the Parseval-Steklov theorem, we obtain q (xq) := ζq(xq)(cid:12)(cid:12)Zq ∈ Φ(Zq) and Φ(Zq) is the Lizorkin space of test functions with Ym<q≤P(cid:0)ζq(xq), φ(q)(xq)(cid:1) = Ym<q≤P(cid:0)F [ζ 0 q (xq) ∈ Φ(Zq) belongs to one of the spaces Dlq q ](ξq), F [φ(q)(xq)](ξq)(cid:1). It is clear that ζ 0 isfies condition (7.2). Then in view of (2.7) and Lemma 7.1, F [ζ 0 and supp F [ζ 0 Nq(Zq), lq ≤ Nq ≤ 0, and sat- −lq (Zq), q ] ⊂ B−lq \B−Nq, 0 ≤ −Nq ≤ −lq. At the same time, F [φ(q)](ξq) = φ(q)(ξq). q ] ∈ Ψ(Qq)∩D−Nq cα = c(bk;bj ba) =(cid:0)eζ(x′),eΨα(x′)(cid:1) = Y2≤q≤P(cid:0)ζq(xq), ψ(q) αq (xq)(cid:1)× αq (xq)(cid:1)Ym<q(cid:0)φ(q)(xq), φ(q)(xq)(cid:1). × YP <q≤m(cid:0)φ(q)(xq), ψ(q) ZZq αq (xq) dxq = 0, ψ(q) for P < q ≤ m, According to (7.7) and (7.8), (7.9), we have 34 A. YU. KHRENNIKOV, A. V. KOSYAK, AND V. M. SHELKOVICH 3. Let m = P . In this case, using the Parseval-Steklov theorem, one can rewrite the q ](ξq), F [φ(q)(xq)](ξq)(cid:1) = 0, Thus supp F [ζ 0 m < q ≤ P . That is in this case cα = 0. product (8.2) in the form q ] ∩ supp φ(q) = ∅ and, consequently, (cid:0)F [ζ 0 cα = c(bk;bj ba) =(cid:0)eζ(x′),eΨα(x′)(cid:1) = Y2≤q≤m(cid:0)ζq(xq), ψ(q) αq (xq)(cid:1) = Y2≤q≤m(cid:0)F [ζq](ξq), F [ψ(q) αq ](ξq)(cid:1). (2.7), any function ζq(xq) ∈ Φ(Qq) belongs to one of the spaces Dlq condition (7.2), i.e., F [ζq] ∈ Ψ(Qq)∩D−Nq Let ζq ∈ Φ(Qq) for 2 ≤ q ≤ p. According to Definition 7.1, Lemma 7.1, and relation Nq (Qq) and satisfies −lq (Qq), and supp F [ζq] ⊂ B−lq\B−Nq , 2 ≤ q ≤ p. For the wavelet function (5.2), αq (xq) = ψ(q) ψ(q) kq; jqaq (xq) F [ψ(q) q (q−jqxq − aq)(cid:17)Ω(cid:0)q−jqxq − aqq(cid:1), + qjqξq(cid:12)(cid:12)(cid:12)q(cid:17), = qjq/2χq(cid:16)kq kq; jqaq ](ξq) = q−jq/2χq(cid:0)qjqaqξq(cid:1)Ω(cid:16)(cid:12)(cid:12)(cid:12) (cid:0)F [ζq](ξq), F [ψ(q) kq q xq ∈ Qq, jq ∈ Z+. (8.3) we have (8.4) Taking into account that supp F [ζq] ⊂ B−lq \ B−Nq and using formula (8.4), one can conclude that (8.5) only if q−Nq ≤ ξqq ≤ q−lq and kq ξq = q−jq(cid:0)ηq − kq Now we consider the scalar product(cid:0)F [ζq](ξq), F [ψ(q) αq ](ξq)(cid:1) 6= 0, q(cid:12)(cid:12)q = qjq+1, one can see that the product (8.5) αq ](ξq)(cid:1), where ζq ∈ Φ(Zq) and q(cid:1) and ξqq = qjq(cid:12)(cid:12)ηq − kq q + qjqξq = ηq ∈ Zq for any 2 ≤ q ≤ p. Since is nonzero only for finite number of indices jq such that xq ∈ Zq, is a wavelet function given by (5.3), p < q ≤ m. Using the identity [64, VII.1], [39] j ≤ j′, kq; jqaq (xq) = qjq/2χq(cid:16)kq xp − a′ q−Nq < ξqq = qjq+1 ≤ q−lq , αq (xq) = ψ(q) ψ(q) 2 ≤ q ≤ p. (8.6) Ω(cid:0)pjxp − app(cid:1)Ω(cid:0)pj ′ by explicit calculation we obtain (q−jqxq − aq)(cid:17)Ω(cid:0)q−jqxq − aqq(cid:1), pp(cid:1) = Ω(cid:0)pjxp − app(cid:1)Ω(cid:0)pj ′−jap − a′ pp(cid:1), + qjqξq(cid:12)(cid:12)(cid:12)q(cid:17), Next, using (8.7) and repeating the above calculation almost word for word, we find that jq ∈ Z+. kq q (8.7) F [ψ(q) αq(cid:12)(cid:12)Zq ](ξq) = q−jq/2Ω(qjqaqq)χq(cid:0)qjqaqξq(cid:1)Ω(cid:16)(cid:12)(cid:12)(cid:12) ](ξq)(cid:1) 6= 0, wavelet functions eΨα(x′), in the product (8.2) we have m = P . (8.8) only for finite number of indices jq, p < q ≤ m. αq(cid:12)(cid:12)Zq (cid:0)F [ζq](ξq), F [ψ(q) Thus cα = c(bk;bj ba) 6= 0 for finite number of indices jq, 2 ≤ q ≤ m. Moreover, for all q WAVELET ANALYSIS ON ADELES AND PSEUDO-DIFFERENTIAL OPERATORS 35 Now we consider the product (8.2) again cα = c(bk;bj ba) =(cid:0)eζ(x′),eΨα(x′)(cid:1) = Y2≤q≤m(cid:0)ζq(xq), ψ(q) αq (xq)(cid:1), where according to the above calculation, m = P . Now let calculate the expression αq (xq) is given by (8.3). Using identity (8.6) we have αq (xq)(cid:1). Here the function ζq(xq) ∈ Φ(Qq) belongs to one of the spaces (cid:0)ζq(xq), ψ(q) Dlq Nq (Qq) and ψ(q) Ω(cid:0)q−jqxq − aqq(cid:1)Ω(cid:0)qNqxqq(cid:1) =( Ω(cid:0)q−jqxq − aqq(cid:1)Ω(cid:0)qN +jqaqq(cid:1), −jq ≤ Nq, Ω(cid:0)qNqxqq(cid:1)Ω(cid:0)aqq(cid:1), αq (xq)(cid:1) 6= 0 at least for qN +jqaq ∈ Zq (−jq ≤ Nq) or This relation implies that(cid:0)ζq(xq), ψ(q) aq = 0 (−jq > Nq). Since the set of indices jq is finite, the set of the above indices aq is also finite. Thus products (8.5), (8.8) are nonzero only for finite number of aq ∈ Iq. (cid:3) Corollary 8.1. Any test function ζ ∈ Φ(A) can be represented in the form of a finite sum −jq > Nq. ζ(x) =Xα∈ eΛ cα(x∞)eΨα(x′), x = (x∞, x′) ∈ A, (8.9) (8.10) where eΛ is the set of indices (4.33); cα(x∞) ∈ Φ(R) are some Lizorkin test functions; eΨα(x′) = eΨ(bk;bj ba)(x′) are wavelet functions (5.6) which do not contain factors ψ(q) φ(q)(xq − aq), ap ∈ Ip, (see (5.2), (5.3)) for q ≤ m. αq (xq) = Using standard results from the book [56], we obtain the following assertion. Proposition 8.1. Any distribution f ∈ Φ′(A) can be realized in the form of an infinite sum: f (x) =Xα∈ eΛ bα(x∞)eΨα(x′), x = (x∞, x′) ∈ A, (5.6). where bα(x∞) ∈ Φ′(R) are some real Lizorkin distributions; eΨα(x′) are wavelet functions Here any distribution f ∈ Φ′(A) is associated with representation (8.10), where the coefficients (8.11) bα(x∞) And vice versa, taking into account Corollary 8.1 and the orthonormality of the wavelet basis (5.6), any infinite sum (8.10) is associated with the distribution f ∈ Φ′(A), whose action on a test function ζ(x) ∈ Φ(A) is defined as def = (cid:10)f (x∞, x′),eΨα(x′)(cid:11), α ∈eΛ. cα(x∞)eΨα(x′)E bβ(x∞)eΨβ(x′),Xα∈ eΛ =Xα∈ eΛ(cid:10)bα(x∞), cα(x∞)(cid:11), (cid:10)f, ζ(cid:11) =DXβ∈ eΛ (8.12) where the latter sum is finite. 36 A. YU. KHRENNIKOV, A. V. KOSYAK, AND V. M. SHELKOVICH Proposition 8.2. Any distribution ef ∈ Φ′(eA) can be realized as an infinite sum of the form (8.13) ef (x′) =Xp Xα∈ eΛ x′ ∈eA, where bα are constants; eΨα(x′) are wavelet functions (5.6). bαeΨα(x′), wavelet functions (6.27) instead of wavelet functions (5.6). It is clear that in Lemma 8.1, Corollary 8.1, and Propositions 8.1, 8.2 one can use 9. Pseudo-differential operators on adeles 9.1. Real and p-adic fractional operators. Let us introduce a distribution from S ′(R) (we keep the notation · = · ∞.) (9.1) s = 0, 1, 2, . . . , , α 6= −2s, α 6= 1 + 2s, α (x) = xα−1 κ(∞) γ1(α) x ∈ R, called the Riesz kernel , where xα is a homogeneous distribution of degree α defined in [54, Lemma 2.9.], [55, (25.19)], [23, Ch.I,§3.9.], and γ1(α) = 2απ . The Riesz One can define the Riesz kernel (9.1) in the real Lizorkin space of distributions Φ′(R) kernel is an entire function of the complex variable α. 1 2 Γ( α 2 ) Γ( 1−α 2 ) (see [45], [54, Lemma 2.13.], [55, Lemma 25.2.]): xα−1 = (9.2) κ(∞) α (x) = xα−1Γ( 1−α 2 ) 2α√πΓ( α 2 ) (−1)s+1x2s log x (−1)sδ(2s)(x), π(2s)!  2Γ(α) cos( πα 2 ) , α 6= −2s, α 6= 1 + 2s, , α = 1 + 2s, α = −2s, s ∈ Z+. According to [23, Ch.II,§3.3.,(2)], [54, Lemma 2.13.], [55, Lemma 25.2.], (9.3) F [κ(∞) α (x)](ξ) = x−α, x ∈ R. Define the real Riesz fractional operator Dα ∞ on the Lizorkin space Φ(R) as a convo- lution (9.4) where κ(∞) α (9.5) (cid:0)Dα ∞ϕ(cid:1)(x) =(cid:0)κ(∞) −α ∗ ϕ(cid:1)(x) =(cid:10)κ(∞) is given by (9.2). It is clear that [55, (25.2)] −α (·), ϕ(x − ·)(cid:11), ϕ ∈ Φ(R), (cid:0)Dα ∞ϕ(cid:1)(x) = F −1[ξαF [ϕ](ξ)](x), ϕ ∈ Φ(R). Lemma 9.1. The Lizorkin spaces of test functions Φ(R) and distributions Φ′(R) are invariant under the Riesz fractional operator (9.5) and Dα ∞(Φ′(R)) = Φ′(R). The p-adic fractional operator Dα in [58], [59, III.4.]. In [4], the fractional operator Dα of distributions Φ′(Qp) for all α ∈ C by the following relations: (9.6) p was introduced on the space of distributions D′(Qp) p was defined in the Lizorkin space f ∈ Φ′(Qp), α ∈ C. Representation (9.6) can be rewritten as a convolution (cid:0)Dα p f(cid:1)(x) = F −1(cid:2) · α (cid:0)Dα p f(cid:1)(x) =(cid:0)κ(p) −α ∗ f(cid:1)(x) = hκ(p) p F [f ](·)(cid:3)(x), (9.7) −α(·), f (x − ·)i, f ∈ Φ′(Qp), α ∈ C, WAVELET ANALYSIS ON ADELES AND PSEUDO-DIFFERENTIAL OPERATORS 37 where (see [4]) (9.8) is the Riesz kernel, and κ(p) α (x) = p xα−1 Γp(α) , δ(x), log p log xp, − 1−p−1 = ZQp xα−1 def Γp(α) p χp(x) dx = 1 − pα−1 1 − p−α α 6= 0, 1, α = 0, α = 1, x ∈ Qp, is the p-adic Γ-function (see [59, III,Theorem (4.2)], [64, VIII,(4.4)]). Lemma 9.2. ([4]) The Lizorkin spaces of test functions Φ(Qp) and distributions Φ′(Qp) are invariant under the fractional operator (9.6) and Dα p (Φ′(Qp)) = Φ′(Qp). 9.2. Adelic fractional operators on the Lizorkin spaces. Let us introduce on the space Φ(A) the adelic fractional operator Dbγ of order bγ = (γ∞, γ2, . . . , γp, . . . ) ∈ C∞, which is defined by its projection on any subspace of test functions Φ[m](A) ⊂ Φ(A) (for details, see Remark 4.1): Dbγ(cid:12)(cid:12)Φ[m](A) def = Dbγ (m) = Op∈{∞,2,3,...,m} Dγp p ⊗Op>m Idp, m ∈ VQ, ∞ and Dγp where Dγ∞ the identity operator in Φ(Qp). Here the fractional operator Dbγ product of one-dimensional operators. p are defined by (9.4), (9.5) and (9.6), (9.7), respectively, Idp is (m) is an infinite tensor (9.9) If (9.10) where (9.11) ζ(x) = ζ∞(x∞) Yp∈VQ\∞ ζp(xp) ∈ Φ[m](A) is an elementary function, i.e., ζp(xp) = φ(p)(ξp) = Ω(xpp) for all p > m(ζ), then taking into account Definition (9.9) and (9.5), (9.6), we obtain (cid:0)Dbγ(cid:12)(cid:12)Φ[m](A)ζ(cid:1)(x) =(cid:0)Dbγ (m)ζ(cid:1)(x) = F −1(cid:2) · bγ (m) = Yp∈{∞,2,3,...,m} p Yp>m ξbγ ξpγp φ(p)(ξp), (m)F [ζ](·)(cid:3)(x), ξ = (ξ∞, ξ2, . . . ξp, . . . ) ∈ A, is a symbol of the operator Dbγ (m), m ∈ VQ. It is clear that ξbγ (m) for all bγ = (γ∞, γ2, . . . , γp, . . . ) ∈ C∞ such that for all p > m, γp = 0 is a multiplier in the space of test functions Ψ[m](Qp), m ∈ VQ. According to (9.4), (9.7), (2.10), relation (9.10) can be rewritten as (cid:0)Dbγ (m)ζ(cid:1)(x) =(cid:0)κ−bγ;(m) ∗ ζ(cid:1)(x) =(cid:10)κ−bγ;(m)(·), ζ(x − ·)(cid:11), ζ ∈ Φ(A), (9.12) where κbγ;(m)(ξ) = Yp∈{∞,2,3,...,m} κ(p) γp (ξp)Yp>m φ(p)(ξp) is the adelic Riesz kernel, and κ(∞) γ∞ and κ(p) γp are given by (9.2) and (9.8), respectively. For f ∈ Φ′(A) we define the distribution Dbγf by the relation (9.13) hDbγf, ζi def = hf, Dbγ (m)ζi, ∀ ζ ∈ Φ[m](A), m ∈ VQ \ ∞. 38 A. YU. KHRENNIKOV, A. V. KOSYAK, AND V. M. SHELKOVICH It is easy to see that Lemmas 9.1, 9.2 imply the following statements. Lemma 9.3. The Lizorkin spaces of test functions Φ(A) and distributions Φ′(A) are invariant under the fractional operator Dbγ (m)(Φ′(A)) = Φ′(A). (m) and Dbγ Proposition 9.1. The family of operators {Dbγ :bγ ∈ C∞} on the space of distributions Φ′(A) forms an abelian group: if f ∈ Φ′(A) then bβDbγf = Dbγ+ bβf, bβf = D DbγD DbγD−bγf = f, bγ, bβ ∈ C∞. For the case bγ = (γ∞, γ2, . . . , γp, . . . ) ∈ C∞ such that γ∞ = γ2 = · · · = γp = · · · = γ ∈ C we shall write Dγ instead of Dbγ. Similarly to (9.9), the adelic fractional operator Dγ of order γ ∈ C∞ is defined by its projection on any subspace Φ[m](A) ⊂ Φ(A) (see (7.5)): (9.14) Dγ(cid:12)(cid:12)Φ[m](A) def = Dγ (m) = Op∈{∞,2,3,...,m} Dγ p ⊗Op>m Idp, m ∈ VQ, p are defined by (9.4), (9.5) and (9.6), (9.7), respectively, Idp is the where Dγ identity operator in Φ(Qp). If ∞ and Dγ ζ(x) = ζ∞(x∞) Yp∈VQ\∞ ζp(xp) ∈ Φ[m](A) is an elementary function, i.e., ζp(xp) = φ(p)(ξp) = Ω(xpp) for all p > m(ζ), then taking into account Definition (9.14) and (9.5), (9.6), we obtain (9.15) where (9.16) (cid:0)Dγ(cid:12)(cid:12)Φ[m](A)ζ(cid:1)(x) =(cid:0)Dγ (m)ζ(cid:1)(x) = F −1(cid:2) · γ p Yp>m (m) = Yp∈{∞,2,3,...,m} ξγ ξpγ φ(p)(ξp), (m)F [ζ](·)(cid:3)(x), ξ = (ξ∞, ξ2, . . . ξp, . . . ) ∈ A. is the symbol of the operator Dγ (m), m ∈ VQ. 9.3. One class of adelic pseudo-differential operators. On the adelic Lizorkin space of distributions Φ′(A) we introduce a pseudo-differential operators A with the symbol Ap(ξp), ξ ∈ A, A(ξ) = Yp∈VQ where A∞(ξ∞) ∈ C ∞(R \ {0}) and(cid:12)(cid:12) ds hood of ξ∞ = 0; Ap(ξp) ∈ E(Qp \ {0}), p = 2, 3, . . . . of test functions Φ[m](A) ⊂ Φ(A) (see (7.5)): (9.17) dξs def = A(m) = F −1 A(m) F, m ∈ VQ, ∞A∞(ξ∞)(cid:12)(cid:12) ≤ Msξ∞−ms (ms ≥ 0) in a neighbor- The adelic pseudo-differential operator A is defined by its projection on any subspace where (9.18) A(m)(ξ) = A∞(ξ∞) Yp∈{2,3,...,m} Ap(ξp)Yp>m φ(p)(ξp), ξ ∈ A, A(cid:12)(cid:12)Φ[m](A) WAVELET ANALYSIS ON ADELES AND PSEUDO-DIFFERENTIAL OPERATORS 39 is the symbol of the pseudo-differential operator A(m). If ζ(x) = ζ∞(x∞) Yp∈VQ\∞ ζp(xp) ∈ Φ[m](A) is an elementary function, i.e., ζp(xp) = φ(p)(ξp) = Ω(xpp) for all p > m(ζ), then taking into account Definition (9.17), (9.18), we obtain (9.19) A(cid:12)(cid:12)Φ[m](A)ζ = A(m)ζ =(cid:0)F −1 A(m) F(cid:1)ζ, k=1 ζ r ∈ Φ[m](A), where ζ r are elementary functions, the operator A is defined by the relation where the symbol A(m) is defined by (9.18). For any test function ζ = Pr (A(cid:12)(cid:12)Φ[m](A)ζ)(x) = (A(m)ζ)(x) = rXk=1 (A(m)ζ r)(x). Now we define a conjugate pseudo-differential operator AT on Φ(A). For any function ζ ∈ Φ[m](A): (AT(cid:12)(cid:12)Φ[m](A)ζ)(x) def = (AT (m)ζ)(x) = F −1(cid:2)A(m)(−ξ) F [ζ](ξ)(cid:3)(x). Then the operator A in the Lizorkin space of distributions Φ′(A) is defined in the usual way: for f ∈ Φ′(A) we have (9.20) ∀ ζ ∈ Φ[m](A), m ∈ VQ \ ∞. hAf, ζi = hf, AT (m)ζi, It follows from the latter relation and (2.8) that (9.21) Af =(cid:0)F −1 A F(cid:1)f, f ∈ Φ′(A). Lemma 9.4. The Lizorkin spaces Φ(A) and Φ′(A) are invariant under the operator (9.21). Proof. It is sufficient to prove this lemma for elementary functions. Let ζ ∈ Φ[m](A) be an elementary function. Since the function A(m)(ξ) defined by (9.18) is a multiplier in the space Ψ[m](Qp), m ∈ VQ, both functions F [ζ](ξ) and A(m)(ξ)F [ζ](ξ) belong to Ψ[m](A) and, consequently, Aζ ∈ Φ[m](A), m ∈ VQ. Thus the pseudo-differential operator A (see (9.17), (9.19)) is well defined and the Lizorkin space Φ(A) is invariant under its action. Therefore, if f ∈ Φ′(A), then according to (9.20), (9.21), Af = (F −1 A F )f ∈ Φ′(A), i.e., the Lizorkin space of distributions Φ′(A) is invariant under A. (cid:3) The fractional operators (9.10) and (9.15) are particular cases of operator (9.19). Similarly to the above constructions, we introduce the corresponding operators on the Lizorkin spaces Φ(eA) and Φ′(eA) : the pseudo-differential operator (9.22) A0ζ =(cid:0)F −1 A0) F(cid:1)ζ, ζ ∈ Φ(eA), with the symbol (9.23) Ap(ξp), A0(ξ′) = Yp∈VQ\∞ (cid:0)Dbγ 0 ζ(cid:1)(x′) = F −1(cid:2) · bγ ξ′ = (ξ2, . . . ξp, . . . ) ∈eA; 0F [ζ](·)(cid:3)(x′), ζ ∈ Φ(eA), the fractional operator (9.24) 40 A. YU. KHRENNIKOV, A. V. KOSYAK, AND V. M. SHELKOVICH and the fractional operator with the symbol (9.25) (9.26) with the symbol (9.27) ξpγp p , 0 = Yp∈VQ\∞ ξ′bγ 0 ζ(cid:1)(x′) = F −1(cid:2) · γ (cid:0)Dγ 0 = Yp∈VQ\∞ ξ′γ ξpγ p, ξ′ = (ξ2, . . . ξp, . . . ) ∈eA; 0F [ζ](·)(cid:3)(x′), ζ ∈ Φ(eA), ξ′ = (ξ2, . . . ξp, . . . ) ∈eA. 9.4. On the eigenfunctions of adelic pseudo-differential operators. Theorem 9.1. Let A0 be a pseudo-differential operator (9.22) with symbol (9.23), and eΨα(x′) = eΨ(bk;bj ba)(x′) = O2≤q≤m ψ(q) αq (xq) ⊗Om<q be wavelet functions (5.6) which do not contain factors ψ(q) φ(q)(xq), α ∈eΛ, x′ ∈eA, αq (xq) = φ(q)(xq − aq), ap ∈ Ip, (9.28) (see (5.2), (5.3)) for q ≤ m; eΛ is the indexes set (4.33). Then eΨα is an eigenfunction of the operator A0 if and only if Yq∈{2,3,...,m} Aq(cid:0)q−jq(−q−1kq + ηq)(cid:1) = Yq∈{2,3,...,m} Aq(cid:0) − q−jq−1kq(cid:1), holds for all ηq ∈ Zq, q = 2, 3, . . . , m. The corresponding eigenvalue is the following jq ∈ Z+, def Proof. In view of (9.23) and (2.10), we have A0(cid:12)(cid:12)Φ[m](eA)eΨα(x′) λ =Qq∈{2,3,...,m} Aq(cid:0) − qjq−1kq(cid:1), i.e., = A0(m)eΨα(x′) =(cid:16) Yq∈{2,3,...,m} Aq(cid:0) − qjq−1kq(cid:1)(cid:17)eΨα(x′), α ∈eΛ, A0(cid:12)(cid:12)Φ[m](eA)eΨα(x′) = A0(m)eΨα(x′) = F −1(cid:2)A0(m)(ξ′)F [eΨα(ξ′)(cid:3)(cid:3)(x′) αq(cid:3)(cid:3)(xq)Yq>m = Y2≤q≤p F −1(cid:2)AqF [ψ(q) kq; jqaq (xq) = qjq/2χq(cid:0) kq where ψ(q) (5.3)), and xq ∈ Qq, 2 ≤ q ≤ p; xq ∈ Zq, p < q ≤ m. αq(cid:3)(cid:3)(xq) Yp<q≤m F −1(cid:2)AqF [ψ(q) If condition (9.28) holds, then using formulas (8.4), (8.7), we obtain αq (xq) = ψ(q) def q (q−jqxq − aq)(cid:1)Ω(cid:0)q−jqxq − aqq(cid:1) (see (5.2), x′ ∈eA. φ(q)(xq), A0(cid:12)(cid:12)Φ[m](eA)eΨα(x′) = A0(m)eΨα(x′) = Y2≤q≤p × Yp<q≤m q−jq/2F −1hAq(ξq)χq(cid:0)qjqaqξq(cid:1)Ω(cid:16)(cid:12)(cid:12)(cid:12) q−jq/2F −1hAq(ξq)Ω(qjqaqq)χq(cid:0)qjqaqξq(cid:1)Ω(cid:16)(cid:12)(cid:12)(cid:12) kq q + qjqξq(cid:12)(cid:12)(cid:12)q(cid:17)i(xq)× + qjqξq(cid:12)(cid:12)(cid:12)q(cid:17)i(xq)× ×Yq>m φ(q)(xq). kq q WAVELET ANALYSIS ON ADELES AND PSEUDO-DIFFERENTIAL OPERATORS 41 kq q Making the change of variables ξq = q−jq(ηq − q−1kq) and using formulas (2.8), (2.10), we have for 2 ≤ q ≤ p kq q αq (xq), xq ∈ Zq. kq q αq (xq) q αq (xq), xq ∈ Qq. In the same way, we obtain for p < q ≤ m χq(cid:0)(−q−1kq + ηq)(−q−jqxq + aq)(cid:1)Ω(ηqq) dηq The latter expression is nonzero only if qjqaq ∈ Zq and xq − qjq aqq ≤ q−jq. This implies that xq ∈ Zq. Thus, + qjqξq(cid:12)(cid:12)(cid:12)q(cid:17)i(xq) = Aq(cid:0) − qjq−1kq(cid:1)ψ(q) + qjqξq(cid:12)(cid:12)(cid:12)q(cid:17)i(xq) (q−jqxq − aq)(cid:17)Ω(cid:0)q−jqxq − aqq(cid:1). q−jq/2F −1hAq(ξq)χq(cid:0)qjqaqξq(cid:1)Ω(cid:16)(cid:12)(cid:12)(cid:12) = Aq(cid:0) − qjq−1kq(cid:1)qjq/2ZQq q−jq/2F −1hAq(ξq)Ω(qjqaqq)χq(cid:0)qjqaqξq(cid:1)Ω(cid:16)(cid:12)(cid:12)(cid:12) = Aq(cid:0) − qjq−1kq(cid:1)Ω(qjqaqq)ψ(q) = Aq(cid:0) − qjq−1kq(cid:1)Ω(qjqaqq)qjq/2χq(cid:16)kq + qjqξq(cid:12)(cid:12)(cid:12)q(cid:17)i(xq) q−jq/2F −1hAq(ξq)Ω(qjqaqq)χq(cid:0)qjqaqξq(cid:1)Ω(cid:16)(cid:12)(cid:12)(cid:12) = Aq(cid:0) − qjq−1kq(cid:1)ψ(q) Consequently, A0eΨα(x′) = A0(m)eΨα(x′) = λeΨα(x′), where Aq(cid:0) − qjq−1kq(cid:1), x′ ∈eA. Conversely, if A0(cid:12)(cid:12)Φ[m](eA)eΨα(x′) = A0(m)eΨα(x′) = λeΨα(x′), then taking the Fourier trans- + qjqξq(cid:12)(cid:12)(cid:12)q(cid:17) = 0, q(cid:1)q−jq for 2 ≤ q ≤ m. Thus, q + qjqξq = ηq, ηq ∈ Zq, then ξq =(cid:0)ηq − kq λ = A0(m)(cid:16)(cid:0)η2 − 2−1k2(cid:1)2−j2, . . . ,(cid:0)ηm − m−1km(cid:1)m−jm(cid:17), In particular, λ = A0(m)(cid:0)2−j2−1k2, . . . , m−jm−1km(cid:1) and consequently (9.28) holds. p Qq>m φ(q)(ξq), ξ′ = (ξ2, . . . ξp, . . . ) ∈ eA. A0(m)(ξ′) = ξ′bγ is easy to see that the symbol A0(m)(ξ′) = ξ′bγ Yq∈{2,3,...,m} form of both left- and right-hand sides of this identity and using formulas (8.4), (8.7), we obtain According to (9.24), (9.25), the adelic fractional operator Dbγ 0(m) satisfies condition (9.28): 0(m) = Qq∈{2,3,...,m} ξqγq Aq(cid:0)q−jq(−q−1kq + ηq)(cid:1) = Yq∈{2,3,...,m}(cid:12)(cid:12)(cid:12)q−jq(cid:0) − q−1kq + ηq(cid:1)(cid:12)(cid:12)(cid:12) = Yq∈{2,3,...,m} qjq(cid:12)(cid:12) − q−1kq + ηq(cid:12)(cid:12)γq = Yq∈{2,3,...,m} q = Yq∈{2,3,...,m} Aq(cid:0) − q−jq−1kq(cid:1), ∀ ηq ∈ Zq, qγq(jq+1) (cid:0)A0(ξ′) − λ(cid:1) Y2≤q≤m 0(m) has the symbol It λ = Yq∈{2,3,...,m} ∀ ηq ∈ Zq, 2 ≤ q ≤ m. γq q 2 ≤ q ≤ m. If now kq kq q Ω(cid:16)(cid:12)(cid:12)(cid:12) ξ′ ∈eA. (cid:3) 42 A. YU. KHRENNIKOV, A. V. KOSYAK, AND V. M. SHELKOVICH Thus according to Theorem 9.1, we have Corollary 9.1. The wavelet function (5.6) which does not contain the factors ψ(q) eigenfunction of the adelic fractional operator Dbγ 0 : αq (xq) = φ(q)(xq − aq), ap ∈ Ip for q ≤ m, is an Corollary 9.2. The wavelet function (5.6) (9.29) Dbγ eΨα(x′) = eΨ(bk;bj ba)(x′) = O2≤q≤m 0(cid:12)(cid:12)Φ[m](eA)eΨα(x′) = Dbγ eΨα(x′) = eΨ(bk;bj ba)(x′) = O2≤q≤m 0(cid:12)(cid:12)Φ[m](eA)eΨα(x′) = Dγ Dγ (9.30) ψ(q) αq (xq) ⊗Om<q φ(q)(xq), ψ(q) φ(q)(xq), 0(m)eΨα(x′) =(cid:16) Yq∈{2,3,...,m} αq (xq) ⊗Om<q 0(m)eΨα(x′) =(cid:16) Yq∈{2,3,...,m} Acknowledgements Λm. Λm. x′ ∈eA, α ∈ Λ = [m∈VQ\∞ qγq(jq+1)(cid:17)eΨα(x′), x′ ∈eA. x′ ∈eA, α ∈ Λ = [m∈VQ\∞ qjq+1(cid:17)γ eΨα(x′), x′ ∈eA. which does not contain the factors ψ(q) eigenfunction of the adelic fractional operator Dγ 0 : αq (xq) = φ(q)(xq − aq), ap ∈ Ip, for q ≤ m, is an The authors are greatly indebted to A.N. Kochubei, V.I. Polischook, O.G. Smolyanov for fruitful discussions. References [1] S. Albeverio, S. Evdokimov, M. Skopina, p-Adic multiresolution analysis and wavelet frames, Journal of Fourier Analysis and Applications, 16, no. 5, (2010), 693 -- 714. [2] S. Albeverio, S. Evdokimov, M. Skopina, p-Adic non-orthogonal wavelet bases, Proc. Steklov Inst. Math., v. 265, Moscow, 2009, 1 -- 12. [3] S. Albeverio, A. Yu. Khrennikov, V. M. Shelkovich, Theory of p-adic distributions: Linear and nonolinear models, London Mathematical Society Lecture Note Series (No. 370), Cambridge Uni- versity Press, Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, Sao Paulo, Delhi, Dubai, Tokyo, 2010. [4] S. Albeverio, A. Yu. Khrennikov, V. M. Shelkovich, Harmonic analysis in the p-adic Lizorkin spaces: fractional operators, pseudo-differential equations, p-adic wavelets, Tauberian theorems, Journal of Fourier Analysis and Applications, Vol. 12, Issue 4, (2006), 393 -- 425. [5] S. Albeverio, A. Yu. Khrennikov, and V. M. Shelkovich, p-Adic semi-linear evolutionary pseudo- differential equations in the Lizorkin space Dokl. Ross. Akad. Nauk, 415, no. 3, (2007), 295 -- 299. English transl. in Russian Doklady Mathematics, 76, no. 1, (2007), 539 -- 543. [6] S. Albeverio, A. Yu. Khrennikov, V. M. Shelkovich, The Cauchy problems for evolutionary pseudo- differential equations over p-adic field and the wavelet theory, Journal of Mathematical Analysis and Applications, 375, (2011), 82 -- 98. [7] S. Albeverio, S. V. Kozyrev, Multidimensional ultrametric pseudodifferential equations, Proc. Steklov Inst. Math., v. 265, Moscow, 2009, 13 -- 29. [8] V. Anashin and A. Yu. Khrennikov, Applied algebraic dynamics, De Gruyter, Berlin, New York, 2009. [9] I. Ya. Aref′eva, B. G. Dragovic, and I. V. Volovich On the adelic string amplitudes, Phys. Lett. B, 209, no. 4, (1998), 445 -- 450. [10] J. J. Benedetto, and R. L. Benedetto, A wavelet theory for local fields and related groups, The Journal of Geometric Analysis, 3, (2004), 423 -- 456. [11] R. L. Benedetto, Examples of wavelets for local fields, Wavelets, Frames, and operator Theory, (College Park, MD, 2003), Am. Math. Soc., Providence, RI, (2004), 27 -- 47. WAVELET ANALYSIS ON ADELES AND PSEUDO-DIFFERENTIAL OPERATORS 43 [12] Yu.M. Berezanskii, Selfadjoint operators in spaces of functions of infinitely many variables, Trans- lated from the Russian by H. H. McFaden. Translation edited by Ben Silver. Translations of Math- ematical Monographs, 63. American Mathematical Society, Providence, RI, 1986. [13] A. Borel, Some finiteness properties of adele groups over number fields, Publications math´ematiques de l'I.H.´E.S., tome 16, (1963), 5 -- 30. [14] L. Brekke and P. G. O. Freund, p-Adic numbers in physics, Phys. Rep., 233, 1, (1993). [15] J. W. S. Cassels and A. Frohlich, Algebraic Number Theory, Academic Press, 1967. [16] C. Chevalley, G´en´eralisation de la th´eorie du corps de classes pour les extensions infinies, J. Math. Pures Appl., 15, (1936), 359 -- 371. [17] K. Conrad, The classical and adelic point of view in number theory, (2010), http://www.mccme.ru/ium/postscript/f10/conrad-lecturenotes.pdf [18] I. Daubechies, Ten Lectures on wavelets, CBMS-NSR Series in Appl. Math., SIAM, 1992. [19] B. Dragovich, On Generalized Functions in Adelic Quantum Mechanics, Integral Transform. Spec. Funct. 6, (1998), 197 -- 203. [20] B. Dragovich, Adelic model of harmonic oscillator, (Russian), Teoret. Mat. Fiz., 101, (1994), 349 -- 359; translation in Theoret. and Math. Phys., 101, no. 3, (1994), 1404 -- 1412. [21] B. Dragovich, p-Adic and adelic quantum mechanics, Proc. V.A. Steklov Inst. Math. 245, (2004), 72 -- 85. [22] B. Dragovich, Ya. V. Radyno, and A. Khrennikov, Distributions on adeles, Journal of Mathematical Sciences, 142, no. 3, (2007), 2105 -- 2112. [23] I. M. Gel'fand and G. E. Shilov, Generalized Functions, vol. 1, Properties and Operations. New York, Acad. Press, 1964. [24] I. M. Gel′fand, M. I. Graev, and I. I. Piatetskii-Shapiro, Generalized functions. vol 6: Represen- tation theory and automorphic functions. Nauka, Moscow, 1966. Translated from the Russian by K.A. Hirsch, Published in 1990, Academic Press (Boston). [25] W. Karwowski, M. Vilela, Hierarchical structures and asymmetric stochastic processes on p-adics and adeles, Journal of Mathematical Physics, Volume 35, Issue 9, (1994), 4637 -- 4650. [26] A. Khrennikov, p-Adic valued distributions in mathematical physics, Kluwer Academic Publ., Dor- drecht, 1994. [27] A. Khrennikov, Non-archimedean analysis: quantum paradoxes, dynamical systems and biological models, Kluwer Academic Publ., Dordrecht, 1997. [28] A. Yu. Khrennikov, Information dynamics in cognitive, psychological, social and anomalous phe- nomena, Kluwer Academic Publ., Dordrecht, 2004. [29] A. Yu. Khrennikov, and S. V. Kozyrev, Wavelets on ultrametric spaces, Applied and Computational Harmonic Analysis, 19, (2005), 61 -- 76. [30] A. Yu. Khrennikov and M. Nilsson, p-Adic Deterministic and Random Dynamics, Kluwer, Dor- drecht, 2004. [31] A. Khrennikov, Ya. V. Radyno, On adelic analog of Laplacian, Proceedings of Jangjeon Mathe- matical Society, 6, no. 1, (2003), 1 -- 18. [32] A.Yu. Khrennikov, V.M. Shelkovich, p-Adic multidimensional wavelets application http://arxiv.org/abs/math-ph/0612049 p-adic to pseudo-differential operators, (2006), Preprint at and the their url: [33] A. Yu. Khrennikov, V. M. Shelkovich, Non-Haar p-adic wavelets and their application to pseudo- differential operators and equations, Applied and Computational Harmonic Analysis, 28, (2010), 1 -- 23. [34] A.Yu. Khrennikov, V.M. Shelkovich, M. Skopina, p-Adic refinable functions and MRA-based wavelets, Journal of Approximation Theory, 161, (2009), 226 -- 238. [35] A.Yu. Khrennikov, V.M. Shelkovich, M. Skopina, p-Adic orthogonal wavelet bases, p-Adic Numbers, Ultrametric Analysis and Applications, 1, no. 2, (2009), 145 -- 156. [36] A. N. Kochubei, Parabolic equations over the field of p-adic numbers, Izv. Akad. Nauk SSSR Ser. Mat., 55, no. 6 (1991), 1312 -- 1330. English transl. in Math. USSR Izv., 39, (1992), 1263 -- 1280 [37] A. N. Kochubei, Schrodinger-type operator over the p-adic number field, Teoret. Mat. Fiz., 86, (1991), no. 3, 323 -- 333; translation in Theoret. and Math. Phys., 86, (1991), no. 3, 221 -- 228. [38] A. N. Kochubei, Pseudo-differential equations and stochastics over non-archimedean fields, Marcel Dekker. Inc. New York, Basel, 2001. 44 A. YU. KHRENNIKOV, A. V. KOSYAK, AND V. M. SHELKOVICH [39] S. V. Kozyrev, Wavelet analysis as a p-adic spectral analysis, Izvestia Ross. Akademii Nauk, Seria Math., 66 no. 2, 149 -- 158; English transl. in Izvestiya: Mathematics, 66 no. 2, 367 -- 376. [40] S. V. Kozyrev, Methods and applications of ultrametric and p-adic analysis: from wavelet Issue 12. Steklov Inst. Math., Moscow, 2008. theory to biophysics, Sovrem. Probl. Mat. http://www.mi.ras.ru/spm/pdf/012.pdf. [41] S. V. Kozyrev, p-Adic pseudodifferential operators and p-adic wavelets, Theor. Math. Physics, 138, no. 3, 1 -- 42; English transl. in Theoretical and Mathematical Physics, 138, no. 3, 322 -- 332. [42] A.V. Kosyak, Regular, quasiregular and induces representations of infinite-dimensional groups, EMS, 453 p. (in preparation). [43] P. I. Lizorkin, Generalized Liouville differentiation and the functional spaces Lp r(En). Imbedding theorems, (Russian) Mat. Sb. (N.S.) 60(102), (1963), 325 -- 353. [44] P. I. Lizorkin, Generalized Liouville differentiation and the multiplier method in the theory of imbeddings of classes of differentiable functions, (Russian) Trudy Mat. Inst. Steklov. Vol. 105, 1969 89 -- 167. [45] P. I. Lizorkin, Operators connected with fractional differentiation, and classes of differentiable functions, (Russian) Studies in the theory of differentiable functions of several variables and its applications, IV. Trudy Mat. Inst. Steklov. Vol. 117, (1972), 212 -- 243. [46] S. Mallat, Multiresolution representation and wavelets, Ph. D. Thesis, University of Pennsylvania, Philadelphia, PA. 1988. [47] S. Mallat, An efficient image representation for multiscale analysis, In: Proc. of Machine Vision Conference, Lake Taho. 1987. [48] Ju. I. Manin, Non-Archimedean integration and p-adic Jacquet-Langlands L-functions, (Russian) Uspehi Mat. Nauk, 31(187), no. 1, (1976), 5 -- 54. Russian Mathematical Surveys, 31:1, (1976), 5 -- 57 [49] Yu. I. Manin and A. A. Panchishkin, Introduction to modern number theory fundamental problems, Ideas and theories, 2nd Edition, Spinger-Verlag, Berlin Heidelberg, 2005. [50] Y. Meyer, Ondelettes and fonctions splines, S´eminaire EDP. Paris. December 1986. [51] Y. Meyer, Principe d'incertitude, bases hilbertiennes et alg`ebres d'op´erateurs, S´eminaire N. Bour- baki, 1985 -- 1986, exp. no. 662, p.209 -- 223. [52] J. von Neumann, On infinite direct products, Compositio mathematica, Vol. 15, no. 1, (1939), 1 -- 77. [53] I. Novikov, V. Protasov, and M. Skopina, Wavelet Theory, Translations of Mathematical Mono- graphs, vol. 239, AMS, 2011. [54] S. G. Samko, Hypersingular integrals and their applications. Taylor & Francis, New York, 2002. [55] S. G. Samko, A. A. Kilbas, and O. I. Marichev, Fractional Integrals and Derivatives and Some of Their Applications. Minsk, Nauka i Tekhnika, 1987 (in Russian). [56] H. H. Schaefer, Topological vector spaces, The Macmillan Co., New York; Collier-Macmillan Ltd., London, 1966. [57] V. M. Shelkovich, M. Skopina, p-Adic Haar multiresolution analysis and pseudo-differential oper- ators, Journal of Fourier Analysis and Applications, Vol. 15, Issue 3, (2009), 366 -- 393. [58] M. H. Taibleson, Harmonic analysis on n-dimensional vector spaces over local fields. I. Basic results on fractional integration, Math. Annalen, 176, (1968), 191 -- 207. [59] M. H. Taibleson, Fourier analysis on local fields, Princeton University Press, Princeton, 1975. [60] V. S. Vladimirov, and T. M. Zapuzhak, Adelic formulas for string amplitudes in fields of algebraic numbers, Lett. Math. Phys., 37, (1996), 232 -- 242. [61] V. S. Vladimirov, Adelic Formulas for Gamma and Beta Functions of Completions of Algebraic Number Fields and Their Applications to String Amplitudes, Izv. Ross. Akad. Nauk, Ser. Mat., 60, (1996), no. 1, 63 -- 86 (in Russian); English translation in Izvestiya: Mathematics, 60, (1996), no. 1, 67 -- 90. [62] V. S. Vladimirov, p-Adic analysis and p-adic quantum mechanics, Ann. of the NY Ac. Sci.: Sym- posium in Frontiers of Math., 1988. [63] V. S. Vladimirov, On the spectrum of some pseudo-differential operators over p-adic number field, Algebra and analysis, 2, no. 6, 1990, 107 -- 124. [64] V. S. Vladimirov, I. V. Volovich, and E. I. Zelenov, p-Adic analysis and mathematical physics. World Scientific Publishing, River Edge, NJ, 1994. [65] A. Weil, Adeles and algebraic groups, Progress in Mathematics, vol. 23, Birkhauser Boston, Mass., 1982, With appendices by M. Demazure and Takashi Ono. WAVELET ANALYSIS ON ADELES AND PSEUDO-DIFFERENTIAL OPERATORS 45 [66] K. Yasuda, Markov Processes on the Adeles and Representations of Euler Products, J. Theoret. Probab. 23, (2010), 748 -- 769. International Center for Mathematical Modelling in Physics and Cognitive Sciences MSI, Vaxjo University, SE-351 95, Vaxjo, Sweden. E-mail address: [email protected] Institute of Mathematics, Ukrainian National Academy of Sciences, 3 Tereshchenkivs'ka, Kyiv, 01601, Ukraine. E-mail address: [email protected] Department of Mathematics, St.-Petersburg State Architecture and Civil Engineer- ing University, 2 Krasnoarmeiskaya 4, 190005, St. Petersburg, Russia. E-mail address: [email protected]
1207.2025	1	1207	2012-07-09T12:54:41	Infinitely divisible metrics and curvature inequalities for operators in the Cowen-Douglas class	[ "math.FA" ]	The curvature $\mathcal K_T(w)$ of a contraction $T$ in the Cowen-Douglas class $B_1(\mathbb D)$ is bounded above by the curvature $\mathcal K_{S^*}(w)$ of the backward shift operator. However, in general, an operator satisfying the curvature inequality need not be contractive. In this note, we characterize a slightly smaller class of contractions using a stronger form of the curvature inequality. Along the way, we find conditions on the metric of the holomorphic Hermitian vector bundle $E_T$ corresponding to the operator $T$ in the Cowen-Douglas class $B_1(\mathbb D)$ which ensures negative definiteness of the curvature function. We obtain a generalization for commuting tuples of operators in the class $B_1(\Omega)$, for a bounded domain $\Omega$ in $\mathbb C^m$.	math.FA	math	INFINITELY DIVISIBLE METRICS AND CURVATURE INEQUALITIES FOR OPERATORS IN THE COWEN-DOUGLAS CLASS SHIBANANDA BISWAS, DINESH KUMAR KESHARI, AND GADADHAR MISRA Abstract. The curvature KT (w) of a contraction T in the Cowen-Douglas class B1(D) is bounded above by the curvature KS ∗(w) of the backward shift operator. However, in gen- eral, an operator satisfying the curvature inequality need not be contractive. In this note, we characterize a slightly smaller class of contractions using a stronger form of the curvature in- equality. Along the way, we find conditions on the metric of the holomorphic Hermitian vector bundle ET corresponding to the operator T in the Cowen-Douglas class B1(D) which ensures negative definiteness of the curvature function. We obtain a generalization for commuting tuples of operators in the class B1(Ω), for a bounded domain Ω in Cm. 2 1 0 2 l u J 9 ] Let H be a complex separable Hilbert space and L(H) denote the collection of bounded linear operators on H. The following important class of operators was introduced in [3]. 1. Introduction Definition 1.1. For a connected open subset Ω of C and a positive integer n, let . A F h t a m [ 1 v 5 2 0 2 . 7 0 2 1 : v i X r a Bn(Ω) = (cid:8) T ∈ L(H) Ω ⊂ σ(T ), ran (T − w) = H for w ∈ Ω, ker(T − w) = H, _w∈Ω dim ker(T − w) = n for w ∈ Ω(cid:9), where σ(T ) denotes the spectrum of the operator T . We recall (cf. [3]) that an operator T in the class Bn(Ω) defines a holomorphic Hermitian vector bundle ET in a natural manner. It is the sub-bundle of the trivial bundle Ω × H defined by ET = {(w, x) ∈ Ω × H : x ∈ ker(T − w)} with the natural projection map π : ET → Ω, π(w, x) = w. It is shown in [3, Proposition 1.12] that the mapping w −→ ker(T − w) defines a rank n holomorphic Hermitian vector bundle ET over Ω for T ∈ Bn(Ω). In [3], it was also shown that the equivalence class of the holomorphic Hermitian vector bundle ET and the unitary equivalence class of the operator T determine each other. Theorem 1.2. The operators T and eT in Bn(Ω) are unitarily equivalent if and only if the corresponding holomorphic Hermitian vector bundles ET and EeT 2000 Mathematics Subject Classification. 47B32, 47B35. Key words and phrases. Cowen-Douglas class, curvature inequality, holomorphic Hermitian vector bundle, ker- are equivalent. nel function, infinite divisibility. The work of S. Biswas was supported by a postdoctoral fellowship funded by the Skirball Foundation via the Center for Advanced Studies in Mathematics at Ben-Gurion University of the Negev. The work of D.K. Keshari was supported in the form of a CSIR Research Fellowship at the Indian Institute of Science. The work of G. Misra was supported in part by UGC - SAP and by a grant from the Department of Science and Technology. 1 2 BISWAS, KESHARI, AND MISRA In general, it is not easy to decide if two holomorphic Hermitian vector bundles are equivalent except when the rank of the bundle is 1. In this case, the curvature K(w) = − ∂2 log k γ(w) k2 ∂w∂w , of the line bundle E, defined with respect to a non-zero holomorphic section γ of E, is a complete invariant. The definition of the curvature is independent of the choice of the section γ: If γ0 is another holomorphic section of E, then γ0 = φγ for some non-zero holomorphic function φ defined on an open subset Ω0 of Ω, consequently the harmonicity of log φ, completes the verification. Let T ∈ B1(Ω). Fix w ∈ Ω and let γ be a holomorphic section of the line bundle ET . From [3, Lemma 1.22], it follows that the vectors γ(w) and ∂γ(w) from a basis of ker(T − w)2. Let NT (w) = T ker(T −w)2 and {γ1(w), γ2(w)} be the basis obtained by applying Gram-Schmidt ortho- normalization to the vectors γ(w) and ∂γ(w). The linear transformation NT (w) has the matrix representation w (cid:19) , NT (w) =(cid:18)w hT (w) 0 where hT (w) =(cid:0) − KT (w)(cid:1)− 1 2 , with respect to the orthonormal basis {γ1(w), γ2(w)}. The curvature KT (w) of an operator T in B1(Ω) is negative. To see this, recall that the curvature may also be expressed (cf.[3, page - 195]) in the form (1.1) KT (w) = − kγ(w)k2kγ′(w)k2 − hγ′(w), γ(w)i2 kγ(w)k4 . Applying Cauchy - Schwarz inequality, we see that the numerator is positive. Let {e0, e1} be an orthonormal set of vectors. Suppose N is a nilpotent linear transformation defined by the rule e1 → a e0, e0 → 0, a ∈ C. Then a determines the unitary equivalence class of N . 0 hT (w) > 0 since we have shown that the curvature KT (w) is negative. Hence the curvature KT (w) is an invariant for the operator T . The non-trivial converse of this statement follows from 0 (cid:17) of the operator T in B1(Ω) is nilpotent. Now, The localization NT (w) − wI2 = (cid:16) 0 hT (w) Theorem 1.2. Thus the operators T and eT in B1(Ω) are unitarily equivalent if and only if NT (w) is unitarily equivalent to NeT (w) for w in Ω. Note that if T ∈ B1(D) is a contraction, that is, kT k ≤ 1, then NT (w) is a contraction for each w ∈ D. Observe that ( a c 0 b ) is a contraction if and only if a ≤ 1 and c2 ≤ (1 − a2)(1 − b2). (1−w2)2 , w ∈ D. The adjoint S∗ of the unilateral Thus kNT (w)k ≤ 1 if and only if KT (w) ≤ − shift operator S is in B1(D). It is easy to see that γS∗(w) = (1, w, . . . , wn, . . .) ∈ ℓ2 +, w ∈ D, is a holomorphic section for the corresponding holomorphic Hermitian line bundle ES∗. The norm kγS∗(w)k2 of the section γS∗ is (1 − w2)−1 and hence the curvature KS∗(w) of the operator S∗ is given by the formula − (1−w2)2 , w ∈ D. We have therefore proved: 1 1 Proposition 1.3. If T is a contractive operator in B1(D), then the curvature of T is bounded above by the curvature of the backward shift operator S∗. We think of the operator S∗ as an extremal operator within the class of contractions in B1(D). This is a special case of the curvature inequality proved in [8]. The curvature inequality is equivalent to contractivity of the operators NT (w), w ∈ Ω, while the contractivity of the operator T is global in nature. So, it is natural to expect that the validity of the inequality Indeed, J. Agler had KT (w) ≤ − (1−w2)2 , w ∈ D, need not force T to be a contraction. 1 INFINITE DIVISIBILITY AND CURVATURE INEQUALITY 3 communicated the existence of an operator T , kT k > 1, in B1(D) with KT (w) ≤ KS∗(w) (cf. [8, Note added in proof]) to G. Misra. Unfortunately, there is a printing error in [8, Note added in proof], which should be corrected by reversing the inequality sign. No explicit example has been written down of this phenomenon. We provide such an example here. The main point of this note is to investigate additional conditions on the curvature, apart from the inequality we have discussed above, which will ensure contractivity. We give an alternative proof the curvature inequality. A stronger inequality becomes apparent from this proof. It is this stronger inequality which, as we will show below, admits a converse. An operator T in the class B1(Ω), as is well-known (cf. [3, pp. 194 ]), is unitarily equivalent to the adjoint M ∗ of the multiplication operator M by the co-ordinate function on some Hilbert space HK of holomorphic functions on Ω∗ := {z ∈ C : ¯z ∈ Ω} possessing a reproducing kernel K. The kernel K is a complex valued function defined on Ω∗ × Ω∗ which is holomorphic in the first variable and anti-holomorphic in the second. In consequence, the map ¯w → K(·, w), w ∈ Ω∗, is holomorphic on Ω. We have K(z, w) = K(w, z) making it Hermitian. It is positive definite in the sense that the n × n matrix (cid:0)(cid:0) K(wi, wj)(cid:1)(cid:1)n i,j=1 is positive definite for every subset {w1, . . . , wn} of Ω∗, n ∈ N. Finally, the kernel K reproduces the value of functions in HK, that is, for any fixed w ∈ Ω∗, the holomorphic function K(·, w) belongs to HK and f (w) = hf, K(·, w)i, f ∈ HK, w ∈ Ω∗. The correspondence between the operator T in B1(Ω) and the operator M ∗ on the Hilbert space HK is easy to describe (cf. [3, pp. 194 ]). Let γ be a non-zero holomorphic section (for bounded domain in C, by Grauert's Theorem, a global section exists) for the operator T acting on the Hilbert space H. Consider the map Γ : H → O(Ω∗), where O(Ω∗) is the space of holomorphic functions on Ω∗, defined by Γ(x)(z) = hx, γ(¯z)i, z ∈ Ω∗. Transplant the inner product from H on the range of Γ. The map Γ is now unitary from H onto the completion of ran Γ. Define K to be the function K(z, w) = Γ(cid:0)γ( ¯w)(cid:1)(z) = hγ( ¯w), γ(¯z)i, z, w ∈ Ω∗. Set Kw(·) := K(·, w). Thus Kw is the function Γ(cid:0)γ( ¯w)(cid:1). It is then easily verified that K has the reproducing property, that is, hΓ(x)(z), K(z, w)iran Γ = h(cid:0)hx, γ(¯z)i(cid:1),(cid:0)hγ( ¯w), γ(¯z)(cid:1)iran Γ = hΓx, Γ(γ( ¯w))iran Γ = hx, γ( ¯w)iH = Γ(x)(w), x ∈ H, w ∈ Ω∗. It follows that kKw(·)k2 = K(w, w), w ∈ Ω∗. Also, Kw(·) is an eigenvector for the operator Γ T Γ∗ with eigenvalue ¯w in Ω: Γ T Γ∗(Kw(·)) = Γ T Γ∗(cid:0)Γ(γ( ¯w))(cid:1) = Γ T γ( ¯w) = Γ ¯w γ( ¯w) = ¯w Kw(·), w ∈ Ω∗. Since the linear span of the vectors {Kw : w ∈ Ω∗} is dense in HK, it follows that Γ T Γ∗ is the adjoint M ∗ of the multiplication operator M on HK . We therefore assume, without loss of generality, that an operator T in B1(Ω) has been realized as the adjoint M ∗ of the multiplication operator M on some Hilbert space HK of holomorphic functions on Ω∗ possessing a reproducing kernel K. 4 BISWAS, KESHARI, AND MISRA Remark 1.4. The contractivity of the adjoint M ∗ of the multiplication operator M on some reproducing kernel Hilbert space HK is equivalent to the requirement that K ‡(z, w) := (1 − z ¯w)K(z, w) is positive definite on D (cf. [1, Corollary 2.37] and [7, Lemma 1]). Suppose that the operator M ∗ is in B1(D). Here is a second proof of the curvature inequality: We have ∂2 ∂w∂ ¯w log K(w, w) = ∂2 ∂w∂ ¯w log 1 (1 − w2) + ∂2 ∂w∂ ¯w log K ‡(w, w), w ∈ D, which we rewrite as KM ∗(w) = KS∗(w) − ∂2 ∂w∂ ¯w log K ‡(w, w), w ∈ D. Recalling that we conclude that ∂2 ∂w∂ ¯w log K ‡(w, w) must be positive (see (1.1)) as long as K ‡ is positive definite, KM ∗(w) ≤ KS∗(w), w ∈ D. The fibre at ¯w of the holomorphic bundle EM ∗ for M ∗ in B1(Ω) is the one-dimensional kernel at ¯w of the operator M ∗ spanned by Kw(·), w ∈ Ω∗. In general, there is no obvious way to define an inner product between the two vectors Kw(·) and ( ∂ ∂ ¯w Kw)(·). However since these [5, Lemma 4.3]), in our special case, there is a vectors belong to the same Hilbert space (cf. natural inner product defined between them. This ensures, via the Cauchy-Schwarz inequality, the negativity of the curvature KT . The reproducing kernel function K of the Hilbert space HK encodes the mutual inner products of the vectors {Kw(·) : w ∈ Ω∗}. The Cauchy-Schwarz inequality, in turn, is just the positivity of the Gramian of the two vectors Kw(·) and ( ∂ ∂ ¯w Kw)(·), w ∈ Ω∗. The positive definiteness of K is a much stronger positivity requirement involving all the derivatives of the holomorphic section Kw(·) defined on Ω∗. We exploit this to show that the curvature function (cid:0) ∂2 ∂z∂ ¯w log K(cid:1)(z, w) is actually negative definite not just negative, whenever K t is assumed to be positive definite for all t > 0. We now construct an example of an operator which is not contractive but its curvature is dominated by the curvature of the backward shift. Expanding the function K(z, w) = 8+8z ¯w−z2 ¯w2 in z ¯w, we see that it has the form 8 + 16z ¯w + 15 z2 ¯w2 1−z ¯w . Therefore, it defines a positive definite kernel on the unit disk D. The monomials 15 for n ≥ 2 forms an orthonormal basis in the corresponding Hilbert space HK . The multiplication zn+1 operator M maps zn kzn+1k . Hence it corresponds to a weighted shift operator W with 15 , 1, 1, . . .}. Evidently, it is not a contraction. (This is the same as saying that the function K ‡(z, w) = 8 + 8z ¯w − z2 ¯w2 is not positive definite.) The operator W is similar to the forward shift S. Since the class B1(D) is invariant under similarity and S ∈ B1(D), it follows that W is in it as well. However, the weight sequence {q 1 zn kznk with k1k2 = 1 16 and kznk2 = 1 kznk to kzn+1k 8 , kzk2 = 1 kznk 2 ,q 16 1−z ¯w − ∂2 ∂w∂ ¯w log K ‡(w, w) = − 8(8 − 4w2 − w4) (8 + 8w2 − w4)2 , w ∈ D, is negative for w < 1. Hence we have shown that KM ∗(w) = − ∂2 w ∈ D, although M ∗ is not a contraction. ∂w∂ ¯w log K(w, w) ≤ KS∗(w), This is not an isolated example, it is easy to modify this example to produce a family of examples parameterized by a real parameter. In the following section, we discuss the case of a commuting tuple T = (T1, . . . , Tm) of operators in B1(Ω), Ω ⊆ Cm, m ≥ 1. Even in this case, as before, it is possible to associate a holomorphic Hermitian bundle ET to the operator tuple T such that the equivalence class of the commuting m-tuple T determines the equivalence class of the bundle ET and conversely. We show that the INFINITE DIVISIBILITY AND CURVATURE INEQUALITY 5 co-efficient matrix KT (w) of the curvature (1, 1) form KT of the holomorphic Hermitian vector bundle ET is negative definite for each w ∈ Ω. The negativity of the curvature provides an alternative proof of the curvature inequality given in [7]. In the third section, we show that the curvature KT is negative definite, that is,(cid:0)(cid:0)KT (wi, wj)(cid:1)(cid:1) is negative-definite for all finite subsets {w1, . . . , wn} of Ω if we impose the additional condition of "infinite divisibility" on the reproducing kernel K. The infinite divisibility of the kernel K requires K t to be positive-definite for all t > 0. In the final section, we give several applications of the positive definiteness of the curvature function to contractivity of operators in the Cowen-Douglas class. 2. Negativity of the curvature in general Let Ω be a bounded domain in Cm. Let T = (T1, . . . , Tm) be a m-tuple of commuting operators on a separable complex Hilbert space H. For x ∈ H, let DT : H → H ⊕ . . . ⊕ H be the operator defined by DT (x) = (T1x, . . . , Tmx). For w = (w1, . . . , wm) ∈ Ω, let T − w denote the operator tuple (T1 − w1, . . . , Tm − wm). The joint kernel of T − w is ∩m j=1 ker(Tj − wj), which is also the kernel of the operator DT −w. Following [5], we say that the commuting tuple T belongs to the Cowen-Douglas class Bn(Ω) if ran DT −w is closed, dim ker DT −w = n for all w in Ω, and the span of {ker DT −w : w ∈ Ω} is dense in H. The class of the corresponding holomorphic Hermitian vector bundle ET = {(w, x) ∈ Ω × H : x ∈ ker DT −w} determines the class of the operator tuple T . As before, if n = 1, then the curvature of ET (cf. [4, 5]) determines the unitary equivalence class of T . If γ is a non-zero holomorphic section of the holomorphic Hermitian line bundle ET defined on some open subset Ω0 ⊆ Ω, then the curvature of the line bundle ET is the (1, 1) form ∂2 log kγ(w)k2 ∂wi∂ ¯wj dwi ∧ d ¯wj, w ∈ Ω0, ∂2 log k γ(w) k2 ∂wi∂ ¯wj (cid:1)(cid:1)m i,j=1, w ∈ Ω0, defined on Ω0. Let (2.1) KT (w) = − mXi,j=1 KT (w) =(cid:0)(cid:0) − denotes the curvature matrix. In general, for a holomorphic Hermitian vector bundle, there are two well-known notions of positivity due to Nakano and Griffiths (cf. [6, page - 338]). These two notions coincide in the case of a line bundle, and one talks of positive line bundle in an unambiguous manner. The following Proposition shows that the line bundle corresponding to a commuting tuple of operators in B1(Ω) is negative. Proposition 2.1. For an operator T in B1(Ω∗), the matrix KT (w) is negative definite for each w ∈ Ω∗. 1 , . . . , M ∗ [5]), it follows that T can be realized as M ∗ = First Proof. Fix w0 ∈ Ω. As before (cf. (M ∗ m) where Mi is the multiplication operator by the co-ordinate function zi on the Hilbert space HK of holomorphic functions on Ω0 ⊆ Ω, w0 ∈ Ω0, possessing a reproducing kernel K with K(w, w) 6= 0 for w ∈ Ω0. The function K0(z, w) = K(w0, w0) 1 2 ϕ(z)−1K(z, w)ϕ(w)−1K(w0, w0) 1 2 is defined on some open neighborhood U × U of (w0, w0), where U is the open set on which K(z, w0) is non-zero and ϕ(z) = K(z, w0) is holomorphic on U . The kernel K0 is said to be normalized at w0([5]). The operator of multiplication by the holomorphic function ϕ−1 then defines a unitary operator from the Hilbert space HK determined by the kernel function K to 6 BISWAS, KESHARI, AND MISRA the Hilbert space HK0 determined by the normalized kernel function K0. This unitary operator intertwines the two multiplication operators on HK and HK0 respectively. Thus KM ∗(w0) is equal to the curvature KM (0) ∗(w0) [5, Lemma 3.9], where M (0) is the m-tuple of multiplication operator by the co-ordinate function zi on the Hilbert space HK0. Let aIJ (z − w0)I ( ¯w − ¯w0)J , z, w ∈ U, I, J ∈ Zm + , be the power series expansion of K0 around the point (w0, w0). Since K0(z, w0) = 1, we have that a00 = 1 and aI0 = 0 for all I with I > 0. Similarly, K0(w0, z) = K0(z, w0) shows that a0J = 0 for all J with J > 0. Also note that if bIJ (z − w0)I ( ¯w − ¯w0)J , z, w ∈ U, I, J ∈ Zm + , then b00 = 1 and bI0 = 0 = b0J for all I, J with I, J > 0. Since γ(w) = K0(·, ¯w), w ∈ U ∗ := ∗ over U ∗, we have K0(z, w) =XI,J K0(z, w)−1 =XI,J {¯z : z ∈ U } is a section of the holomorphic Hermitian line bundle EfM ∂2 log k γ(w) k2 (cid:12)(cid:12)w=w0 K0( ¯w, ¯w))(cid:12)(cid:12)w=w0 (K0( ¯w, ¯w)−1 ∂ ∂wi bIJ ( ¯w − ¯w0)I (w − w0)J )( XI≥1,J ≥0 {(1 + XI,J ≥1 ∂wi∂ ¯wj ∂ ∂ ¯wj ∂ ∂ ¯wj = aεj εi = = aIJ +εi (Ji+1)( ¯w − ¯w0)I (w − w0)J )}(cid:12)(cid:12)w=w0 where εi is the standard unit vector in Cm with 1 at the i-th co-ordinate and 0 elsewhere. On the other hand, we have aεj εi = ∂2K0( ¯w, ¯w) ∂wi∂ ¯wj ∂ ∂wi K0(·, ¯w), ∂ ∂wj Thus for any complex constants α1, . . . , αm, − mXi,j=1 αi ¯αj ∂2 log k γ(w) k2 ∂wi∂ ¯wj = −k αi ∂ ∂wi mXi=1 = h (cid:12)(cid:12)w=w0 (cid:12)(cid:12)w=w0 K0(·, ¯w)i(cid:12)(cid:12)w=w0 K0(·, ¯w)k2(cid:12)(cid:12)w=w0 . ≤ 0. This completes the proof. Second Proof. We show that −KT (w) is the Gramian of a set of n vectors which is explicitly exhibited below. These vectors are ei(w) = Kw ⊗ ∂ ∂ ¯wi Kw − ∂ ∂ ¯wi Kw ⊗ Kw, 1 ≤ i ≤ n, in HK ⊗ HK . Then hei(w), ej (w)i = hKw ⊗ = 2(K(w, w) Thus ∂ ∂ ¯wi Kw − ∂ ∂ ¯wi ∂2K(w, w) ∂wi∂ ¯wj Kw ⊗ Kw, Kw ⊗ − ∂ ∂wi K(w, w) ∂ ¯wj ∂ ∂ ¯wj ∂ Kw − ∂ ∂ ¯wj Kw ⊗ Kwi K(w, w)). ∂2 log k γ(w) k2 ∂wi∂ ¯wj (cid:12)(cid:12)w=w0 = = K(w, w) ∂2 K(w,w) ∂wi∂ ¯wj K(w, w) ∂ ∂ ¯wj − ∂ ∂wi K(w, w)2 K(w, w) hei(w0), ej (w0)i 2K(w0, w0)2 . (cid:12)(cid:12)w=w0 INFINITE DIVISIBILITY AND CURVATURE INEQUALITY This completes the proof. 7 (cid:3) A commuting tuple of operators T = (T1, . . . , Tm) is said to be a row contraction ifPm I. The following characterization of row contractions is well known (cf. [7, Corollary 2]). i=1 TiT ∗ i ≤ Lemma 2.2. Let Bm be the unit ball in Cm and M = (M1, . . . , Mm) be m-tuples of multiplication operator on reproducing kernel Hilbert space with reproducing kernel K. Then M is a row contraction if and only if (1 − hz, wi)K(z, w) is positive definite. 1 Let R∗ m be the adjoint of the joint weighted shift operator on the Drury-Arveson space H 2 m. This is the commuting tuple (M ∗ m which is determined by the reproducing kernel 1−hz,wi , z = (z1, . . . , zm), w = (w1, . . . , wm) ∈ Bm. As in Remark 1.4, using Proposition 2.1 and Lemma 2.2, we obtain a version of curvature inequality for the multi-variate case. It appeared earlier in [7] with a different proof. 1 , . . . , M ∗ m) on H 2 Corollary 2.3. If T = (T1, . . . , Tm) is a row contraction in B1(Bm), then KR∗ positive for each w in the unit ball Bm. m(w) − KT (w) is 3. Infinite divisibility and curvature inequality Starting with a positive definite kernel K on a bounded domain Ω in Cm, it is possible to construct several new positive definite kernel functions. For instance, if K is positive definite then the kernel K n, n ∈ N, is also positive definite. Indeed, a positive definite kernel K is said to be infinitely divisible if for all t > 0, the kernel K t is also positive definite. While the Bergman kernel for the Euclidean ball is easily seen to be infinitely divisible, it is not infinitely divisible for the unit ball (with respect to the operator norm) of the n × n matrices. We give the details for n = 2 in the final Section of this note. The following Lemma shows that if K is positive Lemma 3.1. For any bounded domain Ω in Cm, if K defines a positive definite kernel on Ω, definite then the matrix valued kernel(cid:0)(cid:0) then −K(z, w) =(cid:0)(cid:0) K(z, w)(cid:1)(cid:1)m ∂zi ∂ ¯wj i,j=1 ∂2 ∂2 ∂zi ∂ ¯wj K(z, w)(cid:1)(cid:1)m i,j=1 is positive definite as well. is a positive definite kernel on Ω. Proof. Let ξi = (ξi(1), . . . , ξi(m)), 1 ≤ i ≤ m, be vectors in Cm and u1, . . . , un be an arbitrary set of n points in Ω. Since ¯∂iKw belongs to HK, as shown in [5], it follows that h−K(ui, uj)ξj, ξiiCm = nXi,j = ∂2 ∂wk ∂ ¯wl K(cid:1)(ui, uj)ξj(l)ξi(k) h ∂ ∂ ¯wl Kuj , ∂ ∂ ¯wk KuiiHK ξj(l)ξi(k) nXi,j mXk,l (cid:0) mXk,l nXi,j nXi mXk = k ≥ 0 ξi(k) ∂ ∂ ¯wk Kuik2 HK This completes the proof. (cid:3) Remark 3.2. Even in the case of one variable, the proof of the Lemma given is interesting. In fact, this motivates the proof of the main theorem (Theorem 3.6) in one direction. In particular, it says that if K is a positive definite kernel on a bounded domain Ω ⊂ C, then (cid:0) ∂2 let K(z, w) = P∞ is also a positive definite kernel on Ω. Without loss of generality, assume that 0 is in Ω and m,n amnzm ¯wn be the power series expansion of K around 0. It is shown in [5, ∂z ∂ ¯w K(cid:1)(z, w) 8 BISWAS, KESHARI, AND MISRA Lemma 4.1 and 4.3] that the positivity of the kernel K is equivalent to the positivity of the matrix of Taylor co-efficients of K at 0, namely, a00 a01 a02 a10 a11 a12 ... ... an0 an1 an2 ... · · · · · · . . . · · · a0n a1n ... ann  (m + 1)(n + 1)a(m+1)(n+1)zm ¯wn. Hn(0; K) := ∞Xm,n=0 for each n ∈ Z+. The function ∂2 ∂z ∂ ¯w K(z, w) admits the expansion Therefore, for n ∈ N, Hn−1(0; ∂2 ∂z ∂ ¯w Clearly, for n ∈ N, we have na1n 2na2n ... a11 2a21 ... 2a12 4a22 ... · · · · · · . . . · · · n2ann K) = ∂z ∂ ¯w K)(cid:17) = D(cid:0)Hn(0; K)(cid:1)D, nan1 2nan2 .  01×n 0n×1 Hn−1(0; ∂2 (cid:16) 01×1 where D : Cn+1 → Cn+1 is the linear map which is diagonal and is determined by the sequence ∂2 ∂z ∂ ¯w K) is positive definite for all n ∈ Z+. {0, 1, . . . , k, . . . , n}. It therefore follows that Hn(0; Consequently, ∂2 ∂z ∂ ¯w K is a positive definite kernel. The following Lemma encodes a way to extract scalar valued positive definite kernel from the matrix valued one. Lemma 3.3. If K is a n×n matrix valued positive definite kernel on a bounded domain Ω ⊂ Cm, then for every ζ ∈ Cn, hK(z, w)ζ, ζiCn is also a positive definite kernel on Ω. Proof. Let Kζ(z, w) = hK(z, w)ζ, ζiCn . Let u1, . . . , ul be l points in Ω and αi, 1 ≤ i ≤ l, be scalars in C. From [5], it follows that αiKζ(ui, uj)¯αj = lXi,j lXi,j lXj = k ≥ 0 αi ¯αjhK(·, uj)ζ, K(·, ui)ζiHK ¯αjK(·, uj)ζk2 HK This completes the proof. (cid:3) Definition 3.4. Let G be a real analytic function of w, ¯w for w in some open connected subset Ω in the first variable and anti-holomorphic in the second and restricts to G on the diagonal set {(w, w) : w ∈ Ω}, that is, G(w, w) = G(w, w), w ∈ Ω. If the function G is also positive definite, of Cn. Polarizing G, we obtain a (unique) new function eG defined on Ω × Ω which is holomorphic that is, the n × n matrix(cid:0)(cid:0)eG(wi, ¯wj)(cid:1)(cid:1) is positive definite for all finite subsets {w1, . . . , wn} of Ω, then we say that G is a positive definite function on Ω. INFINITE DIVISIBILITY AND CURVATURE INEQUALITY 9 The curvature K of a line bundle is a real analytic function. We have shown that −K(w), w ∈ Ω ⊂ Cm, is positive definite. However, the following example shows that −K(w) need convention that the positive definiteness of the real analytic function −K(w) is the same as the not be a positive definite function, that is, −eK(w) need not be positive definite! We adopt the positive definiteness of the Hermitian function −eK(w). Example 3.5. Let K(z, w) = 1 +P∞ ∞Xi=1 i=1 aizi ¯wi be a positive definite kernel on the unit disc D. The kernel K then admits a power series expansion some small neighborhood of 0. Consequently, we have log K(z, w) = log(1 + aizi ¯wi) = aizi ¯wi − ∞Xi=1 = a1z ¯w + (a2 − a2 1 i=1 aizi ¯wi)2 (P∞ 3 2 )z2 ¯w2 + (a3 − a1a2 + a3 (P∞ + 2 i=1 aizi ¯wi)3 − · · · 1 3 )z3 ¯w3 + . . . It follows that (cid:0) ∂2 ∂z ∂ ¯w log K(cid:1)(z, w) = a1 + 4(a2 − a2 Thus if we choose 0 < ai, i ∈ N, such that a2 < a2 that ∂2 ∂z ∂ ¯w log K is not positive definite. 1 2 )z ¯w + 9(a3 − a1a2 + a3 1 3 )z2 ¯w2 + . . . 1 2 , then from [5, Lemma 4.1 and 4.3], it follows However we note, for instance, that if K is the function 1 + z ¯w + 1 K t(z, w) = 1 + tz ¯w + t(2t − 1) 4 z2 ¯w2 + · · · is not positive definite for t < 1 2 . 4 z2 ¯w2 +P∞ i=3 zi ¯wi, then It is therefore natural to ask if assuming that K is infinitely divisible is both necessary and sufficient for positive definiteness of the curvature function −K. The following Theorem provides an affirmative answer. Theorem 3.6. Let Ω be a domain in Cm and K be a positive real analytic function on Ω × If K is infinitely divisible then there exist a domain Ω0 ⊆ Ω such that the curvature Ω. is positive definite function on Ω0. Conversely, if the function ∂2 and a infinitely divisible kernel K on Ω0 × Ω0 such that K(w, w) = K(w, w), for all w ∈ Ω0. is a positive definite on Ω, then there exist a neighborhood Ω0 ⊆ Ω of w0 matrix (cid:0)(cid:0) (cid:0)(cid:0) ∂wi ∂ ¯wj ∂2 ∂wi ∂ ¯wj log K(cid:1)(cid:1)m log K(cid:1)(cid:1)m i,j=1 i,j=1 Proof. For each t > 0, assume that K t is positive definite on Ω. This is the same as the positive definiteness of exp t log K, t > 0. Clearly t−1(exp t log K − 1) is conditionally positive definite (An Hermitian kernel L is said to be conditionally positive definite if for every n ∈ N and for i=1 αi = 0, the inequality i,j=1 αi ¯αjL(wi, wj) ≥ 0 holds). By letting t tend to 0, it follows that log K is conditionally every choice n points w1, . . . , wn and complex scalars α1, . . . , αn withPn Pn positive definite. Hence at an arbitrary point in Ω, in particular at w0, the kernel Lw0(z, w) = log K(z, w) − log K(z, w0) − log K(w0, w) + log K(w0, w0) ∂2 ∂wi ∂ ¯wj is positive definite. This is essentially the Lemma 1.7 in [9]. From Lemma 3.1, it follows that the w0 such that log K(z, w0) is holomorphic on Ω0. Hence from the equation above, the curvature Lw0(cid:1)(cid:1) is positive definite on Ω. Note that there exist a neighborhood Ω0 ⊆ Ω of log K(cid:1)(cid:1) is positive definite on Ω0. This proves the Theorem in the forward matrix (cid:0)(cid:0) matrix (cid:0)(cid:0) direction. ∂wi ∂ ¯wj ∂2 10 BISWAS, KESHARI, AND MISRA For the other direction, without loss of generality, assume that w0 = 0. Let K(z, w) be the function obtained by polarizing the real analytic m × m matrix valued function ∂2 ∂wi∂ ¯wj (cid:0)(cid:0) log K(w, w)(cid:1)(cid:1)m i,j=1 defined on some bounded domain Ω in Cm. Suppose that log K has the power series expansion m , ¯wJ = P aIJ zI ¯wJ , where the sum is over all multi-indices I, J of length m and zI = zi1 1 · · · ¯wjm 1 · · · zim m . Then ¯wj1 K(z, w) =XI,J aIJ(cid:0)(cid:0)AIJ (k, ℓ)zI−ǫk ¯wJ−ǫℓ(cid:1)(cid:1), where AIJ (k, ℓ) =(cid:0)(cid:0)ik jℓ(cid:1)(cid:1)m and the sum is again over all multi-indices I, J of size m. Clearly, AIJ can be written as the product D(I) Em D(J), where D(I) and D(J) be the m × m diagonal matrices with (i1, . . . , im) and (j1, . . . , jm) on their diagonal respectively, and Em is the m × m matrix all of whose entries are 1. k,ℓ=1 Let D(z) be the holomorphic function on Ω taking values in the m × m diagonal matrices which has zi in the (i, i) position for z := (z1, z2, . . . , zm) ∈ Ω. If the function K is assumed to be positive definite then eK(z, w) := D(z) K(z, w) D( ¯w) =XI,J aIJ D(I) Em D(J)zI ¯wJ is positive definite on Ω0. Let Λ(I) = {k : 1 ≤ k ≤ m and ik 6= 0}. Consider the m × m matrix E(I, J) defined below: E(I, J)ij =(1 if i ∈ Λ(I) and j ∈ Λ(J), 0 otherwise. Note that if Λ(I) = Λ(J) = {1, . . . , m}, then E(I, J) = Em. Consider the function on Ω0 × Ω0, defined by beK(z, w) = XI,J6=0 aIJ E(I,J) c(I)c(J) zI ¯wJ , kernel on Ω0. To facilitate the proof, we need to fix some notations. where c(I) denotes the cardinality of the set Λ(I). We will prove that beK is a positive definite Let δ be a multi-index of size m. Also let p(δ) =Qm j=1(δj + 1) which is the number of multi- indices I ≤ δ, that is, il ≤ δl, 1 ≤ l ≤ m. As par the notation in [5], given a function L on a domain U × U which is holomorphic in the first variable and antiholomorphic in the second, let Hδ(w0; L) be the p(δ) × p(δ) matrix whose (I, J)-entry is ∂I ¯∂J L(w0,w0) , 0 ≤ I, J ≤ δ. For convenience, one uses the colexicographic order to write down the matrix, that is, I ≤c J if and only if (im < jm) or (im = jm and im−1 < jm−1) or · · · or (im = jm and . . . i2 = j2 and i1 < j1) or I = J. I!J! Let D(I)♯ be the diagonal matrix with the diagonal entry D(I)♯ ℓ ℓ equal to 1 iℓ or 0 according as iℓ is non-zero or zero. Using this notation, we have D(I)♯D(I) Em D(J)D(J)♯ = E(I, J). Let Aδ be the block diagonal matrix, written in the colexicographic ordering, of the form Therefore, in this setup, for any multi-index δ, we have if I = J(6= 0) otherwise. D(I)♯ c(I) 0 (Aδ)IJ = Hδ(0;beK) = AδHδ(0;eK)A∗ δ. INFINITE DIVISIBILITY AND CURVATURE INEQUALITY 11 Clearly Hδ(0;beK) is positive definite since Hδ(0;eK) is so by [5, Lemma 4.1]. Thus from [5, Lemma 4.3], it follows that beK is a positive definite kernel. Let K0 be the scalar function on Ω0 × Ω0 defined by where the sum is over all pairs (I, J) excluding those of the form (I, 0) and (0, J). From Lemma 3.3, it follows that the function K0 is positive definite since it is of the form hbeK(z, w)1, 1i, 1 = (1, . . . , 1). It is evident that K0(z, w) :=X aIJ zI ¯wJ , ∂2 ∂wi∂ ¯wj (cid:16)(cid:16)(cid:0) K0(cid:1)(w, w)(cid:17)(cid:17) = K(w, w), that is, ∂2 ∂wi ∂ ¯wj (log K − K0(w, w)) = 0, 1 ≤ i, j ≤ m, w ∈ Ω0. Therefore, log K − K0 is a real pluriharmonic function on Ω0 and hence there exist a holomor- phic function φ such that log K(w, w) − K0(w, w) = (ℜφ)(w) := φ(w) + φ(w) 2 . (Alternatively, since log K is real analytic, it follows that XI,J aIJ wI ¯wJ = XI,J ¯aIJ wJ ¯wI Equating coefficients, we get aIJ = ¯aJ I for all multi-indices I, J. aI0 = a0I for all multi-indices I. The power series (a00/2) +XI aI0zI In particular, we have defines a holomorphic function ψ on Ω0 such that log K(w, w) − K0(w, w) = ψ(w) + ψ(w).) Thus K(w, w) = exp( φ(w) 2 ) exp(K0(w, w)) exp( φ(w) 2 ), w ∈ Ω0. Let K : Ω0 × Ω0 → C be the function defined by the rule K(z, w) = exp( φ(z) 2 ) exp(K0(z, w)) exp( φ(w) 2 ). For t > 0, we then have K t(z, w) = exp(t φ(z) 2 ) exp(tK0(z, w)) exp(t φ(w) 2 ), z, w ∈ Ω0. By construction K(w, w) = K(w, w), w ∈ Ω0. Since K0 is a positive definite kernel as shown above, it follows from [9, Lemma 1.6] that exp(tK0) is a positive definite kernel for all t > 0 and therefore K t is a positive definite on Ω0 for all t > 0 completing the proof of the converse. (cid:3) 4. Applications Let M ∗ be the adjoint of the commuting tuple of multiplication operators acting on the Hilbert space HK ⊆ O(Ω). Fix a positive definite kernel K on Ω. Let us say that M is infinitely divisible with respect to K if K(z, w)−1K(z, w) is infinitely divisible in some open subset Ω0 of Ω. As an immediate application of Theorem 3.6, we obtain : 12 BISWAS, KESHARI, AND MISRA Theorem 4.1. Assume that the adjoint M ∗ of the multiplication operator M on the reproducing kernel Hilbert space HK belongs to B1(Ω). The function (cid:0)(cid:0) positive definite, if and only if the multiplication operator M is infinitely divisible with respect to K. log(cid:0)K(w, w)−1K(w, w)(cid:1)(cid:1)(cid:1) is ∂2 ∂wi ∂ ¯wj If K is a positive definite kernel on D such that (1 − z ¯w)K(z, w) is infinitely divisible then we say that M ∗ on HK is a infinitely divisible contraction. Here is an example showing that a contraction need not be infinitely divisible. Take K(z, w) = (1 − z ¯w)−1(cid:0)1 + z ¯w + 1 4 z2 ¯w2 + ∞Xn=3 zn ¯wn(cid:1) = 1 + 2z ¯w + (n + 1 4 )zn ¯wn. ∞Xn=2 Clearly K defines the positive definite kernel on D. Since (1−z ¯w)K(z, w) is also positive definite, it follows that the adjoint of multiplication operator M ∗ on HK is contractive. But (cid:0)(1 − z ¯w)K(z, w)(cid:1)t = 1 + tz ¯w + t(2t − 1) 4 z2 ¯w2 + · · · 2 as was pointed out earlier. Hence M ∗ is not infinitely divisible is not positive definite for t < 1 contraction on HK . The following Corollary is a characterization of infinitely divisible contractions in the Cowen- Douglas class B1(D) completing the study begun in [8]. Here, for two real analytic functions G1 and G2 on a domain Ω ⊂ Cm, G1(w) (cid:22) G2(w), w ∈ Ω, means G2 − G1 is a positive definite function on Ω. Corollary 4.2. Let K be a positive definite kernel on the open unit disc D. Assume that the adjoint M ∗ of the multiplication operator M on the reproducing kernel Hilbert space HK belongs to B1(D). The function ∂w ∂ ¯w log(cid:0)(1 − w2)K(w, w)(cid:1) is positive definite, or equivalently KT (w) (cid:22) KS∗(w), w ∈ D, ∂2 if and only if the multiplication operator M is an infinitely divisible contraction. Proof. Recall Theorem 3.6, which says that the positive definiteness of ∂2 ∂w ∂ ¯w log(cid:0)(1 − w2)K(w, w)(cid:1) positive definite for all t ≥ 0. is equivalent to infinite divisibility of the kernel (1 − z ¯w)K(z, w), that is, (cid:0)(1 − z ¯w)K(z, w)(cid:1)t is contraction if (1 − hz, wi)K(z, w) is infinitely divisible, that is,(cid:0)(1 − hz, wi)K(z, w)(cid:1)t is positive We say that a commuting tuple of multiplication operators M is an infinitely divisible row Recall that R∗ m. The following theorem is a characterization of infinitely divisible row contractions. m is the adjoint of the joint weighted shift operator on the Drury-Arveson space definite for all t > 0. H 2 (cid:3) Corollary 4.3. Let K be a positive definite kernel on the Euclidean ball Bm. Assume that the adjoint M ∗ of the multiplication operator M on the reproducing kernel Hilbert space HK belongs , w ∈ Bm, is positive definite, ∂2 to B1(Bm). The function (cid:0)(cid:0) or equivalently ∂wi ∂ ¯wj log(cid:0)(1 − hw, wi)K(w, w)(cid:1)(cid:1)(cid:1)m i,j=1 KM ∗ (w) (cid:22) KR∗ m(w), w ∈ Bm, if and only if the multiplication operator M is an infinitely divisible row contraction. INFINITE DIVISIBILITY AND CURVATURE INEQUALITY 13 We give one last example, namely that of the polydisc Dm. In this case, we say a commuting tuple M of multiplication by the co-ordinate functions acting on the Hilbert space HK is infinitely i=1(1 − zi ¯wi)−1, z, w ∈ Dm, is positive definite for all t > 0. Every commuting tuple of contractions M ∗ need not be infinitely divisible. Let Sm be the commuting m - tuple of the joint weighted shift defined on the Hardy space H 2(Dm). divisible if(cid:0)S(z, w)K(z, w)(cid:1)t, where S(z, w) :=Qm Corollary 4.4. Let K be a positive definite kernel on the polydisc Dm. Assume that the adjoint M ∗ of the multiplication operator M on the reproducing kernel Hilbert space HK belongs to , w ∈ Dm, is positive definite, or ∂2 B1(Dm). The function (cid:0)(cid:0) equivalently ∂wi ∂ ¯wj log(cid:0)S(z, w)K(w, w)(cid:1)(cid:1)(cid:1)m i,j=1 KM ∗(w) (cid:22) KS∗ m(w), w ∈ Dm, if and only if the multiplication operator M is an infinitely divisible m-tuple of contractions. For a second application of these ideas, assume that K is a positive definite kernel on Dm with the property: Ki(z, w) = (1 − zi ¯wi)mK(z, w), 1 ≤ i ≤ m, is infinitely divisible. Then It now follows that Ki(z, w). mYi=1 Ki(z, w)(cid:1) 1 K m(z, w) =(cid:0) mYi=1 (1 − zi ¯wi)(cid:1)−m K(z, w) = S(z, w)(cid:0) mYi=1 i=1 Ki(z, w)(cid:1) 1 m . Let M be the commuting tuple of multiplication operators on the Hilbert space HK , which is contractive since K admits the Szego kernel S as a factor. Clearly, the infinite divisibility of m is positive definite. As pointed out in [7], in Ki, 1 ≤ i ≤ m, implies that (cid:0)Qm consequence, for any polynomial p in m - variables, p(M1, . . . , Mm) = PSp(Sm)S, where S is the invariant subspace of functions vanishing on the diagonal of the Hilbert space H 2 m ⊗ HK ⊆ O(Dm × Dm) and PS is the projection onto the subspace S. A basic question raised in the paper of Cowen and Douglas [3, Section 4] is the determination of non-degenerate holomorphic curves in the Grassmannian. Clearly, a necessary condition for this is the positive definiteness of the curvature function. Thus we have the following corollary to Theorem 3.6. Corollary 4.5. Let E be a holomorphic Hermitian vector bundle of rank 1 over a bounded domain Ω ⊂ Cm. If the curvature K of E is negative definite, then there exists a Hilbert space H and a holomorphic map γ : Ω0 → H, Ω0 open in Ω, such that E is isomorphic to Eγ , where Eγ is the pullback, by the holomorphic map γ : Ω0 → Gr(1, H), of the tautological bundle defined over Gr(1, H). Moreover, the real analytic function hγ(z), γ(w)i defined on Ω0 × Ω0 is infinitely divisible. We finish with an amusing application of the Lemma 1.7 in [9] which is a key ingredient in the proof of Theorem 3.6. Let K be the function on the unit ball B2×2 (with respect to the operator norm) of 2 × 2 matrices, given by the formula K(Z, W ) := det(I − ZW ∗)−1, Z, W ∈ B2×2. The kernel K is normalized at 0 by definition. For δ = (1, 0, 0, 3), the matrix (cid:0)(cid:0) ∂α ¯∂β log K(0, 0) α!β! (cid:1)(cid:1)0≤α,β≤δ 14 BISWAS, KESHARI, AND MISRA ¯∂1 ¯∂3 4 4 log K(0,0) 3!3! is diagonal with ∂1∂3 = −1 < 0 (in fact for δ ≤ 3, the corresponding matrix is diagonal with non-negative entries). Here, δ ≥ µ if and only if δi ≥ µi for all i ∈ {1, . . . , m} and the matrix is written with respect to the colexicographic ordering. From [5, Lemma 4.1 and 4.3], it follows that log K is not positive definite. Hence Theorem 3.6 shows that the function det(I − ZW ∗)−t cannot be positive definite for all t > 0. Of course, a lot more is known. Indeed, the set {0 < t : K(Z, W )t is positive definite} is explicitly determined in [2, Corollary 4.6]. References 1. J. Agler and J. E. McCarthy, Pick interpolation and Hilbert function spaces, Graduate Studies in Mathematics, vol. 44, American Mathematical Society, Providence, RI, 2002. MR 1882259 (2003b:47001) 2. J. Arazy, A survey of invariant Hilbert spaces of analytic functions on bounded symmetric domains, Multivari- able operator theory (Seattle, WA, 1993), Contemp. Math., vol. 185, Amer. Math. Soc., Providence, RI, 1995, pp. 7 -- 65. MR 1332053 (96e:46034) 3. M. J. Cowen and R. G. Douglas, Complex geometry and operator theory, Acta Math. 141 (1978), no. 3-4, 187 -- 261. MR MR501368 (80f:47012) 4. , Operators possessing an open set of eigenvalues, Functions, series, operators, Vol. I, II (Budapest, 1980), Colloq. Math. Soc. János Bolyai, vol. 35, North-Holland, Amsterdam, 1983, pp. 323 -- 341. MR MR751007 (85k:47033) 5. R. E. Curto and N. Salinas, Generalized Bergman kernels and the Cowen-Douglas theory, Amer. J. Math. 106 (1984), no. 2, 447 -- 488. MR MR737780 (85e:47042) 6. J. P Demailly, Complex analytic and differential geometry, http://www-fourier.ujf-grenoble.fr/ de- mailly/manuscripts/agbook.pdf. 7. R. G. Douglas, G. Misra, and J. Sarkar, Contractive Hilbert modules and their dilations, to appear, Israel J. Math. 8. G. Misra, Curvature inequalities and extremal properties of bundle shifts, J. Operator Theory 11 (1984), no. 2, 305 -- 317. MR 749164 (86h:47057) 9. K. R. Parthasarathy and K. Schmidt, Positive definite kernels, continuous tensor products, and central limit the- orems of probability theory, Lecture Notes in Mathematics, Vol. 272, Springer-Verlag, Berlin, 1972. MR 0622034 (58 #29849) (Biswas) Department of Mathematics, Ben Gurion University of the Negev, Be'er Sheva 84105, Israel E-mail address: [email protected] (Keshari) Department of Mathematics, Indian Institute of Science, Bangalore 560012, India E-mail address: [email protected] (Misra) Department of Mathematics, Indian Institute of Science, Bangalore 560012, India E-mail address: [email protected]
1202.4858	2	1202	2012-02-26T14:05:47	Completeness of eigenfunctions of Sturm-Liouville problems with discontinuities at three points	[ "math.FA", "math.CA" ]	In this work, we study discontinuous Sturm-Liouville type problems with eigenparameter dependent boundary condition and transmission conditions at three interior points. A self-adjoint linear operator A is defined in a suitable Hilbert space H such that the eigenvalues of such a problem coincide with those of A. We show that the eigenvalues of the problem are analytically simple, and the eigenfunctions of A are complete in H.	math.FA	math	Completeness of eigenfunctions of Sturm-Liouville problems with discontinuities at three points Erdogan S¸en Department of Mathematics, Faculty of Science and Letters, Namık Kemal University, 59030 Tekirdag, Turkey. e-mail: [email protected] MSC (2010): 34L10, 47E05 Keywords : Sturm-Liouville problems; eigenparameter-dependent bound- ary conditions; transmission conditions; completeness; eigenvalues; eigenfunc- tions. ABSTRACT. In this work, we study discontinuous Sturm-Liouville type problems with eigenparameter dependent boundary condition and transmission conditions at three interior points. A self-adjoint linear operator A is defined in a suitable Hilbert space H such that the eigenvalues of such a problem coincide with those of A. We show that the eigenvalues of the problem are analytically simple, and the eigenfunctions of A are complete in H. 1 Introduction It is well-known that many topics in mathematical physics require the investiga- tion of eigenvalues and eigenfunctions of Sturm-Liouville type boundary value problems. In recent years, more and more researches are interested in the dis- continuous Sturm-Liouville problems with eigenparameter-dependent boundary conditions (see [1 − 4]). The literature on this subject is voluminous and we refer to [5 − 10]. Various physics applications of this kind problem are found in many literatures, including some boundary value problem with transmission conditions that arise in the theory of heat and mass transfer (see [2, 10, 11]). The study of the structure of the solution in the matching region of the layer with the basis solution in the plate leads to consideration of an eigenvalue problem for a second order differential operator with piecewise continuous coefficients and transmission conditions [12]. Sturm-Liouville problems with transmission conditions have been studied by many authors (see [3, 4, 8, 13]). Adjoint and self-adjoint boundary value problems with interface conditions have been stutied in [14, 15]. Such problems with point interactions are also studied in [16]. In this study, we also deal with the class of problems (1) − (9), by means of 1 a combination of the methods [5],[13],[17] and[18]. In Section 1, a self-adjoint linear operator A is defined in a suitable Hilbert space H such that the eigen- values of the problem (1) − (9) coincide with those of A. In Section 2, we prove that the eigenvalues of the problem (1) − (9) are analytically simple. In Section 3, we prove that the eigenfunctions of A are complete in H. Note that each eigenfunction of the original problem and a real number. In this study, we consider a discontinuous eigenvalue problem which consists of Sturm-Liouville equation τ u :=(cid:16)−p (x) u ′ (x)(cid:17)′ + q(x)u (x) = λu (x) (1) on I = [−1, h1) ∪ (h1, h2) ∪ (h2, h3) ∪ (h3, 1] , where p (x) = p2 1 for x ∈ [−1, h1); p (x) = p2 4, for x ∈ (h3, 1] ; p1, p2, p3, p4 are nonzero real constants, q (x) ∈ L1 (I, R) and λ ∈ C is the eigenparameter; with the boundary condition 3, for x ∈ (h2, h3) and p (x) = p2 2, for x ∈ (h1, h2) , p (x) = p2 τ1u := α1u (−1) + α2u ′ (−1) = 0, the eigenparameter-dependent boundary condition τ2u := λhβ ′ 1u (1) − β ′ ′ 2u (1)i +hβ1u (1) − β2u ′ (1)i = 0, and the transmission conditions τ3u := u (h1 + 0) − α3u (h1 − 0) − β3u ′ (h1 − 0) = 0, ′ τ4u := u (h1 + 0) − α4u (h1 − 0) − β4u ′ (h1 − 0) = 0, τ5u := u (h2 + 0) − α5u (h2 − 0) − β5u ′ (h2 − 0) = 0, ′ τ6u := u (h2 + 0) − α6u (h2 − 0) − β6u ′ (h2 − 0) = 0, τ7u := u (h3 + 0) − α7u (h3 − 0) − β7u ′ (h3 − 0) = 0, ′ τ8u := u (h3 + 0) − α8u (h3 − 0) − β8u ′ (h3 − 0) = 0, (2) (3) (4) (5) (6) (7) (8) (9) where the coefficients αi, βi and β out this paper, we assume that θ =(cid:12)(cid:12)(cid:12)(cid:12) and α1 + α2 6= 0. ′ α5 β5 j(cid:0)i = 1, 8, j = 1, 2(cid:1) are real numbers. Through- α6 β6 (cid:12)(cid:12)(cid:12)(cid:12) > 0, ξ =(cid:12)(cid:12)(cid:12)(cid:12) 2 β2 (cid:12)(cid:12)(cid:12)(cid:12) > 0, α8 β8 (cid:12)(cid:12)(cid:12)(cid:12) > 0, α7 β7 1 β1 α3 β3 α4 β4 (cid:12)(cid:12)(cid:12)(cid:12) > 0, γ =(cid:12)(cid:12)(cid:12)(cid:12) ρ =(cid:12)(cid:12)(cid:12)(cid:12) ′ ′ β β 2 2 Operator formulation The relation between a symmetric linear operator A defined in a suitable Hilbert space H and the problem (1) − (9) has been introduced in [13]. Here, we repeat the definition and prove that the operator A is self-adjoint, not only symmetric. We define the inner product in L2 (I) as 1 p2 1 h1Z −1 f1g1 + 1 p2 2θ h2Z f2g2 + h1 1 p2 3θγ 1Z h2 f3g3 + 1 p2 4θγξ 1Z h2 f4g4, ∀f, g ∈ L2 (I) , hf, gi1 = where x ∈ [−1, h1), limx→h1+0 f (x), limx→h1−0 f (x), x = h1, f1 (x) :=(cid:26) f (x), f3 (x) :=  It is easy to verify that (cid:0)L2 (I) , h·, ·i1(cid:1) is a Hilbert space. For simplicity, it is f2 (x) :=  f4 (x) :=(cid:26) limx→h3+0 f (x), limx→h3−0 f (x), limx→h2−0 f (x), limx→h2+0 f (x), denoted by H1. x ∈ (h2, h3) , x = h3, x = h2, f (x) , f (x), f (x), x = h1, x ∈ (h1, h2) , x = h2, x = h3, x ∈ (h3, 1] The inner product in H := H1 ⊕ C is defined by hF, Gi = hf, gi1 + 1 ρθγξ hk for F = (f (x) , h), G = (g (x) , k) ∈ H, where f, g ∈ H1, h, k ∈ C. We define the operator A in H as follows: D (A) :=n(f (x) , h) ∈ H f1, f 3 ∈ ACloc ((h2, h3)) , f4, f ′ ′ f3, f ′ 1 ∈ ACloc ((−1, h1)) , f2, f ′ 2 ∈ ACloc ((h1, h2)) , 4 ∈ ACloc ((h3, 1)) , τ f ∈ H1, τ1f = τ3f = τ4f = τ5f = τ6f, h = β ′ 1f (1) − β ′ ′ 2f (1) ,o AF =(cid:16)τ f, −(cid:16)β1f (1) − β2f ′ (1)(cid:17)(cid:17) for F =(cid:16)f, β ′ 1f (1) − β ′ 2f (1)(cid:17) ∈ D (A). Note that by our assumption on q (x) and Theorem 3.2 in [19], for each 2 are continuous on 4 are continuous on [h3, 1]. 3 are continuous on [h2, h3] and f4, f 1 are continuous on [−1, h1], f2, f (f, h) ∈ D (A), f1, f [h1, h2] , f3, f For simplicity, for (f, h) ∈ D (A) , set ′ ′ ′ ′ N (f ) = β1f (1) − β2f ′ ′ (1) , N (f ) = β ′ 1f (1) − β ′ ′ 2f (1) . So, we can study the problem (1)−(9) in H by considering the operator equation AF = λF . Obviously, we have 3 Lemma 1.1 The eigenvalues of the boundary value problem (1) − (9) co- incide with those of A, and its eigenfunctions are the first components of the corresponding eigenfunctions of A. Lemma 1.2 The domain D (A) is dense in H. Proof. Suppose that F ∈ H is orthogonal to all G ∈ D (A) with respect to the 0 denote the inner product h·, ·i, where F = (f (x) , h), G = (g (x) , k). Let eC∞ set of functions φ (x) = ϕ1 (x) , x ∈ [−1, h1) , ϕ2 (x) , x ∈ (h1, h2) , ϕ3 (x) , x ∈ (h2, h3) , ϕ4 (x) , x ∈ (h3, 1] ,   0 ⊕0 ⊂ D (A) (0 ∈ C), any U = (u (x) , 0) ∈ eC∞ 0 (h1, h2) , ϕ3 (x) ∈ C∞ 0 (h2, h3) and 0 ⊕0 0 [−1, h1), ϕ2 (x) ∈ C∞ where ϕ1 (x) ∈ C∞ ϕ4 (x) ∈ C∞ is orthogonal to F , namely, 0 (h3, 1]. Since eC∞ hF, U i = 1 p2 1 h1Z −1 f (x) u (x)dx + 1 p2 2θ h2Z f (x) u (x)dx + h1 1 p2 3θγ h3Z f (x) u (x)dx h2 + 1 p2 4θγξ 1Z h3 f (x) u (x)dx = hf, ui1 . This implies that f (x) is orthogonal to eC∞ hF, Gi = 1 ρθγξ hk = 0. Thus, h = 0 since k = N So, F = (0, 0). Therefore, D (A) is dense in H. 0 in H1 and hence vanishes. So, (g) can be chosen arbitrarily. ′ Theorem 1.1. The linear operator A is self-adjoint in H. Proof. For all F, G ∈ D (A), (2) implies that f (−1) g and direct calculations using (4) and (5) then yield that ′ (−1)−f ′ (−1) g (−1) = 0, hAF, Gi = hF, AGi + W (f, g; h1 − 0) − W (f, g; −1) + 1 θ W (f, g; h2 − 0) − 1 θ W (f, g; h1 + 0) + 1 θγ W (f, g; 1) − 1 θγξ = hF, AGi , 1 θγξ W (f, g; h3 + 0) − 1 θγ 1 W (f, g; h3 − 0) − W (f, g; h2 + 0) + ρθγξ (cid:16)N (f ) N ′ (g) − N ′ (f ) N (g)(cid:17) where (f, g; x) denotes the Wronskians f (x) g metric. ′ ′ (x) − f (x) g (x). So, A is sym- It remains to show that if hAF, W i = hF, U i for all F = (f, N ′ (f )) ∈ D (A), then W ∈ D (A) and AW = U , where W = (w (x) , h) and U = (u (x) , k), i.e., (i) w1, w 3 ∈ ACloc ((h2, h3)) w4, w (w) = (1) ; (iii) τ1w = τ3w = τ4w = τ5w = τ6w = τ7w = τ8w = 0; (iv) 1w (1) − β β u (x) = τ w; (v) k = −N (w) = −β1w (1) + β2w 4 ∈ ACloc ((h3, 1)) and τ w ∈ H1; (ii) h = N 1 ∈ ACloc ((−1, h1)), w2, w 2 ∈ ACloc ((h1, h2)), w3, w (1). 2w ′ ′ ′ ′ ′ ′ ′ ′ ′ 4 0 ⊕ 0 ⊂ D (A), we obtain (τ f ) wdx + 1 p2 3θγ h3Z h2 (τ f ) wdx + 1 p2 4θγξ 1Z h3 (τ f ) wdx For all F ∈ eC∞ h1Z (τ f ) wdx + 1 p2 1 −1 1 p2 2θ h1 h2Z h2Z h1 = 1 p2 1 h1Z −1 f udx + 1 p2 2θ f udx + 1 p2 3θγ h3Z f udx + h2 1 p2 4θγξ 1Z h3 f udx, namely, hτ f, wi1 = hf, ui1. Hence, by standart Sturm-Liouville theory, (i) and (iv) hold. By (iv), the equation hAF, W i = hF, U i, ∀F ∈ D (A), becomes 1 p2 1 h1Z −1 (τ f ) wdx + 1 p2 2θ h2Z h1 (τ f ) wdx + 1 p2 3θγ h3Z h2 (τ f ) wdx + 1 p2 4θγξ 1Z h3 (τ f ) wdx − N (f ) h ρθγ 1 p2 1 h1Z −1 f (τ w) dx + = So, h2Z f (τ w) dx + 1 p2 2θ h1 1 p2 3θγ h3Z f (τ w) dx + h2 1 p2 4θγξ 1Z h3 f (τ w) dx + ′ N (f ) k ρθγξ . hτ f, wi1 = hf, τ wi1 + ′ N (f ) k + N (f ) h ρθγξ . However, hτ f, wi1 = 1 p2 1 −1 h1Z (cid:16)−p2 h3Z (cid:16)−p2 h1Z h2 −1 1 p2 3θγ = 1 p2 1 ′′ 1f ′′ 3f + q (x) f(cid:17) wdx + + q (x) f(cid:17) wdx + 1 p2 2θ 1 p2 4θγξ f (τ w) dx + h2Z f (τ w) dx + 1 p2 2θ h1 ′′ h1 2f h2Z (cid:16)−p2 1Z (cid:16)−p2 h3Z 1 p2 3θγ h3 h2 4f ′′ + q (x) f(cid:17) wdx+ + q (x) f(cid:17) wdx 1Z 1 p2 4θγξ h3 f (τ w) dx + f (τ w) dx +W (f, w; h1 − 0) − W (f, w; −1) + 1 θ W (f, w; h2 − 0) − 1 θ W (f, w; h1 + 0) + 1 θγ W (f, w; h3 − 0) − 1 θγ W (f, w; h2 + 0) + 1 θγξ W (f, w; 1) − 1 θγξ W (f, w; h3 + 0) = hf, τ wi1 + W (f, w; h1 − 0) − W (f, w; −1) + W (f, w; h2 − 0) − 1 θ 1 θγξ W (f, w; 1) − 1 θγξ 1 θ W (f, w; h1 + 0) W (f, w; h3 + 0) . + 1 θγ W (f, w; h3 − 0) − 1 θγ W (f, w; h2 + 0) + 5 Hence, ′ N (f ) k + N (f ) h ρθγξ = W (f, w; h1 − 0) − W (f, w; −1) + 1 θ W (f, w; h2 − 0) 1 θ W (f, w; h1 + 0) + 1 θγ W (f, w; h3 − 0) − 1 θγ W (f, w; h2 + 0) − + − + − −f −f W (f, w; 1) − W (f, w; h2 + 0) 1 θγξ =(cid:16)f (h1 − 0) w 1 θγξ ′ ′ 1 1 (−1) w (−1)(cid:17) + θ (cid:16)f (h1 + 0) w θγ (cid:16)f (h3 − 0) w θγ (cid:16)f (h2 + 0) w (1) w (1)(cid:17) − 1 ′ ′ ′ ′ (h1 − 0) − f ′ 1 θ (cid:16)f (h2 − 0) w (h1 + 0) − f ′ (h3 − 0) − f (h2 + 0) − f 1 θγξ (cid:16)f (h3 + 0) w (h1 − 0) w (h1 − 0)(cid:17) −(cid:16)f (−1) w ′ (−1) (h2 − 0) w (h2 − 0)(cid:17) ′ (h2 − 0) − f ′ ′ (h1 + 0) w (h1 + 0)(cid:17) (h3 − 0) w (h3 − 0)(cid:17) (h2 + 0) w (h2 + 0)(cid:17) + ′ ′ (h3 + 0) − f ′ ′ 1 θγξ (cid:16)f (1) w (h3 + 0) w (h3 + 0)(cid:17) . (1) (8) ′ ′ ′ By Naimark's Patching Lemma [20], there is an F ∈ D (A) such that f (−1) = (−1) = f (h1 − 0) = f (h1 + 0) = f (h2 − 0) = f (h3 − 0) = f (h3 + 0) = (h2 − 0) = f (h2 + 0) = f f (h3 + 0) = 0, f (1) = β (f ) = 0. Thus, f (1). Namely, (ii) holds. Similarly, one from (8) we obtain h = β proves (v). (h1 − 0) = f (h1 + 0) = f (h2 + 0) = f (h3 − 0) = f 2 and f 1w (1) − β 1. For such an F , N (1) = β 2w ′ ′ ′ ′ ′ ′ ′ ′ ′ ′ ′ ′ ′ ′ ′ ′ ′ ′ ′ ′ ′ ′ ′ (h3 − 0) = α7, f (−1) = 0. Let F ∈ D (A) satisfies f (1) = f (h1 − 0) = f (h2 − 0) = f (−1) = −α1. N It remains to show that (iii) holds. Choose F ∈ D (A) so that f (1) = (1) = f (h1 − 0) = f (h2 − 0) = f (h3 − 0) = (f ) = N (f ) = 0. From (h3 − 0) = 0, f (−1) = α2 and f (1) = (−1) = f (h1 + 0) = f (h2 + 0) = f (h3 + 0) = 0, f (h1 − 0) = (h2 − 0) = (h3 + 0) = ξ. (f ) = 0. By (8), we have w (h1 + 0) = α3w (h1 − 0) + (h2 − 0) and w (h3 + 0) = (1) = (h3 + 0) = 0, f (h1 − 0) = (h2 − 0) = −α6, f (h3 − 0) = β8, (h3 − 0) = −α8 f (h1 + 0) = θ, f (h2 + 0) = γ and f (h3 + 0) = ξ. Then f f (8) , we get α1w (−1) + α2w f (−1) = f −β3, f (h2 − 0) = −β5, f (h3 − 0) = −β7, f α5, f Then N (f ) = N β3w α7w (h3 − 0)+ β7w f (−1) = f β4, f f N (f ) = N w β8w (h3 − 0) . Finally, choose F ∈ D (A) so that f (1) = f (h1 − 0) , w (h2 + 0) = α5w (h2 − 0) + β5w (h1 − 0) = −α4, f (h2 − 0) = β6, f (h2 + 0) = α6w (h2 − 0) + β6w (h1 + 0) = α4w (h1 − 0)+β4w (f ) = 0. From (8) , we obtain w (h3 + 0) = α8w (h3 − 0) + (h2 + 0) = γ and f (h1 − 0) = α3, f (h1 + 0) = θ, f (h2 − 0) and w (h1 + 0) = f (h2 + 0) = f (−1) = f (h3 − 0) . ′ ′ ′ ′ ′ ′ ′ ′ ′ ′ ′ ′ ′ ′ ′ ′ ′ ′ ′ ′ ′ ′ (h1 − 0), Corollary 1.1 The eigenvalues of (1) − (9) are real, and if λ1 and λ2 are two different eigenvalues of (1) − (9), then the corresponding eigenfunctions 6 f (x) and g (x) are orthogonal in the sense of f g + f g + 1 p2 2θ h2Z h1 −1 1 p2 1 h1Z ρθγξ (cid:16)β 1 ′ 1f (1) − β ′ ′ 2f 1g (1) − β ′ ′ 2g 1 p2 3θγ h3Z (1)(cid:17)(cid:16)β h2 ′ f g + 1 p2 4θγξ h3 f g+ 1Z (1)(cid:17) = 0. 3 Simplicity of eigenvalues We consider the initial-value problem ′′ (cid:26) −p2 1u (x) + q(x)u (x) = λu (x) , x ∈ [−1, h1) , u (−1) = α2, u ′ (−1) = −α1. In terms of existence and uniqueness in ordinary differential equation theory, the initial-value problem has a unique solution ϕ1 (x, λ) for every λ ∈ C. Similarly, the initial-value problem has a unique solution ϕ2 (x, λ) for every λ ∈ C.The initial-value problem ′′ −p2 2u (x) + q(x)u (x) = λu (x) , u (h1) = α3ϕ1 (h1, λ) + β3ϕ (h1) = α4ϕ1 (h1, λ) + β4ϕ u ′ ′ 1 (h1, λ) , 1 (h1, λ) ′ x ∈ (h1, h2) , ′′ −p2 3u (x) + q(x)u (x) = λu (x) , u (h2) = α5ϕ2 (h2, λ) + β5ϕ (h2) = α6ϕ2 (h2, λ) + β6ϕ u ′ ′ 2 (h2, λ) , 2 (h2, λ) ′ x ∈ (h2, h3) ,       has a unique solution ϕ3 (x, λ) for every λ ∈ C. Similarly, the initial-value problem ′′ −p2 4u (x) + q(x)u (x) = λu (x) , u (h3) = α7ϕ2 (h2, λ) + β5ϕ (h3) = α8ϕ2 (h2, λ) + β6ϕ u ′ ′ 2 (h2, λ) , 2 (h2, λ) ′ x ∈ (h3, 1] , has a unique solution ϕ4 (x, λ) for every λ ∈ C. For each given x ∈ [−1, h1) , ϕ1 (x, λ) is an entire function of λ; for every x ∈ (h1, h2) , ϕ2 (x, λ) is an entire function of λ; for every x ∈ (h2, h3) , ϕ3 (x, λ) is an entire function of λ and for every x ∈ (h3, 1] , ϕ4 (x, λ) is an entire function of λ. Now we define a function φ (x, λ) on x ∈ [−1, h1) ∪ (h1, h2) ∪ (h2, h3) ∪ (h3, 1] by φ (x, λ) =   ϕ1 (x, λ) , x ∈ [−1, h1) , ϕ2 (x, λ) , x ∈ (h1, h2) , ϕ3 (x, λ) , x ∈ (h2, h3) , ϕ4 (x, λ) , x ∈ (h3, 1] . 7 Obviously φ (x, λ) satisfies (1) , (2) and (4)−(9) . Similarly, we define the function χ (x, λ) = which satisfies (1) , (3) − (9) . χ1 (x, λ) , x ∈ [−1, h1) , χ2 (x, λ) , x ∈ (h1, h2) , χ3 (x, λ) , x ∈ (h2, h3) , χ4 (x, λ) , x ∈ (h3, 1] ,   The Wronskian W (ϕi (x, λ) , χi (x, λ)) (i = 1, 2, 3, 4) are independent of the variable x. Let wi (λ) = W (ϕi (x, λ) , χi (x, λ)) and w (λ) = w1 (λ) , and then we obtain w2 (λ) = θw (λ) , w3 (λ) = θγw (λ) and w4 (λ) = θγξw (λ) . Lemma 2.1 [10] The eigenvalues of the problem (1) − (9) coincide with the zeros of the entire function w (λ). Definition 2.1 The analytic multiplicity of an eigenvalue λ of (1) − (9) is its order as a root of the characteristic equation w (λ) = 0. The geometric multiplicity of an eigenvalue is the dimension of its eigenspace, i.e., the number of its linearly independent eigenfunctions. ′ For convenience, set φ = φ (x, λ) , χ1λ = ∂χ1 Theorem 2.1 The eigenvalues of (1) − (9) are analytically simple. ∂λ , etc. 1λ = ∂χ ∂λ , χ 1 ′ Proof. Let λ = s + it, where s, t ∈ R and i2 = −1. We differentiate the equation τ χ = λχ with respect to λ and have By integration by parts, we get τ χλ = λχλ + χ. hτ χλ, φi1 − hχλ, τ φi1 =(cid:16)χ1λϕ1 θγ (cid:16)χ3λϕ3 3λϕ3(cid:17) h3 h2 + − χ 1 ′ ′ 1 θγξ (cid:16)χ4λϕ4 ′ − χ ′ 1λϕ1(cid:17) h1 −1 + 1 θ (cid:16)χ2λϕ2 ′ − χ ′ 4λϕ4(cid:17) 1 h3 . ′ − χ ′ 2λϕ2(cid:17) h2 h1 + (9) Substituting τ χλ = λχλ + χ and τ φ = λφ into the left side of (9) , we have λ hχλ, φi1 + hχ, φi1 − hχλ, λφi1 = hχ, φi1 + 2it hχλ, φi1 . Moreover, + ′ (cid:16)χ1λϕ1 θγ (cid:16)χ3λϕ3 1 ′ − χ − χ −1 + ′ 1λϕ1(cid:17) h1 3λϕ3(cid:17) h3 ′ h2 + − χ ′ 1 θ (cid:16)χ2λϕ2 θγξ (cid:16)χ4λϕ4 1 ′ ′ h1 2λϕ2(cid:17) h2 4λϕ4(cid:17) 1 ′ − χ h3= 8 χ1λ (h1, λ) ϕ1 ′ ′ χ 1λ (−1, λ) ϕ1 (−1, λ) + ′ (h1, λ) − χ 1λ (h1, λ) ϕ1 (h1, λ) − χ1λ (−1, λ) ϕ1 1 ′ ′ (−1, λ) + (h2, λ) − χ 2λ (h2, λ) ϕ2 (h2, λ) ′ θ (cid:16)χ2λ (h2, λ) ϕ2 2λ (h1, λ) ϕ2 (h1, λ)(cid:17) + ′ −χ2λ (h1, λ) ϕ2 ′ (h1, λ) + χ −χ ′ 3λ (h3, λ) ϕ3 (h3, λ) − χ3λ (h2, λ) ϕ3 ′ (h2, λ) + χ ′ 1 θγ (cid:16)χ3λ (h3, λ) ϕ3 3λ (h2, λ) ϕ3 (h2, λ)(cid:17) + ′ (h3, λ) (1, λ) − χ ′ 4λ (1, λ) ϕ4 (1, λ) − χ4λ (h3, λ) ϕ4 ′ (h3, λ) 1 θγξ (cid:16)χ4λ (1, λ) ϕ4 ′ + χ ′ 4λ (h3, λ) ϕ4 (h3, λ) ′ = α1χ1λ (−1, λ) + α2χ ′ ′ 1 χ + θ (cid:16)χ2λ (h2, λ) ϕ2 2λ (h1, λ) ϕ2 (h1, λ)(cid:17) + θγ (cid:16)χ3λ (h2, λ) ϕ3 θγξ (cid:16)β (1, λ) − β 2ϕ4 − 1 1 ′ ′ ′ 1λ (−1, λ) + χ1λ (h1, λ) ϕ1 ′ (h1, λ) − χ ′ 1λ (h1, λ) ϕ1 (h1, λ) (h2, λ) − χ ′ 2λ (h2, λ) ϕ2 (h2, λ) − χ2λ (h1, λ) ϕ2 ′ (h1, λ) + 1 θγ (cid:16)χ3λ (h3, λ) ϕ3 ′ (h3, λ) − χ3λ (h3, λ) ϕ3 (h3, λ)(cid:17) (h2, λ) − χ ′ 1ϕ4 (1, λ)(cid:17) − ′ ′ 1 3λ (h2, λ) ϕ3 (h2, λ)(cid:17) + θγξ (cid:16)χ4λ (h3, λ) ϕ4 θγξ (cid:16)β 2ϕ4 1 ′ ′ ′ (1, λ) − β ′ 4λ (h3, λ) ϕ4 (h3, λ)(cid:17) (h3, λ) − χ ′ 1ϕ4 (1, λ)(cid:17) . = α1χ1λ (−1, λ) + α2χ 1λ (−1, λ) + Note that ′ w (λ) = α2χ ′ 1λ (−1, λ) + α1χ1λ (−1, λ) Therefore, (9) becomes ′ w (λ) = hχ, φi1 + 2it hχλ, φi1 − 1 θγξ (cid:16)β ′ ′ 2ϕ4 (1, λ) − β ′ 1ϕ4 (1, λ)(cid:17) . (10) Now we consider the simplicity of the eigenvalues of (1)−(9). Let µ be arbitrary zero of w (λ). By Corollary 1.1, µ is real. SInce w (µ) =(cid:12)(cid:12)(cid:12)(cid:12) ϕ1 (x, µ) χ1 (x, µ) ϕ 1 (x, µ) χ ′ ′ 1 (x, µ) (cid:12)(cid:12)(cid:12)(cid:12) = 0, we have ϕ1 (x, µ) = c1χ1 (x, µ) (c1 6= 0) , ϕ2 (x, µ) = c2χ2 (x, µ) (c2 6= 0) , ϕ3 (x, µ) = c3χ3 (x, µ) (c3 6= 0) and ϕ4 (x, µ) = c4χ4 (x, µ) (c4 6= 0) where c1, c2, c3, c4 ∈ C. 9 From ′ ϕ ϕ2 (h1, µ) = c1(cid:16)α3χ1 (h1, µ) + β3χ 2 (h1, µ) = c1(cid:16)α4χ1 (h1, µ) + β4χ ϕ3 (h2, µ) = c2(cid:16)α5χ2 (h2, µ) + β5χ 3 (h2, µ) = c2(cid:16)α6χ2 (h2, µ) + β6χ ϕ4 (h3, µ) = c3(cid:16)α7χ3 (h3, µ) + β7χ 4 (h3, µ) = c3(cid:16)α8χ3 (h3, µ) + β8χ ϕ ϕ ′ ′ ′ ′ ′ ′ ′ 2 (h1, µ) , 1 (h1, µ)(cid:17) = c1χ2 (h1, µ) , 1 (h1, µ)(cid:17) = c1χ 2 (h2, µ)(cid:17) = c2χ3 (h2, µ) , 2 (h2, µ)(cid:17) = c2χ 3 (h3, µ)(cid:17) = c3χ4 (h3, µ) , 3 (h3, µ)(cid:17) = c3χ 3 (h2, µ) , 4 (h3, µ) , ′ ′ ′ ′ we get c1 = c2 = c3 = c4 6= 0. Thus, simple calculations using (10) and the initial values of χ4 at x = 1 give ′ w p2 1 (µ) = c1 h1Z  1 1Z 1 p2 4θγξ + −1 h3 1 p2 2θ h1 h2Z .  χ1 (x, µ)2 dx + χ4 (x, µ)2 dx + ρ θγξ χ2 (x, µ)2 dx + 1 p2 3θγ h3Z h2 χ3 (x, µ)2 dx Note that ρ > 0, θ > 0, γ > 0, ξ > 0 and c1 6= 0, so w analytic multiplicity of µ is one. By Lemma 2.1, the proof is completed. Theorem 2.2 All eigenvalues of (1) − (9) are geometrically simple. (µ) 6= 0. Hence, the ′ Proof. If f and g are two eigenfunctions for an eigenvalue λ∗ of (1) − (9), then (2) implies that f (−1) = Kg (−1) and f (−1) for some constant K ∈ R. By the uniqueness theorem for solutions of ordinary differential equation and the transmission conditions (4)−(9) , we have that f = Kg on [−1, 1] . Thus the geometric multiplicity of λ∗ is one. (−1) = Kg ′ ′ 4 Completeness of eigenfunctions Theorem 3.1 The operator A has only point spectrum, i.e., σ (A) = σρ (A) . Proof. It suffices to prove that if η is not an eigenvalue of A, then η ∈ ρ (A) . Since A is self-adjoint, we only consider a real η. We investigate the equation (A − η) Y = F ∈ H, where F = (f, h) . 10 Let us consider the initial-value problem τ y − ηy = f, x ∈ I, α1y (−1) + α2y ′ (−1) = 0, ′ ′ ′ y (h1 + 0) = α3y (h1 − 0) + β3y (h1 + 0) = α4y (h1 − 0) + β4y y y (h2 + 0) = α5y (h2 − 0) + β5y (h2 + 0) = α6y (h2 − 0) + β6y y y (h3 + 0) = α7y (h3 − 0) + β7y (h3 + 0) = α8y (h3 − 0) + β8y y ′ ′ ′ ′ ′ ′ (h1 − 0) , (h1 − 0) , (h2 − 0) , (h2 − 0) , (h3 − 0) , (h3 − 0) , (11)   Let u (x) be the solution of the equation τ u − ηu = 0 satisfying u (−1) = α2, u ′ (−1) = −α1, u (h1 + 0) = α3u (h1 − 0) + β3u ′ u (h1 + 0) = α4u (h1 − 0) + β4u u (h2 + 0) = α5u (h2 − 0) + β5u ′ u (h2 + 0) = α6u (h2 − 0) + β6u u (h3 + 0) = α7u (h3 − 0) + β7u ′ u (h3 + 0) = α8u (h3 − 0) + β8u ′ ′ ′ ′ ′ ′ (h1 − 0) , (h1 − 0) , (h2 − 0) , (h2 − 0) , (h3 − 0) , (h3 − 0) . In fact, u (x) = u1 (x) , x ∈ [−1, h1) , u2 (x) , x ∈ (h1, h2) , u3 (x) , x ∈ (h2, h3) , u4 (x) , x ∈ (h3, 1] ,   where u1 (x) is the unique solution of the initial-value problem (cid:26) −p2 ′′ + q(x)u = ηu, 1u u (−1) = α2, u x ∈ [−1, h1) , ′ (−1) = −α1; u2 (x) is the unique solution of the problem u3 (x) is the unique solution of the problem     ′′ −p2 + q(x)u = ηu, 2u u2 (h1) = α3u1 (h1) + β3u 2 (h1) = α4u1 (h1) + β4u u ′ ′ 1 (h1) , 1 (h1) ; ′ x ∈ (h1, h2) , ′′ −p2 + q(x)u = ηu, 3u u3 (h2) = α5u2 (h2) + β5u 3 (h2) = α6u2 (h2) + β6u u ′ ′ 2 (h2) , 2 (h2) . ′ x ∈ (h2, h3) , 11 and u4 (x) is the unique solution of the problem   Let ′′ −p2 4u + q(x)u = ηu, x ∈ (h3, 1] , u3 (h3) = α7u3 (h3) + β7u 3 (h3) = α8u3 (h3) + β8u u ′ ′ 3 (h3) , 3 (h3) . ′ w (x) = w1 (x) , x ∈ [−1, h1) , w2 (x) , x ∈ (h1, h2) , w3 (x) , x ∈ (h2, h3) , w4 (x) , x ∈ (h3, 1] ,   be a solution of τ w − ηw = f satisfying α1w (−1) + α2w ′ (−1) = 0, ′ w (h1 + 0) = α3w (h1 − 0) + β3w (h1 + 0) = α4w (h1 − 0) + β4w w w (h2 + 0) = α5w (h2 − 0) + β5w (h2 + 0) = α6w (h2 − 0) + β6w w w (h3 + 0) = α7w (h3 − 0) + β7w (h3 + 0) = α8w (h3 − 0) + β8w w ′ ′ ′ ′ ′ ′ ′ ′ (h1 − 0) , (h1 − 0) , (h2 − 0) , (h2 − 0) , (h3 − 0) , (h3 − 0) . Then, (11) has the general solution y (x) =   du1 + w1, x ∈ [−1, h1) , du2 + w2, x ∈ (h1, h2) , du3 + w3, x ∈ (h2, h3) , du4 + w4, x ∈ (h3, 1] , where d ∈ C. Since η is not an eigenvalue of the problem (1) − (7) , we have ηhβ ′ 1u2 (1) − β ′ 2u ′ 2(1)i +hβ1u2 (1) − β2u ′ 2(1)i 6= 0. The second component of (A − η) Y = F involves the equation −N (y) − ηN ′ (y) = h, namely, Substituting (12) into (14), we get h−β1y (1) + β2y (cid:16)β2u ′ ′ (1)i − ηhβ ′ 1y (1) − β ′ ′ 2y (1)i = h. 1u2 (1)(cid:17) d ′ ′ ′ 2 (1) − β1u2 (1) + ηβ ′ 2u ′ 2 (1) − ηβ = h + β1w2 (1) − β2w 2 (1) + ηβ 1w2 (1) − ηβ 2w 2 (1) ′ ′ (12) (13) (14) In view of (13), we know that d is uniquely solvable. Therefore, y is uniquely determined. 12 The above arguments show that (A − ηI)−1 is defined on all of H, where I is identity matrix. We obtain that (A − ηI)−1 is bounded by Theorem 1.1 and the Closed Graph Theorem. Thus, η ∈ ρ (A) . Therefore, σ (A) = σρ (A) . Lemma 3.1 [10] The eigenvalues of the boundary value problem (1) − (9) are bounded below, and they are countably infinite and can cluster only at ∞. For every δ ∈ R \ σρ (A) , we have the following immediate conclusion. Lemma 3.2 Let λ be an eigenvalue of A−δI, and V a corresponding eigen- function. Then, λ−1 is an eigenvalue of (A − δI)−1 , and V is a corresponding eigenfunction. The converse is also true. On the other hand, if µ is an eigenvalue of A and U is a corresponding eigenfunction, then µ − δ is an eigenvalue of A − δI, and U is a corresponding eigenfunction. The converse is also true. Accordingly, the discussion about the completeness of the eigenfunctions of A is equivalent to considering the corresponding property of (A − δI)−1. By Lemma 1.1, Lemma 3.1 and Corollary 1.1, we suppose that {λn; n ∈ N} is the real sequence of eigenvalues of A, then {λn − δ; n ∈ N} is the sequence of eigenvalues of A − δI. We may assume that λ1 − δ ≤ λ2 − δ ≤ ... ≤ λn − δ ≤ ... → ∞. Let {µn; n ∈ N} be the sequence of eigenvalues of (A − δI)−1 . Then µn = (λn − δ)−1 and µ1 ≥ µ2 ≥ ... ≥ µn ≥ ... → 0. Note that 0 is not an eigenvalue of (A − δI)−1 . Theorem 3.2 The operator A has compact resolvents, i.e, for each δ ∈ R \ σρ (A) , (A − δI)−1 is compact on H. Proof. Let {µ1, µ2, ...} be the eigenvalues of (A − δI)−1 , and {P1, P2, ...} the orthogonal projections of finite rank onto the corresponding eigenspaces. Since {µ1, µ2, ...} is a bounded sequence and all Pn′s are mutually orthogonal, we n=1 µnPn is strongly convergent to the bounded operator (A − δI)−1 , n=1 µnPn. Because for every α > 0, the number of µn′s satisfying µn > α is finite, and all Pn′s are of finite rank, we obtain that (A − δI)−1 is compact. have P∞ i.e., (A − δI)−1 = P∞ In terms of the above statements and the spectral theorem for compact operators, we obtain the following theorem. Theorem 3.3 The eigenfunctions of the problem (1) − (9) , augmented to become eigenfunctions of A, are complete in H, i.e., if we let nΦn =(cid:16)φn (x) , N ′ (φn)(cid:17) ; n ∈ No be a maximum set of orthonormal eigenfunctions of A, where {φn (x) ; n ∈ N} are eigenfunctions of (1) − (9) , then for all F ∈ H, F =P∞ n=1 hF, Φni Φn. 13 References [1] C. T. Fulton, Two-point boundary value problems with eigenvalue parameter contained in the boundary conditions. Proc. Royal Soc. Edinburgh 77A (1977), 293-308. [2] A. V. Likov, Y. A. Mikhalilov, The Theory of Heat and Mass Transfer, Qosener- gaizdat, 1963. (In Russian). [3] E. S¸en, A. Bayramov, Calculation of eigenvalues and eigenfunctions of a discontin- uous boundary value problem with retarded argument which contains a spectral parameter in the boundary condition, Mathematical and Computer Modelling, 54 (2011) 3090-3097. [4] E. S¸en, A. Bayramov, On calculation of eigenvalues and eigenfunctions of a Sturm-Liouville type problem with retarded argument which contains a spectral parameter in the boundary condition, Journal of Inequalities and Applications, Vol. 2011 (1) (2011) 1-9. [5] P. A. Binding, P. J. Browne, B. A. Watson, Sturm-Liouville problems with bound- ary conditions rationally dependent on the eigenparameter, II, J. Comput. Appl. Math. 148 (2002), 147-168. [6] P. A. Binding, Patrick. J. Browne, Oscillation theory for indefinite Sturm- Liouville problems with eigenparameter dependent boundary conditions. Proc. Royal Soc. Edinburg 127A (1997), 1123-1136. [7] P. A. Binding, R. Hryniv, H. Langer, Elliptic eigenvalue problems with eigen- parameter dependent boundary conditions. J. Differential Equations 174 (2001), 30-54. [8] M. Demirci, Z. Akdogan, O. Sh. Mukhtarov, Asymptotic behavior of eigenvalues and eigenfunctions of one discontinuous boundary-value problem. International J. Computational Cognition 2(3) (2004), 101-113. [9] D. B. Hinton, An expansion theorem for an eigenvalue problem with eigenvalue parameter in the boundary condition. Quart. J. Math. (Oxford) 30 (1979), 33-42. [10] M. Kadakal, O.Sh. Mukhtarov, Sturm-Liouville problems with discontinuities at two points, Comput. Math. Appl.,54 (2007) 1367-1379. [11] O. Sh. Mukhtarov, S. Yakubov, Problems for differential equations with trans- mission conditions, Applicable Anal. 81 (2002), 1033-1064. [12] I. Titeux, Y. Yakubov, Completeness of root functions for thermal conduction in a strip with piecewise continuous coefficients, Math. Models Methods Appl. Sc. 7 (1997), 1035-1050. [13] O. Sh. Mukhtarov, M. Kadakal, Spectral properties of one Sturm-Liouville type problem with discontinuous weight, (Russian) Sibirsk. Math. Zh., 46 (4) (2005), 681-694. 14 [14] Y. S. Li, J. Sun, Z. Wang, On the complete description of self-adjoint boundary conditions of the Schrodinger operator with a δ (x)or δ (x)interaction, Sympo- sium of the Fifth Conference of Mathematics Society of Inner Mongolia, Inner Mongolia Univ. Press, 1995, 27-30. ′ [15] A. Zettl, Adjoint and self-adjoint boundary value problems with interface condi- tions, SIAM J. Appl. Math. 16 (1968), 851-859. [16] F. Gesztesy, W. Kirsch, One-dimensional Schrodinger operators with interactions on a discerete set, J. Reine Angew. Math. 362 (1985), 28-50. [17] A. Wang, J. Sun, P. Gao, Completeness of Eigenfunctions of Sturm-Liouville Problems with Transmission Conditions, J. Spectral Math. Appl. (2006) [18] A. Wang, J. Sun, X. Hao, S. Yao, Completeness of Eigenfunctions of Sturm- Liouville Problems with Transmission Conditions, Methods and Application of Analysis 16 (3) (2009), 299-312. [19] J. Weidmann, Spectral Theory of Ordinary Differential Operators, Lecture Notes in Math. 1258, Springer-Verlag, Berline, 1987. [20] M. A. Naimark, Linear Differential Operators, part II. Harrap, London, 1968. 15
1601.00612	1	1601	2015-12-30T16:48:29	A note on the superadditive and the subadditive transformations of aggregation functions	[ "math.FA" ]	We expand the theoretical background of the recently introduced superadditive and subadditive transformations of aggregation functions $A$. Necessary and sufficient conditions ensuring that a transformation of a proper aggregation function is again proper are deeply studied and exemplified. Relationships between these transformations are also studied.	math.FA	math	A note on the superadditive and the subadditive transformations of aggregation functions Alexandra Siposov´a Slovak University of Technology Faculty of Civil Engineering Department of Mathematics and Descriptive Geometry Radlinsk´eho 11, 810 05 Bratislava Slovakia [email protected] Abstract We expand the theoretical background of the recently introduced superadditive and subadditive transformations of aggregation functions A. Necessary and sufficient conditions ensuring that a transformation of a proper aggregation function is again proper are deeply studied and exemplified. Relationships between these transforma- tions are also studied. Keywords: aggregation function, subadditive transformation, superadditive transfor- mation 1 Introduction Motivated by applications in economics, subadditive and superadditive transformations of aggregation functions on R+ = [0,∞[ have been recently introduced in [4]. Formally, both these transformations can be introduced on the improper real interval [0,∞]. Definition 1 A mapping A : [0,∞]n → [0,∞] is called an (n-ary) aggregation function if A(0, . . . , 0) = 0 and A is increasing in each coordinate. Further, A is called a proper (n-ary) aggregation function if it satisfies the following two additional constraints: (i) A(x) ∈ ]0,∞[ for some x ∈ ]0,∞[n, (ii) A(x) < ∞ for all x ∈ [0,∞[n. Though for real applications we only need proper aggregation functions (in fact, their restriction to the domain [0,∞[n), a broader framework of all (n-ary) aggregation function is of advantage in a formal description of our results, making formulations and expressions more transparent. Observe that our framework is broader than the concept of aggregation functions on [0,∞] as introduced in [1], [3], which does not cover Sugeno integral based 1 A∗(x) = inf { k(cid:88) A∗(x) = sup { (cid:96)(cid:88) k(cid:88) (cid:96)(cid:88) aggregation functions, for example. We denote the class of all n-ary aggregation function by An, and the class of all n-ary proper aggregation function by Pn. The next definition was motivated by optimization tasks treated in linear programming area and related areas [2], as well as by recently introduced concepts of concave [5] and convex [6] integrals. Definition 2 For every A ∈ An the subadditive transformation A∗ : [0,∞]n → [0,∞] of A is given by (1) Similarly, for every A ∈ An the superadditive transformation A∗ : [0,∞]n → [0,∞] of A is defined by i=1 i=1 A(y(i)) y(i) ≥ x} A(y(j)) y(j) ≤ x} . (2) j=1 j=1 n, where K∗ Observe that the transformation (1) was originally introduced in [4] for A ∈ Kn∗ , where Kn∗ is the class of all n-ary proper aggregation functions (restricted to [0,∞[n) such that also A∗ is proper, that is, A∗ ∈ Pn. Similarly, A∗ given by (2) was originally introduced in [4] only for A ∈ K∗ n is the class of all A ∈ Pn (restricted to [0,∞[n), so that A∗ ∈ Pn as well. Theorem 2 in [4] gives a necessary and sufficient condition ensuring that a function A ∈ Pn has also the property that A ∈ K∗ n. We develop this result, giving an equivalent condition. Moreover, we also characterize all the functions A ∈ Pn such that A ∈ Kn∗ . Our approach is based on a deep study of transformations (1) and (2) on unary aggregation functions that belong to P1. Our approach allows to show that for any A ∈ Pn we have the inequality (A∗)∗ ≤ (A∗)∗. 1 and K1∗ are completely described, showing that the properties in a neighbourhood of 0 are important for characterization of elements of these classes. In Section 3, necessary and sufficient conditions for a function A ∈ Pn to belong to K∗ n, or to Kn∗ , are given. Section 4 is devoted to the study of relationships of transformations (A∗)∗ and (A∗)∗. Finally, some concluding remarks are added. In the next section, the classes K∗ The paper is organized as follows. 2 The one-dimensional case We begin with basic results which show how the values of the subadditive and superadditive transformations of one-dimensional aggregation functions depend on the behavior of the functions near zero. 2 Theorem 1 Let h be an unary aggregation function on [0,∞] with lim inf t→0+ h(t)/t = a and lim supt→0+ h(t)/t = b, where 0 ≤ a ≤ b ≤ ∞. Then, for every x ∈ ]0,∞[ we have h∗(x) ≤ ax and h∗(x) ≥ bx. Proof. Let x > 0. By definitions of h∗ and h∗, for every positive integer n we have h∗(x) ≤ nh(x/n) ≤ h∗(x), that is, h∗(x) ≤ x · h( x n) x n ≤ h∗(x) . (3) Since h is increasing, for every t such that n we have x n+1 ≤ t ≤ x ≤ h( x n) x ≤ h(t) t . n+1 h( x n+1) x n Applying the limits inferior and superior to these inequalities as t → 0+ and n → ∞ (with n → 1) shows that n+1 lim inf n→∞ h( x n) x n ≤ lim inf t→0+ h(t) t and lim sup n→∞ h( x n) x n ≥ lim sup t→0+ h(t) t . Combining (3) with (4) now gives h∗(x) ≤ x · lim inf t→0+ h(t) t = ax and h∗(x) ≥ x · lim sup t→0+ h(t) t = bx for every x > 0, which completes the proof. (4) 2 h(t) t and lim supt→0+ h(t) t The values of lim inf t→0+ can be interpreted as the ‘lower’ and ‘upper’ slope of h at the point x = 0. The previous result may therefore be interpreted by saying that the values of h∗ and h∗ are to a large extent influenced by the values of the lower and upper slopes of h at 0. Corollary 1 Suppose that h is an unary aggregation function on [0,∞] such that the (cid:48) derivative h (0+) exists and is equal to c ∈ [0,∞[. (1) If h is convex on [0,∞[, then h∗(x) = cx and h∗(x) = h(x) for every x ≥ 0. (2) If h is concave on [0,∞[, then h∗(x) = h(x) and h∗(x) = cx for every x > 0. Proof. For (1) it is sufficient to realize that h(x) ≥ ax for every x ≥ 0, the claim then follows from Theorem 1 regarding h∗ and from [4] regarding h∗. The proof of (2) is similar 2 and therefore omitted. 3 Figure 1: A schematic drawing of a function h from the proof of Corollary 2. Corollary 2 For any real a, b such that 0 < a < b < ∞ there is an infinite number of smooth unary aggregation functions h on [0,∞] such that h∗(x) = ax and h∗(x) = bx for each x ∈ [0,∞[. Proof. Let q be a positive real number such that q < a/b < 1; note that bq2j < aq2j−1 for every positive integer j. Results from calculus now imply the existence of infinitely many smooth increasing functions h(x) defined on [0,∞[ such that ax ≤ h(x) ≤ bx for every x ∈ [0, +∞[, h(q2j−1) = aq2j−1 and h(q2j) = bq2j for every positive integer j. Since ax ≤ h(x) ≤ bx for x ∈ [0,∞[, we obviously have ax ≤ h∗(x) and h∗(x) ≤ bx for every x ≥ 0. But we also have lim inf t→0+ h(t)/t = a and lim supt→0+ h(t)/t = b, because of the values of h at points in the sequences (q2j−1)∞ j=1, respectively. By Theorem 1 we have h∗(x) ≤ ax and h∗(x) ≥ bx for each x ≥ 0, completing the proof. 2 j=1 and (q2j)∞ Observe that the functions h from Corollary 2 have the property that (h∗)∗(x) = ax < bx = (h∗)∗(x) for all x > 0. Corollary 3 There is an infinite number of smooth aggregation functions h on [0,∞] such that h∗(x) = 0 for every x < ∞ and h∗(x) = ∞ for every x > 0. Proof. For every positive integer k let xk = 2−2k. For x ≥ 0 let f (x) = x5/4 and g(x) = x3/4; a straightforward calculation shows that g(x2j) < f (x2j−1) for every positive integer j. By known results from calculus there exists an infinite number of smooth increasing functions h on [0,∞[ such that h(x2j−1) = f (x2j−1) and h(x2j) = g(x2j) for all positive integers j. Since for our function h we have lim inf t→0+ h(t)/t = 0 and lim supt→0+ h(t)/t = +∞ j=1, the result is again 2 due to the values of h at points in the sequences (x2j−1)∞ a consequence of Theorem 1. j=1 and (x2j)∞ 4 bxaxqqqq2j - 12j 2j + 12j + 20.51.01.551015202530Figure 2: A schematic drawing of a function h from the proof of Corollary 3. The functions h from Corollary 3 have even a more striking property that (h∗)∗(x) = 0 for every x ≥ 0 while (h∗)∗(x) = ∞ for all x > 0. To conclude this section we underscore the fundamental role of Theorem 1 by the following complete characterization of one-dimensional degeneracies. Theorem 2 Let h be a one-dimensional aggregation function on [0,∞[. The following conditions are equivalent: (a) There exists an x > 0 for which h∗(x) = ∞. (b) h∗(x) = ∞ for every x > 0. (c) lim supt→0+ (d) sup{ h(t) t ∈]0, x]} = ∞ for some x > 0. Similarly, the following statements are equivalent: t = ∞. h(t) t (a’) There exists an x > 0 for which h∗(x) = 0. (b’) h∗(x) = 0 for every x ≥ 0. (c’) lim inf t→0+ (d’) inf{ h(t) t ∈]0, x]} = 0 for some x > 0. h(t) t = 0. t Proof. Clearly, the statements (c) and (d) are equivalent, and so are (c’) and (d’). By Theorem 1, (c) implies (b) and (c’) implies (b’). Trivially, (b) implies (a) and (b’) implies (a’), and so one only has to prove that (a) implies (d) and (a’) implies (d’). To show that (a) implies (d), we prove the contrapositive. Assume that sup{ h(t) t ∈ ]0, x]} = b < ∞ for every x > 0. This means that h(t) ≤ bt for every t ∈]0, x]. Thus, i=1 xi ≤ x we have for every n-tuple (x1, x2, . . . , xn) of positive real numbers such that(cid:80)n (cid:80)n i=1 h(xi) ≤ b(cid:80)n i=1 xi ≤ bx. It follows that h∗(x) ≤ bx < ∞ for every x > 0. t 5 gfxxxx2j - 12j 2j + 12j + 20.10.20.30.40.50.61234Similarly, to show that (a’) implies (d’) we again proceed by proving the contrapositive. Suppose that inf{ h(t) t ∈]0, x]} = b > 0 for every x > 0. This means that h(t) ≥ bt for every t ∈]0, x]. Thus, for every n-tuple (x1, x2, . . . , xn) of real numbers from ]0, x] such i=1 xi ≥ bx. It follows that h∗(x) ≥ bx > 0 for i=1 xi ≥ x we have(cid:80)n i=1 h(xi) ≥ b(cid:80)n that(cid:80)n t every x > 0. This completes the proof. 2 3 The multidimensional case Based on the results in the one-dimensional case proved in Section 2 we continue by ex- hibiting examples of aggregation functions A with the property that the values of (A∗)∗ are smaller than the values of (A∗)∗ for all non-zero vectors x ∈ [0,∞[n. The method is based on the observation that the values of A∗ and A∗ on the entire space [0,∞[n are influenced by the behavior of the one-dimensional diagonal function A(x, . . . , x) in an arbitrarily small neighbourhood of zero. We present details only regarding extensions of Corollary 3 to arbitrary dimensions. Let µx be the arithmetic mean of the entries in x, that is, if x = (x1, x2, . . . , xm), then µx = (x1 + x2 + . . . + xm)/m. Theorem 3 There are infinitely many aggregation functions A defined on [0,∞]n such that (A∗)∗(x) = 0 for all x ∈ [0,∞]n while (A∗)∗(x) = ∞ for all non-zero x ∈ [0,∞[n. Proof. It is sufficient to take any function h from Corollary 3 and define A for every x ∈ [0,∞[n by A(x) = h(µx). Letting x = µx, for any non-zero x we have, exactly as in the proof of Theorem 1, h(t) A∗(x) ≤ nh(x/n) ≤ x · lim inf t→0+ t A∗(x) ≥ nh(x/n) ≥ x · lim sup t→0+ h(t) t . , and The claim now follows from Corollary 3. 2 Clearly, one can use numerous other compositions of functions h from Corollary 3 with simple aggregation functions (such as weighted average, geometric means, etc.) to provide examples for Theorem 3. We also prove different sufficient conditions for the values of A∗ and A∗ to exhibit the extreme behavior described in Theorem 3. Proposition 1 Let A be an aggregation function on [0,∞]n and let A(x) = A(x, x, . . . , x) for every x ≥ 0. 6 (1) If there exists a divergent series (cid:80)∞ the series(cid:80)∞ (2) If there exists a convergent series(cid:80)∞ the series(cid:80)∞ j=1 j=1 A(aj) converges, then A∗(x) = 0 for every x ∈ [0,∞[n. j=1 aj with decreasing positive terms such that A(aj) diverges, then A∗(x) = +∞ for every non-zero x ∈ [0,∞]n. j=1 aj with decreasing positive terms such that 2 Proof. Let h(x) = A(x, x, . . . , x) for x ≥ 0. We show that the assumption of (1) implies j aj would imply divergence of (cid:80) divergence of (cid:80) h(aj) ≤ (c + )aj for all but finitely many j’s. But then convergence of(cid:80) convergence of(cid:80) that lim inf j→∞ h(aj)/aj = 0 and (2) implies that lim supj→∞ h(aj)/aj = +∞. Indeed, suppose that lim inf j→∞ h(aj)/aj = c > 0. This means that for every > 0 we have h(aj) ≥ (c − )aj for all but a finite number of positive integers j. But then, j h(aj), contrary to the assumption of (1). Similarly, if lim supj→∞ h(aj)/aj = c < +∞, then for every > 0 we would have j aj would imply This shows that lim inf t→0+ h(t)/t = 0 in the case (1) and lim supt→0+ h(t)/t = +∞ in j h(aj), a contradiction. the case (2). The result now follows from Theorems 1 and 3. or f (x) = x/ ln2(x), and take for (cid:80)∞ an arbitrarily small λ > 0, or g(x) = x ln2(x), and take for(cid:80)∞ As examples, one can take the function f (x) = x1+λ for some arbitrarily small λ > 0, j=1 aj the harmonic series to construct aggregation functions A such that A(x) = f (x) on an arbitrarily small interval (0, δ); the result of (1) then gives A∗(x) = 0 on [0,∞[n. Similarly, one can take the function g(x) = x1−λ for j=1 aj the series with aj = g−1(1/j) to obtain aggregation functions A such that A(x) = g(x) on an arbitrarily small interval (0, δ); the result of (2) then shows that A∗(x) = +∞ for all non-zero x ∈ [0,∞[n. n and Kn∗ by looking at the one-dimensional case. For an aggregation function A on [0,∞]n let A be defined on [0,∞] by letting A(x) = A(x, x, . . . , x) and, for every i ∈ {1, 2, . . . , n} let Ai be defined on [0,∞] by Ai(x) = A(xei), where ei is the i-th unit vector. Theorem 4 Let A be an aggregation function on [0,∞]n. Then, With the help of the results of Section 2 we can also decide membership in K∗ (i) A[0,∞[n∈ K∗ (ii) A[0,∞[n∈ Kn∗ if and only if Ai[0,∞[∈ K1∗ for some i ∈ {1, 2, . . . , n}. n if and only if A[0,∞[∈ K∗ 1, and . Proof. (i): We only need to show that A∗(x) = ∞ for some non-zero vector x ∈ [0,∞[n if and only if ( A)∗(x) = ∞ for some x > 0. For the direct implication, by monotonicity it is sufficient to take, for a given non-zero vector x, the value x equal to the maximum of the coordinates of x; the reverse implication follows by taking, for a given x > 0, the vector x = (x, x, . . . , x). (ii): Here we only need to show that A∗(x) = 0 for some non-zero vector x ∈ [0,∞[n if and only if (Ai)∗(x) = 0 for some x > 0 and some i ∈ {1, 2, . . . , n}. For the direct 7 implication it is sufficient to take an i ∈ {1, 2, . . . , n} for which the i-th coordinate of x has a non-zero value x; by monotonicity we have (Ai)∗(x) = 0. The reverse implication 2 follows by simply taking x = xei. As an example of this result consider the aggregation function A defined for any x ∈ It follows [0,∞]3 by taking A(x) to be the median of the coordinates of the vector x. (cid:51) but it does not belong to K(cid:51) immediately from Theorem 4 that A[0,∞[3∈ K∗ ∗ . 4 Comparing (A∗)∗ with (A∗)∗ In this section we prove that the values of (A∗)∗ can never be larger than the values of (A∗)∗ for an aggregation function A defined on [0,∞]n. The proof strongly depends on the fact that A is defined for non-zero vectors of [0,∞]n that are arbitrarily close to the zero vector. Theorem 5 Let A : [0,∞]n → [0,∞] be an aggregation function. Then (A∗)∗ ≤ (A∗)∗. Proof. For our given aggregation function A and for every i ∈ {1, . . . , n} let Ai : [0,∞[→ [0,∞[ be the marginal function of A considered before the proof of Theorem 4, given by Ai(x) = A(xei) = A(0, ..., x, ..., 0), where x appears only in the i-th coordinate. It is not difficult to check that A∗(x1, ..., xn) ≤ n(cid:88) (A∗)∗(x1, ..., xn) ≤ n(cid:88) (Ai)∗(xi) and A∗(x1, ..., xn) ≥ n(cid:88) ((Ai)∗)∗(xi) and (A∗)∗(x1, ..., xn) ≥ n(cid:88) i=1 i=1 It follows that Let now i=1 i=1 ai = lim inf t→0+ Ai(t) t and bi = lim sup t→0+ Ai(t) t . By Theorem 1 from Section 2 we have (Ai)∗(x) ≤ aix ≤ bix ≤ (Ai)∗(x) (Ai)∗(xi) . ((Ai)∗)∗(xi) . which implies that ((Ai)∗)∗(x) ≤ aix ≤ bix ≤ ((Ai)∗)∗(x) . Consequently, for every x = (x1, x2, . . . , xn) ∈ [0,∞[n we have (A∗)∗(x) ≤ n(cid:88) ((Ai)∗)∗(xi) ≤ n(cid:88) aixi ≤ n(cid:88) bixi ≤ n(cid:88) ((Ai)∗)∗(xi) = (A∗)∗(x) , i=1 i=1 i=1 i=1 8 which completes the proof. 2 We also include an observation about the other extreme, that is, when (A∗)∗ = (A∗)∗. A characterization of aggregation functions A on [0,∞]n for which this equality holds is still an open problem. Examples of such non-linear functions can be constructed, for instance, using Corollary 1 in conjunction with Theorem 3. However, if (A∗)∗ = (A∗)∗ we can at least say that the result is an additive function, i.e., a function D satisfying D(x + y) = D(x) + D(y) for every x, y ∈ [0,∞[n. We formulate and prove the result in a slightly larger generality. Corollary 4 Let B and C be aggregation functions on [0,∞]n such that B∗(x) = C∗(x) for every x ∈ [0,∞[n. Then, the function D = B∗ = C∗ is additive, D(x) = wixi for some w = (w1, ..., wn) ∈ [0,∞]n. n(cid:80) i=1 Proof. By the results of [4], B∗ is sub-additive and C∗ is super-additive; it follows that 2 D is both sub- and super-additive and hence additive. n(cid:80) i=1 Example 1 (i): Define A ∈ An by A(x1, . . . , xn) = max(x1, . . . , xn). Then A ∈ Pn and A = A∗, A∗(x1, . . . , xn) = xi, and (A∗)∗ = (A∗)∗ = A∗. (ii): Consider B, C ∈ An given by B(x1, . . . , xn) = ln(Πn exi) − n. Then B∗ = B, C∗ = C and B∗ = C∗ is given by B∗(x1, . . . , xn) = i=1(1+xi)) and C(x1, . . . , xn) = n(cid:80) xi. n(cid:80) ( i=1 i=1 5 Concluding remarks Our main aim was to characterize proper aggregation functions A with the property that A∗ (and, similarly, A∗) is proper as well. In the course of our investigation we obtained a num- ber of related results on the behavior of the subadditive and superadditive transformations A∗ and A∗. Regarding iterations of these transformations we proved that (A∗)∗ ≤ (A∗)∗ and that these two functions may be arbitrarily far from each other. In a nutshell and in a somewhat more abstract setting, we clarified constraints for aggre- gation functions admitting superadditive/subadditive transformations without a complete loss of information. The transformations considered are related to optimization problems, for example to production functions in economical problems. For a deeper discussion we recommend [4]. Acknowledgement. The author acknowledges support from the projects VEGA 1/0420/15 and APVV 0013/14. 9 References [1] Beliakov, G., Pradera, A., Calvo, T.: Aggregation Functions: A Guide for Prac- titioners. Springer, Berlin, 2007 ISBN: 978-3-540-73720-9 (Print) 978-3-540-73721-6 (Online) [2] Denardo, Eric V.: Linear Programming and Generalizations. Springer, Berlin, 2011, ISBN 978-1-4419-6491-5 [3] Grabisch, M., Marichal, J.- L., Mesiar, R., Pap, E.: Aggregation Functions (Encyklo- pedia of Mathematics and its Applications). Cambridge University Press, 2009 [4] Greco, S., Mesiar, R., Rindone, F., Sipeky, L.: The superadditive and the subaddi- tive transformations of integrals and aggregation functions. Fuzzy Sets and Systems, DOI:10.1016/j.fss.2015.08.006 [5] Lehrer, E.: A new integral for capacities., Economic Theory, Springer, 2009, vol. 39 (1), p. 157-176 [6] Mesiar, R., Li, J., Pap, E.: Superdecomposition integrals. Fuzzy Sets and Systems 259 (2015) p. 3-11 10
1807.06692	1	1807	2018-07-17T22:24:03	A characterization of superreflexivity through embeddings of lamplighter groups	[ "math.FA", "math.GR", "math.MG" ]	We prove that finite lamplighter groups $\{\mathbb{Z}_2\wr\mathbb{Z}_n\}_{n\ge 2}$ with a standard set of generators embed with uniformly bounded distortions into any non-superreflexive Banach space, and therefore form a set of test-spaces for superreflexivity. Our proof is inspired by the well known identification of Cayley graphs of infinite lamplighter groups with the horocyclic product of trees. We cover $\mathbb{Z}_2\wr\mathbb{Z}_n$ by three sets with a structure similar to a horocyclic product of trees, which enables us to construct well-controlled embeddings.	math.FA	math	A characterization of superreflexivity through embeddings of lamplighter groups Mikhail I. Ostrovskii and Beata Randrianantoanina July 19, 2018 Abstract We prove that finite lamplighter groups {Z2 ≀ Zn}n≥2 with a standard set of generators embed with uniformly bounded distortions into any non- superreflexive Banach space, and therefore form a set of test-spaces for super- reflexivity. Our proof is inspired by the well known identification of Cayley graphs of infinite lamplighter groups with the horocyclic product of trees. We cover Z2 ≀ Zn by three sets with a structure similar to a horocyclic product of trees, which enables us to construct well-controlled embeddings. 2010 Mathematics Subject Classification. Primary: 46B85; Secondary: 05C12, 20F65, 30L05. Keywords. distortion of a bilipschitz embedding, horocyclic product of trees, lamplighter group, Lipschitz map, metric embedding, Ribe program, superreflexivity, word metric 1 Introduction One of the important directions in metric geometry is to find purely metric charac- terizations of interesting classes of Banach spaces. For classes of spaces determined by finite-dimensional subspaces, this direction is a part of the Ribe program which was described by Bourgain [4] who proved the first metric characterization of su- perreflexivity. See [16] for more information on the Ribe program. The goal of this paper is to prove that lamplighter groups are test spaces for superreflexivity. Definition 1.1 ([21]). Let P be a class of Banach spaces and let T = {Tα}α∈A be a set of metric spaces. We say that T is a set of test spaces for P if the following two conditions are equivalent: (1) X /∈ P. (2) The spaces {Tα}α∈A admit bilipschitz embeddings into X with uniformly bounded distortions. Several different sets of test-spaces for superreflexivity are known: (1) binary trees [4, 14, 2, 9], (2) binary diamond and Laakso graphs [8, 15], (3) multibranching 1 diamond and Laakso graphs [24]. See [23] for a survey on this matter written in 2014. In this paper we add one more item to this list: (4) the set of Cayley graphs of finite lamplighter groups. Moreover, we observe that the Cayley graph of an infinite lamplighter group also is a test space for superreflexivity. Some of the characterizations (1)–(3) are independent in the sense that the cor- responding families of test spaces do not admit bilipschitz embeddings into each other with uniformly bounded distortions. In non-obvious cases this was shown in [22, 11] for finite binary trees and diamond graphs, and in [18] for diamond and Laakso graphs. Lamplighter groups are a very interesting class of groups which has been a rich source of important examples in geometric group theory. In 2008, Naor and Peres [17, Section 4] proved that the finite lamplighter groups Z2 ≀ Zn, with metric defined as a word length with respect to natural sets of generators, are embeddable into L1 with uniformly bounded distortions. The first goal of this paper is to strengthen this result and to prove their embeddability into an arbitrary nonsuperreflexive Banach space with uniformly bounded distortions. As a consequence we get a new metric characterization of superreflexivity, see Corollary 1.7. We consider the following special case of the general wreath product construction (see, for example, [6, p. 214] for the general definition). Definition 1.2 ([5, p. 129]). Let H and L be two groups and LH be the set of all L-valued finitely supported functions on H. Then the wreath product L≀H is defined as the set LH × H equipped with the multiplication ((xh)h∈H, g) · ((yh)h∈H, k) := ((xh · ygh)h∈H, gk). Our main interest will be the wreath product Z2 ≀ Zn, for n ∈ N, n ≥ 2, which we denote by Gn, this group is called the lamplighter group, see [27] for a nice introduction including an explanation for the name. We identify ZZn with {j ∈ Zn : xj = 1}, the group operation on ZZn From now on we will abuse notation and treat an element x ∈ ZZn Zn. 2 with the family of all subsets of Zn by identifying x = (xk)k∈Zn is the symmetric difference. 2 as a subset of 2 Considering an element (x, k) ∈ Gn, we call x the set of positions where the lamp is on, and its complement - the set of positions where the lamp is off. The number k ∈ Zn is called the location of the lamplighter. It is easy to see that the elements a = ({0}, 0) and t = (∅, 1) generate Gn. Observe that multiplication by a = ({0}, 0) on the right is the act of changing the lamp at the current location of the lamplighter and multiplication by t = (∅, 1) on the right is the act of the lamplighter moving one position to the next lamp in the 'positive' direction around the circle. We consider the metric ρ on Gn defined as the metric of the left-invariant Cayley graph with respect to the set of generators S = {t, ta}. This means that x is adjacent 2 to y if and only if x = ys or y = xs (i.e x = ys−1) for one of the generators s ∈ S. Observe that the generator ta acts by first moving one step in the 'positive' direction, and then changing the state of the lamp at the final location of the lamplighter. See Section 2.2 for a more detailed description and an equivalent formula (2.5) for the metric. The first main result of this paper is Theorem 1.3. For any nonsuperreflexive Banach space X the Cayley graphs of Z2 ≀ Zn (n ≥ 2) corresponding to the set S = {t, ta} admit embeddings into X with uniformly bounded distortions. Remark 1.4. In [17], when proving embeddability into L1, Naor and Peres considered Z2 ≀ Zn with the set of generators equal to {t, a} instead of {t, ta} as we do. However, it is easy to see that the metrics induced by these two generating sets are bilipschitz equivalent to each other with a distortion 4. In [17] Naor and Peres constructed two embeddings into L1, one based on ir- reducible representations of finite lamplighter groups, and the other motivated by what they refer to as a 'direct geometric reasoning'. Our embedding technique is quite different from either of the embeddings in [17]. Our approach is inspired by the description of the Cayley graph of the infinite lamplighter group Zq ≀ Z, for q ∈ N, q ≥ 2, as a horocyclic product of two trees, introduced by Bartholdi and Woess [1, 28], and by the analysis of the metric structure of horocyclic products of trees by Stein and Taback in [26] who proved, among other results, Theorem 1.5. [26, Corollary 10] For any q ∈ N, q ≥ 2, the Cayley graph of Zq ≀ Z with the set of generators {t, ta, . . . , taq−1} admits an embedding into an ℓ1-sum of two q-branching trees with distortion bounded by 4. We refer the reader to [27] and [29] for very nice presentations of horocyclic products of trees and their applications to lamplighter groups. We give a more detailed overview of our method in Section 2.1. Remark 1.6. It follows from Theorem 1.5, [4], and [19] that for all q ∈ N, q ≥ 2, the Cayley graph of Zq ≀ Z admits an embedding into any non-superreflexive space with distortion independent of q, as was shown in a similar case of hyperbolic groups in [22, Section 2] (cf. also Section 2.4 below). The bilipschitz embeddability of the infinite group Z2 ≀ Z into any non-superre- flexive space can also be derived from Theorem 1.3 and results of [4, 19]. In [13] it is proved that there exists a constant c > 0 so that for every n ∈ N a complete binary tree of depth cn embeds with constant distortion into Gn. In Sec- tion 3 below we show an alternative simple proof of this fact using our construction. As a consequence, and by Theorem 1.3, Remark 1.6, and [4], we obtain 3 Corollary 1.7. The sequence of Cayley graphs for {Z2 ≀ Zn}n≥2 with respect to the set of generators S = {t, ta} is a set of test-spaces for superreflexivity. The Cayley graph of Z2 ≀ Z with respect to any finite set of generators is a test space for superreflexivity. Remark 1.8. We note that Theorem 1.5 is valid for Zq ≀ Z for all q ∈ N, q ≥ 2. As we elaborate in Section 2.1, our proof of Theorem 1.3 is inspired by the methods of Theorem 1.5, even though we do not apply its conclusion for our argument. In our statement and proof of Theorem 1.3, for greater clarity of the presentation, we focused our attention on Z2 ≀ Zn, but it only requires straightforward adjustments of the proof to obtain the same conclusion as in Theorem 1.3 for Zq ≀ Zn with the set of generators {t, ta, . . . , taq−1}, for all q ∈ N, q ≥ 2. The only difference is that one needs to use an ℓ∞-sum of two q-branching trees in place of binary trees. Similarly as in [26], the value of q affects the branching of the trees, but not the number of the summands. Since for all q, a q-branching tree embeds almost isometrically into a binary tree, and since we always use an ℓ∞-sum of two trees, the uniform bound on distortions does not depend on q ∈ N. Since, for all q ∈ N, q ≥ 2, the Cayley graphs of Zq ≀ Zn and Zq ≀ Z with the set of generators {t, ta, . . . , taq−1} isometrically contain the Cayley graphs of Z2 ≀ Zn and Z2 ≀ Z with generators {t, ta}, respectively, it follows from Corollary 1.7 and the adjusted version of Theorem 1.3, that for any q ∈ N, q ≥ 2, the sequence of Cayley graphs of Zq ≀ Zn with the set of generators {t, ta, . . . , taq−1} is a set of test-spaces for superreflexivity. Similarly, for any q ∈ N, q ≥ 2, the Cayley graph of Zq ≀ Z, with respect to any finite generating set, is a test space for superreflexivity. 2 Proof of Theorem 1.3 2.1 Outline of the proof To simplify notation, we assume that n is divisible by 6. It is clear that the same arguments work in general, but the formulas will be somewhat more complicated. We cover the Cayley graph of Zn with respect to the generating set {±1} by three overlapping paths P1, P2, P3 of length 2 3 n each in such a way that each pair of points in Zn belongs to at least one of the paths (paths that exclude three mutually disjoint thirds of the cycle Zn work). We consider the following three subsets of Gn for i = 1, 2, 3, Pi,n def= {(x, k) ∈ Gn x ⊆ Zn, k ∈ Pi}. (2.1) We equip Pi,n with the metric inherited from the Cayley graph of Gn with respect to S = {t, ta}. We note that the union of the three sets Pi,n covers Gn. Our approach is inspired by the well-known description of infinite lamplighter groups Z2 ≀Z as a horocyclic product of two infinite trees [1, 28] and by the bilipschitz 4 embedding of the word metric on Z2 ≀ Z into an ℓ1-sum of two trees [26] (see Theo- rem 1.5 above). Our first goal is to cover Z2 ≀ Zn by three sets with a structure similar to a horocyclic product of trees, which will enable us to construct well-controlled embeddings. The sets Pi,n are defined so that for all locations k, l ∈ Pi, the length of the path that is contained in Pi and connects k and l is at most twice the length of the shortest path from k to l in Zn. For this reason, the metric structure of the sets Pi,n sufficiently resembles the metric structure of a subset of Z2 ≀ Zn and we are able to construct sets Wn, which we think of as analogs of horocyclic products of trees, and which are bilipschitz equivalent to the sets Pi,n. The sets Wn are defined as specific subsets of the ℓ∞-sum of two trees of depth n. To identify each element of Pi,n with an element of the Cartesian product Tn × Tn of two trees of depth n, we first mark a vertex v0 in Zn which is the midpoint of the complement of the path Pi, that is v0 is at the distance at least n/6 from any element k ∈ Pi. Modelling our description on the identification of Z2 ≀ Zn with the horocyclic product of trees (cf. [28, p. 419]), for every (x, k) ∈ Pi,n we describe the set x as the union of x ∩ Ik,+ and x ∩ Ik,−, where Ik,+, Ik,− are two disjoint arcs in Zn both with the endpoints v0 and k. Each set x ∩ Ik,+ and x ∩ Ik,− is encoded by a sequence of 0s and 1s of length equal to the number of vertices in Ik,+ and Ik,−, respectively. This naturally encodes each element (x, k) ∈ Pi,n by two elements of a binary tree whose levels (in a tree) add up to n. We verify in Section 2.2 that this encoding is metrically faithful on Pi,n, that is, we construct bilipschitz embeddings ϕi,n : Pi,n → Tn ⊕∞ Tn, with uniformly bounded distortions, see Section 2.2 for details. This completes the first and main step of our proof. The next step of our proof is a routine application of the well-known theory of Lipschitz retracts to conclude that, for each i ∈ {1, 2, 3}, the bilipschitz map ϕi,n : Pi,n → Tn ⊕∞ Tn, constructed in the first step, can be extended to a Lipschitz map ¯ϕi,n from the entire Gn into an ℓ∞-sum of two metric trees of depth n, see Section 2.3 for details. In the final step of our proof we define the map Φn from Gn into an ℓ∞-sum of six metric trees of depth n by Φn(x, k) def= ( ¯ϕ1,n(x, k), ¯ϕ2,n(x, k), ¯ϕ3,n(x, k)). Clearly, the maps Φn are Lipschitz with the same Lipschitz constants as those of the maps ¯ϕi,n, for i = 1, 2, 3. Since the paths P1, P2, P3 were chosen in such a way that for any two elements (x, k), (y, l) ∈ Gn, there exists i ∈ {1, 2, 3} so that (x, k), (y, l) ∈ Pi,n, it follows that the map Φn is co-Lipschitz with same constant as the map ϕi,n (in our construction all maps {ϕi,n}3 i=1 have the same co-Lipschitz constant), see Section 2.4. 5 Hence to finish the proof of Theorem 1.3, it is enough to verify that for all n ∈ N, the ℓ∞-sum of six metric trees of depth n embeds into any non-superreflexive Banach space X with uniformly bounded distortions. This follows readily by known tech- niques and results on bilipschitz embeddability of trees into any non-superreflexive Banach space, and on extension of bilipschitz embeddings into Banach spaces from vertex sets to graphs to the corresponding 1-dimensional complexes. In fact it even suffices to prove the existence of Lipschitz maps that satisfy slightly weaker require- ments, see Section 2.4 for details. 2.2 Step 1 We define the paths P1, P2 and P3, to be arcs of lengths 2 [ 1 6 n, 5 by 6). 3 n in Zn with endpoints 2n], respectively (recall that we assumed that n is divisible 6n], [ 1 2n, 1 6 n] and [ 5 6n, 1 Let θn : Zn → Zn be the 'rotation' of Zn by an arc of length n/3, i.e. for each 3 , where the addition is in the sense of Zn. The rotation θn k ∈ Zn, θn(k) = k + n induces the isometry θn of Gn onto itself, defined by θn(x, k) = (θn(x), θn(k)), where, as usual, θn(x) denotes the image of the set x under the action of θn. Note, that the paths P1, P2, P3, satisfy P2 = θn(P1) and P3 = θ2 n(P1), and the metric spaces {Pi,n}3 i=1, defined by (2.1), are isometric to each other, specifically P2,n = θn(P1,n), P3,n = θn 2 (P1,n). (2.2) Let Tn be a binary tree of depth n, that is, Tn is the graph whose vertices are labelled by sequences of 0s and 1s of lengths ≤ n with the usual graph distance dT . We consider the ℓ∞-sum Tn ⊕∞ Tn defined as the Cartesian product Tn ×Tn endowed with the metric d∞((A1, A2), (B1, B2)) def= max{dT (A1, B1), dT (A2, B2)}. We define Wn def= (cid:26)(A1, A2) ∈ Tn ⊕∞ Tn : A1 + A2 = n, A1, A2 ∈(cid:20)1 6 n, 5 6 n(cid:21)(cid:27) , where A denotes the length of the sequence A ∈ Tn. It will be convenient to also use the distance d1 on Wn which is 2-equivalent with d∞. d1((A1, A2), (B1, B2)) def= dT (A1, B1) + dT (A2, B2). We will show that there exist bijections ϕ1,n from {(P1,n, d∞)}n onto {(Wn, d∞)}n which have uniformly bounded distortions. 6 We define a map ϕ1,n : P1,n → Wn as follows: for any (x, k) ∈ P1,n, the element ϕ1,n(x, k) def= (A1, A2), where A1 = (a1,0, . . . , a1,k−1) is a sequence of length k and A2 = (a2,1, . . . , a2,n−k) is a sequence of length n − k, defined by a1,j =(1 0 if j ∈ x, if j /∈ x, a2,i =(1 0 if n − i ∈ x, if n − i /∈ x. It is clear that the map ϕ1,n is one-to-one and onto. We will show that ϕ1,n is a bilipschitz isomorphism of (P1,n, ρ) and (Wn, d∞). For any A, B ∈ Tn, we denote by lgca(A, B) the length of the greatest common ancestor of A and B in Tn. In this notation, dT (A, B) = (A − lgca(A, B)) + (B − lgca(A, B)), and for (A1, A2), (B1, B2) ∈ Wn, d1((A1, A2), (B1, B2)) = dT (A1, B1) + dT (A2, B2) = 2(n − (lgca(A1, B1) + lgca(A2, B2))). In particular, for all (A1, A2), (B1, B2) ∈ Wn we have d1((A1, A2), (B1, B2)) ≤ 2n. (2.3) (2.4) To continue we need to estimate the distance ρ((x, k), (y, l)), where (x, k), (y, l) ∈ Gn and ρ is the distance in the Cayley graph of Gn with the generating set {t, ta}. We use the following observation: to get from (x, k) to (y, l) we need • to traverse at least one of the two paths from k to l on the n-cycle (the graph of Zn with respect to the generating set {±1}); • to visit all positions j ∈ Zn which are not on the selected path, but which belong to x△y, and to change the state of all lamps at these positions. Denote by p1 and p2 the lengths of the two distinct paths from k to l on Zn, and by g1 and g2 – the sizes of the largest "gaps" in these paths, that is, the largest distances between distinct vertices for which there is no element of x△y in between (observe that g1 and g2 are at least 1 each). With this notation it is easy to see the validity of the leftmost inequality in min{p1 + 2(p2 − g2), p2 + 2(p1 − g1)} ≤ ρ((x, k), (y, l)) ≤ min{p1 + 2(p2 − g2), p2 + 2(p1 − g1)} + 2. (2.5) The rightmost inequality in (2.5) holds because discrepancies with the equality in (2.5) can occur only at one of the endpoints of the 'interval' on Zn consisting 7 of all vertices that are visited by an optimal tour from k to l that establishes the distance between (x, k) and (y, l). One of the cases when the distance exceeds the minimum by 2 is the following: the distance from k to l in the positive direction is significantly smaller than in the negative direction and x△y = {k, l}. In this case, if we start at k we need to do one step in the negative direction in order to change the status of the lamp at k, and then head back in the positive direction. Note that the position l is reached by the step ta in order to change the status of the lamp there, so no additional steps are needed at this endpoint. For the rest of the proof we fix (x, k), (y, l) ∈ P1,n, and (A1, A2) = ϕ1,n(x, k), (B1, B2) = ϕ1,n(y, l) in Wn. Observation 2.1. The sum lgca(A1, B1) + lgca(A2, B2) is equal to the number of vertices of the set E constructed as a union of two, possibly empty, 'intervals' in Zn. One of the intervals starts at 0 ∈ Zn, goes in the 'positive' direction and ends at the first vertex which belongs to x△y ∪ {k, l}, excluding this vertex, in particular if 0 ∈ x△y∪{k, l}, then the interval is empty. The other interval starts at (n−1) ∈ Zn, goes in the 'negative' direction and ends at the first vertex which is in x△y ∪ {k, l}, this interval includes its end if and only if it does not belong to x△y. Since the number of vertices in the interval E is equal to either its length E, if E is empty, or to E + 1, otherwise, we obtain E ≤ lgca(A1, B1) + lgca(A2, B2) ≤ E + 1. (2.6) Note that (2.3) and (2.6) immediately imply that d1(ϕ1,n(x, k), ϕ1,n(y, l)) ≥ 2(n − g), where g is the number of vertices in the largest 'interval' I in Zn which is disjoint with {k, l} ∪ x△y. Since it is clear that there exists a 'tour' of the lamplighter that travels from k to l, passes through every vertex of the difference x△y, visits each vertex at most twice, and stays outside I except possibly one vertex (see the paragraph after (2.5)), for all pairs (x, k), (y, l) ∈ P1,n we get the inequality ρ((x, k), (y, l)) ≤ 2 + 2(n − g) ≤ 2 + d1(ϕ1,n(x, k), ϕ1,n(y, l)) ≤ 2 + 2d∞(ϕ1,n(x, k), ϕ1,n(y, l)). To get the inequality in the other direction we observe that ρ((x, k), (y, l)) is at least τ − 1, where τ is the number of vertices in the smallest 'interval' J on Zn that contains all elements of {k, l} ∪ x△y. Note that J is nonempty, and thus τ = J + 1. If the interval J does not contain 0, then J and the interval E constructed in Observation 2.1 have at most one point in common, and the union J ∪ E covers Zn. 8 Thus the length of J is at least n − E − 1. In this case, by (2.3) and (2.6), we get ρ((x, k), (y, l)) ≥ τ − 1 ≥ n − E − 1 ≥ ≥ 1 2 1 2 (d1(ϕ1,n(x, k), ϕ1,n(y, l))) − 1 (d∞(ϕ1,n(x, k), ϕ1,n(y, l))) − 1. If the interval J contains 0, then it contains at least 1 3 n − 1 vertices, because by the definition of P1,n we have k, l ∈ P1 =(cid:2) 1 ρ((x, k), (y, l)) ≥ n − 2 6 n, 5 6 n(cid:3). So in this case, by (2.4), we get 1 3 1 6 1 6 ≥ ≥ max{d1((A1, A2), (B1, B2)) : (A1, A2), (B1, B2) ∈ Wn} − 2 (d∞(ϕ1,n(x, k), ϕ1,n(y, l))) − 2. This ends the proof that the bijections ϕ1,n : (P1,n, d∞) → (Wn, d∞) have uni- formly bounded distortions. Since, by (2.2), the spaces P2,n and P3,n are isometric to P1,n, the maps ϕ2,n def= ϕ1,n ◦ θn, and ϕ3,n are bijections from P2,n and P3,n, respectively, onto (Wn, d∞), and they have the same Lipschitz and co-Lipschitz constants as the maps ϕ1,n. def= ϕ1,n ◦ θn 2 2.3 Step 2 We will apply the well-known theory of Lipschitz retracts. For basic theory of Lipschitz retracts in metric spaces see e.g. [3, Chapter 1] and [10, Propositions 2.1, 2.2, and the comment at the top of page 303]. For the convenience of the reader, we briefly recall the definitions and results that we use. A metric space M is called injective if for every metric space B, every A ⊆ B, and every Lipschitz function f : A → M, there exists a Lipschitz extension of f , that is, a function ¯f : B → M, so that ¯f A = f and Lip( ¯f ) = Lip(f ). A metric space M is called a λ-absolute Lipschitz retract (where 1 ≤ λ < ∞) provided that whenever X is isometrically contained in a metric space Y , there exists a retraction, r, from Y onto X, with Lip(r) ≤ λ. Theorem 2.2 ([10, Proposition 2.2]). A metric space M is injective if and only if it is an absolute 1-Lipschitz retract. A metric space M is said to have the binary intersection property if every collec- tion of mutually intersecting closed balls in M has a common point. A metric space M is said to be metrically convex if for every x0, x1 ∈ M and for every 0 < t < 1 there is a point xt ∈ M such that d(x0, xt) = td(x0, x1) and d(x1, xt) = (1 − t)d(x0, x1). 9 Proposition 2.3 ([3, Proposition 1.4]). A metric space M is an absolute 1-Lipschitz retract if and only it is metrically convex and has the binary intersection property. A metric space M is called a metric tree if it is complete, metrically convex, and for any pair of vertices there is a unique continuous curve joining them in M. Given a tree T in Graph Theory sense, one can construct the corresponding metric tree by attaching between any two adjacent vertices mutually disjoint curves isometric to the interval (0, 1). Proposition 2.4 (Corollary of [7, Lemma 2.1]). A metric binary tree of any height is a 1-absolute Lipschitz retract. Observation 2.5. The binary intersection property and metric convexity are pre- served under ℓ∞-sums. Proof sketch. Binary intersection property: If balls {B((xi, yi), ri)}i in M1 ⊕∞ M2 are mutually intersecting, then the same happens for projected balls {B((xi, ri)}i in M1 and {B((yi, ri)}i in M2. Thus each of the projected collections of balls has a nonempty intersection. Let x and y be some points in the intersections. Using the definition of the ℓ∞-sum we get that {B((xi, yi), ri)}i are Cartesian products of {B((xi, ri)}i and {B((yi, ri)}i. Thus (x, y) is in the intersection of {B((xi, yi), ri)}i. Metric convexity: Let (x0, y0) and (x1, y1) be two points in the ℓ∞-sum. Let {xt} be a suitable family for M1 and {yt} be a suitable family for M2. Then {(xt, yt)} is a suitable family for the ℓ∞-sum. Corollary 2.6. Let Πn be the ℓ∞-sum of two metric binary trees of depth n each. Then Πn is a 1-absolute Lipschitz retract. As an immediate consequence we obtain the objective of Step 2. Corollary 2.7. For i = 1, 2, 3, let ϕi,n : Pi,n → Wn be the maps defined in Step 1. Since Wn ⊆ Tn ⊕∞ Tn ⊆ Πn, there exist Lipschitz extensions of ϕi,n to maps ¯ϕi,n : Gn → Πn which have uniformly bounded Lipschitz constants. 2.4 The Final Step of the proof of Theorem 1.3 By [4] binary trees {Tn}∞ n=1 admit bilipschitz embeddings into any nonsuperreflexive space X with uniformly bounded distortions (see also [25] and [23]), and, without loss of generality, we may assume that these embeddings do not decrease any distances. Using Mazur's method of constructing basic sequences (see [12, pp. 4-5]), as in [22, Section 3], one can construct embeddings jn, which do not decrease distances and have uniformly bounded distortions, of an ℓ∞-sum of six copies of Tn into any non-superreflexive space X jn : Tn ⊕∞ Tn ⊕∞ Tn ⊕∞ Tn ⊕∞ Tn ⊕∞ Tn → X. 10 Using [21, Lemma 3.3] or a direct argument, one can extend embeddings jn to bilipschitz embeddings of Πn⊕∞Πn⊕∞Πn into X with uniformly bounded distortions (recall that Πn denotes the ℓ∞-sum of two metric binary trees of depth n each). However, for our purpose, it suffices to extend embeddings jn to Lipschitz maps from Πn ⊕∞ Πn ⊕∞ Πn into X with uniformly bounded Lipschitz constants. This can be done directly: we extend the maps from a binary tree to the metric edges using 'linear interpolation': a point ut (0 < t < 1) on the edge joining vertices u0 and u1 in the metric tree corresponding to Tn with d(u0, ut) = t and d(ut, u1) = 1 − t is mapped onto the corresponding convex combination of the images of u0 and u1. We denote the obtained Lipschitz maps by En : Πn ⊕∞ Πn ⊕∞ Πn → X. By construction, the maps {En}n are non-contractive on Tn ⊕∞ Tn ⊕∞ Tn ⊕∞ Tn ⊕∞ Tn ⊕∞ Tn, and their Lipschitz constants are bounded by a universal constant C. We are now ready to define the embeddings Fn : Gn → X. We put Fn(x, k) def= En( ¯ϕ1,n(x, k), ¯ϕ2,n(x, k), ¯ϕ3,n(x, k)). It follows from Step 1 and our construction that the Lipschitz constants of the maps Fn are uniformly bounded . To prove that maps Fn are uniformly co-Lipschitz, we observe that for any (x, k), (y, l) ∈ Gn, there exists i ∈ {1, 2, 3} such that both (x, k) and (y, l) are in Pi,n. By Step 1 and since the restriction of En to Tn ⊕∞ Tn ⊕∞ Tn ⊕∞ Tn ⊕∞ Tn ⊕∞ Tn is noncontractive, we get kFn(x, k) − Fn(y, l)kX ≥ d∞(ϕi,n(x, k), ϕi,n(y, l)) ≥ 1 4 ρ((x, k), (y, l)), which completes the proof of Theorem 1.3. 3 Proof of Corollary 1.7 For finite groups. As we mentioned in the Introduction, it is enough to prove that for all n, the Cayley graph of Gn contains a subset that is bilipschitz equivalent with an absolute constant to the binary tree of depth n/2. As mentioned earlier, this fact was proved in [13] for trees of depth n/c, for some c > 0, but below we obtain it as a direct consequence of our construction in Section 2.2. Indeed, let Wn be the set defined in Section 2.2, and let U be the subset of Wn consisting of all pairs (A, B) ∈ Wn so that A ≤ n 2 , and B is a sequence consisting of n − A zeroes. The distance in (Wn, d∞) between any two elements of the set U is bounded below by the tree distance of the corresponding As, and is bounded above by twice the the tree distance of the corresponding As. Thus U is 2-equivalent to the binary tree of depth n/2. Since, by Step 1, P1,n ⊆ Gn is bilipschitz equivalent with Wn, and by [4], the proof for finite groups is complete. 11 For the infinite group, by Remark 1.6, we only need to show that the bilipschitz embeddability of Z2 ≀ Z into a Banach space X implies that X is nonsuperreflexive. This follows similarly as in the case of finite groups. By [28], the Cayley graph of Z2 ≀ Z with the generating set S = {t, ta} coincides with the horocyclic product of two infinite binary trees, i.e. infinite trees whose every vertex has degree 3, and the identification is obtained by mapping every element (x, k) ∈ Z2 ≀ Z to an element (A1, A2, k) of two infinite sequences of 0s and 1s with an (arbitrary) finite number of nonzero terms, cf. also [27]. By Theorem 1.5, the metric on Z2 ≀ Z is bilipschitz equivalent with the metric inherited from the ℓ1-sum of two tree metrics. Thus, as in the finite case, we see that Z2 ≀ Z contains a bilipschitz copy of an infinite rooted binary tree by taking the set of all elements of the form (A1, A2, k), where k = 0, 1, 2, . . . , A1 is any sequence of 0s and 1s so that all terms with indices larger than k are equal to 0, and all terms in the sequence A2 are equal to 0. The fact that bilipschitz embeddability of a binary tree into X implies nonsuper- reflexivity of X follows from [4], see also [2]. Acknowledgements: We would like to thank Florent Baudier for suggesting the problem on lamplighter groups to us. The first named author was supported by the National Science Foundation under Grant Number DMS–1700176. Both authors thank the Fields Institute (Toronto) for partial funding to attend the Workshop on Large Scale Geometry and Applications, where we started our work on lamplighter groups. References [1] L. Bartholdi, W. Woess, Spectral computations on lamplighter groups and Diestel- Leader graphs. J. Fourier Anal. Appl. 11 (2005), no. 2, 175–202. [2] F. Baudier, Metrical characterization of super-reflexivity and linear type of Banach spaces, Archiv Math., 89 (2007), no. 5, 419–429. [3] Y. Benyamini, J. Lindenstrauss, Geometric nonlinear functional analysis. Vol. 1. American Mathematical Society Colloquium Publications, 48. American Mathemat- ical Society, Providence, RI, 2000. [4] J. Bourgain, The metrical interpretation of superreflexivity in Banach spaces, Israel J. Math., 56 (1986), no. 2, 222–230. [5] C. Drutu, M. Kapovich, Geometric group theory. With an appendix by Bogdan Nica. American Mathematical Society Colloquium Publications, 63. American Mathemat- ical Society, Providence, RI, 2018. [6] P. de la Harpe, Topics in geometric group theory. Chicago Lectures in Mathematics. University of Chicago Press, Chicago, IL, 2000. 12 [7] W. B. Johnson, J. Lindenstrauss, D. Preiss, G. Schechtman, Lipschitz quotients from metric trees and from Banach spaces containing ℓ1. J. Funct. Anal. 194 (2002), no. 2, 332–346. [8] W. B. Johnson, G. Schechtman, Diamond graphs and super-reflexivity, J. Topol. Anal., 1 (2009), no. 2, 177–189. [9] B. Kloeckner, Yet another short proof of the Bourgain's distortion estimate for em- bedding of trees into uniformly convex Banach spaces, Israel J. Math., 200 (2014), no. 1, 419–422. [10] U. Lang, Injective hulls of certain discrete metric spaces and groups. J. Topol. Anal. 5 (2013), no. 3, 297–331. [11] S. L. Leung, S. Nelson, S. Ostrovska, M. I. Ostrovskii, Distortion of embeddings of binary trees into diamond graphs. Proc. Amer. Math. Soc. 146 (2018), no. 2, 695– 704. [12] J. Lindenstrauss, L. Tzafriri, Classical Banach spaces. I. Sequence spaces. Ergebnisse der Mathematik und ihrer Grenzgebiete, Vol. 92. Springer-Verlag, Berlin-New York, 1977. [13] R. Lyons, R. Pemantle, Y. Peres, Random walks on the lamplighter group. Ann. Probab. 24 (1996), no. 4, 1993–2006. [14] J. Matousek, On embedding trees into uniformly convex Banach spaces, Israel J. Math., 114 (1999), 221–237. [15] M. Mendel, A. Naor, Markov convexity and local rigidity of distorted metrics, J. Eur. Math. Soc. (JEMS), 15 (2013), no. 1, 287–337; Conference version: Computational geometry (SCG'08), 49–58, ACM, New York, 2008. [16] A. Naor, An introduction to the Ribe program, Jpn. J. Math., 7 (2012), no. 2, 167– 233. [17] A. Naor, Y. Peres, Embeddings of discrete groups and the speed of random walks. Int. Math. Res. Not. IMRN 2008, Art. ID rnn 076, 34 pp. [18] S. Ostrovska, M. I. Ostrovskii, Nonexistence of embeddings with uniformly bounded distortions of Laakso graphs into diamond graphs, Discrete Math., 340 (2017), no. 2, 9–17. [19] M. I. Ostrovskii, Embeddability of locally finite metric spaces into Banach spaces is finitely determined, Proc. Amer. Math. Soc., 140 (2012), 2721–2730. [20] M. I. Ostrovskii, Metric Embeddings: Bilipschitz and Coarse Embeddings into Banach Spaces, de Gruyter Studies in Mathematics, 49. Walter de Gruyter & Co., Berlin, 2013. 13 [21] M. I. Ostrovskii, Different forms of metric characterizations of classes of Banach spaces, Houston. J. Math., 39 (2013), no. 3, 889–906. [22] M. I. Ostrovskii, Metric characterizations of superreflexivity in terms of word hyper- bolic groups and finite graphs. Anal. Geom. Metr. Spaces 2 (2014), 154–168. [23] M. I. Ostrovskii, Metric characterizations of some classes of Banach spaces, in: Har- monic Analysis, Partial Differential Equations, Complex Analysis, Banach Spaces, and Operator Theory, Celebrating Cora Sadosky's life, M. C. Pereyra, S. Marcan- tognini, A. M. Stokolos, W. U. Romero (Eds.), Association for Women in Mathematics Series, Vol. 4, pp. 307–347, Springer-Verlag, Berlin, 2016. [24] M. I. Ostrovskii, B. Randrianantoanina, A new approach to low-distortion embeddings of finite metric spaces into non-superreflexive Banach spaces. J. Funct. Anal. 273 (2017), no. 2, 598–651. [25] G. Pisier, Martingales in Banach spaces, Cambridge Studies in Advanced Mathemat- ics 155. Cambridge, Press Cambridge University Press, 2016. [26] M. Stein, J. Taback, Metric properties of Diestel-Leader groups. Michigan Math. J. 62 (2013), no. 2, 365–386. [27] J. Taback Lamplighter groups. in: M. Clay, D. Margalit (Editors), Office hours with a geometric group theorist, 310–330, Princeton Univ. Press, Princeton, NJ, 2017. [28] W. Woess, Lamplighters, Diestel-Leader graphs, random walks, and harmonic func- tions. Combin. Probab. Comput. 14 (2005), no. 3, 415–433. [29] W. Woess, What is a horocyclic product, and how is it related to lamplighters? Internat. Math. Nachrichten of the Austrian Math. Soc. 224 (2013), 1–27; see also a corrected version on the author's website. Department of Mathematics and Computer Science, St. John's Uni- versity, 8000 Utopia Parkway, Queens, NY 11439, USA E-mail address: [email protected] Department of Mathematics, Miami University, Oxford, OH 45056, USA E-mail address: [email protected] 14
1712.04338	1	1712	2017-12-09T15:32:15	Liftings for ultra-modulation spaces, and one-parameter groups of Gevrey type pseudo-differential operators	[ "math.FA" ]	We deduce one-parameter group properties for pseudo-differential operators $\operatorname{Op} (a)$, where $a$ belongs to the class $\Gamma ^{(\omega _0)}_$ of certain Gevrey symbols. We use this to show that there are pseudo-differential operators $\operatorname{Op} (a)$ and $\operatorname{Op} (b)$ which are inverses to each others, where $a\in \Gamma ^{(\omega _0)}_$ and $b\in \Gamma ^{(1/\omega _0)}_*$. We apply these results to deduce lifting property for modulation spaces and construct explicit isomorpisms between them. For each weight functions $\omega ,\omega _0$ moderated by GRS submultiplicative weights, we prove that the Toeplitz operator (or localization operator) $\operatorname{Tp} (\omega _0)$ is an isomorphism from $M^{p,q}_{(\omega )}$ onto $M^{p,q}_{(\omega /\omega _0)}$ for every $p,q \in (0,\infty ]$.	math.FA	math	LIFTINGS FOR ULTRA-MODULATION SPACES, AND ONE-PARAMETER GROUPS OF GEVREY TYPE PSEUDO-DIFFERENTIAL OPERATORS AHMED ABDELJAWAD, SANDRO CORIASCO, AND JOACHIM TOFT Abstract. We deduce one-parameter group properties for pseudo-differential operators Op(a), where a belongs to the class Γ(ω0) of certain Gevrey symbols. We use this to show that there are pseudo-differential operators Op(a) and Op(b) which are inverses to each others, where a ∈ Γ(ω0) and b ∈ Γ(1/ω0) . ∗ ∗ ∗ We apply these results to deduce lifting property for modulation spaces and construct explicit isomorpisms between them. For each weight functions ω, ω0 moderated by GRS submultiplicative weights, we prove that the Toeplitz operator (or localization operator) Tp(ω0) is an isomorphism from M p,q (ω) onto M p,q (ω/ω0) for every p, q ∈ (0, ∞]. 0. Introduction The topological vector spaces V1 and V2 is said to possess lifting property if there exists a "convenient" homeomorphisms (a lifting) between them. For example, for any weight ω on Rd, p ∈ (0, ∞] and s ∈ R the convenient mappings f 7→ ω · f and f 7→ (1 − ∆)s/2f are homeomorphic from the (weighted) Lebesgue space Lp (ω) and the Sobolev space H p 0 , with inverses f 7→ ω−1 · f and f 7→ (1 − ∆)−s/2f , respectively. (Cf. [30] and Section 1 for notations.) Hence, these spaces possess lifting properties. s , respectively, into Lp = H p It is often uncomplicated to deduce lifting properties between (quasi-)Banach spaces of functions and distributions, if the definition of their norms only differs by a multiplicative weight on the involved distributions, or on their Fourier transforms, which is the case in the previous homeomorphisms. Here recall that multiplications on the Fourier transform side are linked to questions on differentiation of the in- volved elements. A more complicated situation appear when there are some kind of interactions between multiplication and differentiation in the definition of the involved vector spaces. An example where such interactions occur concerns the extended family of Sobolev spaces, introduced by Bony and Chemin in [3] (see also [34]). More pre- cisely, let ω, ω0 be suitable weight functions and g a suitable Riemannian met- ric, which are defined on the phase space W ≃ T ∗Rd ≃ R2d. Then Bony and Chemin introduced in [3] the generalised Sobolev space H(ω, g) which fits the Hormander-Weyl calculus well in the sense that H(1, g) = L2, and if a belongs to the Hormander class S(ω0, g), then Weyl operator Opw(a) with symbol a is con- tinuous from H(ω0ω, g) to H(ω, g). Moreover, they deduced group algebras, from which it follows that to each such weight ω0, there exist symbols a and b such that Opw(a) ◦ Opw(b) = Opw(b) ◦ Opw(a) = I, a ∈ S(ω0, g), b ∈ S(1/ω0, g). (0.1) 2010 Mathematics Subject Classification. Primary: 35S05, 47D06, 46B03, 42B35, 47B35, 46F05; Secondary: 46G15, 47L80. Key words and phrases. Gelfand-Shilov spaces, Banach Function spaces, ultra-distributions, confinement of symbols, comparable symbols. 1 2 AHMED ABDELJAWAD, SANDRO CORIASCO, AND JOACHIM TOFT Here I is the identity operator on S ′. In particular, by the continuity properties of Opw(a) it follows that H(ω0ω, g) and H(ω, g) possess lifting properties with the homeomorphism Opw(a), and with Opw(b) as its inverse. The existence of a and b in (0.1) is a consequence of solution properties of the evolution equation (∂ta)(t, · ) = (b + log ϑ)#a(t, · ), a(0, · ) = a0 ∈ S(ω, g), ϑ ∈ S(ϑ, g), (0.2) which involve the Weyl product # and a fixed element b ∈ S(1, g). It is proved that (0.2) has a unique solution a(t, · ) which belongs to S(ωϑt, g) (cf. [3, Theorem 6.4] or [34, Theorem 2.6.15]). The existence of a and b in (0.1) will follow by choosing ω = a0 = 1, t = 1 and ϑ = ω0. If g is the constant euclidean metric on the phase space R2d, then S(ω0, g) equals S(ω0)(R2d), the set of all smooth symbols a which satisfies ∂αa . ω0. We notice that also for such simple choices of g, (0.1) above leads to lifting properties that are not trivial. In fact, let ω and ω0 be polynomially moderate weight on the phase space, and let B be a suitable translation invariant BF-space. Then it is observed in [26] that the continuity results for pseudo-differential operators on modulation spaces in [48,50] imply that Opw(a) in (0.1) is continuous and bijective from M (ω0ω, B) to M (ω, B) with continuous inverse Opw(b). In particular, by choosing B to be the mixed norm space Lp,q(R2d) of Lebesgue type, then M (ω, B) is equal to the classical modulation space M p,q (ω), introduced by Feichtinger in [12]. Consequently, Opw(a) above lifts M p,q (ω0ω) into M p,q (ω). An important class of operators in quantum mechanics and time-frequency anal- ysis concerns Toeplitz, or localisation operators. The main issue in [26, 27] is to show that the Toeplitz operator Tp(ω0) lifts M p,q (ω) for suitable ω0. The assumptions on ω0 in [26] is that it should be polynomially moderate and satisfies ω0 ∈ S(ω0). In [27], it is only assumed that ω0 is moderated by a GRS weight, but instead it is here required that ω0 is radial in each phase shift, i. e. ω0 should satisfy (ω0ω) into M p,q ω0(x1, . . . , xd, ξ1, . . . , ξd) = ϑ(r1, . . . , rd), rj = (xj , ξj), for some weight ϑ. The approaches in [26, 27] are also different. In [27], the lifting properties for Tp(ω0) is reached by using the links between modulation spaces and Bargmann- Foch spaces in combination of suitable estimates for a sort of generalised gamma- functions. The approach in [26] relies on corresponding lifting properties for pseudo- differential operators, as follows: (1) Tp(ω0) = Opw(c) for some c ∈ S(ω0); (2) by the definitions, it follows by straightforward computations that if ϑ = ω 1 2 0 , then Tp(ω0) is a homeomorphism from M p,q (ϑ) to M p,q (ϑ); (3) combining (0.1) with Wiener's lemma for (S(1), #) ensures that the inverse of Tp(ω0) in (2) is a pseudo-differential operator Opw(b) with the symbol b in S(1/ω0); (4) by (1), (3) and duality, T1 ≡ Opw(b) ◦ Tp(ω0) and T2 ≡ Tp(ω0) ◦ Opw(b) are both the identity operator on S ′(Rd), since T1 is the identity operator on M p,q (ϑ), T2 is the identity operator on M p,q (1/ϑ), and S ⊆ M p,q (5) by (4), T1 = T2 = Opw(1) is the identity operator on each M p,q (ϑ) ∩ M p,q (ω). Since (1/ϑ). Tp(ω0) = Opw(c) : M p,q (ω0ω) → M p,q (ω) and Opw(b) : M p,q (ω) → M p,q (ω0ω) LIFTINGS FOR MODULATION SPACES, AND ONE-PARAMETER GROUPS OF ΨDO 3 are continuous (cf. [48, 50]) and inverses to each other, it follows that they are homeomorphisms. In the first part of the paper we deduce an analog of (0.1) for the Gevrey type symbol classes Γ(ω0) s and Γ(ω0) 0,s of orders s ≥ 1, the set of all a ∈ C∞ such that ∂αa(X) . hαα!sω(X) for some h > 0, respectively for every h > 0, considered in [4]. That is, in Section 3 we show that there exist symbols a and b such that Opw(a) ◦ Opw(b) = Opw(b) ◦ Opw(a) = I, a ∈ Γ(ω0 , b ∈ Γ(1/ω0 s , (0.3) s are replaced by Γ(ω0) s and Γ(1/ω0) and similarly when Γ(ω0) For general ω0 it is clear that Γ(ω0) 0,s , Γ(ω2) weights ω1, ω2 and ω3 in Γ(ω1) to PE,s(R2d), P 0 , respectively. s ⊆ S(ω0). On the other hand, for the and S(ω3) we always assume that they belong E,s(R2d) and P(R2d), respectively. That is, they should satisfy 0,s ⊆ Γ(ω0) 0,s and Γ(1/ω0) 0,s s s ω1(X + Y ) . ω1(X)er1Y 1 s , ω2(X + Y ) . ω2(X)er2Y 1 s , and ω3(X + Y ) . ω3(X)(1 + Y )N , for some r1 > 0 and N > 0, and every r2 > 0. Since it is clear that P ⊆ P 0 E,s ⊆ PE,s, it follows by straightforward computations that there are admissible a1 ∈ Γ(ω1) and S(ω0), and admissible a2 ∈ Γ(ω2) 0,s which are not contained in any admissible Γ(ω0) which are not contained in any S(ω0). s s As in [3], (0.3) is obtained by proving that the evolution equation (∂ta)(t, · ) = (b + log ϑ)#a(t, · ), , ϑ ∈ Γ(ϑ) , analogous to (0.2), has a unique solution a(t, · ) which belongs to Γ(ωϑt) ilarly when the Γ(ω) Section 3. -spaces are replaced by corresponding Γ(ω) a(0, · ) = a0 ∈ Γ(ω) s s s s (and sim- 0,s -spaces), given in (0.4) In Sections 4 and 5 we use the framework in [26] in combination with (0.3) to extend the lifting properties in [26] in such ways that the involved weights are allowed to belong to P 0 E,s or in PE,s instead of the smaller set P which is the assumption in [26]. Our main result, which is similar to [26, Theorem 0.1], can be stated as follows. Theorem 0.1. Let s ≥ 1, ω, ω0 ∈ P 0 Then the Toeplitz operator Tpφ(ω0) is an isomorphism from M p,q E,s(R2d), p, q ∈ (0, ∞] and let φ ∈ Ss(Rd). (ω)(R2d) onto M p,q (ω/ω0)(R2d). We note that, in contrast to [26, 27], our lifting properties also hold for modula- tion spaces which may fail to be Banach spaces, since p and q in Theorem 0.1 are allowed to be smaller than 1. We will establish several related result. Firstly, the window function may be chosen in certain modulation spaces that are much larger than the Gelfand-Shilov space Ss. Secondly, the theorem holds for a more general family of modulation spaces that includes the classical modulation spaces. Finally, we also establish isomorphisms given by pseudo-differential operators rather than Toeplitz operators. In contrast to [27], we do not impose in Theorem 0.1 and in its related results in Section 5 that ω0 should be radial in each phase shift (cf. e. g. [27, Theorem 4.3]). Summing up, our lifting results in Section 5 extend the lifting results in [26, 27]. The paper is organised as follows. In Section 1 we introduce some notation, and discuss modulation spaces and Gelfand-Shilov spaces of functions and distributions, 4 AHMED ABDELJAWAD, SANDRO CORIASCO, AND JOACHIM TOFT and pseudo-differential calculus. In Section 2 we introduce and discuss basic prop- erties for confinements of symbols in Γ(ω0) 0,s . These considerations are related to the discussions in [3,34], but here adapted to symbols that possess Gevrey regularity, e. g. when the symbols belong to Γ(ω0) and in Γ(ω0) s s or Γ(ω0) 0,s . In contrast to the classical Hormander symbol classes Sr 1,0 and the SG-classes SGm,µ 1,1 , techniques on asymptotic expansions are absent for symbols in the classes Γ(ω0) 0,s , and might be absent for symbols in the general Hormander class s S(m, g). The approach with confinements is, roughly speaking, a sort of stand-in of these absent asymptotic expansion techniques. or in Γ(ω0) In Section 3 we show that the (0.4) has a unique solution with the requested properties, which leads to (0.3). In Sections 4 and 5 we use the results from Sec- tion 3 to deduce lifting properties for modulation spaces under pseudo-differential operators and Toeplitz operators with symbols in Γ(ω0) or in Γ(ω0) 0,s s 1. Preliminaries In this section we recall some basic facts on modulation spaces, Gelfand-Shilov spaces of functions and distributions and pseudo-differential calculus (cf. [11 -- 16,18, 21, 25, 30, 31, 34, 38, 42, 45, 47 -- 51]). 1.1. Weight functions. A weight on Rd is a positive function ω ∈ L∞ that 1/ω ∈ L∞ v-moderate, if loc(Rd) such loc(Rd). If ω and v are weights on Rd, then ω is called moderate or ω(x + y) ≤ Cω(x)v(y), (1.1) for some constant C. The set of all moderate weights on Rd is denoted by PE(Rd). The weight v on Rd is called submultiplicative, if it is even and (1.1) holds for ω = v and C = 1. From now on, v always denote a submultiplicative weight if nothing else is stated. In particular, if (1.1) holds and v is submultiplicative, then it follows by straightforward computations that x, y ∈ Rd, C−1 ω(x) v(y) ≤ ω(x + y) ≤ Cω(x)v(y), (1.2) v(x + y) ≤ v(x)v(y) and v(x) = v(−x) ≥ 1, x, y ∈ Rd. If ω is a moderate weight on Rd, then there is a submultiplicative weight v on Rd such that (1.1) and (1.2) hold (cf. [47,48,52]). Moreover if v is submultiplicative on Rd, then (1.3) for some constant r > 0 (cf. [24]). Here and in what follows we write A(θ) . B(θ), θ ∈ Ω, if there is a constant c > 0 such that A(θ) ≤ cB(θ) for all θ ∈ Ω. In particular, if ω is moderate, then 1 . v(x) . erx ω(x + y) . ω(x)ery and e−rx ≤ ω(x) . erx, x, y ∈ Rd (1.4) for some r > 0. Next we introduce suitable subclasses of PE. Definition 1.1. Let s > 0. The set PE,s(Rd) (P 0 PE(Rd) such that E,s(Rd)) consists of all ω ∈ ω(x + y) . ω(x)ery 1 s , x, y ∈ Rd; (1.5) holds for some (every) r > 0. By (1.4) it follows that P 0 E,s1 = PE,s2 = PE when s1 < 1 and s2 ≤ 1. For convenience we set P 0 E(Rd) = P 0 E,1(Rd). LIFTINGS FOR MODULATION SPACES, AND ONE-PARAMETER GROUPS OF ΨDO 5 1.2. Gelfand-Shilov spaces. We let F be the Fourier transform given by (F f )(ξ) = bf (ξ) ≡ (2π)−d/2ZRd f (x)e−ihx,ξi dx when f ∈ L1(Rd). Here h · , · i denotes the usual scalar product on Rd. Definition 1.2. The Gelfand-Shilov space Sσ Beurling type), σ > 0, s > 0, consists of all f ∈ S (Rd) such that s (Rd) of Roumieu type (Σσ s (Rd) of f (x) . e−rx 1 s and x, ξ ∈ Rd (1.6) 1 (F f )(ξ) . e−rξ σ , s and Σs = Σs s. for some r > 0 (for all r > 0). We set Ss = Ss The classes Sσ s (Rd) and related generalizations were widely studied, and used in the applications to partial differential equations, see for example [2, 6, 9, 28, 35, 38]. We recall the following characterisations of Sσ s (Rd). Proposition 1.3. Let s, σ > 0, p ∈ [1, ∞] and let f ∈ S (Rd). Then the following conditions are equivalent: (1) f ∈ Sσ (2) for some (every) h > 0 it holds s (Rd) (f ∈ Σσ s (Rd)); kxαf kLp . hαα!s and (3) for some (every) h > 0 it holds kξβbf kLp . hββ!σ, α, β ∈ Nd; kxαf kLp . hαα!s and k∂βf kLp . hββ!σ, α, β ∈ Nd; (4) for some (every) h > 0 it holds kxα∂βf (x)kLp . hα+βα!s β!σ, α, β ∈ Nd; (5) for some (every) h, r > 0 it holds ker · 1 s ∂αf kLp . hα(α!)σ α ∈ Nd. Remark 1.4. Let h, s, s0, σ, σ0 ∈ R+ be such that s + σ > 1 and s0 + σ0 ≥ 1, and let Sσ s,h(Rd) be the set of all f ∈ C∞(Rd) such that kf kS σ s,h ≡ sup xβ∂αf (x) hα+βα!σβ!s is finite. Here the supremum is taken over all α, β ∈ Nd and x ∈ Rd. Obviously Sσ s,h(Rd) is a Banach space which increases as h, s and σ increase, and is contained in S (Rd), the set of Schwartz functions on Rd. Furthermore Sσ s,h(Rd) and [h>0 Sσ0 s0,h(Rd) are dense in S (Rd). Hence, the dual (Sσ which contains S ′(Rd). s,h)′(Rd) of Sσ s,h(Rd) is a Banach space The spaces Sσ s (Rd) and Σσ s (Rd) are the inductive and projective limits, respec- tively, of Sσ s,h(Rd) with respect to h. This implies that s,h(Rd) and Σσ Sσ s,h(Rd), Sσ (1.7) Sσ s (Rd) = [h>0 s (Rd) = \h>0 and that the topology for Sσ sion map from Sσ (2) -- (5) in Proposition 1.3 induce the same topology for Sσ s (Rd) is the strongest possible one such that each inclu- s (Rd) is continuous. Moreover, any of the conditions s,h(Rd) to Sσ s (Rd) and Σσ s (Rd). 6 AHMED ABDELJAWAD, SANDRO CORIASCO, AND JOACHIM TOFT The Gelfand-Shilov distribution spaces (Sσ s )′(Rd) and (Σσ s )′(Rd) are the projec- tive and inductive limit respectively of (Sσ s,h)′(Rd). Hence (Sσ s )′(Rd) = \h>0 s )′(Rd) = [h>0 (Sσ s,h)′(Rd) and (Σσ (Sσ s,h)′(Rd). (1.7)′ We have that (Sσ (see [37]). s )′ and (Σσ s )′ are the topological duals of Sσ s and Σσ s , respectively Remark 1.5. Let s, σ > 0. Then Σσ h > 0. Moreover, Sσ only if s + σ ≥ 1 and (s, σ) 6= ( 1 2 , 1 s (Rd) 6= {0} if and only if s + σ ≥ 1, and Σσ s (Rd) is a Fr´echet space with seminorms k · kS σ , s (Rd) 6= {0} if and s,h Σσ s (Rd) ⊆ Sσ If in addition (s, σ) 6= ( 1 s+ε (Rd) ⊆ S (Rd) ⊆ S ′(Rd) ⊆ (Σσ+ε s (Rd) ⊆ Σσ+ε 2 , 1 2 ), then 2 ). Moreover, if ε > 0 and s + σ ≥ 1, then s+ε )′(Rd) ⊆ (Sσ s )′(Rd). (Sσ s )′(Rd) ⊆ (Σσ s )′(Rd). The Gelfand-Shilov spaces are invariant or possess convenient mapping prop- erties under several basic transformations. For example they are invariant under translations, dilations, and under (partial) Fourier transformations. The Fourier transform F on S (Rd) extends uniquely to homeomorphisms on s(Rd), and restricts to homeomorphisms on Ss(Rd) and S ′(Rd), S′ Σs(Rd), and to a unitary operator on L2(Rd). s(Rd) and Σ′ We also recall some mapping properties of Gelfand-Shilov spaces under short- time Fourier transforms. Let φ ∈ S (Rd) be fixed. For every f ∈ S ′(Rd), the short-time Fourier transform Vφf is the distribution on R2d defined by the formula (Vφf )(x, ξ) = F (f φ( · − x))(ξ) = (f, φ( · − x)eih · ,ξi). (1.8) We recall that if T (f, φ) ≡ Vφf when f, φ ∈ S1/2(Rd), then T is uniquely extendable to sequentially continuous mappings T : S′ s(Rd) × Ss(Rd) → S′ T : S′ s(Rd) × S′ s(Rd) → S′ s(R2d), s(R2d)\ C∞(R2d), and similarly when Ss and S′ S ′, respectively (cf. [7, 52]). We also note that Vφf takes the form s are replaced by Σs and Σ′ s, respectively, or by S and Vφf (x, ξ) = (2π)−d/2ZRd f (y)φ(y − x)e−ihy,ξi dy (1.8)′ (ω)(Rd) for some ω ∈ PE(Rd), φ ∈ Σ1(Rd) and p ≥ 1. Here Lp when f ∈ Lp when p ∈ (0, ∞] and ω ∈ PE(Rd), is the set of all f ∈ Lp Lp(Rd). (ω)(Rd), loc(Rd) such that f · ω ∈ 1.3. Suitable function classes with Gelfand-Shilov regularity. The next re- sult shows that for any ω ∈ PE(Rd) one can find an equivalent weight ω0 which satisfies similar regularity estimates as functions in Gelfand-Shilov spaces. Proposition 1.6. Let ω ∈ PE(Rd) and s > 0. Then there is an ω0 ∈ PE(Rd) ∩ C∞(Rd) such that the following is true: (1) ω0 ≍ ω; (2) ∂αω0(x) . ω0(x)hαα!s ≍ ω(x)hαα!s for every h > 0. where the last inequality follows from (1.4) and the fact that φ is bounded by a super exponential function. This gives the first part of (2). The equivalences in (1) follows in the same way as in [52]. More precisely, by (1.4) we have In the same way, (1.4) gives ω0(x) =Z ω(y)φ(x − y) dy =Z ω(x + (y − x))φ(x − y) dy . ω(x)Z ecx−yφ(x − y) dy ≍ ω(x). ω0(x) =Z ω(y)φ(x − y) dy =Z ω(x + (y − x))φ(x − y) dy & ω(x)Z e−cx−yφ(x − y) dy ≍ ω(x), LIFTINGS FOR MODULATION SPACES, AND ONE-PARAMETER GROUPS OF ΨDO 7 Proof. We may assume that s < 1. It suffices to prove that (2) should hold for some h > 0. Let φ0 ∈ Σs 1−s(Rd), giving that 1−s(Rd) \ 0, and let φ = φ02. Then φ ∈ Σs ∂αφ(x) . hαe−cx for every h > 0 and c > 0. Now let ω0 = ω ∗ φ. 1 1−s α!s, We have ∂αω0(x) =(cid:12)(cid:12)(cid:12)(cid:12)Z ω(y)(∂αφ)(x − y) dy(cid:12)(cid:12)(cid:12)(cid:12) . hαα!sZ ω(y)e−cx−y . hαα!sZ ω(x + (y − x))e−cx−y . hαα!sω(x)Z e− c 1−s dy 1 1 1 1−s dy 2 x−y 1−s dy ≍ hαα!sω(x), and (1) as well as the second part of (2) follow. (cid:3) A weight ω0 which satisfies Proposition 1.6 (2) is called elliptic or s-elliptic. Definition 1.7. Let s ≥ 0 and ω ∈ PE(Rd). The class Γ(ω) consists of all f ∈ C∞(Rd) such that s (Rd) (Γ(ω) 0,s (Rd)) Dαf (x) . ω(x) hαα!s, x ∈ Rd, for some h > 0 (for every h > 0). Evidently, by Proposition 1.6 it follows that the family of symbol classes in Definition 1.7 is not reduced when the assumption ω ∈ PE(R2d) is replaced by the more restrictive assumption ω ∈ PE,s(R2d) or by ω ∈ P 0 E,s(R2d). By similar arguments as in the proof of Proposition 1.6 we get the following analog of Proposition 2.3.16 in [33]. The details are left for the reader. Proposition 1.8. Let s > 1/2, ω ∈ PE(R2d), and φ ∈ Σs(R2d). Then ω ∗ φ belongs to Γ(ω) 0,s . The following definition is motivated by Lemma 2.6.13 in [33]. Definition 1.9. Let s ≥ 1, ω ∈ PE(Rd) and ϑ0 = 1 + log ω. Then c is called comparable to ω with respect to s ≥ 1 if 8 AHMED ABDELJAWAD, SANDRO CORIASCO, AND JOACHIM TOFT (1) kc − log ωkL∞ < ∞ (2) c ∈ Γ(ϑ0) (Rd) and ∂αc ∈ Γ(1) s s (Rd), when α = 1. Proposition 1.10. Let ω, v ∈ PE(Rd) be such that v is submultiplicative and ω is v-moderate. Also let v1(x) ≡ 1 + log v(x) and ω1(x) ≡ 1 + log ω(x). Then v1 is submultiplicative and ω1 is v1-moderate, satisfying (1.1) with 1+log C ≥ 1 in place of C. Proof. By (1.2) we get v1(x + y) = 1 + log v(x + y) ≤ 1 + log v(x) + log v(y) ≤ (1 + log v(x)) (1 + log v(y)) = v1(x) v1(y), and ω1(x + y) = 1 + log ω(x + y) ≤ 1 + log C + log ω(x) + log v(y) ≤ (1 + log C)(1 + log ω(x)) (1 + log v(y)) ≤ (1 + log C) ω1(x) v1(y), as claimed. (cid:3) Lemma 1.11. Let s ≥ 1, ω ∈ PE(Rd) and ϑ0 = 1 + log ω. Then the following is true: (1) there exists an elliptic weight ω0 ∈ PE(Rd) ∩ Γ(ω) s (Rd) such that ω ≍ ω0 and 1 + log ω0 ∈ PE(Rd) ∩ Γ(ϑ0) s (Rd); (2) there exists an element c which is comparable with ω0. Proof. If ω1(x) is equal to ω(x) when ω(x) ≥ e or ω(x) ≤ e−1, and 3 otherwise, then ω1 is equivalent to ω. The result now follows from Proposition 1.6 and its proof, with ω1 in place of ω. (cid:3) 1.4. Modulation spaces. Let φ ∈ Σ1(Rd) \ 0, p, q ∈ (0, ∞] and ω ∈ PE(R2d) be fixed. Then the modulation space M p,q (ω)(Rd) consists of all f ∈ Σ′ 1(Rd) such that kf kM p,q (ω) ≡(cid:16)Z (cid:16)Z Vφf (x, ξ)ω(x, ξ)p dx(cid:17)q/p dξ(cid:17)1/q < ∞ (1.9) (with the obvious modifications when p = ∞ and/or q = ∞). We set M p and if ω = 1, then we set M p,q = M p,q (ω) and M p = M p (ω). (ω) = M p,p (ω), The following proposition is a consequence of well-known facts in [12, 20, 21, 54]. Here and in what follows, we let p′ denotes the conjugate exponent of p, i. e. ∞ p p − 1 when p ∈ (0, 1] when p ∈ (1, ∞) 1 when p = ∞ . p′ = Proposition 1.12. Let p, q, pj, qj, r ∈ (0, ∞] be such that r ≤ min(1, p, q), j = 1, 2, let ω, ω1, ω2, v ∈ PE(R2d) be such that ω is v-moderate, φ ∈ M r (v)(Rd) \ 0, and let f ∈ Σ′ 1(Rd). Then the following is true: (1) f ∈ M p,q (ω)(Rd) if and only if (1.9) holds, i. e. M p,q of the choice of φ. Moreover, M p,q (1.9), and different choices of φ give rise to equivalent norms; (ω)(Rd) is independent (ω) is a Banach space under the norm in LIFTINGS FOR MODULATION SPACES, AND ONE-PARAMETER GROUPS OF ΨDO 9 (2) if p1 ≤ p2, q1 ≤ q2 and ω2 ≤ Cω1 for some constant C, then Σ1(Rd) ⊆M p1,q1 (ω1) (Rd) ⊆M p2,q2 (ω2) (Rd) ⊆ Σ′ 1(Rd). Proposition 1.12 (1) allows us to be rather vague about to the choice of φ ∈ (v) \ 0 in (1.9). For example, if C > 0 is a constant and Ω is a subset of Σ′ M r 1, then ≤ C for every a ∈ Ω, means that the inequality holds for some choice of kakM p,q φ ∈ M r (v) \ 0, a similar inequality is true although C may have to be replaced by a larger constant, if necessary. (v) \ 0 and every a ∈ Ω. Evidently, for any other choice of φ ∈ M r (ω) Let s, t ∈ R. Then the weights (x, ξ) 7→ hξishxit and (x, ξ) 7→ h(x, ξ)is, (1.10) are common in the applications. It follows that they belong to P(R2d) for every If ω ∈ P(R2d), then ω is moderated by any of the weights in (1.10) s, t ∈ R. provided s and t are chosen large enough. Remark 1.13. For modulation spaces of the form M p,q (ω) with fixed p, q ∈ [1, ∞] the norm equivalence in Proposition 1.12(1) can be extended to a larger class of windows. In fact, assume that ω, v ∈ PE(R2d) with ω being v-moderate and x, ξ ∈ Rd, 1 ≤ r ≤ min(p, p′, q, q′) . Let φ ∈ M r to M p,q φ ∈ M r in [51].) (ω)(Rd), if and only if Vφf ∈ Lp,q (v)(Rd) \ {0} in kVφf kLp,q (ω) (v)(Rd) \ {0}. Then a Gelfand-Shilov distribution f ∈ Σ′ 1(Rd) belongs (ω)(R2d). Furthermore, different choices of give rise to equivalent norms. (Cf. Theorem 3.1 1.5. A broader family of modulation spaces. As announced in the introduc- tion we consider in Section 2 mapping properties for pseudo-differential operators when acting on a broader class of modulation spaces, which are defined by impos- ing certain types of translation invariant solid BF-space norms on the short-time Fourier transforms. (Cf. [11 -- 16, 18].) Definition 1.14. Let B ⊆ Lr loc(Rd) be a quasi-Banach of order r ∈ (0, 1], and let v ∈ PE(Rd). Then B is called a translation invariant Quasi-Banach Function space on Rd, or invariant QBF space on Rd, if there is a constant C such that the following conditions are fulfilled: (1) if x ∈ Rd and f ∈ B, then f ( · − x) ∈ B, and kf ( · − x)kB ≤ Cv(x)kf kB; (1.11) (2) if f, g ∈ Lr loc(Rd) satisfy g ∈ B and f ≤ g, then f ∈ B and kf kB ≤ CkgkB. If v belongs to PE,s(Rd) (P 0 E,s(Rd)) , then B in Definition 1.14 is called an invariant BF-space of Roumieu type (Beurling type) of order s. We notice that the quasi-norm k · kB in Definition 1.14 should satisfy kf + gkB ≤ 2 1 r −1(kf kB + kgkB) f, g ∈ B. (1.12) By Aoki and Rolewi´c in [1, 39] it follows that there is an equivalent quasi-norm to the previous one which additionally satisfies kf + gkr B ≤ kf kr B + kgkr B f, g ∈ B. (1.13) From now on we suppose that the quasi-norm of B has been chosen such that both (1.12) and (1.13) hold true. 10 AHMED ABDELJAWAD, SANDRO CORIASCO, AND JOACHIM TOFT It follows from (2) in Definition 1.14 that if f ∈ B and h ∈ L∞, then f · h ∈ B, and kf · hkB ≤ Ckf kBkhkL∞. (1.14) If r = 1, then B in Definition 1.14 is a Banach space, and the condition (2) means that a translation invariant QBF-space is a solid BF-space in the sense of (A.3) in [15]. The space B in Definition 1.14 is called an invariant BF-space (with respect to v) if r = 1, and Minkowski's inequality holds true, i. e. kf ∗ ϕkB ≤ Ckf kBkϕkL1 (v) , f ∈ B, ϕ ∈ Σ1(Rd) (1.15) for some constant C which is independent of f ∈ B and ϕ ∈ Σ1(Rd). Example 1.15. Assume that p, q ∈ [1, ∞], and let Lp,q f ∈ L1 loc(R2d) such that 1 (R2d) be the set of all if finite. Also let Lp,q 2 (R2d) be the set of all f ∈ L1 loc(R2d) such that kf kLp,q 1 kf kLp,q 2 ≡(cid:16)Z (cid:16)Z f (x, ξ)p dx(cid:17)q/p ≡(cid:16)Z (cid:16)Z f (x, ξ)q dξ(cid:17)p/q dξ(cid:17)1/q dx(cid:17)1/p is finite. Then it follows that Lp,q respect to v = 1. 1 and Lp,q 2 are translation invariant BF-spaces with For translation invariant BF-spaces we make the following observation. Proposition 1.16. Assume that v ∈ PE(Rd), and that B is an invariant BF- space with respect to v such that (1.15) holds true. Then the convolution mapping 0 (Rd) × B to B extends uniquely to a continuous mapping (ϕ, f ) 7→ ϕ ∗ f from C∞ (v)(Rd). from L1 (v)(Rd) × B to B, and (1.15) holds true for any f ∈ B and ϕ ∈ L1 The result is a straightforward consequence of the fact that C∞ (v). 0 The quasi-Banach space B above is usually a mixed quasi-normed Lebesgue space, given as follows. Let E be a non-degenerate parallelepiped in Rd which is spanned by the ordered basis κ(E) = {e1, . . . , ed}. That is, is dense in L1 E = { x1e1 + · · · + xded ; (x1, . . . , xd) ∈ Rd, 0 ≤ xk ≤ 1, k = 1, . . . , d }. The corresponding lattice, dual parallelepiped and dual lattice are given by ΛE = { j1e1 + · · · + jded ; (j1, . . . , jd) ∈ Zd }, E′ = { ξ1e′ 1 + · · · + ξde′ d ; (ξ1, . . . , ξd) ∈ Rd, 0 ≤ ξk ≤ 1, k = 1, . . . , d }, and Λ′ E = ΛE ′ = { ι1e′ 1 + · · · + ιde′ d ; (ι1, . . . , ιd) ∈ Zd }, 1, . . . , e′ respectively, where the ordered basis κ(E′) = {e′ d} of E′ satisfies hej, e′ ki = 2πδjk for every j, k = 1, . . . , d. Note here that the Fourier analysis with respect to general biorthogonal bases has recently been developed in [41]. The basis e′ 1, . . . , e′ d is called the dual basis of e1, . . . , ed. We observe that there d are the images of the standard is a matrix TE such that e1, . . . , ed and e′ basis under TE and TE ′ = 2π(T −1 1, . . . , e′ E )t, respectively. In the following we let max q = max(q1, . . . , qd) and min q = min(q1, . . . , qd) when q = (q1, . . . , qd) ∈ (0, ∞]d, and χΩ be the characteristic function of Ω. LIFTINGS FOR MODULATION SPACES, AND ONE-PARAMETER GROUPS OF ΨDO 11 Definition 1.17. Let E be a non-degenerate parallelepiped in Rd spanned by the ordered set κ(E) ≡ {e1, . . . , ed} in Rd, p = (p1, . . . , pd) ∈ (0, ∞]d and r = min(1, p). If f ∈ Lr loc(Rd), then kf kLp κ(E) ≡ kgd−1kLpd (R) where gk(zk), zk ∈ Rd−k, k = 0, . . . , d − 1, are inductively defined as g0(x1, . . . , xd) ≡ f (x1e1 + · · · + xded), and gk(zk) ≡ kgk−1( · , zk)kLpk (R), k = 1, . . . , d − 1. loc(Rd) such that kf kLp κ(E)(Rd) consists of all f ∈ Lr The space Lp is called E-split Lebesgue space (with respect to p and κ(E)). Definition 1.18. Let E0 ⊆ Rd be a non-degenerate parallelepiped with dual parallelepiped E′ 0, and E ⊆ R2d be a parallelepiped spanned by the ordered set κ(E) ≡ {e1, . . . , e2d}. Then E is called a phase-shift split parallelepiped (with re- spect to E0) if E is non-degenerate and d of the vectors in {e1, . . . , e2d} span E0 and the other d vectors is the corresponding dual basis which span E′ 0. is finite, and κ(E) Next we consider the extended class of modulation spaces which we are interested in. Definition 1.19. Assume that B is a translation invariant QBF-space on R2d, ω ∈ PE(R2d), and that φ ∈ Σ1(Rd) \ 0. Then the set M(ω) = M (ω, B) consists of all f ∈ Σ′ 1(Rd) such that kf kM(ω) = kf kM(ω,B) ≡ kVφf ωkB is finite. Obviously, we have M p,q (ω)(Rd) = M (ω, B) when B = Lp,q 1 (R2d) (cf. Example 1.15). It follows that many properties which are valid for the classical modulation spaces also hold for the spaces of the form M (ω, B). For example we have the following proposition, which shows that the definition of M (ω, B) is independent of the choice of φ when B is a Banach space. We omit the proof since it follows by similar arguments as in the proof of Proposition 11.3.2 in [21]. (See also [36] for topological aspects of M (ω, B).) Proposition 1.20. Let B be an invariant BF-space with respect to v0 ∈ PE(R2d) for j = 1, 2. Also let ω, v ∈ PE(R2d) be such that ω is v-moderate, M (ω, B) is the same as in Definition 1.19, and let φ ∈ M 1 1(Rd). Then M (ω, B) is a Banach space, and f ∈ M (ω, B) if and only if Vφf ω ∈ B, and different choices of φ gives rise to equivalent norms in M (ω, B). (v0v)(Rd) \ 0 and f ∈ Σ′ We refer to [11 -- 16, 18, 20, 21, 40, 54] for more facts about modulation spaces. 1.6. Pseudo-differential operators. We use the notation M(d, Ω) for the set of d×d-matrices with entries in the set Ω. Let s ≥ 1/2, a ∈ Ss(R2d), and A ∈ M(d, R) be fixed. Then, the pseudo-differential operator OpA(a) is the linear and continuous operator on Ss(Rd) given by (OpA(a)u)(x) = (2π)−dZZ a(x − A(x − y), ξ) f (y) eihx−y,ξi dydξ (1.16) when f ∈ Ss(Rd). For general a ∈ S′ OpA(a) is defined as the continuous operator from Ss(Rd) to S′ tribution kernel given by s(R2d), the pseudo-differential operator s(Rd) with dis- Ka,A(x, y) = (2π)−d/2(F −1 2 a)(x − A(x − y), x − y). (1.17) 12 AHMED ABDELJAWAD, SANDRO CORIASCO, AND JOACHIM TOFT Here F2F is the partial Fourier transform of F (x, y) ∈ S′ y variable. This definition makes sense, since the mappings s(R2d) with respect to the F2 and F (x, y) 7→ F (x − A(x − y), y − x) (1.18) s(R2d). In particular, the map a 7→ Ka,A is a homeo- are homeomorphisms on S′ morphism on S′ s(R2d). Remark 1.21. For any K ∈ S′ ping from Ss(Rd1) to S′ s(Rd2), defined by the formula s(Rd1+d2), let TK be the linear and continuous map- (TKf, g)L2(Rd2 ) = (K, g ⊗ f )L2(Rd1+d2 ). (1.19) It is well-known that if t ∈ R, then it follows from e. g. [5, 32] that the Schwartz kernel theorem also holds in the context of Gelfans-Shilov spaces. That is, the mappings K 7→ TK and a 7→ Opt(a) are bijective from S′ s(R2d) to the set of linear s(Rd). Similar facts hold true if Ss and and continuous mappings from Ss(Rd) to S′ s are replaced by Σs and Σ′ S′ As a consequence of Remark 1.21 it follows that for each a1 ∈ S′ s, respectively (or by S and S ′, respectively). s(R2d) and s(R2d) such that OpA1(a1) = OpA2 (a2). A1, A2 ∈ M(d, R), there is a unique a2 ∈ S′ The relation between a1 and a2 is given by OpA1(a1) = OpA2 (a2) ⇔ a2(x, ξ) = eih(A1−A2)Dξ,Dxia1(x, ξ). (1.20) (Cf. [30].) Note here that the right-hand side makes sense, since it is equivalent to S′ s when A ∈ M(d, R). ba2(ξ, x) = ei(A1−A2)hx,ξiba1(ξ, x), and that the map a 7→ eihAx,ξia is continuous on s(R2d) be fixed. Then a is called a rank-one element with respect to A, if the corresponding pseudo-differential operator is of rank-one, i. e. Let A ∈ M(d, R) and a ∈ S′ for some f1, f2 ∈ S′ fulfilled, if and only if a = (2π)d/2W A defined by the formula f1,f2 OpA(a)f = (f, f2)f1, (1.21) s(Rd). By straightforward computations it follows that (1.21) is it the A-Wigner distribution , where W A f ∈ Ss(Rd), f1,f2 W A f1,f2(x, ξ) ≡ F (f1(x + A · )f2(x − (I − A) · ))(ξ), (1.22) which takes the form W A f1,f2(x, ξ) = (2π)−d/2Z f1(x + Ay)f2(x − (I − A)y)e−ihy,ξi dy, when f1, f2 ∈ Ss(Rd). Here I ∈ M(d, R) is the identity matrix. By combining these facts with (1.20) it follows that W A2 f1,f2 = eih(A1−A2)Dξ,DxiW A1 f1,f2 , (1.23) for each f1, f2 ∈ S′ important, we set W A )Wigner distribution of f1 and f2. s(Rd) and A1, A2 ∈ M(d, R). Since the Weyl case is particularly 2 I, i. e. Wf1,f2 is the usual (cross- f1,f2 = Wf1,f2 when A = 1 For future references we note the link (OpA(a)f, g)L2(Rd) = (2π)−d/2(a, W A g,f )L2(R2d), a ∈ S′ s(R2d) and f, g ∈ Ss(Rd) (1.24) between pseudo-differential operators and Wigner distributions, which follows by straightforward computations (see also e. g. [55]). Next we discuss the Weyl product, the twisted convolution and related objects. Let s ≥ 1/2 and let a, b ∈ S′ s(R2d). Then the Weyl product a#b between a and b is the function or distribution which fulfills Opw(a#b) = Opw(a) ◦ Opw(b), provided LIFTINGS FOR MODULATION SPACES, AND ONE-PARAMETER GROUPS OF ΨDO 13 the right-hand side makes sense as a continuous operator from Ss(Rd) to S′ More generally, if A ∈ M(d, R), then the product #A is defined by the formula s(Rd). OpA(a#Ab) = OpA(a) ◦ OpA(b), (1.25) provided the right-hand side makes sense as a continuous operator from Ss(Rd) to S′ s(Rd), in which case a and b are called suitable or admissible. We also use the notation # instead of #A when A = 1 2 I (i. e. in the Weyl case). The Weyl product can also, in a convenient way, be expressed in terms of the symplectic Fourier transform and the twisted convolution. More precisely, let s ≥ 1/2. Then the symplectic Fourier transform for a ∈ Ss(R2d) is defined by the formula (Fσa)(X) = π−dZ a(Y )e2iσ(X,Y ) dY, where σ is the symplectic form given by σ(X, Y ) = hy, ξi − hx, ηi, X = (x, ξ) ∈ R2d, Y = (y, η) ∈ R2d. We note that Fσ = T ◦(F ⊗(F −1)), when (T a)(x, ξ) = a(ξ, x). In particular, Fσ is continuous on Ss(R2d), and extends uniquely to a homeomorphism on S′ s(R2d), and to a unitary map on L2(R2d), since similar facts hold for F . Furthermore, F 2 σ is the identity operator. Let s ≥ 1/2 and a, b ∈ Ss(R2d). Then the twisted convolution of a and b is defined by the formula (a ∗σ b)(X) =(cid:18) 2 π(cid:19) d 2Z a(X − Y )b(Y )e2iσ(X,Y ) dY. The definition of ∗σ extends in different ways. For example, it extends to a con- tinuous multiplication on Lp(R2d) when p ∈ [1, 2], and to a continuous map from s(R2d), then a#b makes sense if and only S′ s(R2d) × Ss(R2d) to S′ s(R2d). If a, b ∈ S′ (1.26) (1.27) (1.28) if a ∗σbb makes sense, and then a#b = (2π)− d 2 a ∗σ (Fσb). We also remark that for the twisted convolution we have Fσ(a ∗σ b) = (Fσa) ∗σ b = a ∗σ (Fσb), where a(X) = a(−X) (cf. [46, 51, 53]). A combination of (1.27) and (1.28) gives (1.29) We now define the subspace of symbols in S ′(R2d) which give rise to L2(Rd)- 2 (Fσa) ∗σ (Fσb). Fσ(a#b) = (2π)− d bounded pseudo-differential operators, which will be useful in the sequel. Definition 1.22. The set sw is linear and continuous on L2(Rd), and we set ∞(R2d) consists of all a ∈ S ′(R2d) such that Opw(a) kaksw ∞ ≡ k Opw(a)kL2(Rd)→L2(Rd). Remark 1.23. By Remark 1.21 it follows that the map a 7→ Opw(a) is an isometric bijection from sw ∞(R2d) to the set of linear continuous operators on L2(Rd). s and s > 1 2 are replaced by Σs, Σ′ Remark 1.24. We remark that the relations in this subsection hold true after Ss, s and s ≥ 1 S′ Next we recall some algebraic properties and characterisations of Γ(ω) Γ(ω) 0,s (R2d), and begin with the following. We refer to [4] for its proof. Proposition 1.25. Let s ≥ 1, ωj ∈ P 0 let ω0,r(X, Y ) = ω0(X)e−rY E,s(R2d), Aj ∈ M(d, R) for j = 1, 2, and s when r > 0. Then the following is true: 2 respectively, in each place. (R2d) and s 1 14 AHMED ABDELJAWAD, SANDRO CORIASCO, AND JOACHIM TOFT (1) If a1, a2 ∈ Σ′ s(R2d) satisfy OpA1(a1) = OpA2 (a2), then a1 ∈ Γ(ω0) s (R2d) if (R2d). (2) Γ(ω1) (3) Γ(ω0) and only if a2 ∈ Γ(ω0) s ⊆ Γ(ω1ω2) s #Γ(ω2) s . s = [r>0 s M ∞,1 (1/ω0,r) = [r≥0 M∞,1 (1/ω0,r). Proposition 1.26. Let s ≥ 1, ωj ∈ PE,s(R2d), Aj ∈ M(d, R) for j = 1, 2, and let ω0,r(X, Y ) = ω0(X)e−rY s when r > 0. Then the following is true: 1 (1) If a1, a2 ∈ Σ′ s(R2d) satisfy OpA1(a1) = OpA2 (a2), then a1 ∈ Γ(ω0) 0,s (R2d) if (2) Γ(ω1) (3) Γ(ω0) and only if a2 ∈ Γ(ω0) 0,s ⊆ Γ(ω1ω2) 0,s #Γ(ω2) . 0,s (R2d). 0,s = \r>0 0,s M ∞,1 (1/ω0,r) = \r≥0 M∞,1 (1/ω0,r). In time-frequency analysis one also considers mapping properties for pseudo- differential operators between modulation spaces or with symbols in modulation spaces. Especially we need the following two results, where the first one is a gen- eralisation of [44, Theorem 2.1] by Tachizawa, and the second one is a weighted version of [21, Theorem 14.5.2]. We refer to [56] for the proof of the first two propositions and to [56] for the proof of the third one. E,s(R2d), Proposition 1.27. Assume that A ∈ M(d, R), s ≥ 1, ω, ω0 ∈ P 0 a ∈ Γ(ω) (R2d), and that B is an invariant BF-space on R2d of Beurling type. Then OpA(a) is continuous from M (ω0ω, B) to M (ω0, B), and also continuous on Ss(Rd) and on S′ s s(Rd). Proposition 1.28. Assume that A ∈ M(d, R), s ≥ 1, ω, ω0 ∈ PE,s(R2d), a ∈ Γ(ω) 0,s (R2d), and that B is an invariant BF-space on R2d of Roumieu type. Then OpA(a) is continuous from M (ω0ω, B) to M (ω0, B), and also continuous on Σs(Rd) and on Σ′ s(Rd). Proposition 1.29. Assume that p, q ∈ (0, ∞], r ≤ min(p, q, 1), ω ∈ PE(R2d ⊕ R2d) and ω1, ω2 ∈ PE(R2d) satisfy ω2(X − Y ) ω1(X + Y ) ≤ Cω(X, Y ), X, Y ∈ R2d, (1.30) for some constant C. If a ∈ M∞,r tinuous map from M p,q (ω1)(Rd) to M p,q (ω2)(Rd). (ω) (R2d), then Opw(a) extends uniquely to a con- Finally we need the following result concerning mapping properties of modulation spaces under the Weyl product. The result is a special case of Theorem 0.3 in [10]. Proposition 1.30. Assume that ωj ∈ PE(R2d ⊕ R2d) for j = 0, 1, 2 satisfy ω0(X, Y ) ≤ Cω1(X − Y + Z, Z)ω2(X + Z, Y − Z), (1.31) for some constant C > 0 independent of X, Y, Z ∈ R2d. Then the map (a, b) 7→ a#b from Σ1(R2d) × Σ1(R2d) to Σ1(R2d) extends uniquely to a mapping from M∞,1 (ω1)(R2d) × M∞,1 In the proof of our main theorem we will need the following consequence of (ω2)(R2d) to M∞,1 (ω0)(R2d). Proposition 1.30. LIFTINGS FOR MODULATION SPACES, AND ONE-PARAMETER GROUPS OF ΨDO 15 Proposition 1.31. Let s ≥ 1, ω0, v0, v1 ∈ P 0 v0-moderate. Set ϑ = ω1/2 , and 0 E,s(R2d ⊕ R2d) be such that ω0 is ω1(X, Y ) = v0(2Y )1/2v1(2Y ) ϑ(X + Y )ϑ(X − Y ) , ω2(X, Y ) = ϑ(X − Y )ϑ(X + Y )v1(2Y ) , v2(X, Y ) = v1(2Y ) . Then Γ(1/ϑ) s #M∞,1 (ω1)#Γ(1/ϑ) s Γ(1/ϑ) s #M∞,1 (v2)#Γ(1/ϑ) s ⊆ M∞,1 (v2) , ⊆ M∞,1 (ω2) . The same holds with PE,s and Γ(1/ϑ) 0,s at each occurrence. in place of P 0 E,s and Γ(1/ϑ) s (1.32) (1.33) (1.34) respectively, Proof. Since Γ(1/ϑ) it suffices to argue with the symbol class M∞,1 of Γ(1/ϑ) =Sr≥0 M∞,1 . s s (ϑr) for some sufficiently large r instead (ϑr) with ϑr(X, Y ) = ϑ(X)erY 1 s (Proposition 1.25(3)), Introducing the intermediate weight ω3(X, Y ) = v1(2Y )ϑ(X + Y ) ω0(X − Y ) . we will show that for suitable r ω3(X, Y ) ≤ Cω1(X − Y + Z, Z)ϑr(X + Z, Y − Z) v1(2Y ) ≤ Cϑr(X − Y + Z, Z)ω3(X + Z, Y − Z) . (1.35) (1.36) Proposition 1.30 applied to (1.35) shows that M∞,1 implies that Γ(1/ϑ) Since ϑ is v1/2 -moderate and v0 ∈ P 0 (v2), and (1.33) follows. (ω3) ⊆ M∞,1 #M∞,1 E,s, we have 0 s (ω1)#Γ(1/ϑ) s ⊆ M∞,1 (ω3), and (1.36) ϑ(X − Y )−1 ≤ v0(2Z)1/2ϑ(X − Y + 2Z)−1 and ϑ(X + Y ) ≤ ϑ(X + Z)erY −Z 1 s for suitable r > 0. Using these inequalities repeatedly in the following, a straight- forward computation yields ω3(X, Y ) = v1(2Y )ϑ(X + Y ) ϑ(X − Y )2 ≤ C1 v0(2Z)1/2v1(2Z)ϑ(X + Z)erY −Z 1 s ϑ(X − Y + 2Z)ϑ(X − Y ) = C1ω1(X − Y + Z, Z)ϑr(X + Z, Y − Z), for some C1 > 0 and r > 0. 16 AHMED ABDELJAWAD, SANDRO CORIASCO, AND JOACHIM TOFT Likewise we obtain v1(2Y ) = ϑ(X − Y )v1(2Y )ϑ(X − Y ) ϑ(X − Y )2 ≤ C1 ≤ C2 ϑ(X − Y )v0(2Y )1/2v1(2Y )ϑ(X + Y ) ϑ(X − Y )2 ϑ(X − Y + Z)erZ 1 s v0(2(Y − Z))1/2v1(2(Y − Z))ϑ(X + Y ) ϑ(X − Y + 2Z)2 = C2ϑr(X − Y + Z, Z)ω3(X + Z, Y − Z) . The twisted convolution relation (1.34) is proved similarly. Let be the intermediate weight. Then the inequality ω4(X, Y ) = ϑ(X − Y )v1(2Y ) ω4(X, Y ) = ϑ(X − Y )v1(2Y ) ≤ Cϑ(X − Y + Z)erZ 1 s v1(2(Y − Z)) = Cϑr(X − Y + Z, Z)v2(X + Z, Y − Z) implies that Γ(1/ϑ) s #M∞,1 (v2) ⊆ M∞,1 (ω4). Similarly we obtain ω2(X, Y ) ≤ Cϑ(X − Y )v1(2Z)ϑ(X + Z)erZ−Y 1 s = Cω4(X − Y + Z, Z)ϑ(X + Z)erZ−Y 1 s = Cω4(X − Y + Z, Z)ϑr(X + Z, Y − Z), and thus M∞,1 (ω4)#Γ(1/ϑ) s ⊆ M∞,1 (ω2). The case PE,s and Γ(1/ϑ) 0,s in place of P 0 E,s and Γ(1/ϑ) s occurrence, is treated in similar ways and is left for the reader. respectively, at each (cid:3) 1.7. The Wiener Algebra Property. As a further crucial tool in our study of the isomorphism property of Toeplitz operators we need to combine these continu- ity results with convenient invertibility properties. The so-called Wiener algebra property of certain symbol classes asserts that the inversion of a pseudo-differential operator preserves the symbol class and is often referred to as the spectral invariance of a symbol class. Proposition 1.32. Let A ∈ M(d, R). Then the following are true: (1) If s > 1, a ∈ Γ(1) 0,s(R2d) and OpA(a) is invertible on L2(Rd), then OpA(a)−1 = OpA(b) for some b ∈ Γ(1) 0,s(R2d). (2) If s ≥ 1, a ∈ Γ(1) s (R2d) and OpA(a) is invertible on L2(Rd), then OpA(a)−1 = OpA(b) for some b ∈ Γ(1) s (R2d). (3) If s ≥ 1, v0 ∈ P 0 E,s(R2d) is submultiplicative, v(X, Y ) ≡ v0(Y ), X, Y ∈ (v) (R2d) and OpA(a) is invertible on L2(Rd), then OpA(a)−1 = R2d, a ∈ M ∞,1 OpA(b), for some b ∈ M ∞,1 (v) (R2d). (4) If s > 1, v0 ∈ PE,s(R2d) is submultiplicative, v(X, Y ) ≡ v0(Y ), X, Y ∈ (v) (R2d) and OpA(a) is invertible on L2(Rd), then OpA(a)−1 = R2d, a ∈ M ∞,1 OpA(b), for some b ∈ M ∞,1 (v) (R2d). LIFTINGS FOR MODULATION SPACES, AND ONE-PARAMETER GROUPS OF ΨDO 17 Proof. The results follows essentially from [22, Corollary 5.5] or [23]. More precisely, Suppose s > 1, a ∈ Γ(1) s (R2d), OpA(a) is invertible on L2(Rd), and let vr(X, Y ) = erY f rac1s when r ≥ 0. Then a ∈ M ∞,1 (vr) (R2d) for some r > 0. By [22, Corollary 5.5], Op(M ∞,1 (vr) (R2d)) is a Wiener algebra, giving that Op(a)−1 = Op(b) for some b ∈ M ∞,1 s (R2d). This gives (2) in the case s > 1. (vr) (R2d) ⊆ Γ(1) If instead s = 1, then by [17, Theorem 4.4] there is an r0 > 0 such that Op(a)−1 = Op(b) for some b ∈ M ∞,1 (vr0 )(R2d) ⊆ Γ(1) 1 (R2d), and (2) follows for general s ≥ 1. By similar arguments, (1), (3) and (4) follow. The details are left for the reader. (cid:3) 1.8. Toeplitz Operators. Fix a symbol a ∈ S (R2d) and a window ϕ ∈ S (Rd). Then the Toeplitz operator Tpϕ(a) is defined by the formula (Tpϕ(a)f1, f2)L2(Rd) = (a Vϕf1, Vϕf2)L2(R2d) , (1.37) when f1, f2 ∈ S (Rd). Obviously, Tpϕ(a) is well-defined and extends uniquely to a continuous operator from S ′(Rd) to S (Rd). The definition of Toeplitz operators can be extended to more general classes of windows and symbols by using appropriate estimates for the short-time Fourier transforms in (1.37). We state two possible ways of extending (1.37). The first result follows from [8, Corollary 4.2] and its proof, and the second result is a special case of [57, Theorem 3.1]. We use the notation L(V1, V2) for the set of linear and continuous mappings from the topological vector space V1 into the topological vector space V2. We also set ω0,t(X, Y ) = v1(2Y )t−1ω0(X) for X, Y ∈ R2d . (1.38) Proposition 1.33. Let 0 ≤ t ≤ 1, p, q ∈ [1, ∞], and ω, ω0, v1, v0 ∈ P(R2d) be such that v0 and v1 are submultiplicative, ω0 is v0-moderate and ω is v1-moderate. Set v = vt 1v0 and ϑ = ω1/2 , 0 and let ω0,t be as in (1.38). Then the following are true: (1) The definition of (a, ϕ) 7→ Tpϕ(a) from S (R2d)×S (Rd) to L(S (Rd), S ′(Rd)) extends uniquely to a continuous map from M∞ L(S (Rd), S ′(Rd)). (1/ω0,t)(R2d) × M 1 (v)(Rd) to (2) If ϕ ∈ M 1 (v)(Rd) and a ∈ M∞ a continuous map from M p,q (ϑω)(Rd) to M p,q (ω/ϑ)(Rd). (1/ω0,t)(R2d), then Tpϕ(a) extends uniquely to Proposition 1.34. Let ω, ω1, ω2, v ∈ P(R2d) be such that ω1 is v-moderate, ω2 is v-moderate and ω = ω1/ω2. Then the following are true: (1) The mapping (a, ϕ) 7→ Tpϕ(a) extends uniquely to a continuous map from (ω)(R2d) × M 2 L∞ (v)(Rd) to L(S (Rd), S ′(Rd)). (2) If a ∈ L∞ (1/ω)(R2d) and ϕ ∈ M 2 (v)(Rd), then Tpϕ(a) extends uniquely to a continuous operator from M 2 (ω1)(Rd) to M 2 (ω2)(Rd). 1.9. Weyl formulation of Toeplitz operators. We finish this section by recall- ing some important relations between Weyl operators, Wigner distributions, and Toeplitz operators. For instance, the Weyl symbol of a Toeplitz operator is the con- volution between the Toeplitz symbol and a Wigner distribution. More precisely, if a ∈ S (R2d) and ϕ ∈ S (Rd), then Tpϕ(a) = (2π)−d/2 Opw(a ∗ Wϕ,ϕ) . (1.39) 18 AHMED ABDELJAWAD, SANDRO CORIASCO, AND JOACHIM TOFT Our analysis of Toeplitz operators is based on the pseudo-differential operator representation given by (1.39). Furthermore, any extension of the definition of Toeplitz operators to cases which are not covered by Propositions 1.33 and 1.34 is based on this representation. Here we remark that this leads to situations were certain mapping properties for the pseudo-differential operator representation make sense, whereas similar interpretations are difficult or impossible to make in the framework of (1.37) (see Remark 5.7 in Section 2). We refer to [50] or Section 2 for extensions of Toeplitz operators in context of pseudo-differential operators. 2. Confinement of the symbol classes Γ(ω) s (Rd) and Γ(ω) 0,s (Rd) In this section we introduce and discuss basic properties for confinements for symbols in Γ(ω0) 0,s . These considerations follow lines similar to the discussions in [3,34], but are here adapted to symbols that possess Gevrey regularity. In particular, this requires the deduction of various types of delicate estimates, concerning the compositions of symbols that are confined in different ways. and in Γ(ω0) s We recall that if σ is the (standard) symplectic form on R2d, namely, σ(X, Y ) = hy, ξi − hx, ηi, X = (x, ξ) ∈ R2d, Y = (y, η) ∈ R2d, then the symplectic Fourier transform Fσ is defined by and the twisted convolution ∗σ is given by (Fσa)(X) =ba(X) ≡ π−dZ a(Y )e2iσ(X,Y ) dY, π(cid:19) d (a ∗σ b)(X) ≡(cid:18) 2 2Z a(X − Y )b(Y )e2iσ(X,Y ) dY, 1/2(R2d). The twisted convolution is linked to the Weyl product (a#b)(X) = π−dZ a(X − Y )bb(Y )e2iσ(X,Y ) dY. σ is the identity map. We also note that F 2 In what follows we let aY = a( · − Y ) when a ∈ S′ analogous ways, bY , φY , ϕY , ψY etc. are defined when b, φ, ϕ, ψ ∈ S′ admissible a and b we have 1/2(R2d) and Y ∈ R2d, and in 1/2(R2d). For (a#b)Y = aY #bY . (2.1) We also recall that if ϕ ∈ Ss(R2d), then there are functions φ, ψ ∈ Ss(R2d) such that ϕ = φ#ψ (cf. [5, 58]). The same is true if Ss is replaced by Σs or by S . In particular, by choosing ϕ such thatR ϕ(X) dX = 1, (2.1) gives the following. 2 . Then there are φ, ψ ∈ Ss(R2d) such that Proposition 2.1. Let s ≥ 1 ψY #φY dY = 1. (2.2) ZR2d The same holds true with Σs or S in place of Ss, provided s > 1 2 . For independent translations in Weyl products we have the following. Proposition 2.2. Let s ≥ 1 2 and let φ, ψ ∈ Ss(R2d). Then (φY #ψZ)(X) = Ψ(X − Y, X − Z) for some Ψ ∈ Ss(R2d × R2d). The same holds true with Σs or S in place of Ss. for suitable a, b ∈ S′ by the formula hence, a#b = (2π)− d 2 a ∗σ (Fσb), LIFTINGS FOR MODULATION SPACES, AND ONE-PARAMETER GROUPS OF ΨDO 19 Proof. We only prove the result when φ, ψ ∈ Ss(R2d). The other cases follow by similar arguments and are left for the reader. We have (φY #ψZ )(X) = π−dZ φ(X − Y − Y1)bψ(Y1)e2iσ(Y1,Z)e2iσ(X,Y1) dY1 Ψ(X, Z) = π−dZ φ(X − Y1)bψ(Y1)e2iσ(Z,Y1) dY1. = π−dZ φ((X − Y ) − Y1)bψ(Y1)e2iσ(X−Z,Y1) dY1 = Ψ(X − Y, X − Z), We note that where Ψ = (Fσ,2 ◦ T )(φ ⊗ bψ), where (T Φ)(X, Z) = Φ(X − Z, Z) when Φ ∈ Ss(R2d × R2d), and Fσ,2Φ is the partial symplectic Fourier transform of Φ(X, Z) with respect to the Z variable. Since (φ, ψ) 7→ φ⊗bψ is continuous from Ss(R2d)×Ss(R2d) to Ss(R2d ×R2d), and T and Fσ,2Φ are continuous on Ss(R2d × R2d), it follows that Ψ ∈ Ss(R2d × R2d). (cid:3) Since Ψ in Proposition 2.2 belongs to similar types of spaces as φ and ψ, it follows that estimates of the form DαΨ(X, Y ) . hαα!se−(X 1 s +Y 1 s )/h hold true. In particular, the following is an immediate consequence of Proposition 2.2 and some standard manipulations in Gelfand-Shilov theory. Corollary 2.3. Let s ≥ 1 2 . If φ, ψ ∈ Ss(R2d) (φ, ψ ∈ Σs(R2d)), then Dα X Dβ Y Dγ Z(φY #ψZ )(X) . hα+β+γ(α!β!γ!)se−(X−Y 1 s +X−Z 1 s )/h (2.3) for some h > 0 (for every h > 0). Proof. Obviously, as stated in the previous Proposition 2.2, we can write φY #ψZ as Ψ(X − Y, X − Z) for some Ψ ∈ Ss(R2d × R2d). Thus Indeed, using the fact that Pδ≤α(cid:0)α (n + k)! ≤ 2n+kn!k!, then δ(cid:1) = 2α and that n! ≤ 2n(n − k)!k!, which implies (α + β + γ)! = (αj + βj + γj)! ≤ 4αj+βj +γj αj!βj!γj! = 4α+β+γα!β!γ!. dYj=1 dYj=1 Thus, inequality (2.3) holds with 2 · 4sh in place of h. (cid:3) The next fundamental result is a consequence of Theorem 4.12 in [4]. Dα X Dβ Y Dγ 2 1 Dγ 1 Dγ+α−δ ZΨ(X − Y, X − Z) =(cid:12)(cid:12)(cid:12)Dα X(cid:16)Dβ ≤Xδ≤α(cid:18)α δ(cid:19)(cid:12)(cid:12)(cid:12)(cid:16)Dβ+δ Ψ(cid:17) (X − Y, X − Z)(cid:12)(cid:12)(cid:12) ≤ hα+β+γXδ≤α(cid:18)α Xδ≤α(cid:18)α 2 Ψ(cid:17) (X − Y, X − Z)(cid:12)(cid:12)(cid:12) δ(cid:19) ((β + δ)!(γ + α − δ)!)s e−r(cid:16)X−Y δ(cid:19) ((β + δ)!(γ + α − δ)!)s ≤ 2α4sα+β+γ (α!β!γ!)s . 1 s +X−Z 1 s (cid:17). Moreover, we have 20 AHMED ABDELJAWAD, SANDRO CORIASCO, AND JOACHIM TOFT Proposition 2.4. Let s ≥ 1 continuous from Σs(R2d) × Γ(ϑ) 2 and ϑ ∈ PE(R2d).Then, the map (φ, a) 7→ φ#a is s (R2d) to Ss(R2d). The next lemma concerns uniform estimates of the Weyl product between ele- ments in sets Ωj,Yj , Yj ∈ R2d, j = 1, 2, such that related sets Ω∪ j , given by Ω∪ j ≡ [Y ∈R2d { a( · + Y ) ; a ∈ Ωj,Y }, (2.4) are bounded in Ss(R2d) or in Σs(R2d), j = 1, 2. j be as in (2.4), j and choose Yj ∈ R2d such that aj ∈ Ωj,Yj , j = 1, 2. Then the following is 2 and let Ωj,Y ⊆ S (R2d), Y ∈ R2d, Ω∪ Lemma 2.5. Let s ≥ 1 aj ∈ Ω∪ true: (1) if Ω∪ j are bounded in Ss(R2d), then there are constants C > 0 and h > 0 which are independent of Yj ∈ R2d and aj, j = 1, 2, and such that (a(α1) 1 #a(α2) 2 )(X) ≤ Chα1+α2(α1!α2!)se−(X−Y1 1 s +X−Y2 1 s +Y1−Y2 1 s )/h and Dα(a1#a2)(X) ≤ Chαα!se−(X−Y1 1 s +X−Y2 1 s +Y1−Y2 1 s )/h (2.5) (2.6) (2) if Ω∪ j are bounded in Σs(R2d), then for every h > 0, there is a constant C > 0 which is independent of Yj ∈ R2d and aj, j = 1, 2, and such that (2.5) and (2.6) hold. Proof. We only prove (2). The assertion (1) follows by similar arguments and is left for the reader. We have DX(a1#a2) = (DX a1)#a2+a1#(DX a2), which implies that the Leibnitz rule is valid for the Weyl product. Hence, (2.6) follows for every h > 0 if we prove that (2.5) holds for every h > 0. Let Y = Y1, Z = Y2, a = a1, b = a2, aY = a( · + Y ) and bZ = b( · + Z). Then a1#a2(X) = π−dZ aY ((X − Y ) − Y1)Fσ(bZ( · − Z))(Y1)e2iσ(X,Y1) dY1 = π−dZ aY ((X − Y ) − Y1)Fσ(bZ )(Y1)e2iσ(X−Z,Y1) dY1 = Fσ(aY (X1 − · )(FσbZ))(X2), where X1 = X − Y and X2 = X − Z. That is a1#a2(X) = G(X1, X2), where G = GY,Z = (Fσ,2 ◦ T )(aY ⊗ bZ). Here Fσ,2F is the partial symplectic Fourier transform of F ( · , X2) with respect to the variable X2 ∈ R2d, and (T F )(X, Y ) = F (X − Y, Y ). Since both Fσ,2 and T are continuous on Σs(R2d × R2d), it follows from the bound- edness of the sets Ω∪ j that for every h > 0, there is a constant Ch which is indepen- dent of aj ∈ Ω∪ j such that Dα X1 Dβ X2 G(X1, X2) ≤ Chhα+β(αβ)se−(X1 1 s +X2 1 s )/h. By straightforward computations it follows that Dα X1 Dβ X2 G(X1, X2) = (a(α)#b(β)), LIFTINGS FOR MODULATION SPACES, AND ONE-PARAMETER GROUPS OF ΨDO 21 since Y1 − Y2 1 s ≤ c(X − Y1 1 s + X − Y2 1 s ) for some constant c which only depends on s. The estimate (2.5) follows from these relations. (cid:3) Lemma 2.6. Let s ≥ 1 ψZ = ψ( · − Z). Then the following is true: 2 , φ, ψ ∈ Σs(R2d), ω, ϑ ∈ PE(R2d), φY = φ( · − Y ), and (1) if a ∈ Γ(ω) s (R2d) (a ∈ Γ(ω) 0,s (R2d)), then Dα X Dβ Y (φY a)(X) . hα 1 hβ 2 (α!β!)se−X−Y 1 s /h1 min(ω(X), ω(Y )) (2.7) and Dα X Dβ Y (φY #a)(X) . hα 1 hβ 2 (α!β!)se−X−Y 1 s /h1 min(ω(X), ω(Y )), (2.8) for some h1 > 0 (for every h1 > 0) and every h2 > 0; (2) if a1 ∈ Γ(ω) s (R2d) and a2 ∈ Γ(ϑ) s then (R2d) (a1 ∈ Γ(ω) 0,s (R2d) and a2 ∈ Γ(ϑ) 0,s (R2d)), Dα X Dβ Y Dγ Z((φY a1)#(ψZ a2))(X) . hα+β 1 hγ 2 (α!β!γ!)se−(X−Y 1 s +X−Z 1 s +Y −Z 1 s )/h1 for some h1 > 0 (for every h1 > 0) and every h2 > 0. min X1,X2∈{X,Y,Z}(cid:0)ω(X1)ϑ(X2)(cid:1), Proof. We only consider the case when a ∈ Γ(ω) other cases follow by similar arguments and are left for the reader. 0,s (R2d) and b ∈ Γ(ϑ) 0,s (R2d). The Let By Leibniz rule we get Ψ(X, Y ) = φ(X − Y )a(X). Dα X Dβ . 2α sup Y Ψ(X, Y ) ≤Xγ≤α(cid:18)α γ(cid:19)φ(α+β−γ)(X − Y )a(γ)(X) γ≤α(cid:18)hα+β((α + β − γ)!γ!)se−X−Y 1 s /hω(X)(cid:19) ≤ (21+sh)α+β(α!β!)se−X−Y 1 s /hω(X) . (21+sh)α+β(α!β!)se−X−Y 1 s /(2h)ω(Y ), for every h > 0 which is chosen small enough. Here we have used the fact that for some c > 0 ω(X) . ω(Y )ecX−Y . ω(Y )eX−Y 1 s /(2h), since ω is a moderate function. This gives (2.7). Next we prove (2). Let Ω1,Y =( X 7→ Ω2,Z =(cid:26) X 7→ Dβ Y (φY a) hββ!sω(Y ) Dγ Z(ψZ b) hγγ!sϑ(Z) ; β ∈ Nd) ; γ ∈ Nd(cid:27) , 22 AHMED ABDELJAWAD, SANDRO CORIASCO, AND JOACHIM TOFT and let Ω∪ in Σs(R2d). Hence, Lemma 2.5 shows that 2 be as in (2.4). By (2.7) it follows that Ω∪ 1 and Ω∪ 1 and Ω∪ 2 are bounded (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Dα X Dβ hββ!sω(Y )! #(cid:18) Dγ Y (φY a) Z(ψZ b) hγγ!sϑ(Z)(cid:19)! (X)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) . hαα!se−(X−Y 1 s +X−Z 1 s +Y −Z 1 s )/h for every h > 0, or equivalently, Dα X Dβ Y Dγ Z((φY a)#(ψZ b))(X) . hα+β+γ(α!β!γ!)se−(X−Y 1 s +X−Z 1 s +Y −Z 1 s )/hω(Y )ϑ(Z). The assertion now follows from the latter estimate and the fact that ω and ϑ are moderate weights, giving that ω(Y ) . ω(X)eX−Y 1 s /(2h) . ω(Z)e(X−Y 1 s +X−Z 1 s )/(2h), and similarly for ϑ. Finally, the estimate in (2) also holds for (φY b)#(ψZa) in place of (φY a)#(ψZ b), and (2.8) follows from this estimate by letting γ = 0, b = 1, ϑ = 1, and then integrate with respect to Z. The proof is complete. (cid:3) Lemmas 2.5 and 2.6 imply the following characterisation of Γ(ω) s (R2d). 1(R2d), φ ∈ Σs(R2d) have Proposition 2.7. Let s > 1/2, ω ∈ PE(R2d), a ∈ Σ′ non-vanishing integrals, and let φY = φ( · − Y ). Then the following conditions are equivalent: s ; (1) a ∈ Γ(ω) (2) φY a is smooth and satisfies (2.7) for some h1 > 0 and every h2 > 0; (3) φY #a is smooth and satisfies (2.8) for some h1 > 0 and every h2 > 0; (4) Dα X(φY a)(X) . hα 1 α!se−X−Y 1 s /h1 min(ω(X), ω(Y )), (2.9) for some h1 > 0; (5) Dα X (φY #a)(X) . hα 1 α!se−X−Y 1 s /h1 min(ω(X), ω(Y )), (2.10) for some h1 > 0. Lemma 2.8. Let s ≥ 1 are bounded families in Ss(R2d). Then, for any X, Y, Z ∈ R2d, α ∈ N2d, 2 . Assume that {a1( · +Y, Y )}Y ∈R2d and {a2( · +Z, Z)}Z∈R2d Dα X (a1( · + Y, Y )#a2( · + Z, Z))(X) . hαα!se−r(X−Y 1 s +X−Z 1 s ), (2.11) for some h > 0. Proof. The result is a straightforward consequence of Corollary 2.3 and its proof, since the involved constants on the right hand side of (2.3) depend continuously on φ and ψ in Ss(R2d). (cid:3) We notice that, by straightforward computation, for some other h, r > 0, (2.11) gives, for any X, Y, Z ∈ R2d, α ∈ N2d, Dα X (a1( · +Y, Y )#a2( · +Z, Z))(X) . hαα!se−r(X−Y 1 s +X−Z 1 s +Y −Z 1 s ). (2.12) LIFTINGS FOR MODULATION SPACES, AND ONE-PARAMETER GROUPS OF ΨDO 23 and Γ(1) 0,s. Let IR = [−R, R] and E0 = E0 ∞(R2d)), with the symbol subspace sw h,s}∞ 2.1. A family related to Γ(1) s L∞(IR × R2d; sw We shall consider suitable decreasing family {En aim, let Gn = {(Y, T1, . . . , Tn) ∈ R2d(n+1) : Y, Tj ∈ R2d with Tj ≤ 1, j = 1, . . . , n}, n ∈ Z+. We define En h,s = ∞(R2d) from Definition 1.22. n=0 of Banach spaces. To this h,s, n ≥ 1, as the set of all a ∈ E0 such that kak(n) = sup 1≤k≤n sup t∈IR sup (Y,T1,...,Tk)∈Gk khT1, DXi · · · hTk, DX ia(t, Y, · )ksw ∞ hk(k!)s < ∞, with the norm kakEn h,s ≡ kakE0 + kak(n). We also let E∞ h,s be the set of all a ∈ \n≥0 En h,s kakE∞ h,s ≡ sup n≥0 kakEn h,s such that is finite. Lemma 2.9. Let n ≥ 0. Then En h,s is a Banach space. Proof. Let {aj}j≥0 be a Cauchy sequence in En h,s, n ≥ 1. By definition, this sequence clearly has a limit a ∈ E0, and for some X 7→ bk(t, Y, T1, . . . , Tk, X) ∈ sw ∞(R2d) we have lim j→∞ sup khT1, DX i · · · hTk, DX iaj(t, Y, · ) − bk(t, Y, T1, . . . , Tk, · )ksw ∞ hk(k!)s = 0, where the supremum is taken over all k ∈ {1, . . . , n}, We need to prove that a ∈ En h,s, and aj → a in En h,s. t ∈ IR and (Y, T1, . . . , Tk) ∈ Gk. The conditions here above are equivalent to lim j→∞ sup t∈IR sup Y ∈R2d kaj(t, Y, · ) − a(t, Y, · )ksw ∞ = 0 (2.13) and lim j→∞ sup k(−1)khT1, Di · · · hTk, Diaj(t, Y, · ) − bk(t, Y, T1, . . . , Tk, · )ksw ∞ hk(k!)s = 0, (2.14) where the latter supremum should be taken over all k ∈ {1, . . . , n}, t ∈ IR and (Y, T1, . . . , Tk) ∈ Gk. Since sw (2.14) that ∞(R2d) is continuously embedded in S ′(R2d), it follows from (2.13) and has the limit X 7→ (−1)khT1, DXi · · · hTk, DXiaj (t, Y, X) X 7→ (−1)khT1, DX i · · · hTk, DX ia(t, Y, X) in S ′(R2d), and the limit X 7→ bk(t, Y, T1, . . . , Tk, X) in sw ∞(R2d), and thereby in S ′(R2d), as j tends to ∞. Hence bk(t, Y, T1, . . . , Tk, X) = (−1)khT1, DX i · · · hTk, DX ia(t, Y, X) 24 AHMED ABDELJAWAD, SANDRO CORIASCO, AND JOACHIM TOFT and it follows that En n ≥ 0. h,s is a Banach space for every h > 0, s > 0 and integer (cid:3) The spaces E∞ h,s can be related to Γ(1) s and Γ(1) 0,s, as the following lemma shows. The details are left for the reader. Lemma 2.10. Let a ∈ L∞(IR × R2d; sw uniformly bounded family in Γ(1) s (R2d) ( Γ(1) ∞(R2d)). Then {a(t, Y, · )}t∈IR,Y ∈R2d is a 0,s(R2d)), if and only if for some h > 0 (for every h > 0). kakE∞ h,s < ∞ Later on we also need the following result of differential equations with functions depending on a real variable with values in E∞ h,s. The proof is omitted since the result can be considered as a part of the standard theory of ordinary differential equations of first order in Banach spaces. Lemma 2.11. Suppose s ≥ 0 and T > 0, and let K be an operator from E∞ E∞ h,s for every h > 0 such that h,s to kKakE∞ h,s ≤ CkakE∞ h,s , a ∈ E∞ h,s, (2.15) for some constant C which only depend on h > 0. Then dc(t) dt = K(c(t)), c(0) ∈ E∞ h,s, has a unique solution t 7→ c(t) from [−T, T ] to E∞ h,s which satisfies where C is the same as in (2.15). kc(t)kE∞ h,s ≤ kc(0)kE∞ h,s eCT , 3. One-parameter group of elliptic symbols in the classes Γ(ω) s (Rd) s In this section we show that for suitable s and ω0, there are elements a ∈ Γ(ω0) and b ∈ Γ(1/ω0) such that a#b = b#a = 1. This is essentially a consequence of The- orem 3.8, where it is proved that the evolution equation (0.4) has a unique solution a(t, · ) which belongs to Γ(ωϑt) , thereby deducing needed semigroup properties for scales of pseudo-differential operators. Similar facts hold for corresponding Beurling type spaces (cf. Theorem 3.9). s s First we have the following result on certain logarithms of weight functions. Theorem 3.1. Let ω ∈ PE(R2d) ∩ Γ(ω) that ω is v-moderate, ϑ(X) = 1 + log v(X) and let s0 (R2d), s0 ∈ (0, 1], v ∈ PE(R2d), be such c(X, Y ) = log ω(X + Y ) ω(Y ) . Then, • {c( · , Y )}Y ∈R2d is a uniformly bounded family in Γ(ϑ) • for α 6= 0, {(∂α (R2d), s ≥ 1; X c)( · , Y )}Y ∈R2d is a uniformly bounded family in Γ(1) s s (R2d), s ≥ 1. For the proof of Theorem 3.1 we will need the following multidimensional version of the well-known Fa`a di Bruno formula for the derivatives of composed functions. It can be found, e.g., setting q = p = 1, n = 2d, in equations (2.3) and (2.4) in [29]. LIFTINGS FOR MODULATION SPACES, AND ONE-PARAMETER GROUPS OF ΨDO 25 Lemma 3.2. Let f : R → R, g : R2d → R. Then, for any α ∈ N2d, α 6= 0, ∂αf (g(x)) α! = X1≤k≤α f (k)(g(x)) k! Xβ1+···+βk=α βj6=0, j=1,...,k Y1≤j≤k (∂βj g)(x) βj! . (3.1) We will also need the next factorial estimate, for expressions involving decompo- sitions of α ∈ N2d, α 6= 0, into the sum of k nontrivial multiindeces βj, j = 1, . . . , k, as in (3.1), and corresponding products of (powers of) factorials. Lemma 3.3. Let s0 ∈ (0, 1], α ∈ N2d, α 6= 0. Then, for a suitable C0 > 0, depending only in d, X1≤k≤α 1 k Xβ1+···+βk=α βj6=0, j=1,...,k Y1≤j≤k (βj!)s0−1 . Cα 0 . (3.2) The proof of Lemma 3.3 can be found in the Appendix. Proof of Theorem 3.1. We have to show that c( · , Y ) satisfies Γ(ϑ) estimates, uni- formly with respect to Y ∈ R2d. If c(X, Y ) ≥ 0, then it follows by submultiplica- tivity of ω, that s c(X, Y ) = log ω(Y + X) − log ω(Y ) . log ω(Y ) + log v(X) − log ω(Y ) ≤ ϑ(X), for any Y ∈ R2d. Again by moderateness, when c(X, Y ) ≤ 0, recall that ω(X+Y ) ≥ ω(Y ) v(X) , so that c(X, Y ) & log ω(Y ) v(X) − log ω(Y ) ≥ − log v(X) ≥ −ϑ(X), and we can conclude c(X, Y ) . ϑ(X), X ∈ R2d. Now, for α ∈ N2d, α 6= 0, (3.1) gives ∂α X c(X, Y ) = α! X1≤k≤α (−1)k+1 k [ω(X + Y )]k Xβ1+···+βk=α βj 6=0, j=1,...,k Y1≤j≤k (∂βj ω)(X + Y ) βj! We can then estimate, in view of (3.2), ∂α 1 X c(X, Y ) . α! X1≤k≤α = hα α! X1≤k≤α k [ω(X + Y )]k Xβ1+···+βk=α k Xβ1+···+βk=α βj 6=0, j=1,...,k Y1≤j≤k βj 6=0, j=1,...,k Y1≤j≤k 1 ω(X + Y )hβj(βj!)s0 βj! (βj !)s0−1 . (C0h)α(α!)s, which concludes the proof. (cid:3) Proposition 3.4. Assume s > 1 for some r > 0. Let {a( · , Y )}Y ∈R2d be a uniformly bounded family in Σs(R2d) and {c( · , Z)}Z∈R2d be a bounded family in Γ(ω) 2 and ω(X) . erX (R2d). Then, s 1 s {a( · , Y )#c( · , Z)}Y,Z∈R2d and {c( · , Z)#a( · , Y )}Y,Z∈R2d are bounded families in Ss(R2d). 26 AHMED ABDELJAWAD, SANDRO CORIASCO, AND JOACHIM TOFT Proof. An immediate consequence of (2.8) in Lemma 2.6 is the following Dα X (φ#a)(X) ≤ Chαα!se−rX 1 s , (3.3) which implies that, for φ ∈ Σs and a ∈ Γ(ω) , (3.3) holds if and only if φ#a belongs to Ss. Then by the proof of (3.3) it follows that the constant C, h and r can be chosen depending continuously on φ ∈ Σs(R2d) and a ∈ Γ(ω) (R2d). Hence if Ω1 is a bounded family in Σs(R2d) and Ω2 is a bounded family in Γ(ω) (R2d), then {φ#a}φ∈Ω1,a∈Ω2 is a bounded family in Ss(R2d). (cid:3) s s s The following result can be found, e.g., in [46]. Lemma 3.5. Let a ∈ S ′(R2d). Then, kaksw ∞ ≤ C Xα≤d+1 kakL∞ ≤ C Xα≤2d+1 k∂αakL∞ k∂αaksw ∞ , (3.4) (3.5) and for some constant C > 0 depending only on the dimension d. Proposition 3.6. Let a ∈ S ′(R2d), s ≥ 1 α, β ∈ N2d. Then the following conditions are equivalent: 2 and set bαβ(X) = ∂α(X βa(X)) when (1) a ∈ Ss(R2d); (2) for some h > 0 it holds kbαβkL∞ . hα+β(α!β!)s, α, β ∈ N2d; (3) for some h > 0 it holds kbαβksw ∞ . hα+β(α!β!)s, α, β ∈ N2d. Proof. The equivalence between (1) and (2) is well-known. The proof of the equiv- alence of (2) and (3) follows by a straightforward application of Lemma 3.5. In fact, assume that (2) holds true. Then (3.4) gives kbαβksw ∞ ≤ C Xγ≤d+1 = hα+β(α!β!)s Xγ≤d+1 k∂γbαβkL∞(R2d) . Xγ≤d+1 hγ(γ!)s(cid:18) (α + γ)! α! γ! (cid:19)s . (2sh)α+β(α!β!)s, hα+β+γ((α + γ)!β!)s with a constant depending only on d and h, since Xγ≤d+1 hγ(γ!)s(cid:18) (α + γ)! α! γ! (cid:19)s ≤ C1 · 2s(α+d+1) ≤ C22sα+β, where the constants C1 and C2 only depend on d and h. Then, (3) holds true, as claimed. The proof of the converse follows by similar argument, employing (3.5). (cid:3) We have now the following. Theorem 3.7. Let a ∈ S ′(R2d) and s > 0. Then the following conditions are equivalent: (1) a ∈ Γ(1) s (R2d); LIFTINGS FOR MODULATION SPACES, AND ONE-PARAMETER GROUPS OF ΨDO 27 (2) there exists h > 0 such that k∂αakL∞(R2d) . hα(α!)s, for all α ∈ Z2d + ; (3) there exists h > 0 such that k∂αaksw ∞ . hα(α!)s, for all α ∈ Z2d + ; (4) there exists h > 0 such that khT1, DX i · · · hTm, DX iaksw ∞ . hm(m!)s, for any T1, . . . , Tm ∈ R2d such that Tj ≤ 1, j = 1, . . . , m, m ≥ 1. Proof. The equivalence between (1) and (2) is well known. The equivalence of (2) and (3) is proved by an argument completely similar to the one employed in the proof of Proposition 3.6, using Lemma 3.5. It remains to prove only the equivalence with (4). Assume that (3) holds true, and let for the standard basis (el)l=1,...,d of Rd and the dual basis (εl)l=1,...,d. Recall that so that the symbol hT1, DX i · · · hTm, DX ia is in the span of symbols of the form hTk, DX ia = tkl ∂a ∂xl τkl ∂a ∂ξl , Tj = tjlel + τjlεl, dXl=1 dXl=1 mXj=1 dXl=1 + dXl=1 mXj=1  mYj=1 j τ γj tβj x ∂γ ξ a, j  ∂β βj = β, γj = γ, β + γ = m, where the summation contains at most (2d)m terms. Since Tj ≤ 1, j = 1, . . . , m, by (3.4) we find khT1, DX i · · · hTm, DX iaksw ∞ ≤ (2d)m sup α=m k∂αaksw ∞ hα+γ ((α + γ)!)s . sup α=m Xγ≤d+1 = sup α=m hα(α!)s Xγ≤d+1 . (2s+1h)m(m!)s, hγ(γ!)s(cid:18) (α + γ)! α! γ! (cid:19)s which proves that (4) holds true. As in the proof of Proposition 3.6, the converse implication is obtained by a completely similar argument, employing (3.5). (cid:3) We are now ready to state and prove the first main result of this section, the following Theorem 3.8. It deals with the existence of one-parameter groups of pseudo-differential operators, obtained as solutions to suitable evolution equations. Theorem 3.8. Let s ≥ 1, ω, ϑ ∈ P 0 ϑ ∈ Γ(ϑ) smooth map (t, X) 7→ a(t, X) ∈ C such that a(t, · ) ∈ Γ(ω ϑt) and (R2d) and s (R2d). Then, there exists a unique (R2d) for all t ∈ R, E,s(R2d) be such that ω ∈ Γ(ω) (R2d), and let a0 ∈ Γ(ω) (R2d), b ∈ Γ(1) s s s s ((∂ta)(t, · ) = (b + log ϑ)#a(t, · ) a(0, · ) = a0. (3.6) 28 AHMED ABDELJAWAD, SANDRO CORIASCO, AND JOACHIM TOFT If ω ≡ a0 ≡ 1, then a(t, X) also satisfies ((∂ta)(t, · ) = a(t, · )#(b + log ϑ) a(0, · ) = a0, (3.7) and a(t1, · )#a(t2, · ) = a(t1 + t2, · ), a(t, · ) ∈ Γ(ϑt) s (R2d), t, t1, t2 ∈ R. (3.8) Proof. First suppose that a solution a(t, X) of (3.6) exists. Then, with a(t, X) = a0(X) +Z t 0 c(u, X) du c(t, · ) = (b + log ϑ)#a(t, · ) ∈ Γ(ωhlog ϑiϑt) s (R2d), in view of Theorem 3.1 and the properties of the Weyl product in the Γ(ω) (R2d) classes, see [4]. This implies that the map t 7→ a(t, · ) is C1 from [−R, R] into the symbol space s Γ(ωhlog ϑi(ϑ+ϑ−1)R) s (R2d). Choose s0 < s, and φ, ψ ∈ Ss0 (R2d) such that (2.2) holds true. Let c1(t, Y, · ) = ω(Y )−1ϑ(Y )−t φY #a(t, · ). (3.9) By Lemma 2.6 (1) we have that, for any Y ∈ R2d, t 7→ c1(t, Y, · ) is a C1 map from [−R, R] into Ss(R2d). Moreover, ∂tc1(t, Y, · ) = ω(Y )−1ϑ(Y )−t φY #(b + log ϑ)#a(t, · ) − c1(t, Y, · ) log ϑ(Y ). Let Then, with and f (Y, X) = b(X) + log ϑ(X) ϑ(Y ) . (∂tc1)(t, Y, · ) = ω(Y )−1 ϑ(Y )tZ φY #f (Y, · )#ψZ #φZ #a(t, · ) dZ =Z KY,Z(t, · )#c1(t, Z, · ) dZ KY,Z(t, · ) = ω(Z) ϑ(Z)t ω(Y ) ϑ(Y )t φY #f (Y, · )#ψZ, (3.10) c1(0, Y, · ) = ω(Y )−1φY #a0. We also need to consider the similar equation where f (Y, · ) is replaced by f (Z, · ), that is ∂tc2(t, Y, · ) =Z eKY,Z(t, · )#c2(t, Z, · ) dZ, eKY,Z(t, · ) = ω(Z) ϑ(Z)t ω(Y ) ϑ(Y )t φY #f (Z, · )#ψZ , c2(0, Y, · ) = c1(0, Y, · ) = ω(Y )φY #a0. (3.11) (3.12) where and We consider the operators K and eK on E0, defined by (Ka)(t, Y, X) =Z (KY,Z(t, · )#a(t, Z, · ))(X) dZ, LIFTINGS FOR MODULATION SPACES, AND ONE-PARAMETER GROUPS OF ΨDO 29 and and show that (eKa)(t, Y, X) =Z (eKY,Z(t, · )#a(t, Z, · ))(X) dZ, kKakEn h,s ≤ C(n + 1)kakEn h,s , kKakE∞ h,s ≤ CkakE∞ h,s , ≤ C(n + 1)kakEn h,s and h,s ≤ CkakE∞ h,s (3.13) for some constant C, which is independent of h, n and s. In order to prove this, it is convenient to let Pk be the family of all subsets of {1, . . . , k}, k ≥ 1. For each P ∈ Pk, a ∈ sw ∞(R2d), we set h,s keKakE∞ when P = ∅, keKakEn H(a, P ) =(a hTj1, DX i · · · hTjl, DX ia when P = {j1 < · · · < jl}, l ≤ k. We now estimate k(hT1, DX i · · · hTk, DX iKa)(t, Y, · )ksw ∞(R2d) hk(k!)s . when a ∈ En h,s. Since (hT1, DX i · · · hTk, DX iKa)(t, Y, X) = hT1, DX i · · · hTk, DX iZ (KY,Z(t, · )#a(t, Z, · ))(X) dZ = XP ∈PkZ (H(KY,Z(t, · ), P )#H(a(t, Z, · ), P c))(X) dZ, we find k(hT1, DXi · · · hTk, DX iKa)(t, Y, · )ksw ∞ hk(k!)s · kH(a(t, Z, · ), P c)ksw ∞ hk−l((k − l)!)s dZ ≤ kXl=0 XP =l(cid:18)k l(cid:19)−sZ kH(KY,Z(t, · ), P )ksw kXl=0 XP =l kakEk−l hl l!s ≤ ∞ h,s(cid:18)k kXl=0 XP =l(cid:18)k l(cid:19)−sZ kH(KY,Z(t, · ), P )ksw hl l!s ∞ dZ l(cid:19)−1Z kH(KY,Z(t, · ), P )ksw hl l!s ∞ . kakEn h,s dZ, (3.14) where the last inequality follows from the fact that s ≥ 1 and kakEn with n. h,s increases We have now to estimate kH(KY,Z(t, · ), P )ksw ties on the right-hand side of (3.10). Since ω and ϑ belong to P 0 for every r > 0, ∞ , and study the different quanti- E,s, it follows that ω(Z) ϑ(Z)t ω(Y ) ϑ(Y )t = ω(Z) ω(Y )(cid:18) ϑ(Z) ϑ(Y )(cid:19)t . erY −Z 1 s (cid:18)erY −Z 1 s(cid:19)t . er(1+t)Y −Z Y, Z ∈ R2d. (3.15) 1 s , 30 AHMED ABDELJAWAD, SANDRO CORIASCO, AND JOACHIM TOFT For the Weyl product in (3.10) we have φY #f (Y, · ) = φ( · − Y )#(cid:16)b( · ) + log =(cid:16)φ#b( · + Y )(cid:17)Y =(cid:16)φ#b( · + Y )(cid:17)Y ϑ( · ) ϑ( · ) ϑ(Y )(cid:17) + φ( · − Y )#(cid:16) log +(cid:16)φ# log and nφ# log ϑ(Y )(cid:17) ϑ(Y ) (cid:17)Y ϑ(Y ) oY ∈R2d ϑ( · + Y ) ϑ( · + Y ) . nφ#b( · + Y )oY ∈R2d By Theorem 3.1 and Proposition 3.4, are uniformly bounded families in Ss(R2d). Note that (3.16) a2(Z, X) = ψZ (X) ⇒ {a2(Z, · + Z)}Z∈R2d = {ψ}Z∈R2d, which is evidently a uniformly bounded family in Ss(R2d). Combining this last observation with the computations on φY #f (Y, · ) above, using Lemma 2.8 and (2.12), we finally obtain, for some h, r0 > 0, Dα X (φY #f (Y, · )#ψZ)(X) . hαα!se−r0(X−Y 1 s +X−Z 1 s +Y −Z 1 s ), X, Y, Z ∈ R2d, α ∈ N2d. (3.17) By Theorem 3.7, (3.15) and (3.17) we get for all P ∈ Pk Y, Z ∈ R2d and some r0, h > 0 that kH(KY,Z(t, · ), P )ksw ∞ ≤ Chll!se−r0Y −Z 1 s , l = P , where C is independent of k. Hence, (3.14) gives kKa(t, Y, · )ksw ∞ . kakEn h,s , and khT1, DX i · · · hTk, DXiKa(t, Y, · )ksw ∞ hk(k!)s ≤ C1kakEn h,s 1 s dZ kXl=0 XP =l(cid:18)k l(cid:19)−1Z e−r0Y −Z kXl=0 XP =l(cid:18)k l(cid:19)−1 = C2kakEn h,s = C2(k + 1)kakEn , 1 ≤ k ≤ n, h,s as claimed, where C1 and C2 are independent of k, n and h. In fact, by similar arguments that lead to (3.16) it follows that By a completely similar argument, an analogous result can be obtained for eK. {b( · + Z)#ψ}Z∈R2d and (cid:26)log ϑ( · + Z) ϑ(Z) #ψ(cid:27)Z∈R2d are bounded in Ss(R2d), given that (3.17) holds with f (Z, · ) in place of f (Y, · ). This gives the first and third inequalities in (3.13). From these estimates we get kKakEn+1 h,s ≤ CkakEn h,s ≤ CkakEn h,s , and keKakEn+1 h,s which give the other inequalities in (3.17). We have proven that for any T > 0, then kKtkEn h,s→En h,s ≤ C(n + 1), t ≤ T, LIFTINGS FOR MODULATION SPACES, AND ONE-PARAMETER GROUPS OF ΨDO 31 where C is independent of n. As a consequence, since ω(Y )−1φY #a0 belongs to E∞ h,s, the equation dc1 dt = Kc1, c1(0) = ω(Y )−1φY #a0 (3.18) has a unique solution on R belonging to E∞ sition 2.7 it follows that c1(t, Y, · ) ∈ Γ(1) t. h,s, in view of Lemma 2.11. By Propo- s (R2d), uniformly in Y and for bounded In order to prove the uniqueness of the solution a of (3.6), first we assume the existence and by what we have proven above i.e. that c1(t, Y, · ) in (3.9) satisfies (3.18) which implies the uniqueness of the solution of (3.6), since ω(Y )ϑ(Y )tψY #c1(t, Y, · ) dY. (3.19) a(t, · ) =ZR2d ψY #φY #a(t, · ) dY =ZR2d a(t, · ) =ZR2d To prove the existence of a solution of (3.6), we consider the solution c2(t, Y, · ) of (3.11) with the initial data (3.12), and we let ω(Y )ϑ(Y )tψY #c2(t, Y, · ) dY. (3.20) By Propositions 2.7 and 1.8, the family {ψY #c2(t, Y, · )}Y ∈R2d belongs to Ss and a(t, · ) belongs to Γ(wϑt) . Moreover, s da(t, · ) dt ω(Y )ϑ(Y )t log ϑ(Y )ψY #c2(t, Y, · ) dY = (b + log ϑ)#a(t, · ), with the initial data a(0, · ) =ZR2d ω(Y )ψY #(ω(Y )−1φY #a0) dY = a0, which provide a solution of (3.6). In order to prove the last part we consider the unique solution a(t, · ) of (3.6) with the initial data a(0, · ) ≡ 1. If ω ≡ 1, then for u ∈ R the mappings t 7→ a(t + u, · ) and t 7→ a(t, · )#a(u, · ) are both solutions of (3.6) with value a(u, · ) at t = 0, and a(t + u, · ) = a(t, · )#a(u, · ), (3.21) by the uniqueness property for the solution of (3.6). Using (3.21) we have for all t ∈ R, a(t, · )#a(−t, · ) = 1. Taking the derivative we get d dt 0 = (a(t, · )#a(−t, · )) = (b+log ϑ)#a(t, · )#a(−t, · )−a(t, · )#(b+log ϑ)#a(−t, · ). That is (b + log ϑ) = a(t, · )#(b + log ϑ)#a(−t, · ), implying the commutation for the sharp product of a(t, · ) with (b + log ϑ), and the result follows. (cid:3) By similar argument as for the previous result we get the following. =ZR2d =ZR2d =ZR2d +ZR2dZR2d +ZR2dZR2d ω(Y )ϑ(Y )tψY #eKY,Z(t, · )#c2(t, Z, · ) dY dZ ω(Z)ϑ(Z)t log ϑ(Z)ψZ #c2(t, Z, · ) dZ ω(Z) ϑ(Z)tψY #φY #f (Z, · )#ψZ #c2(t, Z, · ) dY dZ ω(Z)ϑ(Z)t(b + log ϑ)ψZ #c2(t, Z, · ) dZ 32 AHMED ABDELJAWAD, SANDRO CORIASCO, AND JOACHIM TOFT (R2d), and let a0 ∈ Γ(ω) Theorem 3.9. Let s ≥ 1, ω, ϑ ∈ PE,s(R2d) be such that ω ∈ Γ(ω) ϑ ∈ Γ(ϑ) smooth map (t, X) 7→ a(t, X) ∈ C such that a(t, · ) ∈ Γ(ω ϑt) and a(t, · ) satisfies (3.6). (R2d) and 0,s(R2d). Then, there exists a unique (R2d) for all t ∈ R, 0,s (R2d), b ∈ Γ(1) s s s Moreover, if ω ≡ a0 ≡ 1, then a(t, X) also satisfies (3.7) and a(t1, · )#a(t2, · ) = a(t1 + t2, · ), a(t, · ) ∈ Γ(ϑt) 0,s (R2d), t, t1, t2 ∈ R. 4. Lifting of pseudo-differential operators and Toeplitz operators on modulation spaces In this section we apply the group properties in Theorems 3.8 and 3.9 to deduce lifting properties of pseudo-differential operators on modulation spaces. Thereafter we combine these results by the Wiener property of certain pseudo-differential op- erators with symbols in suitable modulation spaces to get lifting properties for Toeplitz operators with weights as their symbols. We begin to apply Theorems 3.8 and 3.9 to get the following. E,s(R2d), and let Theorem 4.1. Let s ≥ 1, p ∈ (0, ∞]2d, A ∈ M(d, R), ω ∈ P 0 B be an invariant BF-space on R2d, or B = Lp,E(R2d) for some phase-shift split parallelepiped E in R2d. Then the following are true: (1) There exist a ∈ Γ(ω) s (R2d) and b ∈ Γ(1/ω) s (R2d) such that OpA(a) ◦ OpA(b) = OpA(b) ◦ OpA(a) = IdS ′ s(Rd) . (4.1) Furthermore, OpA(a) is an isomorphism from M (ω0, B) onto M (ω0/ω, B), for every ω0 ∈ P 0 E,s(R2d). (2) Let a0 ∈ Γ(ω) s (ω1/ω)(Rd) for some ω1 ∈ P 0 (R2d) be such that OpA(a0) is an isomorphism from M 2 (ω1)(Rd) E,s(R2d). Then OpA(a0) is an iso- to M 2 E,s(R2d). morphism from M (ω2, B) to M (ω2/ω, B), for every ω2 ∈ P 0 Furthermore, the inverse of OpA(a0) is equal to OpA(b0) for some b0 ∈ Γ(1/ω) s (R2d). Theorem 4.2. Let s > 1, p ∈ (0, ∞]2d, A ∈ M(d, R), ω ∈ PE,s(R2d), and let B be an invariant BF-space on R2d, or B = Lp,E(R2d) for some phase-shift split parallelepiped E in R2d. Then the following are true: (1) There exist a ∈ Γ(ω) 0,s (R2d) and b ∈ Γ(1/ω) 0,s (R2d) such that OpA(a) ◦ OpA(b) = OpA(b) ◦ OpA(a) = IdΣ′ s(Rd) . (4.2) Furthermore, OpA(a) is an isomorphism from M (ω0, B) onto M (ω0/ω, B), for every ω0 ∈ PE,s(R2d). (2) Let a0 ∈ Γ(ω) 0,s (R2d) be such that OpA(a0) is an isomorphism from M 2 (ω1)(Rd) to M 2 (ω1/ω)(Rd) for some ω1 ∈ PE,s(R2d). Then OpA(a0) is an iso- morphism from M (ω2, B) to M (ω2/ω, B), for every ω2 ∈ PE,s(R2d). Furthermore, the inverse of OpA(a0) is equal to OpA(b0) for some b0 ∈ Γ(1/ω) 0,s (R2d). We only prove Theorem 4.2. Theorem 4.1 follows by similar arguments and is left for the reader. Proof of Theorem 4.2. The existence of a ∈ Γ(ω) (R2d) such that (4.2) holds is guaranteed by Theorem 3.9. By [56, Theorems 2.3 and 2.6] it 0,s (R2d) and b ∈ Γ(1/ω) 0,s LIFTINGS FOR MODULATION SPACES, AND ONE-PARAMETER GROUPS OF ΨDO 33 follows that and OpA(a) : M (ω0, B) →M (ω0/ω, B) OpA(b) : M (ω0/ω, B) →M (ω0, B) (4.3) (4.4) are continuous. By (4.2) and the fact that M (ω0, B) and M (ω0/ω, B) are contained in Σ′ s(R2d), it follows that (4.3) and (4.4) are homeomorphic, and (1) follows. (2) It suffices to prove the result in the Weyl case, A = 1 2 I, in view of Proposition 1.25. By (1), we may find a1 ∈ Γ(ω1) 0,s , b1 ∈ Γ(1/ω1) 0,s , a2 ∈ Γ(ω1/ω) 0,s , b2 ∈ Γ(ω/ω1) 0,s satisfying the following properties: • Opw(aj) and Opw(bj) are inverses to each others on Σ′ • For arbitrary ω2 ∈ PE,s(R2d), the mappings s(Rd) for j = 1, 2; Opw(a1) : M 2 Opw(b1) : M 2 Opw(a2) : M 2 Opw(b2) : M 2 (ω2/ω1), (ω2ω1), (ω2ω/ω1), (ω2) → M 2 (ω2) → M 2 (ω2) → M 2 (ω2) → M 2 (ω2ω1/ω) (4.5) are isomorphisms. In particular, Opw(a1) is an isomorphism from M 2 (ω1) to L2, and Opw(b1) is an isomorphism from L2 to M 2 (ω1). Now set c = a2#a#b1. Then by [4, Theorem 4.14], the symbol c satisfies c = a2#a#b1 ∈ Γ(ω1/ω) s #Γ(ω) s #Γ(1/ω1) s ⊆ Γ(1) s . Furthermore, Opw(c) is a composition of three isomorphisms and consequently Opw(c) is boundedly invertible on L2. By Proposition 1.32 (2), Opw(c)−1 = Opw(c1) for some c1 ∈ Γ(1) 0,s. Hence, by (1) it follows that Opw(c) and Opw(c1) are isomorphisms on M (ω2, B), for each ω2 ∈ PE,s(R2d). Since Opw(c) and Opw(c1) are bounded on every M (ω, B), the factorization of the identity Opw(c) Opw(c1) = Id is well-defined on every M (ω, B). Consequently, Opw(c) is an isomorphism on M (ω, B). Using the inverses of a2 and b1, we now find that Opw(a) = Opw(b2) ◦ Opw(c) ◦ Opw(a1) is a composition of isomorphisms from the domain space M (ω2, B) onto the im- age space M (ω2/ω, B) (factoring through some intermediate spaces) for every ω2 ∈ PE,s(R2d) and every invariant BF-space B. This proves the isomorphism assertions for Opw(a). Finally, the inverse of Opw(a) is given by Opw(b1) ◦ Opw(c1) ◦ Opw(a2). which is a Weyl operator with symbol in Γ(1/ω) 0,s , and the result follows. (cid:3) 34 AHMED ABDELJAWAD, SANDRO CORIASCO, AND JOACHIM TOFT 5. Mapping properties for Toeplitz operators In this section we study the isomorphism properties of Toeplitz operators be- tween modulation spaces. We first state results for Toeplitz operators that are well-defined in the sense of (1.37) and Propositions 1.33 and 1.34. Then we state and prove more general results for Toeplitz operators that are defined only in the framework of pseudo-differential calculus. We start with the following result about Toeplitz operators with smooth symbols. (R2d) Theorem 5.1. Let s ≥ 1 ω, ω0, v ∈ P 0 and that ω0 is v-moderate, and let B be an invariant BF-space on R2d or B = Lp,E(R2d) for some phase-shift split parallelepiped E in R2d. If φ ∈ M 1 (v)(Rd), then Tpφ(ω0) is an isomorphism from M (ω, B) to M (ω/ω0, B). E,s(R2d) be such that ω0 ∈ Γ(ω0) s In the next result we relax our restrictions on the weights but impose more restrictions on B. Theorem 5.2. Let s > 1, 0 ≤ t ≤ 1, p, q ∈ [1, ∞], and ω, ω0, v0, v1 ∈ PE,s(R2d) 1v0, ϑ = ω1/2 be such that ω0 is v0-moderate and ω is v1-moderate. Set v = vt and let ω0,t be the same as in (1.38). If φ ∈ M 1 (v)(Rd) and ω0 ∈ M∞ (1/ω0,t)(R2d), then Tpφ(ω0) is an isomorphism from M p,q (ϑω)(Rd) to M p,q (ω/ϑ)(Rd). 0 Before the proofs of Theorems 5.1 and 5.2 we state the following consequence of Theorem 5.2 which was the original goal of our investigations. Corollary 5.3. Let s ≥ 1, ω, ω0, v1, v0 ∈ PE,s(R2d) and that ω0 is v0-moderate and ω is v1-moderate. Set v = v1v0 and ϑ = ω1/2 (v)(Rd), then Tpφ(ω0) is an isomorphism from M p,q (ω/ϑ)(Rd) simultaneously for all p, q ∈ [1, ∞]. (ϑω)(Rd) to M p,q . If φ ∈ M 1 0 Proof. Let ω1 ∈ PE,s(R2d) ∩ Γ(ω1) 0,s (R2d) be such that C−1 ≤ ω1/ω0 ≤ C, for some constant C. Hence, ω/ω0 ∈ L∞ ⊆ M ∞. By Theorem 2.2 in [50], it follows that (ω2)(R2d), when ω2(x, ξ, η, y) = 1/ω0(x, ξ). The result ω = ω1 · (ω/ω1) belongs to M ∞ now follows by setting t = 1 and q0 = 1 in Theorem 5.2. (cid:3) In the proofs of Theorems 5.1 and 5.2 we consider Toeplitz operators as defined by an extension of the form (1.37). Later on we present extensions of these theorems (cf. Theorems 5.1′ and 5.2′ below) for those readers who accept to use pseudo- differential calculus to extend the definition of Toeplitz operators. Except for the interpretation of Tpφ(ω) the proofs of Theorems 5.1 and 5.2 are identical to those of Theorems 5.1′ and 5.2′. We need some preparations and start with the following lemma. Lemma 5.4. Let s ≥ 1, ω, v ∈ PE,s(R2d) be such that ϑ = ω1/2 is v-moderate. Assume that φ ∈ M 2 (ϑ)(Rd) onto M 2 (v). Then Tpφ(ω) is an isomorphism from M 2 (1/ϑ)(Rd). Proof. Recall from Remark 1.13 that for φ ∈ M 2 ϑkL2 defines an equivalent norm on M 2 to φ are well defined. (v)(Rd) \ {0} the expression kVφf · (ϑ). Thus the occurring STFTs with respect Since Tpφ(ω) is bounded from M 2 (ϑ) to M 2 k Tpφ(ω)f kM 2 (1/ϑ) holds for all f ∈ M 2 (ϑ). Next we observe that (1/ϑ) by Proposition 1.34, the estimate . kf kM 2 (5.1) (ϑ) (Tpφ(ω)f, g)L2(Rd) = (ωVφf, Vφg)L2(R2d) = (f, g)M 2,φ (ϑ) , (5.2) LIFTINGS FOR MODULATION SPACES, AND ONE-PARAMETER GROUPS OF ΨDO 35 for f, g ∈ M 2 sition 1.12(3)) now yields the following identity: (ϑ)(Rd) and φ ∈ M 2 (v)(Rd). The duality of modulation spaces (Propo- kf kM 2 (ϑ) ≍ ≍ sup kgkM 2 (ϑ) sup kgkM 2 (ϑ) =1 =1 (f, g)M 2 (ϑ) (Tpφ(ω)f, g)L2 ≍ k Tpφ(ω)f kM 2 (1/ϑ) . (5.3) A combination of (5.1) and (5.3) shows that kf kM 2 equivalent norms on M 2 (ϑ) (ϑ). and k Tpφ(ω)f kM 2 (1/ϑ) are In particular, Tpφ(ω) is one-to-one from M 2 (1/ϑ) with closed range. Since Tpφ(ω) is self-adjoint with respect to L2, it follows by duality that Tpφ(ω) has dense range in M 2 (1/ϑ). By Banach's theorem Tpφ(ω) is an isomorphism from M 2 (cid:3) (1/ϑ). Consequently, Tpφ(ω) is onto M 2 (ϑ) to M 2 (ϑ) to M 2 (1/ϑ). We need a further generalization of Proposition 1.33 to more general classes of symbols and windows. Set ω1(X, Y ) = v0(2Y )1/2v1(2Y ) ω0(X + Y )1/2ω0(X − Y )1/2 . (5.4) Proposition 1.33′. Let s ≥ 1, 0 ≤ t ≤ 1, p, q, q0 ∈ [1, ∞], and ω, ω0, v0, v1 ∈ PE,s(R2d) be such that v0 and v1 are submultiplicative, ω0 is v0-moderate and ω is v1-moderate. Set r0 = 2q0/(2q0 − 1), v = vt 1v0 and ϑ = ω1/2 0 , and let ω0,t and ω1 be as in (1.38) and (5.4). Then the following are true: (1) The definition of (a, φ) 7→ Tpφ(a) from Σs(R2d)×Σs(Rd) to L(Σs(Rd), Σ′ s(Rd)) extends uniquely to a continuous map from M∞,q0 L(Σs(Rd), Σ′ s(Rd)). (1/ω0,t)(R2d) × M r0 (v)(Rd) to (2) If φ ∈ M r0 a0 ∈ M∞,1 M p,q (ϑω)(Rd) to M p,q (ω/ϑ)(Rd). (v)(Rd) and a ∈ M∞,q0 (1/ω0,t)(R2d), then Tpφ(a) = Opw(a0) for some (ω1)(R2d), and Tpφ(a) extends uniquely to a continuous map from For the proof we need the following result, which follows from [49, Proposition 2.1] and its proof. Lemma 5.5. Assume that s ≥ 1, q0, r0 ∈ [1, ∞] satisfy r0 = 2q0/(2q0 − 1). Also assume that v ∈ PE,s(R2d) is submultiplicative, and that κ, κ0 ∈ PE,s(R2d ⊕ R2d) satisfy κ0(X1 + X2, Y ) ≤ Cκ(X1, Y ) v(Y + X2)v(Y − X2) X1, X2, Y ∈ R2d, (5.5) for some constant C > 0. Then the map (a, φ) 7→ Tpφ(a) from Σs(R2d) × s(Rd)) extends uniquely to a continuous mapping from Σs(Rd) to L(Σs(Rd), Σ′ M∞,q0 (v)(Rd) and a ∈ M∞,q0 (κ) (R2d), then Tpφ(a) = Opw(b) for some b ∈ M∞,1 (κ0). s(Rd)). Furthermore, if φ ∈ M r0 (v)(Rd) to L(Σs(Rd), Σ′ (ω) (R2d) × M r0 Proof of Proposition 1.33 ′. We show that the conditions on the involved parame- ters and weight functions satisfy the conditions of Lemma 5.5. First we observe that vj(2Y ) ≤ Cvj (Y + X2)vj (Y − X2), j = 0, 1 36 AHMED ABDELJAWAD, SANDRO CORIASCO, AND JOACHIM TOFT for some constant C which is independent of X2, Y ∈ R2d, because v0 and v1 are submultiplicative. Refering back to (5.4) this gives ω1(X1 + X2, Y ) = v0(2Y )1/2v1(2Y ) ω0(X1 + X2 + Y )1/2ω0(X1 + X2 − Y )1/2 ≤ C1 v0(2Y )1/2v1(2Y )v0(X2 + Y )1/2v0(X2 − Y )1/2 ω0(X1) = C1v1(2Y )1−t v0(2Y )1/2v1(2Y )tv0(X2 + Y )1/2v0(X2 − Y )1/2 ω0(X1) Hence ≤ C2v1(2Y )1−t v1(X2 + Y )tv1(X2 − Y )tv0(X2 + Y )v0(X2 − Y ) ω0(X1) . ω1(X1 + X2, Y ) ≤ C v1(2Y )1−tv(X2 + Y )v(X2 − Y ) ω0(X1) . (5.6) By letting κ0 = ω1 and κ = 1/ω0,t, it follows that (5.6) agrees with (5.5). The result now follows from Lemma 5.5. (cid:3) In the remaining part of the paper we interpret Tpφ(a) as the extension of a Toeplitz operator provided by Proposition 1.33′. (See also Remark 5.7 below for more comments.) Proposition 1.33′ can be applied to Toeplitz operators with smooth weights as symbols. Proposition 5.6. Assume that s ≥ 1, ω0 ∈ P 0 that v ∈ P 0 M 2 (v)(Rd), then Tpφ(ω0) = Opw(b) for some b ∈ Γ(ω0) E,s(R2d) is submultiplicative, and that ω1/2 0 (R2d). s is v-moderate. If φ ∈ E,s(R2d) be such that ω0 ∈ Γ(ω0) s (R2d), Proof. By Propositions 1.25 and 1.26 with t = 0 we have ω0 ∈ M∞,1 (1/ω0,r0 )(R2d) s . Furthermore, by letting 1 for some r0 ≥ 0, where ω0,r0(X, Y ) = ω0(X)e−r0Y v1(Y ) = er0Y s , and ω1 in (5.4) we have 1 ω1(X, Y ) & er02Y 1 s v(2Y )1/2 ω0(X + Y )1/2ω0(X − Y )1/2 1 s er0Y ω0(X) . & Proposition 1.33′ implies that existence of some b ∈ M∞,1 (1/ω0,r0 )(R2d) ⊆ Γ(ω0) s (R2d). (cid:3) The following generalization of Theorem 0.1 is an immediate consequence of Theorem 4.1, Lemma 5.4 and Proposition 5.6, since it follows by straightforward computations that Ss ⊆ M 2 (v) when v satisfies the hypothesis in Proposition 5.6. Theorem 0.1′. Let s ≥ 1, ω, ω0 ∈ P 0 E,s(R2d), p ∈ (0, ∞]2d, B be an invariant BF-space on R2d or B = Lp,E(R2d) for some phase-shift split parallelepiped E in R2d, and let φ ∈ Ss(Rd). Then the Toeplitz operator Tpφ(ω0) is an isomorphism from M (ω, B) onto M (ω/ω0, B). Remark 5.7. As remarked and stated before, there are different ways to extend the definition of a Toeplitz operator Tpφ(a) (from φ ∈ Σ1(Rd) and a ∈ Σ1(R2d)) to more general classes of symbols and windows. For example, Propositions 1.33 and 1.34 are based on the "classical" definition (1.37) of such operators and a straightforward extension of the L2-form on Σ1. Proposition 5.6 interprets Tpφ(ω) LIFTINGS FOR MODULATION SPACES, AND ONE-PARAMETER GROUPS OF ΨDO 37 as a pseudo-differential operator. Let us emphasize that in this context the bilinear E,s(R2d) form (1.37) may not be well defined, even when φ ∈ M 2 satisfies ω ∈ Γ(ω) (R2d). (See also [26, Remark 3.8].) (v)(Rd) and ω ∈ P 0 s In Theorems 5.1′ and 5.2′ below, we extend the definition of Toeplitz opera- tors within the framework of pseudo-differential calculus and we interpret Toeplitz operators as pseudo-differential operators. With this understanding, the stated mapping properties are well-defined. The reader who is not interested in full generality or does not accept Toeplitz operators that are not defined directly by an extension of (1.37) may only consider the case when the windows belong to M 1 (v) and stay with Theorems 5.1 and 5.2. For the more general window classes in Theorems 5.1′ and 5.2′ below, however, one should then interpret the involved operators as "pseudo-differential operators that extend Toeplitz operators". The following generalization of Theorem 5.1 is an immediate consequence of Theorem 4.1, Lemma 5.4 and Proposition 5.6. (R2d) Theorem 5.1′. Let s ≥ 1, ω, v, v0 ∈ P 0 and that ω0 is v-moderate, and let B be an invariant BF-space on R2d or B = (v)(Rd), Lp,E(R2d) for some phase-shift split parallelepiped E in R2d. If φ ∈ M 2 then Tpφ(ω0) is an isomorphism from M (ω, B) to M (ω/ω0, B). E,s(R2d) be such that ω0 ∈ Γ(ω0) s Theorem 5.1′ holds only for smooth weight functions. In order to relax the conditions on the weight function ω0, we use the Wiener algebra property of M∞,1 (v) instead of Γ(1) s . On the other hand, we have to restrict our results to modulation spaces of the form M p,q (ω) instead of M (ω, B). Theorem 5.2′. Let s > 1, 0 ≤ t ≤ 1, p, q, q0 ∈ [1, ∞] and ω, ω0, v0, v1 ∈ PE,s(R2d) be such that ω0 is v0-moderate and ω is v1-moderate. Set r0 = 2q0/(2q0 − 1), (v)(Rd) and v = vt ω0 ∈ M∞,q0 (ω/ϑ)(Rd). 1v0, ϑ = ω1/2 (1/ω0,t), then Tpφ(ω0) is an isomorphism from M p,q and let ω0,t be the same as in (1.38). If φ ∈ M r0 (ϑω)(Rd) to M p,q 0 Proof. First we note that the Toeplitz operator Tpφ(ω0) is an isomorphism from (1/ϑ) in view of Lemma 5.4. With ω1 defined in (5.4), Proposition 1.33′ M 2 implies that there exist b ∈ M∞,1 (ϑ) to M 2 s(R2d) such that (ω1) and c ∈ S′ Tpφ(ω0) = Opw(b) and Tpφ(ω0)−1 = Opw(c) . Let ω2 be the "dual" weight defined as ω2(X, Y ) = ϑ(X − Y )ϑ(X + Y )v1(2Y ). (5.7) We shall prove that c ∈ M∞,1 proved the existence of such a symbol c. Then we may proceed as follows. (ω2)(R2d). Let us assume for now that we have already After checking (1.30), we can apply Proposition 1.29 and find that each of the mappings Opw(b) : M p,q (ωϑ) → M p,q (ω/ϑ) and Opw(c) : M p,q (ω/ϑ) → M p,q (ωϑ) (5.8) is well-defined and continuous. In order to apply Proposition 1.30, we next check condition (1.31) for the weights ω1, ω2, and ω3(X, Y ) = ϑ(X + Y ) ϑ(X − Y ) . 38 AHMED ABDELJAWAD, SANDRO CORIASCO, AND JOACHIM TOFT In fact we have ω1(X − Y + Z, Z)ω2(X + Z, Y − Z) = (cid:16) v0(2Z)1/2v1(2Z) ϑ(X − Y + 2Z)ϑ(X − Y )(cid:17) ·(cid:0)ϑ(X − Y + 2Z)ϑ(X + Y )v1(2(Y − Z))(cid:1) v0(2Z)1/2v1(2Z)v1(2(Y − Z)) ϑ(X + Y ) = & ϑ(X − Y ) ϑ(X + Y ) ϑ(X − Y ) = ω3(X, Y ) . (ϑ) to M 2 (ω3)(R2d), or equivalently, b#c ∈ M∞,1 Therefore Proposition 1.30 shows that the Weyl symbol of Opw(b) ◦ Opw(c) belongs to M∞,1 (ω3). Since Opw(b) is an isomorphism from (1/ϑ) with inverse Opw(c), it follows that b#c = 1 and that the constant M 2 symbol 1 belongs to M∞,1 (ω3). By similar arguments it follows that c#b = 1. Therefore the identity operator Id = Opw(b) ◦ Opw(c) on M p,q (ω/ϑ), and thus Opw(b) = Tpφ(ω0) is an isomorphism from M p,q (ω/ϑ) with inverse Opw(c). This proves the assertion. (ωϑ) factors through M p,q (ωϑ) onto M p,q It remains to prove that c ∈ M∞,1 Section 3 (cf. Theorem 3.9), we choose a ∈ Γ(1/ϑ) that the map 0,s (ω2)(R2d). Using once again the basic result in 0,s (R2d) such (R2d) and a0 ∈ Γ(ϑ) Opw(a) : L2(Rd) → M 2 (ϑ)(Rd) is an isomorphism with inverse Opw(a0). By Propositions 1.25 and 1.26, Opw(a) is also bijective from M 2 (1/ϑ)(Rd) to L2(Rd). Furthermore, by Theorem 4.2 it follows that a ∈ M∞,1 (ϑr) when r ≥ 0, where ϑr(X, Y ) = ϑ(X)erY 1 s . Let b0 = a#b#a. From Proposition 1.31 we know that b0 ∈ M∞,1 (v2)(R2d), where v2(X, Y ) = v1(2Y ) (5.9) is submultiplicative and depends on Y only. Since Opw(b) is bijective from M 2 (ϑ) to M 2 (1/ϑ) by Lemma 5.4 (2), Opw(b0) is bijective and continuous on L2. Since v2 is submultiplicative and in PE,s(R2d), M∞,1 (v2) is a Wiener algebra by Proposition 1.32. Therefore, the Weyl symbol c0 of the inverse to the bijective operator Opw(b0) on L2 belongs to M∞,1 (v2)(R2d). Opw(c0) = Opw(b0)−1 = Opw(a)−1 Opw(b)−1 Opw(a)−1, Since we find Opw(a0) = Opw(b)−1 = Opw(a) Opw(c0) Opw(a), or equivalently, a0 = a#c0#a, where a ∈ Γ(1/ϑ) 0,s and c0 ∈ M∞,1 (v2) . (5.10) The definitions of the weights are chosen such that Proposition 1.31 implies that a0 ∈ M∞,1 (cid:3) (ω2), and the result follows. LIFTINGS FOR MODULATION SPACES, AND ONE-PARAMETER GROUPS OF ΨDO 39 6. Examples on bijective pseudo-differential operators on modulation spaces In this section we construct explicit isomorphisms between modulation spaces with different weights. Applying the results of the previous sections, these may be in the form of pseudo-differential operators or of Toeplitz operators. E,s(R2d), and let B be an invariant BF- Proposition 6.1. Let s ≥ 1, ω0, ω ∈ P 0 space on R2d or B = Lp,E(R2d) for some phase-shift split parallelepiped E in R2d. For λ = (λ1, λ2) ∈ R2 + let Φλ be the Gaussian Φλ(x, ξ) = Ce−(λ1x2+λ2ξ2) . (1) ω0 ∗ Φλ belongs to P 0 E,s(R2d) ∩ Γ(ω0) 0,1 for all λ ∈ R2 + and ω0 ∗ Φλ ≍ ω0. (2) If λ1 · λ2 < 1, then there exists ν ∈ R2 + and a Gauss function φ on Rd such that Opw(ω0∗Φλ) = Tpφ(ω0∗Φν) is bijective from M (ω, B) to M (ω/ω0, B) for all ω ∈ PE,s(R2d). (3) If λ1·λ2 ≤ 1 and in addition ω0 ∈ Γ(ω0) s (R2d), then Opw(ω0∗Φλ) = Tpφ(ω0) is bijective from M (ω, B) to M (ω/ω0, B) for all ω ∈ PE,s(R2d). Proof. The assertion (1) follows easily from the definitions. (2) Choose µj > λj such that µ1 · µ2 = 1. Then the Gaussian Φµ is a multiple of a Wigner distribution, precisely Φµ = cW (φ, φ) with φ(x) = e−µ1x2/2. By the semigroup property of Gaussian functions (cf. e.g., [19, 21]) there exists another Gaussian, namely Φν, such that Φλ = Φµ ∗ Φν . Using (1.39), this factorization implies that the Weyl operator with symbol ω0 ∗ Φλ is in fact a Toeplitz operator, namely Opw(ω0 ∗ Φλ) = Opw(ω0 ∗ Φν ∗ Φµ) = Opw(ω0 ∗ Φν ∗ cW (φ, φ)) = c(2π)d/2Tpφ(ω0 ∗ Φν ). By (1) ω0 ∗Φν ∈ P 0 0,1 (R2d) is equivalent to ω0. Hence Theorem 5.1′ shows that Opw(ω0 ∗ Φλ) is bijective from M (ω, B) to M (ω/ω0, B). This proves (2). E,s(R2d)∩Γ(ω0) (3) follows from (2) in the case λ1 · λ2 < 1. Φλ = cW (φ, φ) for φ(x) = e−λ1x2/2 and thus If λ1 · λ2 = 1, then as above Opw(ω0 ∗ Φλ) = Tpw φ (ω0) is bijective from M (ω, B) to M (ω/ω0, B), since ω0 ∈ P 0 E,s(R2d) ∩ Γ(ω0) s (R2d). (cid:3) Appendix A. Proof of Lemma 3.3 Lemma A.1. Let α = (α1, . . . , αd) ∈ Nd. Then the number of elements in the set Ωk,α ≡ { (β1, . . . , βk) ∈ Nkd ; β1 + · · · + βk = α } (A.1) is equal to dYj=1(cid:18)αj + k k (cid:19). 40 AHMED ABDELJAWAD, SANDRO CORIASCO, AND JOACHIM TOFT For the proof we recall the formula kXj=0(cid:18)n + j j (cid:19) =(cid:18)n + k + 1 (cid:19), k which follows by a standard induction argument. (A.2) Proof. Let N be the number of elements in the set (A.1), which is the searched number, and let Nj be the number of elements of the set { (β0 1, . . . , β0 k) ∈ Nk ; β0 1 + · · · + β0 k = αj }, j = 1, . . . , d By straightforward computations it follows that N = N1 · · · Nd, and it suffices to prove the result in the case d = 1, and then α = α1. In order to prove the result for d = 1, let γ ∈ N, S1(γ) = γXβ=0 1 = γ + 1, and define inductively By straightforward computations it follows that N = N1 = Sk(α). We claim Sj(β), j = 1, 2, . . . . Sj+1(γ) = γXβ=0 Sj(γ) =(cid:18)γ + j j (cid:19), j = 1, 2, . . . . (A.3) In fact, (A.3) is obviously true for j = 1. Assume that (A.3) holds for j = n, and consider Sn+1(γ). Then (A.2) gives Sn+1(γ) = Sn(β) = γXβ=0 γXβ=0(cid:18)β + n n (cid:19) γXβ=0(cid:18)β + n = β (cid:19) =(cid:18)γ + n + 1 γ (cid:19) =(cid:18)γ + n + 1 n + 1 (cid:19), which gives (A.3) when j = n + 1. This proves (A.3), and the result follows. (cid:3) Lemma A.2. Let α ≥ 1 be an integer, s0 ∈ (0, 1], and let Ωk,α be the same as in (A.1). Then there is a constant C which is independent of α such that Proof. By Lemma A.1 and the fact that s0 − 1 < 0 we get (β!)s0−1 ≤ 16α. αXk=1 αXk=1 1 k Xβ∈Ωk,α  Xβ∈Ωk,α αXk=1 1 k Xβ∈Ωk,α β!s0−1 ≤ 1 = αXk=1(cid:18)α + k k (cid:19) ≤ (1 + α)22α ≤ 16α. (cid:3) References [1] T. Aoki Locally bounded topological spaces, Proc. Japan Acad. 18 (1942), 588 -- 594. [2] H.A. Biagioni,T. Gramchev Fractional derivative estimates in Gevrey classes, global regu- larity and decay for solutions to semilinear equations in Rd. J. Differential Equations 194 (2003), 140 -- 165. [3] J. M. Bony, J. Y. Chemin Espaces Functionnels Associ´es au Calcul de Weyl-Hormander, Bull. Soc. math. France 122 (1994), 77 -- 118. LIFTINGS FOR MODULATION SPACES, AND ONE-PARAMETER GROUPS OF ΨDO 41 [4] M. Cappiello, J. Toft Pseudo-differential operators in a Gelfand-Shilov setting, Math. Nachr. 290 (2017), 738 -- 755. [5] Y. Chen, M. Signahl, J. Toft Factorizations and singular value estimates of operators with Gelfand-Shilov and Pilipovi´c kernels, J. Fourier Anal. Appl. (appeared online 2017). [6] J. Chung, S.Y. Chung, D. Kim Characterization of the Gelfand-Shilov spaces via Fourier transforms. Proc. Amer. Math. Soc. 124 (1996), 2101 -- 2108. [7] E. Cordero, S. Pilipovi´c, L. Rodino, N. Teofanov Quasianalytic Gelfand-Shilov spaces with applications to localization operators, Rocky Mt. J. Math. 40 (2010), 1123-1147. [8] E. Cordero, K. Grochenig Time-Frequency Analysis of Localization Operators, J. Funct. Anal. (1) 205 (2003), 107 -- 131. [9] E. Cordero, S. Pilipovi´c, L. Rodino, N. Teofanov Localization operators and exponential weights for modulation spaces. Mediterr. J. Math. 2 (2005), 381 -- 394. [10] E. Cordero, J. Toft, P. Wahlberg Sharp results for the Weyl product on modulation spaces, J. Funct. Anal. 267 (2014), 3016 -- 3057. [11] H. G. Feichtinger Banach convolution algebras of Wiener's type, in: Proc. Functions, Se- ries, Operators in Budapest, Colloquia Math. Soc. J. Bolyai, North Holland Publ. Co., Amsterdam Oxford NewYork, 1980. [12] H. G. Feichtinger Modulation spaces on locally compact abelian groups. In Proceedings of "International Conference on Wavelets and Applications" 2002, pages 99 -- 140, Chennai, India, 2003. Updated version of a technical report, University of Vienna, 1983. [13] H. G. Feichtinger Atomic characterizations of modulation spaces through Gabor-type repre- sentations, in: Proc. Conf. on Constructive Function Theory, Rocky Mountain J. Math. 19 (1989), 113 -- 126. [14] H. G. Feichtinger Wiener amalgams over Euclidean spaces and some of their applications, in: Function spaces (Edwardsville, IL, 1990), Lect. Notes in pure and appl. math., 136, Marcel Dekker, New York, 1992, pp. 123 -- 137. [15] H. G. Feichtinger Modulation spaces: Looking back and ahead, Sampl. Theory Signal Image Process. 5 (2006), 109 -- 140. [16] H. G. Feichtinger and K. H. Grochenig Banach spaces related to integrable group represen- tations and their atomic decompositions, I, J. Funct. Anal. 86 (1989), 307 -- 340. [17] C. Fernandez, A. Galbis, J. Toft Spectral invariance for matrix algebras, J. Fourier Anal. Appl. 20 (2014), 362 -- 383. [18] H. G. Feichtinger and K. H. Grochenig Banach spaces related to integrable group represen- tations and their atomic decompositions, II, Monatsh. Math. 108 (1989), 129 -- 148. [19] G. B. Folland Harmonic analysis in phase space, Princeton Univ. Press, Princeton, 1989. [20] Y. V. Galperin, S. Samarah Time-frequency analysis on modulation spaces M p,q m , 0 < p, q ≤ ∞, Appl. Comput. Harmon. Anal. 16 (2004), 1 -- 18. [21] K. H. Grochenig Foundations of Time-Frequency Analysis, Birkhauser, Boston, 2001. [22] K. H. Grochenig Composition and spectral invariance of pseudodifferential operators on modulation spaces, J. Anal. Math. 98 (2006), 65 -- 82. [23] K. H. Grochenig Time-frequency analysis of Sjostrand's class, Rev. Mat. Iberoam. 22 (2006), 703 -- 724. [24] K. Grochenig Weight functions in time-frequency analysis in: L. Rodino, M. W. Wong (Eds) Pseudodifferential Operators: Partial Differential Equations and Time-Frequency Analysis, Fields Institute Comm., 52 2007, pp. 343 -- 366. [25] K. Grochenig and T. Strohmer. Pseudodifferential operators on locally compact abelian groups and Sjostrand's symbol class, J. Reine Angew. Math., 613:121 -- 146, 2007. [26] K. Grochenig, J. Toft Isomorphism properties of Toeplitz operators and pseudo-differential operators between modulation spaces, J. Anal. Math. 114 (2011), 255 -- 283. [27] K. Grochenig, J. Toft The range of localization operators and lifting theorems for modulation and Bargmann-Fock spaces, Trans. Amer. Math. Soc. 365 (2013), 4475 -- 4496. [28] K.H. Grochenig, G. Zimmermann Spaces of test functions via the STFT. J. Funct. Spaces. Appl. 2 (2004), 25 -- 53. [29] H. Gzyl Multidimensional Extension of Fa`a di Bruno's formula, J. Math. Anal. Appl. 116 (1986), 450 -- 455. [30] L. Hormander The Weyl calculus of pseudo-differential operators, Comm. Pure Appl. Math. 32 (1979), 359 -- 443. [31] L. Hormander The Analysis of Linear Partial Differential Operators, vol I, III, Springer- Verlag, Berlin Heidelberg NewYork Tokyo, 1983, 1985. [32] Z. Lozanov-Crvenkovi´c, D. Perisi´c, M. Taskovi´c Gelfand-Shilov spaces structural and kernel theorems, (preprint), arXiv:0706.2268v2. [33] N. Lerner The Wick calculus of pseudo-differential operators and some of its applications, CUBO, 5 (2003), 213 -- 236. 42 AHMED ABDELJAWAD, SANDRO CORIASCO, AND JOACHIM TOFT [34] N. Lerner Metrics on the Phase Space and Non-Selfadjoint Pseudo-Differential Operators, Birkhauser Springer, (2010). [35] B.S. Mitjagin Nuclearity and other properties of spaces of type S , Amer. Math. Soc. Transl., Ser. 93 (1970), 49 -- 59. [36] C. Pfeuffer, J. Toft Compactness properties for modulation spaces (under preparation). [37] S. Pilipovi´c Tempered ultradistributions, Boll. Un. Mat. Ital., VII Ser., 2 (1988), 235 -- 251. [38] S. Pilipovi´c, N. Teofanov Pseudodifferential operators on ultramodulation spaces, J. Funct. Anal. 208 (2004), 194 -- 228. [39] S. Rolewicz On a certain class of linear metric spaces, Bull. Acad. Polon. Sci. S´er. Schi. Math. Astronom. Phys. 5 (1957), 471 -- 473. [40] M. Ruzhansky, M. Sugimoto, N. Tomita, J. Toft Changes of variables in modulation and Wiener amalgam spaces, Math. Nachr. 284 (2011), 2078 -- 2092. [41] M. Ruzhansky, N. Tokmagambetov Nonharmonic analysis of boundary value problems, Int. Math. Res. Notices 12 (2016), 3548 -- 3615. [42] J. Sjostrand An algebra of pseudodifferential operators, Math. Res. L. 1 (1994), 185 -- 192. [43] J. Sjostrand. Pseudodifferential operators and weighted normed symbol spaces, Serdica Math. J. 34 (2008), no. 1, 1 -- 38. [44] K. Tachizawa The boundedness of pseudo-differential operators on modulation spaces, Math. Nachr. 168 (1994), 263 -- 277. [45] N. Teofanov Modulation spaces, Gelfand-Shilov spaces and pseudodifferential operators, Sampl. Theory Signal Image Process, 5 (2006), 225 -- 242. [46] J. Toft Continuity Properties for non-Commutative Convolution Algebras with Applications in Pseudo-Differential Calculus, Bull. Sci. Math. (2) 126 (2002), 115 -- 142. [47] J. Toft Continuity properties for modulation spaces with applications to pseudo-differential calculus, II, Ann. Global Anal. Geom., 26 (2004), 73 -- 106. [48] J. Toft Continuity and Schatten properties for pseudo-differential operators on modulation spaces in: J. Toft, M. W. Wong, H. Zhu (Eds) Modern Trends in Pseudo-Differential Op- erators, Operator Theory: Advances and Applications 172, Birkhauser Verlag, Basel, 2007, pp. 173 -- 206. [49] J. Toft Continuity and Schatten properties for Toeplitz operators on modulation spaces in: J. Toft, M. W. Wong, H. Zhu (Eds) Modern Trends in Pseudo-Differential Operators, Operator Theory: Advances and Applications 172, Birkhauser Verlag, Basel, 2007, pp. 313 -- 328. [50] J. Toft Pseudo-differential operators with smooth symbols on modulation spaces, Cubo. 11 (2009), 87 -- 107. [51] J. Toft Multiplication properties in pseudo-differential calculus with small regularity on the symbols, J. Pseudo-Differ. Oper. Appl. 1 (2010), 101 -- 138. [52] J. Toft The Bargmann transform on modulation and Gelfand-Shilov spaces, with applica- tions to Toeplitz and pseudo-differential operators, J. Pseudo-Differ. Oper. Appl. 3 (2012), 145 -- 227. [53] J. Toft Multiplication properties in Gelfand-Shilov pseudo-differential calculus in: S. Mola- hajlo, S. Pilipovi´c, J. Toft, M. W. Wong (eds) Pseudo-Differential Operators, Generalized Functions and Asymptotics, Operator Theory: Advances and Applications 231, Birkhauser, Basel Heidelberg NewYork Dordrecht London, 2013, pp. 117 -- 172. [54] J. Toft Gabor analysis for a broad class of quasi-Banach modulation spaces, in: S. Pilipovi´c, J. Toft (eds) Pseudo-Differential Operators and Generalized Functions, Operator Theory: Advances and Applications Vol 245, Birkhauser, Basel Heidelberg NewYork Dordrecht Lon- don, 2015, pp. 249 -- 278. [55] J. Toft Matrix parameterized pseudo-differential calculi on modulation spaces in: M. Ober- guggenberger, J. Toft, J. Vindas, P. Wahlberg (eds), Generalized functions and Fourier analysis, Operator Theory: Advances and Applications Vol 260, Birkhauser/Springer, Basel, 2017, pp. 215 -- 235. [56] J. Toft Continuity of Gevrey-Hormander pseudo-differential operators on modulation spaces, (preprint) arXiv:1710.11366 (2017). [57] J. Toft, P. Boggiatto, Schatten classes for Toeplitz operators with Hilbert space windows on modulation spaces, Adv. Math. 217 (2008), no. 1, 305 -- 333. [58] J. Toft, A. Khrennikov, B. Nilsson, S. Nordebo Decompositions of Gelfand-Shilov kernels into kernels of similar class, J. Math. Anal. Appl. 396 (2012), 315 -- 322. LIFTINGS FOR MODULATION SPACES, AND ONE-PARAMETER GROUPS OF ΨDO 43 Dipartimento di Matematica "G. Peano", Universit´a degli Studi di Torino E-mail address: [email protected] Dipartimento di Matematica "G. Peano", Universit´a degli Studi di Torino E-mail address: [email protected] Department of Mathematics, Linnaeus University, Sweden E-mail address: [email protected]
1309.3146	1	1309	2013-09-12T13:09:06	Further results on regular Fredholm pairs and chains	[ "math.FA" ]	The main objective of the present article is to characterize regular Fredholm pairs and chains in terms of Fredholm operators.	math.FA	math	FURTHER RESULTS ON REGULAR FREDHOLM PAIRS AND CHAINS ENRICO BOASSO The main objective of the present article is to characterize regular Fredholm pairs and chains in terms of Fredholm operators. AMS 2000 Subject Classification: Primary 47A13; Secondary 47A53. Key words: Fredholm pairs and chains, regular operators. 1. INTRODUCTION In multiparameter operator theory there exist objects that in the context of Hilbert spaces can be characterized using linear and bounded maps, to mention only some of the works related to this subject, see for example [9, 10, 11, 12]. Since the adjoint plays a key role in this relationship between one and several variable operator theory, it is not possible to directly translate these results to the frame of Banach spaces. However in [7] R. Harte and W. Y. Lee showed that generalized inverses could be used to reformulate the Hilbert space situation for regular Fredholm Banach space chains and complexes. On the other hand, Fredholm pairs and chains have been recently intro- duced and their main properties have been studied, see [1, 2, 4, 5, 8] and the monograph [3]. These objects consist in generalizations of the notions of Fredholm operators and Fredholm Banach space complexes respectively. In particular, in [4] Fredholm pairs and chains in Hilbert spaces were charac- terized using the ideas mentioned in the previous paragraph. Furthermore, in [5] regular Fredholm pairs, that is Fredholm pairs whose operators admit generalized inverses, were characterized and classified. The main objective of the present article is to characterize regular Fredholm pairs and chains in Banach spaces using Fredholm operator extending to these objects the characterizations of [4] using the approach of R. Harte and W. Y. Lee in [7]. Naturally, as in [7], since the adjoint can not be considered any more, suitable generalized inverses must be defined to prove the main results of this work. The article is organized as follows. In the next section the definitions of the objects under consideration will be recalled. In section 3 two character- izations of regular Fredholm pairs will be given. Finally, in section 4 the results of section 3 will be applied to regular Fredholm chains to obtain two characterizations of them. 1 2 ENRICO BOASSO 2. PRELIMINARY DEFINITIONS From now on X and Y will denote two Banach spaces and L(X, Y ) will stand for the algebra of all linear and continuous operators defined on X with values in Y . As usual, when X = Y , L(X, X) is denoted by L(X). For every S ∈ L(X, Y ), the null space of S is denoted by N (S) = {x ∈ X : S(x) = 0}, and the range of S by R(S) = {y ∈ Y : ∃ x ∈ X such that y = S(x)}. Next follow the definitions of Fredholm pairs and chains, see for instance [1, 3, 8]. Definition 2.1. Let X and Y be two Banach spaces and let S ∈ L(X, Y ) and T ∈ L(Y, X) be such that the following dimensions are finite: (i) a : = dim N (S)/(N (S) ∩ R(T )), b : = dim R(T )/(N (S) ∩ R(T )), (ii) c : = dim N (T )/(N (T ) ∩ R(S)), d : = dim R(S)/(N (T ) ∩ R(S)). A pair (S, T ) with the above properties is said to be a Fredholm pair. Let P (X, Y ) denote the set of all Fredholm pairs. If (S, T ) ∈ P (X, Y ), then the index of (S, T ) is defined by the equality ind(S, T ) : = a − b − c + d. In particular, if (S, T ) ∈ P (X, Y ) is such that ST = 0 and T S = 0, that is if b = d = 0, then (S, T ) and (T, S) are Fredholm chains in the sense of [6, section 10.6] and [7]. Definition 2.2. A Fredholm chain (Xp, δp)p∈Z is a sequence of Banach spaces Xp and bounded operators δp ∈ L(Xp, Xp−1) such that there is a natural number n with the property Xp = 0, p < 0 and p ≥ n + 1, δp = 0, p ≤ 0 and p > n, and N (δp)/(N (δp) ∩ R(δp+1)) and R(δp+1)/(N (δp) ∩ R(δp+1)) are finite dimensional subspaces of Xp, p ∈ Z. Given a Fredholm chain, it is possible to associate an index to it. In fact, if (Xp, δp)p∈Z is such an object, then define ind(Xp, δp)p∈Z = X (−1)pdp, n p=0 where dp = dim N (δp)/(N (δp) ∩ R(δp+1)) − dim R(δp+1)/(N (δp) ∩ R(δp+1)), see [8]. Recall that in [8] the more general concept of semi-Fredholm chain was in- troduced. However, since the main concern of this article consists in Fredholm objects, only Fredholm chains will be considered. Remark 2.3. There is a natural relationship between Fredholm pairs and In fact, given as in Definition 2.2 a sequence of spaces and maps chains. (Xp, δp)p∈Z , consider the Banach spaces X = M Xp, Y = M Xp, p=2k p=2k+1 FREDHOLM PAIRS AND CHAINS 3 and the Banach space operators S ∈ L(X, Y ) and T ∈ L(Y, X) defined by S = M δp, p=2k T = M δp, p=2k+1 where Xp = 0, p < 0 and p ≥ n + 1, δp = 0, p ≤ 0 and p > n, and n is a natural number. Now well, it is not difficult to prove that dim R(ST ) and dim R(T S) are finite dimensional if and only if dim R(δp+1)/(N (δp) ∩ R(δp+1)) are finite dimensional, p = 0, . . ., n. Furthermore, a straightforward calculation shows that necessary and sufficient for N (S)/(N (S)∩R(T )) and N (T )/(N (T )∩R(S)) to be finite dimensional is the fact that N (δp)/(N (δp) ∩ R(δp+1)) are finite dimensional, p = 0, . . ., n. Consequently, (Xp, δp)p∈Z is a Fredholm chain if and only if (S, T ) is a Fredholm pair. Finally, in this case, ind(Xp, δp)p∈Z = ind(S, T ), see [4, Remark 2.4]. On the other hand, recall that an operator T ∈ L(X, Y ) is called regular, if there is S ∈ L(Y, X) such that T = T ST . The map S is said to be a generalized inverse of T . In addition, if T is also a generalized inverse of S, that is if S = ST S, then S is said to be a normalized generalized inverse of T . Note that if T is regular, then T always has a normalized generalized inverse. In fact, if S is a generalized inverse of T , then S ′ = ST S is a normalized generalized inverse of T . In the following definition, the notions of regular Fredholm pairs and chains will be recalled, see [5]. Definition 2.4. Let X and Y be two Banach spaces and consider S ∈ L(X, Y ) and T ∈ L(Y, X) such that (S, T ) ∈ P (X, Y ). The pair (S, T ) will be said to be a regular Fredholm pair, if S and T are regular operators. Similarly, given a Fredholm chain (Xp, δp)p∈Z , then (Xp, δp)p∈Z will be said to be a regular Fredholm chain, if the Fredholm pair defined by (Xp, δp)p∈Z is regular. Note that given (S, T ) ∈ P (X, Y ), several statements equivalent to the fact that (S, T ) is a regular Fredholm pair were considered in [5, Proposition 2.4]. Concerning regular Fredholm chains, see [5, Remark 2.7]. 3. CHARACTERIZATIONS OF REGULAR FREDHOLM PAIRS First of all, a preliminary remark is presented. Remark 3.1. Let X and Y be two Banach spaces, and let S ∈ L(X, Y ) and T ∈ L(Y, X) be two operators such that R(ST ) and R(T S) are finite dimensional subspaces of Y and X respectively. Then, define the Banach spaces X = X/R(T S) and Y = Y /R(ST ), and the linear and bounded maps S ∈ L(X , Y) and T ∈ L(Y, X ) as the factorization of S and T through the respective invariant subspaces. Clearly, S T = 0 and T S = 0, that is ( S, T ) and ( T , S) are chains in the sense of [7]. Moreover, the regularity of S and T is equivalent to the one of S and T . 4 ENRICO BOASSO THEOREM 3.2. In the same conditions of Remark 3.1, the following state- ments are equivalent. (i) S and T are regular operators. (ii) S and T are regular operators. Proof. Since S ∈ L(X, Y ) is a regular operator, according to [6, Theorem 3.8.2], there is M a linear subspace of Y such that R(S) ⊕ M = Y . Conse- quently, R( S) + πY (M ) = Y, where πY : Y → Y is the quotient map. Next suppose that there is x ∈ X such that S(x) = m ∈ πY (M ). In particular, there exist x ∈ X, m ∈ M , and y ∈ Y such that πX(x) = x, πY (m) = m, and S(x) − m = ST (y), where πX : X → X is the quotient map. However, in this case S(x − T (y)) = m. Then, m ∈ M ∩ R(S) = 0, and S(x) = m = 0. Therefore, R( S) ⊕ πY (M ) = Y. On the other hand, according to [6, Theorem 3.8.2], there exists N, a vector subspace of X such that N (S) ⊕ N = X. However, since (N (S) + R(T ))/N (S) ∼= R(T ))/(N (S) ∩ R(T )) ∼= R(ST ), there is a finite dimensional subspace X1 such that N (S)⊕X1 = N (S)+R(T ). Then, a straightforward calculation proves that there exists a closed vector subspace R of X such that (N (S) + R(T )) ⊕ R = X. Now well, according to [1, Remark 2.1], N ( S) = πX(N (S) + R(T )). particular, N ( S) + πX (R) = X . In Next suppose that there are n ∈ N (S), y ∈ Y , and r ∈ R such that πX(n + T (y)) = πX(r). Then, there is x0 ∈ X such that n + T (y) − r = T S(x0). In particular, n + T (y − S(x0)) = r. Thus r = 0, which implies that πX(r) = πX(n + T (x)) = 0. Therefore, N ( S) ⊕ πX(R) = X . As a result, according to what has been proved and to [6, Theorem 3.8.2], the operator S is regular. Interchanging S with T and S with T , it can be proved that T is regular. In order to prove the converse implication, suppose that S and T are regular operators. Consider V, a closed vector subspace of Y, such that R( S) ⊕ V = Y. Let V1 = π−1 Y (V)∩R(ST ). Since V1 has finite dimension, there is a vector subspace W1 such V1 ⊕ W1 = π−1 Y (V). Moreover, since V1 ⊆ R(ST ) and πY is surjective, V = πY (W1). Consequently, πY (R(S) + W1) = R( S) + V = Y, which implies that R(S)+W1 +R(ST ) = Y . However, since R(ST ) ⊆ R(S), R(S)+W1 = Y . Next define L = R(S) ∩ W1. Then, πY (L) ⊆ R( S) ∩ V = 0. Thus, L ⊆ R(ST ). However, since W1 ∩ R(ST ) = 0, L = 0 and R(S) ⊕ W1 = X. Similarly, suppose that U is a closed subspace of X such that N ( S) ⊕ U = X (U ) ∩ R(T S). Since U1 is a finite dimensional subspace, X (U ). Furthermore, since X . Let U1 = π−1 there exists a subspace Z1 such that U1 ⊕ Z1 = π−1 πX : X → X is a surjecive map, πX(Z1) = U . FREDHOLM PAIRS AND CHAINS 5 Now well, according to [1, Remark 2.1], πX(N (S)+R(T )+Z1) = πX (N (S)+ R(T )) + πX (Z1) = N ( S) + U = X . Therefore, since R(T S) ⊆ R(T ), N (S) + R(T ) + Z1 = X. Let P = (N (S) + R(T )) ∩ Z1. According to [1, Remark 2.1], πX(P ) ⊆ N ( S)∩U = 0. Thus, P ⊆ R(T S). Consequently, since Z1 ∩R(T S) = 0, P = 0, and (N (S) + R(T )) ⊕ Z1 = X. However, since N (S) ⊕ X1 = N (S) + R(T ), where X1 is a finite dimensional vector subspace of X, N (S) ⊕ (X1 + Z1) = X, which, according to what has been proved and to [6, Theorem 3.8.2], implies that S is a regular operator. operator. Interchanging S with T and S with T , it can be proved that T is a regular (cid:3) In order to prove the first characterization concerning regular Fredholm pairs, some preparation is needed. Remark 3.3. Let X, Y , S, T , X , Y, S, and T be as in Remark 3.1. Recall that S T = 0 and T S = 0. Furthermore, according to Theorem 3.2, necessary and sufficient for S and T to be regular bounded and linear maps is the fact that ( S, T , S) is a regular chain in the sense of [7, p. 283-284]. In addition, according again to [7, p. 284], there exist S ′ ∈ L(Y, X ) and T ′ ∈ L(X , Y) such that ( T ′, S ′, T ′) is a regular chain in the sense of [7] with the property that S ′ (respectively T ′) is a normalized generalized inverse of S (respectively T ). Now well, since R(ST ) and R(T S) are finite dimensional subspaces of Y and X respectively, X and Y are isomorphic to finite codimensional closed subspaces of X and Y respectively. Consequently, using Banach space iso- morphisms, S ′ and T ′ can be thought of operators defined on and to finite codimensional closed subspaces of X and Y . Donote by S ′ ∈ L(Y, X) and T ′ ∈ L(X, Y ) the extension of S ′ and T ′ respectively, such that S ′ = 0 on R(T S) and T ′ = 0 on R(ST ). Note that S ′ and T ′ are regular maps. Next follows the first characterization of regular Fredholm pairs. THEOREM 3.4. In the same conditions of Remark 3.3, the following state- ments are equivalent. (i) (S, T ) is a regular Fredholm pair. (ii) S + T ′ ∈ L(X, Y ) is a Fredholm operator. (iii) T + S ′ ∈ L(Y, X) is a Fredholm operator. Futhermore, in this case ind(S, T ) = ind(S + T ′) = − ind(T + S ′). Proof. According to [1, Remark 2.1] and to Theorem 3.2, (S, T ) is a regular Fredholm pair if and only if ( S, T , S) is a Fredholm regular chain in the sense of [7]. Then, according to [7, Theorem 5], the first statement of the Theorem is equivalent to the fact that S + T ′ ∈ L(X , Y) is a Fredholm operator. Now well, since as in Remark 3.3 X and Y can be thought of finite codi- mensional closed subspaces of X and Y respectively, it is possible to define S1 ∈ L(Y, X) as the extension of S with the property S1 = 0 on R(T S). Con- sequently, since R(ST ) and R(T S) are finite dimensional vector subspaces, 6 ENRICO BOASSO according to what has been proved, the first statement of the Theorem is equivalent to the fact that S1 + T ′ ∈ L(Y, X) is a Fredholm operator. How- ever, since S − S1 is a compact operator, the first and the second statements of the Theorem are equivalent. As regard the index, according to [1, Remark 2.1] and to [7, Theorem 5] ind(S, T ) − dim R(T S) + dim R(ST ) = ind( S, T ) = ind( S + T ′). ind(S, T ) = ind(S + T ′). On the other hand, note that ind(S + T ′) = ind(S1 + T ′) = ind( S + T ′) + dim R(T S) − dim R(ST ). Therefore, A similar argument proves that the first and the third statements are equiv- (cid:3) The second characterization of regular Fredholm pairs is more general in the sense that the generalized inverses can be chosen more freely, however, the formula regarding the index can not be considered. On the other hand, to state this characterization, a Banach space operator is introduced. alent as well as the relationship between the indexes. Remark 3.5 Let X, Y , S, T , X , Y, S, and T be as in Remark 3.3. Since S, and T are regular operators, there exist S ′ ∈ L(Y, X ) and T ′ ∈ L(X , Y) generalized inverses of S and T respectively. In addition, proceeding as in Remark 3.3, denote by S ′ ∈ L(Y, X) and T ′ ∈ L(X, Y ) any extensions of S ′ and T ′ respectively. On the other hand, define the map V ∈ L(X⊕Y ) as follows: V X ∈ L(X, Y ), V X = S + T ′, and V Y ∈ L(Y, X), V Y = T + S ′. THEOREM 3.6. In the same conditions of Remark 3.5, the following state- ments are equivalent. (i) (S, T ) is a regular Fredholm pair. (ii) S ′S + T T ′ and T ′T + SS ′ are Fredholm operators. (iii) V is a Fredholm operator. Proof. As in Theorem 3.4, according to [1, Remark 2.1] and to Theorem 3.2, the first statement is satisfied if and only if ( S, T , S) is a Fredholm regular chain in the sense of [7], which, according to [7, Theorem 3], is equivalent to the fact that S ′ S + T T ′ ∈ L(X ) and T ′ T + S S ′ ∈ L(Y) are Fredholm operators. Now well, since R(T S) and R(ST ) are finite dimensional subspaces of X and Y respectively, it is not difficult to prove that what has been proved is equivalent to the second statement of the Theorem. On the other hand, V is a Fredholm operator if and only if V 2 also is Fredholm. Now well, a straightforward calculation proves that there exists a finite range operator F ∈ L(X ⊕ Y ) such that V 2 − F is a diagonal operator with entries S ′S + T T ′ ∈ L(X) and SS ′ + T T ′ ∈ L(Y ), which clearly implies that the second and the third statements are equivalent. (cid:3) FREDHOLM PAIRS AND CHAINS 7 4. CHARACTERIZATIONS OF REGULAR FREDHOLM CHAINS In order to prove the main results of this section, some preliminary facts will be considered. Remark 4.1 Let (Xp, δp)p∈Z be a sequence of spaces and maps such that Xp = 0, p < 0 and p ≥ n + 1, and δp = 0, p ≤ 0 and p > n, where n is a natural number. In addition, suppose that R(δpδp−1) is finite dimensional and δp ∈ L(Xp, Xp−1) is a regular operator, p ∈ Z. Then, if X, Y , S ∈ L(X, Y ), and T ∈ L(Y, X) are as in Remark 2.3, the properties of (Xp, δp)p∈Z are equivalent to the fact that R(ST ) and R(T S) are finite dimensional and S and T are regular operators. Now well, if X , Y, S and T are as in Remark 3.3, note that X = M Xp, p=2k S = M δp, p=2k Y = M Xp, p=2k+1 T = M δp, p=2k+1 where Xp = Xp/R(δp+1δp+2) and δp : Xp → Xp−1 is the quotient map induced by δp : Xp → Xp−1, p ∈ Z. Since S T = 0 and T S = 0, the sequence of spaces and maps (Xp, δp)p∈Z is a finite complex of Banach spaces. In addition, ac- cording to Theorem 3.2, S and T are regular operators, which is equivalent to the fact that δp : Xp → Xp−1 is regular, p ∈ Z. A straightforward calcula- tion proves that there exists a sequence of linear and bounded maps (δ′ p)p∈Z, p is a normalized generalized inverse of δp and p : Xp−1 → Xp such that δ′ δ′ (Xp, δ′ p)p∈Z is a finite complex of Banach spaces. Consequently, if S ′ = M δ′ p, p=2k T ′ = M δ′ p, p=2k+1 then ( T ′, S ′, T ′) is regular chain in the sense of [7] such that S ′ and T ′ are normalized generalized inverses of S and T . Finally, as in Remark 3.3, consider S ′ ∈ L(Y, X) and T ′ ∈ L(X, Y ), the extensions of S ′ and T ′ such that S ′ = 0 on R(T S) and T ′ = 0 on R(ST ), and denote by δ′ p : Xp−1 → Xp the extension of δ′ p such that δ′ THEOREM 4.2. In the same conditions of Remark 4.1, the following state- p = 0 on R(δpδp−1), p ∈ Z. ments are equivalent. (i) (Xp, δp)p∈Z is a regular Fredholm chain. (ii) ⊕p=2k(δp + δ′ p+1) is a Fredholm operator. (iii) ⊕p=2k+1(δp + δ′ Furthermore, in this case, p+1) is a Fredholm operator. ind(Xp, δp)p∈Z = ind ⊕p=2k (δp + δ′ p+1) = − ind ⊕p=2k+1 (δp + δ′ p+1). Proof. Apply Remark 2.3, Theorem 3.4, and Remark 4.1. (cid:3) 8 ENRICO BOASSO Remark 4.3 In the same conditions of Remark 4.1, consider the same spaces and maps but, as in Remark 3.5, S ′ and T ′ denote any generalized inverses of S and T . In addition, S ′ and T ′ can be any extentions of S ′ and T ′. Similarly, p is any extension of δ′ δ′ THEOREM 4.4. In the conditions of Remark 4.3, the following statements p, p ∈ Z. are equivalent. (i) (Xp, δp)p∈Z is a regular Fredholm chain. (ii) δp+1δ′ pδp ∈ L(Xp) is a Fredholm operator, p ∈ Z. p+1 + δ′ Proof. Apply Remark 2.3, Theorem 3.6, and Remarks 4.1 and 4.3. (cid:3) References [1] C.-G. Ambrozie, On Fredholm index in Banach spaces, Integral Equations Operator Theory 25 (1996), 1-34. [2] C.-G. Ambrozie, The Euler characteristic is stable under compact perturbations, Proc. Amer. Math. Soc. 124 (7) (1996), 2041-2050. [3] C.-G. Ambrozie and F.-H. Vasilescu, Banach Space Complexes, Kluwer Academic Plub- lishers, Dordrecht - Boston - London, 1995. [4] E. Boasso, Characterizations of Fredholm pairs and chains in Hilbert spaces, Rev. Roumaine Math. Pures Appl. 51 (2) (2006), 151-165. [5] E. Boasso, Regular Fredholm pairs, J. Operator Theory 55 (2) (2006), 311-337. [6] R. Harte, Invertibility and Singularity for Bounded Linear Operators, Marcel Dekker, Inc., New York and Basel 1988. [7] R. Harte and W. Y. Lee, An index formula for chains, Studia Math. 116 (3) (1995), 283-294. [8] V. Muller, Stability of index for semi-Fredholm chains, J. Operator Theory 37 (1997), 247-261. [9] F.-H. Vasilescu A characterization of the joint spectrum in Hilbert spaces, Rev. Roumaine Math. Pures Appl. 22 (1977), 1003-1009. [10] F.-H. Vasilescu On pairs of commuting operators, Studia Math. 62 (1978), 203-207. [11] F.-H. Vasilescu Analytic perturbations of the δ-operator and integral representation fromulas in Hilbert spaces, J. Operator Theory 1 (1979), 187-205. [12] F.-H. Vasilescu The stability of the Euler characteristic for Hilbert complexes, Math. Ann. 248 (1980), 109-116. Enrico Boasso E-mail address: enrico [email protected]
1710.05286	2	1710	2017-10-27T07:43:29	Generalization of The Results on Fixed Point For Coupling on Metric Spaces	[ "math.FA" ]	The purpose of this paper is to introduce the concept of self-cyclic maps, g-coupling and Banach type g-coupling which is the generalization of couplings introduced by Choudhury et al. In our main result we prove the existence theorem of coupled coincidence point for Banach type g-couplings which extends the results by Choudhury et al. We give an example in support of our result.	math.FA	math	Generalization of The Results on Fixed Point For Couplings on Metric Spaces Tawseef Rashid 1 and Q. H. Khan 2 The purpose of this paper is to introduce the concept of self-cyclic maps, g-coupling, Banach type g-coupling which is the generalization of couplings in- troduced by Choudhury et al.[5]. In our main result we prove the existence theorem of coupled coincidence point for Banach type g-couplings which extends the results by choudhury et al. [5]. We give an example in support of our result. Keywords : g-coupling, coupled coincidence point, strong coupled coinci- dence point, self-cyclic maps, Banach type g-coupling. 2010MSC: 47H10, 54H25. 1. Introduction and Preliminaries The concept of coupled fixed point was introduced in the work of Guo et al. [10]. Then after the coupled contraction mapping theorem by Bhaskar and Lak- shmikantham [9], and the introduction of coupled coincidence point by Laksh- mikantham.V and ´Ciri´c.L [13], the coupled fixed and coupled coincidence point results reported in papers [2, 3, 4, 6, 7, 11, 13, 14, 15, 16, 17, 18, 19, 20]. Kirk et al. [12] gave the concept of cyclic mappping. Recently Choudhury et al.[5] introduced the concept of couplings which are actually coupled cyclic mappings with respect to two given subsets of a metric space. Choudhury et al. [5] proved the existence of strong coupled fixed point for a coupling ( w.r.t. subsets of a complete metric space). H. Aydi et al. [1] proved the theorem for existence and uniqueness of strong coupled fixed point in partial metric spaces. In this paper we extend the concept of Banach type coupling to Banach type g-coupling by extending the concept of coupling to g-coupling. We have intro- duced self-cyclic maps and generalize the results by Choudhury et al [5]. Now we recall some definitions. Definition 1.1 (Coupled fixed point) [9]. An element (x, y) ∈ X ×X, where X is any non-empty set, is called a coupled fixed point of the mapping F : X×X → X if F (x, y) = x and F (y, x) = y. Definition 1.2 (Strong coupled fixed point) [8]. An element (x, y) ∈ X × X, where X is any non-empty set, is called a strong coupled fixed point of the mapping F : X × X → X if (x, y) is coupled fixed point and x = y; that is if F (x, x) = x. 1Department of Mathematics, Aligarh Muslim University, Aligarh-202002, India. Email : [email protected] 2Department of Mathematics, Aligarh Muslim University, Aligarh-202002, India. Email : [email protected] 1 Definition 1.3 (Coupled Banach Contraction Mapping) [9]. Let (X, d) be a metric space. A mapping F : X × X → X is called coupled Banach contraction if there exists k ∈ (0, 1) s.t ∀ (x, y), (u, v) ∈ X × X, the following inequality is satisfied: d(F (x, y), F (u, v)) ≤ k 2 [d(x, u) + d(y, v)]. Definition 1.4 (Cyclic mapping) [12]. Let A and B be two non-empty subsets of a given set X. Any function f : X → X is said to be cyclic (with respect to A and B) if f (A) ⊂ B and f (B) ⊂ A. Definition 1.5 (Coupling) [8]. Let (X, d) be a metric space and A and B be two non-empty subsets of X. Then a function F : X × X → X is said to be a coupling with respect to A and B if F (x, y) ∈ B and F (y, x) ∈ A whenever x ∈ A and y ∈ B. Definition 1.6 (Banach type coupling) [5]. Let A and B be two non-empty subsets of a complete metric space (X, d). A coupling F : X × X → X is called a Banach type coupling with respect to A and B if it satisfies the following inequality: d(F (x, y), F (u, v)) ≤ k 2 [d(x, u) + d(y, v)]. where x, v ∈ A, y, u ∈ B and k ∈ (0, 1). Definition 1.7 (Coupled coincidence point of F and g) [13]. An element (x, y) ∈ X × X is called a coupled coincidence point of the mappings F : X × X → X and g : X → X if F (x, y) = g(x) and F (y, x) = g(y). Definition 1.8 (Commutative mappings) [13]. For any set X, we say mappings F : X × X → X and g : X → X are commutative if g(F (x, y)) = F (g(x), g(y)), ∀ x, y ∈ X. 2. Main Result Here we give some definitions in support of our main result. Definition 2.1 (Self-cyclic mapping). Let A and B be two non-empty subsets of any set X. Then a mapping g : X → X is said to be self-cyclic (with respect to A and B) if g(A) ⊆ A and g(B) ⊆ B. Definition 2.2 (Strong coupled coincidence points of F and g). A coupled coincidence point (x, y) ∈ X × X of F : X × X → X and g : X → X is said to be strong coupled coincidence point of F and g, if x = y i.e. F (x, x) = g(x). 2 Definition 2.3 (g-coupling). Let (X, d) be a metric space and A and B be two non-empty subsets of X. Let functions F and g are such that F : X × X → X and g : X → X. Then F is said to be g-coupling (with respect to A and B) if F (x, y) ∈ g(A) ∩ B and F (y, x) ∈ g(B) ∩ A, whenever x ∈ A and y ∈ B. Remark 2.4: It should be noted that every g-coupling is a coupling but converse is not true in general. Proof : Let F : X × X → X be g-coupling, where g : X → X. Then by Definition (2.3), we have F (x, y) ∈ (g(A) ∩ B) ⊂ B and F (y, x) ∈ (g(B) ∩ A) ⊂ A. this shows that F is a coupling (w.r.t. A and B). Clearly converse is not true in general. If F is a coupling (w.r.t A and B), it doesn't imply that F is a g-coupling for any g : X → X (w.r.t. A and B) as F (x, y) ∈ B doesn't imply that F (x, y) ∈ g(A), and similarly for F (y, x). Definition 2.5 (Banach type g-coupling). Let A and B be two non-empty subsets of a complete metric space (X, d). Then a g-coupling F : X × X → X is said to be Banach type g-coupling (with respect to A and B) if the following inequality holds: d(F (x, y), F (u, v)) ≤ k 2 [d(gx, gu) + d(gy, gv)]. y, u ∈ B and k ∈ (0, 1). where g :X → X is a self-cyclic whenever x, v ∈ A, mapping (with respect to A and B). Note : If g = I (the identity mapping) which is also self-cyclic, then Banach type g-coupling becomes Banach type coupling. Theorem 2.6: Let A and B be any two subsets of a complete metric space (X, d). If there exists a Banach type g-coupling F : X × X → X (with respect to A andB), where g : X → X is self-cyclic (with respect to A and B). If g(A) and g(B) are closed subsets of (X, d). Then (i) g(A) ∩ g(B) 6= ∅, (ii) F and g have a coupled coincidence point in A × B, i.e. there exists (a, b) ∈ A × B, s.t F (a, b) = g(a) and F (b, a) = g(b). Proof : Here F : X × X → X is given to be Banach type g-coupling, i.e. (1) d(F (x, y), F (u, v)) ≤ k 2 where x, v ∈ A, y, u ∈ B and k ∈ (0, 1). Also as g : X → X is self-cyclic (w.r.t. A and B), so [d(gx, gu) + d(gy, gv)]. gx, gv ∈ g(A) ⊆ A and gy, gu ∈ g(B) ⊆ B Let x0 ∈ A and y0 ∈ B, then by defnition of g-coupling, we have F (x0, y0) ∈ g(A) ∩ B and F (y0, x0) ∈ g(B) ∩ A in particular F (x0, y0) ∈ g(A) and F (y0, x0) ∈ g(B). If F (x0, y0) = g(x0) and F (y0, x0) = g(y0), 3 then (x0, y0) is the coupled coincidence point of F and g, so we are done in this case. Otherwise ∃ x1 ∈ A and y1 ∈ B , s.t. F (x0, y0) = g(x1) and F (y0, x0) = g(y1). Now if, F (x1, y1) = g(x1) and F (y1, x1) = g(y1), then (x1, y1) is a coupled coincidence point of F and g, and we are through otherwise ∃ x2 ∈ A and y2 ∈ B s.t F (x1, y1) = g(x2) and F (y1, x1) = g(y2). continuing in this way, we get sequences {gxn} and {gyn} in g(A) and g(B) respectively, such that g(xn+1) = F (xn, yn) and g(yn+1) = F (yn, xn). (2) Now using (1) and (2), we get and d(gx1, gy2) = d(F (x0, y0), F (y1, x1)) ≤ k 2 [d(gx0, gy1) + d(gy0, gx1)] d(gy1, gx2) = d(F (y0, x0), F (x1, y1)) ≤ k 2 [d(gy0, gx1) + d(gx0, gy1)]. from above two inequalities, we have d(gx1, gy2) + d(gy1, gx2) ≤ k 2 [d(gx0, gy1) + d(gy0, gx1) + d(gy0, gx1) + d(gx0, gy1)] = k[d(gx0, gy1) + d(gy0, gx1)]. or, d(gx1, gy2) + d(gy1, gx2) 2 ≤ k 2 [d(gx0, gy1) + d(gy0, gx1)]. (3) using (1), (2) and (3), we have d(gx2, gy3) = d(F (x1, y1), F (y2, x2)) ≤ ≤ k 2 k2 2 [d(gx1, gy2) + d(gy1, gx2)] [d(gx0, gy1) + d(gy0, gx1)], 4 and d(gy2, gx3) = d(F (y1, x1), F (x2, y2)) ≤ ≤ k 2 k2 2 [d(gy1, gx2) + d(gx1, gy2)] [d(gx0, gy1) + d(gy0, gx1)]. Let for some integer n, d(gxn, gyn+1) ≤ kn 2 [d(gx0, gy1) + d(gy0, gx1)]. kn 2 Now using (1), (4) and (5), we have d(gyn, gxn+1) ≤ [d(gx0, gy1) + d(gy0, gx1)]. d(gxn+1, gyn+2) = d(F (xn, yn), F (yn+1, xn+1)) [d(gxn, gyn+1) + d(gyn, gxn+1)] kn[d(gx0, gy1) + d(gy0, gx1)] ≤ ≤ = k 2 k 2 kn+1 2 [d(gx0, gy1) + d(gy0, gx1)]. similarly as above we can show that d(gyn+1, gxn+2) ≤ kn+1 2 [d(gx0, gy1) + d(gy0, gx1)]. Thus (4) and (5) remains also true for n + 1, Hence by principle of mathematical induction, we have ∀ n ≥ 1, d(gxn, gyn+1) ≤ d(gyn, gxn+1) ≤ kn 2 kn 2 [d(gx0, gy1) + d(gy0, gx1)]. [d(gx0, gy1) + d(gy0, gx1)]. Again by (1) and (2), we have d(gx1, gy1) = d(F (x0, y0), F (y0, x0)) ≤ k 2 [d(gx0, gy0) + d(gy0, gx0)] = kd(gx0, gy0). (4) (5) (6) (7) that is, d(gx1, gy1) ≤ kd(gx0, gy0). (8) 5 Then from (1), (2) and (8), we have d(gx2, gy2) = d(F (x1, y1), F (y1, x1)) ≤ k 2 [d(gx1, gy1) + d(gy1, gx1)] = kd(gx1, gy1) ≤ k2d(gx0, gy0). Let for some integer n, we have d(gxn, gyn) ≤ knd(gx0, gy0). (9) Then from (1), (2) and (9), we get d(gxn+1, gyn+1) = d(F (xn, yn), F (yn, xn)) ≤ k 2 [d(gxn, gyn) + d(gyn, gxn)] = kd(gxn, gyn) ≤ kn+1d(gx0, gy0). This shows that (9) remains also true for n + 1, thus by principle of mathemtical induction, we say that d(gxn, gyn) ≤ knd(gx0, gy0), ∀ n ≥ 1. (10) Now by (6), (7) and (10) and by triangular inequality, we have ∀ n ≥ 1 d(gxn, gxn+1) + d(gyn, gyn+1) ≤ d(gxn, gyn) + d(gyn, gxn+1) + d(gyn, gxn) + d(gxn, gyn+1) = 2d(gxn, gyn) + [d(gyn, gxn+1) + d(gxn, gyn+1)] ≤ 2knd(gx0, gy0) + kn[d(gx0, gy1) + d(gy0, gx1)]. since k ∈ (0, 1), it follows that P d(gxn, gxn+1) + P d(gyn, gyn+1) < ∞. Thus sequences {gxn} and {gyn} are Cauchy sequences in g(A) and g(B). As g(A) and g(B) are closed subsets of complete metric space (X, d), so {g(xn)} and {g(yn)} are convergent in g(A) and g(B) respectively. Therefore ∃ u ∈ g(A) and v ∈ g(B), s.t. gxn → u and gyn → v as n → ∞. (11) using (10), as k ∈ (0, 1), we have d(gxn, gyn) → 0 as n → ∞. therefore from (11), we have u = v. (12) As u ∈ g(A) and v ∈ g(B) ⇒ u ∈ g(A) ∩ g(B) This proves part (i) that g(A) ∩ g(B) 6= ∅. 6 Now, since u ∈ g(A) and v ∈ g(B), therefore ∃ a ∈ A and b ∈ B, s.t u = g(a) and v = g(b). then from (11) and (12), we have and gxn → g(a) and gyn → g(b). g(a) = g(b). (13) (14) Now by (1), (2), (13), (14) and triangular inequality, we have d(g(a), F (a, b)) ≤ d(g(a), gyn+1) + d(gyn+1, F (a, b)) = d(g(a), gyn+1) + d(F (yn, xn), F (a, b)) ≤ d(g(a), gyn+1) + [d(gyn, g(a)) + d(gxn, g(b))] k 2 n → ∞. k 2 n → ∞. → 0 as thus, we have F (a, b) = g(a). (15) Again from(1), (2), (13), (14) and traingular inequality, we have d(g(b), F (b, a)) ≤ d(g(b), gxn+1) + d(gxn+1, F (b, a)) = d(g(b), gxn+1) + d(F (xn, yn), F (b, a)) ≤ d(g(b), gxn+1) + [d(gxn, g(b)) + d(gyn, g(a))] → 0 as thus, we get F (b, a) = g(b). (16) Hence from (15) and (16) we get, F (a, b) = g(a) and F (b, a) = g(b), where a ∈ A and b ∈ B. Thus (a, b) ∈ A × B is the coupled coincidence point of F and g. Remark 2.7. It is worth noting that the above theorem also gives the condition for the existence of symmetric point of F in A × B, i.e. ∃ (a, b) ∈ A × B s.t F (a, b) = F (b, a). As from (14) g(a) = g(b) so, F (a, b) = F (b, a). Existence and Uniqueness of Strong coupled coincidence point of F and g. Theorem 2.8: If in addition to above condition in Theorem 2.6, g is one-one, then (i) A ∩ B 6= ∅, and (ii) F and g have unique strong coupled coincidence point in A ∩ B. Proof: since g is given to be one-one then from (14) of Theorem 2.6, we get a = b. 7 since, a ∈ A and b ∈ B ⇒ a = b ∈ A ∩ B. Hence A ∩ B 6= ∅, this proves our part (i). Also from (15) of Theorem 2.6, we have F (a, a) = g(a). which shows that F and g have strong coupled coincidence point in A ∩ B. Uniqueness: Let us suppose if possible there exists two strong coupled coinci- dence points l, m ∈ A ∩ B of F and g then by definition we have, F (l, l) = g(l) and F (m, m) = g(m). (17) then from (1) of Theorem 2.6, we have d(g(l), g(m)) = d(F (l, l), F (m, m)) ≤ k 2 [d(g(l), g(m)) + d(g(l), g(m))] = k[d(g(l), g(m))]. Which is a contradiction as k ∈ (0, 1) and is only possible if d(g(l), g(m)) = 0, i.e. g(l) = g(m) or l = m because g is one-one. Hence F and g have a unique strong coupled coincidence point in A ∩ B. Corrollary 2.9 : If F : X × X → X and g : X → X are commutative and g is one-one. if (gx, gy) is a coupled coincidence point of F and g, where x, y ∈ X, then (x, y) is also a coupled coincidence point of F and g. Proof : As (gx, gy) is a coupled coincidence point of F and g, we have F (gx, gy) = gx, F (gy, gx) = gy. (18) (19) Now by (18) and commutavity of F and g, we have g(F (x, y)) = F (gx, gy) = gx. as g is one-one, we have from above Now again by (19) and commutavity of F and g, we have F (x, y) = x. (20) g(F (y, x)) = F (gy, gx) = gy. as g is one-one, we have from above F (y, x) = y. (21) 8 Thus from (20) and (21), we get F (x, y) = x and F (y, x) = y. 10 , where x, y ∈ X. i.e. (x, y) is the coupled coincidence point of F and g. The following example illustrates our results. Example 2.10. Let X = R with the metric defined as d(x, y) = x − y , where x, y ∈ X. Let A = [0, 2] and B = [0, 3]. Let F be defined as F (x, y) = x+y and let g : X → X is defined by g(x) = x 2 . Then g(A) = [0, 1] and g(B) = [0, 3 X. Also g(A) ⊆ A and g(B) ⊆ B, so g is self-cyclic. Now we show that F is g-coupling. As g(A) ∩ B = [0, 1] and g(B) ∩ A = [0, 1], so ∀ x ∈ A and y ∈ B, we have 0 ≤ F (x, y) ≤ 1 i.e. F (x, y) ∈ g(A) ∩ B and F (y, x) ∈ g(B) ∩ A, which shows that F is g-coupling with respect to A and B. Again for x, v ∈ A and y, u ∈ B, we have 2 ], so g(A) and g(B) are closed subsets of 2 and 0 ≤ F (y, x) ≤ 1 2 . d(F (x, y), F (u, v)) = = ≤ = = 1 5 1 5 1 5 k 2 [ [ x + y u + v − 10 x − u 10 y − v + 2 x − u 2 − u 2 x 2 2 y − v + + ] v 2 ] 2 − y 2 [d(gx, gu) + d(gy, gv)], where k = 2 5 ∈ (0, 1). which shows that F is a Banach type g-coupling (w.r.t. A and B). Thus all the conditions of Theorem (2.6) are satisfied, therefore ∃ (a, b) ∈ A×B s.t. F (a, b) = g(a) and F (b, a) = g(b). i.e. a + b = and a 2 b + a 10 = b 2 (22) or 10 a + b 5 = a = b. which is only possible with a = 0 and b = 0. thus (0, 0) is the unique strong g-coupled coincidence point of F and g. Note : The uniqueness of the strong g-coupled coincidence point in the above example is because of g is one-one. 9 Corrollary 2.11: Let g : X → X is a Banach contraction, Then every Banach type g-coupling (w.r.t. A and B) is a Banach type coupling (w.r.t. A and B), where A and B are subsets of X. Proof : since g is a contraction, therefore ∃ α ∈ (0, 1), s.t d(gx, gy) ≤ αd(x, y) ∀ x, y ∈ X. (23) Let F : X × X → X be a Banach type g-coupling (w.r.t. A and B), i.e. k 2 d(F (x, y), F (u, v)) ≤ [d(gx, gu)+d(gy, gu)], where x, v ∈ A, y, u ∈ B and k ∈ (0, 1). (24) Also F is a g-coupling (w.r.t. A and B) and hence by remark (2.4), F is a coupling (w.r.t. A and B). Now using (23) in (24), we get d(F (x, y), F (u, v)) ≤ ≤ k 2 αk 2 [αd(x, u) + αd(y, v)] [d(x, u) + d(y, v)]. as k ∈ (0, 1) and α ∈ (0, 1), theref ore kα = k1 ∈ (0, 1), so d(F (x, y), F (u, v)) ≤ k1 2 [d(x, u) + d(y, v)]. where x, v ∈ A, y, u ∈ B and k1 ∈ (0, 1). Hence F is a Banach type coupling (w.r.t. A and B). 3. Conclusion In this paper we introduce some new results that extend the concept of coupling and Banach type coupling introduced by Choudhury et al. [5] to g-coupling and Banach type g-coupling resp. and proved the existence and uniqueness theorem for coupled coincidence point and strong coupled coincidence point. References [1] H. Aydi, M. Barakat, A. Felhi, H. Isik, On φ-contraction type couplings in partial metric spaces, Journal of Mathematical Analysis, 8(2017), 78-89. [2] H. Aydi, Coupled fixed point results in ordered partial metric spaces, Sel¸cuk J. Appl. Math. 13(2012), 23-33. [3] V. Berinde, Generalized coupled fixed point theorems in partially ordered metric spaces and applications, Nonlin. Anal. 74(2011), 7347-7355. [4] N. Bilgili, I. M. Erham, E. Karapinar, D. Turkoglu, A note on coupled fixed point theorems for mixed g-monotone mappings in partially ordered metric spaces, Fixed Point Theory Appl. 2014(2014):120, 6 pages. 10 [5] B.S. Choudhury, P. Maity, P. Konar, Fixed point results for couplings on metric spaces, U. P. B. Sci. Bull., series A, 79(1)(2017). 1-12. [6] B.S. Choudhury, A. Kundu, A coupled coincidence result in partially or- dered metric spaces for compatible mappings, Nonlinear Anal. 73(2010), 2524-2531. [7] B.S. Choudhury, A. Kundu, Two coupled weak contraction theorems in partially ordered metric spaces,Revista de la Real Academia de ciencias Exactas, Fisicas y Naturales. Series A.Mathematicas.108(2014), 335-351. [8] B.S. Choudhury, P. Maity, Cyclic coupled fixed point result using Kannan type contractions, Journal Of Operators. 2014(2014), Article ID 876749, 5 pages. [9] T. Gnana Bhaskar, V. Lakshmikantham, Fixed point theorems in partially ordered metric spaces and applications, Nonlin. Anal. 65(2006), 1379-1393. [10] D. Guo, V. Lakshmikantham, Coupled fixed points of non-linear operators with applications, Nonlin. Anal. 11(1987), 623-632. [11] Harjani. J, Lop´ez. B, Sadarangani. K, Fixed point theorems for mixed monotone operators and applications to integral equations, Nonlinear Anal. 74(2011), 1749-1760. [12] W.A. Kirk, P.S. Srinivasan and P. Veeramani, Fixed points for mappings satisfying cyclical contractive conditions, Fixed Point Theory. 4(2003), 78- 89. [13] Lakshmikantham. V, L. ´Ciri´c, Coupled fixed point theorems for nonlinear contractions in partially ordered metric spaces, Nonlinear Anal. 70(2009), 4341-4349. [14] Luong. NV, Thuan. NX, Coupled fixed point theorems in partially ordered metric spacesdepended on another function, Bull. Math. Anal. 3(2011), 129- 140. [15] N. V. Luong and N. X. Thuan, Coupled points in ordered generalized metric spaces and application to integro-dierential equations, Analele Universitatii "Ovidius" Constanta Seria Matematica, vol. 21, no. 3, pp. 155-180, 2013. [16] Mujahid Abbas, Bashir Ali, Yusuf I.Suleiman, Generalized coupled common fixed point results in partially ordered A-metric spaces, Fixed Point Theory and Application, (2015) 2015:64, 24 pages. [17] Rosouli. SH, Bahrampour. M, A remark on the coupled fixed point theorems for mixed monotone operators in partially ordered metric spaces, J. Math. Comput. Sci. 3(2011), 246-261. 11 [18] W. Shatanawi, H.K. Nashine, N. Tahat, Generalization of some coupled fixed point results on partial metric spaces, International Journal of Math- ematics and Mathematical Sciences, 2012(2012), Article ID 686801, 10 pages. [19] W. Shatanawi, B. Samet, M. Abbas, Coupled fixed point theorems for mixed monotone mappings in ordered partial metric spaces, Mathematical and Computer Modelling, 55(2012), 680-687. [20] Wasfi Shatanawi, Mujahid Abbas, Hassen Aydi, Nedal Tahat, Common coupled coincidence and coupled fixed points in G-metric spaces, Nonlinear Anal. and Application, 2012(2012), Article ID jnaa-00162, 16 pages. 12
1505.01945	1	1505	2015-05-08T07:28:30	Composition operators on vector-valued analytic function spaces: a survey	[ "math.FA" ]	We survey recent results about composition operators induced by analytic self-maps of the unit disk in the complex plane on various Banach spaces of analytic functions taking values in infinite-dimensional Banach spaces. We mostly concentrate on the research line into qualitative properties such as weak compactness, initiated by Liu, Saksman and Tylli (1998), and continued in several other papers. We discuss composition operators on strong, respectively weak, spaces of vector-valued analytic functions, as well as between weak and strong spaces. As concrete examples, we review more carefully and present some new observations in the cases of vector-valued Hardy and BMOA spaces, though the study of composition operators has been extended to a wide range of spaces of vector-valued analytic functions, including spaces defined on other domains. Several open problems are stated.	math.FA	math	COMPOSITION OPERATORS ON VECTOR-VALUED ANALYTIC FUNCTION SPACES: A SURVEY JUSSI LAITILA AND HANS-OLAV TYLLI Abstract. We survey recent results about composition operators in- duced by analytic self-maps of the unit disk in the complex plane on various Banach spaces of analytic functions taking values in infinite- dimensional Banach spaces. We mostly concentrate on the research line into qualitative properties such as weak compactness, initiated by Liu, Saksman and Tylli (1998), and continued in several other papers. We discuss composition operators on strong, respectively weak, spaces of vector-valued analytic functions, as well as between weak and strong spaces. As concrete examples, we review more carefully and present some new observations in the cases of vector-valued Hardy and BM OA spaces, though the study of composition operators has been extended to a wide range of spaces of vector-valued analytic functions, including spaces defined on other domains. Several open problems are stated. 1. Introduction Let D = {z ∈ C : z < 1} be the open unit disk in the complex plane C, and let ϕ : D → D be a fixed analytic map. The classical study of the analytic composition operators Cϕ, where f 7→ Cϕ(f ) = f ◦ ϕ, originates from the work of Ryff (1966) and Nordgren (1968). For instance, they observed that any Cϕ defines a bounded operator H p → H p as a conse- quence of the Littlewood subordination principle. Recall that for 1 ≤ p < ∞, the analytic function f : D → C belongs to the Hardy space H p if kf kp H p = sup 0≤r<1ZT f (rξ)pdm(ξ) < ∞ where T = ∂D = [0, 2π] and dm(eit) = dt/2π. The space H ∞ consists of the bounded analytic functions. Subsequently an extensive literature has emerged, where a very wide variety of properties of analytic composition op- erators has been addressed on a large number of spaces of analytic functions. Date: 20.3.2014. 2010 Mathematics Subject Classification. 47B33, 46E15, 46E40, 47B07. Key words and phrases. Analytic functions, Banach space, compactness, composition operator, weak compactness, vector-valued function. 1 2 JUSSI LAITILA AND HANS-OLAV TYLLI We refer to [47] and [11] for comprehensive accounts of the theory until ca. 1995. This survey reviews more recent results about composition operators on various Banach spaces of vector-valued analytic functions including the vec- tor-valued Hardy space H p(X), where X is a complex Banach space. Let f : D → X be a vector-valued analytic function and let 1 ≤ p < ∞. Then f ∈ H p(X), if kf kp H p(X) = sup 0≤r<1ZT kf (rξ)kp X dm(ξ) < ∞. Moreover, f ∈ H ∞(X), if kf kH∞ = supz∈D kf (z)kX < ∞. In this notation H p = H p(C). Above the analyticity of f : D → X means that the scalar- valued function x∗◦f is analytic D → C for any functional x∗ ∈ X ∗ (that is, f is weakly analytic). This is equivalent to the requirement that the X-valued derivative f ′(z) exists for all points z ∈ D (that is, f is strongly analytic). For the basics of vector-valued analytic functions, see for example [20]. Qualitative properties of the vector-valued composition operators f 7→ f ◦ϕ on H p(X) and certain other spaces were first systematically studied by Liu, Saksman and Tylli [35]. Independently Hornor and Jamison [21] considered the operators f 7→ f ◦ ϕ on H p(X) with different aims, and Sharma and Bhanu [49] looked at some of their basic operator properties on H 2(X), where X is a Hilbert space. We mostly concentrate on qualitative properties, such as weak compact- ness, of composition operators on several Banach spaces of vector-valued analytic functions of both strong and weak type defined on D. Weak type spaces were introduced into this context by Bonet, Domanski and Lindström [4], and in this case the techniques differ from those of the strong type spaces. In section 2 we introduce a general framework for vector-valued composition operators in order to provide a convenient general perspective into the study, and we review results that illustrate both similarities and differences com- pared to the scalar-valued case X = C. We also highlight new phenomena that do not have any counterparts for scalar composition operators. For in- stance, composition operators can be studied between a weak and a strong space. In the final section we briefly discuss attempts to generalize the larger class of weighted composition operators to the vector-valued setting. Some vector-valued arguments are sketched, but we mostly assume that the basic scalar theory is known from [47] and [11]. Composition operators of different nature occur in various other settings. For instance, there is a well-developed theory of the composition operators S 7→ A ◦ S ◦ B, where A and B are fixed bounded operators, on spaces of linear operators, see e.g. the survey [44]. Properties of such composition operators will actually be required in section 5 below. SURVEY OF COMPOSITION OPERATORS 3 2. a general framework We first introduce a flexible general framework for the study of qualita- tive properties of vector-valued composition operators, which will facilitate a discussion of some common features. Suppose that A is a Banach space of analytic functions D → C and let A(X) be an associated vector-valued Banach space of analytic functions D → X, where X is a complex Banach space. Assume that the following properties hold for the pair (A, A(X)) for all Banach spaces X: (AF1) The constant maps f (z) ≡ c belong to A for all c ∈ C. (AF2) f 7→ f ⊗ x defines a bounded linear operator Jx : A → A(X) for any x ∈ X, where (f ⊗ x)(z) = f (z)x for z ∈ D. (AF3) g 7→ x∗ ◦ g defines a bounded linear operator Qx∗ : A(X) → A for any x∗ ∈ X ∗. (AF4) The point evaluations δz, where δz(f ) = f (z) for f ∈ A(X), are bounded A(X) → X for all z ∈ D. It follows from (AF1) and (AF2) that the vector-valued constant maps z 7→ fx(z) ≡ x, that is fx = 1 ⊗ x, belong to A(X) for all x ∈ X. It is easy to check that (AF1) -- (AF4) are satisfied for the pair (H p, H p(X)) for any X. Note that for Banach spaces of analytic functions defined on other domains, such as a half-plane or the plane C, condition (AF1) may not be relevant and the above framework cannot be applied in this form. Suppose that A and B are Banach spaces of analytic functions D → C so that (A, A(X)) and (B, B(X)) satisfy (AF1) -- (AF4), where A(X) and B(X) are X-valued Banach spaces of analytic functions on D associated with A, respectively B. Let ϕ : D → D be a given analytic self-map, and suppose In order to distinguish between composition operators acting on different spaces we will in the sequel use Cϕ : A → B for the composition operator f 7→ f ◦ ϕ in the scalar-valued setting, that is, in that the vector-valued composition operator fCϕ is bounded A(X) → B(X), where f 7→ fCϕ(f ) = f ◦ ϕ. the case X = C, and fCϕ for its vector-valued version A(X) → B(X). The following general formulation is partly motivated by [4, Prop. 1]. Proposition 1. The following factorizations hold. (F1) Let x ∈ X, x∗ ∈ X ∗ be norm-1 vectors so that hx∗, xi = 1. Then A(X) Jx A fCϕ Cϕ / B(X) Qx∗ / B commutes. / O O / 4 JUSSI LAITILA AND HANS-OLAV TYLLI (F2) Let j(x) = fx for x ∈ X, where fx(z) ≡ x for all z ∈ D. Then A(X) j X fCϕ IX / B(X) δ0 / X commutes, where IX is the identity operator on X. The above factorizations place some inherent restrictions on possible qual- Proof. Note towards (F1) that x∗(fCϕ(f ⊗ x)) = x∗((f ◦ ϕ) ⊗ x) = Cϕ(f ) for f ∈ A, while δ0(fCϕ(fx)) = δ0(fx) = x for x ∈ X. itative properties of the vector-valued operators fCϕ. Roughly speaking, part (3) below states that fCϕ : A(X) → B(X) cannot have any qualitative prop- properties of X also influence (qualitative) properties of fCϕ. erties inherited under composition of linear operators that are not shared by Cϕ : A → B and the identity operator IX : X → X. Thus Banach space Corollary 2. Let X be a complex Banach space. (cid:3) X is finite-dimensional. In particular, if X is infinite-dimensional, (1) If fCϕ is bounded A(X) → B(X), then Cϕ is bounded A → B. (2) If fCϕ : A(X) → B(X) is compact, then Cϕ is compact A → B and (3) Let I be an operator ideal in the sense of Pietsch [43]. If fCϕ : A(X) → In fact, by (F1) and (F2) the compactness of fCϕ : A(X) → B(X) implies then fCϕ is never compact A(X) → B(X). that both Cϕ : A → B and IX are compact, that is, X is finite-dimensional. Part (3) is verified in a similar fashion. For a converse of (2), see Proposition 7. B(X) belongs to I, then IX as well as Cϕ : A → B belong to I. 3. Weak compactness on H 1(X) and other vector-valued spaces Let ϕ : D → D be any analytic map. It was observed independently in [35] and [21] that fCϕ is bounded on H p(X), while [49] contains the case H 2(X), where X is a Hilbert space. Boundedness can be verified in the following manner by a small modification of an argument for scalar H p spaces. Note first that z 7→ kf (z)kX is a subharmonic map on D for any analytic function f : D → X, since kf (z)kX = sup kx∗k≤1 hx∗, f (z)i, z ∈ D. Consequently, if ϕ(0) = 0, then the Littlewood inequality [11, Thm. 2.22] yields that kfCϕ(f )kH p(X) ≤ kf kp H p(X) for f ∈ H p(X). / O O / SURVEY OF COMPOSITION OPERATORS 5 For the general case let σa : D → D be the Möbius transformation defined by σa(z) = a−z 1−az for a ∈ D. If ϕ(0) 6= 0 , let ψ = σϕ(0) ◦ ϕ, so that ψ(0) = 0 and Cψ is a contraction H p(X) → H p(X). Since σ−1 ϕ(0) = σϕ(0), we get that Cϕ = Cψ ◦ Cσϕ(0) is bounded on H p(X) once we have checked that Cφ is bounded on H p(X) for any Möbius transformation φ. This can be verified by the change of variables w = φ(z) inside the integral ZT kf (rφ(ξ))kp X dm(ξ) defining kCφ(f (r·))kH p(X) for 0 < r < 1 and letting r → 1. The minor point of difference between the above scalar- and vector-valued arguments for the Hardy spaces relates to the potential absence of radial In fact, it is well known that any f ∈ H p has a.e. radial limits limits. f (ξ) = limr→1− f (rξ) on T, but this is not always true for functions in H p(X): the bounded analytic function f : D → c0, where f (z) = (zn), z ∈ D, does not have radial limits anywhere on T. In fact, the existence of a.e. radial limits for any f ∈ H p(X), and any fixed 1 ≤ p ≤ ∞, characterizes the analytic Radon-Nikodým property (ARNP) of the complex Banach space X. The ARNP and the above result by Bukhvalov and Danilevich (1982) is not needed here, but the reader may keep in mind that e.g. every reflexive Banach space has the ARNP. See, e.g., [28, p. 723] for references and a discussion of the ARNP. Let X be an infinite-dimensional Banach space. According to Corollary 2.(2) there are no compact compositions fCϕ : H p(X) → H p(X). This raises the general question of which are the relevant qualitative properties for com- position operators on vector-valued spaces such as H p(X). In [35] the au- thors considered weak compactness, and related properties, for which there are satisfactory results. Let X and Y be Banach spaces. Recall that the bounded linear operator U : X → Y is weakly compact if there is a weakly convergent subsequence (U xnk ) for any bounded sequence (xn) ⊂ X. If X and Y are non-reflexive spaces, then weakly compact U : X → Y are relatively small operators. A fundamental result of Shapiro [46] (see also [47] and [11]) says that for 1 ≤ p < ∞ the composition operator Cϕ is compact H p → H p if and only if lim w→1 N (ϕ, w) log(1/w) = 0. (1) Above N (ϕ, w) =Pz∈ϕ−1(w) log(1/z), where w ∈ D \ {ϕ(0)}, is the Nevan- linna counting function of ϕ. Several other equivalent criteria for the com- pactness of Cϕ : H p → H p are known in the literature, but (1) suffices for our purposes. Littlewood's inequality implies that N (ϕ, w) ≤ C · log(1/w) 6 JUSSI LAITILA AND HANS-OLAV TYLLI as w → 1 for some constant C = C(ϕ) for any analytic map ϕ : D → D, see [47, 10.4]. Shapiro's condition (1) is interpreted as a little-oh condition describing the rate of decrease of the affinity of ϕ for the values w as w → 1. H 1(X) and the compactness of Cϕ on H 1. Note that Corollary 2.(3) implies There is a precise connection between the weak compactness of fCϕ on that X is reflexive, that is, IX is weakly compact, whenever fCϕ is weakly compact on H p(X). Hence only p = 1 or p = ∞ are interesting for weak compactness, since H p(X) is itself reflexive if 1 < p < ∞ and X is reflexive, because H p(X) is then a closed subspace of the reflexive space Lp(T, X). The vector-valued part of the following result comes from [35]. Theorem 3. Let X be a complex reflexive Banach space, and ϕ : D → D be an analytic map. Then the following conditions are equivalent. (1) fCϕ : H 1(X) → H 1(X) is weakly compact (2) Cϕ : H 1 → H 1 is weakly compact (3) Cϕ : H 1 → H 1 is compact (4) Shapiro's condition (1) holds. Proof. The implication (1) ⇒ (2) follows from Corollary 2.(3). Sarason [45] proved that the weak compactness of Cϕ : H 1 → H 1 actually yields the com- pactness of Cϕ : H 1 → H 1, in other words that (2) ⇒ (3). The equivalence of (3) and (4) is contained in Shapiro's theorem. There remains to show condition. We outline the proof of the implication (4) ⇒ (1). The argument is based that fCϕ is weakly compact H 1(X) → H 1(X) whenever ϕ satisfies Shapiro's on a Littlewood-Paley type formula for kfCϕ(f )kH 1(X) derived from a for- mula of Stanton for continuous subharmonic maps. His formula [51, Thm. 2] implies that kf ◦ ϕkH 1(X) = kf (0)kX + 1 2π Z N (ϕ, w)d[∆(kf kX )](w) (2) for f ∈ H 1(X), where d[∆(kf kX )] denotes the distributional Laplacian as- sociated to the subharmonic map z 7→ kf (z)kX on D. Define de la Vallee-Poussin operators Vn for any n ∈ N by Vnf (z) = nXk=0 fkzk + 2n−1Xk=n+1 2n − k n fkzk k=0 P∞ for analytic functions f : D → X having the Fourier expansion f (z) = fkzk. Then (Vn) is a uniformly bounded sequence of operators on H 1(X) and Vn : H 1(X) → H 1(X) are weakly compact for any n if X is a re- flexive Banach space (in fact, Vn factors through a finite direct sum of copies of X). Moreover, given ε > 0 and 0 < r < 1 there is n0 = n0(ε, r) ∈ N so SURVEY OF COMPOSITION OPERATORS that kf (z) − Vn0f (z)kX ≤ ε · kf kH 1(X) holds for all z ≤ r and f ∈ H 1(X). 7 (3) If Shapiro's condition (1) holds and ε > 0 is arbitrary, then there is r ∈ (0, 1) such that N (ϕ, w) ≤ ε · log(1/w) for all w > r. Fix n0 as in (3) corresponding to ε and r. By applying (2) to f − Vn0f we get for f ∈ H 1(X) that 1 2π Z{r<z<1} N (ϕ, w)d[∆(kf − Vn0f kX)](w) kfCϕ(f ) − fCϕ(Vn0f )kH 1(X) = 1 2π Z{z≤r} + N (ϕ, w)d[∆(kf − Vn0f kX)](w) ≡ I1 + I2. The choice of r ∈ (0, 1) and (2) applied to ψ(z) = z give I1 ≤ ε 2π ZD log(1/w)d[∆(kf − Vn0f kX)](w) = εkf − Vn0f kH 1(X) ≤ C · εkf kH 1(X) for a uniform constant C. Moreover, it can be shown that I2 ≤ 4εkf kH 1(X) by using the estimates N (ϕ, w) ≤ log(1/w) and (3). For this it is convenient to introduce a cut-off function ψ ∈ C ∞ 0 (D) satisfying 0 ≤ ψ ≤ 1, ψ = 1 on {z : z ≤ r} and ψ = 0 on {z ∈ D : z ≥ (1 + r)/2}. We refer to [35, Prop. 2 and Thm. 3] for the complete technical details. Thus kfCϕ −fCϕVn0k ≤ C ′ · ε, where C ′ does not depend on ε, so that fCϕ is well approximated by the weakly compact operators fCϕVn0 for suitable n0. This means that fCϕ is weakly compact H 1(X) → H 1(X). Theorem 3 corresponds to the following template for many results about weak compactness, as well as other qualitative properties, of analytic com- position operators on vector-valued spaces. (cid:3) Proposition 4. Let A be a Banach space of analytic functions on D and A(X) a vector-valued version of A, such that (A, A(X)) satisfies (AF1) -- (AF4). Suppose that X is a complex reflexive Banach space and ϕ : D → D that the following conditions hold: (C1) if Cϕ is weakly compact A → A, then Cϕ is compact A → A, and is an analytic map, so that fCϕ is bounded A(X) → A(X). Assume moreover (C2) if Cϕ is compact A → A, then the vector-valued composition fCϕ is (C) fCϕ is weakly compact A(X) → A(X) ⇔ Cϕ is compact A → A. Then one has the characterization weakly compact A(X) → A(X). 8 JUSSI LAITILA AND HANS-OLAV TYLLI We stress that the above general scheme is only a guiding principle and in practice the techniques for establishing (C2) depend on A and its vector- valued extension A(X). Moreover, the criteria for the compactness of the operator Cϕ : A → A usually depend on A. It is straightforward to modify the scheme of Proposition 4 to apply to vector-valued compositions fCϕ : A(X) → B(X) between different spaces, where (A, A(X)) and (B, B(X)) satisfy the properties (AF1) -- (AF4). Condition (C1) is a problem of independent interest for composition op- erators A → A. Recently Lefevre, Li, Queffelec and Rodriguez-Piazza [31] constructed the first example of a Banach space A of complex-valued analytic functions on D, where (C1) fails for some symbol ϕ, see Example 10 below. We next look at cases where Proposition 4 apply. Let vα(z) = (1 − z2)α for z ∈ D and α > −1. The analytic function f : D → X belongs to the weighted Bergman space Aα p (X) if kf kp Aα p (X) =ZD kf (z)kp X vα(z)dA(z) < ∞, where dA is the area Lebesgue measure normalized by A(D) = 1 and 1 ≤ p < ∞. The classical Bergman space Ap(X) is obtained for α = 0. The following result was established in [4], but the special case A1(X) was already contained in [35]. Theorem 5. Let X be a complex reflexive Banach space, ϕ : D → D an analytic map and α > −1. Then the following conditions are equivalent. (1) fCϕ : Aα 1 (X) → Aα 1 → Aα (2) Cϕ : Aα (3) ϕ satisfies the condition 1 is compact 1 (X) is weakly compact lim sup w→1 Nα+2(ϕ, w) (log(1/w))α+2 = 0. Above Nβ(ϕ, ·) is the generalized Nevanlinna counting function defined for β > 0 by Nβ(ϕ, w) = Xz∈ϕ−1(w) (log(1/z)β , w ∈ D \ {ϕ(0)}, so that N (ϕ, ·) = N1(ϕ, ·). Actually, [4, Thm. 8] contains the estimate dist(fCϕ, W (Aα 1 (X))) ≤ C · lim sup w→1 Nα+2(ϕ, w) (log(1/w))α+2 , where W (Aα pact operators Aα 1 (X)) denotes the linear subspace consisting of the weakly com- 1 (X) → Aα 1 (X) and C is an absolute constant. We list some further Banach spaces A(X) for which the characterization (C) for weak compactness of composition operators are known to hold. We SURVEY OF COMPOSITION OPERATORS 9 emphasize that the arguments establishing (C1) and (C2) usually are specific for A(X), and the relevant compactness conditions for Cϕ depend on A. One verifies by inspection that these pairs (A, A(X)) satisfy (AF1) -- (AF4). • Let v : D → (0, ∞) be a bounded continuous weight function, and f : D → X be an analytic function. Recall that f ∈ H ∞ v (X) if kf kH∞ v (X) = sup z<1 v(z)kf (z)kX < ∞. v = H ∞ v (C). The case v ≡ 1 gives the classical spaces H ∞(X) Let H ∞ and H ∞ of bounded analytic functions. It was shown in [35] that (C) holds on H ∞(X) and this was extended to the case H ∞ v (X) by different means in [4]. • The vector-valued Bloch space B(X) [35]. Recall that f ∈ B(X) if (1 − z2)kf ′(z)kX < ∞. sup z∈D See [4] for an alternative approach via weak spaces (section 5 below). • The space CT (X) of vector-valued Cauchy transforms [28]. The ar- gument proceeds via composition operators on the vector-valued har- monic Hardy space h1(X). The compactness criterion for CT is due to Bourdon, Cima and Matheson. • Vector-valued BM OA(X)-spaces, see section 4. • Weak vector-valued versions of the above spaces, see section 5. A modified general scheme as in Proposition 4 also applies to other op- erator ideal properties, namely, just replace weak compactness by the rele- vant ideal property in (C1) and (C2). We state two results of this kind for H 1(X), respectively A1 α(X), from [35, Thm. 7] and [4, Cor. 9]. The opera- tor U : X → Y is called weakly conditionally compact if (U xn) has a weak Cauchy subsequence (U xnk ) for any bounded sequence (xn) ⊂ X. Recall that by Rosenthal's ℓ1-theorem, see [33, 2.e.5], IX is weakly conditionally compact if and only if X that does not contain any subspaces linearly isomorphic to ℓ1. By Proposition 1 this is the relevant class of spaces here. Theorem 6. Suppose that the Banach space X does not contain any sub- spaces linearly isomorphic to ℓ1, and ϕ : D → D is an analytic map. Let A = H 1 or A = A1 A(X) → A(X) if and only if Cϕ is compact A → A. α for α > −1. Then fCϕ is weakly conditionally compact We mention for completeness that the cases dim(X) < ∞ are similar to the scalar case. Proposition 7. Suppose that dim(X) < ∞, (A, A(X)) satisfies (AF1) -- (AF4), and ϕ : D → D is an analytic map. Then fCϕ is compact A(X) → A(X) if and only if Cϕ is compact A → A. 10 JUSSI LAITILA AND HANS-OLAV TYLLI Proof. Let n = dim(X) and fix a biorthogonal system {(xr, x∗ n} for X, so that x = Pn be written as f (z) =Pn Consequently, if Cϕ is compact A → A, then f 7→ fCϕ(f ) = Pn s) : 1 ≤ r, s ≤ k(x)xk for x ∈ X. Hence any f ∈ A(X) can k ◦ f ∈ A for k = 1, . . . , n. k=1 Cϕ(fk)xk is compact A(X) → A(X). For the converse note that Section 2 applies to this setting. (cid:3) k=1 fk(z)xk, where fk = x∗ k=1 x∗ Other results. Hornor and Jamison [21] characterized the isometrically equivalent compositions fCϕ and fCψ on H p(X), respectively Sp(X), for p 6= 2 and X a Hilbert space. Here f ∈ Sp(X) if the derivative f ′ ∈ H p(X). Sharma and Bhanu [49] studied e.g. normal and unitary compositions on H 2(X), where X is a Hilbert space. Bonet and Friz [5] characterized the weakly compact compositions on weighted vector-valued locally convex spaces of analytic functions on D. Composition operators on the Hilbert space- valued Fock space of entire functions on C were considered by Ueki [53]. See also [54] for results on the vector-valued Nevanlinna class. 4. Vector-valued BM OA-spaces In this section we discuss in more detail the case of composition operators on vector-valued BM OA spaces, since there are several natural vector-valued versions of BM OA, and condition (C1) is a problem of independent interest. Recall that the analytic function f : D → C belongs to BM OA, the space of analytic functions of bounded mean oscillation, if f ∗ = sup a∈D kf ◦ σa − f (a)kH 2 < ∞, where σa(z) = (a − z)/(1 − az) for z ∈ D. The Banach space BM OA is equipped with the norm kf kBM OA = f (0)+f ∗. BM OA is often considered as a Möbius-invariant version of H 2, but its Banach space structure is comp- licated, see e.g. [41]. Recall also that (H 1)∗ ≈ BM OA There are by now several equivalent characterizations of compact com- positions Cϕ : BM OA → BM OA, see [27] for a list. The following double criterion due to W. Smith [50] is the most relevant one for our purposes: Let ϕ : D → D be an analytic map. Then Cϕ is compact BM OA → BM OA if and only if lim ϕ(a)→1 0<w<1 sup w2N (σϕ(a) ◦ ϕ ◦ σa, w) = 0, lim t→1 sup {a : ϕ(a)≤R} m(cid:0){ζ ∈ ∂D : (ϕ ◦ σa)(ζ) > t}(cid:1) = 0. Subsequently it was observed in [26] that (S1) can be restated as lim ϕ(a)→1 kσϕ(a) ◦ ϕ ◦ σakH 2 = 0. (S1) (S2) (L) SURVEY OF COMPOSITION OPERATORS 11 Let f : D → X be an analytic function. We say that f ∈ BM OA(X) if f ∗,X = sup a∈D kf ◦ σa − f (a)kH 2(X) < ∞, and let kf kBM OA(X) = kf (0)kX + f ∗,X. There are also other natural possi- bilities. By departing from a well-known characterization of BM OA in terms of Carleson measures, see [18, Thm. VI.3.4], let f ∈ BM OAC(X) if f C,X = sup kf ′(z)k2 X (1 − σa(z)2)dA(z) < ∞. a∈DZD The norm in BM OAC(X) is kf kBM OAC(X) = kf (0)kX + f C,X. Blasco [3] showed that BM OA(X) = BM OAC(X), with equivalent norms, if and only if X is linearly isomorphic to a Hilbert space. Thus BM OA(X) and BM OAC(X) are different vector-valued versions of BM OA. (In section 5 we will meet yet another vector-valued version of BM OA.) Laitila [24], [25] initiated the study of composition operators on vector- BM OA(X) and BM OAC(X) → BM OAC(X) for any self-map ϕ : D → D. Moreover, if X is a reflexive Banach space and ϕ satisfies conditions (S1) valued BM OA-spaces. He observed that fCϕ is bounded BM OA(X) → and (S2), then fCϕ is weakly compact both BM OA(X) → BM OA(X) and BM OAC(X) → BM OAC(X). In order to obtain a complete characterization following Proposition 4 one has to verify condition (C1) for BM OA. This was actually a problem stated by Tjani in her Ph.D. thesis [52] and Bourdon, Cima and Matheson [7], which was eventually solved in [27] as follows. Theorem 8. The following conditions are equivalent for ϕ : D → D: (1) Cϕ : BM OA → BM OA is compact (2) Cϕ : BM OA → BM OA is weakly compact (3) (S1) holds (alternatively, (L) holds) It is part of the solution that condition (S2) is redundant in Smith's char- acterization above. The combination of Theorem 8 with [24], [25] completes the following result for these vector-valued BM OA-spaces. Theorem 9. Let X be a reflexive Banach space, and ϕ : D → D an analytic function. Then the following conditions are equivalent. (1) fCϕ is weakly compact BM OA(X) → BM OA(X) (2) fCϕ is weakly compact BM OAC(X) → BM OAC(X) (3) Cϕ : BM OA → BM OA is compact, that is, condition (S1) holds. The argument for Theorem 8 is quite intricate. It applies measure density ideas for the radial limits of ϕ combined with a criterion due to Leibov (1986), respectively Müller and Schechtman (1989), which allows to extract copies of the unit vector basis in c0 from bounded sequences in the subspace V M OA of BM OA. We refer to [27] for the full technical details. 12 JUSSI LAITILA AND HANS-OLAV TYLLI The results of sections 3 and 4 might suggest that the weak compactness of Cϕ : A → A always implies its compactness A → A for any Banach space A of analytic functions on D. However, this is not the case [31, Thm. 4.1]: Example 10. Let ϕ be the lens map ϕ(z) = (1 + z)1/2 − (1 − z)1/2 (1 + z)1/2 + (1 − z)1/2 , z ∈ D. Then there is an Orlicz function ψ so that Cϕ is weakly compact H ψ → H ψ, but not compact, where H ψ is the non-reflexive Hardy-Orlicz space of analytic functions of D defined by ψ. Question 11. Characterize the weakly compact compositions on the space H ψ(X) above. The operator Cϕ in Example 10 factors through H 4 by con- struction, and fCϕ through the reflexive space H 4(X) for reflexive spaces X. Thus fCϕ is weakly compact H ψ(X) → H ψ(X), so that (C) cannot hold for H ψ(X). To the best of our knowledge the first example of a weakly compact ana- lytic composition operator which is not compact was obtained in the context of uniform algebras defined on infinite-dimensional domains. Let UE be the open unit ball of the Tsirelson space E and ϕ : UE → UE the map x 7→ x/2. It was shown in [1, Example 3] that the composition operator f 7→ f ◦ ϕ is weakly compact, but non-compact, H ∞(UE) → H ∞(UE). Here H ∞(UE) is the uniform algebra of bounded scalar-valued analytic functions UE → C. 5. Weak vector-valued spaces Bonet, Domanski and Lindström [4] introduced the class of weak spaces of vector-valued analytic functions into the study of vector-valued composition operators. One of their aims was to provide an alternative approach to [35], but the weak spaces are in general different from the spaces considered in sections 3 and 4. On the other hand, for the class of weak spaces there are some general results concerning vector-valued compositions. Suppose that E is a Banach space of analytic functions f : D → C satisfy- ing the following conditions: (W1) E contains the constant functions, (W2) the closed unit ball BE is compact in the compact open topology τco of D. Recall that the vector-valued function f : D → X is analytic if and only if x∗ ◦ f is analytic D → C for all x∗ ∈ X ∗. This fact suggests to define f ∈ wE(X) if kf kwE(X) = sup kx∗kX ∗ ≤1 kx∗ ◦ f kE < ∞. SURVEY OF COMPOSITION OPERATORS 13 By the closed graph theorem kf kwE(X) is finite if and only if x∗ ◦f ∈ E for all x∗ ∈ X ∗. We will say that wE(X) is the weak space of vector-valued analytic functions D → X modelled on E. The spaces appearing in sections 3 and 4, whose norms involve pointwise norm quantities such as kf (z)kX , will in the sequel be called strong spaces. Such a distinction between strong and weak spaces is not precise, since e.g. wH ∞(X) = H ∞(X). If E is a Banach space of harmonic functions on D which satisfies (W1) and (W2), then one may similarly define the weak space wE(X) of vector-valued harmonic functions D → X, see [28]. Weak type spaces first appeared in the theory of vector measures, see e.g. [13, chap. 13]. The weak Hardy spaces wH p(X), and in particular their harmonic versions whp(X), have been studied by Blasco [2], as well as in [15, 16]. Weak spaces wE(X) have a dual nature, since they also admit a canonical isometrically isomorphic representation as certain spaces of bounded opera- tors. This general fact was observed in [4]. Note first that if E satisfies (W1) and (W2), then (W3) the evaluation maps δz ∈ E∗ for z ∈ D, where δz(f ) = f (z) for f ∈ E. The Dixmier-Ng theorem [42] implies that E = V ∗, where V = {u∗ ∈ E∗ : u∗ is τco-continuous on BE}. The identification of f ∈ E with u∗ 7→ u∗(f ) gives the isometric isomorphism E → V ∗. In addition, V = [δz ∈ E∗ : z ∈ D] by Hahn-Banach, where [B] denotes the closed linear span of the subset B ⊂ E∗. We next formulate the general linearization result from [4], which also implies that wE(X) is a Banach space. An analogue holds for weak harmonic spaces, see [28]. Theorem 12. Suppose that E satisfies (W1) and (W2), let V = [δz ∈ E∗ : z ∈ D] and X be a complex Banach space. Then there is an isometric isomorphism χ : L(V, X) → wE(X), so that (χ(T ))(z) = T (δz), (χ−1(f ))(δz) = f (z), hold for T ∈ L(V, X), f ∈ wE(X) and z ∈ D. Special cases and variants of this linearization result were known earlier. The closest precursor is the general results of Mujica [40] that apply to the case E = H ∞. An explicit operator representation was obtained by Blasco [2] for the weak harmonic spaces whp(X), where 1 ≤ p ≤ ∞. The study of composition operators between weak spaces of analytic func- tions was initiated by Bonet, Domański and Lindström [4], and this was extended in [28] to weak spaces of vector-valued harmonic functions. Propo- sition 11 of [4] contains the result on weak compactness stated below in The- orem 14 for the weak spaces, but [4] only explicitly discusses the weighted wH ∞ v (X)-spaces and the weak Bloch space wB(X) as examples. However, this approach applies to a large class of weak spaces of analytic functions, such 14 JUSSI LAITILA AND HANS-OLAV TYLLI as the weak Hardy and weak Bergman spaces, as well as to wBM OA(X), see the discussion below as well as in [28, 24, 25]. Let ϕ : D → D be an analytic map. The vector-valued composition fCϕ is bounded wE(X) → wE(X) if and only if Cϕ is bounded E → E. In fact, if x∗ ∈ X ∗ then kx∗ ◦ (fCϕf )kE = kCϕ(x∗ ◦ f )kE ≤ kCϕk · kx∗ ◦ f kE, so that kfCϕk ≤ kCϕk. For the converse it is worthwhile to point out that the framework from section 2 applies to the weak spaces. Lemma 13. If E satisfies (W1) and (W2), then the pair (E, wE(X)) satis- fies (AF1) -- (AF4) for any Banach space X Proof. Conditions (AF1) -- (AF3) are obvious. Towards (AF4) note that x∗(eδz(f )) = δz(x∗ ◦ f ), f ∈ wE(X), x∗ ∈ X ∗, where we momentarily use eδz for the vector-valued evaluations f 7→ f (z) taking wE(X) to X. (cid:3) We stress that the following basic weak compactness result from [4] for vector-valued compositions holds on all weak spaces wE(X). The proof uses different tools compared to the analytic arguments in Sections 3 and 4. Recall again from Corollary 2 that fCϕ is never compact wE(X) → wE(X) whenever X is infinite dimensional. Theorem 14. Suppose that E is a Banach space of analytic functions on D that satisfy (W1) and (W2). Let ϕ : D → D be an analytic map and X a reflexive Banach space. If Cϕ : E → E is compact, then fCϕ is weakly compact wE(X) → wE(X). Proof. Assume that Cϕ : E → E is compact. Its adjoint (Cϕ)∗ : E∗ → E∗ satisfies (Cϕ)∗(δz) = δϕ(z), z ∈ D, so that (Cϕ)∗(V ) ⊂ V . We obtain the factorization fCϕ = χ ◦ Uϕ ◦ χ−1, where Uϕ is the operator composition map T 7→ IX ◦ T ◦ (Cϕ)∗V ; L(V, X) → L(V, X), and χ is the isometric isomorphism L(V, X) → wE(X) from Theorem 12. Since (Cϕ)∗V is a compact operator V → V by duality, and IX is weakly compact, it follows from a general result of Saksman and Tylli, see [44, Prop. 2.3], that the operator composition Uϕ is weakly compact L(V, X) → (cid:3) L(V, X). Consequently fCϕ is weakly compact wE(X) → wE(X). Theorem 14 verifies condition (C2) from Proposition 4 for the weak spaces wE(X). The following observation includes many examples. SURVEY OF COMPOSITION OPERATORS 15 Proposition 15. Suppose that E is a Banach space of analytic functions on D that satisfy (W1) and (W2), let ϕ : D → D be an analytic map so that Cϕ is bounded E → E and X a reflexive Banach space. Suppose moreover: (C1) if Cϕ is weakly compact E → E, then Cϕ is compact E → E. Then one has the characterization (C) fCϕ is weakly compact wE(X) → wE(X) ⇔ Cϕ is compact E → E. Moreover, (C1) holds e.g. if E is one of the following spaces: H 1, A1 α > −1, BM OA, H ∞ The preceding examples cover results for wH ∞ α for v , where v is a bounded continuous weight on D, or B. v (X) and wB(X) from [4], wH 1(X) [28], and wBM OA(X) (combine Theorem 8 with [24, 25]). The results of sections 3 -- 5 raise the question of what is the precise connection between these strong and weak spaces of vector-valued analytic functions. Clearly wH 2(ℓ2) ≈ L(ℓ2) by Theorem 12, whereas H 2(ℓ2) is a separable Hilbert space, so the difference can be huge. On the other hand, [4] observed that wH ∞ v (X) (equal norms) and wB(X) ≈ B(X) (equivalent norms). It is evident that e.g. H p(X) ⊂ wH p(X), and v (X) = H ∞ kf kwH p(X) ≤ kf kH p(X), f ∈ H p(X), where 1 ≤ p < ∞ and X is any Banach space. Blasco [2] observed that h1(C(T)) wh1(C(T)) and hp(Lp′ ) for 1 < p < ∞. Subse- quently Freniche, García-Vázquez and Rodríguez-Piazza [15, 16] exhibited functions f ∈ whp(X) \ hp(X) and g ∈ wH p(X) \ H p(X) for 1 ≤ p < ∞ and any X. Fairly concrete functions of this kind were provided in [24, 28], and [30] contains the analogous results for the weak vs. strong Bergman norms. In fact, the norms ) whp(Lp′ k · kwH p(X) ≁ k · kH p(X) are non-equivalent on H p(X). Strict inclusions BM OA(X) wBM OA(X) and BM OAC(X) wBM OA(X) for any infinite-dimensional X were ob- tained in [24, 25]. A common feature of these examples for arbitrary X is the use of Dvoretzky's ℓn 2 -theorem to transfer from the Hilbert space setting to X. The linearization from Theorem 14 can also be used for other purposes. The following result from [4] concern weak conditional compactness on the spaces wE(X). Theorem 16. Suppose that E is a Banach space of analytic functions on D that satisfy (W1) and (W2). Let ϕ : D → D be an analytic map and X a Banach space that does not contain any subspaces linearly isomorphic to ℓ1. wE(X) → wE(X). If Cϕ : E → E is compact, then fCϕ is weakly conditionally compact The proof is analogous to that of Theorem 14, but instead apply [34] to deduce the weak conditional compactness of the operator composition Uϕ. 16 JUSSI LAITILA AND HANS-OLAV TYLLI Since wH 2(ℓ2) ≈ L(ℓ2) is non-reflexive one may also look for a characteri- ϕ satisfies Shapiro's condition (1). observation is not included in Proposition 15 since H 2 is reflexive. zation of weakly compact fCϕ : wH 2(ℓ2) → wH 2(ℓ2). Note that the following Proposition 17. fCϕ is weakly compact wH 2(ℓ2) → wH 2(ℓ2) if and only if ness of fCϕ : wH 2(ℓ2) → wH 2(ℓ2) implies condition (1). As in the proof of Proof. In view of Theorem 14 there remains to show that the weak compact- Theorem 14 let Uϕ be the operator composition map S 7→ S ◦ (Cϕ)∗V , L(V, ℓ2) → L(V, ℓ2), where V = [δz : z ∈ D] = H 2. We get that Uϕ = χ−1 ◦ fCϕ ◦ χ is weakly compact L(H 2, ℓ2) → L(H 2, ℓ2). It is known, see e.g. [44, Example 2.6], that for such operator compositions this yields the compactness of (Cϕ)∗V on V . Hence Cϕ is compact H 2 → H 2, so that (1) holds. (cid:3) The corresponding picture for the general class wE(X) is quite compli- cated for reflexive E, and remains open, since the spaces wE(X) can also be reflexive. For instance, the weak Hardy spaces wH 2(ℓp) ≈ L(H 2, ℓp) = K(H 2, ℓp) are reflexive for 1 < p < 2 by Pitt's theorem [33, Prop. 2.c.3] and [23, Sect. 2, Cor. 2]. Here K(X, Y ) denotes the space of compact operators X → Y . 6. Compositions from weak to strong spaces A different line of study concerns the mapping properties of composition operators from weak to strong spaces of analytic functions on D, such as wH p(X) → H p(X). This line was initiated by Laitila, Tylli and Wang in [30] for the Hardy and Bergman spaces, and subsequently the approach has been extended to weighted Bergman and Dirichlet spaces by Wang [55, 56, 57]. The question which motivated [30] came from S. Kaijser for X = ℓ2. Recall from section 5 that e.g. wH 2(ℓ2) ≈ L(ℓ2) while H 2(ℓ2) is a separable Hilbert space, so that boundedness of a composition operator wH 2(ℓ2) → H 2(ℓ2) entails strong compression. Somewhat surprisingly, the boundedness of Cϕ : wH p(X) → H p(X) for 2 ≤ p < ∞ is related to composition operators in the Hilbert-Schmidt class on H 2. Recall from [48, 11] that Cϕ is a Hilbert-Schmidt operator on H 2 precisely when kCϕk2 HS =ZT 1 1 − ϕ(ζ)2 dm(ζ) < ∞. The following result is taken from [30], which also contains a formally similar result for the vector-valued Bergman spaces. Note that results of this type have no counterparts in the scalar-valued theory. SURVEY OF COMPOSITION OPERATORS 17 Theorem 18. Let X be any infinite-dimensional complex Banach space. (1) If kCϕkHS < ∞, then Cϕ is bounded wH p(X) → H p(X) for any p satisfying 1 ≤ p < ∞. (2) The norm kCϕ : wH p(X) → H p(X)k is equivalent to kCϕk2/p HS for 2 < p < ∞. (3) kCϕ : wH 2(X) → H 2(X)k = kCϕkHS. Parts (2) and (3) are obtained by explicit computations for X = ℓ2. The 2 -theorem extension to arbitrary Banach spaces X is based on Dvoretzky's ℓn and coefficient multiplier results [14] corresponding to bounded operators Xk akzk 7→Xk λkakzk, H 2 → H p, where (λk) is a fixed sequence. By contrast to Theorem 18, fCϕ is bounded wBM OA(ℓ2) → BM OA(ℓ2) if and only if the scalar-valued operator Cϕ is bounded B → BM OA, see [30, Example 4.1], where B is the Bloch space. Question 19. (a) Does part (2) of Theorem 18 extend to 1 ≤ p < 2? The corresponding coefficient multiplier theorems H 2 → H p for 1 ≤ p < 2 are not readily useful. (b) Characterize the weakly compact compositions fCϕ from wH 1(X) to H 1(X) if X is reflexive. It is possible to show that if 0<r<1Zϕ(rζ)>s sup lim s→1 1 1 − ϕ(ζ)2 dm(ζ) = 0, then fCϕ is weakly compact from wH 1(X) to H 1(X) (details omitted). Note that fCϕ is never compact wH 1(X) → H 1(X) for infinite-dimensional X by section 2. One may also consider the composition operators f : 7→ f ◦ ϕ from strong to weak spaces, e.g. as acting H p(X) → wH p(X), but this case does not produce new qualitative phenomena. This follows from the factorization H p(X) cCϕ wH p(X) fCϕ %❏❏❏❏❏❏❏❏❏ J 9ssssssssss H p(X) H p(X) → wH p(X), and for p = 1 and reflexive spaces X one obtains that where cCϕ denotes the composition operator acting H p(X) → wH p(X) and J : H p(X) → wH p(X) is the continuous inclusion. Hence, any cCϕ is bounded cCϕ : H 1(X) → wH 1(X) is weakly compact if and only if ϕ satisfies (1), that is, Cϕ : H 1 → H 1 is compact. (For the "only if" -part note that section 2 applies here.) / / % 9 18 JUSSI LAITILA AND HANS-OLAV TYLLI 7. Operator-weighted composition operators In the final section we briefly discuss extensions of weighted composition operators to the vector-valued setting. Let ψ : D → C and ϕ : D → D be given analytic maps. The weighted composition operator Wψ,ϕ : f 7→ ψ · (f ◦ ϕ) defines a linear map H(D) → H(D), where H(D) denotes the linear space of analytic functions D → C. Clearly Wψ,ϕ = Mψ ◦ Cϕ, where Mψ is the pointwise multiplier defined by Mψf = ψ · f on H(D). Thus W1,ϕ = Cϕ and Wψ,id = Mψ. Weighted composition operators Wψ,ϕ have been extensively studied on a range of complex-valued analytic function spaces, and characterizations of e.g. boundedness and compactness are known for many classical spaces. The case of the weighted spaces H ∞ v was resolved by Contreras and Hernández- Díaz in [9] and Montes-Rodríguez in [39]. For 1 < p < ∞ and H p there is a Carleson measure characterization in [10], and the analogous results for the Bergman space Ap are found in [12]. The case of BM OA can be found in [26]. We refer to e.g. [17] and [19] for other types of function-theoretic conditions for the boundedness of Wψ,ϕ on H p. Moreover, by [10] all weakly compact weighted compositions Wψ,ϕ : H 1 → H 1 are compact. Independently Manhas [37] and the authors [29] proposed the following natural analogue of weighted composition operators in the vector-valued set- ting. Let ϕ : D → D be an analytic self-map and ψ : D → L(X, Y ) an ana- lytic operator-valued map, where X and Y are complex Banach spaces. Here L(X, Y ) denotes the space of bounded linear operators X → Y . Define the operator-weighted composition operator Wψ,ϕ by f 7→ ψ(f ◦ ϕ), that is, (Wψ,ϕ(f ))(z) = ψ(z)(f (ϕ(z)), z ∈ D, for analytic functions f : D → X. Note that z 7→ ψ(z)(f (ϕ(z)) is an analytic map D → Y , so that Wψ,ϕ is a linear map H(D, X) → H(D, Y ). Again Wψ,ϕ = Mψ ◦ Cϕ, where Mψ denotes the operator-valued pointwise multi- plier defined by (Mψf )(z) = ψ(z)(f (z)) from H(D, X) to H(D, X). Thus the operator-weighted composition operators form a much larger class com- pared to its scalar-valued relative. Operator-weighted compositions appear naturally: for a large class of Banach spaces X all linear onto isometries H ∞(X) → H ∞(X) have the form Wψ,ϕ, where ψ(z) ≡ U is a fixed onto isometry of X and ϕ is an automorphism of D, see [32] and [8]. The above definition of Wψ,ϕ is analogous to that of weighted compositions on spaces C(K, X) of continuous functions, see e.g. [22], where K is a compact Haus- dorff space. The present knowledge of operator-weighted compositions is fairly rudi- v (X) spaces and their mentary and most of the results deal with weighted H ∞ SURVEY OF COMPOSITION OPERATORS 19 locally convex variants. Characterizations of boundedness and (weak) com- pactness of Wψ,ϕ between H ∞ v (X) spaces were obtained in [29], and results for certain locally convex spaces of analytic vector-valued functions are found in [37] and [6]. Boundedness of the operator multipliers Mψ in related set- tings has been considered e.g. in [36, 37, 38, 57]. We next state the main results from [29], which are vector-valued exten- sions of scalar results from [9] and [39]. For a bounded continuous weight function v : D → (0, ∞), put v(z) = (sup{f (z) : kf kH∞ v ≤ 1})−1. If ψ : D → L(X, Y ) is a given analytic operator-valued map define the auxil- iary linear map Tψ by x 7→ ψ(·)x. It follows that Tψ is bounded X → H ∞ w (Y ) whenever Wψ,ϕ is bounded H ∞ v (X) → H ∞ w (Y ). Theorem 20. (1) Let v and w be weight functions. Then kWψ,ϕ : H ∞ v (X) → H ∞ w (Y )k = sup z∈D w(z) v(ϕ(z)) kψ(z)kL(X,Y ). (2) Assume that v and w are radial weight functions. Then Wψ,ϕ is com- pact (respectively, weakly compact) H ∞ v (X) → H ∞ w (Y ) if and only if both lim sup ϕ(z)→1 w(z) v(ϕ(z)) kψ(z)kL(X,Y ) = 0, (4) and Tψ is compact (respectively, weakly compact) X → H ∞ w (Y ). Parts of Theorem 20 were independently obtained in [37] together with other results. Clearly the case X = Y and ψ(z) ≡ IX yields the bound- edness and weak compactness results from [35, 4] for fCϕ, since Tψ is then (essentially) IX. Theorem 20 points to some striking differences between scalar- and vector- valued weighted compositions as well as between operator-weighted and stan- dard composition operators. For instance, the auxiliary operators Tψ play no role for scalar weighted compositions. Moreover, Wψ,ϕ can easily be compact H ∞ w (Y ) for infinite-dimensional spaces X and Y . For instance, let kϕk∞ < 1 and ψ(z) ≡ U ∈ K(X, Y ), whence (4) holds and Tψ is compact X → H ∞ w (X) do not always factor through IX as in Proposition 1, but (F2) is replaced by the following factorization for any z ∈ D: w (Y ). Operator-weighted compositions Wψ,ϕ : H ∞ v (X) → H ∞ v (X) → H ∞ H ∞ v (X) Wψ,ϕ / H ∞ w (Y ) j X ψ(z) δz / Y / O O / 20 JUSSI LAITILA AND HANS-OLAV TYLLI There are also examples where Wψ,ϕ is weakly compact H ∞ w (X) but not compact. For this one may use the fact [29, Thm. 4.4] that if w is a radial weight and ψ ∈ H 0 w(L(X, Y ))), then Tψ is compact (respectively, weakly compact) if and only if v (X) → H ∞ ψ(D) ⊂ K(X, Y ) (respectively, ψ(D) ⊂ W (X, Y )). Here H 0 nomials in H ∞ Y . There are further differences between H ∞ ting as studied in [6]. w(L(X, Y ))) denotes the closure of the analytic L(X, Y )-valued poly- w (L(X, Y ))), and W (X, Y ) the weakly compact operators X → v (X) and the locally convex set- The following problem stated in [29] appears quite challenging. Question 21. Characterize boundedness and (weak) compactness of operator- weighted compositions Wψ,ϕ : H p(X) → H p(Y ) as well as Ap(X) → Ap(Y ) for 1 ≤ p < ∞. The argument from [10] based on Carleson measure tech- niques does not readily extend to the vector-valued setting. References [1] R. Aron, P. Galindo and M. Lindström, Compact homomorphisms between alge- bras of analytic functions, Studia Math. 123 (1997), 235 -- 247. [2] O. Blasco, Boundary values of vector-valued harmonic functions considered as operators, Studia Math. LXXXVI (1987), 19 -- 33. [3] O. Blasco, Remarks on vector-valued BMOA and vector-valued multipliers, Posi- tivity 4 (2000), 339 -- 356. [4] J. Bonet, P. Domański and M. Lindström, Weakly compact composition operators on analytic vector-valued function spaces, Ann. Acad. Sci. Fenn. Math. 26 (2001), 233 -- 248. [5] J. Bonet and M. Friz, Weakly compact composition operators on locally convex spaces, Math. Nachr. 245 (2002), 26 -- 44. [6] J. Bonet, M.C. Gómez-Collado, D. Jornet and E. Wolf, Operator-weighted compo- sition operators between weighted spaces of vector-valued analytic functions, Ann. Acad. Sci. Fenn. Math. 37 (2012), 319 -- 338. [7] P.S. Bourdon, J.A. Cima and A.L. Matheson, Compact composition operators on BMOA, Trans. Amer. Math. Soc. 351 (1999), 2183 -- 2196. [8] M. Cambern and K. Jarosz, Multipliers and isometries in H ∞ E , Bull. London Math. Soc. 22 (1990), 463 -- 466. [9] M. D. Contreras and A.G. Hernández-Díaz, Weighted composition operators in weighted Banach spaces of analytic functions, J. Austral. Math. Soc. 69 (2000), 41 -- 60. [10] M.D. Contreras and A.G. Hernández-Díaz, Weighted composition operators on Hardy spaces, J. Math. Anal. Appl. 263 (2001), 224 -- 233. [11] C.C. Cowen and B.D. MacCluer, Composition Operators on Spaces of Analytic Functions, CRC Press, Boca Raton 1995. [12] Z. Cucković and R. Zhao, Weighted composition operators on the Bergman space, J. London Math. Soc. 70 (2004), 499 -- 511. [13] N. Dinculeanu, Vector Measures, Pergamon Press, 1967. [14] P.L. Duren, On the multipliers of H p spaces, Proc. Amer. Math. Soc. 22 (1969), 24 -- 27. SURVEY OF COMPOSITION OPERATORS 21 [15] F.J. Freniche, J.C. García-Vázquez and L. Rodríguez-Piazza, The failure of Fa- tou's theorem on Poisson integrals of Pettis integrable functions, J. Funct. Anal. 160 (1998), 28-41. [16] F.J. Freniche, J.C. García-Vázquez and L. Rodríguez-Piazza, Operators into Hardy spaces and analytic Pettis integrable functions, in: Recent Progress in Functional Analysis, North-Holland Mathematical Studies vol. 189 (2001), North- Holland, pp. 349 -- 362. [17] E.A. Gallardo-Gutiérrez and J.R. Partington, The role of BMOA in the bounded- ness of weighted composition operators, J. Funct. Anal. 258 (2010), 3593 -- 3603. [18] J. Garnett, Bounded Analytic Functions, Academic Press, 1981. [19] G. Gunatillake, Compact weighted composition operators on the Hardy space, Proc. Amer. Math. Soc. 136 (2008), 2895 -- 2899. [20] E. Hille and R.S. Phillips, Functional analysis and semi-groups, Amer. Math. Soc. Colloq. Publ. 31, Amer. Math. Soc., Providence, 1975. [21] W.E. Hornor and J.E. Jamison, Isometrically equivalent composition operators on spaces of analytic vector-valued functions, Glasgow Math. J. 41 (1999), 441 -- 451. [22] J.E. Jamison and M. Rajagopalan, Weighted composition operators on C(K, E), J. Operator Theory 19 (1988), 307 -- 317. [23] N.J. Kalton, Spaces of compact operators, Math. Ann. 208 (1974), 267 -- 278. [24] J. Laitila, Weakly compact composition operators on vector-valued BMOA, J. Math. Anal. Appl. 308 (2005), 730 -- 745. [25] J. Laitila, Composition operators and vector-valued BMOA, Integral Equ. Opera- tor Theory 58 (2007), 487 -- 502. [26] J. Laitila, Weighted composition operators on BMOA, Comput. Methods Funct. Theory 9 (2009), 27 -- 46. [27] J. Laitila, P. Nieminen, E. Saksman and H.-O. Tylli, Compact and weakly com- pact composition operators on BMOA, Complex Anal. Oper. Theory 7 (2013), 163 -- 181. [28] J. Laitila and H.-O. Tylli, Composition operators on vector-valued harmonic func- tions and Cauchy transforms, Indiana Univ. Math. J. 55 (2006), 719 -- 746. [29] J. Laitila and H.-O. Tylli, Operator-weighted composition operators on vector- valued analytic function spaces, Illinois J. Math. 53 (2009), 1019 -- 1032. [30] J. Laitila, H.-O. Tylli and M. Wang, Composition operators from weak to strong spaces of vector-valued analytic functions, J. Operator Theory 62 (2009), 281 -- 295. [31] P. Lefevre, D. Li, H. Queffelec and L. Rodriguez-Piazza, Some new properties of composition operators associated with the lens map, Israel J. Math. 195 (2013), 801 -- 824. [32] P.-K. Lin, The isometries of H ∞(E), Pacific J. Math. 143 (1990), 69 -- 77. [33] J. Lindenstrauss and L. Tzafriri, Classical Banach spaces I, Springer-Verlag, 1977. [34] M. Lindström and G. Schlüchtermann, Composition of operator ideals, Math. Scand. 84 (1999), 284 -- 296. [35] P. Liu, E. Saksman and H.-O. Tylli, Small composition operators on analytic vector-valued function spaces, Pacific J. Math. 184 (1998), 295 -- 309. [36] J. S. Manhas, Multiplication operators on weighted locally convex spaces of vector- valued analytic functions, Southeast Asian Bull. Math. 27 (2003), 649 -- 660. [37] J.S. Manhas, Weighted composition operators on weighted spaces of vector-valued analytic functions, J. Korean Math. Soc. 45 (2008), 1203 -- 1220. [38] J.S. Manhas, Weighted composition operators between weighted spaces of vector- valued holomorphic functions on Banach spaces, Appl. Math. Comput. 218 (2011), 929 -- 934. 22 JUSSI LAITILA AND HANS-OLAV TYLLI [39] A. Montes-Rodríguez, Weighted composition operators on weighted Banach spaces of analytic functions, J. London Math. Soc. 61 (2000), 872 -- 884. [40] J. Mujica, Linearization of bounded holomorphic mappings on Banach spaces, Trans. Amer. Math. Soc. 324 (1991), 295 -- 309. [41] P.F.X. Müller, Isomorphisms between H 1-spaces, Monografie Matematyczne vol. 66, Birkhäuser, 2005. [42] K. Ng, On a theorem of Dixmier, Math. Scand. 29 (1971), 279 -- 280. [43] A. Pietsch, Operator Ideals, North-Holland, 1980. [44] E. Saksman and H.-O. Tylli, Multiplications and elementary operators in the Ba- nach space setting, in: Methods in Banach space theory, London Math. Soc. Lecture Note Ser. vol. 337 (2006), Cambridge University Press, Cambridge, pp. 253-292. [45] D. Sarason, Weak compactness of holomorphic composition operators on H 1, in: Functional Analysis and Operator Theory, Lecture Notes in Mathematics 1511 (1992), Springer-Verlag, pp. 75 -- 79. [46] J.H. Shapiro, The essential norm of a composition operator, Ann. Math. 125 (1987), 375 -- 404. [47] J.H. Shapiro, Composition Operators and Classical Function Theory, Springer- Verlag, New York, 1993. [48] J.H. Shapiro and P.D. Taylor, Compact, nuclear and Hilbert-Schmidt composition operators on H 2, Indiana Univ. Math. J. 23 (1973/74), 471 -- 496. [49] S.D. Sharma and U. Bhanu, Composition operators on vector-valued Hardy spaces, Extracta Math. 14 (1999), 31 -- 39. [50] W. Smith, Compactness of composition operators on BMOA, Proc. Amer. Math. Soc. 127 (1999), 2715 -- 2725. [51] C. Stanton, Counting functions and majorization theorems for Jensen measures, Pacific J. Math. 125 (1986), 459 -- 468. [52] M. Tjani, Compact composition operators on some Möbius invariant Banach spaces, Ph.D. thesis, Michigan State University, 1996. [53] S.-I. Ueki, Composition operators on the Fock space of vector-valued analytic func- tions, Ars Combin. 100 (2011), 161 -- 167. [54] M. Wang, Composition operators on analytic vector-valued Nevanlinna classes, Acta Math. Sci. 25 (2005), 771 -- 780. [55] M. Wang, Composition operators between vector-valued weighted Bergman spaces, Far East J. Math. Sci. 26 (2007), 419 -- 431. [56] M. Wang, Composition operators between different Dirichlet spaces, Complex Var. Elliptic Equ. 53 (2008), 1093 -- 1102. [57] M. Wang, Weighted composition operators between Dirichlet spaces, Acta Math. Sci. Ser. B Engl. Ed. 31 (2011), 641 -- 651. Jussi Laitila, Department of Biosciences, P.O. Box 65, FI-00014 University of Helsinki, Helsinki, Finland E-mail address: [email protected] Hans-Olav Tylli, Department of Mathematics and Statistics, P.O. Box 68, FI-00014 University of Helsinki, Helsinki, Finland E-mail address: [email protected]
1301.4819	1	1301	2013-01-21T11:08:30	Smoothing properties of the discrete fractional maximal operator on Besov and Triebel--Lizorkin spaces	[ "math.FA" ]	Motivated by the results of Korry and Kinnunen and Saksman, we study the behaviour of the discrete fractional maximal operator on fractional Hajlasz spaces, Hajlasz-Besov and Hajlasz-Triebel-Lizorkin spaces on metric measure spaces. We show that the discrete fractional maximal operator maps these spaces to the spaces of the same type with higher smoothness. Our results extend and unify aforementioned results. We present our results in general setting, but they are new already in the Euclidean case.	math.FA	math	SMOOTHING PROPERTIES OF THE DISCRETE FRACTIONAL MAXIMAL OPERATOR ON BESOV AND TRIEBEL -- LIZORKIN SPACES TONI HEIKKINEN AND HELI TUOMINEN Abstract. Motivated by the results of Korry and Kinnunen and Saks- man, we study the behaviour of the discrete fractional maximal operator on fractional Haj lasz spaces, Haj lasz -- Besov and Haj lasz -- Triebel -- Lizorkin spaces on metric measure spaces. We show that the discrete fractional maximal operator maps these spaces to the spaces of the same type with higher smoothness. Our results extend and unify aforementioned results. We present our results in general setting, but they are new already in the Euclidean case. 1. Introduction Maximal functions are standard tools in harmonic analysis. They are usually used to estimate absolute size, but recently there has been interest in studying their regularity properties, see [1], [2], [3], [11], [12], [13], [15], [17], [18], [19], [21], [22], [23], [25], [26], [28]. A starting point was [17], where Kinnunen observed that the Hardy-Littlewood maximal operator is bounded on W 1,p(Rn) for 1 < p ≤ ∞. In [22] and [23] Korry extended this result by showing that the maximal operator preserves also fractional Sobolev spaces as well as Besov and Triebel -- Lizorkin spaces. Another kind of extension was given in [20], where Kinnunen and Saksman showed that the fractional maximal operator Mα, defined by rα B(x, r) ZB(x,r) Mα u(x) = sup r>0 u(y) dy, is bounded from W 1,p(Rn) to W 1,p∗(Rn), where p∗ = np/(n − αp), and from Lp(Rn) to W 1,q(Rn), where q = np/(n−(α −1)p) and W 1,q(Rn) is the homoge- nous Sobolev space. These results indicate that Mα has similar smoothing properties as the Riesz potential. It is natural to ask whether these results can be seen as special cases of the behaviour of the fractional maximal operator on Besov and Triebel -- Lizorkin spaces. In this paper we show that this is indeed the case, and that all these results can be obtained by the same rather simple method. Instead of the Date: May 31, 2018. 2010 Mathematics Subject Classification. 42B25, 46E35. Key words and phrases. Besov space, fractional maximal function, fractional Sobolev space, Triebel -- Lizorkin space, metric measure space. The research was supported by the Academy of Finland, grant no. 135561. 1 2 TONI HEIKKINEN AND HELI TUOMINEN standard fractional maximal operator, we consider its variant, the so-called discrete fractional maximal operator M∗ α. This allows us to present our results in a setting of doubling metric measure spaces. In this generality, the standard fractional maximal operator behaves quite badly. Indeed, one can construct spaces, where the fractional maximal function of a Lipschitz function fails to be continuous, see [3] and [15]. Since M∗ α and Mα are comparable, for practical purposes it does not matter which one we choose. The discrete fractional maximal operator was introduced in [18] and further studied in [21], [1] and [13]. Among the many possible definitions of Besov and Triebel -- Lizorkin spaces, the most suitable for our purposes is the one based on Haj lasz type pointwise inequalities. This approach, introduced by Koskela, Yang and Zhou in [24], provides a new point of view to the classical Besov and Triebel -- Lizorkin spaces. On the other hand, it allows these spaces to be defined in the setting of metric measure spaces. By employing this definition, we can prove very general results using only simple "telescoping" arguments and Poincar´e type inequalities. As special cases, we obtain versions of the results of Kinnunen and Saksman as well as those of Korry, see Remark 3.4 and Theorems 4.5 and 4.6. We prove our results in doubling metric measure spaces but they are new even in Euclidean spaces. Our main results (Theorems 4.3 and 4.4) imply that if α ≥ 0 and 0 < s+α < 1, then M∗ p,q (Rn) for n/(n + s) < p, q < ∞ and from Bs p,q (Rn) for n/(n + s) < p < ∞, 0 < q < ∞, see Section 4 for the definition of Triebel -- Lizorkin and Besov spaces. α is bounded from F s p,q(Rn) to Bs+α p,q(Rn) to F s+α 2. Preliminaries and notation We assume that X = (X, d, µ) is a metric measure space equipped with a metric d and a Borel regular outer measure µ, which satisfies 0 < µ(U) < ∞ whenever U is nonempty, open and bounded. We assume that the measure is doubling, that is, there exists a fixed constant cd > 0, called the doubling constant, such that (2.1) µ(B(x, 2r)) ≤ cdµ(B(x, r)) for every ball B(x, r) = {y ∈ X : d(y, x) < r}. The doubling condition implies that (2.2) µ(B(y, r)) µ(B(x, R)) ≥ C(cid:16) r R(cid:17)Q for every 0 < r ≤ R and y ∈ B(x, R) for some C and Q > 1 that only depend on cD. In fact, we may take Q = log2 cd. For the boundedness of the fractional maximal operator in Lp, we have to assume, in Theorems 2.1 and 3.3.(b), that the measure µ satisfies the lower bound condition (2.3) µ(B(x, r)) ≥ clrQ THE FRACTIONAL MAXIMAL OPERATOR ON BESOV AND T -- L SPACES 3 with some constant cl > 0 for all x ∈ X and r > 0. Throughout the paper, C will denote a positive constant whose value is not necessarily the same at each occurrence. The fractional maximal function. Let α ≥ 0. The fractional maximal function of a locally integrable function u is (2.4) Mα u(x) = sup r>0 rαZB(x,r) u dµ, where uB = RB u dµ = 1 α = 0, we have the usual Hardy-Littlewood maximal function µ(B) RB u dµ is the integral average of u over B. For M u(x) = sup r>0ZB(x,r) u dµ. The following Sobolev type inequality for the fractional maximal operator fol- lows easily from the boundedness of the Hardy-Littlewood maximal operator in Lp, for the proof, see [4], [6] or [15]. Theorem 2.1. Assume that the measure lower bound condition holds. If p > 1 and 0 < α < Q/p, then there is a constant C > 0, depending only on the doubling constant, constant in the measure lower bound, p and α, such that for every u ∈ Lp(X) with p∗ = Qp/(Q − αp). k Mα ukLp∗ (X) ≤ CkukLp(X), Remark 2.2. If u is only locally integrable, then Mα u may well be identically infinite. However, if Mα u(x0) < ∞ for some x0 ∈ X, then Mα u(x) < ∞ for almost every x. This follows from the estimate rαZB(x,r) u dµ ≤ µ(B(x0, r + d(x, x0))) µ(B(x, r)) Mα u(x0) combined with the doubling condition and the fact that lim r→0 rαZB(x,r) u dµ < ∞, whenever x is a Lebesgue point of u. The discrete fractional maximal function. We begin the construction of the discrete maximal function with a covering of the space. Let r > 0. Since the measure is doubling, there are balls B(xi, r), i = 1, 2, . . . , such that X = ∞ [i=1 B(xi, r) and ∞ Xi=1 χB(xi,6r) ≤ N < ∞, where χB(xi,6r) is the characteristic function of the ball B(xi, 6r). This means that the dilated balls B(xi, 6r), i = 1, 2, . . . , are of bounded overlap. The constant N depends only on the doubling constant and, in particular, it is independent of r. 4 TONI HEIKKINEN AND HELI TUOMINEN Then we construct a partition of unity subordinate to the covering B(xi, r), i = 1, 2, . . . , of X. Indeed, there is a family of functions ϕi, i = 1, 2, . . . , such that 0 ≤ ϕi ≤ 1, ϕi = 0 in X \ B(xi, 6r), ϕi ≥ ν in B(xi, 3r), ϕi is Lipschitz with constant L/r with ν and L depending only on the doubling constant, and ∞ Xi=1 ϕi(x) = 1 for every x ∈ X. The discrete convolution of a locally integrable function u at the scale 3r is ∞ for every x ∈ X, and we write uα ur(x) = ϕi(x)uB(xi,3r) Xi=1 r = rαur. Let rj, j = 1, 2, . . . be an enumeration of the positive rationals and let balls B(xi, rj), i = 1, 2, . . . be a covering of X as above. The discrete fractional maximal function of u in X is M∗ uα α u(x) = sup rj(x) j for every x ∈ X. For α = 0, we obtain the Hardy-Littlewood type discrete maximal function M∗ studied in [18], [21] and [1]. The discrete fractional maximal function is easily seen to be comparable to the standard fractional maximal function, see [13]. 3. Fractional Haj lasz spaces Let u be a measurable function and let s ≥ 0. A nonnegative measurable function g is an s-Haj lasz gradient of u if there exists E ⊂ X with µ(E) = 0 such that for all x, y ∈ X \ E, u(x) − u(y) ≤ d(x, y)s(g(x) + g(y)). (3.1) The collection of all s-Haj lasz gradients of u is denoted by Ds(u). A homoge- neous Haj lasz space M s,p(X) consists of measurable functions u such that kuk M s,p(X) = inf g∈Ds(u) kgkLp(X) is finite. The Haj lasz space M s,p(X) is norm M s,p(X) ∩ Lp(X) equipped with the kukM s,p(X) = kukLp(X) + kuk M s,p(X). The space M 1,p(X), a counterpart of a Sobolev space in metric measure space, was introduced in [9], see also [10]. The fractional spaces M s,p(X) were introduced in [30] and studied for example in [16] and [14]. Notice that M 0,p(X) = Lp(X). The pointwise definition of the Haj lasz spaces implies the validity of Sobolev- Poincar´e type inequalities without the assumption that the space admits any weak Poincar´e inequality. THE FRACTIONAL MAXIMAL OPERATOR ON BESOV AND T -- L SPACES 5 Lemma 3.1 ([7]). Let s ∈ [0, ∞) and let p ∈ (0, Q/s). There exists a constant C such that for all measurable functions u with g ∈ Ds(u), all x ∈ X and r > 0, (3.2) inf c∈R(cid:18)ZB(x,r) u(y) − cp∗(s) dµ(y)(cid:19)1/p∗(s) ≤ Crs(cid:18)ZB(x,2r) gp dµ(cid:19)1/p , where p∗(s) = Qp/(Q − sp). Moreover, if p ≥ Q/(Q + s) and g ∈ Ds(u) ∩ Lp(X), then (3.2) implies that u is locally integrable and that (3.3) ZB(x,r) u − uB(x,r) dµ ≤ Crs(cid:18)ZB(x,2r) gp dµ(cid:19)1/p . For the case s = 1, see [9] and [10]. In the next theorem, we use the following simple result. If ui, i ∈ N, are measurable functions with a common s-Haj lasz gradient g and u = supi ui is finite almost everywhere, then g is an s-Haj lasz gradient of u. Theorem 3.2. Assume that M∗ s-Haj lasz gradient of u. α u 6≡ ∞. Let t ≥ Q/(Q + s) and let g be an a) If 0 < s + α ≤ 1, then there exists a constant C > 0 such that is an (s + α)-Haj lasz gradient of M∗ g = C(cid:0)M gt(cid:1)1/t α u. b) If s + α > 1, then there exists a constant C > 0 such that is a 1-Haj lasz gradient of M∗ α u. g = C(cid:0)Mt(s+α−1) gt(cid:1)1/t Proof. We begin by proving the claims for uα let x, y ∈ X. r . Let r > 0, let g ∈ Ds(u) and Assume first that r ≥ d(x, y). Let Ixy be a set of indices i for which x or y belongs to B(xi, 6r). Then, for each i ∈ Ixy, B(xi, 3r) ⊂ B(x, 10r) ⊂ B(xi, 17r). This together with the doubling condition, the properties of the functions ϕi, the fact that there are bounded number of indices in Ixy and Poincar´e inequality (3.3) implies that uα r (x) − uα r (y) ≤ rα Xi∈Ixy ϕi(x) − ϕi(y)uB(xi,3r) − uB(x,10r) (3.4) u − uB(xi,3r) dµ ZB(x,10r) ≤ Crα−1d(x, y) Xi∈Ixy ≤ Crα−1d(x, y)ZB(x,10r) ≤ Crs+α−1d(x, y)(cid:18)ZB(x,20r) u − uB(x,10r) dµ gt dµ(cid:19)1/t . 6 TONI HEIKKINEN AND HELI TUOMINEN If 0 < s + α ≤ 1, then by (3.4) and the assumption r ≥ d(x, y), we have that uα r (x) − uα r (y) ≤ Cd(x, y)s+α(cid:0) M gt(x)(cid:1)1/t. If s + α > 1, then by (3.4), uα r (x) − uα r (y) ≤ Cd(x, y)(cid:0) Mt(s+α−1) gt(x)(cid:1)1/t. This shows that Haj lasz gradient inequality (3.1) with desired exponent holds when r ≥ d(x, y). Assume then that r < d(x, y). Let R = d(x, y). Then B(y, r) ⊂ B(x, 2R) and (3.5) uα r (x) − uα r (y) ≤ rα(cid:16)Xi∈Ix +Xi∈Iy ϕi(x)uB(xi,3r) − uB(x,9R) ϕi(y)uB(xi,3r) − uB(x,9R)(cid:17), where Ix is a set of indices i for which x belongs to B(xi, 6r) and Iy the corresponding set for y. Let k ∈ N be the smallest integer such that 2kr ≥ R. Assume first that 0 < s + α ≤ 1. If i ∈ Ix, then (3.6) uB(xi,3r) − uB(x,9R) ≤ uB(xi,3r) − uB(x,9r) + + uB(x,2k9r) − uB(x,9R). k Xi=1 uB(x,2i9r) − uB(x,2i−19r) By the doubling condition and Poincar´e inequality (3.3), we have rαuB(xi,3r) − uB(x,9r) ≤ CrαZB(x,9r) u − uB(x,9r) dµ gt dµ(cid:19)1/t ≤ Crs+α(cid:18)ZB(x,18r) ≤ CRs+α(cid:0) M gt(x)(cid:1)1/t , and, by the doubling condition, Poincar´e inequality (3.3), the fact that r ≤ 2i9r for all i, and the selection of k, uB(x,2i9r) − uB(x,2i−19r) ≤ Crα u − uB(x,2i9r) dµ (3.7) rα k Xi=1 (3.8) k Xi=1ZB(x,2i9r) (2i9r)s+α(cid:18)ZB(x,2i+19r) Xi=1 k ≤ C ≤ CRs+α(cid:0) M gt(x)(cid:1)1/t . gt dµ(cid:19)1/t THE FRACTIONAL MAXIMAL OPERATOR ON BESOV AND T -- L SPACES 7 Similarly we obtain that (3.9) rαuB(x,2k9r) − uB(x,9R) ≤ CRs+α(cid:18)ZB(x,36R) ≤ CRs+α(cid:0) M gt(x)(cid:1)1/t . gt dµ(cid:19)1/t If i ∈ Iy, we use balls B(y, 2i9r) instead of balls B(x, 2i9r) in (3.6). Estimates corresponding (3.7) and (3.8) are as above (x replaced by y) and, corresponding to (3.9), rαuB(y,2k9r) − uB(x,9R) ≤ CRs+α(cid:18)ZB(x,38R) gt dµ(cid:19)1/t (3.10) Now, by (3.5)-(3.10) and the fact R = d(x, y), we have ≤ CRs+α(cid:0) M gt(x)(cid:1)1/t. rαuB(xi,3r) − uB(x,R) ≤ Cd(x, y)s+α(cid:16)(cid:0) M gt(x)(cid:1)1/t + M gt(y)(cid:1)1/t(cid:17). If s + α > 1, then similar estimates as above show that if i ∈ Ix ∪ Iy, then (3.11) rαuB(xi,3r) −uB(x,R) ≤ Cd(x, y)(cid:16)(cid:0) Mt(s+α−1) gt(x)(cid:1)1/t +(cid:0) Mt(s+α−1) gt(y)(cid:1)1/t(cid:17). These estimates together with (3.5) and the fact that there are bounded num- ber of indices in Ix and Iy imply that Haj lasz gradient inequality (3.1) with desired exponent holds when r < d(x, y). The claim for uα r follows from the estimates above and for M∗ discussion before the theorem. Theorem 3.3. Let Q/(Q + s) < p < ∞. a) If 0 < s + α ≤ 1, there exists a constant C > 0, such that k M∗ α uk M s+α,p(X) ≤ Ckuk M s,p(X) α u from the (cid:3) for all u ∈ M s,p(X) with M∗ α u 6≡ ∞. b) If 1 < s + α ≤ 1 + Q/p and the measure lower bound condition holds, there exists a constant C > 0 such that k M∗ α uk M 1,q(X) ≤ Ckuk M s,p(X), where q = Qp/(Q − (s + α − 1)p), for all u ∈ M s,p(X) with M∗ α u 6≡ ∞. Proof. a) Let Q/(Q + s) ≤ t < p. By Theorem 3.2, the function C(M gt)1/t is an (s + α)-gradient of M∗ α u. Since g ∈ Lp(X), the claim follows from the boundedness of the Hardy -- Littlewood maximal operator in Lq(X) for q > 1. b) Let Q/(Q + s) ≤ t < p. By Theorem 3.2, the function (Mt(s+α−1) gt)1/t α u. Since g ∈ Lp(X), the claim follows from Theorem (cid:3) is a 1-gradient of M∗ 2.1. 8 TONI HEIKKINEN AND HELI TUOMINEN Remark 3.4. In the cases s = 0 and s = 1 of Theorem 3.3.b), we obtain counterparts of the results of Kinnunen and Saksman. Remark 3.5. As a special case of Theorems 3.2 and 3.3 we obtain boundedness results for the discrete maximal operator M∗ in M s,p(X). If 0 < s ≤ 1, then g = C(M gt)1/t is an s-Haj lasz gradient of M∗ u for all t ≥ Q/(Q + s) and k M∗ uk M s,p(X) ≤ Ckuk M s,p(X) for all u ∈ M s,p(X), p > Q/(Q + s). If 1 < s ≤ 1 + Q/p, then g = C(Mt(s−1) gt)1/t is a 1-Haj lasz gradient of M∗ u for all t ≥ Q/(Q + s) and k M∗ uk M 1,q(X) ≤ Ckuk M s,p(X), where q = Qp/(Q − (s − 1)p), for all u ∈ M s,p(X). Moreover, when s = 1, we obtain boundedness results for the discrete max- imal operator M∗ in (homogeneous) Haj lasz spaces M 1,p(X), proved earlier for M 1,p(X) in [18] and [21]. 4. Haj lasz -- Besov and Haj lasz -- Triebel -- Lizorkin spaces Let u be a measurable function and let s ∈ (0, ∞). Following [24], we say that a sequence of nonnegative measurable functions (gk)k∈Z is a fractional s-Haj lasz gradient of u if there exists E ⊂ X with µ(E) = 0 such that u(x) − u(y) ≤ d(x, y)s(gk(x) + gk(y)) for all k ∈ Z and all x, y ∈ X \ E satisfying 2−k−1 ≤ d(x, y) < 2−k. The collection of all fractional s-Haj lasz gradients of u is denoted by Ds(u). For p ∈ (0, ∞), q ∈ (0, ∞] and a sequence (fk)k∈Z of measurable functions, we write and k(fk)k∈ZkLp(X, lq) = (cid:13)(cid:13)k (fk)k∈Z klq(cid:13)(cid:13)Lp(X) k(fk)k∈Zklq(Lp(X)) = (cid:13)(cid:13)(cid:0)kfkkLp(X)(cid:1)k∈Z(cid:13)(cid:13)lq , where k(fk)klq = (Pk∈Z fkq)1/q if 0 < q < ∞ and k(fk)kl∞ = supk∈Z fk. The homogeneous Haj lasz -- Triebel -- Lizorkin space M s p,q(X) consists of mea- surable functions u such that kuk M s p,q(X) = inf (gk)∈Ds(u) k(gk)kLp(X, lq) is finite. The Haj lasz -- Triebel -- Lizorkin space M s equipped with the norm p,q(X) is M s p,q(X) ∩ Lp(X) kukM s p,q(X) = kukLp(X) + kuk M s p,q(X). The homogeneous Haj lasz -- Besov space N s p,q(X) consists of measurable func- tions u such that kuk N s p,q(X) = inf (gk)∈Ds(u) k(gk)klq(Lp(X)) THE FRACTIONAL MAXIMAL OPERATOR ON BESOV AND T -- L SPACES 9 is finite and the Haj lasz -- Besov space N s with the norm p,q(X) is N s p,q(X) ∩ Lp(X) equipped kukN s p,q(X) = kukLp(X) + kuk N s p,q(X). M s Notice that p,∞(X) is the homogeneous fractional Haj lasz space M s,p(X), for the simple proof, see [24, Prop. 2.1]. The homogeneous Haj lasz -- Triebel -- M s p,q(Rn) coincides with the classical homogeneous Triebel- Lizorkin space Lizorkin space F s p,q(Rn) for s ∈ (0, 1), p ∈ (n/(n + s), ∞) and q ∈ (n/(n + p,q(Rn) coincides with the classical homogeneous Besov s), ∞]. Similarly, space Bs p,q(Rn) for s ∈ (0, 1), p ∈ (n/(n + s), ∞) and q ∈ (0, ∞] by [24, Thm 1.2]. For the definitions of F s p,q(Rn), see [29]. p,q(Rn) and Bs N s If X supports a (weak) (1, p)-Poincar´e inequality with p ∈ (1, ∞), then for p,q(X) are trivial, that is, they contain all q ∈ (0, ∞), the spaces M 1 p,q(X) and N 1 only constant functions, see [7, Thm 4.1]. Lemma 4.1 ([7]). Let s ∈ (0, ∞) and p ∈ (0, Q/s). Then for every ε, ε′ ∈ (0, s) with ε < ε′ there exists a constant C > 0 such that for all measurable functions u with (gj) ∈ Ds(u), x ∈ X and k ∈ Z, (4.1) inf c∈R(cid:18)ZB(x,2−k) ≤ C2−kε′ Xj≥k−2 u(y) − cp∗(ε) dµ(y)(cid:19)1/p∗(ε) 2−j(s−ε′)(cid:18)ZB(x,2−k+1) gp j dµ(cid:19)1/p , where p∗(ε) = Qp/(Q − εp). If p ≥ Q/(Q + ε), then (4.1) implies that (4.2) ZB(x,2−k) u − uB(x,2−k) dµ ≤ C2−kε′ Xj≥k−2 2−j(s−ε′)(cid:18)ZB(x,2−k+1) gp j dµ(cid:19)1/p . We are now ready to state and prove our main results. Theorem 4.2 below gives a formula for an (s + α)-Haj lasz gradient of M∗ α in terms of an s-Haj lasz gradient of u. This easily implies the desired boundedness results for M∗ α in homogeneous Haj lasz -- Besov and Haj lasz -- Triebel -- Lizorkin spaces. For related results concerning Riesz potentials in the metric setting, see [31] Theorem 4.2. Assume that M∗ α u 6≡ ∞ and that (gk) ∈ Ds(u). Let 0 < s + α < 1, 0 < δ < 1 − s − α, 0 < ε < ε′ < s and t ≥ Q/(Q + ε). Then there is a constant C > 0, indepent of u and (gk), such that (C gk), where (4.3) gk = k Xj=−∞ ∞ 2(j−k)δ(cid:0)M gt j(cid:1)1/t + Xj=k−7 2(k−j)(s−ε′)(cid:0)M gt j(cid:1)1/t , is a fractional (s + α)-Haj lasz gradient of M∗ α u. 10 TONI HEIKKINEN AND HELI TUOMINEN Proof. Let k ∈ Z and let x, y ∈ X such that 2−k−1 ≤ d(x, y) < 2−k. We will show that uα r (x) − uα r (y) ≤ Cd(x, y)s+α(gk(x) + gk(y)), where C is independent of r and k. Assume first that d(x, y) > r. Then ur(x) − ur(y) ≤ ur(x) − uB(x,2−k+4) + ur(y) − uB(x,2−k+4) ≤ Xi∈Ix +Xi∈Iy ϕi(x)uB(xi,3r) − uB(x,2−k+4) ϕi(y)uB(xi,3r) − uB(x,2−k+4), where Ix is a set of indices i for which x belongs to B(xi, 6r) and Iy the corresponding set for y. Let m ∈ Z be such that 2−m−1 < 9r ≤ 2−m. Since r < d(x, y) < 2−k, it follows that m ≥ k − 4. If i ∈ Ix, we obtain uB(xi,3r) − uB(x,2−k+4) ≤ uB(xi,3r) − uB(x,2−m) + m−1 Xl=k−4 uB(x,2−l) − uB(x,2−l−1) ≤ C m Xl=k−4 ZB(x,2−l) u − uB(x,2−l) dµ and hence Poincar´e inequality (4.2) implies that uB(xi,3r) − uB(x,2−k+4) ≤ C 2−lε′ ∞ Xl=k−4 ∞ ∞ Xj=l−2 j+2 2−lε′ = C Xj=k−6 j(x)(cid:1)1/t 2−j(s−ε′)(cid:0)M gt j(x)(cid:1)1/t Xl=k−4 j(x)(cid:1)1/t 2−j(s−ε′)(cid:0)M gt j(x)(cid:1)1/t 2(k−j)(s−ε′)(cid:0)M gt 2−j(s−ε′)(cid:0)M gt Xj=k−6 Xj=k−6 ≤ C2−ksgk(x). ≤ C2−kε′ = C2−ks ∞ ∞ Similarly, if i ∈ Iy, then uB(xi,3r) − uB(x,2−k+4) ≤ uB(xi,3r) − uB(y,2−m) + m−1 Xl=k−4 uB(y,2−l) − uB(y,2−l−1) + uB(y,2−k+5) − uB(x,2−k+4) ≤ C m Xl=k−5 ZB(y,2−l) u − uB(y,2−l) dµ, THE FRACTIONAL MAXIMAL OPERATOR ON BESOV AND T -- L SPACES 11 which implies that It follows that uB(xi,3r) − uB(x,2−k+4) ≤ C2−ksgk(y). uα r (x) − uα r (y) ≤ Crα2−ks(gk(x) + gk(y)) ≤ Cd(x, y)s+α(gk(x) + gk(y)). Suppose then that d(x, y) ≤ r. Let Ixy be a set of indices i for which x or y belongs to B(xi, 6r). Let l be such that 2−l−1 < 10r ≤ 2−l. Using the doubling condition, the properties of the functions ϕi, the fact that there are bounded number of indices in Ixy and Poincar´e inequality (4.2), we have that (4.4) uα r (x) − uα r (y) ≤ rα ∞ Xi=1 ϕi(x) − ϕi(y))uB(xi,3r) − uB(x,2−l) ≤ Cd(x, y)rα−12−lε′ ∞ Xj=l−2 2−j(s−ε′)(cid:0)M gt j(x)(cid:1)1/t . Using the assumptions 0 < δ < 1 − α − s, r ≥ d(x, y) and d(x, y) < 2−k, we have that d(x, y)rα−12−lε′ ≤ Cd(x, y)rs+α+δ−1 2l(s−ε′+δ) ≤ Cd(x, y)s+α+δ 2l(s−ε′+δ) ≤ Cd(x, y)α+s2(l−k)δ+l(s−ε′). This together with (4.4) implies that uα r (x) − uα r (y) ≤ Cd(x, y)s+α ∞ Xj=l−2 2(l−k)δ+(l−j)(s−ε′)(cid:0)M gt j(x)(cid:1)1/t . By splitting the sum in two parts and using the estimates l ≤ j + 2 and l ≤ k, we obtain ∞ = k−1 Xj=l−2 Xj=l−2 ≤ C(cid:16) 2(l−k)δ+(l−j)(s−ε′)(cid:0)M gt 2(l−k)δ+(l−j)(s−ε′)(cid:0)M gt Xj=−∞ 2(j−k)δ(cid:0)M gt j(x)(cid:1)1/t j(x)(cid:1)1/t + Xj=k j(x)(cid:1)1/t + k−1 ∞ ∞ Xj=k 2(l−k)δ+(l−j)(s−ε′)(cid:0)M gt j(x)(cid:1)1/t(cid:17), 2(k−j)(s−ε′)(cid:0)M gt j(x)(cid:1)1/t which implies the claim for uα proof of Theorem 3.2. r . The claim for M∗ α u follows similarly as in the (cid:3) Theorem 4.3. Let 0 < s + α < 1 and Q/(Q + s) < p, q < ∞. Then there exists a constant C > 0 such that k M∗ α uk M s+α p,q (X) ≤ Ckuk M s p,q(X) for all u ∈ M s p,q(X) with M∗ α u 6≡ ∞. 12 TONI HEIKKINEN AND HELI TUOMINEN 2(1 − (s + α)), ε = 1 Proof. Let δ = 1 2(ε + s), where r = min{p, q}, and let t = Q/(Q + ε). Then 0 < ε < ε′ < s and Q/(Q + s) < t < min{p, q}. By Theorem 4.2, (C gk) defined by (4.3) is a fractional (s + α)- Haj lasz gradient of M∗ 2 max{s, s + Q−Qr r }, ε′ = 1 It suffices to show that (gk) ∈ Lp(X, lq). We estimate the Lp(X, lq) norm of α u. k Xj=−∞ (cid:16) 2(j−k)δ(cid:0)M gt j(cid:1)1/t(cid:17)k∈Z , the other part can be estimated similarly. If q ≥ 1, we have, by the Holder inequality, that Xk∈Z(cid:16) k Xj=−∞ 2(j−k)δ(cid:0) M gt j(cid:1)1/t(cid:17)q k j(cid:1)q/t ∞ Xj=−∞ ≤ CXk∈Z ≤ CXj∈Z (cid:0)M gt ≤ CXj∈Z (cid:0)M gt 2(j−k)δ(cid:0)M gt j(cid:1)q/t j(cid:1)q/t Xk=j . 2(j−k)δ If q < 1, we obtain the same estimate by using the elementary inequality (Pj aj)q ≤ Pj aq j for aj ≥ 0. By the Fefferman -- Stein vector valued maximal function theorem from [5] (for a metric space version, see for example [27] or [8]), we obtain now the desired estimate k Xj=−∞ (cid:13)(cid:13)(cid:13)(cid:16) 2(j−k)δ(cid:0)M gt j(cid:1)1/t(cid:17)k∈Z(cid:13)(cid:13)(cid:13)Lp(X, lq) 1/t Lp/t(X, lq/t) ≤ C(cid:13)(cid:13)(cid:0)M gt k(cid:1)k∈Z(cid:13)(cid:13) ≤ Ck(gt k)k∈Zk1/t Lp/t(X, lq/t) = Ck(gk)k∈ZkLp(X, lq). (cid:3) Theorem 4.4. Let 0 < s + α < 1, Q/(Q + s) < p < ∞ and 0 < q < ∞. Then there exists a constant C > 0 such that k M∗ α uk N s+α p,q (X) ≤ Ckuk N s p,q(X) for all u ∈ N s p,q(X) with M∗ α u 6≡ ∞. 2(1 − (s + α)), ε = 1 Proof. Let δ = 1 2(ε + s), and let t = Q/(Q + ε). Then 0 < ε < ε′ < s and Q/(Q + s) < t < p. Then (C gk) defined by (4.3) is a fractional (s + α)-Haj lasz gradient of M∗ α u by Theorem 4.2. 2 max{s, s + Q−Qp p }, ε′ = 1 THE FRACTIONAL MAXIMAL OPERATOR ON BESOV AND T -- L SPACES 13 It suffices to show that k(gk)klq(Lp(X)) ≤ Ck(gk)klq(Lp(X)). By the Hardy -- Littlewood maximal theorem, k Xj=−∞ (cid:13)(cid:13)(cid:13) 2(j−k)δ(cid:0)M gt j(cid:1)1/t(cid:13)(cid:13)(cid:13)Lp(X) ≤ ≤ k Xj=−∞ Xj=−∞ k 2(j−k)δ(cid:13)(cid:13)(cid:0)M gt j(cid:1)1/t(cid:13)(cid:13)Lp(X) 2(j−k)δkgjkLp(X). If q ≥ 1, we have by the Holder inequality, k Xj=−∞ kgjkq k Xj=−∞ Xk∈Z(cid:16) 2(j−k)δkgjkLp(X)(cid:17)q ≤ CXk∈Z ≤ CXj∈Z ≤ CXj∈Z If q < 1, we use the inequality (Pj aj)q ≤ Pj aq ity. The second part of (gk) can be estimated similarly. 2(j−k)δkgjkq Lp(X) Lp(X) ∞ Xk=j 2(j−k)δ kgjkq Lp(X). j instead of the Holder inequal- (cid:3) Theorems 4.3, 4.4 and the Hardy -- Littlewood maximal theorem imply the following results for the discrete maximal operator. Theorem 4.5. Let 0 < s < 1. a) If Q/(Q + s) < p, q < ∞, then there exist a constant C > 0 such that k M∗ uk M s p,q(X) ≤ Ckuk M s p,q(X), whenever u ∈ M s p,q(X) and M∗ u 6≡ ∞. b) If 1 < p, q < ∞, then there exist a constant C > 0 such that k M∗ ukM s p,q(X) ≤ CkukM s p,q(X), for all u ∈ M s p,q(X). Theorem 4.6. Let 0 < s < 1. a) If Q/(Q + s) < p < ∞ and 0 < q < ∞, there exist a constant C > 0 such that k M∗ uk N s p,q(X) ≤ Ckuk N s p,q(X) for all u ∈ N s p,q(X) with M∗ u 6≡ ∞. b) If 1 < p < ∞ and 0 < q < ∞, there exist a constant C > 0 such that k M∗ ukN s p,q(X) ≤ CkukN s p,q(X) for all u ∈ N s p,q(X). 14 TONI HEIKKINEN AND HELI TUOMINEN References [1] D. Aalto and J. Kinnunen, The discrete maximal operator in metric spaces, J. Anal. Math. 111 (2010), 369 -- 390. [2] J. M. Aldaz and J. P´erez L´azaro, Functions of bounded variation, the derivative of the one dimensional maximal function, and applications to inequalities, Trans. Amer. Math. Soc. 359 (2007), 2443-2461. [3] S.M. Buckley, Is the maximal function of a Lipschitz function continuous?, Ann. Acad. Sci. Fenn. Math. 24 (1999), 519 -- 528. [4] D. Edmunds, V. Kokilashvili, A. Meskhi, Bounded and Compact Integral Operators, Mathematics and its Applications, vol. 543, Kluwer Academic Publishers, Dordrecht, Boston, London, 2002. [5] C. Fefferman and E. M. Stein, Some maximal inequalities, Amer. J. Math. 93 (1971), 107 -- 115. [6] I. Genebashvili, A. Gogatishvili, V. Kokilashvili and M. Krbec, Weight Theory for Inte- gral Transforms on Spaces of Homogeneous Type, Addison Wesley Longman Limited, 1998. [7] A. Gogatishvili, P. Koskela and Y. Zhou, Characterizations of Besov and Triebel -- Lizorkin Spaces on Metric Measure Spaces, to appear in Forum math. [8] L. Grafakos, L. Liu and D. Yang, Vector-valued singular integrals and maximal func- tions on spaces of homogeneous type, Math. Scand. 104 (2009), 296 -- 310. [9] P. Haj lasz, Sobolev spaces on an arbitrary metric space, Potential Anal. 5 (1996), 403 -- 415. [10] P. Hajlasz, Sobolev spaces on metric-measure spaces, (Heat kernels and analysis on manifolds, graphs, and metric spaces (Paris, 2002)), 173 -- 218, Contemp. Math., 338, Amer. Math. Soc., Providence, RI, 2003. [11] P. Haj lasz and J. Maly, On approximate differentiability of the maximal function, Proc. of AMS., 138 (2010), no. 1, 165 -- 174. [12] P. Haj lasz and J. Onninen, On boundedness of maximal functions in Sobolev spaces, Ann. Acad. Sci. Fenn. Math. 29 (2004), 167 -- 176. [13] T. Heikkinen, J. Kinnunen, J. Nuutinen and H. Tuominen, Mapping properties of the discrete fractional maximal operator in metric measure spaces, to appear in Kyoto J. Math. [14] T. Heikkinen, P. Koskela and H. Tuominen, Sobolev-type spaces from generalized Poincar´e inequalities, Studia Math. 181 (2007), 1 -- 16. [15] T. Heikkinen, J. Lehrback, J. Nuutinen and H. Tuominen, Regularity of the fractional maximal function in metric measure spaces, preprint. [16] J. Hu, A note on Haj lasz-Sobolev spaces on fractals, J. Math. Anal. Appl. 280 (2003), 91-101. [17] J. Kinnunen, The Hardy-Littlewood maximal function of a Sobolev-function, Israel J.Math. 100 (1997), 117-124. [18] J. Kinnunen and V. Latvala, Lebesgue points for Sobolev functions on metric spaces, Rev. Mat. Iberoamericana 18 (2002), no. 3, 685 -- 700. [19] J. Kinnunen and P. Lindqvist, The derivative of the maximal function, J. Reine Angew. Math. 503 (1998), 161-167. [20] J. Kinnunen and E. Saksman, Regularity of the fractional maximal function, Bull. London Math. Soc. 35 (2003), no. 4, 529 -- 535. [21] J. Kinnunen and H. Tuominen, Pointwise behaviour of M 1,1 Sobolev functions, Math. Z. 257 (2007), no. 3, 613 -- 630. [22] S. Korry, A class of bounded operators on Sobolev spaces, Arch. Math. (Basel) 82 (2004), no. 1, 40 -- 50. THE FRACTIONAL MAXIMAL OPERATOR ON BESOV AND T -- L SPACES 15 [23] S. Korry, Boundedness of Hardy-Littlewood maximal operator in the framework of Lizorkin-Triebel spaces, Rev. Mat. Complut. 15 (2002), no. 2, 401 -- 416. [24] P. Koskela, D. Yang and Y. Zhou, Pointwise Characterizations of Besov and Triebel- Lizorkin Spaces and Quasiconformal Mappings, Adv. Math. 226 (2011), no. 4, 3579 -- 3621. [25] H. Luiro, Continuity of the Hardy-Littlewood maximal operator in Sobolev spaces, Proc. of AMS., 135 (2007), no. 1, 243 -- 251. [26] H. Luiro, On the regularity of the Hardy-Littlewood maximal operator on subdomains of Rn, Proc. Edinb. Math. Soc. (2) 53 (2010), no. 1, 211 -- 237. [27] Y. Sawano, Sharp estimates of the modified Hardy-Littlewood maximal operator on the nonhomogeneous space via covering lemmas, Hokkaido Math. J. 34 (2005), no. 2, 435 -- 458. [28] H. Tanaka, A remark on the derivative of the one-dimensional Hardy-Littlewood max- imal function, Bull. Austral. Math. Soc. 65 (2002), 253 -- 258. [29] H. Triebel, Theory of Function Spaces, Birkhauser Verlag, Basel, 1983. [30] D. Yang, New characterizations of Haj lasz-Sobolev spaces on metric spaces, Sci. China Ser. A 46 (2003), 675-689. [31] D. Yang, Riesz Potentials in Besov and Triebel -- Lizorkin Spaces over Spaces of Homo- geneous Type, Potential Anal. 19 (2003), 193 -- 210. T.H., Department of Mathematics, P.O. Box 11100, FI-00076 Aalto Uni- versity, Finland [email protected] H.T., Department of Mathematics and Statistics, P.O. Box 35, FI-40014 University of Jyvaskyla, Finland [email protected]
1709.04968	1	1709	2017-09-14T20:26:59	Toeplitz Quantization and Convexity	[ "math.FA", "math.SG" ]	Let $T^m_f $ be the Toeplitz quantization of a real $ C^{\infty}$ function defined on the sphere $ \mathbb{CP}(1)$. $T^m_f $ is therefore a Hermitian matrix with spectrum $\lambda^m= (\lambda_0^m,\ldots,\lambda_m^m)$. Schur's theorem says that the diagonal of a Hermitian matrix $A$ that has the same spectrum of $ T^m_f $ lies inside a finite dimensional convex set whose extreme points are $\{( \lambda_{\sigma(0)}^m,\ldots,\lambda_{\sigma(m)}^m)\}$, where $\sigma$ is any permutation of $(m+1)$ elements. In this paper, we prove that these convex sets "converge" to a huge convex set in $L^2([0,1])$ whose extreme points are $ f^\circ \phi$, where $ f^$ is the decreasing rearrangement of $ f$ and $ \phi $ ranges over the set of measure preserving transformations of the unit interval $ [0,1]$.	math.FA	math	Toeplitz Quantization and Convexity O. El Hadrami, Mohamed Lemine, Department of Mathematics King Khalid Universtiy, Abha, September 14, 2017 1 Abstract Let T m 0 , . . . , λm f be the Toeplitz quantization of a real C∞ function defined on the sphere CP(1). T m is therefore a Hermitian matrix with spec- f trum λm = (λm m). Schur's theorem says that the diagonal of a Hermitian matrix A that has the same spectrum of T m lies inside a fi- f nite dimensional convex set whose extreme points are {(λm σ(m))}, where σ is any permutation of (m + 1) elements. In this paper, we prove that these convex sets "converge" to a huge convex set in L2([0, 1]) whose extreme points are f ∗ ◦ φ, where f ∗ is the decreasing rearrangement of f and φ ranges over the set of measure preserving transformations of the unit interval [0, 1]. σ(0), . . . , λm 1 Introduction and background In their papers [3, 4], the authors have described similarities between the infinite and finite dimensional Lie groups. They have strengthen the idea that the set SDif f (CP(1)) of area preserving diffeomorphisms of the Riemann sphere is an infinite dimensional analog of SU (n) by proving an infinite version of the SU (n) Schur-Horn convexity theorem . In the present paper, we want to show that the convex sets in the two versions (finite and infinite) of Schur-Horn convexity theorem are related. In order to do that, we will show first that each permutation of (m + 1) letters determine a measure preserving transformation of the interval [0, 1]. Secondly we use the fact that the eigenvalues of the Toeplitz quantization of f determine the decreasing rearrangement of f . But since the two convex sets are defined by inequalities arising from the theory of majorization in Rm+1 developed in [11, 12] and its generalization to L1([0, 1]) by J.Ryff [13, 14, 15], we start by summarizing briefly here the main points. Majorization is a partial ordering in Rm+1 defined as it follows: 12010 Mathematics Subject Classification. 70G65, 37A15. Key words and phrases. Toeplitz quantization, measure preserving transformations. 1 For x ∈ Rm+1, let x∗ denote the vector obtained by rearranging the components of x in non-increasing order. We say that x majorizes y, written y ≺ x, if y∗ 0 + y∗ 1 + · · · + y∗ k ≤ x∗ 0 + x∗ 1 + · · · + x∗ k, 0 ≤ k ≤ m − 1 k=mXk=0 y∗ k = k=mXk=0 x∗ k Now we can state Schur' theorem for hermitian matrices. Theorem 1.1 Let λm = (λ0, · · · , λm) ∈ Rm+1 be the eigenvalues of a hermi- tian matrix A . Let diag(A) = (a00, · · · , amm) be the diagonal of A then λm majorizes diag(A). i.e diag(A) ≺ λm (1.1) Before going on, let us first fix some notations: 1. Let x = (x0, · · · , xm) ∈ Rm+1. Pm x is the orbit of x under the symmetric group of (m + 1) letters, i.e the collection of points (xσ(0), · · · , xσ(m)), where σ ranges over all (m + 1)! permutations. 2. For C ⊂ E where E is a vector space over R, co(C) denote the convex hull of C. Majorization and convexity are closely related as it is shown by the following theorem. Theorem 1.2 Let x ∈ Rm+1 • ( Rado's theorem) • y ≺ x if and only if Pi=m whose domain contains all the numbers xi, yi, 0 ≤ i ≤ m. {y ∈ Rm+1, y ≺ x} = co(Xm i=0 f (yi) ≤ Pi=m x) i=0 f (xi) for any convex function Schur-Horn' theorem can be therefore restated in the following terms : Theorem 1.3 (Schur-Horn's theorem) Let λm = (λ0, · · · , λm) ∈ Rm+1. Let Hλm be the set of all hermitian matrices whose spectrum is λm. Let pm : Hλm −→ Rm+1 be the map that picks the diagonal of a matrix. Then the image of the map pm is the convex set co(Pm λm). The convex set co(Pm λm) plays a very important role in symplectic geometry: It is the image of a moment map [1, 8]. The concepts of majorization is also extended to integrable functions in the 2 following sense. Let (X, µ) be a finite measurable space. For f measurable function on X, the distribution function of f is the function Ff defined by Ff (t) = µ({ω, f (ω) < t}). Let df (t) = µ(X) − Ff (t). Definition 1.1 The decreasing rearrangement of f is the function f ∗ defined by: f ∗(s) = inf{t > 0, df (t) < s}. For f, g ∈ L1((X, µ)), let f ∗ and g∗ be their decreasing rearrangement We say that f majorizes g ( written g ≺ f ) if 0 Z s Z 1 0 g∗(z) dz ≤Z s g∗(z) dz =Z 1 0 0 f ∗(z) dz, 0 ≤ s < 1 f ∗(z) dz To stay conform with the notation of [4], let X = CP(1) be the Riemann sphere , µ the measure defined by the the Fubini-Study symplectic form which in the local coordinate [1, w] is given by Ω = i (1 + ww)2 dw ∧ dw. Set w = r exp(iθ) and introduce the real variable z ∈ [0, 1[ by z = r2\(1 + r2). The symplectic form Ω becomes Ω = 2dz ∧ dθ. The infinite version of Schur theorem is Theorem 1.4 (Schur-type Theorems,[3]) Let L2(CP(1)) be the set of square integrable functions on the sphere. Let P : L2(CP(1)) −→ L2[0, 1] be the projec- tion P (f ) = f (z, θ)dθ. 1 2π Z 2π 0 Then f ∗ majorizes P (f ). i.e P (f ) ≺ f ∗ We have also the equivalent of Rado's theorem in ∈ L2([0, 1]). Theorem 1.5 Let f ∈ L2([0, 1]). The set Ω(f ) = {g ∈ L2([0, 1]), g ≺ f } is weakly compact and convex. Its set of extreme points is {f ∗ ◦ φ φ is a measure preserving transformation of [0, 1]}. The paper is organized in three sections: In §2, we have reviewed the topology of the set of measure preserving transformations of the unit interval[0, 1] and showed that the groups,Pm of permutations of (m+1) letters, can be identified 3 with a dense subset of the set of all invertible measure preserving transforma- tions of [0, 1] endowed with strong operator topology. In §3 we use Toeplitz quantization to show that the weak closure of the topo- logical lim sup of the sets co (Pm ·λm) is the set Ω(f ∗) = co ({f ∗ ◦ φ, φ measure preserving transformation of [0, 1]}). 2 measure preserving transformations of [0,1] The Lebesgue measure on the unit interval I = [0, 1] will be always denoted by · . A map φ from [0, 1] to itself is a measure preserving transformation if φ−1(A) = A, for Borel set A The set of all (non necessary invertible ) measure preserving transformation of the unit interval will be denoted Smeas(I). The invertible ones will be denoted by Imeas(I). Each S ∈ Smeas(I) determine a bounded linear operator PS on L2([0, 1]) by PS(f ) = f ◦ S. In this way, Smeas(I) can be identified to a subset of the set of bounded linear operators of L2[0, 1]) and the strong operator topology induces a topology on Smeas(I). Evidently, a sequence Sn converges to S in the strong operator topology if for every function f, f ◦ Sn converges to f ◦ S in L2([0, 1]). σ(k) by ordinary translation. k to the To state our first main result, we need to define dyadic permutations. Let I m k = [k\(m + 1), (k + 1)\(m + 1)) , k = 0, 1, · · · , m; m = 0, 1, · · · . invertible measure preserving transformation that sends the interval I m interval I m Let Pm be the group of permutations of (m + 1) letters. For σ ∈Pm, bσ is the We call bσ a permutation of rank m. In this way, we can identified the group of permutations Pm with a subgroup of Imeas(I). Halmos ([10]) shows that the set of all permutations of different rank is dense in Imeas(I) for the strong operator topology. Also Brown, in [6] has proved that Smeas(I) is the closure of Imeas(I) for the strong operator topology. We can summarize these results in the following theorem. Theorem 2.1 Let Pm be the symmetric group of (m + 1) letters. There exists a one to one group homomorphism Ψm : Pm −→ Imeas(I) that sends σ to bσ and if we identify Pm with Ψm(Pm), then SmPm is a dense set in Smeas(I) for the strong operator topology. 4 3 Toeplitz Quantization and Convexity 3.1 Toeplitz Quantization Consider the Riemann sphere CP(1) with the Fubini-Study symplectic form in the local coordinate [1, w] Ω = i (1 + ww)2 dw ∧ dw and the standard hyperplane bundle L. The tensor power L⊗m has (m + 1) lin- early independent sections which in the local coordinate w are just 1, w, ..., wm. The bundle L⊗m comes equipped with the Hermitian metric hs1, s2i(w) = 1 (1 + w2)m s1(w)s2(w). 2 be the space of square-integrable sections of L⊗m and Γm Now let Γm hol the space of holomorphic sections ( the span of 1, w, ..., wm).The orthogonal projection Γm 2 −→ Γm : Γm hol → Γm hol is denoted by P m.The Toeplitz quantization of f is the map T m hol defined by f T m f = P m ◦ Mf ◦ P m where Mf is multiplication by f . We refer the interested reader to [5] for a detailed exposition on Toeplitz quantization. The crucial result is the following theorem Theorem 3.1 (Distribution of the Eigenvalues of Toeplitz Quantization ) Let λm = (λm order and let Λm(s) be the real step function defined on the interval [0, 1[ by f arranged in non-increasing m) be the eigenvalues of T m 1 , · · · , λm 0 , λm Λm(cid:18)(cid:20) k m + 1 , k + 1 m + 1(cid:20)(cid:19) = λm k , 0 ≤ k ≤ m. (3.1) Then the sequence (Λm(s))m converges point-wise almost everywhere to the de- creasing rearrangement f ∗(s) of the function f . The proof of this Theorem is based on the the following theorem Theorem 3.2 (Szego-type Theorem,[9] p: 248) Given a smooth real val- f be Toeplitz quantization of f and let µm be its ued function f on CP(1), let T m spectral measure . Then tends weakly to a limiting measure as m tends µm m + 1 to infinity, this limiting measure being µ(φ) = 1 2πZCP(1) φ(f (x))dΩ, for φ ∈ C(R). 0 , λm 1 , · · · , λm m) are the eigenvalues of T m f ordered in non-increasing order, i.e if (λm then lim m−→+∞ k=mXk=0 φ(λk) m + 1 = 1 2πZCP(1) φ(f (x))dΩ. (3.2) 5 If we use the step function Λm defined be (3.1) then (3.2) can be written m−→+∞Z 1 lim 0 φ(Λm) = 1 2πZCP(1) φ(f (x))dΩ. (3.3) But since f and f ∗ are equi-measurable, we have 1 2π ZCP(1) (φ ◦ f )dΩ =Z 1 0 φ ◦ f ∗(t)dt and (3.3) becomes m7→+∞Z 1 lim 0 φ(Λm)(t) dt =Z 1 0 φ(f ∗)(t) dt. (3.4) where Λm is defined by (3.1). Relation ( 3.4 ) is equivalent to: the sequence of step functions Λm converges in distribution to the real function f ∗. (One can see [7] for more details on convergence in distribution.) In general convergence in distribution does not imply convergence point-wise. Nevertheless, there exists another sequence gn with the same distribution as fn that converges point-wise to a function g, that has the same distribution of f . That is the content of Skorokhod's Theorem. 6 Theorem 3.3 (Skorokhod) Let (X, Σ, µ) be a finite measure space,and f, fn : X → R be a sequence of measurable functions such that fn converge in distribu- tion to f . Then on the Lebesgue measure space (I, B, ·), where I = (0, µ(X), B is the Borel σ-algebra of I, and · is the Lebesgue measure, there exists measur- able functions gn, g : I → R such that gn(t) → g(t) a.e[·], and µ({ω : fn(ω) < x}) = {t : gn(t) < x} µ({ω : f (ω) < x}) = {t : g(t) < x} x ∈ R, n ≥ 1. Let Fn and F be the distribution functions of fn and f . We can take gn and g to be just the generalized inverse of Fn and F : gn(t) = inf{x : Fn(x) > t}, g(t) = inf{x : F (x) > t}, 0 < t < 1. It easily seen then that gn(1 − t) = g∗ n(t) = f ∗ n(t), g(1 − t) = g∗(t) = f ∗(t). ( [11] [p141-144] .) Applying Skorokhod's Theorem to the the sequence Λm, we deduce that the generalized inverses of the functions Λm converges point-wise almost everywhere to the generalized inverse of the function f . Consequently , the sequence of the decreasing rearrangements of Λm converges to the decreasing rearrangement f ∗ of f . lim m7−→+∞ (Λm)∗ (s) = f ∗(s). But since Λm(s) is a decreasing function, (Λm)∗ (s) = Λm(s), and therefore we have and that ends the proof of the theorem. lim m7−→+∞ Λm(s) = f ∗(s) 3.2 Convexity Let pm : Rm+1 → L2([0, 1]) be the map that associates to the point (a0, a1, · · · , am) k the step function pm(a0, a1, · · · , am) = Pk=m val (cid:20) pm(co(Pm ·λm)). m + 1(cid:20) and χIm m + 1 k + 1 , Now we are ready to state the main result about convexity. k=0 akχIm k , where I m k is the inter- is the characteristic function of I m k . Set Em = k Theorem 3.4 Let f ∈ C∞(CP(1)). Let f ∗ be the decreasing rearrangement of f . Let T m m) be the f . Let Ω(f ∗) = {g ∈ L2([0, 1], g ≺ f ∗}. Then the closed convex eigenvalues of T m hull of the set Smeas(I) · f ∗ = {f ∗ ◦ φ φ ∈ Smeas(I)} is the weak closure of f be the Toeplitz quantization of f and let λm = (λm 0 , · · · , λm the topological lim sup of the convex sets Em = pm(co(Pm ·λm)). 7 We recall the definition of the closed limit. Definition 3.1 ( [2], p. 114) Let {Em} be a sequence of subsets of a topolog- ical space X. Then, a point x in X belongs to the topological lim sup of Em, denoted LsEm, if for every neighborhood V of x there are infinitely many m with V ∩ Em 6= ∅. Clearly, LsEm is a closed sets and moreover LsEm = m=∞\m=1 m=∞[k=m Ek. In our case, X = L2([0, 1]) and Em = pm(co(Pm ·λm)) . Since L2([0, 1], k·k2) is a normed vector space, every point of L2([0, 1] has a countable basis of neighborhood and LsEm can be characterized in terms of sequences: g ∈ LsEm if and only if there exists a subsequence (gmk )k such that gmk ∈ Emk and gmk converges to g ∈ L2([0, 1]). Now we claim (A) Smeas(I) · f ∗ ⊂ LsEm. (B) Ω(f ∗) ⊂ LsEm weak ⊂ Ω(f ∗). Proof of( A): Let f ∗ ◦ φ ∈ Smeas(I) · f ∗. The set of dyadic permutations is a countable set and from theorem (2.1) is dense in Smeas(I) for the strong operator topology; therefore there exists a sequence (σn)n of permutations that converges to φ , i.e ∀g ∈ L2([0, 1]), ∀ǫ > 0, ∃N0, ∀n ≥ N0, kg ◦ σn − g ◦ φk2 < ǫ 2 . In particular for g = Λm, we have, ∀m, ∀ǫ > 0, ∃N0, ∀n ≥ N0, kΛm ◦ σn − Λm ◦ φk2 < ǫ 2 . (3.5) Now since the sequence Λm converges almost everywhere to f ∗ and ∀m, ∀x ∈ [0, 1], Λm(x) ≤ kf ∗k∞, , it converges also to f ∗ in L2([0, 1]), ie ∀ǫ > 0, ∃M0, ∀m ≥ M0, kΛm ◦ φ − f ∗ ◦ φk2 < ǫ 2 . (3.6) We conclude then from (3.5) and (3.6) that ∀ǫ > 0, ∃M0, ∃N0, ∀n ≥ N0, ∀m ≥ M0 kΛm ◦ σn − f ∗ ◦ φk2 ≤ kΛm ◦ σn − Λm ◦ φk2 + kΛm ◦ φ − f ∗ ◦ φk2 ≤ ǫ. (3.7) Now if σn is of order kn, and if we let m = kn in (3.7) we get 8 ∀ǫ > 0, ∃M0, ∃N0 for every n ≥ N0 such that kn ≥ M0, we have kΛkn ◦ σn − f ∗ ◦ φk2 ≤ ǫ. Since Λkn ◦ σn ∈ Ekn , we conclude therefore that f ∗ ◦ φ ∈ LsEm . Proof of (B): It is shown in [ [3],p 523] that Ω(f ∗) = {g ∈ L2([0, 1]), g ≺ f ∗} is weakly compact and convex. Its extreme points are precisely the elements f ∗ ◦ φ, φ ∈ Smeas(I). It is also indicated in [ [15], p1030] that the set of extreme points is dense in Ω(f ∗) for the weak topology. But we have just seen in part (A) of our claim that Smeas(I) · f ∗ ⊂ LsEm . Taking the closure in the weak topology we get Ω(f ∗) = Smeas(I) · f ∗ weak ⊂ LsEm weak . weak It remains to show that LsEm Let g ∈ LsEm. Then by definition, there exists a subsequence (gmk )k such that gmk ∈ Emk , and (gmk )k converges in L2([0, 1]) to g. Now by Rado's theorem (1.2) , we have ⊂ Ω(f ∗). gmk ∈ Emk ⇐⇒ gmk ≺ Λmk But the sequence (gmk )k being convergent in L2([0, 1])), we can extract a sub- l(cid:17)l sequent (cid:16)gmk that converges pointwise to g a.e. We have then gmk l ≺ Λmk l . And taking limit (simple convergence) of both sides, we get g ≺ f ∗ i.e LsEm ⊂ Ω(f ∗). Taking the weak closure of both sets we conclude therefore that Ω(f ∗) = LsEm weak . Our next goal is to compare the co-adjoint orbits of SU (m + 1) and the co- adjoint orbits of SDiff(CP(1), the group of area preserving diffeomorphisms of the sphere. Acknowledgment I would like thank Professor H. Flaschka for valuable discussions. References [1] M.F.Atiyah, Convexity and commuting Hamiltonians, Bull. London Math. Soc 14, 1-15, (1982) [2] C. D. Aliprantis, K. C. Border, Infinite Dimensional Analysis, a Hitch- hiker's Guide, 3rd edition, Springer (2005). 9 [3] A.Bloch, H.Flaschka, T.Ratiu, A Shur-Horn-Kostant Convexity Theorem for the Diffeomorphism Group of the Annulus, Inv.Math. 113, 511-529, (1993). [4] A. Bloch, M. El Hadrami, H. Flaschka, T. Ratiu, Maximal Tori Of Some Symplectomorphism Groups And Applications To Convexity, Proceedings of Ascona Meeting, June 1996. (D. Sternheimer, J. Rawnsley, S. Gutt, eds.), Mathematical Physics Studies 20, Kluwer Academic Publishers, 201-222, (1997). [5] M. Bordemann, E. Meinrenken and M. Schlichenmaier, Toeplitz Quantiza- tion Of Kahler Manifolds And gl(N ), N → +∞ Limits., Commun. Math. Phys. 165 , 281-296,(1994). [6] J. R. Brown, Approximation theorems for Markov Operators, Pac.J.Maths.16, 13-23, (1966). [7] Richard Durrett Probability:Theory And Examples, Wadsworth & brooks/cole Advanced Books & Software, 1991. [8] V.Guillemin and S. Sternberg, Convexity properties of the momentum map- ping I, Invent. Math, 67, 491-513,(1982). [9] Victor Guillemin, Some Classical Theorems In Spectral Theory Revisited, Seminar on Singulariries of Solutions of Linear Partial Differential Equa- tions (L. H ormander ed) Annals of Mathematics Studies, 91, Princeton University Press and University of Tokyo Press, [1979], 219- 259. [10] P. Halmos Approximation Theorems for Measure Preserving Transforma- tions, Trans.AMS 55, 1-18, (1944). [11] M. M. Rao Measure Theory And Integration, 2nd Edition, Marcel Dekker Inc, (2004). [12] Jean-Michel Rakotoson R´earrangement Relatif, Instrument d'Estimation dans les Probl`emes aux Limites, Math´ematiques & Ap- plications, 64, Springer, (2008). Un [13] J. V. Ryff, On the Representation of Doubly Stochastic Operators, Pac.J.Maths.13, 1379-1386, (1963). [14] J. V. Ryff, Orbits of L1 Functions Under Doubly Stochastic Transforma- tions, Trans.AMS 117, 92-100, (1965). [15] J. V. Ryff, Extreme points of some convex subsets of L1(0, 1), Proc. AMS 18 , 1026-1034, (1967). King Khaled University, Abha, Saudi Arabia. E-mail address: [email protected] 10
1704.01360	2	1704	2018-04-28T14:14:59	Two modified proximal point algorithms in geodesic spaces with curvature bounded above	[ "math.FA", "math.OC" ]	We obtain existence and convergence theorems on two variants of the proximal point algorithm for proper lower semicontinuous convex functions in complete geodesic spaces with curvature bounded above.	math.FA	math	TWO MODIFIED PROXIMAL POINT ALGORITHMS IN GEODESIC SPACES WITH CURVATURE BOUNDED ABOVE YASUNORI KIMURA AND FUMIAKI KOHSAKA Abstract. We obtain existence and convergence theorems for two variants of the proximal point algorithm involving proper lower semicontinuous convex functions in complete geodesic spaces with curvature bounded above. 1. Introduction The aim of this paper is to study the asymptotic behavior of sequences generated by two variants of the proximal point algorithm for proper lower semicontinuous convex functions in admissible complete CAT(1) spaces. We focus not only on the convergence of the sequences to minimizers of functions but also on the equivalence between their boundedness and the existence of minimizers. Applications to convex minimization problems in complete CAT(κ) spaces with a positive real number κ are also included. The proximal point algorithm introduced by Martinet [31] and studied more gen- erally by Rockafellar [36] is an iterative method for finding zero points of maximal monotone operators in Hilbert spaces. Bruck and Reich [12] also obtained some con- vergence theorems for m-accretive operators in Banach spaces. It is known that this algorithm has a wide range of applications including convex minimization problems, variational inequality problems, minimax problems, and equilibrium problems. For a proper lower semicontinuous convex function f of a Hilbert space H into ]−∞,∞], the proximal point algorithm generates a sequence {xn} by x1 ∈ H and (1.1) where {λn} is a sequence of positive real numbers and Jλnf is the resolvent of λnf defined by xn+1 = Jλn f xn (n = 1, 2, . . . ), Jλnf x = argmin y∈H (cid:26)f (y) + 1 2λn ky − xk2(cid:27) for all n ∈ N and x ∈ H. See also [8, 39] for more details on convex analysis in Hilbert spaces. The celebrated theorem by Rockafellar [36, Theorem 1] implies the following existence and weak convergence theorems on the sequence {xn} defined by (1.1). If inf n λn > 0, then {xn} is bounded if and only if the set argminH f of all min- imizers of f is nonempty. Further, in this case, {xn} is weakly convergent to an element of argminH f . Br´ezis and Lions [10, Th´eor`eme 9] also showed that {xn} is weakly convergent to an element of argminH f if P∞n=1 λn = ∞ and argminH f 2010 Mathematics Subject Classification. Primary: 52A41, 90C25; Secondary: 47H10, 47J05. Key words and phrases. CAT(1) space, convex function, fixed point, geodesic space, minimizer, proximal point algorithm, resolvent. 1 2 Y. KIMURA AND F. KOHSAKA is nonempty. Later, Guler [18, Corollary 5.1] and Bauschke, Matouskov´a, and Re- ich [9, Corollary 7.1] found the counterexamples to the strong convergence of {xn}. By assuming the so-called convergence condition, Nevanlinna and Reich [33, Theo- rem 2] obtained a strong convergence theorem for m-accretive operators in Banach spaces. In 2000, Solodov and Svaiter [38] and Kamimura and Takahashi [20] pro- posed two different types of strongly convergent proximal-type algorithms in Hilbert spaces. On the other hand, a CAT(κ) space is a geodesic metric space such that every geodesic triangle in it satisfies the CAT(κ) inequality, where κ is a real number. A complete CAT(0) space is particularly called an Hadamard space. Since the concept of CAT(κ) spaces includes several fundamental spaces, the fixed point theory and the convex optimization theory in such spaces have been increasingly important. See, for instance, [5, 11, 35] for more details in this direction. In the 1990s, Jost [19] and Mayer [32] generalized the concept of resolvents of convex functions to Hadamard spaces. According to [5, Section 2.2], [19, Lemma 2], and [32, Section 1.3], if f is a proper lower semicontinuous convex function of an Hadamard space X into ]−∞,∞], then the resolvent Jf of f given by (1.2) Jf x = argmin y∈X (cid:26)f (y) + 1 2 d(y, x)2(cid:27) for all x ∈ X is a single-valued nonexpansive mapping of X into itself. In this case, the set F (Jf ) of all fixed points of Jf coincides with argminX f . See also [5] on convex analysis in Hadamard spaces. In 2013, Bac´ak [4] generalized the classical theorem by Br´ezis and Lions [10, Th´eor`eme 9] to Hadamard spaces. Some related asymptotic results were found by Ariza-Ruiz, Leu¸stean, and L´opez-Acedo [3, Corollary 6.6] and Bac´ak and Reich [6, Proposition 1.5]. Theorem 1.1 ([4, Theorem 1.4]). Let X be an Hadamard space, f a proper lower semicontinuous convex function of X into ]−∞,∞], Jηf the resolvent of ηf for all η > 0, and {xn} a sequence defined by x1 ∈ X and (1.1), where {λn} is a sequence of positive real numbers such that P∞n=1 λn = ∞. If argminX f is nonempty, then {xn} is ∆-convergent to an element of argminX f . Motivated by [2, 4, 20], the authors [23] recently obtained the following exis- tence and convergence theorems for two variants of the proximal point algorithm in Hadamard spaces. The algorithms (1.3) and (1.4) were originally introduced by Eckstein and Bertsekas [14] and Kamimura and Takahashi [20] for maximal monotone operators in Hilbert spaces, respectively. Theorem 1.2 ([23, Theorem 4.2]). Let X, f , {Jηf}η>0 be the same as in Theo- rem 1.1 and {xn} a sequence defined by x1 ∈ X and (1.3) (n = 1, 2, . . . ), xn+1 = αnxn ⊕ (1 − αn)Jλn f xn where {αn} is a sequence in [0, 1[ and {λn} is a sequence of positive real numbers such that P∞n=1(1 − αn)λn = ∞. Then the following hold. are ∆-convergent to an element x∞ of argminX f . (i) The set argminX f is nonempty if and only if {Jλnf xn} is bounded; (ii) if argminX f is nonempty and supn αn < 1, then both {xn} and {Jλnf xn} TWO MODIFIED PROXIMAL POINT ALGORITHMS 3 Theorem 1.3 ([23, Theorem 5.1]). Let X, f , and {Jηf}η>0 be the same as in Theorem 1.1, v an element of X, and {yn} a sequence defined by y1 ∈ X and (1.4) (n = 1, 2, . . . ), yn+1 = αnv ⊕ (1 − αn)Jλnf yn where {αn} is a sequence in [0, 1] and {λn} is a sequence of positive real numbers such that limn λn = ∞. Then the following hold. (i) The set argminX f is nonempty if and only if {Jλnf yn} is bounded; (ii) if argminX f is nonempty, limn αn = 0, and P∞n=1 αn = ∞, then both {yn} and {Jλnf yn} are convergent to P v, where P denotes the metric projection of X onto argminX f . Theorem 1.4 ([23, Theorem 5.4]). Let X, f , and {Jηf}η>0 be the same as in Theorem 1.1, v an element of X, and {yn} a sequence defined by y1 ∈ X and (1.4), where {αn} is a sequence in ]0, 1] and {λn} is a sequence of positive real numbers such that αn = 0, lim n→∞ ∞ Xn=1 αn = ∞, and inf n λn > 0. If argminX f is nonempty, then both {yn} and {Jλnf yn} are convergent to P v, where P denotes the metric projection of X onto argminX f . In 2015, Ohta and P´alfia [34, Definition 4.1 and Lemma 4.2] showed that the resolvent Jf given by (1.2) is still well defined in a complete CAT(1) space such that diam(X) < π/2, where diam(X) denotes the diameter of X. Using this result, they [34, Theorem 5.1] obtained a ∆-convergence theorem on the proximal point algorithm in such spaces. It should be noted that the condition that diam(X) < π/2 for a complete CAT(1) space X corresponds to the boundedness condition for an Hadamard space. In fact, every sequence in a complete CAT(1) space X such that diam(X) < π/2 has a ∆-convergent subsequence. In 2016, the authors [22] introduced another type of resolvents of convex func- tions in CAT(1) spaces. For a given proper lower semicontinuous convex function f of an admissible complete CAT(1) space X into ]−∞,∞], they [22, Definition 4.3] defined the resolvent Rf of f by (1.5) Rf x = argmin y∈X (cid:8)f (y) + tan d(y, x) sin d(y, x)(cid:9) for all x ∈ X. Following [24], we say that a CAT(1) space X is admissible (1.6) d(w, w′) < π 2 for all w, w′ ∈ X. Recently, the authors [24] obtained the following result on the proximal point al- gorithm in CAT(1) spaces. We note that the ∆-convergence of {xn} in Theorem 1.5 was also found independently by Esp´ınola and Nicolae [16, Theorem 3.2]. Theorem 1.5 ([24, Theorems 1.1 and 1.2]). Let X be an admissible complete CAT(1) space, f a proper lower semicontinuous convex function of X into ]−∞,∞], Rηf the resolvent of ηf for all η > 0, and {xn} a sequence defined by x1 ∈ X and xn+1 = Rλnf xn (n = 1, 2, . . . ), 4 Y. KIMURA AND F. KOHSAKA where {λn} is a sequence of positive real numbers such that P∞n=1 λn = ∞. Then argminX f is nonempty if and only if π 2 d(y, xn) < inf y∈X lim sup n→∞ and sup d(xn+1, xn) < n π 2 . Further, in this case, {xn} is ∆-convergent to an element of argminX f . Motivated by the papers mentioned above, we study the asymptotic behavior of sequences generated by x1, y1, v ∈ X, (1.7) xn+1 = αnxn ⊕ (1 − αn)Rλnf xn (n = 1, 2, . . . ), and yn+1 = αnv ⊕ (1 − αn)Rλnf yn (1.8) where {αn} is a sequence in [0, 1], {λn} is a sequence of positive real numbers, X is an admissible complete CAT(1) space, f is a proper lower semicontinuous convex function of X into ]−∞,∞], and Rλnf is the resolvent of λnf for all n ∈ N given by (1.5). These algorithms correspond to (1.3) and (1.4) in Hadamard spaces, respectively. (n = 1, 2, . . . ), This paper is organized as follows. In Section 2, we recall some definitions and results needed in this paper. In Section 3, we obtain some fundamental proper- ties of resolvents of convex functions in complete CAT(1) spaces. In Sections 4 and 5, we study the asymptotic behavior of sequences generated by (1.7) and (1.8), respectively. The three main results in this paper, Theorems 4.1, 5.1, and 5.2 in admissible complete CAT(1) spaces, correspond to Theorems 1.2, 1.3, and 1.4 in Hadamard spaces, respectively. In Section 6, we deduce three corollaries of our results in complete CAT(κ) spaces with a positive real number κ. 2. Preliminaries Throughout this paper, we denote by N the set of all positive integers, R the set of all real numbers, ]−∞,∞] the set R∪{∞}, R2 the two dimensional Euclidean space with Euclidean metric ρR2 , S2 the unit sphere of the three dimensional Euclidean space R3 with the spherical metric ρS2, H a real Hilbert space with inner product h· , ·i and induced norm k·k, X a metric space with metric d, F (T ) the set of all fixed points of a mapping T of X into itself, and argminX f or argminy∈X f (y) the set of all minimizers of a function f of X into ]−∞,∞]. In the case where argminX f is a singleton {p}, we sometimes identify argminX f with the single point p. We need the following lemma. Lemma 2.1 ([40, Lemma 2.5]; see also [1, Lemma 2.3]). Let {sn} be a sequence of nonnegative real numbers, {αn} a sequence in [0, 1] such that P∞n=1 αn = ∞, and {tn} a sequence of real numbers such that lim supn tn ≤ 0. Suppose that (2.1) for all n ∈ N. Then limn sn = 0. sn+1 ≤ (1 − αn)sn + αntn Saejung and Yotkaew [37] found the following variant of Lemma 2.1. Later, Kimura and Saejung [25] filled in a slight gap in the original proof of this result. Although it was assumed in [25, 37] that αn < 1 for all n ∈ N, the proof in [25, Lemma 2.8] is also valid in the case below. TWO MODIFIED PROXIMAL POINT ALGORITHMS 5 Lemma 2.2 ([25, Lemma 2.8] and [37, Lemma 2.6]). Let {sn} be a sequence of nonnegative real numbers, {αn} a sequence in ]0, 1] such that P∞n=1 αn = ∞, and {tn} a sequence of real numbers. Suppose that (2.1) holds for all n ∈ N and that lim supi tni ≤ 0 whenever {ni} is an increasing sequence in N satisfying Then limn sn = 0. lim sup i→∞ (cid:0)sni − sni+1(cid:1) ≤ 0. Let κ be a nonnegative real number and Dκ the extended real number defined by Dκ = ∞ if κ = 0 and π/√κ if κ > 0. A metric space X is said to be Dκ-geodesic if for each x, y ∈ X with d(x, y) < Dκ, there exists a mapping c of [0, l] into X such that c(0) = x, c(l) = y, and d(cid:0)c(t1), c(t2)(cid:1) = t1 − t2 for all t1, t2 ∈ [0, l], where l = d(x, y). The mapping c is called a geodesic path from x to y. The image of c is denoted by [x, y]c and is called a geodesic segment between x and y. We denote by αx ⊕c (1 − α)y the point given by αx ⊕c (1 − α)y = c(cid:0)(1 − α)l(cid:1) for all α ∈ [0, 1]. A Dκ-geodesic metric space is also called a Dκ-geodesic space. An ∞-geodesic metric space is also called a geodesic metric space or a geodesic space. A subset F of a Dκ-geodesic space X such that d(w, w′) < Dκ for all w, w′ ∈ F is said to be convex if [x, y]c ⊂ F whenever x, y ∈ F and c is a geodesic path from x to y. Although [x, y]c and αx ⊕c (1 − α)y depend on the choice of a geodesic path c from x to y, we sometimes denote them simply by [x, y] and αx ⊕ (1 − α)y, respectively. They are determined uniquely if the space X is uniquely Dκ-geodesic, that is, for each x, y ∈ X with d(x, y) < Dκ, there exists a unique geodesic path from x to y. If H is a Hilbert space, then the unit sphere SH of H is a uniquely π-geodesic complete metric space with the spherical metric ρSH defined by for all x, y ∈ SH . For all distinct x, y ∈ SH with ρSH (x, y) < π, the unique geodesic path c from x to y is given by ρSH (x, y) = arccoshx, yi c(t) = (cos t)x + (sin t) · y − hx, yi x ky − hx, yi xk for all t ∈ [0, ρSH (x, y)]. The space (SH , ρSH ) is called a Hilbert sphere. See [7, 11, 17] for more details on Hilbert spheres. Let (Mκ, dκ) be the uniquely Dκ-geodesic space given by (Mκ, dκ) =((cid:0)R2, ρR2(cid:1) (cid:16)S2, 1 √κ ρS2(cid:17) (κ > 0). (κ = 0); If κ is a nonnegative real number, X is a Dκ-geodesic space, and x1, x2, x3 are points of X satisfying (2.2) then there exist ¯x1, ¯x2, ¯x3 ∈ Mκ such that d(x1, x2) + d(x2, x3) + d(x3, x1) < 2Dκ, d(xi, xj) = dκ(¯xi, ¯xj ) 6 Y. KIMURA AND F. KOHSAKA for all i, j ∈ {1, 2, 3}; see [11, Lemma 2.14 in Chapter I.2]. The two sets ∆ and ¯∆ given by ∆ = [x1, x2] ∪ [x2, x3] ∪ [x3, x1] and ¯∆ = [¯x1, ¯x2] ∪ [¯x2, ¯x3] ∪ [¯x3, ¯x1] are called a geodesic triangle with vertices x1, x2, x3 in X and a comparison triangle for ∆, respectively. A point ¯p ∈ ¯∆ is called a comparison point for p ∈ ∆ if p ∈ [xi, xj], ¯p ∈ [¯xi, ¯xj], and d(xi, p) = dκ(¯xi, ¯p) for some distinct i, j ∈ {1, 2, 3}. A metric space X is said to be a CAT(κ) space if it is Dκ-geodesic and the CAT(κ) inequality d(p, q) ≤ dκ(¯p, ¯q) holds whenever ∆ is a geodesic triangle with vertices x1, x2, x3 ∈ X satisfying (2.2), ¯∆ is a comparison triangle for ∆, and ¯p, ¯q ∈ ¯∆ are comparison points for p, q ∈ ∆, respectively. In this case, the space X is also uniquely Dκ-geodesic. Every CAT(κ) space is a CAT(κ′) space for all κ′ ∈ ]κ,∞[. A complete CAT(0) space is particularly called an Hadamard space. The class of Hadamard spaces includes nonempty closed convex subsets of Hilbert spaces, open unit balls of Hilbert spaces with hyperbolic metric, Hadamard manifolds, and complete R-trees. The class of complete CAT(1) spaces includes Hadamard spaces and Hilbert spheres with spherical metric. We say that a CAT(1) space X is admissible if (1.6) holds for all w, w′ ∈ X. If κ > 0, then (X, d) is a complete CAT(κ) space such that d(w, w′) < Dκ/2 for all w, w′ ∈ X if and only if (X,√κd) is an admissible complete CAT(1) space. See [5, 11, 13] for more details on CAT(κ) spaces. We know that if X is a CAT(1) space, x1, x2, x3 ∈ X satisfy (2.2) for κ = 1, and cos d(cid:0)αx1 ⊕ (1 − α)x2, x3(cid:1) ≥ α cos d(x1, x3) + (1 − α) cos d(x2, x3). We also know the following fundamental inequalities. Lemma 2.3 ([27, Corollary 2.2]). If X is a CAT(1) space, x1, x2, x3 ∈ X sat- isfy (2.2) for κ = 1, and α ∈ [0, 1], then (2.4) cos d(cid:0)αx1 ⊕ (1 − α)x2, x3(cid:1) sin d(x1, x2) ≥ cos d(x1, x3) sin(cid:0)αd(x1, x2)(cid:1) + cos d(x2, x3) sin(cid:0)(1 − α)d(x1, x2)(cid:1). Lemma 2.4 ([28, Lemma 3.1]). If X is an admissible CAT(1) space, x1, x2, x3 ∈ X, and α ∈ [0, 1], then α ∈ [0, 1], then (2.3) cos d(cid:0)αx1 ⊕ (1 − α)x2, x3(cid:1) ≥ (1 − β) cos d(x2, x3) + β · (2.5) where sin d(x1, x2) tan(cid:0) α cos d(x1, x3) , 2 d(x1, x2)(cid:1) + cos d(x1, x2) (x1 6= x2); (x1 = x2). sin(cid:0)(1 − α)d(x1, x2)(cid:1) sin d(x1, x2) 1 − α β =  The concept of ∆-convergence was originally introduced by Lim [30] in metric spaces. Later, Kirk and Panyanak [29] applied it to the study of geodesic spaces. TWO MODIFIED PROXIMAL POINT ALGORITHMS 7 Let X be a metric space and {xn} a sequence in X. The asymptotic center A(cid:0){xn}(cid:1) of {xn} is defined by A(cid:0){xn}(cid:1) =(cid:26)z ∈ X : lim sup n→∞ d(z, xn) = inf y∈X lim sup n→∞ d(y, xn)(cid:27) . The sequence {xn} is said to be ∆-convergent to p ∈ X if A(cid:0){xni}(cid:1) = {p} holds for each subsequence {xni} of {xn}. In this case, {xn} is bounded and its each subsequence is also ∆-convergent to p. If X is a nonempty closed convex subset of a Hilbert space, then the ∆-convergence coincides with the weak convergence. We denote by ω∆(cid:0){xn}(cid:1) the set of all points q ∈ X such that there exists a subsequence of {xn} which is ∆-convergent to q. Following [24], we say that a sequence {xn} in a CAT(1) space X is spherically bounded if inf y∈X We know the following lemmas. lim sup n→∞ d(y, xn) < π 2 . Lemma 2.5 ([15, Proposition 4.1 and Corollary 4.4]). Let X be a complete CAT(1) space and {xn} a spherically bounded sequence in X. Then A(cid:0){xn}(cid:1) is a singleton and {xn} has a ∆-convergent subsequence. Lemma 2.6 ([26, Proposition 3.1]). Let X be a complete CAT(1) space and {xn} a spherically bounded sequence in X. If {d(z, xn)} is convergent for each element z of ω∆(cid:0){xn}(cid:1), then {xn} is ∆-convergent to an element of X. Let X be an admissible CAT(1) space. A function f of X into ]−∞,∞] is said to be proper if f (a) ∈ R for some a ∈ X. It is also said to be convex if f(cid:0)αx ⊕ (1 − α)y(cid:1) ≤ αf (x) + (1 − α)f (y) whenever x, y ∈ X and α ∈ ]0, 1[. We denote by Γ0(X) the set of all proper lower semicontinuous convex functions of X into ]−∞,∞]. The set argminX f It follows from (2.3) that is obviously closed and convex for each f ∈ Γ0(X). − cos d(·, z) belongs to Γ0(X) for all z ∈ X. For a nonempty closed convex subset C of X, the indicator function iC for C, which is defined by iC(x) = 0 if x ∈ C and ∞ otherwise, belongs to Γ0(X). See [21, 41] on convex functions in CAT(1) spaces. A function f of X into ]−∞,∞] is said to be ∆-lower semicontinuous if f (p) ≤ lim inf n f (xn) whenever {xn} is a sequence in X which is ∆-convergent to p ∈ X. A function g of X into [−∞,∞[ is said to be concave if −g is convex. Let X be an admissible complete CAT(1) space and f an element of Γ0(X). It is known [22, Theorem 4.2] that for each x ∈ X, there exists a unique x ∈ X such that f (x) + tan d(x, x) sin d(x, x) = inf y∈X {f (y) + tan d(y, x) sin d(y, x)} . Following [22, Definition 4.3], we define the resolvent Rf of f by Rf x = x for all x ∈ X. In other words, Rf can be defined by (1.5) for all x ∈ X. If f is the indicator function iC for a nonempty closed convex subset C of X, then the 8 Y. KIMURA AND F. KOHSAKA resolvent Rf coincides with the metric projection PC of X onto C, that is, Rf x = argmin tan d(y, x) sin d(y, x) = argmin d(y, x) = PC x y∈C y∈C for all x ∈ X. mapping of X into itself, It is known [22, Theorems 4.2 and 4.6] that Rf is a well-defined and single-valued (2.6) and (2.7) F (Rf ) = argmin X f, x(1 + C 2 (cid:0)C 2 ≥ C 2 x(1 + C 2 y )Cy + C 2 y (1 + C 2 y ) cos d(Rf x, y) + C 2 x)Cx(cid:1) cos d(Rf x, Rf y) y (1 + C 2 x) cos d(Rf y, x) for all x, y ∈ X, where Cz the real number given by Cz = cos d(Rf z, z) for all z ∈ X. The following lemma was recently obtained in [24, Lemma 3.1 and Corollary 3.2]. The inequality (2.9) is a generalization of (2.7) and corresponds to [2, Lemma 3.1] in Banach spaces. Lemma 2.7 ([24, Lemma 3.1 and Corollary 3.2]). Let X be an admissible complete CAT(1) space, f an element of Γ0(X), Rηf the resolvent of ηf for all η > 0, and Cη,z the real number given by Cη,z = cos d(Rηf z, z) (2.8) for all η > 0 and z ∈ X. Then (2.9) λ,x(1 + C 2 (cid:0)λC 2 ≥ λC 2 µ,y)Cµ,y + µC 2 µ,y(1 + C 2 λ,x(1 + C 2 µ,y) cos d(Rλf x, y) + µC 2 λ,x)Cλ,x(cid:1) cos d(Rλf x, Rµf y) µ,y(1 + C 2 λ,x) cos d(Rµf y, x) holds for all x, y ∈ X and λ, µ > 0. Further, π 2 1 C 2 λ,x + 1!(cid:0)Cλ,x cos d(u, Rλf x) − cos d(u, x)(cid:1) ≥ λ(cid:0)f (Rλf x) − f (u)(cid:1) (2.10) and (2.11) hold for all x ∈ X, u ∈ argminX f , and λ > 0. cos d(Rλf x, x) cos d(u, Rλf x) ≥ cos d(u, x), We also know the following results. Lemma 2.8 ([22, Lemma 3.1]). If X is an admissible complete CAT(1) space, then every f ∈ Γ0(X) is ∆-lower semicontinuous. Theorem 2.9 ([24, Theorem 4.1]). Let X be an admissible complete CAT(1) space, {zn} a spherically bounded sequence in X, {βn} a sequence of positive real numbers such that P∞n=1 βn = ∞, and g the real function on X defined by n βk cos d(y, zk) (2.12) g(y) = lim inf n→∞ 1 l=1 βl Pn Xk=1 for all y ∈ X. Then g is a 1-Lipschitz continuous and concave function of X into [0, 1] which has a unique maximizer. TWO MODIFIED PROXIMAL POINT ALGORITHMS 9 It is obvious that if A is a nonempty bounded subset of R, I is a closed subset of R containing A, and f is a continuous and nondecreasing real function on I, then f (sup A) = sup f (A) and f (inf A) = inf f (A). Thus we obtain the following. Lemma 2.10. Let I be a nonempty closed subset of R, {tn} a bounded sequence in I, and f a continuous real function on I. Then the following hold. (i) If f is nondecreasing, then f (lim supn tn) = lim supn f (tn); (ii) if f is nonincreasing, then f (lim supn tn) = lim inf n f (tn). 3. Resolvents of convex functions in CAT(1) spaces In this section, we obtain three fundamental lemmas on the resolvents of convex functions in CAT(1) spaces. The following lemma corresponds to [2, Lemmas 3.5 and 3.6] in Banach spaces. Lemma 3.1. Let X be an admissible complete CAT(1) space, f an element of Γ0(X), Rηf the resolvent of ηf for all η > 0, {λn} a sequence of positive real numbers, p an element of X, and {xn} a sequence in X. Then the following hold. (i) If inf n λn > 0, A(cid:0){xn}(cid:1) = {p}, and limn d(Rλnf xn, xn) = 0, then p is an (ii) if limn λn = ∞, A(cid:0){Rλnf xn}(cid:1) = {p}, and supn d(Rλnf xn, xn) < π/2, then Proof. Let Cη,z be the real number in ]0, 1] given by (2.8) for all η > 0 and z ∈ X. It follows from (2.9) that p is an element of argminX f . element of argminX f ; (cid:0)λnC 2 ≥ λnC 2 λn,xn (1 + C 2 1,p) + C 2 1,p(1 + C 2 λn,xn(1 + C 2 1,p) cos d(Rλn f xn, p) + C 2 λn,xn )(cid:1) cos d(Rλnf xn, Rf p) 1,p(1 + C 2 λn,xn) cos d(Rf p, xn) and hence (3.1) cos d(Rλn f xn, Rf p) ≥ cos d(Rλnf xn, p) + C 2 1,p 1 + C 2 1,p · 1 + C 2 λnC 2 λn,xn λn ,xn (cid:0)cos d(Rf p, xn) − cos d(Rf p, Rλnf xn)(cid:1) for all n ∈ N. and inf n λn > 0, the sequence We first show (i). Suppose that the assumptions hold. Since limn Cλn ,xn = 1 ( 1 + C 2 λnC 2 λn,xn λn,xn ) is bounded. Since t 7→ cos t is 1-Lipschitz continuous and limn d(Rλn f xn, xn) = 0, we have cos d(Rf p, xn) − cos d(Rf p, Rλnf xn) ≤ d(Rf p, xn) − d(Rf p, Rλnf xn) ≤ d(xn, Rλn f xn) → 0 as n → ∞. Taking the lower limit in (3.1), we have cos d(Rλnf xn, Rf p) ≥ lim inf n→∞ lim inf n→∞ cos d(Rλnf xn, p). 10 Y. KIMURA AND F. KOHSAKA It then follows from Lemma 2.10 that cos(cid:18)lim sup n→∞ d(Rλn f xn, Rf p)(cid:19) ≥ cos(cid:18)lim sup n→∞ d(Rλn f xn, p)(cid:19) lim sup n→∞ d(Rλnf xn, Rf p) ≤ lim sup n→∞ d(Rλnf xn, p). and hence (3.2) On the other hand, since limn d(Rλn f xn, xn) = 0, we have (3.3) lim sup n→∞ d(Rλnf xn, y) = lim sup n→∞ d(xn, y) for all y ∈ X. By (3.2) and (3.3), we have lim sup n→∞ d(xn, Rf p) ≤ lim sup n→∞ d(xn, p). Since A(cid:0){xn}(cid:1) = {p}, we obtain Rf p = p. Consequently, it follows from (2.6) that p ∈ argminX f . We next show (ii). Suppose that the assumptions hold. Since d(Rλn f xn, xn) < sup n π 2 , we know that 0 < cos(cid:18)sup n d(Rλn f xn, xn)(cid:19) = inf n Thus it follows from limn λn = ∞ that ≤ 1 + C 2 λnC 2 0 < λn ,xn λn,xn cos d(Rλn f xn, xn) = inf n Cλn,xn . 2 λn (inf m Cλm,xm)2 → 0 as n → ∞. Taking the lower limit in (3.1), we know that (3.2) holds. Then, since A(cid:0){Rλnf xn}(cid:1) = {p}, we obtain Rf p = p and hence p ∈ argminX f . We also need the following two lemmas. (cid:3) Lemma 3.2. Let X be an admissible complete CAT(1) space, f an element of Γ0(X), Rηf the resolvent of ηf for all η > 0, and {λn} a sequence of positive real numbers. If {xn} is a sequence in X such that n→∞(cid:0)cos d(u, Rλnf xn) − cos d(u, xn)(cid:1) = 0 d(u, xn) < sup and lim π 2 n for some u ∈ argminX f , then limn d(Rλn f xn, xn) = 0. Proof. It follows from (2.11) that (3.4) cos d(u, Rλnf xn) ≥ cos d(Rλn f xn, xn) cos d(u, Rλnf xn) ≥ cos d(u, xn) and hence (3.5) and (3.6) inf n cos d(u, Rλnf xn) ≥ inf n cos d(u, xn) = cos(cid:18)sup n d(u, xn)(cid:19) > 0 0 ≤ cos d(u, Rλnf xn)(cid:18)1 − = cos d(u, Rλnf xn) − cos d(u, xn) → 0 cos d(u, Rλnf xn)(cid:19) cos d(u, xn) TWO MODIFIED PROXIMAL POINT ALGORITHMS 11 as n → ∞. Thus it follows from (3.5) and (3.6) that (3.7) cos d(u, xn) = 1. lim n→∞ cos d(u, Rλnf xn) By (3.4) and (3.7), we obtain 1 ≥ cos d(Rλn f xn, xn) ≥ cos d(u, xn) cos d(u, Rλnf xn) → 1 (cid:3) as n → ∞ and hence limn d(Rλnf xn, xn) = 0. Lemma 3.3. Let X be an admissible complete CAT(1) space, F a nonempty closed convex subset of X, P the metric projection of X onto F , and {xn} a spherically bounded sequence in X. If ω∆(cid:0){xn}(cid:1) is a subset of F , then for all v ∈ X. Proof. Let v ∈ X be given. Since {xn} is spherically bounded, Lemma 2.5 ensures that there exists a subsequence {xni} of {xn} which is ∆-convergent to some q ∈ X and cos d(P v, v) ≥ lim sup n→∞ cos d(xn, v) (3.8) lim i→∞ cos d(xni , v) = lim sup n→∞ cos d(xn, v). (3.9) Since ω∆(cid:0){xn}(cid:1) ⊂ F , we have q ∈ F and hence the definition of P gives us that Since − cos d(·, v) belongs to Γ0(X), Lemma 2.8 implies that it is ∆-lower semicon- tinuous. Thus we have cos d(P v, v) ≥ cos d(q, v). (3.10) By (3.8), (3.9), and (3.10), we obtain the conclusion. − cos d(q, v) ≤ lim inf i→∞ (cid:0)− cos d(xni , v)(cid:1) = − lim i→∞ cos d(xni , v). (cid:3) 4. ∆-convergent proximal-type algorithm In this section, using some techniques from [24], we obtain the following theorem, which is a generalization of Theorem 1.5. This is the first one of our three main results in this paper. Theorem 4.1. Let X be an admissible complete CAT(1) space, f an element of Γ0(X), Rηf the resolvent of ηf for all η > 0, and {xn} a sequence defined by x1 ∈ X and (1.7), where {αn} is a sequence in [0, 1[ and {λn} is a sequence of positive real numbers such that P∞n=1(1 − αn)λn = ∞. Then the following hold. (i) The set argminX f is nonempty if and only if {Rλnf xn} is spherically (ii) if argminX f is nonempty and supn αn < 1, then both {xn} and {Rλnf xn} bounded and supn d(Rλnf xn, xn) < π/2; are ∆-convergent to an element x∞ of argminX f . Proof. Let Cη,z be the real number given by (2.8) for all η > 0 and z ∈ X and let {zn} be the sequence in X given by zn = Rλnf xn for all n ∈ N. 12 Y. KIMURA AND F. KOHSAKA We first show the if part of (i). Suppose that {zn} is spherically bounded and supn d(zn, xn) < π/2 and let {βn} and {σn} be the real sequences given by βn = (1 − αn)λnC 2 λn,xn 1 + C 2 λn ,xn and σn = βl n Xl=1 for all n ∈ N. Since αn < 1 and λn > 0 for all n ∈ N and X is admissible, we know that {βn} is a sequence of positive real numbers. Since supn d(zn, xn) < π/2, we also know that 0 < cos(cid:18)sup n d(zn, xn)(cid:19) = inf n cos d(zn, xn) = inf n Cλn,xn . Thus, noting that (1 − αn)λn (inf m Cλm,xm)2 2 βn ≥ and P∞n=1(1 − αn)λn = ∞, we obtain P∞n=1 βn = ∞. According to Theorem 2.9, the real function g on X defined by (2.12) for all y ∈ X has a unique maximizer p ∈ X. It then follows from (2.9) that λkC 2 1 + C 2 λk ,xk λk ,xk λkC 2 1 + C 2 λk ,xk λk ,xk ≥ cos d(Rλk f xk, Rf p) cos d(Rλk f xk, p) + and hence (4.1) βk cos d(zk, Rf p) ≥ βk cos d(zk, p) + C 2 1,p 1 + C 2 1,p(cid:0)cos d(Rf p, xk) − cos d(Rf p, Rλkf xk)(cid:1) (1 − αk)C 2 1,p 1 + C 2 1,p (cid:0)cos d(Rf p, xk) − cos d(Rf p, zk)(cid:1) for all k ∈ N. On the other hand, it follows from (2.3) and the definition of {xn} that (4.2) cos d(Rf p, xk+1) ≥ αk cos d(Rf p, xk) + (1 − αk) cos d(Rf p, zk) and hence, by (4.1) and (4.2), we have βk cos d(zk, Rf p) (4.3) C 2 ≥ βk cos d(zk, p) + 1,p(cid:0)cos d(Rf p, xk) − cos d(Rf p, xk+1)(cid:1) for all k ∈ N. Summing up (4.3) with respect to k ∈ {1, 2, . . . , n}, we obtain 1,p 1 + C 2 βk cos d(zk, Rf p) 1 σn (4.4) n Xk=1 Xk=1 1 σn n C 2 1 1,p 1 + C 2 ≥ βk cos d(zk, p) + σn(cid:0)cos d(Rf p, x1) − cos d(Rf p, xn+1)(cid:1) for all n ∈ N. Since limn σn = ∞, it follows from (4.4) that g(Rf p) ≥ g(p). Since p is the unique maximizer of g, we obtain Rf p = p. Consequently, it follows from (2.6) that p ∈ argminX f . Therefore, the set argminX f is nonempty. 1,p · TWO MODIFIED PROXIMAL POINT ALGORITHMS 13 We next show the only if part of (i). Suppose that argminX f is nonempty and let u ∈ argminX f be given. Then it follows from (2.11) that (4.5) min(cid:8)cos d(u, zn), cos d(zn, xn)(cid:9) ≥ cos d(u, zn) cos d(zn, xn) and hence ≥ cos d(u, xn) max(cid:8)d(u, zn), d(zn, xn)(cid:9) ≤ d(u, xn) (4.6) for all n ∈ N. By (2.3) and (4.5), we have (4.7) cos d(u, xn+1) ≥ αn cos d(u, xn) + (1 − αn) cos d(u, zn) ≥ cos d(u, xn). It then follows from (4.7) and the admissibility of X that (4.8) d(u, xn+1) ≤ d(u, xn) ≤ d(u, x1) < π 2 for all n ∈ N. Thus it follows from (4.6) and (4.8) that lim sup n→∞ d(u, zn) ≤ lim n→∞ d(u, xn) ≤ d(u, x1) < π 2 . This implies the spherical boundedness of {xn} and {zn}. It also follows from (4.6) and (4.8) that supn d(zn, xn) < π/2. We finally show (ii). Suppose that argminX f is nonempty and supn αn < 1. Then we know that (4.5), (4.6), (4.7), and (4.8) hold and that both {xn} and {zn} are spherically bounded. Let u ∈ argminX f be given. It follows from (4.8) that {d(u, xn)} tends to some β ∈ [0, π/2[. By (2.3) and (4.5), we have cos d(u, xn+1) ≥ αn cos d(u, xn) + (1 − αn) cos d(u, zn) ≥ αn cos d(u, xn) + (1 − αn) · = cos d(u, xn) + (1 − αn) cos d(u, xn)(cid:18) cos d(u, xn) cos d(zn, xn) 1 cos d(zn, xn) − 1(cid:19) and hence (4.9) 0 ≤ (1 − αn)(cid:18) 1 cos d(zn, xn) − 1(cid:19) ≤ cos d(u, xn+1) cos d(u, xn) − 1 → cos β cos β − 1 = 0 as n → ∞. Since supn αn < 1, it follows from (4.9) that (4.10) d(zn, xn) = 0. lim n→∞ On the other hand, it follows from (2.10) and (4.10) that there exists a positive real number K such that Kπ (4.11) for all n ∈ N. It then follows from (4.7) and (4.11) that λn(cid:0)f (zn) − f (u)(cid:1) ≤ (1 − αn)λn(cid:0)f (zn) − f (u)(cid:1) ≤ Xn=1 (1 − αn)λn(cid:0)f (zn) − f (u)(cid:1) ≤ 2 (cid:0)cos d(u, zn) − cos d(u, xn)(cid:1) 2 (cid:0)cos d(u, xn+1) − cos d(u, xn)(cid:1) 2 (cid:0)cos β − cos d(u, x1)(cid:1) < ∞. Kπ Kπ ∞ and hence (4.12) 14 Y. KIMURA AND F. KOHSAKA (4.13) Since P∞n=1(1 − αn)λn = ∞, it follows from (4.12) that n→∞ (cid:0)f (zn) − f (u)(cid:1) = 0. By the definitions of {xn} and {zn} and the convexity of f , we also have lim inf −∞ < inf f (X) ≤ f (zn) ≤ f (zn) + 1 λn tan d(zn, xn) sin d(zn, xn) ≤ f (xn) and −∞ < inf f (X) ≤ f (xn+1) ≤ αnf (xn) + (1 − αn)f (zn) ≤ f (xn) for all n ∈ N. Thus {f (xn)} tends to some γ ∈ R and {f (zn)} is bounded. Let {ni} be any increasing sequence in N. Since supn αn < 1, we have a subsequence {nij} of {ni} such that {αnij } tends to some δ ∈ [0, 1[. Then, letting j → ∞ in 1 1 − αnij (cid:16)f(cid:0)xnij +1(cid:1) − αnij f(cid:0)xnij(cid:1)(cid:17) ≤ f(cid:0)znij(cid:1) ≤ f(cid:0)xnij(cid:1), we obtain f(cid:0)znij(cid:1) → γ. Thus {f (zn)} also tends to γ. Consequently, it follows from (4.13) that (4.14) f (xn) = γ = f (u) = inf f (X). lim n→∞ Let z be any element of ω∆(cid:0){xn}(cid:1). Then we have a subsequence {xmi} of {xn} which is ∆-convergent to z. Since f is ∆-lower semicontinuous by Lemma 2.8, it follows from (4.14) that f (z) ≤ lim inf i→∞ f (xmi ) = lim n→∞ f (xn) = inf f (X) and hence z ∈ argminX f . Thus ω∆(cid:0){xn}(cid:1) is a subset of argminX f . It then follows from (4.8) that {d(z, xn)} is convergent for each z ∈ ω∆(cid:0){xn}(cid:1). Thus, Lemma 2.6 ensures that {xn} is ∆-convergent to some x∞ ∈ X. Since f, {x∞} = ω∆(cid:0){xn}(cid:1) ⊂ argmin X we know that x∞ belongs to argminX f . It then follows from (4.10) that A(cid:0){zli}(cid:1) = A(cid:0){xli}(cid:1) = {x∞} for each increasing sequence {li} in N. Consequently, we conclude that both {xn} and {zn} are ∆-convergent to an element x∞ of argminX f . (cid:3) As a direct consequence of Theorem 4.1, we obtain the following corollary. Corollary 4.2. Let (SH , ρSH ) be a Hilbert sphere, X an admissible closed convex subset of SH , f an element of Γ0(X), and {xn} a sequence defined by x1 ∈ X and (1.7), where {αn} is a sequence in [0, 1[ and {λn} is a sequence of positive real numbers such that P∞n=1(1 − αn)λn = ∞. Then argminX f is nonempty if and only if {Rλnf xn} is spherically bounded and supn ρSH (Rλn f xn, xn) < π/2. Further, if argminX f is nonempty and supn αn < 1, then both {xn} and {Rλnf xn} are ∆-convergent to an element x∞ of argminX f . TWO MODIFIED PROXIMAL POINT ALGORITHMS 15 5. Convergent proximal-type algorithm In this section, using some techniques from [28, Theorem 3.2], we first obtain the following theorem. This is the second one of our three main results in this paper. Theorem 5.1. Let X be an admissible complete CAT(1) space, f an element of Γ0(X), Rηf the resolvent of ηf for all η > 0, v an element of X, and {yn} a sequence defined by y1 ∈ X and (1.8), where {αn} is a sequence in [0, 1] and {λn} is a sequence of positive real numbers such that limn λn = ∞. Then the following hold. bounded and supn d(Rλnf yn, yn) < π/2; (i) The set argminX f is nonempty if and only if {Rλnf yn} is spherically (ii) if argminX f is nonempty, limn αn = 0, and P∞n=1 α2 n = ∞, then both {yn} and {Rλnf yn} are convergent to P v, where P denotes the metric projection of X onto argminX f . Proof. Let {zn} be the sequence in X given by zn = Rλn f yn for all n ∈ N. We first show the if part of (i). Suppose that {zn} is spherically bounded and supn d(zn, yn) < π/2. Then Lemma 2.5 implies that there exists p ∈ X such that A(cid:0){zn}(cid:1) = {p}. Since lim n→∞ λn = ∞ and sup n d(zn, yn) < π 2 , it follows from (ii) of Lemma 3.1 that p ∈ argminX f . Thus argminX f is nonempty. We next show the only if part of (i). Suppose that argminX f is nonempty and let P be the metric projection of X onto argminX f . It follows from (2.11) that (5.1) for all n ∈ N. It also follows from (2.3) and (2.11) that max(cid:8)d(P v, zn), d(zn, yn)(cid:9) ≤ d(P v, yn) cos d(P v, yn+1) ≥ αn cos d(P v, v) + (1 − αn) cos d(P v, yn). This implies that for all n ∈ N and hence (5.2) cos d(P v, yn) ≥ min(cid:8)cos d(P v, v), cos d(P v, y1)(cid:9) d(P v, yn) ≤ max(cid:8)d(P v, v), d(P v, y1)(cid:9) < π 2 for all n ∈ N, where the last inequality follows from the admissibility of X. Then, by (5.1) and (5.2), we see that both {yn} and {zn} are spherically bounded and (5.3) d(zn, yn) < sup . n π 2 P∞n=1 α2 and {zn} are spherically bounded. By (2.5) and (5.1), we have We finally show (ii). Suppose that argminX f is nonempty, limn αn = 0, and n = ∞. Then we know that (5.1), (5.2), and (5.3) hold and that both {yn} cos d(cid:0)P v, yn+1(cid:1) ≥ (1 − βn) cos d(P v, yn) + βn · cos d(P v, v) (5.4) , sin d(zn, v) tan(cid:0) αn 2 d(zn, v)(cid:1) + cos d(zn, v) 16 where (5.5) Y. KIMURA AND F. KOHSAKA sin(cid:0)(1 − αn)d(zn, v)(cid:1) sin d(zn, v) 1 − αn (zn 6= v); (zn = v) βn =  for all n ∈ N. Note that if zn 6= v, then it follows from sin(cid:0)(1 − αn)d(zn, v)(cid:1) ≥ αn sin 0 + (1 − αn) sin d(zn, v) = (1 − αn) sin d(zn, v) that αn ≥ βn. Hence we have (5.6) for all n ∈ N. Letting (5.7) αn ≥ βn sn = 1 − cos d(P v, yn) and (5.8) tn = 1 − we have from (5.4) that cos d(P v, v) sin d(zn, v) tan(cid:0) αn sn+1 ≤ (1 − βn)sn + βntn 2 d(zn, v)(cid:1) + cos d(zn, v) , (5.9) for all n ∈ N. Let us show that P∞n=1 βn = ∞. Suppose that zn 6= v. ϕ(t) = sin(γt)/ sin t for all t ∈ ]0, π/2], then we have (cid:0)γ tan t − tan(γt)(cid:1) (cid:16)γ tan t −(cid:0)(1 − γ) tan 0 + γ tan t(cid:1)(cid:17) = 0 cos(γt) cos t cos(γt) cos t ϕ′(t) = sin2 t sin2 t ≥ for all t ∈ ]0, π/2[. Thus the function t 7→ sin(cid:0)(1 − αn)t(cid:1) sin t If γ ∈ [0, 1] and is nondecreasing on ]0, π/2]. Using this property and the inequality 1− t2/4 ≥ cos t on [0, π/2], we have (1 − αn)π βn ≥ 1 − sin = 1 − cos 2 If zn = v, then βn = αn ≥ α2 n ≥ 16−1π2α2 n. Hence the inequality π2 βn ≥ (5.10) 16 α2 n = αnπ 2 ≥ 1 4(cid:16) αnπ 2 (cid:17)2 α2 nπ2 16 . holds for all n ∈ N. It then follows from P∞n=1 α2 We next show that lim supn tn ≤ 0. Since limn αn = 0, we have n = ∞ that P∞n=1 βn = ∞. If lim supn cos d(zn, v) = 0, then we have sin d(zn, v) tan lim n→∞ = 0. αnd(zn, v) 2 cos d(P v, v) lim sup n→∞ tn = 1 − lim inf n→∞ sin d(zn, v) tan(cid:0) αn 2 d(zn, v)(cid:1) + cos d(zn, v) = −∞. TWO MODIFIED PROXIMAL POINT ALGORITHMS 17 If lim supn cos d(zn, v) > 0, then it follows from Lemma 2.10 that tn lim sup n→∞ = 1 − lim inf n→∞ (5.11) = 1 − = 1 − cos d(P v, v) sin d(zn, v) tan(cid:0) αn 2 d(zn, v)(cid:1) + cos d(zn, v) cos d(P v, v) lim supn→∞(cid:16)sin d(zn, v) tan(cid:0) αn lim supn→∞ cos d(P v, v) cos d(zn, v) . 2 d(zn, v)(cid:1) + cos d(zn, v)(cid:17) On the other hand, if q is any element of ω∆(cid:0){zn}(cid:1), then there exists a subsequence {zni} of {zn} which is ∆-convergent to some q ∈ X. Since lim i→∞ λni = ∞, A(cid:0){Rλni f yni}(cid:1) = {q}, and sup i d(Rλni f yni , yni) < π 2 , it follows from (ii) of Lemma 3.1 that q ∈ argminX f . Thus ω∆(cid:0){zn}(cid:1) is a subset of argminX f . It then follows from Lemma 3.3 that (5.12) cos d(zn, v). cos d(P v, v) ≥ lim sup n→∞ By (5.11) and (5.12), we know that lim supn tn ≤ 0. Therefore, Lemma 2.1 yields that limn sn = 0 and hence (5.13) d(P v, yn) = 0. lim n→∞ Consequently, by (5.1) and (5.13), we conclude that both {yn} and {zn} are con- vergent to P v. (cid:3) Xn=1 Using Lemma 2.2, we next obtain the last one of our three main results in this paper. Theorem 5.2. Let X, f , {Rηf}η>0, and v be the same as in Theorem 5.1, and {yn} a sequence defined by y1 ∈ X and (1.8), where {αn} is a sequence in ]0, 1] and {λn} is a sequence of positive real numbers such that inf n αn = 0, λn > 0. and ∞ α2 n = ∞, lim n→∞ If argminX f is nonempty, then both {yn} and {Rλnf yn} are convergent to P v, where P denotes the metric projection of X onto argminX f . Proof. Let {zn} be the sequence in X given by zn = Rλn f yn for all n ∈ N. As in the proof of Theorem 5.1, we can see that (5.1), (5.2), and (5.3) hold and that both {yn} and {zn} are spherically bounded. Let {βn}, {sn}, and {tn} be the real sequences defined by (5.5), (5.7), and (5.8), respectively. Then we can see that (5.6), (5.9), and (5.10) hold for all n ∈ N. Since {αn} is a sequence in ]0, 1], so is {βn}. Since P∞n=1 α2 Let {ni} be any increasing sequence in N such that i→∞ (cid:0)sni − sni+1(cid:1) ≤ 0. n = ∞, it follows from (5.10) that P∞n=1 βn = ∞. lim sup (5.14) Then we show that (5.15) lim sup i→∞ tni ≤ 0. 18 Y. KIMURA AND F. KOHSAKA If lim supi cos d(zni , v) = 0, then we have If lim supi cos d(zni , v) > 0, then it follows from Lemma 2.10 that lim sup i→∞ tni = 1 − lim inf i→∞ (5.16) lim sup i→∞ It follows from (2.3) that cos d(P v, v) 2 d(zni , v)(cid:1) + cos d(zni , v) = −∞. sin d(zni, v) tan(cid:0) αni tni = 1 − cos d(P v, v) lim supi→∞ cos d(zni, v) . sni − sni+1 = cos d(P v, yni+1) − cos d(P v, yni ) ≥ αni cos d(P v, v) + (1 − αni ) cos d(P v, zni) − cos d(P v, yni ). Hence (2.11) yields that (5.17) sni − sni+1 + αni(cid:0)cos d(P v, zni ) − cos d(P v, v)(cid:1) ≥ cos d(P v, zni ) − cos d(P v, yni) ≥ 0. Since limi αni = 0, it follows from (5.14) and (5.17) that On the other hand, it follows from (5.2) that lim i→∞(cid:0)cos d(P v, zni) − cos d(P v, yni )(cid:1) = 0. sup i d(P v, yni ) ≤ sup n d(P v, yn) < π 2 . Thus it follows from Lemma 3.2 that limi d(zni, yni) = 0. Let q be any element of ω∆(cid:0){zni}(cid:1). Then there exists a subsequence {znij } of {zni} which is ∆-convergent to some q ∈ X. Since inf j λnij > 0, A(cid:0){ynij}(cid:1) = {q}, and lim j→∞ d(cid:0)Rλnij f ynij , ynij(cid:1) = 0, it follows from (i) of Lemma 3.1 that q ∈ argminX f . Thus ω∆(cid:0){zni}(cid:1) is a subset of argminX f . Then, by Lemma 3.3, we know that (5.18) cos d(zni, v). cos d(P v, v) ≥ lim sup i→∞ By (5.16) and (5.18), we know that (5.15) holds. Therefore, Lemma 2.2 yields that limn sn = 0 and hence (5.19) d(P v, yn) = 0. lim n→∞ Consequently, by (5.1) and (5.19), we conclude that both {yn} and {zn} are con- vergent to P v. (cid:3) As direct consequences of Theorems 5.1 and 5.2, we obtain the following two corollaries, respectively. Corollary 5.3. Let (SH , ρSH ) be a Hilbert sphere, X an admissible closed convex subset of SH , f an element of Γ0(X), v an element of X, and {yn} a sequence de- fined by y1 ∈ X and (1.8), where {αn} is a sequence in [0, 1] and {λn} is a sequence of positive real numbers such that limn λn = ∞. Then argminX f is nonempty if and only if {Rλnf yn} is spherically bounded and supn ρSH (Rλnf yn, yn) < π/2. Fur- ther, if argminX f is nonempty, limn αn = 0, and P∞n=1 α2 n = ∞, then both {yn} TWO MODIFIED PROXIMAL POINT ALGORITHMS 19 and {Rλnf yn} are convergent to P v, where P denotes the metric projection of X onto argminX f . Corollary 5.4. Let (SH , ρSH ), X, f , and v be the same as in Corollary 5.3 and {yn} a sequence defined by y1 ∈ X and (1.8), where {αn} is a sequence in ]0, 1] and {λn} is a sequence of positive real numbers such that limn αn = 0, P∞n=1 α2 n = ∞, and inf n λn > 0. If argminX f is nonempty, then both {yn} and {Rλnf yn} are convergent to P v, where P denotes the metric projection of X onto argminX f . 6. Results in CAT(κ) spaces with a positive κ In this final section, using Theorems 4.1, 5.1, and 5.2, we deduce three corollaries in CAT(κ) spaces with a positive real number κ. Throughout this section, we suppose the following. • κ is a positive real number and Dκ = π/√κ; • X is a complete CAT(κ) space such that d(w, w′) < Dκ/2 for all w, w′ ∈ X; • f is a proper lower semicontinuous convex function of X into ]−∞,∞]; • Rηf is the mapping of X into itself defined by Rηf x = argmin y∈X (cid:26)f (y) + 1 η tan(cid:0)√κd(y, x)(cid:1) sin(cid:0)√κd(y, x)(cid:1)(cid:27) for all η > 0 and x ∈ X. Since the space (X,√κd) is an admissible complete CAT(1) space, the mapping Rηf is well defined and Theorems 4.1, 5.1, and 5.2 immediately imply the following three corollaries, respectively. Corollary 6.1. Let {xn} be a sequence defined by x1 ∈ X and xn+1 = αnxn ⊕ (1 − αn) Rλnf xn (n = 1, 2, . . . ), where {αn} is a sequence in [0, 1[ and {λn} is a sequence of positive real numbers such that P∞n=1(1 − αn)λn = ∞. Then argminX f is nonempty if and only if d(y, Rλnf xn) < Dκ 2 and sup n d( Rλnf xn, xn) < Dκ 2 . inf y∈X lim sup n→∞ Further, if argminX f is nonempty and supn αn < 1, then both {xn} and { Rλnf xn} are ∆-convergent to an element x∞ of argminX f . Corollary 6.2. Let v be an element of X and {yn} a sequence defined by y1 ∈ X and (6.1) yn+1 = αnv ⊕ (1 − αn) Rλn f yn (n = 1, 2, . . . ), where {αn} is a sequence in [0, 1] and {λn} is a sequence of positive real numbers such that limn λn = ∞. Then argminX f is nonempty if and only if d( Rλn f yn, yn) < d(y, Rλnf yn) < sup and . Dκ 2 n Dκ 2 inf y∈X lim sup n→∞ Further, if argminX f is nonempty, limn αn = 0, and P∞n=1 α2 n = ∞, then both {yn} and { Rλnf yn} are convergent to P v, where P denotes the metric projection of X onto argminX f . 20 Y. KIMURA AND F. KOHSAKA Corollary 6.3. Let v be an element of X, and {yn} a sequence defined by y1 ∈ X and (6.1), where {αn} is a sequence in ]0, 1] and {λn} is a sequence of positive real numbers such that αn = 0, lim n→∞ α2 n = ∞, and inf n λn > 0. ∞ Xn=1 If argminX f is nonempty, then both {yn} and { Rλnf yn} are convergent to P v, where P denotes the metric projection of X onto argminX f . Concluding remarks As we stated in Section 1, it is known [34, Definition 4.1 and Lemma 4.2] that the classical resolvent given by (1.2) is still well defined for any proper lower semi- continuous convex function in a complete CAT(1) space whose diameter is strictly less than π/2. However, it is also known [22, Corollary 3.3] that this diameter condition on the space implies that such a function always has a minimizer. On the other hand, according to [22, Theorem 4.2], we can define another type of resolvent by (1.5) with the perturbation function tan d sin d in an admissible complete CAT(1) space. This makes it possible for us to study the existence of minimizers as well as the convergence to minimizers through the two proximal-type algorithms defined by (1.7) and (1.8). Finally, we point out that it is not clear whether there is any relationship between the two types of resolvents and hence we cannot deduce any result for the classical resolvents from the results obtained in this paper so far. Acknowledgment The authors would like to thank the anonymous referees for their helpful com- ments on the original version of this paper. This work was supported by JSPS KAKENHI Grant Numbers 15K05007 and 17K05372. References [1] K. Aoyama, Y. Kimura, W. Takahashi, and M. Toyoda, Approximation of common fixed points of a countable family of nonexpansive mappings in a Banach space, Nonlinear Anal. 67 (2007), 2350–2360. [2] K. Aoyama, F. Kohsaka, and W. Takahashi, Proximal point methods for monotone operators in Banach spaces, Taiwanese J. Math. 15 (2011), 259–281. [3] D. Ariza-Ruiz, L. Leu¸stean, and G. L´opez-Acedo, Firmly nonexpansive mappings in classes of geodesic spaces, Trans. Amer. Math. Soc. 366 (2014), 4299–4322. [4] M. Bac´ak, The proximal point algorithm in metric spaces, Israel J. Math. 194 (2013), 689– 701. [5] [6] M. Bac´ak and S. Reich, The asymptotic behavior of a class of nonlinear semigroups in , Convex analysis and optimization in Hadamard spaces, De Gruyter, Berlin, 2014. Hadamard spaces, J. Fixed Point Theory Appl. 16 (2014), 189–202. [7] C. Bargetz, M. Dymond, and S. Reich, Porosity results for sets of strict contractions on geodesic metric spaces, Topol. Methods Nonlinear Anal. 50 (2017), 89–124. [8] H. H. Bauschke and P. L. Combettes, Convex analysis and monotone operator theory in Hilbert spaces, Springer, New York, 2011. [9] H. H. Bauschke, E. Matouskov´a, and S. Reich, Projection and proximal point methods: con- vergence results and counterexamples, Nonlinear Anal. 56 (2004), 715–738. [10] H. Br´ezis and P.-L. Lions, Produits infinis de r´esolvantes, Israel J. Math. 29 (1978), 329–345. [11] M. R. Bridson and A. Haefliger, Metric spaces of non-positive curvature, Springer-Verlag, Berlin, 1999. TWO MODIFIED PROXIMAL POINT ALGORITHMS 21 [12] R. E. Bruck and S. Reich, Nonexpansive projections and resolvents of accretive operators in Banach spaces, Houston J. Math. 3 (1977), 459–470. [13] D. Burago, Y. Burago, and S. Ivanov, A course in metric geometry, American Mathematical Society, Providence, RI, 2001. [14] J. Eckstein and D. P. Bertsekas, On the Douglas-Rachford splitting method and the proximal point algorithm for maximal monotone operators, Math. Programming 55 (1992), 293–318. [15] R. Esp´ınola and A. Fern´andez-Le´on, CAT(k)-spaces, weak convergence and fixed points, J. Math. Anal. Appl. 353 (2009), 410–427. [16] R. Esp´ınola and A. Nicolae, Proximal minimization in CAT(κ) spaces, J. Nonlinear Convex Anal. 17 (2016), 2329–2338. [17] K. Goebel and S. Reich, Uniform convexity, hyperbolic geometry, and nonexpansive map- pings, Marcel Dekker, Inc., New York, 1984. [18] O. Guler, On the convergence of the proximal point algorithm for convex minimization, SIAM J. Control Optim. 29 (1991), 403–419. [19] J. Jost, Convex functionals and generalized harmonic maps into spaces of nonpositive cur- vature, Comment. Math. Helv. 70 (1995), 659–673. [20] S. Kamimura and W. Takahashi, Approximating solutions of maximal monotone operators in Hilbert spaces, J. Approx. Theory 106 (2000), 226–240. [21] W. S. Kendall, Convexity and the hemisphere, J. London Math. Soc. (2) 43 (1991), 567–576. [22] Y. Kimura and F. Kohsaka, Spherical nonspreadingness of resolvents of convex functions in geodesic spaces, J. Fixed Point Theory Appl. 18 (2016), 93–115. [23] [24] , Two modified proximal point algorithms for convex functions in Hadamard spaces, Linear Nonlinear Anal. 2 (2016), 69–86. , The proximal point algorithm in geodesic spaces with curvature bounded above, Linear Nonlinear Anal. 3 (2017), 133–148. [25] Y. Kimura and S. Saejung, Strong convergence for a common fixed point of two different generalizations of cutter operators, Linear Nonlinear Anal. 1 (2015), 53–65. [26] Y. Kimura, S. Saejung, and P. Yotkaew, The Mann algorithm in a complete geodesic space with curvature bounded above, Fixed Point Theory Appl. (2013), 2013:336, 1–13. [27] Y. Kimura and K. Sato, Convergence of subsets of a complete geodesic space with curvature bounded above, Nonlinear Anal. 75 (2012), 5079–5085. [28] , Halpern iteration for strongly quasinonexpansive mappings on a geodesic space with curvature bounded above by one, Fixed Point Theory Appl. (2013), 2013:7, 1–14. [29] W. A. Kirk and B. Panyanak, A concept of convergence in geodesic spaces, Nonlinear Anal. 68 (2008), 3689–3696. [30] T. C. Lim, Remarks on some fixed point theorems, Proc. Amer. Math. Soc. 60 (1976), 179– 182. [31] B. Martinet, R´egularisation d'in´equations variationnelles par approximations successives, Rev. Fran¸caise Informat. Recherche Op´erationnelle 4 (1970), 154–158. [32] U. F. Mayer, Gradient flows on nonpositively curved metric spaces and harmonic maps, Comm. Anal. Geom. 6 (1998), 199–253. [33] O. Nevanlinna and S. Reich, Strong convergence of contraction semigroups and of iterative methods for accretive operators in Banach spaces, Israel J. Math. 32 (1979), 44–58. [34] S. Ohta and M. P´alfia, Discrete-time gradient flows and law of large numbers in Alexandrov spaces, Calc. Var. Partial Differential Equations 54 (2015), 1591–1610. [35] S. Reich and I. Shafrir, Nonexpansive iterations in hyperbolic spaces, Nonlinear Anal. 15 (1990), 537–558. [36] R. T. Rockafellar, Monotone operators and the proximal point algorithm, SIAM J. Control Optim. 14 (1976), 877–898. [37] S. Saejung and P. Yotkaew, Approximation of zeros of inverse strongly monotone operators in Banach spaces, Nonlinear Anal. 75 (2012), 742–750. [38] M. V. Solodov and B. F. Svaiter, Forcing strong convergence of proximal point iterations in a Hilbert space, Math. Program. 87 (2000), 189–202. [39] W. Takahashi, Introduction to nonlinear and convex analysis, Yokohama Publishers, Yoko- hama, 2009. [40] H.-K. Xu, Iterative algorithms for nonlinear operators, J. London Math. Soc. (2) 66 (2002), 240–256. 22 Y. KIMURA AND F. KOHSAKA [41] T. Yokota, Convex functions and barycenter on CAT(1)-spaces of small radii, J. Math. Soc. Japan 68 (2016), 1297–1323. (Y. Kimura) Department of Information Science, Toho University, Miyama, Funabashi, Chiba 274-8510, Japan E-mail address: [email protected] (F. Kohsaka) Department of Mathematical Sciences, Tokai University, Kitakaname, Hiratsuka, Kanagawa 259-1292, Japan E-mail address: [email protected]
1312.1059	1	1312	2013-12-04T08:57:22	Fourier--Jacobi expansions in Morrey spaces	[ "math.FA", "math.CA" ]	In this paper we obtain a characterization of the convergence of the partial sum operator related to Fourier--Jacobi expansions in Morrey spaces.	math.FA	math	FOURIER -- JACOBI EXPANSIONS IN MORREY SPACES ALBERTO ARENAS AND ´OSCAR CIAURRI Abstract. In this paper we obtain a characterization of the convergence of the partial sum operator related to Fourier -- Jacobi expansions in Morrey spaces. 1. Introduction and main results For α, β > −1, we consider the Jacobi functions p(α,β) n (x)(1 − x)α/2(1 + x)β/2, (x) = d(α,β) P (α,β) n n x ∈ (−1, 1), n = 0, 1, 2, . . . , where P (α,β) n denotes the Jacobi polynomial of order (α, β) and degree n, and 1 d(α,β) n =(cid:18)Z 1 −1(cid:16)P (α,β) n (x)(cid:17)2 (1 − x)α(1 + x)β dx(cid:19) 1/2 . The system of functions {p(α,β) }n≥0 is orthonormal and complete in L2(−1, 1) with the Lebesgue measure. Given an appropriate function f , its Fourier expansion respect to the Jacobi functions, which we call Fourier -- Jacobi expansion, is given by n f ∼ a(α,β) k (f )p(α,β) k , a(α,β) k p(α,β) k (t)f (t) dt. (f ) =Z 1 −1 ∞ Xk=0 The convergence of the partial sum operator for the Fourier -- Jacobi expansions, given by n Snf = a(α,β) k (f )p(α,β) k , Xk=0 which is equivalent to the uniform boundedness of Sn, has been widely analyzed in different kinds of spaces. In the case of the Fourier series related to the Jacobi polynomials in Lp spaces the first known results are due to Pollard [12] and Wing [18] who treated the case α, β ≥ −1/2. In [10], Muckenhoupt extended the analysis to the whole range α, β > −1 and included some weights. With the sufficient conditions on the weights given in Muckenhoupt's paper and by using the results about the necessary conditions in, for instance, [7] (about this question see [2, 8, 5] also), the result for Fourier -- Jacobi functions establishes that kSnf kLp(−1,1) ≤ Ckf kLp(−1,1) ⇐⇒ 4 3 < p < 4. A complete study of the boundedness properties of the partial sum operator in Lp,∞ spaces can be seen in [4]. Date: February 5, 2018. 2010 Mathematics Subject Classification. Primary 42C10, 43A50. Key words and phrases. Fourier -- Jacobi expansions, Morrey spaces, weighted inequalities. Research of the author supported by grant MTM2012-36732-C03-02 of the DGI. 1 2 A. ARENAS AND ´O. CIAURRI Our target in this paper is the study of the convergence of the Fourier -- Jacobi expansions in Morrey spaces. To this end, for 1 ≤ p < ∞ and 0 ≤ λ < 1 we define the Morrey space Lp,λ(−1, 1) as the set of functions f on (−1, 1) such that kf kp,λ := x∈(−1,1),r>0 1 sup rλ ZB(x,r) f (t)p dt!1/p < ∞, where B(x, r) = {t ∈ (−1, 1) : t − x ≤ r}. It is clear that Lp,λ(−1, 1) are Banach spaces. Morrey spaces can be defined in a more general way but this is enough for our purposes. The Lp(−1, 1) spaces with the Lebesgue measure correspond with the case λ = 0. Morrey spaces were introduced by Morrey, see [9], in the setting of partial differ- ential equations. In the last years, some classical operators from harmonic analysis have been analyzed in the setting of Morrey spaces, see, for instance, [1], [13], [15], [16], and the references therein. Specially relevant for our purposes will be the boundedness with weights of the Hilbert transform given in [15, Theorem 4.7 and Corollary 4.8]. In particular, we will use the following version of that result: if and Hf (x) =Z 1 −1 f (t) x − t dt k wk(x) = x − xjγj , Yj=1 with −1 ≤ x1 < x2 < · · · < xk−1 < xk ≤ 1, then for 1 < p < ∞ and 0 ≤ λ < 1 it holds that (1) kwkHf kp,λ ≤ Ckwkf kp,λ ⇐⇒ λ − 1 p < γj < 1 + λ − 1 p , j = 1, . . . , k. Our result about convergence of Fourier-Jacobi expansions on Morrey spaces is the following. Theorem 1. Let 0 ≤ λ < 3/4, 1 < p < ∞, and α, β ≥ 0. Then if and only if Snf −→ f, in Lp,λ(−1, 1), 4 3 (1 − λ) < p < 4(1 − λ). As a first step to prove Theorem 1 we start by establishing the equivalence between the convergence of the partial sums and the uniform boundedness of the operator Sn in Morrey spaces. Theorem 2. Let 0 ≤ λ < 1, 1 ≤ p < ∞ and α, β > −1. Then Snf −→ f, in Lp,λ(−1, 1), if and only if where C is a constant independent of n and f . kSnf kp,λ ≤ Ckf kp,λ, FOURIER -- JACOBI EXPANSIONS IN MORREY SPACES 3 So, Theorem 1 will follow from the uniform boundedness of the partial sum operator in Morrey spaces. The next theorem contains a characterization of the interval of values of p for which this estimate holds. Theorem 3. Let 0 ≤ λ < 3/4, 1 < p < ∞, and α, β ≥ 0. Then (2) kSnf kp,λ ≤ Ckf kp,λ, where C is a constant independent of n and f , if and only if (3) 4 3 (1 − λ) < p < 4(1 − λ). The region where the partial sum Sn converges in Morrey spaces is the shadowed one shown in Figure 1. Outside and in the border of that region the convergence is not possible. As it occurs in Lp spaces, we think that in the border, with dashed lines in the figure and corresponding with the values p = 4(1 − λ) and p = max{4(1−λ)/3, 1}, some kind of weak boundedness of the partial sum operator should hold. p 4 4 3 1 p = 4 ( 1 − λ ) p = 4 3(1 − λ) 0 1 4 3 4 λ 1 Figure 1: The region where the partial sum operator converges. The proofs of Theorem 2 and Theorem 3 will be contained in Sections 2 and 3, respectively. In the last section, we include the proof of some auxiliary results. 2. Proof of Theorem 2 Theorem 2 is a standard consequence of Banach-Steinhaus Theorem and the following proposition. Proposition 4. Let 1 ≤ p < ∞, 0 ≤ λ < 1, and α, β > −1. Then, span{p(α,β) dense in Lp,λ(−1, 1). n } is Proof. The case λ = 0 is well known and our proof for 0 < λ < 1 relies on it. Let us suppose that span{p(α,β) } is not dense in Lp,λ(−1, 1), for 0 < λ < 1. Then, by a standard consequence of Hahn-Banach Theorem [14, Theorem 5.19], n 4 A. ARENAS AND ´O. CIAURRI there exists a non-zero functional T on Lp,λ(−1, 1) such that T p(α,β) n ≥ 0. n = 0, for all It is easy to check that for p ≤ r and λ ≤ µ, we have Lr,µ(−1, 1) ⊂ Lp,λ(−1, 1) ⊂ Lp,0(−1, 1) = Lp(−1, 1). So, by Hahn-Banach Theorem [14, Theorem 5.16], we can extend the functional T to Lp(−1, 1). Then, there exists a unique function g ∈ Lq(−1, 1), with p−1 + q−1 = 1, such that T f =Z 1 −1 f (x)g(x) dx, for each f ∈ Lp(−1, 1). But the condition 0 = T p(α,β) this is a contradiction because T was non-zero. implies a(α,β) n n (g) = 0 and g = 0. Then T = 0 and (cid:3) Proof of Theorem 2. By the Banach-Steinhaus Theorem [14, Theorem 5.8], the con- vergence implies the uniform boundedness. On the other hand, for each f ∈ Lp,λ(−1, 1) and given ε > 0, by Proposition } such that kf − gkp,λ < ε. Moreover there exists 4, there exists g ∈ span{p(α,β) N > 0 such that Sn(g) = g, for each n ≥ N . Then n kSnf − f kp,λ ≤ kSnf − Sngkp,λ + kg − f kp,λ ≤ (C + 1)kg − f kp,λ ≤ (C + 1)ε, for n ≥ N and Snf −→ f in Lp,λ(−1, 1). (cid:3) 3. Proof of Theorem 3 Before starting with the proof we are going to collect some facts that we will use. Given a function g defined in (−1, 1), for each 1 ≤ p < ∞ and 0 < λ < 1, we define kgk∗ q,λ := inf x∈(−1,1)Z ∞ 0 rλ/p−1kχ(B(x,r))cgkLq(−1,1) dr, where (B(x, r))c = (−1, 1) \ B(x, r) and p−1 + q−1 = 1. With the previous definition we can give an appropriate version of Holder in- equality for Morrey spaces (see [3, Lemma 4.1]). In our case it reads so. Lemma 1. Let 1 ≤ p < ∞, p−1 + q−1 = 1, and 0 < λ < 1. Then the inequality (4) Z 1 −1 f (x)g(x) dx ≤ Ckf kp,λkgk∗ q,λ, holds with a constant C independent of f and g. It is interesting to know some facts about the norm of certain functions in Morrey spaces. From [16, Remark 4.4] we see that, for a ∈ (−1, 1) and νp > −1, the function x − aνχB(x0,r0) ∈ Lp,λ(−1, 1) if and only if νp ≥ λ − 1, with λ > 0, and (5) kx − aνχB(x0,r0)kp,λ ≃ B(x0, r0)(1+νp−λ)/p. With respect to the functions p(α,β) n we prove the following result. FOURIER -- JACOBI EXPANSIONS IN MORREY SPACES 5 Lemma 2. Let 1 ≤ p < ∞ and α, β ≥ 0. Then (6) kp(α,β) n for 0 < λ ≤ 3/4, and (7) kp,λ ≃(1, kp,0 ≃  n kp(α,β) n1/2−2(1−λ)/p, 1, (log n)1/4, n1/2−2/p, if p ≤ 4(1 − λ), if p > 4(1 − λ), if p < 4, if p = 4, if p > 4. The estimates in the previous lemma will be deduced by using a very sharp bound for the Jacobi polynomials arising from a Hilb type formula for Jacobi polynomials (see [17, Theorem 8.21.12]). In fact, we have n n1/2P (α,β) (x) ≤ C(1 − x + n−2)−α/2−1/4(1 + x + n−2)−β/2−1/4, (8) for α, β > −1. Then, for α, β ≥ 0, we deduce in an easy way (note that d(α,β) n1/2) the bounds n ∼ (9) where h±,n(x) = (1 ± x + n−2)−1/4, and n p(α,β) (x) ≤ C(h+,n(x) + h−,n(x)), (10) for α, β ≥ 0. p(α,β) n (x) ≤ C(1 − x2)−1/4, To complete the boundedness of Sn, we will use Lemma 2 and to that end we q,λ for the functions h±,n. That is the content of the next have to estimate kgk∗ lemma. Lemma 3. Let 1 ≤ p < ∞ and α, β > 0. Then kh±,nk∗ q,λ ≤ C, for p ≥ 4(1 − λ)/3. Moreover, to analyze the norm of some functions in Morrey spaces, we have to recall the following theorem from [7]. Theorem 5. Let dµ = w(x) dx be a measure on (−1, 1). If {pn}n≥0 is the sequence of ortonormal polynomials on L2((−1, 1), dµ), then (cid:16)Z 1 −1(cid:12)(cid:12)(cid:12) g(x) w(x)1/2(1 − x2)1/4(cid:12)(cid:12)(cid:12) p w(x) dx(cid:17)1/p for 0 < p ≤ ∞ and for each measurable function g on (−1, 1). ≤ C lim inf n→∞ (cid:16)Z 1 −1 pn(x)g(x)pw(x) dx(cid:17)1/p , As an immediate consequence of the previous result, we have (11) k(1 − x2)−1/4g(x)kp,λ ≤ C lim inf n→∞ kp(α,β) n (x)g(x)kp,λ. The last tool that we need to complete the proof of the theorem is an easy observation related to the boundedness of the Hilbert transform with weights that will be used to complete the necessity of the conditions (3). Lemma 4. Let H be the Hilbert transform on (−1, 1) and let us suppose that kuHgkp,λ ≤ Ckvgkp,λ, vg ∈ Lp,λ(−1, 1), for two nonnegative weights u and v. Then for r, s ∈ (−1, 1) such that r ≤ s we have the inequality (12) (cid:18)Z r −1 f (t) s − t dt(cid:19) kuχ(r,s)kp,λ ≤ Ckvf χ(−1,r)kp,λ, vf ∈ Lp,λ(−1, 1), 6 A. ARENAS AND ´O. CIAURRI with the same constant C. The proofs of Lemma 2, Lemma 3, and Lemma 4 will be done in the last section. Proof of Theorem 3. Let us start by assuming the conditions (3) to prove the uni- form boundedness of the partial sum operators Sn. They can be written as −1 Snf (x) =Z 1 Xk=0 n Kn(x, t) = Kn(x, t)f (t) dt, p(α,β) k (x)p(α,β) k (t). Now, we consider Pollard decomposition of the kernel, see [11], Kn(x, t) = anT1(n, x, t) + anT2(n, x, t) + bnT3(n, x, t), where T1(n, x, t) = p(α,β) n+1 (x) T2(n, x, t) = T1(n, t, x) and T3(n, x, t) = p(α,β) an and bn are bounded. This leads to (1 − t2)1/2p(α+1,β+1) n (t) , x − t n+1 (x)p(α,β) n+1 (t). Moreover the sequences where with and Snf (x) = anW1,nf (x) − anW2,nf (x) + bnW3,nf (x), W1,nf (x) = p(α,β) W2,nf (x) = (1 − x2)1/2p(α+1,β+1) n+1 (x)H((1 − (·)2)1/2p(α+1,β+1) (x)H(p(α,β) n n+1 f )(x), n f )(x), W3,nf (x) = p(α,β) p(α,β) n+1 (t)f (t) dt. n+1 (x)Z 1 −1 Now, taking into account the estimate (10), the boundedness in Lp,λ(−1, 1) of W1,n and W2,n will follow from the inequalities k(1 − x2)±1/4Hg(x))kp,λ ≤ Ck(1 − x2)±1/4g(x))kp,λ. Then, by using (1), it is enough that λ − 1 p < −1 4 < 1 + λ − 1 p and λ − 1 p < 1 4 < 1 + λ − 1 p , and this is implied by (3). To treat W3,n we start by using (4), then kW3,nf kp,λ ≤ kp(α,β) p(α,β) n+1 (t)f (t) dt ≤ kp(α,β) n+1 kp,λkp(α,β) n+1 k∗ q,λkf kp,λ. n+1 kp,λZ 1 −1 Now, from (8), we have kp(α,β) the estimate it is enough to apply Lemma 2 and Lemma 3. Indeed, q,λ ≤ C(kh−,n+1k∗ q,λ+kh+,n+1k∗ n+1 k∗ q,λ). So, to conclude kW3,nf kp,λ ≤ Ckp(α,β) n+1 kp,λ(kh−,n+1k∗ q,λ + kh+,n+1k∗ q,λ)kf kp,λ ≤ Ckf kp,λ. Let us show that conditions (3) are necessary for the uniform boundedness. We consider the operators Tnf (x) = Snf (x) − Sn−1f (x), then it is clear that Tnf (x) = a(α,β) n (f )p(α,β) n (x). FOURIER -- JACOBI EXPANSIONS IN MORREY SPACES 7 By using the uniform boundedness for Sn, we have (13) kTnf kp,λ = a(α,β) n (f )kp(α,β) n kp,λ ≤ Ckf kp,λ, with a constant C independent of n and f . Taking f = sgn(p(α,β) n ), (13) becomes kp(α,β) n k1,0kp(α,β) n kp,λ ≤ C. So, by using (7), the inequality kp(α,β) implies, by (6), p ≤ 4(1 − λ). n kp,λ ≤ C has to be verified and this fact Now, by considering f = sgn(p(α,β) n )p(α,β) n q−1, with p−1 + q−1 = 1, in (13), we have kp(α,β) n kq q,0kp(α,β) n kp,λkp(α,β) n k−q/p q,λ ≤ C. Analyzing the different cases of the previous inequality for p ≤ 4(1 − λ) with the estimates in Lemma 2, we deduce the restriction p ≥ 4(1 − λ)/3. To conclude the necessity of (3), we have to check that for the cases p = 4(1−λ)/3 and p = 4(1 − λ) the inequality (2) is not possible. For p = 4(1 − λ) the operators W2,n and W3,n are uniformly bounded. Let us see that W1,n is unbounded. We proceed by contradiction. If W1,n is bounded, we have the equivalent estimate kp(α,β) n+1 Hgkp,λ ≤ Ck(1 − (·)2)−1/2(p(α+1,β+1) Now, from (12) with s = 1 and f (t) = χ(0,r)(t)p(α+1,β+1) we have n n )−1gkp,λ. (t)(1 − t2)1/4, for r > 0, p(α+1,β+1) n (t)(1 − t2)1/4 1 − t dtkp(α,β) n χ(r,1)kp,λ ≤ Ck(1 − t2)−1/4χ(0,r)kp,λ. Z r 0 By using (11), we obtain the inequality Z r 0 dt 1 − t k(1 − t)−1/4χ(r,1)kp,λ ≤ Ck(1 − t)−1/4χ(0,r)kp,λ, which, by (5), is equivalent to − log(1 − r) ≤ C and this is impossible. In the case p = 4(1 − λ)/3, W1,n and W3,n are bounded. Let us suppose that W2,n is also bounded. Thus, we have the equivalent inequality k(1 − (·))1/2p(α+1,β+1) n+1 Hgkp,λ ≤ k(p(α,β) n )−1gkp,λ. Then we can consider (12) with s = 1, r = 1/2 and f (t) = (p(α,β) obtain n (t))4(1 − t), to (14) Z 1/2 −1 (p(α,β) n (t))4 dtk(1 − (·)2)1/2p(α+1,β+1) n χ(1/2,1)kp,λ ≤ Ck(1 − t)(p(α,β) n )3χ(−1,1/2)kp,λ. It is clear that k(1 − t)(p(α,β) n )3χ(−1,1/2)kp,λ ≤ Ck(1 + t)1/4χ(−1,1/2)kp,λ ≤ C, where in the last step we have used (5). Following the estimate in [6] to evaluate the Lp-norm of Jacobi polynomials, we have Z 1/2 −1 (p(α,β) n (t))4 dt ∼ log n. 8 A. ARENAS AND ´O. CIAURRI Finally, we observe that, for any ε > 0 and n big enough, k(1 − (·)2)1/2p(α+1,β+1) χ(1/2,1)kp,λ ≥ lim inf k(1 − (·)2)1/2p(α+1,β+1) n n n→ χ(1/2,1)kp,λ − ε by using again (5). In this way, (14) implies log n ≤ C and this is not possible. (cid:3) ≥ k(1 − (·)2)1/4χ(1/2,1)kp,λ − ε ≥ C, 4. Proof of Lemmas Proof of Lemma 2. The case λ = 0 is well known and it can be done, with the proper modifications, following the ideas in [6, Proposition 1], so we omit it. From (9), it will be enough to study the norms kh±,nkp,λ. We will focus on h−,n, the other can be analyzed in the same way. We have kh−,nkp,λ = sup x∈(−1,1),r>0(cid:16)r−λnp/2−2I(cid:17)1/p , with I =R t s (z + 1)−p/4 dz, s = (1 − x − r)n2 and t = (1 − x + r)n2. Note that in our situation we can consider r bounded by a positive value. We will analyze different cases. When r ≤ n−2, with the estimate I ≤ C(t − s) we have r−λnp/2−2I ≤ Cr1−λnp/2 ≤ np/2−2(1−λ). Now, when r > n−2, we consider the cases s ≤ 1 < t and 1 ≤ s < t (when s < t ≤ 1 it is verified that r ≤ n−2 and this case has been already treated). If s ≤ 1 < t and r > n−2, for p ≤ 4(1 − λ), we have I ≤ Ct1−p/4. So, using that in this case 1 − x ≤ Cr, we obtain r−λnp/2−2I ≤ Cr−λ(1 − x + r)1−p/4 ≤ Cr1−λ−p/4 ≤ C. For p > 4(1 − λ), it is clear that I ≤ CtλR t s (1 + z)−p/4−λ dz ≤ Ct−λ and r−λnp/2−2I ≤ Cr−λ(1 − x + r)λnp/2−2(1−λ) ≤ Cnp/2−2(1−λ). Finally, for 1 ≤ s < t and r < n−2, when p ≤ 4(1 − λ) we can check easily that I ≤ C(t1−p/4 − s1−p/4) ≃ t−p/4(t − s). Then r−λnp/2−2I ≤ Cr1−λ(1 − x + r)−p/4 ≤ Cr1−λ−p/4 ≤ C. For p > 4(1 − λ), we use that I ≤ CtλR t tλ−1s1−p/4−λ(t − s). In this way s z−p/4−λ ≃ tλ(s1−p/4−λ − t1−p/4−λ) ≃ r−λnp/2−2I ≤ C(1 − x − r)1−p/4−λ ≤ Cnp/2−2(1−λ), and the proof of the upper bound is completed. To obtain the lower estimate in the case p ≤ 4(1 − λ) we use that, for ε > 0 and n big enough, by (11), kp(α,β) n kp,λ ≥ Ck(1 − t2)−1/4kp,λ − ε. Then, the required estimate follows by (5). In the case p > 4(1 − λ), we will use that P (α,β) n2 < x < 1. This is a well known consequence of the Hilb type formula for the Jacobi polynomials. So, (x) ≥ nα, for 1 − 1 n kp(α,β) n kp,λ ≥ C(cid:16)n2λ+p(α+1/2)Z 1 1−1/n2 (1 − x)pα/2 dx(cid:17)1/p ≥ Cn1/2−2(1−λ)/p, FOURIER -- JACOBI EXPANSIONS IN MORREY SPACES and the proof is finished. Proof of Lemma 3. We analyze h−,n, the other case is similar. It is clear that 9 (cid:3) kh−,nk∗ q,λ ≤ = x∈(1−1/n2,1)Z ∞ inf 0 inf (J1 + J2), x∈(1−1/n2,1) rλ/p−1kχ(B(x,r))ch−,nkLq(−1,1) dr where and 0 J1 =Z 1−x J2 =Z 1+x 1−x rλ/p−1kχ(−1,x−r)∪(x+r,1)h−,nkLq(−1,1) dr rλ/p−1kχ(−1,x−r)h−,nkLq(−1,1) dr. It is easy to check that kh−,nkLq(−1,1) can be controlled by the right hand side in (7) and then J1 ≤ Ckh−,nkLq(−1,1)Z 1/n2 0 rλ/p−1 dr ≤ Cn−2λ/pkh−,nkLq(−1,1) ≤ C, where in the last step we have used the condition p ≥ 4(1 − λ)/3. For J2, when q < 4 we have J2 ≤ C because J2 ≤ Ckh−,nkLq(−1,1)Z 1+x 1−x In the case q > 4, we have rλ/p−1 dr ≤ C. kχ(−1,x−r)h−,nkLq(−1,1) ≤ C((1 + n−2)1/q−1/4 + (1 − x + r + n−2)1/q−1/4) and rλ/p−1((1 + n−2)1/q−1/4 + (1 − x + r + n−2)1/q−1/4) dr 1−x J2 ≤ CZ 1+x ≤ C(cid:18)1 +Z 1+x ≤ C(cid:18)1 +Z 1+x 1−x 1−x rλ/p−1(1 − x + r + n−2)1/q−1/4) dr(cid:19) rλ/p+1/q−5/4 dr(cid:19) ≤ C(1 + (1 + x)λ/p+1/q−1/4 + (1 − x)λ/p+1/q−1/4) ≤ C, where in the last step we have used the condition p ≥ 4(1 − λ)/3. When q = 4, J2 ≤ CZ 1+x 1−x r3λ/4−1 log(cid:18) 2n2 + 1 1 + n2(1 − x + r)(cid:19) dr. Then, after applying integration by parts, we deduce that J2 ≤ C(cid:18)1 + n2(1 − x)3λ/4 +Z 1+x ≤ C(cid:18)1 + n2(1 − x)3λ/4 +Z 1+x 1−x 1−x ≤ C(1 + n2(1 − x)3λ/4 + (1 − x)3λ/4) r3λ/4−1 dr(cid:19) n2r3λ/4 1 + n2(1 − x + r) dr(cid:19) and inf x∈(1−1/n2,1)(J1 + J2) ≤ C. (cid:3) 10 A. ARENAS AND ´O. CIAURRI Proof of Lemma 4. If g(t) = χ(−1,r)(t)f (t), for a function f such that vf ∈ Lp,λ(−1, 1), for r ≤ r ≤ s, we have −1 f (t) x − t Hg(x) =Z r dt ≥Z r Hg(x) ≥ χ(s,r)(x)Z r −1 f (t) s − t dt. −1 f (t) s − t dt and then In this way, the boundedness of the Hilbert transform implies (12). (cid:3) References [1] D. R. Adams and J. Xiao, Morrey spaces in harmonic analysis, Ark. Math. 50 (2012), 201 -- 230. [2] S. Chanillo, On the weak behaviour of partial sums of Legendre series, Trans. Amer. Math. Soc. 268 (1981), 367-376. [3] A. Gogatishvili and R. Ch. Mustafayev, New pre-dual space of Morrey space, J. Math. Anal. Appl. 397 (2013), 678-692. [4] J. J. Guadalupe, M. P´erez, and J. L. Varona, Weighted weak behaviour of Fourier -- Jacobi series, Math. Nachr. 158 (1992), 161 -- 174. [5] J. J. Guadalupe, M. P´erez, F. J. Ruiz, and J. L. Varona, Two notes on convergence and divergence a.e. of Fourier series with respect to some orthogonal systems, Proc. Amer. Math. Soc. 116 (1992), 457-464. [6] J. Levesley and A. K. Kushpel, On the norm of the Fourier-Jacobi projection, Numer. Funct. Anal. Optim. 22 (2001), 941-952. [7] A. M´at´e, G. P. Nevai, and V. Totik, Neccesary conditions for weighted mean convergence of Fourier series in orthogonal polynomials, J. Approx. Theory 46 (1986), 314 -- 322. [8] C. Meaney, Divergent Jacobi polynomial series, Proc. Amer. Math. Soc. 87 (1983), 459-462. [9] C. B. Morrey, On the solutions of quasi-linear elliptic partial differential equations, Trans. Amer. Math. Soc. 43 (1938), 126 -- 166. [10] B. Muckenhoupt, Mean convergence of Jacobi series, Proc. Amer. Math. Soc. 23 (1969), 306 -- 310. [11] H. Pollard, The mean convergence of orthogonal series II, Trans. Amer. Math. Soc. 63 (1948), 355-367. [12] H. Pollard, The mean convergence of orthogonal series III, Duke Math. J. 16 (1949), 189-191. [13] M. Rosenthal and H. Triebel, Calder´on-Zygmund operators in Morrey spaces, Rev. Mat. Complut., to appear. DOI 10.1007/s13163-013-0125-3 [14] W. Rudin, "Real and complex analysis", 3rd edition, McGraw-Hill International Editions, Singapore, 1987. [15] N. Samko, Weighted Hardy and singular operators in Morrey spaces, J. Math. Anal. Appl. 350 (2009), 56-72. [16] N. Samko, On a Muckenhoupt-Type Condition for Morrey Spaces, Mediterr. J. Math. 10 (2013), 941-951. [17] G. Szego, "Orthogonal polynomials", 3rd edition, Amer. Math. Soc. Colloq. Publ., vol. 23. Amer. Math. Soc., Providence, R.I., 1967. [18] G. M. Wing, The mean convergence of orthogonal series, Amer. J. Math. 72 (1950), 792-808. Departamento de Matem´aticas y Computaci´on, Universidad de La Rioja, Calle Luis de Ulloa s/n, 26004 Logrono, Spain E-mail address: [email protected] Departamento de Matem´aticas y Computaci´on, Universidad de La Rioja, Calle Luis de Ulloa s/n, 26004 Logrono, Spain E-mail address: [email protected]
1004.1527	1	1004	2010-04-09T11:01:06	Slowly Changing Vectors and the Asymptotic Finite-Dimensionality of an Operator Semigroup	[ "math.FA" ]	Let $T:X\to X$ be a linear power bounded operator on Banach space. Let $X_0$ is a subspace of vectors tending to zero under iterating of $T$. We prove that if $X_0$ is not equal to $X$ then there exists $\lambda$ in Sp(T) such that, for every $\epsilon>0$, there is $x$ such that $\|Tx-\lambda x\|<\epsilon $ but $\|T^nx\|>1-\epsilon$ for all $n$. The technique we develop enables us to establish that if $X$ is reflexive and there exists a compactum $K$ in $X$ such that for every norm-one $x\in X$ $\rho\{T^nx, K\}<\alpha (T)<1$ for some $n=n_1, n_2,...$ then $codim(X_0)<\infty$. The results hold also for a one-parameter semigroup.	math.FA	math	SLOWLY CHANGING VECTORS AND THE ASYMPTOTIC FINITE-DIMENSIONALITY OF AN OPERATOR SEMIGROUP K. V. STOROZHUK Abstract. Let X be a Banach space and let T : X → X be a linear power bounded operator. Put X0 = {x ∈ X T nx → 0}. We prove that if X0 6= X then there exists λ ∈ Sp(T ) such that, for every ε > 0, there is x such that kT x − λxk < ε but kT nxk > 1 − ε for all n. The technique we develop enables us to establish that if X is reflexive and there exists a compactum K ⊂ X such that lim inf n→∞ ρ{T nx, K} < α(T ) < 1 for every norm-one x ∈ X then codim X0 < ∞. The results hold also for a one-parameter semigroup. 0 1 0 2 r p A 9 ] . A F h t a m [ 1 v 7 2 5 1 . 4 0 0 1 : v i X r a 1. Introduction In this article, X is a complex Banach space, T : X → X is a linear operator whose all powers are bounded by a constant C < ∞. We use the notations: BX is the unit ball in X, Γ is the unit circle in C, X0 = {x ∈ X T nx →n→∞ 0}. A vector x is called an ε-almost eigenvector (or simply an ε-eigenvector) if there exists λ ∈ C such that kT x − λxk < ε. These vectors exist for each λ ∈ Sp(T ) ∩ Γ. In Section 1, we study ε-eigenvectors that do not shorten much under the iterations T n. Definition 1. Suppose that ε > 0. Call a vector x ∈ X ε-slow if ∃λ ∈ Γ kT x − λxk < ε and kT nxk > 1 − ε ∀n = 0, 1, 2 . . . . For example, the eigenvectors x, T x = λx, λ ∈ Γ, are slow. Example 1. T : l2 → l2 is the right shift, T (x1, x2, . . . ) = (0, x1, x2, . . . ). This is an isometry; hence, every norm-one ε-eigenvector is ε-slow. Example 1∗. T : l2 → l2 is the left shift, T (x1, x2, . . . ) = (x2, x3, . . . ). The spectrum includes Γ but T nx → 0 for every x and there are no slow vectors. If a vector x is ε-slow then T nx are Cε-slow for each n, since T (T nx) − λT nx = T n(T x − λx). Thus, the angle between the (complex) lines T nx and T n+1x is slow not only for n = 0 but for all n ∈ N (i.e., the vector changes slowly under the iterations). Remark. Our terminology is by no means connected with the terms "slow vector" and "slow variable" of the classical theory of dynamical systems. In the title, we call slow vectors slowly changing. Definition 2. An operator T has slow vectors if, for every ε > 0, there exist ε-slow vectors. An operator T has many slow vectors if dim X = ∞ and, for every ε > 0 and n < ∞, there exist n-dimensional subspaces in X whose unit spheres consist of ε-slow vectors. If the powers kT nk are bounded below, i.e., there exists a number c such that ckxk ≤ kT nxk for every x, then there are many slow vectors. Indeed, if kxk = 1 and x is a cε-eigenvector then x c is ε-slow. 1 2 K.V. Storozhuk If X0 = X then it is clear that there are no slow vectors. It turns out that the condition X0 = X is the only obstacle to the existence of slow vectors; if codim X0 = ∞ then there are many slow vectors (Theorem 1.1). In Section 2, slow vectors are used in the study of the asymptotic properties of T n. It is known that if there exists an attracting compact set K, i.e., such that ∀x ∈ BX lim n→∞ ρ(T nx, K) = 0 then X = X0 ⊕ L, where L is a finite-dimensional invariant subspace in X. This was proved in [1] for Markov semigroups in L1. For an arbitrary Banach space, this was established in [2, 3]. In [4] it was proved that for the splitting X = X0 ⊕ L, dim L < ∞, it suffices that a compact set K attract only sometimes, i.e., ∀x ∈ BX lim inf n→∞ ρ(T nx, K) = 0. A semigroup T n : X → X or a one-parameter semigroup Tt : X → X is called asymptotically finite-dimensional [5] if codim X0 < ∞. In [6, 1.3.33] the question is posed whether a semigroup is asymptotically finite-dimensional if ∀x ∈ BX lim inf n→∞ ρ(T nx, K) < α(T ) < 1. (∗) Clearly, the condition (∗) follows from the above-listed analogous conditions. In Section 2 of the article, we prove that if X is reflexive then an operator satisfying (∗) has few slow vectors and codim X0 < ∞ (Theorem 2.3). Thus, we give a partial positive answer to the question of [6]. Theorem 2.3 is easily generalized to the case of a one-parameter semigroup of {Tt : X → X, t ≥ 0}. Note that it is in the reflexive case that the condition codim X0 < ∞ for a bounded semigroup implies the splitting X0 ⊕ L [7]. For nonreflexive X, the author does not know the answer to the question of [6, 1.3.33]. 2. Slow Vectors Theorem 1.1. Suppose that X is a Banach space, T : X → X, kT nk < C. If X0 6= X then T has slow vectors. If codim X0 = ∞ then T has many slow vectors. Proof. Without loss of generality, by passing to the equivalent norm kxk := supn{kT nxk}, we may assume that kT k ≤ 1. The scheme of the proof is as follows: (a) X0 = 0 ⇒ slow vectors exist; (b) X0 = 0 ⇒ there are many slow vectors; (c) codim X0 = ∞ ⇒ there are many slow vectors. (a) Introduce the norm kxkp := lim n→∞ kT nxk on X. In this norm, T is an isometry. The norms k kp and k k need not be equivalent but kT kxk ∼k→∞ kT kxkp (1.1) for all x. If the powers of T are not bounded below then the space (X, kkp) is incomplete. Let bX be the completion of X in the norm k kp. Extend the isometry T of (X, k kp) to bX and denote the extension by bT . Take λ ∈ Sp(bT ) ∩ Γ. Suppose that ε > 0 and bx ∈ bX is a k kp-one ε-eigenvector corresponding to λ. Involving the fact that bT : bX → bX is an isometry, we get kbTbx − λbxkp < ε and ∀n kbT nbxkp = kbxkp = 1 > 1 − ε. Slowly Changing Vectors 3 The set X is dense in bX. all strict inequalities of the last formula remain valid. Thus, there exists a vector x ∈ X such that If a vector x ∈ X is sufficiently k kp-close to bx then kT x − λxkp < ε and ∀n kT nxkp > 1 − ε. (1.2) Thus, x is an ε-slow vector of T in the norm k kp. Of course, x need not be slow in the initial norm, since kT x−λxk may be large. However, applying the k kp-isometry T k to (1.2), we infer ∀k kT (T kx) − λT kxkp < ε and ∀n kT n(T kx)kp > 1 − ε. (1.3) By (1.1), starting from some k = k0, inequalities (1.3) also hold for the norm k k, i.e., starting from some k, the vector T kx is as well slow in the initial norm of X. (b) Take a number λ ∈ Γ in the spectrum of the isometry bT of (X, k kp); λ has many ε-eigenvectors (there exist even infinite-dimensional spheres of ε-eigenvectors (see [8, Chapter IV, Theorems 5.33, 5.9]). Let l < ∞ and let W be an l-dimensional subspace in (X, k kp) whose k kp-unit sphere S consists of ε-eigenvectors. Perturbing W slightly, we may assume that W ⊂ X. All vectors in S satisfy (1.2) and (1.3). By (1.1), for each x ∈ S, all the vectors T k0x are slow vectors for the operator T : X → X starting from some k0. The ellipsoid S is compact; therefore, k0 may chosen common for all x ∈ S. Thus, X includes (l − 1)-dimensional ellipsoids of the form T k0(S) consisting of small vectors. (c) Consider the quotient space X/X0. Its elements are [x] := x + X0. The norm k[x]k is as follows: k[x]k = ρ(x, X0) = inf{kx − x0k, x0 ∈ X0}. We have T (X0) ⊂ X0; therefore, the operator [T ] : X/X0 → X/X0 is defined. Clearly, [T ]n = [T n]. It is easy to see that if [x] 6= [0] then [T ]n[x] 6→ 0. By (b), [T ] has many slow vectors, i.e., for each l, in X/X0, there are l-dimensional ellipsoids of slow vectors [x] for [T ] (moreover, we may assume that these ellipsoids have the form [S], where S is an ellipsoid in X): ∀n ≥ 0 k[T ]n[x]k > 1 − ε and k[T ][x] − λ[x]k < ε ∀[x] ∈ [S]. Since T kx0 → 0 for all x0 ∈ X0, we have ∀n ≥ 0 kT n(T kx)k > 1 − ε and kT (T kx) − λT kxk <k→∞ ε (1.4) for every x ∈ [x] = x + X0 ∈ [S]. The compactness of S enables us to assert now that, starting from some k, the (cid:3) ellipsoids T kS ⊂ X consist of slow vectors. In what follows, we will need some properties of slow vectors. Denote by Sm,λ the operator Lemma 1.2. Suppose that δ > 0, m ∈ N. If T has slow vectors then there exists λm(cid:1), the Cesaro mean of T /λ. m+1(cid:0)Pm 1 T m i=0 λ ∈ Γ such that ∃x ∈ BX kSm,λx − xk < δ and ∀n kT n(Sm,λx)k > 1 − δ. (1.5) If T has many slow vectors then there exist subspaces W ⊂ X of an arbitrarily large dimension whose unit spheres S consist of vectors x satisfying the inequality of (1.5). Proof. Involving a compactness of Γ, consider λ ∈ Γ to which there correspond slow vectors. If ε := ε(δ, m) is very small and x is an ε-slow vector corresponding 4 K.V. Storozhuk to λ then Sm,λ(x) ≈ x and Sm,λ(x) satisfies (1.5). The remaining part of the proof is obvious. (cid:3) Remark. Geometrically, Lemma 1.2 means that, for each m, there exist spheres of an arbitrarily large dimension that almost do not flatten under the mappings T n(Sm,λ) for any n. 3. Asymptotic Finite-Dimensionality in the Reflexive Case Throughout the section, we suppose that T : X → X satisfies (∗). We may assume that K is a balanced set. Lemma 2.1. Suppose that x ∈ BX . For each k, there exist vectors a1, . . . , ak ∈ K, numbers m1 > m2 > · · · > mk, and t1, . . . , tk, ti ≤ αi−1, such that kT m1x − [t1T m2a1 + t2T m3a2 + · · · + tk−1T mk ak−1 + tkak]k ≤ αk. (2.1) Proof. We write down some inequalities for k = 1, 2, 3. The first is condition (∗), and the validity of each subsequent inequality is guaranteed by an application (∗) to the preceding inequality multiplied by α: ∃n1 kT n1x − t1a1k ≤ α, t1 ≤ 1, ∃n2 kT n2(T n1x − t1a1) − t2a2k ≤ α2, t2 ≤ α, ∃n3 kT n3(T n2(T n1x − t1a1) − t2a2) − t3a3k ≤ α3, t3 ≤ α2, . . . . To finish, it remains to remove parenthesis and put mj = nj + · · · + nk. The sum of the numbers ti in (2.1) does not exceed h := Pk Hence, the convex hull bK of S∞ ∀x ∈ BX ∀ε > 0 ∃n ∈ N ∃a ∈ bK kT nx − ak < ε. i=0 T i(hK) attracts BX , i.e., We now show that T cannot act by multiplication by a scalar on the subspaces (cid:3) i=1 αi = 1 1−α . (2.2) X whose dimension is rather high. Theorem 2.2. dim ker(T − λI) < ∞ for all λ ∈ Γ. Proof. Choose a finite (1 − α)-net of k vectors for K and consider the subspace Y spanned by the net. By the Kreın -- Krasnosel′skiı -- Milman Theorem [9], in every subspace Z ⊂ X such that dim Z > dim Y , there exists a norm-one vector z such that ρ(z, Y ) = 1. By (∗), ρ(T nz, Y ) < α + (1 − α) = 1 for some n. Therefore, T z cannot have the form λ · z. (cid:3) Theorem 2.3. Suppose that X is reflexive. They T cannot have many slow vectors and so codim X0 < ∞. Proof. It suffices to prove that to no λ ∈ Γ there correspond many slow vectors. We may assume that λ = 1. By the Statistical Ergodic Theorem (see, for example, [10, § 2]), the operator means Sm,1 = Sm = 1 m+1(cid:18) mP0 T k(cid:19) converge to the projection P of X onto ker(I − T ). By Theorem 2.2, dim ker(I − T ) < ∞. On a compact set K, the convergence of Sm − P to zero is uniform (for ex- ample, by Arzel`a's Theorem). The operators Sm commute with T ; therefore, the convergence (Sm − P ) → 0 is also uniform on bK. Hence, starting from some m, kSm(a) − P (a)k is sufficiently small for all a ∈ bK, for example, less than 1 (2.2), for every x ∈ B and large n, the vector T nx is close to bK. Therefore, ∃m∀x ∈ BX k(Sm − P )T nxk = kT n(Smx) − P xk ≤n→∞ 1/3. 3 . But by Slowly Changing Vectors 5 This implies, for example, that, under T n ◦ Sm, every k-dimensional sphere such that k > dim ker(I − T ), "flattens" three times for large n along some radius x (x must be chosen so that P x = 0). This, by Lemma 1.2 and the remark thereto, means that the number λ = 1 does not have many slow vectors. The inequality codim X0 < ∞ follows now from Theorem 1.1. (cid:3) References [1] Lasota A., Li T. Y., and Yorke J. A. Asymptotic periodicity of the iterates of Markov oper- ators.// Trans. Amer. Math. Soc. 1984, v. 286, 2, 751 -- 764. [2] Phong Vu Quoc. Asymptotic almost-periodicity and compactifying representations of semi- groups.// Ukrainian Math. J. 1986, v. 38, 6, 576 -- 579. [3] Sine R. Constricted systems.// Rocky Mountain J. Math. 1991, v. 21, 4, 1373 -- 1383. [4] Storozhuk K. V. An extension of the Vu-Sine theorem and compact-supercyclicity.// J. Math. Anal. Appl., 2007, v. 332, 2, 1365 -- 1370. [5] Emel'yanov E. Yu. and Wolff Manfred P. H. Quasi-constricted linear operators on Banach spaces.// Studia Math. 2001, v. 144, 2, 169 -- 179. [6] Emel'yanov E. Yu., Non-Spectral Asymptotic Analysis of One-Parameter Operator Semi- groups, Birkhauser, Basel (2007) (Oper. Theory, Advances Appl.; 173). [7] Emel'yanov `E. Yu. Some conditions for a C0-semigroup to be asymptotically finite- dimensional.// Siberian Math. J. 2003, v. 44, 5, 793 -- 796. [8] Kato T. Perturbation Theory for Linear Operators. 1995 Springer-Verlag, Berlin etc. [9] Kreın M. G., Krasnosel ′skiı M. A., and Milman V. D. On the defect numbers of linear operators in Banach space and on some geometric properties.// Sb. Tr. Inst. Mat. Akad. Nauk USSR 1948, v. 11, 97 -- 112. [10] Lyubich Yu. I. Introduction to the Theory of Banach Group Representations. 1985 Vishcha Shkola (Izdat. pri Khar′kov. Gos. Univ.), Khar′kov (in Russian) Sobolev Institute of mathematics SB RAS
1812.06704	1	1812	2018-12-17T11:24:43	Exhaustive families of representations of $C^*$-algebras associated to $N$-body Hamiltonians with asymptotically homogeneous interactions	[ "math.FA" ]	We continue the analysis of algebras introduced by Georgescu, Nistor and their coauthors, in order to study $N$-body type Hamiltonians with interactions. More precisely, let $Y$ be a linear subspace of a finite dimensional Euclidean space $X$, and $v_Y$ be a continuous function on $X/Y$ that has uniform homogeneous radial limits at infinity. We consider, in this paper, Hamiltonians of the form $H = - \Delta + \sum_{Y \in S} v_Y$, where the subspaces $Y$ belong to some given family S of subspaces. We prove results on the spectral theory of the Hamiltonian when $S$ is any family of subspaces and extend those results to other operators affiliated to a larger algebra of pseudo-differential operators associated to the action of $X$ introduced by Connes. In addition, we exhibit Fredholm conditions for such elliptic operators. We also note that the algebras we consider answer a question of Melrose and Singer.	math.FA	math	EXHAUSTIVE FAMILIES OF REPRESENTATIONS OF C∗-ALGEBRAS ASSOCIATED TO N -BODY HAMILTONIANS WITH ASYMPTOTICALLY HOMOGENEOUS INTERACTIONS J´ER´EMY MOUGEL AND NICOLAS PRUDHON Abstract. We continue the analysis of algebras introduced by Georgescu, Nistor and their coauthors, in order to study N -body type Hamiltonians with interactions. More precisely, let Y ⊂ X be a linear subspace of a finite di- mensional Euclidean space X, and vY be a continuous function on X/Y that has uniform homogeneous radial limits at infinity. We consider, in this pa- per, Hamiltonians of the form H = −∆ + PY ∈S vY , where the subspaces Y ⊂ X belong to some given family S of subspaces. Georgescu and Nistor have considered the case when S consists of all subspaces Y ⊂ X, and Nis- tor and the authors considered the case when S is a finite semi lattice and Georgescu generalized these results to any families. In this paper, we develop new techniques to prove their results on the spectral theory of the Hamiltonian to the case where S is any family of subspaces also, and extend those results to other operators affiliated to a larger algebra of pseudo-differential operators associated to the action of X introduced by Connes. In addition, we exhibit Fredholm conditions for such elliptic operators. We also note that the algebras we consider answer a question of Melrose and Singer. An new approach in the study of Hamiltonians of N -body type with interactions that are asymptotically homogeneous at infinity on a finite dimensional Euclidean space X was initiated by Georgescu and Nistor [3, 5, 6]. For any finite real vector space Z, we let Z denote its spherical compactification. A function in C(Z) is thus a continuous function on Z that has uniform radial limits at infinity. Let SZ be the set of half-lines in Z, that is SZ := { a, a ∈ Z, a 6= 0 } where a := {ra, r > 0}. We identify SZ = Z r Z . For any subspace Y ⊂ X, πY : X → X/Y denotes the canonical projection. Let (1) H = −∆ + XY ∈S vY , where vY ∈ C(X/Y ) is seen as a bounded continuous function on X via the projec- tion πY : X → X/Y . The sum is over all subspaces Y ⊂ X, Y ∈ S and is assumed to be uniformly convergent. One of the main results of [5, 9] describe the essential spectrum of H extending the celebrated HVZ theorem [13]. The goal of this paper is to explain how these results can be extended to any family of subspaces that contains {0} and to more general operators using C∗-algebras techniques. Let S be a family of subspaces of X with 0 ∈ S. We define the commutative sub- C∗-algebra ES(X) of the commutative C∗-algebra C u b (X) of bounded uniformly 1 2 J. MOUGEL AND N. PRUDHON continous functions on X by (2) ES(X) = hC(X/Y ) , Y ∈ S i ⊂ C u b (X). The algebras ES(X) give an answer to a question of Melrose and Singer [8]. Theorem 1. Let n be an integer. Let S n be the semi-lattice of subspaces of X n generated by S n i ∪ S n ij where S n i = {(x1, . . . , xn) ∈ X n ; xi = 0} S n ij = {(x1, . . . , xn) ∈ X n ; xi = xj} Then the spectrum ΩS n of ES n (X n) is a compactification of X n satisfying the fol- lowing properties : (1) ΩS 1 is the spherical compactification X, (2) The action of the symmetric group Sn on X n extends continuously to ΩSn , (3) The projections pn,k (x1, . . . , xn) = (xi1 , . . . , xik ) extend I continuously to pn,k : X n → X k, pn,k : ΩSn → ΩSk , I I (4) The difference maps δij (x1, . . . , xn) = xi − xj from X n to X extend con- tinuously to the compactifications. Actually, the spectrum ΩS n have very strong connection with the space built by Vasy in [15] and generalized by Kottke in the last section of [7]. The additive group X acts by translation on C u invariant. So a crossed product C∗-algebra is obtained b (X) and the subalgebra ES(X) is ES(X) ⋊ X , (3) which can be regarded as an algebra of operators on L2(X). Thanks to the as- sumption 0 ∈ S, the algebra C0(X) belongs ES(X). Hence C0(X) ⋊ X is contained in ES(X) ⋊ X. It follows from the definition of crossed products algebras that the C∗-algebra ES(X) ⋊ X is generated by two kinds of operators : multiplication operators mf associated to functions f ∈ ES(X), and convolution operators Cφu(x) := ZX φ(y)u(x − y)dy with φ ∈ Cc(X), a continuous compactly supported function. An immediate com- putation shows that mf cφ (resp. cφmf ) is a kernel operator with kernel (4) K(x, y) = f (x)φ(y − x) , (resp. K(x, y) = f (y)φ(y − x)). Proposition 2. (i) The subalgebra C0(X) ⋊ X is the algebra K(X) of compact operators on L2(X). (ii) For f ∈ C(X) and φ ∈ Cc(X) the commutator [mf , cφ] is compact. The point (i) is a consequence of equation (4) because the kernel K has compact support when f does and the result follows by density. Again, thanks to equation (4), one sees that the commutator is a kernel operator with kernel K(x, y) = φ(y − x)(f (x) − f (y)) . Hence, in view of φ ∈ Cc(X), the support of K is contained in a band around the diagonal. The distance between the border of the band and the diagonal is bounded. Moreover, K goes to 0 at infinity because f has radial limits. So the commutator is a limit of Hilbert-Schmidt operators, and hence is compact. ESSENTIAL SPECTRUM 3 Recall that a self-adjoint operator P on L2(X) is said to be affiliated to a C∗- algrebra A of bounded operators on L2(X) if for some (and hence any) function h ∈ C0(R) then h(P ) belongs to A. For example,, it follows from the identity (H + i)−1 = (−∆ + i)−1(cid:0)1 + V (−∆ + i)−1(cid:1)−1 , that H is affiliated to ES(X)⋊X. More generally, for any C∗-algebra A, a morphism h : C0(R) → A is called an operator affiliated to A. Following Connes [2] and Baaj [1] we introduce the C∗-algebra of non positive order pseudo-differential operators ΨDO(ES(X), X) together with the symbol map exact sequence 0 → ES(X) ⋊ X → ΨDO(ES (X), X) σ0−→ C(SX , ES(X)) → 0 . Positive order pseudo-differential operators are examples of operators affiliated to the algebra of non positive order pseudo-differential operators ΨDO(ES(X), X). Let α ∈ SX . For each x ∈ X, we let (Txf )(y) = f (y − x) denote the translation on L2(X). For any operator P on L2(X), we let (5) τα(P ) = lim r→+∞ T ∗ raP Tra , if α = a ∈ SX , whenever the strong limit exists. Lemma 3. For f ∈ C(X/Y ) one has τα(f )(x) = (cid:26) f (x) f (πY (α)) if Y ⊃ α , else. We define Sα = {Y ∈ S ; α ⊂ Y }. It follows from the previous lemma that on ES(X), τα is the projection on the subalgebra ESα(X), τα : ES(X) → ESα(X) . Theorem 4. (1) Let P be a self-adjoint operator affiliated to ΨDO(ES(X), X) and α = a ∈ SX . Then the limit τα(P ) := limr→+∞ T ∗ raP Tra exists and Specess(P ) = ∪α∈SX Spec(τα(P )) . (2) Let P ∈ ΨDO(ES (X), X). Then P is a Fredholm operator if and only if P is elliptic (i.e. σ0(P ) is invertible) and for all α ∈ SX , τα(P ) is invertible. This extends theorems of [5, 9] in the following sense : only operators affiliated to ES(X) ⋊ X are considered there, and the relation is (6) Specess(H) = ∪α∈SX Spec(τα(H)) in [5]. In [9] only finite semi-lattice S are considered. The equation (6) means that the family (τα) is a faithful family of morphism of ES(X) ⋊ X. The stronger result of [9] is obtained by showing that the family (τα ⋊ X)α∈SX is actually an exhaustive family of representations of ES(X) ⋊ X, when S is a finite semi-lattice. In the framework of admissible locally compact group, decomposition of essential spectrum involving exhaustive families can be found in [10] [11]. In fact, by [12, Proposition 3.12], exhaustive families are also strictly spectral families in the following sense. Definition 5. [12, 14] (1) A family (φi)i∈I of morphisms of a C∗-algebra A is said to be exhaustive if any primitive ideal contains at least ker φi for some i ∈ I. 4 J. MOUGEL AND N. PRUDHON (2) A family (φi)i∈I of morphisms of a unital C∗-algebra A is said to be strictly spectral if (∀a ∈ A) Spec(a) = ∪i∈I Spec(φi(a)) Theorem 6. Let S be a family of subspaces of X with 0 ∈ S. Then the family (τα ⋊ X)α∈SX is an exhaustive family of ES(X) ⋊ X/K(X). Let us prove this result. Let π be an irreducible representation of ES(X) ⋊X/K(X). It extends to an irreducible representation of ES(X) ⋊ X as well as to their mul- tipliers algebras M(ES(X) ⋊ X/K(X)) and M(ES(X) ⋊ X). By proposition 2(i), one obtains the following commutative diagram: C(X) ֒→ ES(X) → M(ES(X) ⋊ X) (7) p ↓ C(SX ) φ ↓ → M(ES(X) ⋊ X/K(X)) π → B(Hπ) Lemma 7. The image φ(C(SX )) is central in M(ES(X) ⋊ X/K(X)). In fact it is enough to show that any f ∈ C(X) commutes with any element of ES(X) ⋊ X modulo a compact operator. But the result is true on the generators by Proposition 2(ii), so the lemma follows by density. By the Schur Lemma, we deduce that π ◦ φ is a character of C(SX ). Hence there exists some α ∈ SX such that πC(X) = χαI, where χα is the character of C(X) given by the evaluation at α ∈ SX . Proposition 8. One has ker τα = (ker χα)ES(X). Proof. We need to show that ES(X)/ ker τα = ESα(X) and ES(X)/(ker χα)ES (X) have the same characters. By definition, for any character χ of ESα(X), there exists a unique character χ′ of ES(X) such that χ′ = χ ◦ τα. In view of lemma 3, this is equivalent to the following : (8) (∀Y ∈ S, α 6⊂ Y, ∀u ∈ C(X/Y )) χ(u) = u(πY (α)). In particular, for Y = 0, we see that χC(X) = χα. Reciprocally it follows from [5, Lemma 6.7] that if χC(X) = χα then relation (8) is true. On the other hand, the characters of ES(X)/(ker χα)ES(X) are precisely the characters χ of ES(X) such that χC(X) = χα. So ker τα = (ker χα)ES(X) as claimed. (cid:3) Now if πC(X) = χα, one has ker π ⊃ (ker χα)ES(X) = ker τα. Finally, ker(τα ⋊ X) = (ker τα) ⋊ X ⊂ ker π . It follows that (τα ⋊ X)α∈SX is an exhaustive family of morphisms. Remark 9. The results presented here can easily be extended to pseudo-differential operators with matrix coefficients. For example, Dirac operators DV = D + V , with potentials V as in (1) may be considered and satisfy the condition of Theorem 4. See also [4, Example 6.35] for others physical interesting operators. ESSENTIAL SPECTRUM 5 Acknowledgments. The authors thank Victor Nistor for useful discussions. References [1] S. Baaj. Calcul pseudo-diff´erentiel et produits crois´es de C ∗-alg`ebres. I. C. R. Acad. Sci. Paris S´er. I Math., 307(11):581 -- 586, 1988. [2] A. Connes. C ∗ alg`ebres et g´eom´etrie diff´erentielle. C. R. Acad. Sci. Paris S´er. A-B, 290(13):A599 -- A604, 1980. [3] V. Georgescu. On the structure of the essential spectrum of elliptic operators on metric spaces. J. Funct. Anal., 260(6):1734 -- 1765, 2011. [4] V. Georgescu and A. Iftimovici. Localizations at infinity and essential spectrum of quantum Hamiltonians. I. General theory. Rev. Math. Phys., 18(4):417 -- 483, 2006. [5] V. Georgescu and V. Nistor. On the essential spectrum of N -body Hamiltonians with asymp- totically homogeneous interactions. J. Operator Theory, 77(2):333 -- 376, 2017. [6] Vladimir Georgescu. On the essential spectrum of elliptic differential operators. J. Math. Anal. Appl., 468(2):839 -- 864, 2018. [7] C. Kottke. Functorial compactification of linear spaces. 2017. arXiv :1712.03902. [8] R. Melrose and M. Singer. Scattering configuration spaces. 2008. arXiv :0808.2022. [9] J. Mougel, V. Nistor, and N. Prudhon. A refined HVZ-theorem for asymptotically homo- geneous interactions and finitely many collision planes. Rev. Roumaine Math. Pures Appl., 62(1):287 -- 308, 2017. [10] M. Mantoiu. C ∗-algebras, dynamical systems at infinity and the essential spectrum of gener- alized Schrodinger operators. J. Reine Angew. Math., 550:211 -- 229, 2002. [11] M. Mantoiu. Essential spectrum and Fredholm properties for operators on locally compact groups. J. Operator Theory, 77(2):481 -- 501, 2017. [12] V. Nistor and N. Prudhon. Exhaustive families of representations and spectra of pseudodif- ferential operators. J. Operator Theory, 78(2):247 -- 279, 2017. [13] M. Reed and B. Simon. Methods of modern mathematical physics. IV. Analysis of operators. Academic Press [Harcourt Brace Jovanovich Publishers], New York, 1978. [14] S. Roch. Algebras of approximation sequences: structure of fractal algebras. In Singular integral operators, factorization and applications, volume 142 of Oper. Theory Adv. Appl., pages 287 -- 310. Birkhauser, Basel, 2003. [15] A. Vasy. Propagation of singularities in many-body scattering. Ann. Sci. ´Ecole Norm. Sup. (4), 34(3):313 -- 402, 2001. Universit´e de Lorraine, UFR MIM, 3 rue Augustin Fresnel 57045 METZ, France E-mail address: [email protected] Universit´e de Lorraine, UFR MIM, 3 rue Augustin Fresnel 57045 METZ, France E-mail address: [email protected]
1803.05052	2	1803	2018-06-17T16:36:02	The weighted Property (A) and the greedy algorithm	[ "math.FA" ]	We investigate various aspects of the "weighted" greedy algorithm with respect to a Schauder basis. For a weight w, we describe w-greedy, w-almost-greedy and w-partially-greedy bases, and show some properties of w-semi-greedy bases. To achieve these goals, we introduce and study the w-Property (A).	math.FA	math	THE WEIGHTED PROPERTY (A) AND THE GREEDY ALGORITHM P. M. BERN ´A, S. J. DILWORTH, D. KUTZAROVA, T. OIKHBERG, AND B. WALLIS Abstract. We investigate various aspects of the "weighted" greedy algorithm with respect to a Schauder basis. For a weight w, we describe w-greedy, w-almost-greedy, and w-partially- greedy bases, and examine some properties of w-semi-greedy bases. To achieve these goals, we introduce and study the w-Property (A). 1. Introduction i (ej) = 1 if i = j and e∗ n=1, with biorthogonal functionals (e∗ n)∞ n=1; that is, nk} ≤ supn max{kenk,ke∗ In this paper, we investigate the operation of the "weighted" greedy algorithm, and its efficiency. Throughout, (X,k · k) is a real Banach space with a semi-normalized Schauder basis B = (en)∞ A1) 0 < c1 := inf n min{kenk,ke∗ A2) e∗ A3) X = span[ei : i ∈ N], i=1 aiei) = i=1 aiei. We denote by Kb the least value of K for which the preceding inequality holds, A4) kSmk ≤ K for every m, where (Sm)m are partial sum operators -- that is, Sm(P∞ Pm x =Pj e∗ We will refer to B as a basis. Of course, for every x ∈ X, there exists a unique expansion i (x) 6= 0}, A denotes the cardinality of a set j (x)ej. As usual, supp (x) = {i ∈ N : e∗ i (ej) = 0 for i 6= j, and call it the basis constant. nk} =: c2 < ∞, A and Nm = {A ⊂ N : A = m}, N<∞ = Nm. ∞[m=0 Further notations will be often used: if a and b are functions of some variable, a . b means that there exists a constant c > 0 such that a ≤ c· b; if A and B are subsets of N, A < B means that maxj∈A j < minj∈B j, PA is the projection operator, i.e, if A is a finite set, PA(Pj ajej) = Pj∈A ajej and P c A = I − PA is the complementary projection, 1εA =Pn∈A εnen for εn ∈ {±1} and if εn ≡ 1, we write 1A. Algorithm (TGA): for in x ∈ X we produce the sequence of greedy approximands In 1999, S. V. Konyagin and V. N. Temlyakov introduced in [16] the Thresholding Greedy Gm(x) = e∗ π(n)(x)eπ(n), mXn=1 2000 Mathematics Subject Classification. 46B15, 41A65. Key words and phrases: thresholding greedy algorithm, unconditional basis, Property (A), w-greedy bases. The first author was supported by a PhD fellowship FPI-UAM and the grants MTM-2016-76566-P (MINECO, Spain) and 19368/PI/14 (Fundaci´on S´eneca, Regi´on de Murcia, Spain). The second author was supported by the National Science Foundation under Grant Number DMS -- 1361461. The second and third authors were supported by the Workshop in Analysis and Probability at Texas A&M University in 2017. 1 2 P. M. BERN ´A, S. J. DILWORTH, D. KUTZAROVA, T. OIKHBERG, AND B. WALLIS π(i)(x) ≥ e∗ where π is a greedy ordering, that is, π : {1, 2, ...,supp x} −→ supp x is a bijection such that e∗ k(x)ek, where Am(x) = {π(n) : n ≤ m} is a greedy set of x: inf k∈Am(x) e∗ k(x). positive constant C such that π(j)(x) for i ≤ j. Alternatively we can write Gm(x) = Pk∈Am(x) e∗ Also, they defined in [16] the quasi-greedy bases as those bases such that there exists a k(x) ≥ supk /∈Am(x) e∗ kGm(x)k ≤ Ckxk, ∀x ∈ X,∀m ∈ N. (1) P. Wojtaszczyk proved in [18] that a basis is quasi-greedy if and only if the (TGA) converges -- that is, Of course, (1) is equivalent to the existence of a constant C ′ such that m→∞kx − Gm(x)k = 0, ∀x ∈ X. lim kx − Gm(x)k ≤ C ′kxk, ∀x ∈ X,∀m ∈ N. (2) We denoted by Cq the least constant that satisfies (2), it is called the quasi-greedy constant and we say that B is Cq-quasi-greedy. On the other hand, the (TGA) is a good candidate to obtain the best m-term approximation with regard to B. In this sense, S. V. Konyagin and V. N. Temlyakov defined in [16] the greedy bases as those bases such that there exists a constant C ≥ 1 such that anenk : A ⊂ N,A = m, an ∈ R}. (3) kx − Gm(x)k ≤ C inf{kx −Xn∈A Furthermore, they showed that B is greedy if and only if B is democratic (that is, k1Ak . k1Bk, for all A ≤ B) and unconditional. Some years later, G. Kerkyacharian, D. Picard and V. N. Temlyakov [15] introduced the following extension of the greedy bases: we consider a weight w = (wi)∞ i=1 ∈ (0,∞)N. If A ⊂ N, w(A) =Pi∈A wi denote the w-measure of A. We define the error σw δ (x) as anenk : A ∈ N<∞, w(A) ≤ δ, an ∈ R}. δ (x,B)X = σw σw δ (x) := inf{kx −Xn∈A kx − Gm(x)k ≤ Cσw Definition 1.1. We say that B is w-greedy if there exists a constant C ≥ 1 such that w(Am(x))(x), ∀x ∈ X,∀m ∈ N. (4) We denote by Cg the least constant that satisfies (4) and we say that B is Cg-w-greedy. Roughly, the greedy bases are those where the greedy approximation is "as effective as m- term approximation can possibly be". This generalization was motivated by the work of A. Cohen, R. A. DeVore and R. Hochmuth in [8]. In their recent paper [6], the first author and ´O. Blasco characterize w-greedy bases using the best m-term error in the approximation "with polynomials of constant coefficients". Moreover, [17] characterizes w-greedy bases in terms of their w-democracy and unconditionality. Definition 1.2. We say that B is w-democratic if there exists a constant C ≥ 1 such that (5) for any pair of sets A, B ∈ N<∞ with w(A) ≤ w(B). We denote by Cd the least constant that satisfies (5) and we say that B is Cd-w-democratic. k1Ak ≤ Ck1Bk, THE WEIGHTED PROPERTY (A) AND THE GREEDY ALGORITHM 3 Recall that a basis B in X is unconditional if any rearrangement ofPn e∗ n(x)en converges in norm to x for any x ∈ X. This is equivalent to the uniform boundedness of basis projections: (6) kx − PA(x)k ≤ Kkxk, ∀x ∈ X,∀A ⊂ N. We denote by Ku the least constant that satisfies (6), it is called the (suppression) unconditional constant, and we say that B is Ku-(suppression) unconditional. Other important w-type greedy basis in this context is the w-almost-greedy basis. Definition 1.3. We say that B is w-almost-greedy if there exists a constant C ≥ 1 such that (7) kx − Gm(x)k ≤ C σw w(Am(x)), ∀x ∈ X,∀m ∈ N, where δ (x,B)X = σw σw δ (x) := inf{kx − PA(x)k : A ∈ N<∞, w(A) ≤ δ}. We denote by Cal the least constant that satisfies (7) and we say that B is Cal-w-almost-greedy. Remark 1.4. If w ≡ 1, that is, w(A) = A, we recover the classical definition of almost- greediness (resp. greediness, democracy and Property (A)-see the definition below), and we will say that B is almost-greedy (resp. greedy, democratic, has the Property (A)). In the classical sense, that is, when w ≡ 1, S. J. Dilworth, N. J. Kalton, D. Kutzarova and V. N. Temlyakov gave in [12] a characterization of almost-greedy bases in terms of the quasi- greediness and democracy. Recently, S. J. Dilworth, D. Kutzarova, V. N. Temalykov and B. Wallis, in [13], gave a characterization of w-almost-greedy bases in terms of quasi-greedy and w-democratic bases. It is well known that, even for w ≡ 1, the w-democracy and unconditionality (resp. quasi- greediness), cannot be used to determine whether a given basis is w-greedy (resp. w-almost- greedy) with constant 1. For the weight w ≡ 1, F. Albiac and P. Wojtaszczyk introduced in [4] the so called Property (A) (defined below) in order to obtain finer estimate for the greedy constant Cg (and, in particular, to characterize bases with Cg = 1). The results of [4] were further generalized in [11]; in [2], the Property (A) was used to estimate the almost-greedy constant Cal. Throughout the paper, we will be using a weighted version of Property (A): Definition 1.5. We say that B satisfies the w-Property (A) if there exists a constant C ≥ 1 such that kx + t1εAk ≤ Ckx + t1ηBk, (8) for any x ∈ X, for any A, B ∈ N<∞such that w(A) ≤ w(B), A∩ B = ∅, supp (x)∩ (A∪ B) = ∅, for any ε, η ∈ {±1} and t ≥ supj e∗ j (x). We denote by Ca the least constant that satisfies (8) and we say that B has the Ca-w-Property (A). Remark 1.6. The definition of w-Property (A) was motivated by the "classical" Property (A) (introduced in [4]), which states that kx + t1εAk ≤ Ckx + t1ηBk, whenever A = B < ∞, A ∩ B = ∅, supp (x) ∩ (A ∪ B) = ∅, ε, η ∈ {±1} and t ≥ supj e∗ j (x). Proposition 3.2 shows that the classical Property (A) is equivalent to the w-Property (A) if 0 < inf n wn ≤ supn wn < ∞. 4 P. M. BERN ´A, S. J. DILWORTH, D. KUTZAROVA, T. OIKHBERG, AND B. WALLIS Another way of estimating the efficiency of greedy approximation is to compare the rate of convergence with straightforward Schauder approximation. To this end we consider w- partially-greedy bases. In [12], the authors defined the partially-greedy bases as those satisfying kx − Gm(x)k ≤ Ckx − Sm(x)k, ∀x ∈ X,∀m ∈ N, for some positive and absolute constant C. Moreover, they proved that B is partially-greedy if and only if B is quasi-greedy and conservative (that is, k1Ak . k1Bk for all pair of finite sets A, B such that A < B and A ≤ B). Here, we present the notion of w-partially-greedy bases and we characterize these bases using w-conservative bases. The paper is structured as follows. In Section 2 we describe the w-greedy and w-almost- greedy bases in terms of their other properties (such as w-Property (A), unconditionality, or being quasi-greedy). The main results are Theorems 2.1 and 2.2. In Section 3, we collect basic facts about the w-Property (A). In addition, we consider the w- semi-greedy bases -- that is, the bases where the Chebyshev greedy approximands are optimal. It turns out (Theorem 3.7) such bases necessarily possess the w-Property (A). Section 4 is devoted to properties (C) and (D), which arise naturally in the study of quasi- greedy bases. In particular, it is shown that w-superdemocracy and Property (C) imply w- Property (A) (Proposition 4.2). However, superdemocracy does not imply Property (C) (Ex- ample 4.8). Further, we show that any w-semi-greedy basis has Property (C) if the weight w is equivalent to a constant (Proposition 4.10). In Section 5, we compare the efficiency of greedy approximation with that of the canonical basis projections. This gives rise to the notion of an w-partially-greedy basis; such bases are characterized in Theorem 5.7. Finally, in Section 6 and Section 7 we state some open questions related to our results, and prove some basic lemmas used throughout the paper. We freely use the standard "greedy" terminology. The reader can consult e.g. [17] for more information. 2. Characterization of w-greedy and w-almost-greedy bases In this section we describe the w-(almost)-greediness of a basis in terms of its w-Property (A) and unconditionality (resp. quasi-greediness). The corresponding results for the constant weight w ≡ 1 can be found, for instance, in [17]. Theorem 2.1. Let B be a basis of a Banach space X. a) If B is Cg-w-greedy, then the basis is Ku-unconditional and has the Ca-w-Property (A) b) If B is Ku-unconditional and has the Ca-w-Property (A), then the basis is Cg-w-greedy with constants Ku ≤ Cg and Ca ≤ Cg. with Cg ≤ KuCa. Theorem 2.2. Let B be a basis of a Banach space X. a) If B is Cal-w-almost-greedy, then the basis is Cq-quasi-greedy and has the Ca-w-Property b) If B is Cq-quasi-greedy and has the Ca-w-Property (A), then the basis is Cal-w-almost- (A) with constants Cq ≤ Cal and Ca ≤ Cal. greedy with Cal ≤ CqCa. Later (Proposition 5.11) we will see examples of bases with w-Property (A) for a certain weight w, but failing the "classical" Property (A). For further use, we need the following reformulation of the w-Property (A) (inspired by [2]). THE WEIGHTED PROPERTY (A) AND THE GREEDY ALGORITHM 5 Proposition 2.3. A basis B has the Ca-w-Property (A) if and only if kxk ≤ Cakx − PA(x) + 1ηBk, (9) j (x) ≤ 1, A, B ∈ N<∞, w(A) ≤ w(B), B ∩ supp (x) = ∅ and for any x ∈ X with supj e∗ η ∈ {±1}. The proof requires a technical result. j (y). j (x) = maxj e∗ Lemma 2.4. Suppose D is a finite subset of N, and x ∈ X\{0} satisfies supp (x) ∩ D = ∅. Then for any ε > 0 there exists a finitely supported y ∈ X, so that kx−yk < ε, supp (y)∩D = ∅, and maxj e∗ Proof. It suffices to consider ε < 1/(2c2). By scaling, we can assume that maxj e∗ (then kxk ≥ 1/c2). Clearly PD(x) = 0, and P c D(x) = x. Now set δ = ε/(3c2 span [ej : j ∈ N] is dense in X, there exists a finitely supported z ∈ X so that kx−zk < δ/kP c Let u = P c C = maxj e∗ j (x) = 1 2kxk). As Dk. j (x − u) < c2δ, hence j (y) = 1, and D(z), then kx − uk = kP c D(x − z)k < δ. For every j, e∗ j (x) ∈ (1 − c2δ, 1 + c2δ). Now let y = u/C. Then maxj e∗ kx − yk ≤ kx − uk + 1 − C −1kuk < δ + (kxk + δ) < ε. (cid:3) c2δ 1 − c2δ Proof of Proposition 2.3. By Lemma 2.4, it suffices to restrict our attention to finitely supported vectors x ∈ X only. So, throughout this proof, we assume supp (x) < ∞. proposition with supj e∗ B, we obtain Suppose that B has the Ca-w-Property (A), and x, A, B, ε, η are as in the statement of the j (x) ≤ 1. Applying the definition of w-Property (A) to PAcx, A, and kPAc(x) + 1εAk ≤ CakPAc(x) + 1ηBk = Cakx − PA(x) + 1ηBk. To finish the proof, observe that x belongs to the convex hull of the set(cid:8)PAc(x) + 1εA}ε∈{±1}. Now, suppose (9), and prove that the basis B has the w-Property (A) with the same constant. j (x) ≤ 1, A, B ∈ N<∞ such that w(A) ≤ w(B), A ∩ B = ∅, supp (x) ∩ Take x ∈ X and supj e∗ (A ∪ B) = ∅ and ε, η ∈ {±1}. Define x′ = x + 1εA. Using (9), kx + 1εAk = kx′k ≤ Cakx′ − PA(x′) + 1ηBk = Cakx + 1ηBk. (cid:3) Proof of Theorem 2.1: Assume that B is Cg-w-greedy. where Unconditionality: Let x ∈ X and A ⊂ supp (x). Define y := PAc(x) +Pn∈A(α + e∗ n(x))en, α > sup j∈A e∗ j (x) + sup j∈Ac e∗ j (x). As A is a greedy set of y, kx − PA(x)k = ky − PA(y)k ≤ Cgσw w(A)(y) ≤ Cgky − α1Ak = Cgkxk. w-Property (A): Fix x ∈ X, take t ≥ supn e∗ Thus, the basis is unconditional with constant Ku ≤ Cg. n(x). Consider ε, η ∈ {±1} and finite sets A, B such that A∩ B = ∅, w(A) ≤ w(B), and (A∪ B)∩ supp (x) = ∅. Set y := x + t1εA + (t + δ)1ηB with δ > 0. Hence, kx + t1εAk = ky − GB(y)k ≤ Cgσw w(B)(y) ≤ Cgky − t1εAk = Cgkx + (t + δ)1ηBk. Taking δ → 0, we obtain that the basis satisfies the w-Property (A) with constant Ca ≤ Cg. 6 P. M. BERN ´A, S. J. DILWORTH, D. KUTZAROVA, T. OIKHBERG, AND B. WALLIS w-greedy. Next we prove that if B is Ku-unconditional and has the Ca-w-Property (A), then it is Take x ∈ X and suppose that A is a greedy set of cardinality m for x ∈ X -- that is, w(A)(x) + ε, with supp (y) = B j (x)}, using the the PA(x) = Gm(x). For ε > 0 find y ∈ X such that kx − yk < σw and w(B) ≤ w(A). Then, taking t := min{e∗ reformulation of the w-Property (A) and Lemma 7.1, we obtain that j (x) : j ∈ A} and η ≡ sgn{e∗ kx − Gm(x)k ≤ Cakx − PA(x) − PB\A(x) + t1η(A\B)k = CakP(A∪B)c(x − y) + t1η(A\B)k = CakTt(I − PB)(x)k = CakTt(I − PB)(x − y)k ≤ KuCakx − yk. Consequently, for any greedy set A we have kx − PAxk ≤ KuCaσw Proof of Theorem 2.2. Assume that B is Cal-w-almost-greedy. Quasi-greedy: Since w(A)(x). (cid:3) kx − Gm(x)k ≤ Cal inf{kx −Xn∈B n(x)enk : w(B) ≤ w(Am(x)), B ∈ N<∞}, e∗ we can select B = ∅. Then, we obtain that kx − Gm(x)k ≤ Calkxk, hence the basis is quasi- greedy with constant Cq ≤ Cal. w-Property (A): We can use the same argument as in Theorem 2.1. Now, we will prove that if B is Cq-quasi-greedy and has the Ca-w-Property (A), then it is For x ∈ X, let A be a greedy set of cardinality m. For ε > 0, find B such that kx− PB(x)k < σw w(A)(x) + ε, with w(B) ≤ w(A). Then, taking t := min{e∗ j (x)}, using the reformulation of the w-Property (A) and Lemma 7.1, j (x) : j ∈ A} and η ≡ sgn{e∗ w-almost-greedy. kx − Gm(x)k ≤ CakP(A∪B)c(x − y) + t1η(A\B)k = CakTt(I − PB)(x)k ≤ CqCakx − PB(x)k. This gives that, for any greedy set A, kx − PA(x)k ≤ CqCaσw(A)(x) as desired. (cid:3) Remark 2.5. In this paper, we focus on the situation when B is a Schauder basis. However, the w-Property (A) can be defined for any complete biorthogonal system satisfying the conditions (A1)-(A3); the proof of Proposition 2.3 goes through as well. Moreover, in the definition of the w-Property (A), it suffices to show that (8) holds for with maxj e∗ j (x) = t. More specifically, the following four statements are equivalent: B ∩ supp (x) = ∅, w(A) ≤ w(B) and t ≥ supj e∗ w(B), A ∩ B = ∅, supp (x) ∩ (A ∪ B) = ∅, ε, η ∈ {±1} and s = supj e∗ B ∩ supp (x) = ∅, w(A) ≤ w(B) and s = supj e∗ (a) B satisfies the w-Property (A) (see Definition 1.5). (b) There exists a constant C so that kxk ≤ Ckx− PA(x) + t1ηBk for any η ∈ {±1}, x ∈ X, j (x). (c) There exists a constant C ′ so that kx + s1εAk ≤ Ckx + s1ηBk for any x ∈ X, w(A) ≤ (d) There exists a constant C ′ so that kxk ≤ C ′kx−PA(x)+s1ηBk for any x ∈ X, η ∈ {±1}, Indeed, the implications (a) ⇒ (c) and (b) ⇒ (d) (with C ′ = C) are immediate. The equiv- alence (a) ⇔ (b) (with the same constant C) has been established in Proposition 2.3. Minor adjustments to that argument give us (c) ⇔ (d). j (x). As before, we can assume that x is finitely supported. Find k so that e∗ j (x). By replacing x by To establish (d) ⇒ (b), take x, A, B, η as in (b) and t ≥ supj e∗ k(x) = supj e∗ j (x). j (x). THE WEIGHTED PROPERTY (A) AND THE GREEDY ALGORITHM 7 −x if necessary, we can assume s = e∗ k(x) ≥ 0. Let c = t − s, and consider x′ = x + cek = Xj∈supp (x)\{k} e∗ j (x)ej + tek. Note that kx − x′k ≤ cc2 ≤ tc2. Furthermore, x′ − PA(x′) equals either x − PA(x) (if k ∈ A), or x − PA(x) + cek (if k /∈ A). In either case, kx − PA(x) + t1ηBk ≥ kx′ − PA(x′) + t1ηBk − tc2. By (d), we have kx′k ≤ C ′kx′ − PA(x′) + t1ηBk. By the above, As kx − PA(x) + t1ηBk ≥ tc−1 kxk − tc2 ≤ C ′(kx − PA(x) + t1ηBk + tc2) 2 , we conclude that kxk ≤ (C ′ + 2c2 2)kx − PA(x) + t1ηBk. 3. Some remarks on the w-Property (A) Definition 3.1. Let v = (vn)∞ to w, written v ≈ w, whenever there exist positive real constants 0 < a ≤ b < ∞ satisfying n=1 be weights. We say that v is equivalent n=1 and w = (wn)∞ avn ≤ wn ≤ bvn for all n ∈ N. Proposition 3.2. Let v, w weights and suppose that v ≈ w. Then every basis with the w- Property (A) also has the v-Property (A). Proof. Let x ∈ X with supp (x) < ∞ and supj e∗ v(B), A ∩ B = ∅, supp (x) ∩ (A ∪ B) = ∅ and ε, η ∈ {±1}. We set j (x) ≤ 1, A and B finite satisfying v(A) ≤ Observe that which gives us Γ = {n ∈ A : wn ≥ w(B)} . w(A) ≤ b · v(A) ≤ b · v(B) ≤ b a · w(B), w(A) ≥ w(Γ) ≥ Γ · w(B) ≥ Γ · a b · w(A), and hence Γ ≤ b/a. Next, we give the following partition of A \ Γ: A1 < . . . < Am, so that for each i = 1, . . . , m, the set Ai is a maximal such that w(Ai) ≤ w(B). Due to maximality, w(B) < w(Ai) + w(Ai+1) for all i = 1, . . . , m − 1. Thus, (m − 1) · w(B) < m−1Xi=1 This gives us [w(Ai) + w(Ai+1)] < 2 · w(A \ Γ) ≤ 2 · w(A) ≤ 2b a · w(B). m ≤ 2b a + 1. 8 P. M. BERN ´A, S. J. DILWORTH, D. KUTZAROVA, T. OIKHBERG, AND B. WALLIS Hence, using the bounds of Γ, m and the condition of the w-Property (A), kx + 1εAk ≤ k1Γk + kx + mXi=1 1εAik ≤Xn∈Γ kenk + x m k + 1εAik mXi=1 c2 2b a kx + 1ηBk + Cakx + m1ηBk x m + 1ηBk ≤ ≤ c2 2Γkx + 1ηBk + Camk c2 2b ≤ a kx + 1ηBk + Camkx + 1ηBk + Ca(m − 1)kxk c2 2b a kx + 1ηBk + Camkx + 1ηBk + C 2 ≤ 2b + 2bC 2 a ≤ (cid:18)c2 a (cid:19)kx + 1ηBk. a(m − 1)kx + 1ηBk (cid:3) Remark 3.3. In a similar fashion, one can show that, if the weights w and v are equivalent, then any w-democratic (w-superdemocratic, w-conservative -- for the definitions, see below) basis is also v-democratic (resp. v-superdemocratic or v-conservative). Remark 3.4. The converse to Proposition 3.2 does not hold in general. For example, suppose the weights w, v belong to ℓ1. By [13], the family of w-democratic (or v-democratic) bases consists precisely of those bases which are equivalent to the canonical basis of c0. However, w and v need not be equivalent. The rest of this section is motivated by the recent definition of w-semi-greedy bases introduced in [13]. To give this notion, we need the Chebyshev Greedy Algorithm: for x ∈ X and m ∈ N, if Am(x) is the greedy set of x with cardinality m, we define the Chebyshev Greedy Approximand of order m as any Gm(x) ∈ span{ei : i ∈ Am(x)} such that kx − Gm(x)k = min{kx − Xn∈Am(x) kx − Gm(x)k ≤ Cσw bnenk : bn ∈ R}. Definition 3.5. We say that B is w-semi-greedy if there exists a constant C ≥ 1 such that (10) We denote by Csg the least constant that satisfies (10) and we say that B is Csg-w-semi-greedy. w(Am(x))(x), ∀x ∈ X,∀m ∈ N. By [13], any w-semi-greedy basis is w-superdemocratic. Definition 3.6. We say that B is w-superdemocratic if there exists a constant C ≥ 1 such that k1εAk ≤ Ck1ηBk, (11) for any A, B ∈ N<∞ with w(A) ≤ w(B) and ε, η ∈ {±1}. We denote by Cs the least constant that satisfies (11) and we say that B is Cs-superdemocratic. Here, we show that w − semi-greedy ⇒ w − Property (A) ⇒ w − superdemocracy. Theorem 3.7. If a basis B is w-semi-greedy, then B has the w-Property (A). (12) THE WEIGHTED PROPERTY (A) AND THE GREEDY ALGORITHM 9 We take ε, η, A, B and x in the conditions of the definition of the w-Property (A). In all of Proof. Assume that kx − Gm(x)k ≤ Csgσw w(Am(x)) for any x ∈ N and m ∈ N. the following cases we consider x ∈ X such that supp (x) < ∞ and supn e∗ n(x) ≤ 1. Case 1:P∞ Case 1.1: w(B) > lim supn→∞ wn. Since Pn wn = ∞, we can choose E and n0 ∈ N with min E > max(A ∪ B ∪ supp (x)) and n0 > max E such that n=1 wn = ∞ and supn wn < ∞. Set F := E ∪ {n0}. Then, w(E) ≤ w(B) < w(F ) < 2w(B). w(E) ≤ w(B) < w(E) + wn0 < 2w(B). We define the element z := x + 1εA + (1 + δ)1F . For any scalar sequence (fn)n∈F , we have kx + 1εAk ≤ Kbkx + 1εA +Pn∈F fnenk. As the basis B is w-semi-greedy with constant Csg, and w(A) ≤ w(B) < w(F ), we conclude that inf fn kx + 1εA +Xn∈F fnenk ≤ Csgσw w(F )(z) ≤ Csgkx + (1 + δ)1Fk. Consequently, kx + 1εAk ≤ KbCsgkx + (1 + δ)1Fk. Taking δ → 0, kx + 1εAk ≤ KbCsgkx + 1Fk ≤ KbCsgkx + 1Ek + KbCsgken0k (13) ≤ KbCsgkx + 1Ek + KbCsgc2 ≤ KbCsg(kx + 1ηBk + k1ηBk + k1Ek) + KbCsgc2. Now, we set y := 1ηB + (1 + δ)1F . Reasoning as before, we obtain k1ηBk ≤ Kb inf Sending δ → 0, we obtain cn k1ηB +Xn∈F cnenk ≤ KbCsgσw w(F )(y) ≤ KbCsgk(1 + δ)1Fk. k1ηBk ≤ KbCsgk1Fk ≤ KbCsgk1Ek + KbCsgc2. On the other hand, taking s := x + (1 + δ)1ηB + 1E, bnen + 1Ek ≤ Csg(Kb + 1)σw k1Ek ≤ (Kb + 1)kx +Xn∈B Then, taking δ → 0, k1Ek ≤ Csg(Kb + 1)kx + 1ηBk. w(B)(s) ≤ Csg(Kb + 1)kx + (1 + δ)1ηBk. Finally, using (13), (14) and (15), the basis satisfies the w-Property (A) with constant K = O(C 3 sgK 3 b c2). Case 1.2: w(A) ≤ w(B) ≤ lim supn→∞ wn. Using Proposition 3.5 of [13], max{k1εAk,k1ηBk} ≤ 2KbCsgc2. Since 1 = e∗ j (x + 1ηB) ≤ ke∗ jkkx + 1ηBk ≤ c2kx + 1ηBk for j ∈ B, then kx + 1εAk ≤ kx + 1ηBk + k1ηBk + k1εAk ≤ kx + 1ηBk + 4KbCsgc2 ≤ (4KbCsgc2 2 + 1)kx + 1ηBk. Case 2: IfPn wn < ∞ or supn wn = ∞, using the Proposition 3.5 of [13], B is equivalent to the canonical basis of c0 and the result is trivial. (14) (15) (16) (cid:3) 10 P. M. BERN ´A, S. J. DILWORTH, D. KUTZAROVA, T. OIKHBERG, AND B. WALLIS Proposition 3.8. If B has the Ca-w-Property (A), then B is 2Ca-w-superdemocratic. Proof. Take A, B ∈ N<∞ with w(A) ≤ w(B), and show that, for any choice of signs, k1ηAk ≤ 2Cak1εBk. As in [7, Subsection 4.4], it is enough to prove our inequality for ε ≡ 1 (otherwise, replace B = {en : n ∈ N} by {εnen : n ∈ N}). Since 1ηA ∈ 2S, where S = {PA′⊂A θA′1A′ : PA′⊂A θA′ ≤ 1} (see [10, Lemma 6.4]), it suffices to show that k1A′k ≤ Cak1Bk, ∀A′ ⊂ A. Take A′ ⊂ A. Obviously, 1A′ = 1A′\B + 1B∩A′. Then, using the w-Property (A), k1A′k = k1A′\B + 1B∩A′k ≤ Cak1B\A′ + 1B∩A′k = Cak1Bk. We can apply the w-Property (A) because w(A′) = w(A′\B)+w(A′∩B) ≤ w(A) ≤ w(B) = w(B\A′)+w(B∩A′) ⇒ w(A′\B) ≤ w(B\A′). This completes the proof. (cid:3) With these results, we have proved the implications (12). Remark 3.9. If w ≡ 1, we recover the classical definition of semi-greediness (resp. superdemoc- racy), and we will say that B is semi-greedy (resp. superdemocratic). Improving [13, Proposition 4.5], we prove that, in certain cases, any w-superdemocratic basis has to contain a subsequence equivalent to the canonical basis of c0 (c0-basis henceforth), or even to be equivalent to such basis. Proposition 3.10. Suppose a basis B = (en)∞ n=1 is Cs-w-superdemocratic. i) If A ∈ N<∞ and w(A) ≤ lim supn→∞ wn, then maxε∈{±1} k1εAk ≤ c2Cs. ii) If supn wn = ∞, then B is equivalent to the c0-basis. iii) If inf n wn = 0, then there exist i1 < i2 < . . . so that the sequence (eik )k∈N is equivalent to the c0-basis. Moreover, if limn wn = 0, then for any infinite set A ⊂ N we can select i1, i2, . . . ∈ A with the properties described above. Proof. i) Find n ∈ N\A so that wn > w(A), then k1εAk ≤ Csk1{n}k ≤ c2Cs. ii) By (i), k1εAk ≤ c2Cs for all choices of signs, which yields the desired equivalence. iv) IfPn wn < ∞, then B is equivalent to the c0-basis. iii) Suppose inf n wn = 0, and find i1 < i2 < . . . so that Pk wik < ∞. By convexity, it for any finite set E ⊂ {i1, i2, . . .}. To this end, find N ∈ N so that PN j=1 wj ≥Pk wik < ∞. suffices the existence of a constant K with the property that the inequality k1εEk ≤ K holds Let B = {1, . . . , N}\{i1, i2, . . .}, D = {1, . . . , N} ∩ {i1, i2, . . .}, and A = E\D. Note that B,D ≤ N, hence, for every ε ∈ {−1, 1}N, k1εBk,k1εDk ≤ c2N. Then w(A) ≤ w(B) and hence k1εAk ≤ Csk1εBk ≤ c2NCs. By the triangle inequality, k1εEk ≤ k1εAk + k1εDk ≤ c2N(Cs + 1). If limn wn = 0, then every infinite A ⊂ N contains i1 < i2 < . . . withPk wik < ∞. It remains to invoke the preceding result. iv) The proof proceeds as in (iii). (cid:3) From this we immediately obtain: Corollary 3.11. If the weight w is unbounded, then a basis has the w-Property (A) if and only if it is equivalent to the canonical basis of c0. THE WEIGHTED PROPERTY (A) AND THE GREEDY ALGORITHM 11 4. Properties (C) and (D) Properties (C) and (D) (discussed below) naturally arise in the study of quasi-greedy bases. Definition 4.1. We say that B satisfies the Property (C) if for any x ∈ X, there exists a positive constant C such that j∈Λ e∗ min j (x)k1εΛk ≤ Ckxk, (17) for any greedy set Λ of x and ε ∈ {±1}. We denote by Cu the least constant that satisfies (17) and we say that B has the Property (C) with constant Cu. It is well known any quasy-greedy basis has Property (C) (see [7, Lemma 2.3]). Generalizing [7, Lemma 2.2], we prove that any w-superdemocratic basis with the Property (C) has the w-Property (A). Proposition 4.2. If B is Cs-w-superdemocratic and satisfies the Property (C) with constant Cu, then B has the Ca-w-Property (A) with Ca ≤ 3CuCs. Proof. Take x, A, B, ε, η as in the definition of the w-Property (A) and assume that supj e∗ 1. Then, j (x) ≤ kx + 1εAk ≤ kx + 1ηBk + k1ηBk + k1εAk. (18) Using the w-superdemocracy and w(A) ≤ w(B), we obtain that k1εAk ≤ Csk1ηBk. Now, we only have to estimate k1ηBk. For that, we consider the element y := x + 1ηB. It's clear that 1ηB is a greedy sum for y, so j∈B e∗ min j (y)k1ηBk = k1ηBk ≤ Cukyk = Cukx + 1ηBk. (19) Then, using (18) and (19), kx + 1εAk ≤ kx + 1ηBk + 2CsCukx + 1ηBk ≤ 3CsCukx + 1ηBk. Hence, the basis has the w-Property (A) with constant Ca ≤ 3CuCs. Example 4.3. We next revisit a "pathological" basis constructed in Section 5.5 of [7] (using some ideas from [9, Example 4.8]): a basis which has the Property (A), but fails to be quasi- greedy. The initial proof of the Property (A) was unwieldy. Here we present a streamlined proof that the basis has the Property (C), and then invoke Proposition 4.2. (cid:3) First recall the construction: Dk denote the set of all dyadic intervals I ⊂ [0, 1] with length 1 of all real sequences I = 2−k, and consider D = ∪k≥0Dk. Now, we consider the space fq a = (aI )I∈D such that aIχ(1) I < ∞, q!1/q(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)L1 kakf q 1 XI =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) where χ(1) I = I−1χI. By [14], the canonical basis {eI}I∈D is unconditional and democratic. For every N ≥ 1, we shall pick a subset {k1, ..., kN} ⊂ N0 and look at the finite dimensional space FN consisting of sequences supported in ∪N j=1Dkj . We order the canonical basis by ∪N j=1{eI}I∈Dkj . We also consider in , so we may as well write their elements as a = (aj)dim FN j=1 12 P. M. BERN ´A, S. J. DILWORTH, D. KUTZAROVA, T. OIKHBERG, AND B. WALLIS FN the James norm k(aj)kJq = Now, set in FN a new norm m0=0<m1<...Xk≥0 sup kak = max{kakf q 1 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Xmk<j≤mk+1 ,kakJq}. 1/q q . aj(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Finally, we consider the Banach space X = ⊕ℓ1FN with B the consecutive union of the natural bases in FN . . To show that B 1 is unconditional: take a ∈ X It's possible to show that B is superdemocratic and k1εAk ≈ A ≈ k1εAkfq satisfies the Property (C), we use that the canonical basis in fq and Λ a greedy set of a, then 1 n∈Λ ank1εΛk . min min n∈Λ ank1εΛkfq 1 . kakfq 1 ≤ kak. Hence, the basis satisfies the Property (C). Also, since the basis is superdemocratic, using the Proposition 4.2, the basis satisfies the Property (A). Definition 4.4. We say that B is bidemocratic if there exists a constant C ≥ 1 such that k1εAkk1∗ ηAk∗ ≤ CA, ∀ finite A,∀ε, η ∈ {±1}. Here, k · k∗ is the norm of X∗, and 1∗ Lemma 4.5. If B is bidemocratic, then B satisfies the Property (C). Proof. Here, we prove a stronger condition than Property (C). Take x ∈ X and A ⊂ supp (x). Then, taking η = 1/sgn{e∗ i . ηA =Pi∈A ηie∗ i (x)}, j (x)k1εAk . min j∈A e∗ j (x) j∈A e∗ min ηAk∗ ≤ Pj∈A e∗ k1∗ A k1∗ ηAk∗ j (x) ≤ 1∗ ηA(x) ηAk∗ ≤ kxk. k1∗ (cid:3) Corollary 4.6. All bidemocratic bases satisfy the "classical" Property (A). Proof. If a basis B is bidemocratic, then it is superdemocratic. Now combine the preceding lemma with Proposition 4.2. (cid:3) Relaxing the assumptions of Definition 4.1, we consider: Definition 4.7. We say that B satisfies the Property (D) if there exists a positive constant C such that min n∈A ank1Ak ≤ CkXn∈A anenk, for any finite set A and scalars (an)n∈A. It's clear that if B satisfies the Property (C), then B satisfies the Property (D) as well. Example 4.8. [Example of superdemocratic basis in a Banach space without the Property (D)] Let X = ℓ1 ⊕ c0 and k(x, y)k = kxkℓ1 + kyk∞. Let (en)n be the canonical basis in ℓ1 and (fm)m the canonical basis in c0. We define E2n−1 =(cid:18)1 2 en, −1 2 fn(cid:19) , E2n =(cid:18) 1 4 en, 3 4 fn(cid:19) , n = 1, 2, ..., THE WEIGHTED PROPERTY (A) AND THE GREEDY ALGORITHM 13 and consider B = {En}n = {E2n−1, E2n}n. This basis is normalized. To prove that this basis is superdemocratic, we show the following proposition: Proposition 4.9. D(m) ≈ d(m) ≈ m, where D(m) := sup{k1εAk : A ≤ m, ε ∈ {±1}} and d(m) := inf{k1εAk : A ≥ m, ε ∈ {±1}}. Proof. Of course, d(m) ≤ D(m) ≤ m. We prove that d(m) ≥ 1 finite, we write 8m. To this end, given A ⊂ N A1 = {k ∈ N : 2k ∈ A and 2k − 1 ∈ A}, A2 = {k ∈ N : 2k ∈ A and 2k − 1 6∈ A}, A3 = {k ∈ N : 2k 6∈ A and 2k − 1 ∈ A}. Observe that the sets A1, A2, A3 are mutually disjoint, and 2A1 + A2 + A3 = A. For any choice of signs, k1εAk = kXk∈A1 = kXk∈A1(cid:18)[ + Xk∈A2 ≥ Xk∈A1 ε2k( 1 4 ε2kE2k + ε2k−1E2k−1 +Xk∈A2 1 4 1 2 ε2k + ε2k−1]ek, [ ε2kE2k +Xk∈A3 ε2k−1]fk(cid:19) 1 2 ε2k−1E2k−1k ε2k−1( 3 4 1 4 ε2k − 1 2 ek,− 1 2 + Xk∈A3 fk)k 1 2 . 1 4 ek, 3 4 fk) +Xk∈A3 ε2k−1 +Xk∈A2 1 2 ε2k + Therefore, This finishes the proof. k1εAk ≥ 1 4A1 + 1 4A2 + 1 2A3 ≥ 1 8A. (cid:3) Back to Example 4.8: to see that the basis does not have the Property (D), take z = n=1 E2n−1. Then, PN n=1 2E2n −PN NXn=1 Write z =Pi∈A aiEi. Then, mini∈A ai = 1 and kzk = k (0, 2fn)k = 2. kXi∈A Eik = k E2n + E2n−1k = k NXn=1 NXn=1 ( 3 4 NXn=1 en, 1 4 fn)k = 3 4 N + 1 4 . This shows that the Property (D) fails. Lemma 4.13 of [13] establishes that, if w is equivalent to the constant, and B is w-semi-greedy, then B satisfies the Property (D). Here, we improve this result showing that the condition of being w-semi-greedy implies the Property (C). Proposition 4.10. Assume that w is equivalent to the constant, and the basis B is w-semi- greedy. Then B satisfies the Property (C). Proof. By Theorem 3.7, B has the w-Property (A). By Proposition 3.2, B also has the "classical" Property (A). This, in turn, implies the Property (C). (cid:3) 14 P. M. BERN ´A, S. J. DILWORTH, D. KUTZAROVA, T. OIKHBERG, AND B. WALLIS 5. w-Partially-greedy bases Partially-greedy and conservative bases were introduced in [12], in order to compare the errors of greedy approximation with those of the canonical approximation relative to Schauder basis (the "tails" of the basis expansion). In this section we define w-partially-greedy and w- conservative bases and extend the characterization of partially-greedy bases proved in [12] to this more general setting. Definition 5.1. We say that B is w-partially-greedy if for all m and r such that w({1, ..., m}) ≤ w(Ar(x)), there exists a positive constant such that kx − Gr(x)k ≤ Ck e∗ i (x)eik. (20) ∞Xn=m+1 We denote by Cp the least constant that satisfies (20) and we say that B is Cp-w-partially-greedy. Definition 5.2. We say that B is w-conservative if there exists a positive constant C such that k1Ak ≤ Ck1Bk, (21) for all pair of A, B ∈ N<∞ such that A < B and w(A) ≤ w(B). We denote by Cc the least constant that satisfies (21) and we say that B is Cc-w-conservative. Remark 5.3. If w ≡ 1, we recover the classical definition of partially-greediness (resp. conser- vativeness), and we will say that B is partially-greedy (resp. conservative). Remark 5.4. Note that for some choices of weight w, the property of w-conservativeness can be in some sense trivial. For instance, if w = (2−n)∞ n=1 then every seminormalized basis is w-conservative. This is because there are no nonempty A, B ∈ N<∞ satisfying both A < B and w(A) ≤ w(B). Let us give a simple characterization of weights for which this occurs. Proposition 5.5. Let w be a weight and set sw := sup(cid:8)n ∈ N0 : there exist A ∈ Nn and B ∈ N<∞ such that A < B and w(A) ≤ w(B)(cid:9). Then sw < ∞ if and only if every seminormalized basis is w-conservative. Proof. (=⇒): Suppose sw < ∞. Let (en)∞ n=1 be a seminormalized basis for a Banach space X, and select A, B ∈ N<∞ such that A < B and w(A) ≤ w(B). Observe that k1Ak ≤ c2·A ≤ c2sw. It follows immediately that k1Ak ≤ c2sw ≤ c2 2sw)-w-conservative. n=1 ⊂ N<∞ and (Bn)∞ (⇐=): Suppose sw = ∞. Let's inductively construct sequences (An)∞ 2swk1Bk. Hence, (en)∞ n=1 is (c2 n=1 ⊂ N<∞ satisfying A1 < B1 < A2 < B2 < A3 < B3 < . . . , and also satisfying An ≥ n and w(An) ≤ w(Bn) for all n ∈ N. Let us begin by selecting A1 ∈ N<∞ and B1 ∈ N<∞ with A1 = 1, A1 < B1, and w(A1) ≤ w(B1), which is possible as sw ≥ 1. This is the base case; from now on, we proceed inductively. Since sw = ∞, we may select bAn+1 ∈ N<∞ and Bn+1 ∈ N<∞ with bAn+1 > n + max Bn, bAn+1 < Bn+1, and w(bAn+1) < w(Bn+1). Now set An+1 = bAn+1 \ {1, . . . , max Bn} so that we have An+1 > n, An+1 < Bn+1, and w(An+1) < w(Bn+1). This completes the inductive step, and gives us our THE WEIGHTED PROPERTY (A) AND THE GREEDY ALGORITHM 15 intertwining sequences with the desired properties. We may now define a norm on c00 via the rule k(an)∞ n=1kX = k(an)∞ n=1k∞ ∨ sup an ∀ (an)∞ n=1 ∈ c00, k∈N Xn∈Ak and denote by X the completion of c00 under this norm. It is clear that the standard canon- ical basis for this space form a normalized 1-unconditional basis. However, it fails to be w- conservative as k1AkkX = Ak ≥ k whereas k1BkkX = 1 for all k ∈ N. (cid:3) Proposition 5.6. Let w be a nonincreasing weight, i.e., wn+1 ≤ wn for all n ∈ N. Then every conservative basis in a Banach space is w-conservative with the same constant. Proof. Let (en)∞ satisfying both A < B and w(A) ≤ w(B). Now, n=1 be a conservative basis in a Banach space X, and select any A, B ∈ N<∞ A · wmax A ≤ w(A) ≤ w(B) ≤ B · wmin B ≤ B · wmax A, so that A ≤ B. Theorem 5.7. A basis B is w-partially-greedy if and only if B is quasi-greedy and w-conservative. Proof. Assume that B is Cp-w-partially-greedy. (cid:3) (1) w-conservative: take A and B such that A < B and w(A) ≤ w(B). Let m = max A and define the set D = [1, .., m] \ A. Of course, w({1, ..., m}) = w(A ∪ D) ≤ w(B ∪ D). Define now x := 1A + (1 + δ)1B∪D. Then, k1Ak = kx − GB∪D(x)k ≤ Cpk(1 + δ)1Bk. Taking δ → 0, the basis is w-conservative. (2) Quasi-greedy: here, we consider two cases. 1(x))e1, with t = maxe∗ e∗ a) Assume that the index 1 6∈ Ar(x). Define then x = te1 +P∞ hence x − Gr(x) =P∞ i (x) + δ with δ > 0. Then Gr(x) = te1 + Gr−1(x), i (x)ei − Gr−1(x). Thus, using the triangle inequality and the fact that w({1}) ≤ w(Ar(x)), i (x)ei = x + (t − i=2 e∗ i=2 e∗ kGr−1(x)k ≤ kx − Gr(x)k + k e∗ i (x)eik ≤ Cpk ∞Xi=2 ∞Xi=2 e∗ i (x)eik + k ∞Xi=2 e∗ i (x)eik ≤ (Cp + 1)(1 + Kb)kxk. That's implies that kGr(x)k ≤ ((Cp + 1)(Kb + 1) + c2 b) Assume now that 1 ∈ Ar(x). Taking the same x that in the above case, 2)kxk. Gr(x) = Gr−1(x − e∗ 1(x)e1) + te1, 16 P. M. BERN ´A, S. J. DILWORTH, D. KUTZAROVA, T. OIKHBERG, AND B. WALLIS i=2 e∗ i (x)ei − Gr−1(x − e∗ 1(x)e1). Hence, using the same argument than before, so x − Gr(x) =P∞ kGr−1(x − e∗ 1(x)e1)k ≤ kx − Gr(x)k + k e∗ i (x)eik ∞Xi=2 ∞Xi=2 e∗ ≤ Cpk i (x)eik + k ≤ (Cp + 1)(1 + Kb)kxk. 1(x)e1) + e∗ ∞Xi=2 e∗ i (x)eik Now, kGr−1(x − e∗ 1(x)e1k = kGr(x) − e∗ 1(x)e1k. So, using the triangle inequality, we obtain that kGr(x)k ≤ ((Cp +1)(Kb + 1) + c2 1(x)e1)k = kGr−1(x − e∗ 1(x)e1 − e∗ 2)kxk. Now, assume that B is Cc-w-conservative and Cq-quasi-greedy, and show that B is w-partially- greedy. Take x ∈ X, m, and r as in the definition of w-partially-greedy, and consider the sets D := {ρ(j) : j ≤ r, ρ(j) ≤ m}, B := {ρ(j) : j ≤ r, ρ(j) > m}, A := [1, ..., m] \ D, where ρ is the greedy ordering. Then Ar(x) = B ∪ D, and w(A) = w({1, ..., m}) − w(D) ≤ w(Ar(x)) − w(D) = w(B). x − Gr(x) = e∗ i (x)ei − PB(x) + PA(x). ∞Xi=m+1 i (x)eik. On the other hand, using Lemmas 7.1 and 7.2 with η ≡ sgn (e∗ On the one hand, kPB(x)k ≤ 2CqkP∞ A e∗ kPA(x)k ≤ 4CqCc max j (x)), i=m+1 e∗ i (x)k1ηBk ≤ 4CqCc min B e∗ i (x)k1ηBk q CckPB(x)k ≤ 8C 3 q Cck ∞Xi=m+1 e∗ i (x)eik. ≤ 8C 2 q CckP∞ i=m+1 e∗ i (x)eik. Then, kx − Gr(x)k . C 3 Remark 5.8. Note that if the inequality kx − Gr(x)k ≤ Ckx − Sm(x)k is satisfied for m and r, then it is automatically satisfied -- with a different constant -- for any n < m and the same r (since Ckx − Sm(x)k ≤ (1 + Kb)Ckx − Sn(x)k where Kb is the basis constant). So we only need to check the condition in the definition of w-partially-greedy for the largest m satisfying w([1, . . . , m]) ≤ w(Ar(x)). Using the constant weight w ≡ 1, we recover the usual definition of a partially-greedy basis. Indeed, for w ≡ 1, the largest m satisfying the definition is m = r, which recaptures the original definition of partially-greedy given in [12]. (cid:3) 5.1. Example of conservative and not democratic basis. Define the set √min A}. Observe that S has the spreading property, i.e, if m ∈ N, (fi)m fi ≤ gi for all i = 1, . . . , n, then (gi)m B ∈ S. S = {A ∈ N<∞ : A ≤ i=1 ∈ Nm with i=1 ∈ S. It also hereditary, i.e., if A ∈ S and B ⊂ A then i=1 ∈ S and (gi)m THE WEIGHTED PROPERTY (A) AND THE GREEDY ALGORITHM 17 Now, let X be the Banach space that we define like the completion of c00 under the norm k(an)nk = sup A∈SXn∈A an. Observe that this is a very slight modification of the Schreier space. Of course, the canonical basis (en)n is a normalized 1-unconditional basis. Note that the hereditary property guarantees that k1Ak = sup F ∈S,F ⊆AF. Now, if A < B and A ≤ B, then there is F ∈ S with F ⊆ A such that k1Ak = F. By the spreading property, we can "push out" F to obtain a set G ⊆ B such that G ∈ S and G = F. Hence, k1Ak = F = G ≤ k1Bk. Thus, the basis is conservative with constant 1. To prove that the basis is not democratic, we can select the sets A = {N 2 + 1, ..., N 2 + N} and B = {1, ..., N} . Then, since A ∈ S, k1Ak = N. However, k1Bk ≤ √N, hence the basis is not democratic: to prove this upper estimate, take a set A1 ∈ S such that k1Bk = A1. Then, min A1 ≤ N, so A1 ≤ √N . Hence, k1Bk ≤ √N . Remark 5.9. Of course, since the canonical basis is unconditional (hence, quasi-greedy) and conservative, is partially-greedy, but not almost-greedy because is not democratic. 5.2. Example of a w-greedy basis which is not conservative (hence not greedy). Definition 5.10. Fix 1 ≤ p < q ≤ ∞, and consider (en)∞ bases of ℓp and ℓq (or c0 if q = ∞). Let w = (wn)∞ space Xq,p,w as the closed subspace [fn ⊕ wnen]∞ basis of c0, we can define Xq,s,w similarly as the subspace [fn ⊕ wnsn]∞ n=1 the respective canonical n=1 ∈ (0,∞)N. We define the Rosenthal-Woo n=1 of ℓq ⊕∞ ℓp. For s = (sn)∞ n=1, the summing n=1 and (fn)∞ n=1 of ℓq ⊕∞ c0. It was mentioned in [13] that if w ∈ (0,∞)N satisfies w ∈ c0 \ ℓ1 then the basis formed by completing c00 under the norm k(an)∞ n=1k∞ ∨ ∞Xn=1 an2wn!1/2 , (an)∞ n=1 ∈ c00, forms a normalized 1-w-greedy basis which is not greedy. In fact, this is just the canonical basis of the Rosenthal-Woo space X∞,2,w1/2. More generally, we have the following. Proposition 5.11. Fix 1 ≤ p < ∞ and w ∈ (0,∞)N ∩ (c0 \ ℓ1). Then the canonical basis of X∞,p,w1/p is 1-w-greedy, but it is not conservative. Proof. Clearly it is unconditional with constant 1. To prove that the canonical basis is w- greedy with constant 1, we need to show that it satisfies the w-Property (A) with constant 1 (Theorem 2.1). For that, take x ∈ X j (x) ≤ 1, and consider A, B ⊂ N such that A ∩ B = ∅, supp (x) ∩ (A ∪ B) = ∅, w(A) ≤ w(B). Then, if ε and η are arbitrary choice ∞,p,w1/p with supj e∗ 18 P. M. BERN ´A, S. J. DILWORTH, D. KUTZAROVA, T. OIKHBERG, AND B. WALLIS of signs, kx + 1εAk = 1 ∨ Xn∈supp (x) ≤ 1 ∨ Xn∈supp (x) e∗ e∗ n(x)pwn + w(A) n(x)pwn + w(B) 1/p 1/p = kx + 1ηBk. Then, the basis satisfies the w-Property (A) with constant 1, hence, using that the basis is unconditional with constant 1, the basis is w-greedy with constant 1. To see that it fails to be conservative, fix m ∈ N and set Am = {1, . . . , m} and Bm,k = {k + 1, . . . , k + m} for each k ∈ N. Now observe that w ∈ c0 ensures that w(Bm,k) → 0 when k → ∞ and hence k1Bm,kk∞,p,w1/p = 1 for sufficiently large k. Hence, we may select km ∈ N so that Bm,km > Am and k1Bm,kmk∞,p,w1/p = 1. On the other hand, w /∈ ℓ1 guarantees that w(Am) → ∞. (cid:3) Remark 5.12. The above proof works for X∞,s,w as well, except that the basis is no longer unconditional. It yields an example of a subspace of c0 with a basis which is w-democratic but not conservative, so long as w ∈ (0,∞)N ∩ (c0 \ ℓ1). However, the situation is different for the canonical basis of Xq,p,w1/p when q 6= ∞. These spaces fail to contain any copies of c0, and hence do not admit v-greedy bases for v non- seminormalized. Even when v is seminormalized the canonical basis may not be v-democratic (nor v-greedy), as we will see momentarily. Proposition 5.13. Fix 1 ≤ p < q < ∞, and let w ∈ (0,∞)N be decreasing. Then the canonical basis of Xq,p,w1/p is unconditional and w-conservative with constants 1. Proof. Select A, B ∈ N<∞ with w(A) ≤ w(B) and A < B. Since w is decreasing, we must have A ≤ B. Thus, k1Akq,p,w1/p = (A)1/q ∨ w(A)1/p ≤ (B)1/q ∨ w(B)1/p = k1Bkq,p,w1/p. (cid:3) Proposition 5.14. Fix 1 ≤ p < q < ∞ and 0 < θ < 1 − p n=1 ∈ (0,∞)N be defined by wn = n−θ for n ∈ N. Then the canonical basis for Xq,p,w1/p is not conservative, and not w-democratic. q . Let w = (wn)∞ Proof. First establish that our basis is not conservative. As in the proof of Proposition 5.11, for k, m ∈ N we set Am = {1, . . . , m} and Bm,k = {k + 1, . . . , k + m}, and for each m ∈ N we find km ∈ N large enough that Am < Bm,km and k1Bm,kmkq,p,w1/p = m1/q. Meanwhile, so that, due to 1 − θ − p/q > 0, k1Amkq,p,w1/p ≥ w(Am)1/p = mXn=1 1 − θ (cid:19)1/p k1Bm,kmkq,p,w1/p ≥(cid:18)m1−θ − 1 k1Amkq,p,w1/p ≥(cid:18)Z m t−θ dt(cid:19)1/p n−θ!1/p m−1/q =(cid:18) m1−θ−p/q − m−p/q 1 − θ =(cid:18)m1−θ − 1 1 − θ (cid:19)1/p (cid:19)1/p → ∞. 1 THE WEIGHTED PROPERTY (A) AND THE GREEDY ALGORITHM 19 We next sketch the proof of the lack of w-democracy. To this end, consider the sets An and Bm,k as defined in the preceding paragraph. There exist universal constants c and C so that w(An) ≥ cn1−θ, and k1Ank ≤ Cn1−θ. Then w(Bm,k) < mk−θ, while k1Bm,kk ≥ m1/q. For large values of k, select m ∈ [cn1−θkθ/2, cn1−θkθ]. Then w(Bm,k) ≤ w(An), yet the ratio k1Bm,kk k1Ank ≈ m1/q n1−θ ≈ n(1−θ)/qkθ/q n1−θ = kθ/qn−(1−θ)(1−1/q) can be arbitrarily large (for large k), ruling out the possibility of w-democracy. (cid:3) Corollary 5.15. Fix 1 ≤ p < q < ∞ and 0 < θ < 1 − p n=1 ∈ (0,∞)N be defined by wn = n−θ for n ∈ N. Then the canonical basis for Xq,p,w1/p is not v-democratic for any weight v ∈ (0,∞)N. Proof. Suppose, for the sake of contradiction, that the canonical basis for Xq,p,w1/p is v-democratic. Note that this basis contains no subsequences equivalent to the c0-basis. Proposition 3.10 shows that 0 < inf vn ≤ sup vn < ∞ -- that is, the weight v is equivalent to a constant. Then, by Remark 3.3, the canonical basis for Xq,p,w1/p has to be democratic, hence conservative. This, however, contradicts Proposition 5.14. (cid:3) q . Let w = (wn)∞ Remark 5.16. Consider again the weight w from Proposition 5.14. • It follows from Propositions 5.13 and 5.14 that the canonical basis of Xq,p,w1/p is w- partially-greedy basis, but not w-almost-greedy. However, the space Xq,p,w1/p does have Indeed, we recall from [3, Theorem 10.7.1] that if X has a an almost-greedy basis. complemented subspace with a symmetric basis and finite cotype then X admits an almost-greedy basis. If w ∈ (0,∞)N ∩ (c0 \ ℓ(pq)/(q−p)) then the Woo-Rosenthal spaces Xq,p,w contain complemented copies of ℓp and ℓq (or c0 if q = ∞; see [20, Corollary 3.2]), and hence satisfy this condition. • Just as in Proposition 5.14, one can show that the canonical basis of Xq,s,w is not conservative when 0 < θ < 1− 1 q . However, it is not quasi-greedy, either. For θ = 1− 1 q , this basis becomes quasi-greedy and democratic (for q = 2, this was observed in [3, Example 10.2.9], the argument is valid for all 1 < q < ∞). 6. Questions • Does Property (D) imply Property (C)? • Does Property (A) imply Property (D)? • If a basis has Properties (C) and (A), is it necessarily semi-greedy? • Is it possible to formulate a new property so that every conservative basis with this • If w = (1, 1, ...) in Theorem 5.7, do we get the same constant as in the classic case ([12, property is necessarily democratic? Theorem 3.4])? The purpose of this appendix is to show two basic lemmas. The first one resembles a result 7. Appendix from [7]. For each λ > 0, we define the λ-truncation of z ∈ C by Tλ(z) =(cid:26) λsgn (z) z if z ≥ λ if z ≤ λ 20 P. M. BERN ´A, S. J. DILWORTH, D. KUTZAROVA, T. OIKHBERG, AND B. WALLIS We extend Tλ to an operator on X by Tλ(x) =Xj Tλ(e∗ j (x))ej = Xj∈Λλ λsgn (e∗ j (x))en +Xj∈Λc λ e∗ j (x)ej, where Λλ = {j : λ < e∗ Lemma 7.1. For all λ > 0 and x ∈ X, if B is Cq-quasi-greedy, we have j (x)}. kTλ(x)k ≤ Cqkxk, k(I − Tλ)(x)k ≤ (Cq + 1)kxk, αk1εΛk ≤ 2Cqkxk, where α = minj∈Λ e∗ j (x)). Moreover, if B is Ku-unconditional, for every set A with A < ∞, j (x), Λ is a greedy set of x and ε ≡ sgn (e∗ kTλ(I − PA)(x)k ≤ Kukxk. Proof. Tλ(x) =Z 1 0 "Xj (x) ](s)e∗ λ e∗ j j (x)ej# ds =Z 1 0 χ[0, (I − PΛλ,s)x ds, where Λλ,s = {j : λ Minkowski's integral inequality, s < e∗ j (x)} is a greedy set of x of finite cardinality. Then, using the kTλ(x)k ≤Z 1 0 k(I − PΛλ,s)xkds ≤ Cqkxk. 0 PΛλ,s(x)ds, hence Also, since (I − Tλ)x =R 1 Now, since α1εΛ = Tα(x) − PΛc(x) =R 1 k(I − Tλ)(x)k ≤ (Cq + 1)kxk. 0 (PΛ(x) − PΛα,s(x))ds, αk1εΛk ≤ 2Cqkxk. On the other hand, if A is a general set with A < ∞, Tλ(I − PA)x =Z 1 0 (I − PΛλ,s)(I − PA)x ds =Z 1 0 (I − PA∪Λλ,s)x ds, thus kTλ(I − PA)(x)k ≤ Kukxk. (cid:3) The second lemma involves the concept of w-partially-greedy bases. Lemma 7.2. If B is Cw-w-conservative and Cq-quasi-greedy, then j∈A ajk1ηBk, ajejk ≤ 4CqCw max kXj∈A for any sign η and A, B ∈ N<∞ such that w(A) ≤ w(B), A < B and any collection of scalars (aj)j∈A. THE WEIGHTED PROPERTY (A) AND THE GREEDY ALGORITHM 21 Proof. We prove that k1εAk ≤ 4CqCwk1ηBk for any signs ε and η. First, we can decompose 1εA = 1A+ − 1A−, where A± = {j ∈ A : εj = ±1}. Then, Now, using the condition to be quasi-greedy, it is clear that k1Bk ≤ 2Cqk1ηBk, then k1εAk ≤ k1A+k + k1A−k ≤ 2Cwk1Bk. k1εAk ≤ 4CqCwk1ηBk. Now, using convexity, we are done. (cid:3) Acknowledgments: The first author thanks the University of Murcia for partially support- ing of his research stay in the University of Illinois at Urbana-Champaign in September 2017, where this paper began. References [1] F. Albiac, J. L. Ansorena, Characterization of 1-quasi greedy bases, J. Approx. Theory, 201, 7-12 (2016). [2] F. Albiac, J. L. Ansorena, Characterization of 1-almost greedy bases, Rev. Mat. Complut., 30(1), 13-24 (2017). [3] F. Albiac and N. Kalton, Topics in Banach Space Theory, second edition, ISBN 978-3-319-31555-3 (2016). [4] F. Albiac, P. Wojtaszczyk, Characterization of 1-greedy bases, J. Approx. Theory, 138 (2006), 65-86. [5] P. M. Bern´a, ´O. Blasco, Characterization of greedy bases in Banach spaces, J. Approx. Theory, 205 (2017), 28-39. [6] P. M. Bern´a, ´O. Blasco, The best m-term approximation with respect to polynomials with constant coefficients, Anal. Math., 43 (2) (2017), 119-132. [7] P. M. Bern´a, ´O. Blasco, G. Garrig´os, Lebesgue inequalities for the greedy algorithm in general bases, Rev. Mat. Complut. 30 (2017), 369-392. [8] A. Cohen, R. A. DeVore, R. Hochmuth, Restricted nonlinear approximation, Constr. Approx., 16 (2000), 85-113. [9] S. J. Dilworth, N. J. Kalton, D. Kutzarova, On the existence of almost greedy bases in Banach spaces, Studia Math. 159 (2003), 67 -- 101. [10] S. J. Dilworth, D. Kutzarova, T. Oikhberg, Lebesgue constants for the weak greedy algorithm, Rev. Mat. Complut. 28(2), 393 -- 409 (2015). [11] S. J. Dilworth, D. Kutzarova, E. Odell, T. Schlumprecht, A. Zs´ak, Renorming spaces with greedy bases, J. Approx. Theory 188 (2014), 39-56. [12] S. J. Dilworth, N. J. Kalton, D. Kutzarova, V. N. Temlyakov, The thresholding greedy algorithm, greedy bases, and duality, Constr. Approx. 19 (2003), no.4, 575-597. [13] S. J. Dilworth, D. Kutzarova, V. N. Temlyakov, B. Wallis, Weight-Almost greedy bases. (Preprint) http://arxiv.org/abs/1803.02932v1. [14] G. Garrig´os, E. Hern´andez, T. Oikhberg, Lebesgue-type inequalities for quasi-greedy bases, Constr. Approx. 38 (2013), 447-470. [15] G. Kerkyacharian, D. Picard, V. N. Temlyakov, Some inequalities for the tensor product of greedy bases and weight-greedy bases, East J. Approx., 12 (2006), 103-118. [16] S. V. Konyagin, V. N. Temlyakov, A remark on greedy approximation in Banach spaces, East J. Approx. 5 (1999), 365-379. [17] V. N. Temlyakov, Greedy approximation, Cambridge Monographs on Applied and Computational Math- ematics, vol.20, Cambridge University Press, Cambridge, 2011. [18] P. Wojtaszczyk, Greedy algorithm for general biorthogonal systems, J. Approx.Theory 107 (2) (2000), 293-314. [19] P. Wojtaszczyk, Greedy type bases in Banach spaces, Constructive theory of functions, 136-155, DARBA, Sofia, 2003. [20] J. Y. T. Woo, On a class of universal modular sequence spaces, Israel Journal of Mathematics 20 (1975), 193-215. 22 P. M. BERN ´A, S. J. DILWORTH, D. KUTZAROVA, T. OIKHBERG, AND B. WALLIS Pablo M. Bern´a, Departmento de Matem´aticas, Universidad Aut´onoma de Madrid, 28049 Madrid, Spain E-mail address: [email protected] Stephen J. Dilworth, Department of Mathematics, University of South Carolina, Columbia SC 29208, USA E-mail address: [email protected] Denka Kutzarova, Department of Mathematics, University of Illinois Urbana-Champaign, Urbana, IL 61801, USA; and Institute of Mathematics and Informatics, Bulgarian Academy of Sciences E-mail address: [email protected] Timur Oikhberg, Department of Mathematics, University of Illinois Urbana-Champaign, Ur- bana, IL 61801, USA E-mail address: [email protected] Ben Wallis, Department of Mathematical Sciences, Northern Illinois University, DeKalb, IL 60115-2888, USA E-mail address: [email protected]
1303.4040	1	1303	2013-03-17T07:48:36	Criteria for Hankel operators to be sign-definite	[ "math.FA", "math.SP" ]	We show that total multiplicities of negative and positive spectra of a self-adjoint Hankel operator $H$ with kernel $h(t)$ and of an operator of multiplication by some real function $s(x)$ coincide. In particular, $\pm H\geq 0$ if and only if $\pm s(x)\geq 0$. The kernel $h(t)$ and its "sign-function" $s(x)$ are related by an explicit formula. An expression of $h(t)$ in terms of $s(x)$ leads to an exponential representation of $h(t)$. Our approach directly applies to various classes of Hankel operators. In particular, for Hankel operators of finite rank, we find an explicit formula for the total multiplicity of their negative and positive spectra.	math.FA	math	CRITERIA FOR HANKEL OPERATORS TO BE SIGN-DEFINITE D. R. YAFAEV Abstract. We show that total multiplicities of negative and positive spectra of a self-adjoint Hankel operator H with kernel h(t) and of an operator of multiplication by some real function s(x) coincide. In particular, ±H ≥ 0 if and only if ±s(x) ≥ 0. The kernel h(t) and its "sign-function" s(x) are related by an explicit formula. An expression of h(t) in terms of s(x) leads to an exponential representation of h(t). Our approach directly applies to various classes of Hankel operators. In particular, for Hankel operators of finite rank, we find an explicit formula for the total multiplicity of their negative and positive spectra. 1.1. Hankel operators can be defined as integral operators 1. Introduction (Hf )(t) =Z ∞ 0 h(t + s)f (s)ds (1.1) in the space L2(R+) with kernels h that depend on the sum of variables only. Of course H is symmetric if h(t) = h(t). There are very few cases where Hankel operators can be explicitly diagonalized. We mention classical results by F. Mehler [8], T. Carleman [3], W. Magnus [7] and M. Rosenblum [12]. They are treated in a unified way in [16] where some new examples are also considered. Our goal here is to find necessary and sufficient conditions for the positivity1 of Hankel operators. This question seems to be of particular importance because of its intimate relation to a representation of the function h(t) as the Laplace transform of a (positive) measure. Such representations are continuous analogues of moment problems. According to the Hamburger theorem (see, e.g., the book [1]) the posi- tivity of a discrete Hankel operator is equivalent to the existence of a solution of the corresponding moment problem. 2000 Mathematics Subject Classification. 47A40, 47B25. Key words and phrases. Hankel operators, convolutions, necessary and sufficient conditions for the positivity, the sign-function, operators of finite rank, the Carleman operator and its perturbations. 1We always use the term "positive" instead of a more precise but lengthy term "non-negative." 1 b(ξ − η)g(η)dη, −∞ (Bg)(ξ) =Z ∞ 2πR ∞ R ∞ b(ξ) = 1 0 h(t)t−iξdt 0 e−tt−iξdt . (1.2) (1.3) 2 D. R. YAFAEV 1.2. Our condition of the positivity2 of Hankel operators is quite explicit. Let B, be the operator in the space L2(R) of the convolution with the function Of course b(−ξ) = b(ξ) if h(t) = h(t) so that the operator B is symmetric. Our main result is that the Hankel operator H ≥ 0 if and only if B ≥ 0. We call b(ξ) the b-function of the Hankel operator H (or of the kernel h(t)), and we use the term the sign-function or s-function for its Fourier transform s(x). So our result, roughly speaking, means that a Hankel operator H is positive if and only if its sign-function s(x) is positive. Note that in specific examples we consider, functions s(x) may be of a quite different nature. For instance, s(x) may be a polynomial or, on the contrary, it may be a distribution; for example, it may be a combination of delta functions and their derivatives. Our proofs rely on the identity (Hf, f ) = (Bg, g) (1.4) where g(ξ) = Γ(1/2 + iξ) f (ξ) =: (Ξf )(ξ), (1.5) f (ξ) is the Mellin transform of f (t) and Γ(·) is the gamma function. Actually, the identity (1.4) or, equivalently, (1.6) allows us to find the numbers of positive and negative eigenvalues of a Hankel op- erator H. For a self-adjoint operator A, we denote by N+(A) (by N−(A)) the total mutiplicity of its strictly positive (negative) spectrum. Then the identity (1.6) shows that H = Ξ∗BΞ N±(H) = N±(B). (1.7) This result can be compared with Sylvester's inertia theorem which states the same for Hermitian matrices H and B related by equation (1.6) provided the matrix Ξ is invertible. In contrast to the linear algebra, in our case the operators H and B are of a completely different nature and B (but not H) admits an explicit spectral analysis. Of course the results above can be reformulated in terms of the sign-function s(x) = √2π(Φ∗b)(x) where Φ is the Fourier transform. Let us introduce the operator S = Φ∗BΦ of multiplication by s(x). Then (1.6) means that H = bΞ∗SbΞ where bΞ = Φ∗Ξ, and (1.7) means that N±(H) = N±(S). 2We usually discuss conditions for H ≥ 0, but of course replacing H by −H we obtain conditions for H ≤ 0. CRITERIA FOR HANKEL OPERATORS TO BE SIGN-DEFINITE 3 1.3. The precise meaning of formula (1.3) requires some discussion. Observe that the denominator in (1.3) coincides with the numerator for the special case h(t) = e−t. It equals Γ(1 − iξ) and hence exponentially tends to zero as ξ → ∞. Therefore b(ξ) is a "nice" function of ξ only under very restrictive assumptions on the kernel h(t). Thus to cover natural examples, we have, on the contrary, to extend a class of kernels and to work with distributions h(t). The choice of appropriate spaces of distributions is also very important. In order to be able to divide in (1.3) by an exponentially decaying function, we assume that the numerator belongs to the 0 (R)′. It means that the Fourier transform of the function class of distributions C ∞ 0 (R)′. The corresponding class of functions h(t) θ(x) = exh(ex) should belong to C ∞ 0 (R)′. The will be denoted Z ′ Schwartz space S(R)′ is too restrictive for our purposes which is seen already on the example of finite rank Hankel operators. 0 (R)′ for the validity of the identity (1.4) is very general. It is satisfied for all bounded, but also for a wide class of unbounded, Hankel operators H. More than that, it is not even required that H be defined by formula (1.1) on some dense set. Therefore we work with quadratic forms (Hf, f ) which is more convenient and yields more general results. In this context it is natural to consider distributions h(t) which makes the theory self-consistent. +. Under the assumption h ∈ Z ′ The condition Φθ ∈ C ∞ + we have b ∈ C ∞ If however h ∈ L1 loc(R+), then θ ∈ S(R)′ and hence Φθ ∈ C ∞ 0 (R)′ if Z ∞ 0 h(t)(1 + ln t)−κdt < ∞ (1.8) for some κ. Condition (1.8) is quite general; corresponding Hankel operator be bounded. For example, it admits kernels it also does not require that the h(t) = P (ln t)t−1 (1.9) where P (x) is an arbitrary polynomial. Note that Hankel operators with such kernels are bounded for P (x) = const only. As far as test functions f (t) are concerned, we require that their Mellin transforms f ∈ C ∞ 0 (R). Then both sides of (1.4) are well defined and the identity holds. We note that the distribution b(ξ) is "worse" than the kernel h(t). On the contrary, due to the factor Γ(1/2 + iξ) in (1.5), the test function g(ξ) is "better" than f (t). In the case of bounded operators H, this permits us to extend the main identity (1.4) to all elements f ∈ L2(R+). 1.4. It turns out that the knowledge of the sign-function s(x) allows one to recover the kernel h(t) by the formula where e−tλh♮(λ)λdλ h(t) =Z ∞ h♮(λ) = λ−1s(− ln λ). 0 (1.10) (1.11) 4 D. R. YAFAEV Thus h(t) is the Laplace transform of the distribution λh♮(λ). It is noteworthy that + and the correspondence h 7→ h♮ is a continuous one-to-one mapping of Z ′ h♮ ∈ Z ′ onto itself. Representation (1.10) does not require the positivity of H. If however H ≥ 0, then combining our results with the Bochner-Schwartz theorem, we obtain that λh♮(λ)dλ = dm(λ) where dm(λ) is a positive measure. In this case + h(t) =Z ∞ 0 e−tλdm(λ). (1.12) This representation implies that h(t) is necessarily a completely monotonic function and, in particular, h ∈ C ∞(R+). Note that the converse statement is also true: a completely monotonic function admits representation (1.12). This is one of famous Bernstein's theorems (see his paper [2], or the book by N. Akhiezer [1] or the book by D. V. Widder [13]). In contrast to the Bernstein theorem, we deduce representation (1.12) from the positivity of the Hankel operator with kernel h(t) and show that the measure dm(λ) satisfies for some κ the condition Z ∞ 0 (1 + ln λ)−κλ−1dm(λ) < ∞. (1.13) Hankel operators can also be realized as operators in the space of sequences l2 +. The relation between this discrete representation and the continuous representation we consider is given by the unitary transformation of l2 + onto L2(R+) constructed in terms of the Laguerre polynomials. Thus all our results can in principle be translated into the discrete representation. This hopefully will be discussed in another article on this subject. 1.5. A large part of the paper is devoted to applying the general theory to various classes of Hankel operators although we do not try to cover all possible cases. In some examples the sign definiteness of H can also be verified or refuted with the help of Bernstein's theorems. Note however that our approach yields additionally an explicit formula for the total numbers of negative and positive eigenvalues of H. In Section 5, we present such a formula for Hankel operators of finite rank. Then we consider two specific examples. The first one is given by the formula h(t) = tke−αt, α > 0, k ≥ −1. (1.14) Note that the Hankel operator H with such kernel has finite rank for k ∈ Z+ only. We show that H is positive if and only if k ≤ 0. The second class of kernels is defined by the formula (1.15) It turns out that the corresponding Hankel operator is positive if and only if r ≤ 1. Section 6 is devoted to a study of Hankel operators H with non-smooth ker- nels. In this case both numbers N±(H) are infinite, and we find the asymptotics of eigenvalues of H. h(t) = e−tr r > 0. , CRITERIA FOR HANKEL OPERATORS TO BE SIGN-DEFINITE 5 Finally, in Section 7 we consider perturbations of the Carleman operator, that is, of the Hankel operator with kernel h0(t) = t−1 by various classes of compact Hankel operators. The Carleman operator can be explicitly diagonalized by the Mellin transform. We recall that it has the absolutely continuous spectrum [0, π] of multiplicity 2. The Carleman operator plays the distinguished role in the theory of Hankel operators. In particular, it is important for us that its sign-function s(x) = 1. As was pointed out by J. S. Howland in [6], Hankel operators are to a certain extent similar to differential operators. In terms of this analogy, the Carleman operator H0 plays the role of the "free" Schrodinger operator D2, D = −id/dx, in the space L2(R). Furthermore, Hankel operators H with "perturbed" kernels h(t) = t−1 + v(t) can be compared to Schrodinger operators D2 + V(x). The assumption that v(t) decays sufficiently rapidly as t → ∞ and is not too singular as t → 0 corresponds to a sufficiently rapid decay of the potential V(x) as x → ∞. As shown in [17], the results on the discrete spectrum of the operator H lying above its essential spectrum [0, π] are close in spirit to the results on the discrete spectrum of the Schrodinger operator D2 + V(x). On the contrary, the results on the negative spectrum of the Hankel operator H are drastically different. In particular, contrary to the case of differential operators with decaying coefficients, the finiteness of the negative spectrum of the Hankel operator H is not determined by the behaviour of v(t) at singular points t = 0 and t = ∞. As an example, consider the Hankel operator with kernel h(t) = t−1 − γe−tr , r ∈ (0, 1). Now the kernel of the perturbation is the function which decays faster than any power of t−1 as t → ∞ and it has the finite limit as t → 0. Nevertheless we show that the negative spectrum of H is infinite if γ > γ0 (here γ0 = γ0(r) is an explicit constant) while H is positive if γ ≤ γ0. Such a phenomenon has no analogy for Schrodinger operators with decaying potentials. However it occurs for three-particle Schrodinger operators and is known as the Efimov effect. We also study perturbations of the Carleman operator H0 by Hankel operators V of finite rank. Here we obtain a striking result: the total numbers of negative eigenvalues of the operators H = H0 + V and V coincide. As examples, we consider only bounded Hankel operators in this paper. However, our general results directly apply to a wide class of unbounded operators, such as Hankel operators with kernels (1.9). Moreover, with slight modifications our method works also for kernels (1.14) where α ≥ 0 and k is an arbitrary negative number. In this case condition (1.8) is not satisfied. Hankel operators with kernels (1.9) and (1.14) generalize the Carleman operator, and we call them quasi-Carleman operators. They will be studied elsewhere. 1.6. Let us briefly describe the structure of the paper. We obtain the main iden- tity (1.4) and the reconstruction formula (1.10) in Section 2. Necessary information on bounded Hankel operators (including a continuous version of the Nehari theorem) 6 D. R. YAFAEV is collected in Section 3. In Sections 2 and 3 we do not assume that the function h is real, i.e., the corresponding Hankel operator H is not necessarily symmetric. Spectral consequences of the formula (1.4) and, in particular, criteria of the sign- definiteness of Hankel operators are formulated in Section 4. In Sections 5, 6 and 7 we apply the general theory to particular classes of Hankel operators. Let us introduce some standard notation. We denote by Φ, (Φu)(ξ) = (2π)−1/2Z ∞ −∞ u(x)e−ixξdx, the Fourier transform. The space Z = Z(R) of test functions is defined as the subset of the Schwartz space S = S(R) which consists of functions ϕ admitting the analytic continuation to entire functions in the complex plane C and satisfying bounds ϕ(z) ≤ Cn(1 + z)−ner Im z, ∀z ∈ C, for some r = r(ϕ) > 0 and all n. We recall that the Fourier transform Φ : C ∞ 0 (R) → Z and Φ∗ : Z → C ∞ 0 (R). The dual classes of distributions (continuous antilinear functionals) are denoted S ′, C ∞ 0 (R)′ and Z ′, respectively. In general, for a linear topological space L, we use the notation L′ for its dual space. The Dirac function is standardly denoted δ(·). We use the notation hhh·,·iii and h·,·i for scalar products and duality symbols in L2(R+) and L2(R), respectively. They are always linear in the first argument and antilinear in the second argument. The letter C (sometimes with indices) denotes various positive constants whose precise values are inessential. 2. The main identity 2.1. Let us consider a Hankel operator H defined by equality (1.1) in the space L2(R+). Actually, it is more convenient to work with sesquilinear forms (Hf1, f2) instead of operators. Before giving precise definitions, let us explain our construction at a formal level. It follows from (1.1) that (Hf1, f2) =Z ∞ 0 Z ∞ =Z ∞ h(t)F (t)dt =: hhhh, Fiii, 0 0 h(t + s)f1(s)f2(t)dtds where F (t) =Z t 0 f1(s)f2(t − s)ds =: ( ¯f1 ⋆ f2)(t) (2.1) (2.2) is the Laplace convolution of the functions ¯f1 and f2. Formula (2.1) allows us to consider h as a distribution with the test function F defined by (2.2). Thus the Hankel form will be defined by the relation h[f1, f2] = hhhh, ¯f1 ⋆ f2iii. (2.3) CRITERIA FOR HANKEL OPERATORS TO BE SIGN-DEFINITE 7 Let us introduce the test function and the distribution Ω(x) = F (ex) =: (RF )(x) θ(x) = exh(ex) defined for x ∈ R. Setting in (2.1) t = ex, we see that hhhh, Fiii =Z ∞ −∞ θ(x)Ω(x)dx =: hθ, Ωi. (2.4) (2.5) (2.6) We are going to consider the form (2.6) on pairs F, h such that the corresponding test function Ω defined by (2.4) is an element of the space Z of analytic functions and the corresponding distribution θ defined by (2.5) is an element of the dual space Z ′. The set of all such F and h will be denoted Z+ and Z ′ +, respectively, that is, F ∈ Z+ ⇐⇒ Ω ∈ Z and h ∈ Z ′ Of course, the topology in Z+ is induced by that in Z and Z ′ that h ∈ Z ′ the corresponding function (2.5) satisfies the condition (2.7) + is dual to Z+. Note loc(R+) and integral (1.8) is convergent for some κ. In this case + ⇐⇒ θ ∈ Z ′. + if h ∈ L1 Z ∞ −∞ θ(x)(1 + x)−κdx < ∞, and hence θ ∈ S ′ ⊂ Z ′. Define the unitary operator U : L2(R+) → L2(R) by the equality (Uf )(x) = ex/2f (ex). (2.8) Let the set D consist of functions f (t) such that Uf ∈ Z. Since f (t) = t−1/2(Uf )(ln t) and Z ⊂ S, we see that functions f ∈ D and their derivatives satisfy the estimates f (m)(t) = Cn,mt−1/2−m(1 + ln t)−n for all n and m. Obviously, f ∈ D if and only if ϕ(t) = t1/2f (t) belongs to the class Z+. Let us show that form (2.3) is correctly defined on functions f1, f2 ∈ D. To that end, we have to verify that function (2.2) belongs to the space Z+ or, equivalently, function (2.4) belongs to the space Z. This requires some preliminary study which will also allow us to derive a convenient representation for form (2.3). Recall that the Mellin transform M : L2(R+) → L2(R) is defined by the formula (2.9) f (t)t−1/2−iξdt. (Mf )(ξ) = (2π)−1/2Z ∞ 0 8 D. R. YAFAEV Of course, M = ΦU where Φ is the Fourier transform and U is operator (2.8). Since both Φ and U are unitary, the operator M is also unitary. The inversion of the formula (2.9) is given by the relation f (t) = (2π)−1/2Z ∞ −∞ f (ξ)t−1/2+iξdξ, f = Mf. (2.10) Let Γ(z) be the gamma function. Recall that Γ(z) is a holomorphic function in the right half-plane and Γ(z) 6= 0 for all z ∈ C. According to the Stirling formula the function Γ(z) tends to zero exponentially as z → ∞ parallel with the imaginary axis. To be more precise, we have Γ(α + iλ) = eπi(2α−1)/4(2π/e)1/2λα−1/2eiλ(ln λ−1)e−πλ/2(cid:0)1 + O(λ−1)(cid:1) (2.11) for a fixed α > 0 and λ → +∞. Since Γ(α − iλ) = Γ(α + iλ), this yields also the asymptotics of Γ(α + iλ) as λ → −∞. If fj ∈ D, j = 1, 2, then fj = Mfj = ΦUfj ∈ C ∞ 0 (R) and hence the functions gj(ξ) = Γ(1/2 + iξ) fj(ξ), j = 1, 2, (2.12) also belong to the class C ∞ g1 and g2, 0 (R). Let us introduce the convolution of the functions (g1 ∗ g2)(ξ) =Z ∞ −∞ g1(ξ − η)g2(η)dη, and set (J g)(ξ) = g(−ξ). We have the following result. Lemma 2.1. Suppose that fj ∈ D, j = 1, 2, and define functions gj(ξ) by equality (2.12). Let the function Ω(x) be defined by formulas (2.2) and (2.4). Then (ΦΩ)(ξ) = (2π)−1/2Γ(1 + iξ)−1((J ¯g1) ∗ g2)(ξ). (2.13) Proof. Substituting (2.10) into definition (2.2), we see that F (t) = (2π)−1Z t 0 dsZ ∞ −∞ f1(τ )(t − s)−1/2−iτ dτZ ∞ −∞ f2(σ)s−1/2+iσdσ. Observe that (t − s)−1/2−iτ s−1/2+iσds = ti(σ−τ ) Γ(1/2 − iτ )Γ(1/2 + iσ) . Γ(1 + i(σ − τ )) Then using definition (2.12) we obtain the representation ti(σ−τ )Γ(1 + i(σ − τ ))−1g1(τ )g2(σ)dτ dσ −∞Z ∞ tiξΓ(1 + iξ)−1((J ¯g1) ∗ g2)(ξ)dξ, −∞ 0 Z t F (t) =(2π)−1Z ∞ =(2π)−1Z ∞ −∞ CRITERIA FOR HANKEL OPERATORS TO BE SIGN-DEFINITE 9 whence Ω(x) = (2π)−1Z ∞ −∞ eiξxΓ(1 + iξ)−1((J ¯g1) ∗ g2)(ξ)dξ. This is equivalent to formula (2.13). (cid:3) Observe that the function Γ(1 + iξ)−1 in the right-hand side of (2.13) tends to 0 (R) because (J ¯g1) ∗ g2 ∈ infinity exponentially as ξ → ∞. Nevertheless ΦΩ ∈ C ∞ C ∞ 0 (R) for g1, g2 ∈ C ∞ Corollary 2.2. Let fj ∈ D, j = 1, 2, and let the function Ω(x) be defined by formulas (2.2) and (2.4). Then Ω ∈ Z or, equivalently, F ∈ Z+. 0 (R). Thus we have Now we are in a position to give the precise definition. Definition 2.3. Let h ∈ Z ′ form is defined by the relation (2.3). + and fj ∈ D, j = 1, 2. Then the Hankel sesquilinear We shall see in subs. 2.4 that h ∈ Z ′ Of course definition (2.3) can be rewritten as hhhh, ¯f1 ⋆ f2iii on f1, f2 ∈ D, that is, h = 0 if hhhh, ¯f1 ⋆ f2iii = 0 for all f1, f2 ∈ D. + is determined uniquely by the values where and θ is distribution (2.5). h[f1, f2] = hθ, Ωi Ω(x) = ( ¯f1 ⋆ f2)(ex) (2.14) We sometimes write h[f1, f2] as integral (2.1) keeping in mind that its precise meaning is given by Definition 2.3. 2.2. Our next goal is to show that (2.14) is the sesquilinear form of the convolution operator B, that is, it equals the right-hand side of (1.4). Here the representation of Lemma 2.1 for the function G(ξ) = √2πΓ(1 + iξ)(ΦΩ)(ξ) (2.15) plays the crucial role. Since θ ∈ Z ′, its Fourier transform a = Φθ is correctly defined as an element of 0 (R)′. Formally, C ∞ a(ξ) = (Φθ)(ξ) = (2π)−1/2Z ∞ 0 h(t)t−iξdt, (2.16) that is, a(ξ) is the Mellin transform of the function h(t)t1/2. Let Ω ∈ Z. Passing to the Fourier transforms and using notation (2.15), we see that where G ∈ C ∞ 0 (R) and the distribution b ∈ C ∞ 0 (R)′ is given by the relation hθ, Ωi = ha, ΦΩi = hb, Gi b(ξ) = (2π)−1/2a(ξ)Γ(1 − iξ)−1 which is of course the same as (1.3). Thus we are led to the following (2.17) (2.18) 10 D. R. YAFAEV +. The distribution b ∈ C ∞ 0 (R)′ defined by formulas Definition 2.4. Let h ∈ Z ′ (2.5), (2.16) and (2.18) is called the b-function of the kernel h(t) (or of the Hankel operator H). Its Fourier transform s = √2πΦ∗b ∈ Z ′ is called the s-function or the Recall that the distribution h♮ ∈ Z ′ + was defined by relation (1.11). The following assertion is an immediate consequence of formulas (2.5), (2.16) and (2.18). sign-function. Proposition 2.5. The mappings h 7→ θ 7→ a 7→ b 7→ s 7→ h♮ yield one-to-one correspondences (bijections) 0 (R) → C ∞ Z ′ + → Z ′ → C ∞ 0 (R) → Z ′ → Z ′ +. All of them, as well as their inverse mappings, are continuous. Putting together equalities (2.6) and (2.17), we see that hhhh, Fiii = hb, Gi. (2.19) Combining this relation with Lemma 2.1 and Definitions 2.3, 2.4 and using notation (1.5), we obtain the main identity (1.4). To be more precise, we have the following result. 0 (R)′ be the corresponding b- Theorem 2.6. Suppose that h ∈ Z ′ function. Let fj ∈ D, j = 1, 2, and let the functions gj = Ξfj be defined by formula (2.12). Then gj ∈ C ∞ +, and let b ∈ C ∞ 0 (R) and the representation hhhh, ¯f1 ⋆ f2iii = hb, (J ¯g1) ∗ g2i =: b[g1, g2] (2.20) holds. Passing in the right-hand side of (2.20) to the Fourier transforms and using that Φ∗((J ¯g1) ∗ g2) = (2π)1/2Φ∗g1Φ∗g2, we obtain Corollary 2.7. Let s ∈ Z ′ be the sign-function of h, and let uj = Φ∗gj = Φ∗Ξfj ∈ Z. Then (2.21) hhhh, ¯f1 ⋆ f2iii = hs, ¯u1u2i =: s[u1, u2]. Loosely speaking, equalities (2.20) and (2.21) mean that hhhh, ¯f1 ⋆ f2iii =Z ∞ −∞Z ∞ =Z ∞ −∞ −∞ b(ξ − η)g1(η)g2(ξ)dξdη (2.22) s(x)u1(x)u2(x)dx. CRITERIA FOR HANKEL OPERATORS TO BE SIGN-DEFINITE 11 In the particular case h(t) = t−1, we have θ(x) = 1, a(ξ) = (2π)1/2δ(ξ), b(ξ) = δ(ξ), s(x) = 1, (2.23) and hence (2.22) yields hhhh, ¯f1 ⋆ f2iii =Z ∞ −∞ where g1(ξ)g2(ξ)dξ =Z ∞ −∞ Γ(1/2 + iξ)2 f1(ξ) f2(ξ)dξ, Γ(1/2 + iξ)2 = π cosh(πξ) . (2.24) This leads to the familiar diagonalization of the Hankel operator H with kernel h(t) = t−1. This operator is known as the Carleman operator and will be denoted by C. Let ΓΓΓα : C ∞ 2.3. According to Proposition 2.5 the distribution h♮ determines uniquely the distribution h. Let us now obtain an explicit formula for the mapping h♮ 7→ h. This requires some auxiliary information. 0 (R) → C ∞ 0 (R), α > 0, be the operator of multiplication by the function Γ(α + iξ). Making the change of variables t = e−x in the definition of the gamma function, we see that Γ(α + iλ) =Z ∞ e−ttα+iλ−1dt =Z ∞ e−αxe−ixλdx, α > 0, e−e−x −∞ 0 and hence It follows that −∞ (2π)−1Z ∞ (Φ∗ΓΓΓαΦΩ)(x) =Z ∞ −∞ Let us also introduce the operator Lα: (LαF )(λ) = λαZ ∞ 0 eixλΓ(α + iλ)dλ = e−e−x e−αx. eα(y−x)e−ey−x Ω(y)dy. (2.25) (2.26) e−tλtα−1F (t)dt, λ > 0, α > 0. Obviously, LαF ∈ C ∞(R+) for all bounded functions F (t) and, in particular, for F ∈ Z+. Note that Lα is the Laplace operator L, (LF )(λ) =Z ∞ 0 e−tλF (t)dt, (2.27) sandwiched by the weights λα and tα−1. Recall that the operator R defined by (2.4) is a one-to-one mapping of Z+ onto We need the following result. Z. 12 D. R. YAFAEV Lemma 2.8. For all α > 0, the identity Lα = R−1J Φ∗ΓΓΓαΦR (2.28) holds. In particular, Lα as well as its inverse are the one-to-one continuous mappings of Z+ onto itself. Proof. Putting Ω(y) = (RF )(y) = F (ey) in (2.26) and making the change of vari- ables t = ey, we find that (Φ∗ΓΓΓαΦRF )(x) = e−αxZ ∞ 0 e−e−xttα−1F (t)dt. Now making the change of variables λ = e−x, we arrive at the identity (2.28). Consider the right-hand side of (2.28). All mappings R : Z+ → Z, Φ : Z → C ∞ 0 (R), ΓΓΓα : C ∞ 0 (R) → Z, J : Z → Z are bijections. All of them as well as their inverses are continuous. Therefore the identity (2.28) ensures the same result for the operator Lα : Z+ → Z+. (cid:3) 0 (R) → C ∞ 0 (R), Φ∗ : C ∞ To recover h(t), we proceed from formula (2.19). Passing to the Fourier trans- forms, we can write it as where G is defined by formulas (2.4), (2.15), that is, G = (2π)1/2ΓΓΓ1ΦRF. Therefore using the identity (2.28) for α = 1, we see that hhhh, Fiii = (2π)−1/2hs, Φ∗Gi hhhh, Fiii = hs,J RL1Fi =Z ∞ −∞ s(x)(L1F )(e−x)dx. Making the change of variables λ = e−x in the right-hand side, we obtain the identity Passing here to adjoint operators and using that F ∈ Z+ is arbitrary, we find that (2.29) h = L∗ 1h♮, hhhh, Fiii = hhhh♮, L1Fiii. which gives the precise sense to formula (1.10). Of course formula (1.10) can also be rewritten as h(t) =Z ∞ −∞ e−te−x e−xs(x)dx. Let us state the result obtained. +, and let s ∈ Z ′ be the corresponding sign-functions (see Theorem 2.9. Let h ∈ Z ′ Definition 2.4). Define the distribution h♮ by formula (1.11). Then h♮(λ) belongs to the class Z ′ + and h can be recovered from h♮ by formula (2.29). We emphasize that in the roundabout h 7→ h♮ 7→ h the mappings h 7→ h♮ as well + onto itself. Let us also give a direct expression of u(x) = (Φ∗g)(x) in terms of f (t). as its inverse h♮ 7→ h are one-to-one continuous mappings of the set Z ′ CRITERIA FOR HANKEL OPERATORS TO BE SIGN-DEFINITE 13 Lemma 2.10. Suppose that f ∈ D and put ϕ(t) = t1/2f (t). Let g(ξ) be defined by formula (1.5) and u(x) = (Φ∗g)(x). Then u(x) = (L1/2ϕ)(e−x). (2.30) Proof. Since (Rϕ)(x) = (Uf )(x), it follows from formula (2.28) for α = 1/2 that (R−1J Φ∗ΓΓΓ1/2ΦUf )(λ) = (L1/2ϕ)(λ). The left-hand side here equals (R−1J u)(λ) which after the change of variables λ = e−x yields (2.30). (cid:3) Now we can rewrite identity (2.21) in a slightly different way. Corollary 2.11. Let h ∈ Z ′ (1.11). Then for arbitrary fj ∈ D, j = 1, 2, and ϕj(t) = t1/2fj(t), we have +, and let the distribution h♮ ∈ Z ′ + be defined by formula (2.31) Proof. It suffices to make the change of variables x = − ln λ in the right-hand side of (2.22) and to take equality (2.30) into account. hhhh, ¯f1 ⋆ f2iii = hhhh♮, L1/2ϕ1L1/2ϕ2iii. (cid:3) We emphasize that according to Lemma 2.8, L1/2ϕj ∈ Z+ and hence L1/2ϕ1L1/2ϕ2 ∈ Z+. Thus the right-hand side of (2.31) is correctly defined. + is determined uniquely by the values hhhh, ¯f1 ⋆ f2iii on f1, f2 ∈ D. First we consider convolution operators. Let us introduce the shift in the space L2(R): 2.4. Finally, we check that a distribution h ∈ Z ′ Since (T (τ )g)(ξ) = g(ξ − τ ), τ ∈ R. (g1 ∗ g2)(ξ) =Z ∞ −∞ (T (τ )g1)(ξ)g2(τ )dτ, ∀g1, g2 ∈ C ∞ 0 (R), we have the formula hb, (J ¯g1) ∗ g2i =Z ∞ −∞hb, T (τ )J ¯g1ig2(τ )dτ (2.32) (2.33) where for b ∈ C ∞ 0 (R)′ the function hb, T (τ )J ¯g1i is infinitely differentiable in τ ∈ R. The following assertion is quite standard. 0 (R)′. Suppose that hb, (J ¯g1) ∗ g2i = 0 for all g1, g2 ∈ Lemma 2.12. Let b ∈ C ∞ C ∞ 0 (R). Then b = 0. Proof. If hb, (J ¯g1) ∗ g2i = 0 for all g2 ∈ C ∞ 0 (R), then hb, T (τ )J ¯g1i = 0 for all τ ∈ R according to formula (2.33). In particular, for τ = 0 we have hb,J ¯g1i = 0 whence b = 0 because g1 ∈ C ∞ 0 (R) is arbitrary. (cid:3) Next we pass to Hankel operators. 14 D. R. YAFAEV +. Suppose that hhhh, ¯f1 ⋆ f2iii = 0 for all f1, f2 ∈ D. Proposition 2.13. Let h ∈ Z ′ Then h = 0. Proof. Let b ∈ C ∞ 0 (R)′ be the b-function of h (see Definition 2.4). For arbitrary 0 (R), we can construct f1, f2 ∈ D by formula (2.12). Since hhhh, ¯f1 ⋆f2iii = 0, g1, g2 ∈ C ∞ it follows from the identity (2.20) that hb, (J ¯g1)∗ g2i = 0. Therefore b = 0 according to Lemma 2.12. Now Proposition 2.5 implies that h = 0. (cid:3) 3. Bounded Hankel operators Our main goal here is to show that the condition h ∈ Z ′ bounded Hankel operators H. + is satisfied for all 3.1. In this section we a priori only assume that h ∈ C ∞ Hankel form (2.3) on functions f1, f2 ∈ C ∞ restriction of the shift (2.32) on its invariant subspace L2(R+). Since 0 (R+)′ and consider the 0 (R+). Let T+(τ ) where τ ≥ 0 be the ( ¯f1 ⋆ f2)(t) =Z ∞ 0 (T+(τ ) ¯f1)(t)f2(τ )dτ, ∀f1, f2 ∈ C ∞ 0 (R+), for all h ∈ C ∞ 0 (R+)′ we have the formula hhhh, ¯f1 ⋆ f2iii =Z ∞ 0 hhhh, T+(τ ) ¯f1iiif2(τ )dτ. (3.1) Here the function hhhh, T+(τ ) ¯f1iii is infinitely differentiable in τ ∈ R+, and this function, as well as all its derivatives, have finite limits as τ → 0. In the theory of Hankel operators, formula (3.1) plays the role of formula (2.33) for convolution operators. The proof of the following assertion is almost the same as that of Lemma 2.12. 0 (R+)′. Suppose that hhhh, ¯f1 ⋆ f2iii = 0 for all f1, f2 ∈ Proposition 3.1. Let h ∈ C ∞ C ∞ 0 (R+). Then h = 0. 0 (R+), then hhhh, T+(τ ) ¯f1iii = 0 for all τ ∈ [0,∞) Proof. If hhhh, ¯f1 ⋆f2iii = 0 for all f2 ∈ C ∞ according to formula (3.1). In particular, for τ = 0 we have hhhh, ¯f1iii = 0 which implies that h = 0 because f1 ∈ C ∞ 0 (R+) is arbitrary. (cid:3) Of course Propositions 2.13 and 3.1 differ only by the set of functions on which the Hankel form is considered. Assume now that Then there exists a bounded operator H such that hhhh, ¯f ⋆ fiii ≤ Ckfk2, ∀f ∈ C ∞ 0 (R+). (Hf1, f2) = hhhh, ¯f1 ⋆ f2iii, ∀f1, f2 ∈ C ∞ 0 (R+). We call H the Hankel operator associated to the Hankel form hhhh, ¯f1 ⋆ f2iii. (3.2) (3.3) CRITERIA FOR HANKEL OPERATORS TO BE SIGN-DEFINITE 15 3.2. It is possible to characterize Hankel operators by some commutation rela- tions. Let us define a bounded operator Σ in the space L2(R+) by the equality Note that esf (s)ds. (Σf )(t) = −2e−tZ t Σ = −2Z ∞ 0 T+(τ )e−τ dτ. 0 (3.4) Lemma 3.2. Let assumption (3.2) hold. Then the corresponding Hankel operator H satisfies the commutation relations: HT+(τ ) = T+(τ )∗H, ∀τ ≥ 0, HΣ = Σ∗H. (3.5) (3.6) and Proof. Since (T+(τ ) ¯f1) ⋆ f2 = ¯f1 ⋆ (T+(τ )f2), ∀τ ≥ 0, relation (3.5) directly follows from definition (3.3). By virtue of formula (3.4), relation (3.6) is a consequence of (3.5). (cid:3) Below we need the Nehari theorem; see the original paper [9], or the books [10], +(R) Chapter 1, §1 or [11], Chapter 1, §2. We formulate it in the Hardy space H2 of functions analytic in the upper half-plane. We denote by bΣ the operator of multiplication by the function (µ − i)/(µ + i) in this space. Theorem 3.3 (Nehari). Let ω ∈ L∞(R), and let a bounded operator bH in the space H2 +(R) be defined by the relation (bH f1, f2) =Z ∞ −∞ ω(µ) f1(−µ) f2(µ)dµ, ∀ f1, f2 ∈ H2 Then bHbΣ = bΣ∗bH. Conversely, if bH is a bounded operator in H2 then there exists a function ω ∈ L∞(R) such that representation (3.7) holds. +(R) and bHbΣ = bΣ∗bH, The following assertion can be regarded as a translation of this theorem into the +(R) → L2(R+) is the unitary space L2(R+). Recall that the Fourier transform Φ : H2 operator. Since +(R). (3.7) Z ∞ −∞ µ − i µ + i e−iµtdµ = −4πe−t for t > 0 and this integral is zero for t < 0, we have the relation bΣ = Φ∗ΣΦ. (3.8) 16 D. R. YAFAEV Theorem 3.4. If representation (3.3) holds with h = Φω where ω ∈ L∞(R) (in this case h ∈ S ′ ⊂ C ∞ 0 (R+)′), then estimate (3.2) is true and the corresponding Hankel operator satisfies commutation relation (3.5). Conversely, if a bounded operator H satisfies (3.5), then representation (3.3) holds with h = Φω for some ω ∈ L∞(R). Proof. Since (Φ∗(f1 ⋆ ¯f2))(µ) = (J f1)(µ) f2(µ), ∀f1, f2 ∈ C ∞ 0 (R+), where f1 = Φ∗f1, f2 = Φ∗f2 and (J f1)(µ) = f1(−µ), we have ∀h ∈ S ′. hhhh, ¯f1 ⋆ f2iii = hhhΦ∗h, (J f1) f2iii, (3.9) Therefore estimate (3.2) is satisfied if Φ∗h ∈ L∞(R). Relation (3.5) for the corre- sponding Hankel operator H follows from Lemma 3.2. Conversely, if a bounded operator H satisfies relation (3.5), then by virtue of (3.4) it also satisfies relation (3.6). Hence it follows from (3.8) that bHbΣ = bΣ∗bH where bH = Φ∗HΦ is a bounded operator in the space H2 exists a function ω ∈ L∞(R) such that representation (3.7) holds. It means that ω(µ) f1(−µ) f2(µ)dµ, (Hf1, f2) =Z ∞ +(R). Thus, by Theorem 3.3, there ∀f1, f2 ∈ L2(R+). (3.10) −∞ If h = Φω, then the right-hand sides in (3.9) and (3.10) coincide. This yields representation (3.3). (cid:3) Corollary 3.5. For a bounded operator H, commutation relations (3.5) and (3.6) are equivalent. Proof. As was already noted, (3.6) follows from (3.5) according to formula (3.4). Conversely, if H satisfies (3.5), then representation (3.3) holds according to Theo- rem 3.4. Thus it remains to use Lemma 3.2. (cid:3) A function ω ∈ L∞(R) such that Φω = h is called the symbol of a bounded Hankel operator H with kernel h(t). Of course if ω ∈ H∞ − (R), that is, ω admits an analytic continuation to a bounded function in the lower half-plane, then the corresponding Hankel operator is zero. Therefore the symbol is defined up to a function in the class H∞ − (R). 3.3. Now we are in a position to check that the condition h ∈ Z ′ + is satisfied for all bounded Hankel operators. By definition (2.7) it means that distribution (2.5) belongs to the class Z ′. We shall verify the stronger inclusion θ ∈ S ′. To that end, it suffices to check that, for some N ∈ Z+ and some κ ∈ R, hθ, Ωi ≤ C NXn=0 max x∈R (cid:0)(1 + x)κΩ(n)(x)(cid:1), ∀Ω ∈ C ∞ 0 (R). (3.11) CRITERIA FOR HANKEL OPERATORS TO BE SIGN-DEFINITE 17 Putting F (t) = Ω(ln t), we see that (3.11) is equivalent to the estimate hh, Fi ≤ C NXn=0 max t∈R+(cid:0)(1 + ln t)κtnF (n)(t)(cid:1), F ∈ C ∞ 0 (R+). (3.12) If h ∈ L1 loc(R), then estimate (3.12) for N = 0 is equivalent to the convergence of integral (1.8). If H is Hilbert-Schmidt, that is Z ∞ 0 h(t)2tdt < ∞, then integral (1.8) converges for any κ > 1/2. Similarly, if h(t) ≤ Ct−1, then integral (1.8) converges for any κ > 1. For the proof of (3.12) in the general case, we use the following elementary result. Its proof is given in Appendix A. Lemma 3.6. If F ∈ C ∞ 0 (R+), then for an arbitrary κ > 2 the estimate kΦ∗FkL1(R) ≤ C 2Xn=0 max t∈R+(cid:0)(1 + ln t)κtnF (n)(t)(cid:1) holds. Corollary 3.7. If h = Φω where ω ∈ L∞(R), then estimate (3.12) holds for N = 2 and an arbitrary κ > 2. Since, by Theorem 3.4, for a bounded Hankel operator H, its kernel h = Φω for some ω ∈ L∞(R), we arrive at the following result. 0 (R+)′ and that condition (3.2) is satisfied. Theorem 3.8. Suppose that h ∈ C ∞ Then estimate (3.12) holds for N = 2 and an arbitrary κ > 2; in particular, h ∈ Z ′ +. The following simple example shows that for N = 0 estimate (3.12) is not in general true (for all κ). Example 3.9. Let h(t) = e−it2. Then the corresponding Hankel operator H is bounded because according to the formula e−i(t+s)2 = e−it2e−i2tse−is2, it is a product of three bounded operators. However integral (1.8) diverges at infinity for all κ. In this example condition (3.12) is satisfied for N = 1 and κ = 0. Indeed, integrating by parts, we see that Z ∞ h(t)F (t)dt = −Z ∞ where the function h1(t) = R t 0 e−is2ds is bounded and h1(t) = O(t) as t → 0. Therefore integral (3.13) is bounded by maxt∈R+(cid:0)(1 + ln t)κtF ′(t)(cid:1) for κ > 1. h1(t)F ′(t)dt (3.13) 0 0 18 D. R. YAFAEV Note that the symbol of H equals ω(µ) = (2π)−1e−πi/4eiµ2/4. More generally, one can consider the class of symbols ω(µ) such that ω ∈ C ∞(R), ω(µ) = eiω0µα, ω0 > 0, for large positive µ and ω(µ) = 0 for large negative µ. Of course Hankel operators with such symbols are bounded. Using the stationary phase method, we find that for α > 1 the corresponding kernel h(t) has the asymptotics h(t) ∼ h0tβeiσtγ , (3.14) where β = (1 − α/2)(α − 1)−1, γ = α(α − 1)−1 and h0, σ = ¯σ are some numbers. Moreover, h(t) is a bounded function on all finite intervals. Similarly to Example 3.9, it can be checked that for such kernels condition (3.12) is satisfied for N = 1 but not for N = 0. If α ∈ (0, 1), then h(t) has asymptotics (3.14) for t → 0. t → ∞, 3.4. Here we shall show that, for bounded Hankel operators H, the represen- tations (2.20) and (2.21) extend to all f1, f2 ∈ L2(R+). By Theorem 3.8, we have h ∈ Z ′ +. Let b and s be the corresponding b- and s-functions (see Definition 2.4). Recall that the operator Ξ : L2(R+) → L2(R) was defined by formula (1.5). We de- note by K the operator of multiplication by the functionpcosh(πξ)/π in the space L2(R). It follows from identity (2.24) and the unitarity of the Mellin transform (2.9) that and hence the operator KΞ : L2(R+) → L2(R) is unitary. Therefore in view of the identities (2.20) and (2.21) we have the following result. kKΞfk = kfk, Lemma 3.10. The inequalities (3.2), and hb, (J ¯g) ∗ gi ≤ CkKgk2, ∀g ∈ C ∞ 0 (R), (3.15) (3.16) are equivalent. The Hankel operator corresponding to form (2.3) is bounded if and only if one of equivalent estimates (3.2), (3.15) or (3.16) is satisfied. hs,u2i ≤ CkKΦuk2, ∀u ∈ Z, These estimates can be formulated in a slightly different way. Let us introduce the space E ⊂ L2(R) of exponentially decaying functions with the norm kgkE = kKgk. Then the space W = Φ∗E consists of functions u(x) admitting the analytic continuation u(z) in the strip Im z ∈ (−π/2, π/2); moreover, functions u(x + iy) have limits in L2(R) as y → ±π/2. The identity defines the Hilbert norm on W. We call W the exponential Sobolev space because it is contained in standard Sobolev spaces Hl(R) for all l. The operators Ξ : L2(R+) → E and bΞ := Φ∗Ξ : L2(R+) → W are of course unitary. Obviously, kKgk and kKΦuk in the right-hand sides of (3.15) and (3.16) can be replaced by kgkE and kukW, E = (2π)−1Z ∞ kΦuk2 −∞(cid:0)u(x + iπ/2)2 + u(x − iπ/2)2(cid:1)dx =: kuk2 W CRITERIA FOR HANKEL OPERATORS TO BE SIGN-DEFINITE 19 are equivalent. respectively. Note that the inclusions f ∈ L2(R+), g = Ξf ∈ E and u = bΞf ∈ W If one of the equivalent estimates (3.2), (3.15) or (3.16) is satisfied, then all oper- ators H : L2(R) → L2(R), B : E → E ′ and S : W → W ′ are bounded. Using that relations fn → f in L2(R+), gn = Ξfn → g = Ξf in E and un = Φ∗gn → u = Φ∗g in W are equivalent, we extend (2.20) and (2.21) to all f ∈ L2(R+). Thus we have obtained the following result. Proposition 3.11. If one of equivalent estimates (3.2), (3.15) or (3.16) is satisfied, then the identities (Hf1, f2) = (Bg1, g2) = (Su1, u2), gj = Ξfj, uj = Φ∗gj, are true for all f1, f2 ∈ L2(R+). Let Kl be the operator of multiplication by the function (1+ξ2)l/2. Then estimates (3.15) or (3.16) are satisfied provided hb, (J ¯g) ∗ gi ≤ ClkKlgk2 or hs,u2i ≤ Clkuk2 Hl(R), (3.17) for some l; in this case C = Clπ max ξ∈R (cid:0)(1 + ξ2)l(cosh(πξ))−1(cid:1). We note the following assertion. Proposition 3.12. A Hankel operator H is bounded if its sign-function satisfies the condition s ∈ L1(R) + L∞(R). (3.18) If s ∈ L∞(R) and s(x) → 0 as x → ∞, then H is compact. Proof. The first statement is obvious because under assumption (3.18) the second estimate (3.17) is satisfied for l > 1/2. To prove the second statement, we observe that the operator SΦ∗K −1 is compact because both S and K −1 are operators of multiplication by bounded functions which tends to zero at infinity. It follows that the operator H = Ξ∗Φ(SΦ∗K −1)(KΞ) is also compact. (cid:3) We emphasize, however, that as show already examples of Hankel operators H of finite rank (see formula (5.9) for h(t) = e−αt), condition (3.18) is not necessary for the boundedness of H. 20 D. R. YAFAEV 4. Criteria of the sign-definiteness In this section we suppose that h(t) = h(t) so that the operator H is symmetric. The results of Section 2 allow us to give simple necessary and sufficient conditions for a Hankel operator H to be positive or negative. Moreover, they provide also convenient tools for a evaluation of the total multiplicity of the negative and positive spectra of H. We often formulate results only for the negative spectrum. The corresponding results for the positive spectrum are obtained if H is replaced by −H. + and introduce the b-function b ∈ C ∞ 4.1. Actually, we consider the problem in terms of Hankel quadratic forms rather than Hankel operators. This is both more general and more convenient. As usual, 0 (R)′ we suppose that a distribution h ∈ Z ′ and the s-function s ∈ Z ′ by Definition 2.4. Below we use the following natural notation. Let h[ϕ, ϕ] be a real quadratic form defined on a linear set D. We denote by N±(h) the maximal dimension of linear sets M± ⊂ D such that ±h[ϕ, ϕ] > 0 for all ϕ ∈ M±, ϕ 6= 0. We apply this definition to the forms h[f, f ] = hhhh, ¯f ⋆ fiii defined on D, to b[g, g] = hb,J ¯g ∗ gi defined on C ∞ 0 (R) and to s[u, u] = hs,u2i defined on Z. Of course, if D is dense in a Hilbert space H and h[ϕ, ϕ] is closed on D, then for the self-adjoint operator H corresponding to h, we have N±(H) = N±(h) . Observe that formula (1.5) establishes one-to-one correspondence between the sets 0 (R). Moreover, the Fourier transform establishes one-to-one correspon- 0 (R) and Z. Therefore the following assertion is a direct D and C ∞ dence between the sets C ∞ consequence of Theorem 2.6. Theorem 4.1. Let h ∈ Z ′ +. Then N±(h) = N±(b) = N±(s). In particular, we have Theorem 4.2. Let h ∈ Z ′ only if the form ±hb,J ¯g ∗ gi ≥ 0 for all g ∈ C ∞ all u ∈ Z. +. Then the form ±hhhh, ¯f ⋆ fiii ≥ 0 for all f ∈ D if and 0 (R), or the form ±hs,u2i ≥ 0 for 4.2. In many cases the following consequence of Theorem 4.1 is convenient. According to Proposition 3.12, under the assumptions of Theorem 4.3, H is defined as the bounded self-adjoint operator corresponding to the form hhhh, ¯f ⋆ fiii. Therefore N±(h) = N±(H) is the total multiplicity of the (strictly) positive spectrum for the sign "+" and of the (strictly) negative spectrum for the sign "−" of the operator H. For definiteness, we consider the negative spectrum. Theorem 4.3. Let h ∈ Z ′ +, and let the corresponding sign-function satisfy condition (3.18). If s(x) ≥ 0, then the operator H is positive. If s(x) ≤ −s0 < 0 for almost all x in some interval ∆ ⊂ R, then the operator H has infinite negative spectrum. CRITERIA FOR HANKEL OPERATORS TO BE SIGN-DEFINITE 21 Choose a function ϕ ∈ C ∞ Proof. If s(x) ≥ 0, then H ≥ 0 according to the second relation (2.22). Let s(x) ≤ −s0 < 0 for x ∈ ∆. For an arbitrary N, we shall construct a linear set L ⊂ Z of dimension N such that s[u, u] < 0 for all u ∈ L, u 6= 0. Then the second statement will follow from Theorem 4.1. 0 (R) such that ϕ(x) = 1 for x ∈ [−δ, δ] and ϕ(x) = 0 for x 6∈ [−2δ, 2δ] where δ = δN is a sufficiently small number. Let points αj ∈ ∆, j = 1, . . . , N, be such that αj+1 − αj = αj − αj−1 for j = 2, . . . , N − 1. Set ∆j = (αj − δ, αj + δ), e∆j = (αj − 2δ, αj + 2δ). For a sufficiently small δ, we may suppose that e∆j ⊂ ∆ for all j = 1, . . . , N and that e∆j+1 ∩ e∆j = ∅ for j = 1, . . . , N − 1. We set ϕj(x) = ϕ(x − αj). Since s(x) ≤ −s0 < 0 for x ∈ ∆, we have hs,ϕj2i =Z ∞ −∞ s(x)ϕj(x)2dx ≤ −2δs0 < 0. (4.1) The problem is that ϕj ∈ C ∞ The functions ϕ1, . . . , ϕN have disjoint supports and hence hs,u2i < 0 for an arbi- trary non-trivial linear combination u of the functions ϕj. 0 (R) but ϕj 6∈ Z (and even ϕj 6∈ W). Thus we have j ∈ Z. For a to approximate these functions in the topology of S by functions ϕ(ε) given κ > 1/2 and an arbitrary ε > 0, we can find a function ϕ(ε) ∈ Z such that max x∈R (cid:0)(1 + x)κϕ(x) − ϕ(ε)(x)(cid:1) < ε. (4.2) Now we set ϕ(ε) j (x) = ϕ(ε)(x − αj). independent. Assume that Let us check that for ε small enough, the functions ϕ(ε) 1 , . . . , ϕ(ε) N are linearly NXj=1 λjϕ(ε) j (x) = 0. (4.3) If x ∈ ∆k, then ϕ(ε) the estimate k (x) − 1 < ε and ϕ(ε) j (x) < ε for j 6= k. Therefore (4.3) yields (1 − ε)λk ≤ εXj6=k λj. Summing these estimates over k = 1, . . . , N, we see that NXj=1 Hence λj = 0 for all j = 1, . . . , N if ε(N − 1) < 1 − ε. It follows from (4.1), (4.2) that, for all j = 1, . . . , N, λj ≤ ε(N − 1) (1 − ε) NXj=1 λj. Z ∞ −∞ s(x)ϕ(ε) j (x)2dx ≤ −2δs0 + Cε 22 and that D. R. YAFAEV Thus we have the estimate j 6= k. −∞ (cid:12)(cid:12)λj ¯λkZ ∞ Z ∞ −∞ s(x)ϕ(ε) j (x)ϕ(ε) k (x)dx(cid:12)(cid:12) ≤ Cε(λj2 + λk2), j (x)2dx ≤ − NXj=1 NXj=1 λjϕ(ε) s(x) λj2(2δs0 − Cε). The right-hand side here is negative if ε is small enough and PN Let L be spanned by the functions ϕ(ε) N for sufficiently small ε. Then dimL = N and s[u, u] < 0 for all u ∈ L, u 6= 0 according to (4.4). It follows from Theorem 4.1 that N−(H) ≥ N, and hence N−(H) = ∞. j=1 λj2 6= 0. 1 , . . . , ϕ(ε) (4.4) (cid:3) Theorem 4.3 can be reformulated, although in a weaker form, in terms of the functions b(ξ) and even h(t). Suppose, for example, that (4.5) Then its Fourier transform s(x) is a continuous function which tends to 0 as x → ∞. The operator B defined by formula (1.2) is bounded, self-adjoint and b ∈ L1(R). spec(B) = [min x∈R s(x), max x∈R s(x)]. The result below follows directly from Theorem 4.3. Note that by Proposition 3.12 under assumption (4.5) the operator H is compact. Proposition 4.4. Under assumption (4.5) the Hankel operator H is positive if and only if s(x) ≥ 0. If minx∈R s(x) < 0, then necessarily H has an infinite negative spectrum. In particular, condition (4.5) is satisfied if In this case a = Φθ ∈ C ∞ θ(ln t) h(t) = where 0 (R) and hence b ∈ C ∞ t 0 (R). θ ∈ Z. 4.3. For a proof that a Hankel operator is not sign-definite it is sometimes even not necessary to calculate the sign-function s(x) (the Fourier transform of b(ξ)). It turns out that if b(ξ) grows as ξ → ∞, then the form b[g, g] = hb,J ¯g∗ gi cannot be sign-definite. More precisely, we have the following statement about convolutions with growing kernels b(−ξ) = b(ξ). Theorem 4.5. Let b = b0+b∞ where b0 ∈ C p(R)′ for some p ∈ Z+ and b∞ ∈ L∞ loc(R). Suppose that there exists a sequence of intervals ∆n = (rn − σn, rn + σn) where rn → ∞ (or equivalently rn → −∞) and the sequence σn is bounded such that lim n→∞ σl n min ξ∈∆n Re b∞(ξ) = ∞ or lim n→∞ σl n max ξ∈∆n Re b∞(ξ) = −∞, (4.6) CRITERIA FOR HANKEL OPERATORS TO BE SIGN-DEFINITE 23 where l = 2 if p = 0 or p = 1 and l = p + 1 if p ≥ 2. Then for both signs N±(b) ≥ 1. Proof. Since b can be replaced by −b, we can assume that, for example, the first condition (4.6) is satisfied. Pick a real even function ϕ ∈ C ∞ 0 (R) such that ϕ(ξ) ≥ 0, ϕ(ξ) = 1 for ξ ≤ 1/4, ϕ(ξ) = 0 for ξ ≥ 1/2 and set gn(ξ) = ϕ((ξ − rn/2)/σn) ± ϕ((ξ + rn/2)/σn). (4.7) An easy calculation shows that ((J gn) ∗ gn)(ξ) = 2σnψ(ξ/σn) ± σnψ((ξ − rn)/σn) ± σnψ((ξ + rn)/σn) (4.8) 0 (R). The function ψ(ξ) is also even, ψ(ξ) ≥ 0, ψ(ξ) ≥ 1/8 where ψ = (J ϕ)∗ ϕ ∈ C ∞ for ξ ≤ 1/8 and ψ(ξ) = 0 for ξ ≥ 1. Since hb0, gi ≤ CkgkC p, it follows from (4.8) that hb0, (J gn) ∗ gni ≤ Cσ1−p nZ ∞ hb∞, (J gn) ∗ gni = 2σ2 b∞(σnη)ψ(η)dη ± 2σ2 Moreover, again according to (4.8) we have nZ ∞ n . −∞ −∞ (4.9) Re b∞(σnη + rn)ψ(η)dη. The first term in the right-hand side is O(σ2 estimate (4.10) n). For the second one, we use the 32Z ∞ −∞ Re b∞(σnη + rn)ψ(η)dη ≥ min ξ−rn≤σn Re b∞(ξ). (4.11) Let us first choose the sign " + " in (4.7). Then using representation (4.10) and putting together estimates (4.9) and (4.11), we obtain the lower bound hb, (J gn) ∗ gni ≥ −c(σ1−p n + σ2 n) + σ2 n/16 min Re b∞(ξ). ξ−rn≤σn If p = 0 or p = 1, then under the first condition (4.6) the right-hand side here tends to +∞ as n → ∞. If p ≥ 2, it is bounded from below by Re b∞(ξ)(cid:1) n (cid:0) − c + σl n/16 min ξ−rn≤σn σ1−p where the expression in the brackets tends again to +∞ . Therefore hb, (J gn)∗gni > 0 for sufficiently large n. Similarly choosing the sign " − " in (4.7), we see that hb, (J gn) ∗ gni < 0 for sufficiently large n. Corollary 4.6. Instead of condition (4.6) assume that (cid:3) lim ξ→∞ Re b∞(ξ) = ∞ or lim ξ→∞ Re b∞(ξ) = −∞. Then for both signs N±(b) ≥ 1. 24 D. R. YAFAEV In contrast to Theorem 4.5 there are no restrictions in Corollary 4.6 on the pa- rameter p in the assumption b0 ∈ C p(R)′. On the other hand, condition (4.6) permits Re b(ξ) to tend to ±∞ only on some system of intervals. Moreover, the lengths of these intervals may tend to zero. In this case, however, the growth of Re b(ξ) and the decay of these lengths should be correlated and there are restrictions on admissible values of the parameters p and l. Unlike Theorem 4.3, Theorem 4.5 does not guarantee that N = ∞; see subs. 5.4, for a discussion of various possible cases. 4.4. Theorem 4.2 can be combined with the Bochner-Schwartz theorem (see, e.g., Theorem 3 in §3 of Chapter II of the book [5]). It states that a distribution b ∈ C ∞ 0 (R) (such b are sometimes called distributions of positive type) is the Fourier transform 0 (R)′ satisfying the condition hb,J ¯g ∗ gi ≥ 0 for all g ∈ C ∞ b(ξ) = (2π)−1Z ∞ −∞ e−ixξdM(x) of a positive measure dM(x) such that Z ∞ −∞ (1 + x)−κdM(x) < ∞ (4.12) for some κ (that is, of at most polynomial growth at infinity). In particular, this ensures that b ∈ S ′. distribution b is of positive type. determined by the measure dM(x): Theorem 4.2 implies that if hhhh, ¯f ⋆ fiii ≥ 0 for all f ∈ D, then the corresponding It means that the sign-function s(x) of h(t) is hs, ϕi =Z ∞ −∞ ϕ(x)dM(x), ϕ ∈ S, that is, s(x)dx = dM(x). Let us define the measure dm(λ) = λdM(− ln λ), λ ∈ R+. (4.13) It is easy to see that condition (4.12) is equivalent to condition (1.13) on measure In terms of distribution (1.11), we have λh♮(λ)dλ = dm(λ). Therefore (4.13). Theorem 2.9 leads to the following result. + and hhhh, ¯f ⋆ fiii ≥ 0 for all f ∈ D. Then h(t) admits the Theorem 4.7. Let h ∈ Z ′ representation (1.12) with a positive measure dm(λ), λ ∈ R+, satisfying for some κ condition (1.13). The representation (1.12) is of course a particular case of (1.10). It is much more precise than (1.10) but requires the positivity of hhhh, ¯f ⋆ fiii. Theorem 4.7 shows that the positivity of hhhh, ¯f ⋆ fiii imposes very strong conditions on h(t). Actually, we have CRITERIA FOR HANKEL OPERATORS TO BE SIGN-DEFINITE 25 Corollary 4.8. Let h ∈ Z ′ and + and hhhh, ¯f ⋆ fiii ≥ 0 for all f ∈ D. Then h ∈ C ∞(R+) (4.14) (−1)nh(n)(t) ≥ 0 for all t > 0 and all n = 0, 1, 2, . . . (such functions are called completely monotonic). The function h(t) admits an analytic continuation in the right-half plane Re t > 0 and it is uniformly bounded in every strip Re t ∈ (t1, t2) where 0 < t1 < t2 < ∞. Moreover, for some κ ∈ R and C > 0 the estimate holds: t > 0. (4.15) h(t) ≤ Ct−1(1 + ln t)κ, All these assertions are direct consequences of the representation (1.12). In par- ticular, under condition (1.13) we have h(t) ≤ C max λ≥0 (cid:0)e−tλλ(1 + ln λ)κ(cid:1) which yields (4.15). Note that according to the Bernstein theorem (see, e.g., Theorems 5.5.1 and 5.5.2 in [1]) condition (4.14) implies that the function h(t) admits the representation (1.12) with some measure dm(λ). Of course, condition (4.14) does not impose any restrictions on the measure dm(λ) (except that the integral (1.12) is convergent for all t > 0). In contrast to the Bernstein theorem we deduce the representation (1.12) from the positivity of the Hankel form. In this context condition (1.13) is due to the assumption h ∈ Z ′ +. We mention also a related result of H. Widom. He considered in [14] Hankel operators H with kernels h(t) admitting the representation (1.12) and showed that H is bounded if and only if m([0, λ)) = O(λ) as λ → 0 and as λ → ∞. In this case h(t) ≤ Ct−1 for some C > 0. Thus Theorem 4.7 and Corollary 4.8 can be regarded as an extension of Widoms's results to unbounded operators. Under the positivity assumption the identity (2.21) takes a more precise form. Proposition 4.9. Let h ∈ Z ′ positive measure dM(x) satisfying condition (4.12) for some κ such that + and hhhh, ¯f ⋆ fiii ≥ 0 for all f ∈ D. Then there exists a hhhh, ¯f1 ⋆ f2iii =Z ∞ −∞ u1(x)u2(x)dM(x) for all fj ∈ D, j = 1, 2, and uj = Φ∗Ξfj ∈ Z where the mapping Ξ is defined by (1.5). 5. Applications and examples ******************************************************** 5.1. Consider first self-adjoint Hankel operators H of finite rank. Recall that integral kernels of Hankel operators of finite rank are given (this is the Kronecker 26 D. R. YAFAEV theorem -- see, e.g., Sections 1.3 and 1.8 of the book [10]) by the formula h(t) = MXm=1 Pm(t)e−αmt (5.1) where Re αm > 0 and Pm(t) are polynomials of degree Km. If H is self-adjoint, that is, h(t) = h(t), then the set {α1, . . . , αM} consists of points lying on the real axis and pairs of points symmetric with respect to it. Let Im αm = 0 for m = 1, . . . , M0 and Im αm > 0, αM1+m = ¯αm for m = M0 + 1, . . . , M0 + M1. Thus M = M0 + 2M1; of course the cases M0 = 0 or M1 = 0 are not excluded. The condition h(t) = h(t) also requires that Pm(t) = Pm(t) for m = 1, . . . , M0 and PM1+m(t) = Pm(t) for m = M0 + 1, . . . , M0 + M1. As is well known and as we shall see below, rank H = MXm=1 Km + M =: r. For m = 1, . . . , M0, we denote by pm = ¯pm the coefficient at tKm in the polynomial Pm(t). The following assertion yields an explicit formula for the numbers N±(H). Its proof will be given in [18]. Theorem 5.1. For m = 1, . . . , M0, set − = (Km + 1)/2 if Km is odd N (m) + = N (m) + − 1 = N (m) N (m) N (m) + = N (m) − = Km/2 − − 1 = Km/2 if Km is even and pm > 0 (5.2) if Km is even and pm < 0.   Then the total numbers N±(H) of (strictly) positive and negative eigenvalues of the operator H are given by the formula N±(H) = M0Xm=1 N (m) ± + M0+M1Xm=M0+1 Km + M1. (5.3) Formula (5.2) shows that every pair Pm(t)e−αmt + Pm+M1(t)e−αm+M1 t, m = M0 + 1, . . . , M0 + M1, (5.4) of complex conjugate terms in (5.1) yields Km + 1 positive and Km + 1 negative eigenvalues. The contribution of every real term Pm(t)e−αmt where m = 1, . . . , M0 also consists of the equal numbers (Km + 1)/2 of positive and negative eigenvalues if the degree Km of the polynomial Pm(t) is odd. If Km is even, then there is one more positive (negative) eigenvalue if pm > 0 (pm < 0). In particular, in the question considered, there is no "interference" between different terms Pm(t)e−αmt, m = 1, . . . , M0, and pairs (5.4) in representation (5.1) of the kernel h(t). CRITERIA FOR HANKEL OPERATORS TO BE SIGN-DEFINITE 27 According to (5.3) the operator H cannot be sign-definite if M1 > 0. Moreover, ± = 0 for m = 1, . . . , M0 if and only if Km = 0 and ∓pm > 0. according to (5.2), N (m) Therefore we have the following result. Corollary 5.2. A Hankel operator H of finite rank is positive (negative) if and only if its kernel is given by the formula h(t) = M0Xm=1 pme−αmt where αm > 0 and pm > 0 (pm < 0). Corollary 5.2 admits different proofs which avoid formula (5.3). For example, one can use that although the functions Pm(t)e−αmt are analytic in the right-half plane Re t > 0, they are bounded for t = τ + iσ as σ → ∞ for a constant Pm(t) only. Therefore according to Corollary 4.8 such Hankel operators cannot be posi- tive. Alternatively, using formula (5.16) below, one can deduce Corollary 5.2 from Theorem 4.5. 5.2. Consider now Hankel operators H with kernels (1.14). Since the case k = 0, 1, . . . (finite rank Hankel operators) has been discussed in the previous subsection, here we suppose that k 6= 0, 1, . . .. If k > −1, condition (1.8) is satisfied for all κ, and the operators H are compact (actually, they belong to much better classes of operators). If k = −1, then condition (1.8) is satisfied for κ > 1, and the operators H are bounded but not compact. Let us calculate the corresponding b- and s-functions. If k > −1, then function (2.16) equals a(ξ) = (2π)−1/2Z ∞ 0 and hence function (2.18) equals tke−αtt−iξdt = (2π)−1/2α−1−k+iξΓ(1 + k − iξ), (5.5) b(ξ) = α−1−k+iξ Γ(1 + k − iξ) 2πΓ(1 − iξ) . (5.6) If k = −1, then in accordance with formulas (5.5) and (5.6), we have a(ξ) = (2π)−1/2αiξ lim ε→+0 Γ(ε − iξ), b(ξ) = (2π)−1αiξi(ξ + i0)−1. It yields the expression s(x) = 0, x > β, (5.7) for the function s = √2πΦ∗b. Formula (5.7) remains true for the Carleman operator C (the Hankel operator with kernel h(t) = t−1) when α = 0. Indeed, in this case according to (2.23) the sign-function s(x) = 1. x < β, where β = − ln α, s(x) = 1, 28 D. R. YAFAEV Next, we calculate the Fourier transform of function (5.6). Assume first that k ∈ (−1, 0). Then (see, e.g., formula (1.5.12) in the book [4]) t−k−1(t + 1)−1+iξdt = Γ(−k)Γ(1 + k − iξ) . Γ(1 − iξ) Making here the change of variables t + 1 = α−1e−x, we find that 0 Z ∞ Γ(−k)Z ∞ 1 −∞ (e−x − α)−k−1 + e−ixξdx = α−1−k−iξ Γ(1 + k − iξ) Γ(1 − iξ) . Passing now to the inverse Fourier transform, we see that for k ∈ (−1, 0) the sign- function s(x) = sk(x) of kernel (1.14) equals s(x) = 1 Γ(−k) (e−x − α)−k−1 + . (5.8) Let us verify that this formula remains true for all non-integer k. To that end, we assume that (5.8) holds for some non-integer k > −1 and check it for k1 = k + 1. Since we have Γ(1 + k1 − iξ) = (k1 − iξ)Γ(1 + k − iξ), sk1(x) = α−1(k1 − ∂)sk(x). Substituting here formula (5.8) for sk(x) and differentiating this expression, we obtain formula (5.8) for sk1(x). This concludes the proof of relation (5.8) for all k ≥ −1. Lemma 5.3. Let h(t) be given by formula (1.14) where k 6∈ Z+. Then the sign- function is determined by relation (5.8). Actually, relation (5.8) remains true for k ∈ Z+ if one takes into account that the has poles at integer points. For example, for k = 0 we distribution (e−x − α)−k−1 have + s(x) = α−1δ(x − β). (5.9) Obviously, s(x) = 0 for x > β = − ln α. If k = −1, then s(x) = 1 for x < β. If k ∈ (−1, 0), then s(x) ≥ 0 and s ∈ L1(R). Therefore it follows from Theorem 4.3 that H ≥ 0. If k > 0, then distribution (5.8) does not have a definite sign. Therefore it can be deduced from Theorem 4.2 that the corresponding Hankel operator also is not sign-definite. Alternatively, for the proof of this result we can use Corollary 4.6. Formula (2.11) implies that function (5.6) has the asymptotics b(ξ) = (2π)−1α−1−k−iξ(−iξ)k(1 + O(ξ−1)), ξ → ∞. (5.10) CRITERIA FOR HANKEL OPERATORS TO BE SIGN-DEFINITE 29 Making the dilation transformation in (1.14), we can suppose that α = 1. Then we have Re b(ξ) = (2π)−1 cos(πk/2)ξk + O(ξk−1), (5.11) Since cos(πk/2) 6= 0 unless k is an integer odd number, this expression tends to ±∞ if ± cos(πk/2) > 0. Thus Corollary 4.6 for the case b0 = 0 ensures that the Hankel operator H is not sign-definite. ξ → +∞, Let us summarize the results obtained. Proposition 5.4. The Hankel operator with kernel (1.14) is positive for k ∈ [−1, 0], and it is not sign-definite for k > 0. Explicit formulas for the sign-functions can also be used to treat more complicated Hankel operators. For example, in view of (5.7) the following assertion directly follows from Theorem 4.3. Example 5.5. The Hankel operator with kernel h(t) = t−1(e−α1t − γe−α2t), is positive if and only if α2 ≥ α1 ≥ 0 and γ ≤ 1. γ ≥ 0, 5.3. In this subsection, we consider the Hankel operator H with kernel (1.15). Condition (1.8) is now fulfilled for all κ, and the operator H belongs of course to the Hilbert-Schmidt class (actually, to much better classes). Observe that a(ξ) = (2π)−1/2Z ∞ 0 e−tr t−iξdt = (2π)−1/2r−1Γ((1 − iξ)/r) and define, as usual, the function b(ξ) by formula (2.18) so that b(ξ) = (2πr)−1 Γ((1 − iξ)/r) Γ(1 − iξ) . (5.12) Consider first the case r > 1. It follows from the Stirling formula (2.11) that for all r > 1 the modulus of function (5.12) exponentially grows and the periods of its oscillations tend to zero only logarithmically as ξ → ∞. Therefore Theorem 4.5 implies that the Hankel operator with kernel (1.15) is not sign-definite. The Hankel operator H with kernel h(t) = e−t2 can also be treated (see Appen- dix B) in a completely different way which is perhaps also of some interest. This method shows that both positive and negative spectra of the operator H are infinite. If r = 1, then h(t) = e−t yields a positive Hankel operator of rank 1. Let us now consider the case r < 1. Then, again according to the Stirling formula (2.11), function (5.12) belongs to L1(R) so that its Fourier transform s(x) = (2πr)−1Z ∞ −∞ Γ((1 − iξ)/r) Γ(1 − iξ) eixξdξ =: Ir(x) (5.13) 30 D. R. YAFAEV is a continuous function which tends to 0 as x → ∞. Therefore by Proposi- tion 4.4the corresponding Hankel operator H ≥ 0 if and only if Ir(x) ≥ 0 for all x ∈ R. It turns out that Ir(x) ≥ 0. Surprisingly, we have not found a proof of this fact in the literature, but it follows from our results. Only for r = 1/2, integral (5.13) can be explicitly calculated. Indeed, according to formula (1.2.15) of [4] Γ(2(1 − iξ)) Γ(1 − iξ) = 21−2iξπ−1/2Γ(3/2 − iξ). Therefore it follows from formula (2.25) that I1/2(x) = 2−1π−1/2e3x/2e−ex/4 (5.14) which is of course positive. For an arbitrary r ∈ (0, 1), one can proceed from the Bernstein theorem on completely monotonic functions (see subs. 4.4). Observe that if ψ(t) = t−pe−tr , p ≥ 0, (5.15) then ψ′(t) = −pt−p−1e−tr − rt−p+r−1e−tr . Further differentiations of ψ(t) change the sign and yield sums of terms having the form (5.15). Thus the function h(t) = e−tr satisfies for all n condition (4.14) and hence admits the representation (1.12) with some positive measure dm(λ). It follows from (1.12) that (Hf, f ) =Z ∞ 0 (Lf )(λ)2dm(λ) ≥ 0, ∀f ∈ C ∞ 0 (R+), where L is the Laplace transform (2.27). Since the operator H is bounded, this implies that H ≥ 0. Thus we have obtained the following result. Proposition 5.6. The Hankel operator with kernel (1.15) is positive for r ∈ (0, 1], and it is not sign-definite for r > 1. Putting together this result with Theorem 4.3, we see that integral (5.13) is pos- itive for all r ∈ (0, 1). Our indirect proof of this fact looks curiously enough. 5.4. Let us now discuss convolution operators with growing kernels b(ξ). We emphasize that condition (4.6) does not guarantee that the numbers N±(b) are Indeed, consider the kernel h(t) = tke−αt where k is a positive integer. infinite. Formula (5.6) shows that for Im α = 0 the corresponding function b(ξ) (5.16) has a power asymptotics as ξ → ∞. According to Theorem 5.1 the positive and negative spectra of the Hankel operator H with the kernel h(t) are finite; for example, b(ξ) = (2π)−1α−1−k+iξ(1 − iξ)· · · (k − iξ) CRITERIA FOR HANKEL OPERATORS TO BE SIGN-DEFINITE 31 H has exactly (k + 1)/2 positive and negative eigenvalues if k is odd. Moreover, if Im α 6= 0, then in view of (5.16) the function b(ξ) exponentially grows as ξ → +∞ or ξ → −∞. Nevertheless the Hankel operator H with kernel h(t) = tk(e−αt + e− ¯αt) has exactly k + 1 positive and negative eigenvalues. On the other hand, for kernel (5.12) where r = 2 we have N±(b) = ∞. This follows from Theorem 4.1 because, by Proposition B.1, the Hankel operator with kernel h(t) = e−t2 has infinite number of positive and negative eigenvalues. A similar phenomenon occurs for Hankel operators with non-smooth kernels. This is discussed in the next section. 6. Hankel operators with non-smooth kernels 6.1. Let the symbol of the Hankel operator H be defined by the formula ω(µ) = if t0 ≤ 0, and hence H = 0 in this case. If t0 > 0, then the eit0µ. Evidently, ω ∈ H∞ integral kernel of H equals δ(t − t0) so that − (Hf )(t) = f (t0 − t). Condition (3.12) is now satisfied for N = 0 and κ = 0. The operator H admits an explicit spectral analysis. Indeed, observe first that (Hf )(t) = 0 for t > t0 and hence L2(t0,∞) ⊂ Ker H. Since H 2f = f for f ∈ L2(0, t0), the restriction of H on its invariant subspace L2(0, t0) may have only ±1 as eigenvalues. Obviously, the eigenspace H± of H corresponding to the eigenvalue ±1 consists of all functions f (t) such that f (t) = ±f (t0 − t). Since H+ ⊕ H− ⊕ L2(t0,∞) = L2(R+), the spectrum of H consists of the eigenvalues 0, 1,−1 of infinite multiplicity each. In this example, the b-function equals b(ξ) = t−iξ 0 2πΓ(1 − iξ) . (6.1) Note that all functions (5.12) where r > 1, (6.1) as well as (6.2) below exponentially grow and oscillate at infinity. In these cases the corresponding Hankel operators have infinite positive and negative spectra. 6.2. It follows from Corollary 4.8 that a Hankel operator H can be sign-definite only for kernels h ∈ C ∞(R+). Actually, if h(t) or one of its derivatives h(l)(t) has a jump discontinuity, then H has infinite number of both positive λ(+) and negative −λ(−) n , n = 1, 2, . . . , eigenvalues, and we can even calculate their asymptotics as n → ∞. Positive (negative) eigenvalues are of course enumerated in decreasing (increasing) order with multiplicities taken into account. n Let us start with an explicit 32 D. R. YAFAEV Example 6.1. Let h(t) = (t0 − t)l for some l = 0, 1, . . . if t ≤ t0 and h(t) = 0 if t > t0. Then (Hf )(t) =Z t0−t 0 (t0 − t − s)lf (s)ds, t ∈ (0, t0), and (Hf )(t) = 0 for t ≥ t0. For such h(t), the symbol equals (iµt0)k(cid:1) ω(µ) = l!(iµ)−l−1(cid:0)eiµt0 − lXk=0 1 k! and the b-function equals b(ξ) = l!tl+1−iξ 0 2πΓ(l + 2 − iξ) . Let us consider the spectral problem Hf = λf , that is, Z t0−t (t0 − t − s)lf (s)ds = λf (t), Differentiating this equation k times, we find that 0 t ∈ (0, t0). (−1)kl(l − 1)· · · (l − k + 1)Z t0−t 0 (t0 − t − s)l−kf (s)ds = λf (k)(t) (6.2) (6.3) (6.4) (6.5) (6.6) for k = 1, . . . , l. Differentiating the last equation where k = l once more, we see that Setting t = t0 in (6.3) and (6.4), we obtain the boundary conditions l!f (t) = λ(−1)l+1f (l+1)(t0 − t), t ∈ (0, t0). f (t0) = f ′(t0) = · · · = f (l)(t0) = 0. Conversely, if a function f (t) satisfies equation (6.5) and boundary conditions (6.6), it satisfies also equation (6.3). This leads to the following intermediary result. Lemma 6.2. Let the operator A be defined on the Sobolev class Hl+1(0, t0) by the equation (6.7) Considered with boundary conditions (6.6), it is self-adjoint in the space L2(0, t0), and its eigenvalues αn are linked to eigenvalues λn of the operator H by the equation αn = l!λ−1 n . (Af )(t) = (−1)l+1f (l+1)(t0 − t). 6.3. Clearly, A2 is a differential operator and the asymptotics of its eigenvalues is described by the Weyl formula. However, to find the asymptotics of eigenvalues of the operator A, we have to distinguish its positive and negative eigenvalues. For this reason, it is convenient to introduce an auxiliary operator eA with symmetric (with respect to the point 0) spectrum having the same asymptotics of eigenvalues as A. CRITERIA FOR HANKEL OPERATORS TO BE SIGN-DEFINITE 33 We define the operator eA by the same formula (6.7) as A but consider it on functions in Hl+1(0, t0/2) ⊕ Hl+1(t0/2, t0) satisfying the boundary conditions f (k)(0) = f (k)(t0/2 − 0), f (k)(t0/2 + 0) = f (k)(t0), where k = 0, . . . , l for l even and f (k)(0) = f (k)(t0/2 − 0) = f (k)(t0/2 + 0) = f (k)(t0) = 0, where k = 0, . . . , (l − 1)/2 for l odd. The operator eA is self-adjoint in the space and it is determined by the matrix (6.8) (6.9) (6.10) L2(0, t0) = L2(0, t0/2) ⊕ L2(t0/2, t0), 0 (cid:19) , A1,2 = A∗ eA =(cid:18) 0 A1,2 A2,1 2,1, where A2,1 : L2(0, t0/2) → L2(t0/2, t0). The operator A2,1 is again given by relation (6.7) on functions in Hl+1(0, t0/2) satisfying conditions (6.8) or (6.9) at the points 0 and t0/2 − 0. It follows from formula (6.10) that the spectrum of the operator eA is symmetric with respect to the point 0 and consists of eigenvalues ±an where a2 are eigenvalues of the operator A∗ An easy calculation shows that A is the differential operator A = (−1)l+1∂2l+2 in the space L2(0, t0/2) defined on functions in the class H2l+2(0, t0/2) satisfying the boundary conditions f (k)(0) = f (k)(t0/2) where k = 0, . . . , 2l + 1 for l even and the boundary conditions 2,1A2,1 =: A. n f (k)(0) = f (k)(t0/2) = f (l+1+k)(0) = f (l+1+k)(t0/2) = 0, where k = 0, . . . , (l− 1)/2 for l odd. The asymptotics of eigenvalues a2 by the Weyl formula, that is, n of A is given an = (2πt−1 0 n)l+1(1 + O(n−1)). symmetric operator A0 with finite deficiency indices. For example, A0 can be defined by formula (6.7) on C ∞ functions vanishing in some neighbourhoods of the points 0, Let us now observe that the operators A and eA are self-adjoint extensions of a t0/2 and t0. Therefore the operators A and eA have the same asymptotics of spectra. Taking Lemma 6.2 into account, we obtain the following result. Lemma 6.3. Eigenvalues of the Hankel operator H defined in Example 6.1 have the asymptotics λ(±) n = l!(2π)−l−1tl+1 0 n−l−1(1 + O(n−1)). Remark 6.4. In the case l = 0 we have the explicit formulas n = (2π)−1t0(n − 3/4)−1, λ(+) n = (2π)−1t0(n − 1/4)−1, n = 1, 2, . . . . λ(−) 34 D. R. YAFAEV 6.4. Now we are in a position to obtain the asymptotics of the spectrum of Hankel operators whose kernels (or their derivatives) have jumps of continuity. We combine Lemma 6.3 with the result by V. V. Peller (see Theorem 7.4 in Chapter 6 of his book [10]) which implies that singular numbers sn(V ) of a Hankel operator V satisfy the bound sn(V ) = o(n−l−1) if V has a symbol belonging to the Besov class Bl+1 (l+1)−1(R). Applying the Weyl theorem on the preservation of the power asymptotics for the sum of operators, we can state the following result. Theorem 6.5. Let l ∈ Z+, and let v(t) be the Fourier transform of a function in Bl+1 (l+1)−1(R). Set for t ≤ t0 and h(t) = v(t) for t > t0. Then eigenvalues of the Hankel operator H have the asymptotics h(t) = h0(t0 − t)l + v(t) n = h0l!(2π)−l−1tl+1 λ(±) 0 n−l−1(1 + o(1)) as n → ∞. We emphasize that under the assumptions of this theorem the leading terms of the asymptotics of positive and negative eigenvalues are the same. Of course if h(t) becomes smoother (l increases), then eigenvalues of the Hankel operator H decrease faster as n → ∞. 7. Perturbations of the Carleman operator In this section we consider operators H = H0 + V where H0 is the Carleman operator C (or a more general operator) and the perturbation V belongs to one of the classes introduced in Section 5. Different objects related to the operator H0 will be endowed with the index "0", and objects related to the operator V will be endowed with the index "v". 7.1. For perturbations V of finite rank, we have the following result. Theorem 7.1. Let the sign-function s0(x) of a Hankel operator H0 be bounded and positive. If V is a Hankel operator of finite rank and H = H0 + V , then N−(H) = N−(V ). In particular, H ≥ 0 if and only if V ≥ 0. Applying Theorem 5.1 to the operator V , we get an explicit formula for the total number of negative eigenvalues of the operator H. CRITERIA FOR HANKEL OPERATORS TO BE SIGN-DEFINITE 35 Corollary 7.2. Let the kernel v(t) of V be given by the formula v(t) = MXm=1 Pm(t)e−αmt where Pm(t) is a polynomial of degree Km. Define the numbers N (m) (5.2). Then N−(H) is given by formula (5.3). ± by formula Since for the Carleman operator C the sign-function equals 1, Theorem 7.1 applies to H0 = C. The inequality N−(H) ≤ N−(V ) is of course obvious because H0 ≥ 0. On the contrary, the opposite inequality N−(H) ≥ N−(V ) looks surprising because the operator H0 which may have the continuous spectrum is much "stronger" than the operator V of finite rank. At a heuristic level the equality N−(H) = N−(V ) can be explained by the fact that the supports of the sign-functions s0(x) and sv(x) are essentially disjoint. Very loosely speaking, it means that the operators H0 and V "live in orthogonal subspaces", and hence the positive operator H0 does not affect the negative spectrum of H. The detailed proof of Theorem 7.1 as well as that of Theorem 5.1 will be given in [18]. 7.2. Let C be the Carleman operator, and let V be the Hankel operator with kernel v(t) = tke−αt, α > 0, k > −1. The operator V is compact, and hence the essential spectrum specess(Hγ) of the operator (7.1) (7.2) Hγ = C − γV, γ ∈ R, coincides with the interval [0, π]. Since the sign-function of the operator C equals 1, the sign-function sγ of the operator Hγ equals sγ(x) = 1 − γsv(x) where the function sv(x) is given by formula (5.8). Let first k ∈ (−1, 0). Observe that sv(x) is continuous for x < β = − ln α and sv(x) → +∞ as x → β − 0 but sv ∈ L1(R) Thus the function sγ(x) → −∞ as x → β − 0 for all γ > 0, and hence it follows from Theorem 4.3 that the operator Hγ has an infinite negative spectrum for all γ > 0. In the case k > 0 we use the formula (7.3) and apply Corollary 4.6 with b0(ξ) = δ(ξ) and b∞(ξ) = bv(ξ). Since b0 ∈ C(R)′ and b∞ has asymptotics (5.10), the operator Hγ has a negative spectrum for all γ 6= 0. b(ξ) = δ(ξ) + bv(ξ) Let us summarize the results obtained. 36 D. R. YAFAEV Proposition 7.3. Let Hγ = C − γV where V is the Hankel operator with kernel (7.1). Then: 10 The operator Hγ has an infinite negative spectrum for all γ > 0 if k ∈ (−1, 0). 20 The operator Hγ has negative eigenvalues for all γ 6= 0 if k > 0. 7.3. The result below directly follows from Theorem 4.3. Proposition 7.4. Suppose that the sign-function sv(x) of a Hankel operator V is continuous and sv(x) → 0 as x → ∞. Then the operator Hγ defined by formula (7.2) is positive if and only if γsv(x) ≤ 1, ∀x ∈ R. If this condition is not satisfied, then Hγ has an infinite negative spectrum. We note that, by Proposition 3.12, under the assumption of Proposition 7.4 on the sign-function sv the operator V is compact. Of course this assumption is satisfied if bv ∈ L1(R). Example 7.5. Let v(t) = e−tr where r < 1. We have seen in subs. 5.3 that its sign-function sv(x) = Ir(x) where Ir(x) is integral (5.13). Recall that Ir(x) is a nonnegative continuous function of x ∈ R and Ir(x) → 0 as x → ∞. Set νr = max x∈R Ir(x). r r Then Hγ ≥ 0 if γ ≤ ν−1 γ > ν−1 and the operator Hγ has infinite negative spectrum for all . Using explicit formula (5.14) it is easy to calculate ν1/2 = 3p6/πe−3/2. In the case r > 1 we use formula (7.3). As shown in subs. 5.3, the modulus of the function bv(ξ) exponentially grows and the periods of its oscillations tend to zero only logarithmically as ξ → ∞. Therefore Theorem 4.5 yields the following result. Proposition 7.6. Let v(t) = e−tr where r > 1. Then the operator Hγ has at least one negative eigenvalue for all γ 6= 0. Appendix A. Proof of Lemma 3.6 Set t∈R+(cid:0)hln tiκtnF (n)(t)(cid:1) where for shortness we use the notation hxi = (1 + x). κ = max Let us first consider (Φ∗F )(λ) for λ ∈ (−1, 1) =: I. We have F (n) √2π(Φ∗F )(λ) =Z k 0 F (t)eiλtdt +Z ∞ k F (t)eiλtdt, k = λ−1/2. CRITERIA FOR HANKEL OPERATORS TO BE SIGN-DEFINITE 37 The first integral in the right-hand side is bounded by F (0) L1(I). In the second integral we integrate by parts 0 λ−1/2 which belongs to Z ∞ k F (t)eiλtdt = iλ−1F (k)eiλk + iλ−1Z ∞ k F ′(t)eiλtdt. (A.1) The first term here is bounded by Cλ−1hln λi−κF (0) The second term is bounded by κ which belongs to L1(I) if κ > 1. λ−1Z ∞ k t−1hln ti−κdtF (1) κ ≤ Cλ−1hln λi−κ+1F (1) κ . It belongs to L1(I) if κ > 2. Next, we consider (Φ∗F )(λ) for λ ≥ 1. Integrating by parts, we see that √2π(Φ∗F )(λ) = iλ−1Z k 0 F ′(t)eiλtdt + iλ−1Z ∞ k F ′(t)eiλtdt. (A.2) The first term here is bounded by λ−1Z k 0 t−1hln ti−κdtF (κ) 1 ≤ Cλ−1hln λi−κ+1F (κ) 1 . It belongs to L1(R \ I) if κ > 2. In the second integral in (A.2) we once more integrate by parts, that is, we use formula (A.1) with F (t) replaced by F ′(t). The function λ−2F ′(k) is bounded by λ−2k−1F (1) . For the second term, we use the estimate 0 = λ−3/2F (1) 0 (cid:12)(cid:12)λ−2Z ∞ k F ′′(t)eiλtdt(cid:12)(cid:12) ≤ λ−2Z ∞ k t−2dtF (2) 0 = λ−2k−1F (2) 0 = λ−3/2F (2) 0 . Therefore the second term in (A.2) also belongs to L1(R \ I). Appendix B. The Gaussian kernel Here we return to the Hankel operator H with kernel h(t) = e−t2. Now we proceed from the identity (Hf, f ) = (Qψ, ψ), (B.1) where ψ(t) = e−t2f (t) and Q is the integral operator with real kernel e−2ts in the space L2(R+). We shall use (B.1) essentially in the same way as the main identity (1.4). Observe that operators Q with kernels q(ts) which depend only on the product of variables can be explicitly diagonalized by the Mellin transform (see [15]). Under fairly general assumptions on q the spectrum of Q consists of the interval [−γ, γ] where γ = √2π max ξ∈R (Mq)(ξ) 38 D. R. YAFAEV and M is the Mellin transform (2.9). In particular, for q(t) = e−2t the spectrum of Q is absolutely continuous, simple and coincides with the interval [−pπ/2,pπ/2]. This allows us to check the following assertion. Proposition B.1. The Hankel operator H with kernel h(t) = e−t2 number of positive and negative eigenvalues. has infinite 1 j , . . . , ∆(−) Choose functions ϕ(±) Proof. Choose some µ ∈ (0,pπ/2). For an arbitrary N, (µ,pπ/2) and ∆(−) and kϕ(±) Let ϕ(±) = PN ± (Qϕ(±), ϕ(±)) = ± N ⊂ N ⊂ (−pπ/2,−µ) be closed mutually disjoint intervals. j k = 1, j = 1, . . . N. , . . . , ϕ(±) N . be a linear combination of the functions ϕ(±) j k2 = µkϕ(±)k2. (B.2) αj2(Qϕ(±) j = EQ(∆(±) such that ϕ(±) , . . . , ∆(+) j=1 αjϕ(±) j let ∆(+) )ϕ(±) j Then , ϕ(±) j j 1 1 j NXj=1 ) ≥ µ αj2kϕ(±) NXj=1 0 (R+) such that kψ(±) j ∈ C ∞ For an arbitrary ε > 0, we can choose ψ(±) j −ϕ(±) for all j = 1, . . . N. Since the functions ϕ(±) are orthogonal, the functions ψ(±) j linearly independent if ε is small enough. Moreover, it follows from (B.2) that j k < ε are j ±(Qψ(±), ψ(±)) ≥ 2−1µkψ(±)k2 j j=1 αjψ(±) and ε is small. if ψ(±) =PN Set now f (±)(t) = et2ψ(±)(t). Then f (±) ∈ L2(R+) and according to relation (B.1) we have the inequality ±(Hf±, f±) > 0 on linear subspace of dimension N (except f± = 0). Hence the Hankel operator H with kernel (1.15) has at least N positive and N negative eigenvalues. Since N is arbitrary, this concludes the proof. (cid:3) We emphasize that the operator H is compact while the operator Q has the continuous spectrum. Nevertheless the multiplicities of their positive and negative spectra are the same (infinite). References [1] N. Akhiezer, The classical moment problem and some related questions in analysis, Olyver and Boyd, Edinburgh and London, 1965. [2] S. N. Bernstein, Sur les fonctions absolument monotones, Acta Math. 52 (1929), 1-66. [3] T. Carleman, Sur les ´equations int´egrales singuli`eres `a noyau r´eel et sym´etrique, Almqvist and Wiksell, 1923. [4] A. Erd´elyi, W. Magnus, F. Oberhettinger, F. G. Tricomi, Higher transcendental functions. Vol. 1, 2, McGraw-Hill, New York-Toronto-London, 1953. [5] I. M. Gel'fand and N. Ya. Vilenkin, Generalized functions. Vol. 4, Academic Press, New York and London, 1964. [6] J. S. Howland, Spectral theory of operators of Hankel type. I, II Indiana Univ. Math. J. 41 (1992), no. 2, 409 -- 426, 427 -- 434. [7] W. Magnus, On the spectrum of Hilbert's matrix, Amer. J. Math. 72 (1950), 405-412. CRITERIA FOR HANKEL OPERATORS TO BE SIGN-DEFINITE 39 [8] F. G. Mehler, Uber eine mit den Kugel- und Cylinderfunctionen verwandte Function und ihre Anwendung in der Theorie der Elektricitatsvertheilung, Math. Ann. 18 (1881), 161-194. [9] Z. Nehari, On bounded bilinear forms, Ann. Math. 65 (1957), 153-162. [10] V. V. Peller, Hankel operators and their applications, Springer Verlag, 2002. [11] S. R. Power, Hankel operators on Hilbert space, Pitnam, Boston, 1982. [12] M. Rosenblum, On the Hilbert matrix, I, II, Proc. Amer. Math. Soc. 9 (1958), 137-140, 581-585. [13] D. V. Widder, The Laplace transform, Princeton Univ. Press, New Jersey, 1941. [14] H. Widom, Hankel matrices, Trans. Amer. Math. Soc. 121 (1966), 1-35. [15] D. R. Yafaev, The discrete spectrum in the singular Friedrichs model, in: Advances in Math. Sciences, AMS, v. 189, 255-274, 1999. [16] D. R. Yafaev, A commutator method for the diagonalization of Hankel operators, Funct. Anal. and Appl. 44 (2010), 295-306. [17] D. R. Yafaev, Spectral and scattering theory for perturbations of the Carleman operator, ArXiv: 1210.5709, St.-Petersburg Math. J. 25, N 2. [18] D. R. Yafaev, On finite rank Hankel operators, ArXiv, 2013. IRMAR, Universit´e de Rennes I, Campus de Beaulieu, 35042 Rennes Cedex, FRANCE E-mail address: [email protected]
1804.01730	1	1804	2018-04-05T08:28:43	Hypercyclic algebras	[ "math.FA" ]	We prove the existence of algebras of hypercyclic vectors in three cases: convolution operators, composition operators, and backward shift operators.	math.FA	math	HYPERCYCLIC ALGEBRAS FR ´ED´ERIC BAYART Abstract. We prove the existence of algebras of hypercyclic vectors in three cases: convolution operators, composition operators, and backward shift operators. 1. Introduction When we work with a linear dynamical system (X, T ), it is natural to study how its linear properties influence its dynamical properties. Here, X denotes a topological vector space and T is a continuous linear operator on X. We are mainly interested in hypercyclic operators: there exists x ∈ X, called a hypercyclic vector for T , such that {T nx; n ≥ 0} is dense in X (when X is a second-countable Baire space, this amounts to saying that T is topologically transitive). It is well known that the linear properties of (X, T ) reflects on HC(T ), the set of hypercyclic vectors for T : HC(T )∪{0} always contains a dense subspace ([9]) and there are nice criteria to determine if it contains a closed infinite-dimensional subspace (see [13], [14]). In this paper we assume that X is also a topological algebra and we ask whether HC(T ) ∪ {0} contains a nontrivial algebra; we will call this a hypercyclic algebra for T . We explore three relevant situations. 1.1. Convolution operators. Following the pioneering work of Birkhoff and MacLane, Godefroy and Shapiro have shown in [12] that a nonconstant operator which commutes with all translations τa is hypercyclic on H(C). Such an operator may be written φ(D), where φ is an entire function of exponential type and D is the derivation operator. Bayart and Matheron in [2] and independently Shkarin in [15] have shown that D admits a hypercyclic algebra. The argument of [2], which is based on the Baire category theorem and the fact thatSn ker(Dn) is dense in H(C), was refined by B`es, Conejero and Papathanasiou in [5] to prove that P (D) supports a hypercyclic algebra for all nonzero polynomials P with P (0) = 0 (see also the recent paper [11] for the existence of hypercyclic algebras for weighted backward shifts in various Fr´echet algebras). More recently, in the nice paper [6], the same authors provide further examples of entire functions φ such that φ(D) admits such an algebra. For instance, this holds true for φ(z) = cos(z) which does not satisfy φ(0) = 0 and which is not a polynomial. In stark contrast with this, it was observed in [1] that the orbit of f 2 under λτa (which corresponds to the case φ(z) = λeaz) can only contain functions for which the multiplicities of their zeros is even. Date: April 6, 2018. The author was partially supported by the grant ANR-17-CE40-0021 of the French National Research Agency ANR (project Front). 1 2 FR´ED´ERIC BAYART Our first main theorem characterizes the existence of a hypercyclic algebra for φ(D) when φ(0) < 1, or when φ(0) = 1 and φ has moderate growth. Theorem 1.1. Let φ be a nonconstant entire function with exponential type. (1) Assume that φ(0) < 1. Then the following assertions are equivalent: (i) φ(D) supports a hypercyclic algebra. (ii) φ is not a multiple of an exponential function. (2) Assume that φ(0) = 1 and φ has subexponential growth. If either φ′(0) 6= 0 or φ has order less than 1/2, then φ(D) supports a hypercyclic algebra. In view of the previous result, it is tempting to conjecture that the assumption φ(0) ≤ 1 is a necessary condition for φ(D) to admit a hypercyclic algebra. Surprizingly, this is not the case. Theorem 1.2. Let φ(z) = 2 exp(−z) + sin(z). Then φ(D) supports a hypercyclic algebra. Another natural conjecture is that φ(D) always suppport a hypercyclic algebra as soon as φ is not a multiple of an exponential function. We do not know if this conjecture is true. Nevertheless, we still get an interesting result if we weaken the conclusion. Recall that a vector x ∈ X is a supercyclic vector for T ∈ L(X) provided {λT nx; λ ∈ C, n ≥ 0} is dense in X. If X is a topological algebra, then any subalgebra of X consisting entirely (but zero) of supercyclic vectors for T is said to be a supercyclic algebra. Theorem 1.3. Let φ be a nonconstant entire function with exponential type. The following assumptions are equivalent: (i) φ(D) supports a supercyclic algebra. (ii) There exists f ∈ H(C) such that, for all m ≥ 1, f m is a hypercyclic vector for φ(D). (iii) φ is not a multiple of an exponential function. To our knowledge, the existence of f ∈ H(C) such that f m ∈ HC(cid:0)φ(D)(cid:1) for all m ≥ 1 was only known when φ has subexponential type [3]. 1.2. Composition operators. Let Ω ⊂ C be a simply connected domain and let ϕ be a holomorphic self-map of Ω. The composition operator Cϕ(f ) = f ◦ϕ is a bounded operator on H(Ω) which is hypercyclic if and only if ϕ is univalent and has no fixed point in Ω. Moreover, in that case, if P is a nonconstant polynomial, then P (Cϕ) is hypercyclic (see [4]). It was observed in [6] that Cϕ never supports a hypercyclic algebra and it was asked whether P (Cϕ) can support such an algebra. We provide an affirmative answer. Theorem 1.4. Let Ω ⊂ C be a simply connected domain and let ϕ be a holomorphic self-map of Ω which is univalent and has no fixed point in Ω. Let also P be a nonconstant polynomial which is not a multiple of z and which satisfies P (1) < 1. Then P (Cϕ) supports a hypercyclic algebra. 1.3. Backward shift operators. So far, our examples live only in F -algebras (namely in metrizable and complete topological algebras without assuming that the distance is induced by a norm) and the proofs of Theorems 1.1 and 1.4 depend heavily on the non Banach structure of the ambient space. We provide now an example in the Ba- nach algebra ℓ1(N) = {(un)n≥0; Pn un < +∞} endowed with the convolution product HYPERCYCLIC ALGEBRAS 3 j=0 ujvk−j. It was already sketched in [2] that, denoting by B the back- ward shift operator, HC(2B) ∪ {0} contains a nontrivial algebra. Again, the density of (u ⋆ v)(k) = Pk Sn ker(Bn) was the key for the proof. We go much further (we denote by D the open unit disk and by T its boundary the unit circle). Theorem 1.5. Let P ∈ C[X] be nonconstant and let B be the backward shift operator on ℓ1(N). Then the following assertions are equivalent. (i) P (B) is hypercyclic. (ii) P (B) admits a hypercyclic algebra. (iii) P (D) ∩ T 6= ∅. 1.4. Organization of the paper and strategy for the proofs. The proofs of our results develop a method initiated in [6] and use eigenvalues and eigenvectors of our op- erators. In Section 2, we use eigenvalues of modulus slightly bigger than 1 to prove half of Theorem 1.1 and Theorem 1.4. In Section 3, we use eigenvalues with large modulus to prove the remaining part of Theorem 1.1 whereas in Section 4, we use unimodular eigenvalues to prove Theorem 1.5. There is also a significant difference between the first two cases and the last one: the product of two eigenvectors can or cannot be still an eigenvector. The latter situation is of course more difficult! In Section 5, we come back to convolution operators and study the case φ(0) > 1. As a consequence we prove Theorems 1.2 and 1.3. Of course, if we know that an operator admits a hypercyclic algebra, then it is natural to ask how big it can be. Can it be dense? Can it be infinitely generated? These questions were investigated very recently in [7] where it is shown that for most of the examples exhibited in [6], one can improve the construction to get a dense and infinitely generated hypercyclic algebra. In Section 6, we show how to modify our proofs to obtain a similar improvement. We choose to postpone this in a separate section because the arguments become more technical and we think that the ideas appear more clearly by handling separately the case of singly generated algebras. We end up this introduction with the following lemma taken from [2, Remark 8.28], which gives a criterion for the existence of a hypercyclic algebra. It can be seen as a strong form of the property of topological transitivity. Lemma 1.6. Let T be a continuous operator on some separable F -algebra X. Assume that, for any pair (U, V ) of nonempty open sets in X, for any open neighbourhood W of zero, and for any positive integer m, one can find u ∈ U and an integer N such that T N (un) ∈ W for all n < m and T N (um) ∈ V . Then T admits a hypercyclic algebra. 2. Small eigenvalues 2.1. A general result for operators with small eigenvalues. We shall deduce part of Theorem 1.1 and Theorem 1.4 from the following general result. Theorem 2.1. Let X be an F-algebra and let T ∈ L(X). Assume that there exist a func- tion E : C → X and an entire function φ : C → C satisfying the following assumptions: 4 FR´ED´ERIC BAYART (1) for all λ ∈ C, T E(λ) = φ(λ)E(λ); (2) for all λ, µ ∈ C, E(λ)E(µ) = E(λ + µ); (3) for all Λ ⊂ C with an accumulation point, span (E(λ); λ ∈ Λ) is dense in X; (4) φ is not a multiple of an exponential function; (5) for all m ∈ N, there exist a, b ∈ C such that φ(mb) > 1 and, for all n ∈ {1, . . . , m}, all d ∈ {0, . . . , n}, with (n, d) 6= (m, m), φ(db + (n − d)a) < 1. Then T supports a hypercyclic algebra. We start with a lemma which explains why we have to exclude multiples of exponential functions. Lemma 2.2. Let φ be an entire function which is not a multiple of an exponential function. Then, for any w0 ∈ C with φ(w0) 6= 0 and any δ > 0, there exist w1, w2 ∈ B(w0, δ), w1 6= w2, such that the map [0, 1] → R, t 7→ log φ(tw1 + (1 − t)w2) is stricly convex. Proof. Since φ(w0) 6= 0, there exist some neighbourhood V of w0 and a holomorphic of an exponential function, we know that h is not an affine map. Thus there exists w1 ∈ B(w0, δ) ∩ V such that h′′(w1) 6= 0. Without loss of generality, we assume that function h : V → C such that φ(z) = exp(cid:0)h(z)(cid:1) for all z ∈ V . Since φ is not a multiple w1 = 0 and we write h(z) =P+∞ log φ(z) = ℜe(cid:0)h(z)(cid:1) = ℜe(a0) + (ℜe(a1)x − ℑm(a1)y) k=0 akzk. Then if z = x + iy. Since a2 6= 0, one may find (x0, y0) ∈ R2 with +(cid:0)ℜe(a2)(cid:0)x2 − y2(cid:1) − 2ℑm(a2)xy(cid:1) + o(cid:0)x2 + y2(cid:1) ℜe(a2)(cid:0)x2 0(cid:1) − 2ℑm(a2)x0y0 > 0. 0 − y2 Then g(t) = ℜe (h (t (x0 + iy0))) = b0 + b1t + b2t2 + o(t2) with b2 > 0 is strictly convex around 0. (cid:3) Proof of Theorem 2.1. Let U, V, W be nonempty open subsets of X with 0 ∈ W and let m ≥ 1. By Lemma 1.6, it suffices to find u ∈ U and N ∈ N so that (1) (2) T N (un) ∈ W, n = 1, . . . , m − 1 T N (um) ∈ V. The assumptions give us for this value of m two complex numbers a and b. We set w0 = mb. We then consider δ > 0 very small and w1, w2 ∈ B(w0, δ) such that • t ∈ [0, 1] 7→ log φ (tw1 + (1 − t)w2) is strictly convex; • φ > 1 on [w1, w2]; • for all n ∈ {1, . . . , m}, for all d ∈ {0, . . . , n} with (n, d) 6= (m, m), for all λ1, . . . , λd ∈ [w1, w2] and all γ1, . . . , γn−d ∈ B(a, δ), (3) (cid:12)(cid:12)(cid:12)(cid:12) φ(cid:18) λ1 + · · · + λd m + γ1 + · · · + γn−d(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) < 1. HYPERCYCLIC ALGEBRAS 5 We may ensure this last property because λ1 + · · · + λd m + γ1 + · · · + γn−d = db + (n − d)a + z where the norm of z is controlled by δ. Since B(a, δ) and [w1, w2] have accumulation points, we may find p, q ∈ N, complex numbers a1, . . . , ap, b1, . . . , bq, complex numbers γ1, . . . , γp ∈ B(a, δ) and complex numbers λ1, . . . , λq in [w1, w2] such that p q alE (γl) ∈ U and bjE (λj) ∈ V. Xl=1 Xj=1 For N ≥ 1 and j ∈ {1, . . . , q}, let cj := cj(N ) be any complex number satisfying cm j = bj/(cid:0)φ(λj)(cid:1)N and define u := uN = p q alE (γl) + cjE (λj/m) . Xl=1 Xj=1 We claim that, for N large enough, uN belongs to U and satisfies (1) and (2). That uN belongs to U is clear, since cj(N ) tends to zero as N goes to infinity. In order to prove the other points, we need to compute un for 1 ≤ n ≤ m. To simplify the notations, let Ip = {1, . . . , p} and for a multi-index l ∈ I d p , al will stand for al1 · · · ald with the convention that an empty product is equal to 1. Then we may write un = n Xd=0 Xl∈I n−d p j∈I d q α(l, j, d, n)alcjE(cid:18)γl1 + · · · + γln−d + λj1 + · · · + λjd m (cid:19) for some coefficients α(l, j, d, n) that we do not try to compute, but which does not depend on N . To prove that T N (un) belongs to W for N large enough and 1 ≤ n < m, we only have to prove that, for any d ∈ {0, . . . , n} and any l ∈ I n−d , any j ∈ I d q , p cj(N )(cid:18)φ(cid:18)γl1 + · · · + γln−d + λj1 + · · · + λjd m (cid:19)(cid:19)N −→N→+∞ 0. This follows from (3) and the fact that cj(N ) tends to zero. The case n = m is slightly different. We denote by Dq the diagonal of I m m-uples (j, . . . , j), 1 ≤ j ≤ q. Then we decompose um into q , namely the um = m−1 Xd=0 Xl∈I m−d p j∈I d q α(l, j, d, m)alcjE(cid:18)γl1 + · · · + γln−d + λj1 + · · · + λjd m (cid:19) α(j, m)cjE(cid:18) λj1 + · · · + λjm m (cid:19) q \Dq + Xj∈I m Xj=1 + q cm j E (λj) =: v1 + v2 + v3. 6 FR´ED´ERIC BAYART The same considerations as above show that T N v1 tends to zero as N goes to infinity. That T N v2 tends also to zero follows from a convexity argument. Indeed, for j ∈ I m q \Dq, the strict convexity of the map t 7→ log φ (tw1 + (1 − t)w2) implies that < φ (λj1)1/m · · · φ (λjm)1/m . Moreover, m φ(cid:18) λj1 + · · · + λjm (cid:12)(cid:12)(cid:12)(cid:12) cj(N ) ×(cid:12)(cid:12)(cid:12)(cid:12) φ(cid:18) λj1 + · · · + λjm m (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) N Since the left hand side of this inequality goes to zero, we get that T N v2 tends to zero. We conclude the proof by observing that N . m φ(cid:16) λj1 +···+λjm (cid:17) φ (λj1)1/m · · · φ (λjm)1/m(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ bj1/m(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Xj=1 q T N v3 = bjE (λj) . (cid:3) For the applications, we emphasize two corollaries of Theorem 2.1. Corollary 2.3. Let X be an F-algebra and let T ∈ L(X). Assume that there exist a func- tion E : C → X and an entire function φ : C → C satisfying the following assumptions: (1) for all λ ∈ C, T E(λ) = φ(λ)E(λ); (2) for all λ, µ ∈ C, E(λ)E(µ) = E(λ + µ); (3) for all Λ ⊂ C with an accumulation point, span (E(λ); λ ∈ Λ) is dense in X; (4) φ is not a multiple of an exponential function; (5) for all ρ ∈ (0, 1), there exists w0 ∈ C with φ (w0) > 1 and, for all r ∈ (0, ρ], φ (rw0) < 1. Then T supports a hypercyclic algebra. Proof. We show that Assumption (5) of Theorem 2.1 is satisfied. Let m ≥ 1 and ε ∈ (0, 1/m). Set ρ = m−1 m + mε. We get the existence of w0. We set b = w0/m and a = εw0/m. Then, for any d ≤ m − 1 and any n ≤ m, db + (n − d)a ≤(cid:18)m − 1 m + ε(cid:19) w0 ≤ ρw0 showing that φ(db + (n − d)a) < 1. (cid:3) Corollary 2.4. Let X be an F-algebra and let T ∈ L(X). Assume that there exist a func- tion E : C → X and an entire function φ : C → C satisfying the following assumptions: (1) for all λ ∈ C, T E(λ) = φ(λ)E(λ); (2) for all λ, µ ∈ C, E(λ)E(µ) = E(λ + µ); (3) for all Λ ⊂ C with an accumulation point, span (E(λ); λ ∈ Λ) is dense in X; (4) φ is not a multiple of an exponential function; (5) φ(0) < 1. Then T supports a hypercyclic algebra. HYPERCYCLIC ALGEBRAS 7 Proof. We prove that Assumption (5) of Corollary 2.3 is satisfied. Denote by M (r) = sup{φ(z); z = r} which is a continuous and increasing function of r satisfying M (0) < 1. Let r0 > 0 with M (r0) = 1 and let r1 > r0 with ρr1 < r0. Any w0 ∈ C such that w0 = r1 and φ(w0) = M (r1) > 1 does the job. (cid:3) Remark 2.5. An operator on a Banach space cannot satisfy the assumptions of Theorem 2.1. Indeed, they imply that its spectrum is unbounded. We shall see later how it remains possible to get a hypercyclic algebra in a Banach algebra context. 2.2. Applications to convolution operators. We now show how to deduce the first half of Theorem 1.1 from Corollary 2.4. Thus, let φ be an entire function of exponential type which is not a multiple of an exponential function. Then we let X = H(C) and T = φ(D). The map E is defined by E(λ)(z) = exp(λz); it satisfies (1), (2) and (3) of Corollary 2.4. Corollary 2.3 may also be applied to functions satisfying φ(0) = 1. We give here two examples which are not covered by Theorem 1.1. Example 2.6. ([6]) Let φ(z) = cos(z). Then φ(D) supports a hypercyclic algebra. Proof. Recall that if z = x + iy, then cos(z)2 = cos2 x + sinh2 y. Let us set ψ(t) = cos2(2t) + sinh2(t). Using standard calculus one may prove that there exists t0 > 0 such that ψ is decreasing on (0, t0) and increasing on (t0, +∞). Let t1 > 0 be such that ψ(t1) = 1. It then suffices to consider w0 = (1 + η)(2 + i)t1 for some sufficiently small η > 0. (cid:3) Example 2.7. Let φ(z) = ez − 2. Then φ(D) supports a hypercyclic algebra. Proof. Let t1 > 0 be such that φ(t1) = 1. It suffices to consider w0 = (1 + η)t1 for some sufficiently small η > 0. (cid:3) We now give a surprizing example of an entire function φ with φ(0) > 1 and φ(D) supports a hypercyclic algebra. Example 2.8. Let φ(z) = 2e−z + sin z. Then φ(D) supports a hypercyclic algebra. Proof. We just need to show that φ satisfies Assumption (5) of Theorem 2.1. We let a = kπ for some sufficiently large k and b = kπ + π 2m . Then φ (db + (n − d)a) =(cid:12)(cid:12)(cid:12)(cid:12) sin(cid:18)nkπ + dπ 2m(cid:19) + 2 exp(cid:18)−nkπ − . dπ 2m(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) Provided k is large enough, this is less than 1 as soon as d < m, whereas φ(mb) > 1. (cid:3) 2.3. Applications to composition operators. In this subsection, we prove Theorem 1.4. Recall that given a simply connected domain Ω ⊂ C and ϕ a holomorphic self-map of Ω, Cϕ is hypercyclic if and only if ϕ is univalent and without fixed points. We first prove Theorem 1.4 when Ω = C. In that case ϕ is also entire hence ϕ is a translation ϕ(z) = z + a, a 6= 0. Thus P (Cϕ) = φ(D), where φ(z) = P ◦ exp(az) and the result is a particular case of Theorem 1.1. Otherwise, by the Riemann mapping theorem, we may assume that Ω = D. We simplify the proof by using the linear fractional model (see for instance [10]): since ϕ has no fixed 8 FR´ED´ERIC BAYART points in D, there exists a univalent map σ : D → C and a linear fractional map ψ such that σ ◦ ϕ = ψ ◦ σ. Moreover • either ψ can be taken to be a dilation ψ(z) = rz for some 0 < r < 1; • or ψ can be taken to be a translation ψ(z) = z + a for some a ∈ C\{0} and σ(D) ⊂ {z; ℜe(z) > 0}. The functional equation guarantees that σ(D) =: U is preserved by ψ and that P (Cϕ) acting on H(D) and P (Cψ) acting on H(U ) are quasi-conjugate by Cσ. Since Cσ is a multiplicative map, it is sufficient to prove that P (Cψ) admits a hypercyclic algebra (see [6, Remark 6]). ◮ The translation case. We denote T = P (Cψ) and E(λ)(z) = exp(λz) so that T E(λ) = φ(λ)E(λ) with φ(λ) = P (exp(aλ)). Then the assumptions of Corollary 2.4 are satis- fied provided we are able to prove that, for any Λ ⊂ C with an accumulation point, span (E(λ); λ ∈ Λ) is dense in H(U ). The proof is exactly similar to that for H(C), since the polynomials are dense in H(U ) (recall that U is simply connected) - see for instance [12]. ◮ The dilation case. We still denote T = P (Cψ) but now we set E(λ) = zλ. This defines a holomorphic function on U since U ⊂ {z; ℜe(z) > 0}. Moreover, T E(λ) = φ(λ)E(λ) with φ(λ) = P (exp(λ log r)). Again the assumptions of Corollary 2.4 are satisfied provided that, for all Λ ⊂ C with an accumulation point, span (E(λ); λ ∈ Λ) is dense in H(U ). Let L be a linear form on H(U ) which vanishes on span (E(λ); λ ∈ Λ). By the Riesz representation theorem, there exists K a compact subset of U and µ a complex measure supported in K such that, for all f ∈ H(U ), L(f ) = RK f dµ. The map λ 7→ L(zλ) is holomorphic and has an accumulation point of zeros. Hence it is identically zero. Therefore, L vanishes on each monomial zn, hence on H(U ) since U is simply connected. This shows that span (E(λ); λ ∈ Λ) is dense in H(U ). 3. Large eigenvalues As in Section 2, we shall deduce Part (2) of Theorem 1.1 from a more general statement. Theorem 3.1. Let X be an F-algebra and let T ∈ L(X). Assume that there exist a function E : C → X and a nonconstant entire function φ : C → C satisfying the following assumptions: (1) for all λ ∈ C, T E(λ) = φ(λ)E(λ); (2) for all λ, µ ∈ C, E(λ)E(µ) = E(λ + µ); (3) for all Λ ⊂ C with an accumulation point, span (E(λ); λ ∈ Λ) is dense in X; (4) φ(0) = 1, φ has subexponential growth and either φ′(0) 6= 0 or φ has order less than 1/2. Then T supports a hypercyclic algebra. The proof of Theorem 3.1 shares many similarities with that of Theorem 2.1. Nevertheless, we will now choose the complex numbers λj with φ (λj) very large (instead of being slightly bigger than 1). In this way, because φ has subexponential growth, we will ensure that φ(λj) is bigger than φ(2λj )1/2. Thus, when we will take the powers of u and apply T N , the main term will change. We will also need a more careful interaction between the HYPERCYCLIC ALGEBRAS 9 λj' and the γk'. This is the content of the following key lemma, which uses the fact that we control the growth of φ. Lemma 3.2. Let φ be a nonconstant entire function with subexponential growth and φ(0) = 1. Assume that either φ′(0) 6= 0 or φ has order less than 1/2. Then for all m ≥ 2, there exist z0 ∈ C\{0} and w0 = ρz0 for some ρ > 0 such that • φ < 1 on (0, z0]; • φ(w0) > 1; • φ(w0) > φ(dw0)1/d for all d = 2, . . . , m. • t 7→ φ(w0 + tz0) is increasing on some interval [0, η), η > 0. Proof. We first show the existence of z0, z1 ∈ C with z0 ∈ (0, z1), φ < 1 on (0, z0] and φ(z1) > 1. The proof differs here following the assumptions made on φ. Assume first that φ has order less than 1/2. Write φ(z) = eiθ0 + ρpeiθpzp + o(zp) with ρp > 0. Then, φ(cid:0)te−i(θp−θ0+π)/p(cid:1) = eiθ0 − ρptpeiθ0 + o(tp) has modulus less than 1 provided t is small enough. We then set z0 = te−i(θp−θ0+π)/p for some small t. We then find z1 since any nonconstant entire function of order less than 1/2 cannot be bounded on a half-line (see [8, Theorem 3.1.5]). On the other hand, suppose now that φ′(0) 6= 0. Without loss of generality we may assume that φ(z) = 1 − az + o(z) for some a > 0. Now, a nonconstant entire function with subexponential growth cannot be bounded on a half-plane (see [8, Theorem 1.4.3]). Thus, there exists z1 = r1eiα1 with r1 > 0 and α1 ∈ (−π/2, π/2) such that φ(z1) > 1. It is easy to check that, for t > 0 small enough, φ(tz1) < 1 and we set z0 = tz1 for such a small t > 0. We now proceed with the construction of w0. Without loss of generality we may assume z0 = 1. We set ψ(t) = φ(t)2. We proceed by contradiction and we assume that there does not exist w0 > 0 such that the last three points of the lemma are satisfied. We fix a n=1(1 − εn) > 0. We shall construct two sequences sequence (εn) in (0, 1) such that κ :=Q+∞ (tn)n≥0 and (rn)n≥1 of positive real numbers such that, for all n ≥ 0, ψ(tn) > 1, ψ′(tn) > 0 tn = rntn−1 rn ∈ (1, m]   ψ(tn) ≥ max(cid:0)ψ(tn−1)(1−εn)rn, ψ(tn−1)3/2(cid:1) . First, the existence of z1 leads to some positive real number t0 > 1 such that ψ(t0) > 1 and ψ′(t0) > 0. Next, assume that the construction has been done until step n and let us proceed with step n + 1. Since we assumed that the conclusion of Lemma 3.2 is false, there exists kn+1 ∈ {2, . . . , m} such that ψ(kn+1tn) ≥ ψ(tn)kn+1 > max(cid:16)ψ(tn)(1−εn+1)kn+1, ψ(tn)3/2(cid:17) . If ψ′(kn+1tn) > 0, then we are done by choosing tn+1 = kn+1tn. Otherwise, let τ := sup{t ∈ [tn, kn+1tn]; ψ′(t) > 0}. Then ψ(τ ) ≥ ψ(kn+1tn). 10 FR´ED´ERIC BAYART Thus, there exists tn+1 ∈ [tn, τ ] such that ψ′(tn+1) > 0 and Now, tn+1 = rn+1tn for some rn+1 ∈ (1, kn+1) so that ψ(tn+1) > max(cid:16)ψ(tn)(1−εn+1)kn+1, ψ(tn)3/2(cid:17) . ψ(tn+1) > max(cid:16)ψ(tn)(1−εn+1)rn+1, ψ(tn)3/2(cid:17) as required to prove step n + 1. The sequence (tn) we have just built satisfies, for all n ≤ N , ψ(tn) ≥ (ψ(t0))(3/2)n In particular, (ψ (tn)), hence (tn), go to infinity. Now . ψ(tn) ≥ ψ(t0)Qn k=1(1−εk)rk ≥ ψ(t0)κtn/t0. This is a contradiction since ψ has subexponential growth. (cid:3) Proof of Theorem 3.1. Let w0, z0 6= 0 be given by Lemma 3.2. Then we may find γ1 ∈ (0, z0/m) which is sufficiently close to 0 so that • φ (w0 + (m − 1)γ1) > 1; • φ (w0 + (m − 1)γ1) > φ (dw0 + sγ1)1/d for all d ∈ {2, . . . , m} and for all s ∈ {0, . . . , m − d}; • φ (w0 + (m − 1)γ1) > φ (w0 + sγ1) for all s ∈ {0, . . . , m − 2}. We then fix δ > 0 sufficiently small so that • for all λ ∈ B(w0, δ), φ (λ + (m − 1)γ1) > 1; • for all d ∈ {1, . . . , m}, for all s ∈ {0, . . . , m − d} with (d, s) 6= (1, m − 1), for all λ, λ1, . . . λd ∈ B(w0, δ), for all z ∈ B(0, δ), φ (λ + (m − 1)γ1) > φ (λ1 + · · · + λd + sγ1 + z)1/d . Let p, q be integers, let a1, . . . , ap, b1, . . . , bq be complex numbers, let γ2, . . . , γp ∈ (0, z0/m)∩ B(0, δ/m) and let λ1, . . . , λq ∈ B(w0, δ) be such that alE (γl) ∈ U p Xl=1 q bjE (λj + (m − 1)γ1) ∈ V. Xj=1 Without loss of generality we may assume a1 6= 0. For N ≥ 1 and j ∈ {1, . . . , q}, let cj := cj(N ) be defined by and let us set u := uN = p cj = mam−1 1 bj (cid:0)φ (λj + (m − 1)γ1)(cid:1)N Xl=1 alE (γl) + Xj=1 q cjE (λj) HYPERCYCLIC ALGEBRAS 11 (observe that now we do not divide λj by m). As before, for N large enough, u belongs to U . Moreover, the formula for un is similar: n un = Xd=0 Xl∈I n−d p j∈I d q α(l, j, d, n)alcjE(cid:0)γl1 + · · · + γln−d + λj1 + · · · + λjd(cid:1) . Assume first that n < m and let us show that, for all d ∈ {0, . . . , n}, for all l ∈ I n−d all j ∈ I d q , p and (4) Assume first d ≥ 1 and let s = card {i; li = 1}. Then γl1 + · · · + γln−d + λj1 + · · · + λjd = λj1 + · · · + λjd + sγ1 + z with z < δ and s ≤ m − 2. So, writing N cj(N ) ×(cid:12)(cid:12)φ(cid:0)γl1 + · · · + γln−d + λj1 + · · · + λjd(cid:1)(cid:12)(cid:12) cj(N ) ×(cid:12)(cid:12)φ(cid:0)γl1 + · · · + γln−d + λj1 + · · · + λjd(cid:1)(cid:12)(cid:12) N −→N→+∞ 0. = bj mda1(m−1)d d Yi=1 φ (λj1 + · · · + λjd + sγ1 + z)1/d φ (λji + (m − 1)γ1) !N we observe that (4) is true. If d = 0 (in that case, cj(N ) = 1), (4) remains also true since γl1 + · · · + γln ∈ (0, z0] so that φ (γl1 + · · · + γln) < 1. This yields that T N (un) tends to zero. The case n = m requires small modifications. We now decompose um into um = mam−1 cjE (λj + (m − 1)γ1) q p 1 m + Xj=1 Xd=0 d6=1 Xl∈I m−d + Xl∈I m−1 j∈I d q p j∈I 1 q α(l, j, d, m)alcjE(cid:0)γl1 + · · · + γln−d + λj1 + · · · + λjd(cid:1) α(l, j, d, m)alcjE(cid:0)γl1 + · · · + γln−d + λj1 + · · · + λjd(cid:1) l6=(1,...,1) =: v1 + v2 + v3. With exactly the same argument as above, one shows that T N (v2 + v3) tends to zero. Furthermore, T N (v1) = which closes the argument. bjE (λj + (m − 1)γ1) ∈ V, m Xj=1 (cid:3) 4. Unimodular eigenvalues 4.1. Proof of Theorem 1.5. In this section, we provide a proof for Theorem 1.5. As before, P (B) admits a natural family of eigenvectors: for any λ ∈ D, (λk) is an eigenvector of P (B) associated to P (λ). Nevertheless, we do not have a so simple formula for the product of two eigenvectors. 12 FR´ED´ERIC BAYART Lemma 4.1. Let Θ = (θ1, . . . , θn) ∈ Cn and let µ1, . . . , µr be pairwise distinct complex numbers such that {θ1, . . . , θn} = {µ1, . . . , µr}. For j = 1, . . . , r, let κj = card{k; θk = µj}. Then there exist polynomials PΘ,j, 1 ≤ j ≤ r, with degree less than or equal to κj − 1 such that 1(cid:17) ⋆ · · · ⋆(cid:16)θk (cid:16)θk n(cid:17) = When all the θi are equal to the same λ, then r j(cid:17) . Xj=1(cid:16)PΘ,j(k)µk where deg(Pn) = n − 1. (cid:16)λk(cid:17) ⋆ · · · ⋆(cid:16)λk(cid:17) =(cid:16)Pn(k)λk(cid:17) The statement of this lemma motivates the study of the effect of P (B)N on the vectors (kdλk). Lemma 4.2. Let P ∈ C[X] and let d ≥ 0. Let also λ ∈ D with λP (λ)P ′(λ) 6= 0. There exist complex numbers (Ad,N,s)N ≥0, 0≤s≤d, such that, for all N ≥ 0, (5) (cid:0)P (B)(cid:1)N(cid:16)kdλk(cid:17) = d Xs=0 P (λ)N +s−dAd,N,s(cid:16)ksλk(cid:17) with Ad,N,s ∼N→+∞ ωd,sN d−s for some nonzero ωd,s. We point out that in the statement of the previous lemma, the complex number ωd,s may depend on λ; later we will sometimes denote them ωd,s(λ). We will also need a density lemma. Lemma 4.3. Let Λ ⊂ D with an accumulation point inside D. Then(cid:8)(cid:0)λk(cid:1) ; λ ∈ Λ(cid:9) spans a dense subspace of ℓ1(N). We postpone the proof of these lemmas to give that of Theorem 1.5. Proof of Theorem 1.5. Since the spectrum of a hypercyclic operator has to intersect the unit circle, if P (D) is hypercyclic, then P (D) ∩ T 6= ∅. Hence the only difficult implication is (iii) =⇒ (ii). Thus we start with a nonconstant polynomial satisfying P (D) ∩ T 6= ∅. Let Λ1 ⊂ D with an accumulation point in D such that P (λ) = 1 and λP ′(λ) 6= 0 for all λ ∈ Λ1. Let also Λ2 ⊂ D with an accumulation point in D such that P (λ) < 1 and λP ′(λ) 6= 0 for all λ ∈ Λ2. Let U, V, W be nonempty open subsets of ℓ1 with 0 ∈ W . Let m ≥ 1. We may find p, q ∈ N, complex numbers γ1, . . . , γp ∈ Λ2, λ1, . . . , λq ∈ Λ1, a1, . . . , ap, b1, . . . , bq such that p Xl=1 al(cid:16)γk l(cid:17) ∈ U and q Xj=1 bj(cid:16)λk j(cid:17) ∈ V. We then set, for j = 1, . . . , q, N ≥ 0, cj := cj(N ) =(cid:18) ωm−1,0(λj)N m−1P (λj)N −m+1(cid:19)1/m bj HYPERCYCLIC ALGEBRAS 13 (we take any m-th root) and u := uN = p Xl=1 al(cid:16)γk l(cid:17) + q Xj=1 j(cid:17) cj(cid:16)λk so that, if N is large enough, u belongs to U . As usual, for n ∈ N, un = n Xd=0 Xl∈I n−d p j∈I d q α(l, j, d, n)alcj(cid:16)γk l1(cid:17) ⋆ · · · ⋆(cid:16)γk ln−d(cid:17) ⋆(cid:16)λk jd(cid:17) . j1(cid:17) ⋆ · · · ⋆(cid:16)λk Let us fix for a while n ≤ m, d ∈ {0, . . . , n}, l ∈ I n−d p and j ∈ I d q . Applying Lemma 4.1, we observe that (cid:0)γk µ ∈ {γl1, . . . , λjd} and s ≤ card {l; γl = µ} + card {j; λj = µ} − 1. Moreover, by Lemma 4.2, jd(cid:17) writes as a linear combination of (cid:0)ksµk(cid:1) for some l1(cid:1) ⋆ · · · ⋆(cid:16)λk P (B)N(cid:16)ksµk(cid:17)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13) cj(N ) ×(cid:13)(cid:13)(cid:13) P (B)N(cid:16)ksµk(cid:17)(cid:13)(cid:13)(cid:13) ≤ CN sP (µ)N . ≤ C N s . N d× m−1 m In particular, if µ ∈ (cid:8)γl1, . . . , γln−d(cid:9), so that P (µ) < 1, then cj(N ) ×(cid:13)(cid:13)P (B)N(cid:0)ksµk(cid:1)(cid:13)(cid:13) tends to zero. If µ ∈ {λj1, . . . , λjd}, then (6) Assume first that n < m. Then s d ≤ d − 1 d ≤ n − 1 n < m − 1 m . Hence, the right hand side of (6) goes to zero as N tends to +∞. This implies in particular that P (B)N (un) goes to zero as N tends to +∞. Assume now that n = m. The same we find that j1(cid:17) ⋆ · · · ⋆(cid:16)λk argument shows that cj(N ) ×(cid:13)(cid:13)P (B)N(cid:0)ksµk(cid:1)(cid:13)(cid:13) tends to zero, except if s = d − 1 = m − 1. jd(cid:17)(cid:17) tends to zero except for ln−d(cid:17) ⋆(cid:16)λk Namely, cj(N )P (B)N(cid:16)(cid:0)γk the terms cj(N )P (B)N(cid:16)(cid:16)λk j(cid:17)(cid:17). Applying again Lemma 4.1 and Lemma 4.2, j(cid:17) ⋆ · · · ⋆(cid:16)λk cjP (B)N(cid:16)(cid:16)λk l1(cid:1) ⋆ · · · ⋆(cid:16)γk j(cid:17) ⋆ · · · ⋆(cid:16)λk j(cid:17)(cid:17) = P (λ)N −m+1ωm−1,0(λj)N m−1(λk j(cid:17) ⋆ · · · ⋆(cid:16)λk P (B)N(cid:16)(cid:16)λk j(cid:17)(cid:17) = bj(λk Hence, by the definition of cj, This achieves the proof that P (B)N (um) belongs to V provided N is large enough. j ) + o(N m−1). j ) + o(1). (cid:3) 4.2. Proof of Lemma 4.1. The proof of Lemma 4.1 relies on the following facts and an easy induction. Fact 1. For any d ≥ 0, for any λ ∈ D, there exists a polynomial Qd with deg(Qd) = d + 1 such that (cid:16)kdλk(cid:17) ⋆(cid:16)λk(cid:17) =(cid:16)Qd(k)λk(cid:17) . For the induction step, we write using Fact 1 k uk = λ µ λjµk−j = µ − λ(cid:16)µk(cid:17) . λ − µ(cid:16)λk(cid:17) + Xj=0 (cid:16)kd+1λk(cid:17) = a(cid:16)kdλk(cid:17) ⋆(cid:16)λk(cid:17) +(cid:16)P (k)λk(cid:17) 14 FR´ED´ERIC BAYART Fact 2. For any d ≥ 0, for any λ, µ ∈ D with λ 6= µ, there exist a polynomial Qd,λ,µ with deg(Qd,λ,µ) ≤ d and a complex number Bd,λ,µ such that (cid:16)kdλk(cid:17) ⋆(cid:16)µk(cid:17) =(cid:16)Qd,λ,µ(k)λk(cid:17) + Bd,λ,µ(cid:16)µk(cid:17) . The proof of Fact 1 is easy. Denoting (cid:0)kdλk(cid:1) ⋆(cid:0)λk(cid:1) by (uk), we have k uk = jdλk Xj=0 so that the result is proved with Qd the polynomial of degree d + 1 such that Qd(k) = j=0 jd for all k ∈ N. Pk The proof of Fact 2 can be done by induction on d. For d = 0, we simply write(cid:0)λk(cid:1) ⋆(cid:0)µk(cid:1) as (uk) with with deg(P ) ≤ d (to simplify the notations, we do not write the subscripts on the complex numbers and on the polynomials involved, but they clearly depend on d, λ, and later on µ). Thus, We then apply the induction hypothesis to both (cid:0)kdλk(cid:1) ⋆(cid:0)µk(cid:1) and (cid:0)P (k)λk(cid:1) ⋆(cid:0)µk(cid:1) to get (cid:16)kd+1λk(cid:17) ⋆(cid:16)µk(cid:17) = a(cid:16)λk(cid:17) ⋆(cid:16)(cid:16)kdλk(cid:17) ⋆(cid:16)µk(cid:17)(cid:17) +(cid:16)P (k)λk(cid:17) ⋆(cid:16)µk(cid:17) . (cid:16)kd+1λk(cid:17) ⋆(cid:16)µk(cid:17) = a(cid:16)λk(cid:17) ⋆(cid:16)(cid:16)Q(k)λk(cid:17) + b(cid:16)µk(cid:17)(cid:17) +(cid:16)R(k)λk(cid:17) + c(cid:16)µk(cid:17) with deg(Q), deg(R) ≤ d and b, c ∈ C. We conclude by using again either Fact 1 or the case d = 0. 4.3. Proof of Lemma 4.2. We first isolate the case N = 1. Lemma 4.4. Let P ∈ C[X], d ≥ 0. There exist polynomials (Qd,s), 0 ≤ s ≤ d, such that, for all λ ∈ D, P (B)(cid:16)kdλk(cid:17) = Qd,s(λ)(cid:16)ksλk(cid:17) d Xs=0 where Qd,d(λ) = P (λ) and Qd,d−1(λ) = dλP ′(λ). Proof. We start from Bn(cid:16)kdλk(cid:17) =(cid:16)(k + n)dλk+n(cid:17) = λn(cid:16)Pn,d(k)λk(cid:17) where Pn,d is a monic polynomial of degree d. More precisely, Pn,d(k) = kd + dnkd−1 + · · · . The result follows now from a linear combination of these equalities. Observe that if (cid:3) j=1 αjdjλj = dλP ′(λ). P (X) =Pn j=0 αjX j, then Qd,d−1(λ) =Pn HYPERCYCLIC ALGEBRAS 15 Proof of Lemma 4.2. We first prove by induction on N that the relation (5), which is clearly true for N = 0, holds for all N and we get an induction formula for the complex numbers Ad,N,s. Indeed, assuming (5) is true for N and using Lemma 4.4, we have (P (B))N +1 (kdλk) = = r P (λ)N +r−dAd,N,r Xs=0 Qr,s(λ)(cid:16)ksλk(cid:17) P (λ)N +r−dAd,N,rQr,s(λ)(cid:16)ksλk(cid:17) . Xr=s d Thus, (5) is true for N + 1 with the induction formula (7) Ad,N +1,s = P (λ)r−s−1Ad,N,rQr,s(λ). d d Xr=0 Xs=0 d Xr=s When s = d, using Qd,d(λ) = P (λ), this formula simply writes Ad,N +1,d = Ad,N,d so that Ad,N,d = 1 for all N . When s = d−1, using Qd−1,d−1(λ) = P (λ) and Qd,d−1(λ) = dλP ′(λ), we have Ad,N +1,d−1 = dλP ′(λ)Ad,N,d + Ad,N,d−1 so that Ad,N,d−1 = N dλP ′(λ). Assume now that we have shown that Ad,N,r ∼N→+∞ ωd,rN d−r for r = s + 1, . . . , d and let us prove it for s. Rewriting (7) we get d Ad,N +1,s = Ad,N,s + (s + 1)λP ′(λ)Ad,N,s+1 + P (λ)r−s−1Ad,N,rQr,s(λ). Xr=s+2 We sum these equalities and use that Ad,N,r ∼N→+∞ ωd,rN d−r for r ≥ s + 1 to get Ad,N,s = (s + 1)λP ′(λ) The result follows now easily. N Xn=0 Ad,n,s+1 + O N Xn=0 nd−(s+2)! . (cid:3) 4.4. Proof of Lemma 4.3. Let u ∈ ℓ∞(N) which is orthogonal to all (cid:0)λk(cid:1) for λ ∈ Λ and let F (λ) = hu,(cid:0)λk(cid:1)i. Then F is a holomorphic function in D with an accumulation point of zeros inside D. Therefore, F and u are zero, which means that (cid:8)(cid:0)λk(cid:1) ; λ ∈ Λ(cid:9) spans a dense subspace in ℓ1(N). 5. Convolution operators, the case φ(0) > 1 Example 2.8 relies clearly on the periodicity of the zeros of the sine function. We do not know what happens for other examples of entire functions with φ(0) > 1, for instance for φ(z) = ez − λ with λ − 1 > 1. Nevertheless, we have a general result for the existence of powers of hypercyclic vectors. Theorem 5.1. Let X be an F-algebra and let T ∈ L(X). Assume that there exist a func- tion E : C → X and an entire function φ : C → C satisfying the following assumptions: (1) for all λ ∈ C, T E(λ) = φ(λ)E(λ); (2) for all λ, µ ∈ C, E(λ)E(µ) = E(λ + µ); (3) for all Λ ⊂ C with an accumulation point, span (E(λ); λ ∈ Λ) is dense in X; 16 FR´ED´ERIC BAYART (4) φ is not a multiple of an exponential function. Then there exists a residual set of vectors u ∈ X such that, for all m ≥ 1, um ∈ HC(T ). Proof. Let us set E = {u ∈ X; ∀m ≥ 1, um ∈ HC(T )}. Fixing (Vj)j≥1 a basis of open subsets of X, we set Oj,m =(cid:8)u ∈ X; ∃N ≥ 1, T N (um) ∈ Vj(cid:9) . Then E =Tj,m≥1 Oj,m so that, since each Oj,m is clearly open, one just has to prove that these sets are dense. Thus, let U, V be nonempty open subsets of X and let m ≥ 1. We are looking for a vector u ∈ U and for an integer N ≥ 1 such that T N (um) ∈ V . Let a ∈ C be such that φ(a) < 1. Arguing as in the proof of Corollary 2.4, we may find w0 ∈ C with φ(w0) > 1 and, for any z ∈ C with z − a ≤ m−1 m w0 − a, then φ(z) < 1. Then we fix δ > 0 sufficiently small so that φ(z) < 1 if z − a ≤ m−1 m w0 − a + δ and φ(w) > 1 if w ∈ B(w0, δ). Let finally, as usual(!), w1, w2 ∈ B(w0, δ) such that t ∈ [0, 1] 7→ log φ (tw1 + (1 − t)w2) is strictly convex. One may find p, q ∈ N, complex numbers a1, . . . , ap, b1, . . . , bq, complex numbers γ1, . . . , γp ∈ B(a, δ) and λ1, . . . , λq ∈ [w1, w2] such that p q Xl=1 alE (γl/m) ∈ U and bjE (λj) ∈ V. Xj=1 For N ≥ 1 and j ∈ {1, . . . , q}, let cj := cj(N ) be any complex number satisfying cm j = bj/(cid:0)φ(λj)(cid:1)N and define u := uN = alE (γl/m) + p Xl=1 cjE (λj/m) q Xj=1 so that m um = Xd=0 Xl∈I m−d p j∈I d q α(l, j, d, m)alcjE(cid:18) γl1 + · · · + γlm−d + λj1 + · · · + λjd m (cid:19) . We write γl = a + zl with zl < δ and λj = w0 + z′ j with z′ j < δ. Then γl1 + · · · + γlm−d + λj1 + · · · + λjd m = (m − d)a + dw0 m + Z with Z < δ. Moreover, provided m < d, = m m d m w0 − a ≤ (m − d)a + dw0 − a(cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12) φ(cid:16) γl1 +···+γlm−d +λj1 +···+λjd (cid:17)(cid:12)(cid:12)(cid:12) cj(N )T N E(cid:18) γl1 + · · · + γlm−d + λj1 + · · · + λjd < 1 and m Therefore, (cid:12)(cid:12)(cid:12) m − 1 m w0 − a . (cid:19) −→N→+∞ 0. When d = m, we conclude exactly as we have done before. (cid:3) HYPERCYCLIC ALGEBRAS 17 Proof of Theorem 1.3. That (i) or (ii) implies (iii) is already contained in [1]. The proof of (iii) implies (i) is easy if we observe that, for any λ ∈ C\{0}, a hypercyclic algebra for λφ(D) is a supercyclic algebra for φ(D). Finally, the implication (iii) implies (ii) is a consequence of Theorem 5.1 for T = φ(D). (cid:3) 6. Infinitely generated and dense hypercyclic algebras 6.1. Notations. We use several specific notations for this section. We denote by N the set of nonnegative integers and by N(∞) the set of sequences (α1, α2, . . . ) with αi ∈ N for i=1 αi. If A is a all i and αi = 0 for all large i. For α ∈ N(∞), α stands for the sum P+∞ finite subset of N(∞), A 6= ∅, then L(A) denotes sup{α; α ∈ A}. For f ∈ X N and α ∈ N(∞), with αi = 0 for i > d, the notation f α simply means the product f α1 d . If A is a finite subset of N(∞)\{(0, . . . )}, we will often consider it as 1 a subset of some Nd, since we may choose d ≥ 1 such that αi = 0 for all i ≥ d + 1 and all α ∈ A. · · · f αd 6.2. A criterion `a la Birkhoff for the existence of a dense and infinitely gener- ated hypercyclic algebra. To prove the existence of a dense and infinitely generated algebra of hypercyclic vectors, we need a reinforcement of Lemma 1.6. This is achieved by the following natural proposition, which simplifies a statement of [7] since it does not use the notion of pivot. Proposition 6.1. Let T be a continuous operator on the separable F -algebra X. Let ≺ be a total order on N(∞). Assume that, for any d ≥ 1, for any finite and nonempty subset A ⊂ N(∞)\{(0, . . . )}, for any nonempty open subsets U1, . . . , Ud, V of X, for any neighbourhood W of 0, there exist u = (u1, . . . , ud) ∈ U1 × · · · × Ud and N ≥ 1 such that, setting β = max(α; α ∈ A), T N (uβ) ∈ V T N (uα) ∈ W for all α ∈ A, α 6= β. Then T admits a dense and not finitely generated hypercyclic algebra. Proof. Let (Vk) be a basis of open neighbourhoods of X. For A ⊂ N(∞)\{(0, . . . )}, A 6= ∅, A finite, for s, k ≥ 1, define P (α)zα; P (βA) = 1 and P (α) ≤ s) E(A, s) =(P (z) = Xα∈A A(A, s, k) =nf ∈ X N; ∀P ∈ E(A, s), ∃N ≥ 1, T N (P (f )) ∈ Vko where βA = max(α; α ∈ A). The set A being fixed, A may be considered as a subset of Nd and E(A, s) as a subset of C[X1, . . . , Xd]. Moreover this set E(A, s) is compact. By continuity of the maps (f, P ) 7→ T N (P (f )), this implies that each set A(A, s, k) is open. Moreover, the assumptions of the theorem clearly imply that each such set is dense. Observe also that the set of f in X N that induce a dense algebra in X is residual in X N Hence, G :=TA,s,k A(A, s, k) is a residual subset of X N. (see [7]). Hence we may pick f ∈ X N belonging to TA,s,k A(A, s, k) and inducing a dense algebra in X. 18 FR´ED´ERIC BAYART We show that for all nonzero polynomials P , P (f ) belongs to HC(T ). Let A be the spectrum of P , let β = max(α; α ∈ A), let Q = 1 P and let s ≥ 1 be such that P (β) Q ∈ E(A, s). Since f belongs to Tk A(A, s, k), we conclude that Q(f ), hence P (f ), is a hypercyclic vector for T . It remains to show that the algebra generated by f is not finitely generated. Assume on the contrary that it is generated by a finite number of f α(1), . . . , f α(p). In particular, it is generated by a finite number of f1, . . . , fq. Then there exists a polynomial Q ∈ C[z1, . . . , zq] such that fq+1 = Q(f1, . . . , fq). Define P (z) = zq+1 − Q(z). Then P is a nonzero polynomial. Nevertheless, P (f ) = 0, which contradicts the fact that P (f ) is a hypercyclic vector for T . (cid:3) let Remark 6.2. The last part of the proof may be formulated in the following way: f ∈ X N be such that P (f ) is never zero for any nonzero polynomial P . Then the algebra generated by f is not finitely generated. We could avoid this by using the following lemma, proved in [7]: the set of sequences f ∈ X N whose induced algebra is not finitely generated is residual in X. Nevertheless, the proof of this last statement seems more complicated. 6.3. On the existence of infinitely generated hypercyclic algebras. We now show how to adapt our proofs to the existence of a dense and infinitely generated algebra. We have to pay the price of additional technical difficulties and we restrict ourselves to an analogue of Corollary 2.3. Theorem 6.3. Let X be an F-algebra and let T ∈ L(X). Assume that there exist a func- tion E : C → X and an entire function φ : C → C satisfying the following assumptions: (1) for all λ ∈ C, T E(λ) = φ(λ)E(λ); (2) for all λ, µ ∈ C, E(λ)E(µ) = E(λ + µ); (3) for all Λ ⊂ C with an accumulation point, span (E(λ); λ ∈ Λ) is dense in X; (4) φ is not a multiple of an exponential function; (5) for all ρ ∈ (0, 1), there exists w0 ∈ C with φ (w0) > 1 and, for all r ∈ (0, ρ], φ (rw0) < 1. Then T supports a hypercyclic algebra which is dense and is not finitely generated. Proof. We intend to apply Proposition 6.1. Thus, let d ≥ 1, let A be a finite and nonempty subset of N(∞)\{(0, . . . )}. Enlarging d if necessary, we may and shall assume that A ⊂ Nd. We choose for total order on Nd the lexicographical order and we denote by β the maximal element of A for this order. Without loss of generality, we assume that β1 6= 0. Let also U1, . . . , Ud, V be nonempty open subsets of X and let W be an open neighbourhood of 0. We set Iβ = {i ∈ {2, . . . , d}; βi 6= 0} ΩA = {α ∈ A; α1 = β1} \{β} ∪nα ∈ Nd; ∀i ∈ Iβ, αi ≤ βi and ∃i ∈ Iβ, αi < βio . We first consider the case Iβ 6= ∅ (the other case, which is easier, will be discussed at the end of the proof). Observe that, for any α ∈ ΩA, there exists i0 ∈ Iβ such that, for all HYPERCYCLIC ALGEBRAS 19 i ≤ i0, αi = βi and αi0 < βi0. Therefore it is easy to construct a sequence (ρi)i∈Iβ ⊂ (0, 1) satisfying Let η > 0 be such that Xi∈Iβ ∀α ∈ ΩA, Xi∈Iβ ∀α ∈ ΩA, Xi∈Iβ ρi = 1 ρi × αi βi < 1. ρi × αi βi ≤ 1 − η. We finally choose ε > 0 and ρ ∈ (0, 1) satisfying ρ > (1 − ε) β1 − 1 β1 + L(A)ε ρ > (1 − ε) + (1 − η)ε = 1 − ηε.   This is possible for instance by setting ρ = 1 − ηε/2 for a sufficiently small ε > 0. For this value of ρ, we get w0 ∈ C with φ(w0) > 1 and φ(rw0) < 1 if r ∈ (0, ρ]. We set κ = εw0 and z0 = (1 − ε)w0 = w0 − κ and we summarize some properties of w0, z0 and κ below: (8) (9) (10) φ(w0) > 1 φ(z0 + rκ) < 1 if r ∈ [0, 1 − η] φ(tz0 + sκ) < 1 if t ≤ β1 − 1 β1 and s ≤ L(A). By continuity, there exists δ > 0 such that these properties remain valid respectively in B(w0, δ), B(z0 +rκ, δ), B(tz0 +sκ, δ). Let (as usual!) w1 6= w2 in B(w0, δ/2) such that t 7→ log φ(tw1+(1−t)w2) is stricly convex on [0, 1]. We then choose integers p, q, complex num- bers a1,i, . . . , api,i, b1, . . . , bq, complex numbers γ1,i, . . . , γpi,i ∈ B(0, δ/2L(A)) ∩ (0, ρw0), complex numbers λ1, . . . , λq in [w1, w2] such that ∀i ∈ {1, . . . , d}, pi Xl=1 al,iE(γl,i) ∈ Ui and bjE(λj) ∈ V. q Xj=1 We define for j = 1, . . . , q, zj = λj − κ. We then set u1(N ) = al,1E(γl,1) + cj(N )E(zj /β1), q Xj=1 ui = al,iE(γl,i) + ωE(ρiκ/βi), p1 Xl=1 Xl=1 pi ui = pi Xl=1 al,iE(γl,i) provided i ∈ Iβ, 20 FR´ED´ERIC BAYART otherwise, namely if i /∈ Iβ and i > 1. Above, ω is any positive real number small enough so that all ui belong to Ui for i ≥ 2, and cj(N ) is any complex number such that cj(N )β1 = bj . Pi∈Iβ βi (cid:0)φ(λj)(cid:1)N ω Since cj(N ) goes to zero, u1(N ) belongs to U1 for N large enough. It remains to show that T N (uβ) ∈ V and T N (uα) ∈ W for α ∈ A\{β} and N large enough. Let us first compute uβ. Let us examine uβ1 1 . As in the proof of Theorem 2.1, we can distinguish three different kinds of terms: • the terms cj(N )β1E(zj ), for j = 1, . . . , q; • the terms cj(N )E(cid:16) zj1 +···+zjβ1 • a finite number of terms a(j, γ, N )E(cid:16) zj1 +···+zjl (cid:17) for j ∈ I β1 γ < β1δ/2L(A). β1 β1 q nondiagonal; + γ(cid:17) with j ∈ I l q, l < β1 and q for l > 0) but that they are uniformly bounded in N . Observe that the terms a(j, γ, N ) may depend on N by involving cj(N ) (this happens if j ∈ I l Let us now inspect uβ2 Pi∈Iβ d . For this product, we distinguish two kinds of terms: • the term ω • a finite number of terms b(r, γ′)E(rκ + γ′) with 0 ≤ r ≤ 1 − η and γ′ < (β2 + αiρi/βi for some α ∈ ΩA, which 2 · · · uβd βiE(κ); · · · + βd)δ/2L(A) ≤ δ/2. Here, r is equal to Pi∈Iβ explains why r ≤ 1 − η. Thus, taking the product uβ, we get six kinds of terms: • the terms cj(N )β1ω Pi∈Iβ βiE(λj) for j = 1, . . . , q. But the choice of cj(N ) ensures that cj(N )β1ω Pi∈Iβ q T N Xj=1  βiE(λj )  = bjE(λj). q Xj=1 • the terms cj(N )β1b(r, γ′)E(zj + rκ + γ′), with 1 ≤ j ≤ q, 0 ≤ r ≤ 1 − η and γ′ < δ/2. Since zj + rκ + γ′ − (z0 + rκ) < δ, we deduce from (9) that T N(cid:16)cj(N )β1b(r, γ′)E(zj + rκ + γ′)(cid:17) →N→+∞ 0. • the terms cj(N )ω Pi∈Iβ βiE(cid:18) λj1 +···+λjβ1 β1 (cid:19) for j ∈ I β1 proof of Theorem 2.1, the choice of [w1, w2] ensures that for all j ∈ I β1 q nondiagonal, q nondiagonal. But as in the T N(cid:18)cj(N )ω Pi∈Iβ βiE(cid:18) λj1 + · · · + λjβ1 β1 (cid:19)(cid:19) →N→+∞ 0. • the terms cj(N )b(r, γ′)E(cid:16) zj1 +···+zjβ1 β1 + rκ + γ′(cid:17). As before, (9) ensures that T N(cid:18)cj(N )b(r, γ′)E(cid:18) zj1 + · · · + zjβ1 β1 + rκ + γ′(cid:19)(cid:19) →N→+∞ 0. HYPERCYCLIC ALGEBRAS 21 • the terms a(j, γ, N )ω But writing Pi∈Iβ βiE(cid:16) zj1 +···+zjl β1 + κ + γ(cid:17) with l < β1, γ < β1δ/2L(A). zj1 + · · · + zjl β1 + κ = l β1 z0 + κ + γ′ with γ′ < δ, we deduce from (10) that T N(cid:18)a(j, γ, N )ω Pi∈Iβ βiE(cid:18) zj1 + · · · + zjl β1 + κ + γ(cid:19)(cid:19) →N→+∞ 0. and r ∈ [0, 1 − η]. A similar proof shows that • the terms a(j, γ, N )b(r, γ′)E(cid:16) zj1 +···+zjl T N(cid:18)a(j, γ, N )b(r, γ′)E(cid:18) zj1 + · · · + zjl + γ + rκ + γ′(cid:17) with l < β1, γ+γ′ < δ/2 + γ + rκ + γ′(cid:19)(cid:19) →N→+∞ 0. β1 β1 Hence, as expected, provided N is large enough, T N (u(N )β) ∈ V . Let us now consider α ∈ A with α ≺ β and let us show that T N (u(N )α) goes to zero as N goes to +∞. The analysis is similar but simpler. Either α1 = β1 and in that case in uα2 d appear only terms like b(r, γ)E(rκ + γ) with 0 ≤ r ≤ 1 − η and γ < (α2 + · · · + αd)δ/2L(A). Here, i ≤ αi so that r ≤ 1 − η since α ∈ ΩA. Then applying T N to α′ iρi/βi for some α′ 2 · · · uαd r =Pi∈Iβ each term of uα will lead to a sequence going to zero. Or α1 < β1, and now in uα1 1 appear only terms like a(j, l, N )E(cid:18) zj1 + · · · + zjt with t ≤ α1, namely terms like a(j, γ, N )E(cid:16) t product uα2 + γl1,1 + · · · + γlα1−t,1(cid:19) z0 + γ(cid:17) with γ < δ/2. Expanding the leads to terms like b(r, γ′)E(rκ + γ′) with γ′ < δ/2 and β1 β1 2 · · · uαd d ρi αi βi r ≤ Xi∈Iβ ≤ L(A). A last application of (10) shows that T N(cid:18)a(l, γ, N )b(r, γ′)E(cid:18) t β1 z0 + rκ + γ + γ′(cid:19)(cid:19) →N→+∞ 0. We need finally to consider the case Iβ = ∅. In that case, the proof of Theorem 2.1 works almost mutatis mutandis. Indeed, the situation is simplified because now uβ = uβ1 1 . We then set u1(N ) = ui = p1 pi Xl=1 Xl=1 al,1E(γl,1) + q Xj=1 β1(cid:19) cj(N )E(cid:18) λj al,iE(γl,i), i = 2, . . . , d with cj(N )β1 = bj/ (φ(λj))N and we follow a completely similar proof. Observe in partic- ular that if α ∈ A satisfies α ≺ β, then α1 < β1. (cid:3) 22 FR´ED´ERIC BAYART References [1] R. M. Aron, J. A. Conejero, A. Peris, and J. B. Seoane-Sep´ulveda. Powers of hypercyclic functions for some classical hypercyclic operators. Integral Equations Operator Theory, 58:591–596, 2007. [2] F. Bayart and ´E. Matheron. Dynamics of linear operators, volume 179 of Cambridge Tracts in Math. Cambridge University Press, 2009. [3] L. Bernal-Gonz´alez. On universal functions with zero-free derivatives. Arch. Math., 68:145–150, 1997. [4] J. B`es. Dynamics of composition operators with holomorphic symbol. Rev. Real Acad. Ciencias Exactas Fisicas y Naturales Serie A (RACSAM), 107:437–449, 2013. [5] J. B`es, A. Conejero, and D. Papathanasiou. Convolution operators supporting hypercyclic algebras. J. Math. Anal. Appl., 445:1232–1238, 2017. [6] J. B`es, A. Conejero, and D. Papathanasiou. Hypercyclic algebras for convolution and composition operators. ArXiv e-prints, June 2017. [7] J. B`es and D. Papathanasiou. Algebrable sets of hypercyclic vectors for convolution operators. ArXiv e-prints, June 2017. [8] R. P. Boas, Jr. Entire functions. Academic Press Inc., New York, 1954. [9] P. S. Bourdon. Invariant manifolds of hypercyclic vectors. Proc. Amer. Math. Soc., 118:845–847, 1993. [10] P. S. Bourdon and J. H. Shapiro. Cyclic phenomena for composition operators. Mem. Amer. Math. Soc., 125(596), 1997. [11] J. Falc´o and K-G. Grosse-Erdmann. Algebrability of the set of hypercyclic vectors for backward shift operators. Preprint, 2018. [12] G. Godefroy and J. H. Shapiro. Operators with dense, invariant, cyclic vector manifolds. J. Funct. Anal., 98:229–269, 1991. [13] M. Gonz´alez, F. Le´on-Saavedra, and A. Montes-Rodr´ıguez. Semi-Fredholm theory: hypercyclic and supercyclic subspaces. Proc. London Math. Soc. (3), 81:169–189, 2000. [14] Q. Menet. Hypercyclic subspaces and weighted shifts. Adv. in Math., 255:305–337, 2014. [15] S. Shkarin. On the set of hypercyclic vectors for the differentiation operator. Israel J. Math., 180:271– 283, 2010. Universit´e Clermont Auvergne, CNRS, LMBP, F-63000 Clermont–Ferrand, France. E-mail address: [email protected]
1508.02036	2	1508	2018-07-18T00:06:41	Functional calculus for $C_{0}$-groups using (co)type	[ "math.FA", "math.NA" ]	We study the functional calculus properties of generators of $C_{0}$-groups under type and cotype assumptions on the underlying Banach space. In particular, we show the following. Let $-iA$ generate a $C_{0}$-group on a Banach space $X$ with type $p\in[1,2]$ and cotype $q\in[2,\infty)$. Then $A$ has a bounded $\mathcal{H}^{\infty}$-calculus from $\mathrm{D}_{A}(\tfrac{1}{p}-\tfrac{1}{q},1)$ to $X$, i.e. $f(A):\mathrm{D}_{A}(\tfrac{1}{p}-\tfrac{1}{q},1)\to X$ is bounded for each bounded holomorphic function $f$ on a sufficiently large strip. As a corollary of our main theorem, for sectorial operators we quantify the gap between bounded imaginary powers and a bounded $\mathcal{H}^{\infty}$-calculus in terms of the type and cotype of the underlying Banach space. For cosine functions we obtain similar results as for $C_{0}$-groups. We extend our results to $R$-bounded operator-valued calculi, and we give an application to the theory of rational approximation of $C_{0}$-groups.	math.FA	math	FUNCTIONAL CALCULUS FOR C0-GROUPS USING TYPE AND COTYPE JAN ROZENDAAL Abstract. We study the functional calculus properties of generators of C0- groups under type and cotype assumptions on the underlying Banach space. In particular, we show the following. Let −iA generate a C0-group on a Banach space X with type p ∈ [1, 2] and cotype q ∈ [2, ∞). Then f (A) : → X is bounded for each bounded holomorphic function f (X, D(A)) 1 p q ,1 − 1 on a sufficiently large strip. As a corollary of this result, for sectorial opera- tors we quantify the gap between bounded imaginary powers and a bounded H∞-calculus in terms of the type and cotype of the underlying Banach space. For cosine functions we obtain similar results as for C0-groups. We extend our theorems to R-bounded operator-valued calculi, and we give an application to the theory of rational approximation of C0-groups. 1. Introduction Let −iA generate a C0-group (U (s))s∈R ⊆ L(X) on a Banach space X, and set θ(U ) := infnθ ≥ 0(cid:12)(cid:12)(cid:12)∃M ≥ 1 : kU (s)k ≤ M eθs for all s ∈ Ro . (1.1) For each ω > 0 let Stω := {z ∈ CIm(z) > ω} , (1.2) and let H∞(Stω) be the space of bounded holomorphic functions on Stω with the supremum norm. There is a natural definition of f (A) as an unbounded operator on X for each ω > θ(U ) and each f ∈ H∞(Stω) (see Section 2.1). It was shown by Boyadzhiev and deLaubenfels in [9] that, if X is a Hilbert space, then there exists a constant C ≥ 0 such that f (A) ∈ L(X) with (1.3) for each ω > θ(U ) and f ∈ H∞(Stω). One says that A has a bounded H∞-calculus. It is useful to know that an operator has a bounded H∞-calculus. The theory of H∞-calculus has applications to questions of maximal regularity (see [2, 35, 38]) and played a crucial role in the solution to the Kato square root problem in [4, 6]. Also, functional calculus bounds can be used to determine convergence rates in the numerical approximation theory for solutions to evolution equations (see [10,17,21]). One can obtain nontrivial functional calculus results for C0-groups even when X is not a Hilbert space. For example, it was shown in [25] that if −iA generates a kf (A)kL(X) ≤ C kfkH∞(Stω) 2010 Mathematics Subject Classification. Primary 47A60; Secondary 47D03, 46B20, 42A45. Key words and phrases. Functional calculus, Operator group, Type and cotype, Transference, R-boundedness. 1 2 JAN ROZENDAAL C0-group (U (s))s∈R on a UMD space X, then A has a bounded H∞1 (Stω)-calculus for all ω > θ(U ). Here H∞1 (Stω) consists of all f ∈ H∞(Stω) such that (1.4) kfkH∞ 1 (Stω ) := sup z∈Stω f (z) + (1 + z)f′(z) < ∞. In [29] a similar statement was obtained on general Banach spaces, where one restricts the calculus to real interpolation spaces between X and the domain of A. However, there are several drawbacks to functional calculus theory for function spaces which are strictly contained in the class of H∞-functions. For example, in applications one might be interested in functions f which are not contained in a smaller function space such as H∞1 (Stω). But even when dealing with a function f which is known to yield a bounded operator f (A), it can be of interest to know that (1.3) holds, instead of an estimate for kf (A)kL(X) with respect to a larger function norm. This is the case for numerical approximation schemes for solutions to evolution equations, where one can often improve convergence rates for operators with a bounded H∞-calculus (see [17, 20]). Moreover, a bounded H∞-calculus yields square function estimates that arise frequently in harmonic analysis (see [22, 36, 40]) and that are of use in the theory of stochastic evolution equations [50, 51]. Also, a bounded H∞-calculus can be bootstrapped to yield large classes of R-bounded operators, and the notion of R- boundedness has various applications in the theory of evolution equations (see [38]). It seems that such connections are not available in full generality for subspaces of the class of H∞-functions (if A is self-adjoint then more can be said; see for example [5]). Finally, it seems somewhat unsatisfactory that there is such a rough division between functional calculus properties on Hilbert spaces and on UMD spaces, in the sense that one goes from a bounded H∞-calculus on Hilbert spaces to a bounded H∞1 -calculus on UMD spaces. One might expect a finer division of functional calculus theorems within the class of UMD spaces, and on Lp-spaces one might hope that functional calculus properties improve as p tends to 2. The latter is indeed the case for symmetric contraction semigroups on Lp-spaces, cf. [12, 14]. In the present article we aim to address the issues above. We study H∞-calculus for generators of C0-groups in terms of the type and cotype of the underlying Banach space (see Definition 2.3). Our main result, proved as Theorem 4.1, is as follows. (For the real interpolation space DA( 1 Theorem 1.1. Let −iA generate a C0-group (U (s))s∈R ⊆ L(X) on a Banach space X with type p ∈ [1, 2] and cotype q ∈ [2,∞). Let ω > θ(U ). Then there exists a constant C ≥ 0 such that DA( 1 q , 1) see (1.7)). p − 1 p − 1 q , 1) ⊆ D(f (A)) and kf (A)xkX ≤ C kfkH∞(Stω ) kxkDA( 1 p− 1 q ,1) for all f ∈ H∞(Stω) and x ∈ DA( 1 p − 1 q , 1). Under additional geometric assumptions we improve Theorem 1.1. Indeed, in Theorem 4.3 we show that, if X is isomorphic to a complemented subspace of a p-convex and q-concave Banach lattice for p ∈ [1, 2] and q ∈ [2,∞), then for each λ > ω > θ(U ) there exists a constant C ≥ 0 such that D((λ+ iA)− 1 q ) ⊆ D(f (A)) p + 1 and (1.5) kf (A)xkX ≤ C kfkH∞(Stω) k(λ + iA)− 1 p + 1 q xkX for all f ∈ H∞(Stω) and x ∈ D((λ + iA)− 1 p + 1 q ). FUNCTIONAL CALCULUS FOR C0-GROUPS USING TYPE AND COTYPE 3 One might say that Theorem 1.1 shows that each generator −iA of a C0-group on a Banach space X with type p and cotype q has a bounded H∞-calculus from q , 1) to X, and A has a bounded H∞-calculus from D((λ + iA)− 1 p − 1 DA( 1 q ) to X under additional assumptions on X. Since 1 q = 0 if and only if X is isomorphic to an L2-space, (1.5) recovers as a special case (1.3), the main result of [9]. p − 1 p + 1 The interpolation spaces DA(θ, 1) and the fractional domains D((λ + iA)−θ) increase as θ tends to zero. Therefore the statements in this paper show that many functional calculus properties depend in a quantitative manner on how close the geometry of the underlying space is to that of a Hilbert space. Since each UMD space has non-trivial type and finite cotype, Theorem 1.1 provides a scale of functional calculus results on UMD spaces. p − 1 q ≥ 1 q is attained. For (Ω, µ) a measure space and p ∈ [1,∞), the space Lp(Ω, µ) is a p-convex and p-concave Banach lattice with type min(p, 2) and cotype max(p, 2). Hence our results show that the functional calculus properties of group generators on Lp- spaces improve as p tends to 2. It should be noted that (1.5) holds without any assumptions on the Banach space if 1 2 (see Remark 4.8). Therefore most of the results in Sections 4 and 5 are only of interest on Lp-spaces when p ∈ (1,∞). In Proposition 4.6 we deduce from Theorem 1.1 that each f ∈ H∞(Stω) with polynomial decay of order α > 1 p − 1 q at infinity satisfies f (A) ∈ L(X). Under the assumptions of (1.5) the case α = 1 p − 1 It was shown by McIntosh in [43] that, on Hilbert spaces, sectorial operators with bounded imaginary powers have a bounded sectorial H∞-calculus. It is also known that on general Banach spaces (even on Lp-spaces) an operator with bounded imag- inary powers need not have a bounded H∞-calculus. In Theorem 5.1 we quantify the gap between bounded imaginary powers and a bounded H∞-calculus, by show- ing that each sectorial operator A with bounded imaginary powers has a bounded sectorial H∞-calculus from Dlog(A)( 1 q , 1) to X if the underlying Banach space X has type p ∈ [1, 2] and cotype q ∈ [2,∞). As in the classical case of a bounded H∞-calculus on X, one obtains from this an unconditionality result and square function estimates. For generators of cosine functions we derive similar results as for C0-groups. In particular, in Theorem 5.4 we show that each generator −A of a cosine function on a Banach space X with type p and cotype q has a bounded parabola-type H∞- calculus from DA( 1 We extend Theorem 1.1 and (1.5) to operator-valued functional calculi. Then in Theorem 6.2 we show that A in fact has an R-bounded H∞-calculus from DA( 1 p − 1 q , 1) to X if X additionally has property (α). In Theorem 6.3 we obtain an R- bounded version of (1.5). These results are sharp, as Example 6.5 shows. q ), 1) to X. p − 1 2 ( 1 p − 1 To indicate the use of Theorem 1.1 we give an application to the study of numer- ical approximation methods for the solutions to evolution equations. In Proposition 7.3 we consider a method of rational approximation proposed in [34, 44], and for exponentially stable groups on Lp-spaces we improve the rates of convergence which were obtained in [17]. In Corollary 7.4 we show that many rational approximation methods of exponentially stable C0-groups converge strongly on DA( 1 q , 1) if the underlying space X has type p ∈ [1, 2] and cotype q ∈ [2,∞). To prove Theorem 1.1 we use transference principles going back to [8,11,13]. The transference technique has been applied to functional calculus theory in [14], [30] p − 1 4 JAN ROZENDAAL and [25, 26, 28, 29]. In [29] an interpolation version of the transference principle for unbounded groups from [25] was established. This transference principle was then combined with a theorem about Fourier multipliers on vector-valued Besov spaces. In the present paper we also obtain a transference principle involving vector-valued Besov spaces, but we combine it with a theorem about Fourier multipliers between distinct Besov spaces. For (1.5) we use statements about Fourier multipliers from Lp to Lq which were obtained recently in [47, 48]. We suspect that it is possible to apply these Fourier multiplier theorems to the transference principle from [28] for C0-semigroups, but this matter will not be explored in the present article. This paper is organized as follows. In Section 2 we discuss some of the basics of functional calculus theory, vector-valued Besov spaces and the notions of type and cotype. We then state two Fourier multiplier results which are vital for that which follows. In Section 3 we establish two new transference principles for C0-groups. These are used in Section 4 to prove Theorem 1.1 and (1.5), as well as several corollaries for C0-groups. In Section 5 we obtain results for sectorial operators and generators of cosine functions, and in Section 6 we extend the results from previous sections to R-bounded operator-valued calculi. Finally, in Section 7 we give an application to the theory of rational approximation schemes. 1.1. Notation and terminology. We write N := {1, 2, 3, . . .} for the natural numbers and N0 := N ∪ {0}. We write C+ for the open right half-plane of complex numbers z ∈ C with Re(z) > 0, and C− := C \ C+. We denote nonzero Banach spaces over the complex numbers by X and Y . The space of bounded linear operators from X to Y is L(X, Y ), and L(X) := L(X, X). The domain of a closed operator A on X is D(A), a Banach space with the norm kxkD(A) := kxkX + kAxkX (x ∈ D(A)). The spectrum of A is σ(A) and the resolvent set ρ(A) := C \ σ(A). We write R(λ, A) := (λ − A)−1 for the resolvent operator of A at λ ∈ ρ(A). For p ∈ [1,∞] and (Ω, µ) a measure space, Lp(Ω; X) denotes the Bochner space of equivalence classes of p-integrable X-valued functions on Ω. We often write k·kp = k·kLp(R;C). For n ∈ N0 we let W n,p(R; X) be the Sobolev space of n times weakly differentiable f ∈ Lp(R; X) such that f (n) ∈ Lp(R; X). The Holder conjugate of p is denoted by p′ and is defined by 1 = 1 We let M(R) denote the space of complex Borel measures on R with the total variation norm. For ω ≥ 0 we let Mω(R) be the convolution algebra of µ ∈ M(R) of the form µ(ds) = e−ωsν(ds) for some ν ∈ M(R), with kµkMω(R) := keω·µkM(R). For Ω 6= ∅ open in C, we denote by H∞(Ω) the space of bounded holomorphic functions f : Ω → C, a Banach algebra with the norm p + 1 p′ . kfkH∞(Ω) := sup z∈Ωf (z) (f ∈ H∞(Ω)). The class of X-valued rapidly decreasing smooth functions on R is S(R; X), and the space of X-valued tempered distributions is S′(R; X). The Fourier transform of Φ ∈ S′(R; X) is F Φ. If µ ∈ Mω(R) for ω > 0 then F µ ∈ H∞(Stω) is given by F µ(z) :=ZR e−izsµ(ds) (z ∈ Stω). An interpolation couple is a pair (X, Y ) of Banach spaces which are embedded continuously in a Hausdorff topological vector space Z. The real interpolation FUNCTIONAL CALCULUS FOR C0-GROUPS USING TYPE AND COTYPE 5 space of (X, Y ) with parameters θ ∈ [0, 1] and q ∈ [1,∞] is denoted by (X, Y )θ,q. If T : X + Y → X + Y restricts to a bounded operator on X and Y then (1.6) kTkL((X,Y )θ,q ) ≤ kTk1−θ L(X) kTkθ L(Y ) for all θ ∈ (0, 1) and q ∈ [1,∞]. We mainly consider real interpolation spaces for the interpolation couple (X, D(A)), where A is a closed operator on X. We write (1.7) DA(θ, q) := (X, D(A))θ,q and kxkθ,q := kxkDA(θ,q) (x ∈ DA(θ, q)). For an operator B on X and a continuously embedded space Y ֒→ X, the part of B in Y is the operator BY on Y that satisfies BY y = By for y ∈ D(BY ) := {z ∈ D(B) ∩ Y Bz ∈ Y }. We write Bθ,q := BDA(θ,q) for θ ∈ [0, 1] and q ∈ [1,∞]. 2. Functional calculus and Fourier multipliers In this section we present the background on functional calculus and Fourier multipliers which will be needed for the rest of the article. 2.1. Functional calculus. We assume that the reader is familiar with the basics of the theory of C0-groups from [19]. For more on the functional calculus for generators of C0-groups see [23, Chapter 4]. An operator A on a Banach space X is a strip-type operator of height ω0 ≥ 0 if σ(A) ⊆ Stω0 , where St0 := R, and supλ∈C\Stω kR(λ, A)k < ∞ for all ω > ω0. For ω > 0 set E(Stω) :=(cid:8)g ∈ H∞(Stω)(cid:12)(cid:12)g(z) ∈ O(z−α) for some α > 1 as Re(z) → ∞(cid:9) . The strip-type functional calculus for a strip-type operator A of height ω0 is defined as follows. First, operators f (A) ∈ L(X) are associated with f ∈ E(Stω) for ω > ω0: (2.1) f (z)R(z, A) dz. f (A) := 1 2πiZ∂Stω′ Here ∂Stω′ is the positively oriented boundary of Stω′ for ω′ ∈ (ω0, ω). This proce- dure is independent of the choice of ω′ by Cauchy's theorem, and yields an algebra homomorphism E(Stω) → L(X), f 7→ f (A). The definition of f (A) is extended to a larger class of functions by regularization: f (A) := e(A)−1(ef )(A) (2.2) if there exists an e ∈ E(Stω) with e(A) injective and ef ∈ E(Stω). Then f (A) is a closed unbounded operator on X, and the definition of f (A) is independent of the choice of the regularizer e. Each f ∈ H∞(Stω) is regularizable by the function z 7→ (λ − z)−2 for Im(λ) > ω. Let −iA generate a C0-group (U (s))s∈R ⊆ L(X). Then A is a strip-type operator of height θ(U ), with θ(U ) as in (1.1). Let M ≥ 1 and ω ≥ 0 be such that kU (s)k ≤ M eωs for all s ∈ R, and for µ ∈ Mω(R) set (2.3) U (s)x dµ(s) Uµx :=ZR (x ∈ X). The mapping µ 7→ Uµ is an algebra homomorphism Mω(R) → L(X) called the Hille-Phillips calculus, and the following lemma from [24, Lemma 2.2] shows that this calculus is consistent with the strip-type calculus for A. 6 JAN ROZENDAAL Lemma 2.1. Let −iA generate a C0-group (U (s))s∈R ⊆ L(X) on a Banach space X, and let ω > α > θ(U ). Then each f ∈ E(Stω) satisfies f = F µ and f (A) = Uµ ∈ L(X) for some µ ∈ Mα(R). Conversely, if µ ∈ Mα(R) then f := F µ ∈ H∞(Stα) is such that f (A) = Uµ ∈ L(X). In fact, it is observed in Remark 4.8 that the first statement in Lemma 2.1 holds for all f ∈ H∞(Stω) such that f (z) = O(z−1/2) as Re(z) → ∞. The following is a version of this lemma adapted to our setting. A fundamental result in functional calculus theory is the Convergence Lemma. Lemma 2.2 (Convergence Lemma). Let A be a densely defined strip-type operator of height ω0 ≥ 0 on a Banach space X. Let Y be a Banach space continuously embedded in X such that D(A2) ⊆ Y is dense. Let ω > ω0 and let (fj)j∈J ⊆ H∞(Stω) be a net satisfying the following conditions: • supj∈J kfjkH∞(Stω ) < ∞; • f (z) := limj fj(z) exists for all z ∈ Stω; • supj∈J kfj(A)kL(Y,X) < ∞. Then f ∈ H∞(Stω), f (A) ∈ L(Y, X), fj(A)x → f (A)x for all x ∈ Y and kf (A)kL(Y,X) ≤ lim sup kfj(A)kL(Y,X) . j∈J Proof. The proof is similar to the proofs of [23, Proposition 5.1.7] and [7, Theorem 3.1]. Vitali's theorem for nets from [3, Theorem 2.1] implies that f ∈ H∞(Stω) and that fj → f uniformly on compact subsets of Stω. Let λ > ω. Applying the dominated convergence theorem to (2.1) yields fj(A)x =(cid:18) fj(·) (iλ − ·)2(cid:19) (A)(iλ − A)2x →(cid:18) f (·) (iλ − ·)2(cid:19) (A)(iλ − A)2x = f (A)x and kf (A)xkX ≤ lim sup j kfj(A)kL(Y,X) kxkY for all x ∈ D(A2). The required statements now follow since D(A2) ⊆ Y is dense and f (A) is a closed operator on X. (cid:3) 2.2. Fourier multipliers and Banach space geometry. In this section we treat Fourier multiplier operators under geometric assumptions on the underlying space. For more on the prerequisite notions from Banach space geometry, as well as for proofs of some of the statements below, see e.g. [15, 31, 32, 41]. Recall that a stan- dard complex Gaussian random variable is a random variable γ on a probability space (Ω, P) such that γ = γr +iγi√2 for independent standard real Gaussian random variables γr, γi on Ω. A Gaussian sequence is a sequence of independent standard complex Gaussian random variables. Definition 2.3. Let X be a Banach space, (γk)k∈N a Gaussian sequence on a probability space (Ω, P) and p ∈ [1, 2], q ∈ [2,∞]. • X has (Gaussian) type p if there exists a constant C ≥ 0 such that for all n ∈ N and x1, . . . , xn ∈ X, (2.4) (cid:16)E(cid:13)(cid:13)(cid:13) nXk=1 γkxk(cid:13)(cid:13)(cid:13) 2(cid:17)1/2 ≤ C(cid:16) nXk=1 kxkkp(cid:17)1/p . FUNCTIONAL CALCULUS FOR C0-GROUPS USING TYPE AND COTYPE 7 (2.5) • X has (Gaussian) cotype q if there exists a constant C ≥ 0 such that for all n ∈ N and x1, . . . , xn ∈ X, kxkkq(cid:17)1/q nXk=1 with the obvious modification for q = ∞. ≤ C(cid:16)E(cid:13)(cid:13)(cid:13) (cid:16) nXk=1 2(cid:17)1/2 γkxk(cid:13)(cid:13)(cid:13) , The minimal constants C in (2.4) and (2.5) are called the Gaussian type p con- stant and the Gaussian cotype q constant and will be denoted by τp,X and cq,X . We say that X has nontrivial type if X has type p ∈ (1, 2] and that X has finite cotype if X has cotype q ∈ [2,∞). By the Kahane-Khintchine inequalities, one may replace the exponent 2 in (2.4) and (2.5) by any r ∈ [1,∞). This does not change the properties of type and cotype, only the minimal constants in (2.4) and (2.5). It is common to replace the Gaussian sequence in Definition 2.3 by a Rademacher sequence, i.e. a sequence (rk)k∈N of independent complex random variables on a probability space (Ω, P) that are uniformly distributed on {z ∈ C z = 1}. This does not change the class of spaces under consideration, only the minimal constants in (2.4) and (2.5). We choose to work with Gaussian sequences because the constants τp,X and cq,X occur in Proposition 2.4. Every Banach space X has type p = 1 and cotype q = ∞, with τ1,X = c∞,X = 1. If X has type p and cotype q then it has type r with τr,X ≤ τp,X for all r ∈ [1, p] and cotype s with cs,X ≤ cq,X for all s ∈ [q,∞]. A Banach space X has type p = 2 and cotype q = 2 if and only if X is isomorphic to a Hilbert space, by Kwapie´n's result [39]. A Banach space X with nontrivial type has finite cotype. Each UMD space has nontrivial type. Let X be a Banach space with type p ∈ [1, 2] and cotype q ∈ [2,∞), let r ∈ [1,∞) and let Ω be a measure space. Then Lr(Ω; X) has type min(p, r) and cotype max(q, r) with τmin(p,r),Lr(Ω;X) ≤ Cp,rτp,X and cmax(q,r),Lr(Ω;X) ≤ Cq,rcq,X for constants Cp,r, Cq,r ≥ 0 coming from the Kahane-Khintchine inequalities. Let ψ ∈ C∞(R) be such that ψ ≥ 0, supp(ψ) ⊆ [ 1 ψ(2−ks) = 1 for all s ∈ (0,∞). For k ∈ N and s ∈ R let ϕk(s) := ψ(2−ks) and ϕ0(s) := 1 −P∞k=1 ϕk(s). Let X be a Banach space and let p, q ∈ [1,∞] and r ∈ R. The p,q(R; X) is the space of all f ∈ S′(R; X) such that (inhomogeneous) Besov space Br 2 , 2] and P∞k=−∞ kfkBr p,q(R;X) :=(cid:13)(cid:13)(cid:13)(cid:16)2kr(cid:13)(cid:13)F−1ϕk ∗ f(cid:13)(cid:13)Lp(R;X)(cid:17)k∈N0(cid:13)(cid:13)(cid:13)ℓq < ∞, p,q(R;X). Then Br endowed with the norm k·kBr p,q(R; X) is a Banach space, S(R; X) ⊆ Br p,q(R; X) is dense if p, q < ∞, and a different choice of ψ yields an equivalent norm on Br p,q(R; X). More details on vector-valued Besov spaces can be found in [1, 49]. For X a Banach space and m ∈ L∞(R;L(X)), the Fourier multiplier operator Tm : S(R; X) → S′(R; X) with symbol m is given by Tm(f ) := F−1 (m · F f ) (f ∈ S(R; X)). For each µ ∈ M(R), (2.6) defines a Fourier multiplier operator with symbol F µ ∈ L∞(R). We say that X is a UMD space if the function 1[0,∞) − 1(−∞,0) is the symbol of a bounded Fourier multiplier on L2(R; X). Lµ(f ) := µ ∗ f (f ∈ S(R; X)) 8 JAN ROZENDAAL Let X and Y be Banach spaces and let (rk)k∈N be a Rademacher sequence on a probability space (Ω, P). A collection T ⊆ L(X, Y ) is R-bounded if there exists a constant C ≥ 0 such that, for all n ∈ N, x1, . . . , xn ∈ X and T1, . . . , Tn ∈ T , (2.7) . nXk=1 (cid:16)E(cid:13)(cid:13)(cid:13) rkTkxk(cid:13)(cid:13)(cid:13) 2 Y(cid:17)1/2 ≤ C(cid:16)E(cid:13)(cid:13)(cid:13) nXk=1 rkxk(cid:13)(cid:13)(cid:13) The minimal constant C in (2.7) is denoted by R(T ). If we want to specify the underlying spaces X and Y then we write RX,Y (T ) = R(T ), and we write RX (T ) = R(T ) if T ⊆ L(X). By the Kahane contraction principle, each uniformly bounded collection T ⊆ C is R-bounded as a subset of L(X), with RX (T ) equal to the uniform bound of T . The following result was obtained in [47, Theorem 1.1]. 2 X(cid:17)1/2 Proposition 2.4. Let X be a Banach space with type p ∈ [1, 2] and cotype q ∈ [2,∞]. Let r ∈ R, s ∈ [1,∞] and m ∈ L∞(R;L(X)) be such that {m(s) s ∈ R} ⊆ L(X) is R-bounded. Then Tm ∈ L(Br+1/p−1/q p − 1 q,s(R; X)) and (R; X), Br p,s 1 q,s(R;X)) ≤ 4 q τp,X cq,X RX ({m(s) s ∈ R}). kTmkL(Br+1/p−1/q p,s (R;X),Br Corollary 2.5. Let X be a Banach space with type p ∈ [1, 2] and cotype q ∈ [2,∞], and let µ ∈ M(R). Then Lµ ∈ L(B1/p−1/q (R;X),Lq(R;X)) ≤ 4 q τp,X cq,X kF µkL∞(R) . (R; X), Lq(R; X)) and kLµkL(B1/p−1/q p− 1 p,1 p,1 1 Proof. By the Kahane contraction principle, RX ({F µ(s) s ∈ R}) = kF µkL∞(R). q,1(R; X) is contractively embedded in Lq(R; X), the result follows by ap- Since B0 plying Proposition 2.4 to Lµ = TF µ. (cid:3) Remark 2.6. If in Corollary 2.5 one assumes additionally that X is a UMD space, then Lµ ∈ L(B1/p−1/q p,p (2.8) (R; X), Lq(R; X)) and kLµkL(B1/p−1/q p,p (R;X),Lq(R;X)) ≤ Cτp,X cq,X kF µkL∞(R) for each µ ∈ M(R) and some C ≥ 0 independent of µ. This follows from Proposition 2.4 and the embedding B0 q,p(R; X) ⊆ Lq(R; X) from [52, Proposition 3.1]. We assume that the reader is familiar with the basics of Banach lattices from [41]. Definition 2.7. Let X be a Banach lattice and p, q ∈ [1,∞]. • X is p-convex if there exists a constant C ≥ 0 such that for all n ∈ N and x1, . . . , xn ∈ X, (cid:13)(cid:13)(cid:13)(cid:16) nXk=1 xkp(cid:17)1/p(cid:13)(cid:13)(cid:13) ≤ C(cid:16) nXk=1 kxkkp(cid:17)1/p , • X is q-concave if there exists a constant C ≥ 0 such that for all n ∈ N and with the obvious modification for p = ∞. x1, . . . , xn ∈ X, (cid:16) nXk=1 kxkkq(cid:17)1/q ≤ C(cid:13)(cid:13)(cid:13)(cid:16) nXk=1 xkq(cid:17)1/q(cid:13)(cid:13)(cid:13), with the obvious modification for q = ∞. FUNCTIONAL CALCULUS FOR C0-GROUPS USING TYPE AND COTYPE 9 Every Banach lattice X is 1-convex and ∞-concave. If X is p-convex and q- concave then it is r-convex and s-concave for all r ∈ [1, p] and s ∈ [q,∞]. By [41, Proposition 1.f.3], if X is q-concave then it has cotype max(q, 2), and if X is p-convex and q-concave for some q < ∞ then X has type min(p, 2). For (Ω, µ) a measure space and r ∈ [1,∞], Lr(Ω, µ) is an r-convex and r-concave Banach lattice. A subspace X0 ⊆ Y of a Banach space Y is said to be complemented if there exists a projection P ∈ L(Y ) with P (Y ) = X0. Proposition 2.8. Let X be isomorphic to a complemented subspace of a p-convex and q-concave Banach lattice, for p ∈ [1, 2] and q ∈ [2,∞). Then there exists a constant C ≥ 0 such that Tm ∈ L(Lp(R; X), Lq(R; X)) with (2.9) p − 1 1 q m(s) s ∈ R}) kTmkL(Lp(R;X),Lq(R;X)) ≤ CRX ({s p − 1 1 for each m ∈ L∞(R;L(X)) such that RX ({s Proof. Let X0 be a subspace of a p-convex and q-concave Banach lattice Y , S : X → X0 an isomorphism and P ∈ L(Y ) a projection with P (Y ) = X0. Let m ∈ L∞(R;L(X)) be such that RX ({s q m(s) s ∈ R}) < ∞, and let m0 ∈ L∞(R;L(Y )) be given by m0(s) := Sm(s)S−1P ∈ L(Y ) for s ∈ R. Then q m(s) s ∈ R}) < ∞. p − 1 1 RY ({s 1 p − 1 q m0(s) s ∈ R}) ≤ kSkkS−1kkPkRX({s 1 p− 1 q m(s) s ∈ R}). It follows from [48, Theorem 3.21] that there exists a constant C ≥ 0 independent of m such that Tm0 ∈ L(Lp(R; Y ), Lq(R; Y )) with kTm0kL(Lp(R;Y ),Lq(R;Y )) ≤ CRY ({s This concludes the proof since S−1Tm0S = Tm. q m0(s) s ∈ R}). p− 1 (cid:3) 1 3. Transference principles In this section we establish two new transference principles for C0-groups, both based on the transference principles for unbounded groups from [25] and [29]. For ω ≥ 0 and µ ∈ Mω(R) let µω ∈ M(R) be given by (3.1) and note that (3.2) µω(ds) := cosh(ωs)µ(ds), F µω(s) = F µ(s + iω) + F µ(s − iω) 2 for all s ∈ R. Proposition 3.1. Let ω > ω0 ≥ 0, p ∈ [1, 2] and q ∈ [2,∞). Then there exists a constant C ≥ 0 such that the following holds. Let X be a Banach space with type p and cotype q, and let −iA generate a C0-group (U (s))s∈R ⊆ L(X) such that kU (s)kL(X) ≤ M cosh(ω0s) for all s ∈ R and some M ≥ 1. Then ZR (cid:13)(cid:13)(cid:13)(cid:13) U (s)x µ(ds)(cid:13)(cid:13)(cid:13)(cid:13)X ≤ Cτp,X cq,X M 2kF µωkL∞(R)kxk1/p−1/q,1 for all µ ∈ Mω(R) and all x ∈ DA( 1 p − 1 q , 1). 10 JAN ROZENDAAL p − 1 q ∈ (0, 1) (for 1 Proof. Since DA(0, 1) = {0} we may assume that 1 stronger result is obtained in Proposition 3.3 below). Let √8ω π cosh(ωs) cosh(2ωs) cosh(2ωs) ψ(s) := ϕ(s) := (3.3) and 1 p − 1 q = 0 a for s ∈ R. Define ι : X → Lp(R; X) by ιx(s) := ψ(−s)U (−s)x (3.4) and P : Lq(R; X) → X by P f :=ZR (3.5) ϕ(s)U (s)f (s) ds Then ι is bounded and (x ∈ X, s ∈ R) (f ∈ Lq(R; X)). (3.6) kιkL(X,Lp(R;X)) ≤ M kψ(·) cosh(ω0·)kp . By Holder's inequality, P is bounded and (3.7) kPkL(Lq(R;X),X) ≤ M kϕ(·) cosh(ω0·)kq′ . Let x ∈ D(A). Then ιx ∈ C1(R; X) and (ιx)′(s) = −ψ′(−s)U (−s)x + iψ(−s)U (−s)Ax = −2ω tanh(2ωs) cosh(2ωs) U (−s)x + i 1 cosh(2ωs) U (−s)Ax for all s ∈ R. Hence (ιx)′ ∈ Lp(R; X) and cosh(ω0·) k(ιx)′kp ≤ 2ωM ktanhkL∞(R)(cid:13)(cid:13)(cid:13)(cid:13) cosh(2ω·)(cid:13)(cid:13)(cid:13)(cid:13)pkxkX + M(cid:13)(cid:13)(cid:13)(cid:13) Now (3.6) implies that ιx ∈ W 1,p(R; X), with cosh(ω0·) cosh(2ω·)(cid:13)(cid:13)(cid:13)(cid:13)pkAxkX . cosh(ω0·) Hence ι : D(A) → W 1,p(R; X) is bounded and (3.8) kιxk1,p ≤ M (2ω ktanhkL∞(R) + 1)(cid:13)(cid:13)(cid:13)(cid:13) kιkL(D(A),W 1,p(R;X)) ≤ M (2ω ktanhkL∞(R) + 1)(cid:13)(cid:13)(cid:13)(cid:13) (R; X) =(cid:0)Lp(R; X), W 1,p(R; X)(cid:1) 1 cosh(2ω·)(cid:13)(cid:13)(cid:13)(cid:13)pkxkD(A) . cosh(2ω·)(cid:13)(cid:13)(cid:13)(cid:13)p By equation (5.9) in [1], cosh(ω0·) B1/p−1/q p − 1 q ,1 p,1 . with equivalent norms. Moreover, it follows from a direct sum argument that the constant in the norm equivalence does not depend on X. Hence (1.6), (3.6) and (3.8) imply that ι : DA( 1 (R; X) is bounded with p − 1 kιkL(DA( 1 q , 1) → B1/p−1/q q ,1),B1/p−1/q p− 1 p,1 p,1 (R;X)) ≤ C1M, (3.9) for some constant C1 ≥ 0 independent of A and X. FUNCTIONAL CALCULUS FOR C0-GROUPS USING TYPE AND COTYPE 11 It is shown in [25, Theorem 3.2] that ϕ ∗ ψ(s) = cosh(ωs) for all s ∈ R. Let Uµ ∈ L(X) be as in (2.3). Then the abstract transference principle from [26, Section 2] yields the commutative diagram 1 B1/p−1/q p,1 (R; X) ιx DA( 1 p − 1 q , 1) Lµω−−−−→ Lq(R; X) yP Uµ−−−−→ X of bounded maps. Finally, estimate the norms of P and ι using (3.7) and (3.9) and apply Corollary 2.5 to Lµω . (cid:3) Remark 3.2. If X is a UMD space then one can replace the space DA( 1 by the larger DA( 1 except that one uses (2.8) instead of Corollary 2.5 at the very end of the proof. q , 1) q , p). This follows in the exact same way as Proposition 3.1, p − 1 p − 1 For Banach lattices we establish another transference principle. Recall that each Banach space X with type p = 2 and cotype q = 2 is isomorphic to an L2-space, by [39]. Hence the following proposition deals with the case p = q = 2 in Proposition 3.1. Proposition 3.3. Let ω > ω0 ≥ 0, p ∈ [1, 2] and q ∈ [2,∞). Let X be isomorphic to a complemented subspace of a p-convex and q-concave Banach lattice. Then there exists a constant C ≥ 0 such that the following holds. Let −iA generate a C0-group (U (s))s∈R ⊆ L(X) such that kU (s)kL(X) ≤ M cosh(ω0s) for all s ∈ R and some M ≥ 1. Then (cid:13)(cid:13)(cid:13)(cid:13) ZR U (s)x µ(ds)(cid:13)(cid:13)(cid:13)(cid:13)X ≤ CM 2 sup s∈R (cid:16)s 1 p− 1 q F µω(s)(cid:17)kxkX for all µ ∈ Mω(R) and all x ∈ X. Proof. For µ ∈ Mω(R) let Uµ ∈ L(X) be as in (2.3). Let ι and P be as in (3.4) and (3.5), with ψ and ϕ as in (3.3). As in the proof of Proposition 3.1 one can factorize Uµ as Uµ = P ◦ Lµω ◦ ι. Hence (3.6), (3.7) and Proposition 2.8 conclude the proof. (cid:3) 4. Results for C0-groups In this section we obtain functional calculus results for generators of C0-groups. 4.1. The main result for C0-groups. We now prove our main functional calculus result for C0-groups, already stated in the Introduction as Theorem 1.1. Theorem 4.1. Let −iA generate a C0-group (U (s))s∈R ⊆ L(X) on a Banach space X with type p ∈ [1, 2] and cotype q ∈ [2,∞). Let ω > θ(U ). Then there exists a constant C ≥ 0 such that DA( 1 (4.1) q , 1) ⊆ D(f (A)) and p − 1 q ,1),X) ≤ C kfkH∞(Stω ) p − 1 kf (A)kL(DA( 1 for all f ∈ H∞(Stω). 12 JAN ROZENDAAL Proof. First consider f ∈ E(Stω) and let α ∈ (θ(U ), ω). By Lemma 2.1 there exists a µ ∈ Mα(R) with f = F µ and f (A) = Uµ. By Proposition 3.1, kf (A)xkX ≤ C kF µαkL∞(R) kxk 1 p − 1 q ,1 for all x ∈ DA(1/p− 1/q, 1) and some constant C ≥ 0 independent of f and x. This implies (4.1) since kF µαkL∞(R) ≤ kfkH∞(Stω), as follows from (3.2). For general f ∈ H∞(Stω), define τk(z) := −k2(ik − z)−2 for k ∈ N with k > ω and z ∈ Stω. Then f τk ∈ E(Stω) for all k, k kf τkkH∞(Stω) ≤ kfkH∞(Stω ) sup sup k kτkkH∞(Stω ) < ∞, and (f τk)(z) → f (z) as k → ∞, for all z ∈ Stω. By what we have already shown, kf τk(A)xkX ≤ C kf τkkH∞(Stω ) kxk 1 := C supk kτkkH∞(Stω ). Since D(A2) ⊆ DA( 1 p − 1 for C′ lim supk kτkkH∞(Stω ) = 1, Lemma 2.2 yields f (A) ∈ L(DA( 1 q ,1),X) ≤ C kfkH∞(Stω) , kf (A)kL(DA( 1 p− 1 p − 1 q ,1 ≤ C′ kfkH∞(Stω) kxk 1 p− 1 q ,1 q , 1) is dense and since p − 1 q , 1), X) with which concludes the proof. (cid:3) Remark 4.2. If in Theorem 4.1 one assumes in addition that X is a UMD space, then DA( 1 q , 1) may be replaced by DA( 1 q , p). This follows by using Remark 3.2 instead of Proposition 3.1 in the proof. p − 1 p − 1 Under additional assumptions one can obtain stronger results. Note that (4.2) improves (4.1), since DA(θ, 1) ⊆ D((λ + iA)θ) for each group generator A, λ ∈ R sufficiently large and θ ∈ [0, 1], by [42, Proposition 4.1.7]. Theorem 4.3. Let −iA generate a C0-group (U (s))s∈R ⊆ L(X) on a Banach space X. Suppose that X is isomorphic to a complemented subspace of a p-convex and q-concave Banach lattice, for p ∈ [1, 2] and q ∈ [2,∞). Let λ > ω > θ(U ). Then there exists a constant C ≥ 0 such that D((λ + iA) (4.2) q ) ⊆ D(f (A)) and p− 1 − 1 1 kf (A)kL(D((λ+iA) 1 p q ),X) ≤ C kfkH∞(Stω ) for all f ∈ H∞(Stω). p − 1 Proof. Write θ := 1 q . For f ∈ E(Stω) and α ∈ (θ(U ), ω) there exists a µ ∈ Mα(R) with f = F µα and f (A) = Uµ, by Lemma 2.1. Let x ∈ D((λ + iA)θ). By [23, Corollary 3.3.6], f (A)x = f (A)(λ + iA)−θ(λ + iA)θx = UµUν(λ + iA)θx = Uµ∗ν(λ + iA)θx, where ν ∈ Mω(R) is given by ν(ds) := 1 ν)(s) = f (s)(λ + is)−θ for s ∈ R, (3.2) yields Γ(θ) 1[0,∞)(s)sθ−1e−λsds. Since F (µ ∗ sup s∈R sθF (µ ∗ ν)α(s) ≤ C1 kfkH∞(Stω) for some constant C1 ≥ 0. Now Proposition 3.3 yields a constant C2 ≥ 0 such that kf (A)xkX ≤ C2 sup s∈R sθF (µ ∗ ν)α(s)k(λ + iA)θxkX ≤ C1C2 kfkH∞(Stω ) k(λ + iA)θxkX , FUNCTIONAL CALCULUS FOR C0-GROUPS USING TYPE AND COTYPE 13 which proves (4.2) for f ∈ E(Stω). Proceed as in the proof of Theorem 4.1 to obtain (4.2) for general f ∈ H∞(Stω). (cid:3) Remark 4.4. It follows from the proofs of Theorems 4.1 and 4.3 that the constants in (4.1) and (4.2) depend on A only through the norm kU (s)kL(X) of U (s) for all s ∈ R. In (4.1) the constant C depends on the underlying space X only through the Gaussian type p constant and the Gaussian cotype q constant of X. In particular, the dependence of C on these constants is as in Proposition 2.4. In (4.2) the constant depends on the underlying space only through the constant in (2.9). Note from the proof of Proposition 2.8 that, if (2.9) holds with constant C ≥ 0 on Y and if X is complemented in Y by a projection P , then (2.9) holds on X with constant C kPkL(Y ). This fact will be used in Theorem 6.3. 4.2. More results for C0-groups. Here we derive some additional results for group generators from Theorems 4.1 and 4.3. Proposition 4.5. Let −iA generate a C0-group (U (s))s∈R ⊆ L(X) on a Banach space X with type p ∈ [1, 2] and cotype q ∈ [2,∞). Let ω > θ(U ) and α, λ ∈ C with Re(α) > 1 q and Re(λ) > ω. Then there exists a constant C ≥ 0 such that D((λ + iA)α) ⊆ D(f (A)) and p − 1 (cid:13)(cid:13)f (A)(λ + iA)−α(cid:13)(cid:13)L(X) ≤ C kfkH∞(Stω ) for all f ∈ H∞(Stω). p − 1 If 1 Proof. The case p = q = 2 follows from Theorem 4.3. p − 1 by [42, Propositions 1.1.4 and 4.1.7], D((λ + iA)α) ⊆ DA( 1 Hence the proof is concluded by appealing to Theorem 4.1. q ∈ (0, 1) then, q , 1) continuously. (cid:3) One may equivalently formulate Proposition 4.5 as a statement about bounded- ness on X of the calculus for functions with sufficient decay: Proposition 4.6. Let −iA generate a C0-group (U (s))s∈R ⊆ L(X) on a Banach space X with type p ∈ [1, 2] and cotype q ∈ [2,∞). Let ω > θ(U ) and α, λ ∈ C with Re(α) > 1 q and Re(λ) > ω. Then there exists a constant C ≥ 0 such that the following holds. Let f ∈ H∞(Stω) be such that f (z) ∈ O(z−α) as Re(z) → ∞. Then f (A) ∈ L(X) with (4.3) p − 1 (cid:3) kf (A)kL(X) ≤ C sup z∈Stωλ + izαf (z). Proof. Apply Proposition 4.5 to (λ + i·)αf (·) ∈ H∞(Stω). p − 1 In Propositions 4.5 and 4.6 one may let α = 1 q if X is isomorphic to a complemented subspace of a p-convex and q-concave Banach lattice, as follows from Theorem 4.3. Corollary 4.7. Let ω > 0 and let f ∈ H∞(Stω) be such that f (z) ∈ O(z−1/2) as Re(z) → ∞. Then f = F µ, where µ ∈ Mω′(R) for all ω′ ∈ [0, ω). If in addition f is bounded and holomorphic on {z ∈ C Im(z) > −w} then supp(µ) ⊆ [0,∞). dt with maximal domain on X := L1(R, eω′tdt). Proof. For ω′ ∈ [0, ω), let A := i d Then −iA generates the left translation group (U (s))s∈R ⊆ L(X), and θ(U ) = ω′. By Proposition 4.6, f (A) ∈ L(X). Now [24, Proposition 2.3] implies that f = F µ for some µ ∈ Mω′(R). By uniqueness of the Fourier transform, µ is independent 14 JAN ROZENDAAL of the choice of ω′ ∈ [0, ω). The final statement follows from an application of Liouville's theorem. (cid:3) p − 1 q ≥ 1 Remark 4.8. It follows from Corollary 4.7 that the conclusion of Theorem 4.3 holds without any assumptions on the Banach space X if 1 2 . Indeed, let −iA generate a C0-group (U (s))s∈R ⊆ L(X) on a general Banach space X, and let λ > ω > θ(U ) and f ∈ H∞(Stω). By Corollary 4.7 and Lemma 2.1, f (A)(λ + iA)−1/2 = (f (·)(λ + i·)−1/2)(A) ∈ L(X) and hence f (A) ∈ L(D((λ + iA)1/2), X). Also note that Corollary 4.7 extends [28, Lemma 2.4] from α > 1/2 to α = 1/2. Remark 4.9. Let −iA generate a C0-group (U (s))s∈R ⊆ L(X) on a UMD space X. In [25] it is shown that A has a bounded H∞1 (Stω)-calculus for all ω > θ(U ), where H∞1 (Stω) is defined by (1.4). Since X has type p ∈ (1, 2] and cotype q ∈ [2,∞), one can compare our results with those in [25]. We note that our results do not imply those in [25], nor does [25] imply the results in this article. Indeed, let ω > θ(U ). Then f (z) := (λ + iz)−α defines an element of H∞1 (Stω) for all α > 0 and λ > ω, but Proposition 4.6 does not apply to f if α ∈ (0, 1 q ). Also, the function f ∈ H∞(Stω) given by f (z) := e−iz(λ + iz)−α is not an element of H∞1 (Stω) if α ∈ ( 1 Although in both examples it is clear that f (A) ∈ L(X), the difference between estimating kf (A)kL(X) by (1.4) or using (4.3) is relevant for numerical approxima- tion methods, as is shown in Section 7. p − 1 q at infinity. q , 1) but decays with order α > 1 p − 1 p − 1 We now obtain a version of Theorem 4.1 for other interpolation spaces. Proposition 4.10. Let −iA generate a C0-group (U (s))s∈R ⊆ L(X) on a Banach p + 1 space X with type p ∈ [1, 2] and cotype q ∈ [2,∞). Let ω > θ(U ), r ∈ (0, 1 − 1 q ) and u ∈ [1,∞]. Then there exists a constant C ≥ 0 such that q ,u kf (A)xkr,u ≤ C kfkH∞(Stω) kxkr+ 1 p − 1 q , u). p − 1 for all f ∈ H∞(Stω) and x ∈ DA(r + 1 Proof. Let f ∈ H∞(Stω) and first consider the case where u = 1. Then the part −iAr,1 of −iA in DA(r, 1) generates the C0-group (U (s) ↾DA(r,1))s∈R ⊆ L(DA(r, 1)) which satisfies θ(U ↾DA(r,1)) ≤ θ(U ), by [29, Lemma 2.2]. Hence Theorem 4.1 yields (4.4) kf (Ar,1)xkr,1 ≤ C kfkH∞(Stω ) kxk(DA(r,1),D(Ar,1)) 1 − 1 q ,1 p for all x ∈ (DA(r, 1), D(Ar,1)) 1 q ,1. Moreover, it follows from [42, Proposition p− 1 3.1.5] that DA(r, 1) = (X, D(A2)) r 2 ,1 and D(Ar,1) = (X, D(A2)) r+1 2 ,1 with equiva- lent norms. Hence, by the Reiteration Theorem (see [42, Theorem 1.3.5]) and again by [42, Proposition 3.1.5], (DA(r, 1), D(Ar,1)) 1 p − 1 q ,1 = ((X, D(A2)) r = (X, D(A2)) 1 2 ,1, (X, D(A2)) r+1 2 ,1) 1 q ),1 = DA(r + 1 2 (r+ 1 p − 1 q ,1 p− 1 p − 1 q , 1). Combine this with (4.4), using that f (Ar,1)x = f (A)x for all x ∈ D(f (Ar,1)) by [29, Lemma 2.2], to conclude the proof in the case where u = 1. For general u ∈ [1,∞], the Reiteration Theorem yields DA(r, u) = (DA(θ1, 1), DA(θ2, 1))θ,u FUNCTIONAL CALCULUS FOR C0-GROUPS USING TYPE AND COTYPE 15 and DA(r + 1 p − 1 p − 1 q , 1))θ,u q , u) = (DA(θ1 + 1 q , 1), DA(θ2 + 1 p − 1 p + 1 q ) and θ ∈ (0, 1). Now apply (1.6) to what we have (cid:3) for certain θ1, θ2 ∈ (0, 1 − 1 already shown to conclude the proof. Remark 4.11. For u < ∞ Proposition 4.10 can be proved directly, without using Theorem 4.1. To do so, adapt Proposition 3.1 to the setting of Proposition 4.10, using instead of Corollary 2.5 the more general Proposition 2.4, and then proceed as in the proof of Theorem 4.1. Similarly, Theorem 4.3 extends to other fractional domains. Proposition 4.12. Let −iA generate a C0-group (U (s))s∈R ⊆ L(X) on a comple- mented subspace X of a p-convex and q-concave Banach lattice, where p ∈ [1, 2] and q ∈ [2,∞). Let λ > ω > θ(U ). Then there exists a constant C ≥ 0 such that k(λ + iA)αf (A)xkX ≤ C kfkH∞(Stω ) k(λ + iA)α+ 1 p− 1 q xkX for all α ∈ C, f ∈ H∞(Stω) and x ∈ D((λ + iA)α+ 1 p− 1 q ). Proof. Apply Theorem 4.3 to (λ + iA)αx ∈ D((λ + iA) iA)α+ 1 p − 1 q ). 1 p − 1 q ) for each x ∈ D((λ + (cid:3) 5. Results for sectorial operators and cosine functions In this section we derive from Theorem 4.1 some results for sectorial operators and generators of cosine functions. By Remark 4.2, the results in this section can be improved on UMD spaces, and by Remark 4.8 the statements hold on general Banach spaces for 1 p − 1 q ≥ 1 2 . 5.1. Sectorial operators. For ϕ ∈ (0, π) let Sϕ := {z ∈ Carg(z) < ϕ}. An operator A on a Banach space X is said to be a sectorial operator of angle ϕ if σ(A) ⊆ Sϕ and sup{kzR(z, A)k ψ ∈ C \ Sψ} < ∞ for all ψ ∈ (ϕ, π). as for strip-type operators. Define f (A) ∈ L(X) via a Cauchy-type integral for For sectorial operators one can construct a functional calculus in a similar manner f ∈ H∞0 (Sψ) :=ng ∈ H∞(Sψ)(cid:12)(cid:12)(cid:12)∃C, δ > 0 : g(z) ≤ C zδ 1 + z2δo, set 1(A) := IX and (1 + ·)−1(A) := (1 + A)−1, and then extend linearly and regularize as in (2.2). For details see [23, Chapter 2]. If A is an injective sectorial operator of angle ϕ ∈ (0, π), then log(A) is defined via the sectorial calculus for A, as is f (A) for all ψ ∈ (ϕ, π) and f ∈ H∞(Sψ). A sectorial operator A of angle ϕ ∈ (0, π) has bounded imaginary powers if A is injective and if −i log(A) generates a C0-group (U (s))s∈R ⊆ L(X). Then U (s) = A−is for all s ∈ R, and A is sectorial of angle θA := θ(U ) if θ(U ) ∈ [0, π), by [46, Theorem 2]. We write A ∈ BIP(X). If A ∈ BIP(X) and if there exists an ω ∈ [0, π) such that {e−ωsA−is s ∈ R} ⊆ L(X) is R-bounded, then A has a bounded H∞(Sψ) calculus for all ψ ∈ (ω, π). Moreover, if X has property (α) as below, then a bounded H∞(Sψ)-calculus for A implies that {e−ωsA−is s ∈ R} ⊆ L(X) is R-bounded for ω > ψ. See [36, Theorem 7.5]. Hence the following result is only of use if A does not have R-bounded imaginary powers. 16 JAN ROZENDAAL Theorem 5.1. Let A ∈ BIP(X) with θA < π, where X is a Banach space with type p ∈ [1, 2] and cotype q ∈ [2,∞). Let ψ ∈ (θA, π). Then there exists a constant C ≥ 0 such that Dlog(A)( 1 q , 1) ⊆ D(f (A)) and p − 1 kf (A)kL(Dlog(A)( 1 p− 1 q ,1),X) ≤ C kfkH∞(Sψ) for all f ∈ H∞(Sψ). Proof. By [23, Theorem 4.2.4], (g ◦ log)(A) = g(log(A)) for all g ∈ H∞(Stψ). Since g 7→ g◦log is an isometric isomorphism H∞(Stψ) → H∞(Sψ), Theorem 4.1 concludes the proof. (cid:3) Remark 5.2. It was shown by Dore in [16] that each sectorial operator A with dense range on a general Banach space X has a bounded H∞-calculus on the real interpolation spaces (X, D(A)∩ ran(A))θ,r for all θ ∈ (0, 1) and q ∈ [1,∞]. We note that this statement implies neither Proposition 4.10 nor Theorem 5.1. Indeed, after rotation Proposition 4.10 deals with functions on strips around the imaginary axis, whereas for unbounded group generators [16] only applies to f ∈ H∞(Sψ) for ψ > π 2 . Moreover, for all θ ∈ (0, 1) and q ∈ [1,∞] it holds that (X, D(A) ∩ ran(A))θ,r ⊆ D(log(A)). Since in general D(log(A)) ( Dlog(A)( 1 q , 1), Theorem 5.1 does not follow from [16]. p − 1 It follows from Proposition 4.10 that f (A) : Dlog(A)(r + 1 q , u) → Dlog(A)(r, u) is bounded for each r ∈ (0, 1 − 1 kf (A)kL(Dlog(A)(r+ 1 p + 1 p − 1 p − 1 q ) and each u ∈ [1,∞], and q ,u),Dlog(A)(r,u)) ≤ C kfkH∞(Sψ) . In the same manner one can deduce from Proposition 4.6 that each f ∈ H∞(Sψ) such that f (z) ∈ O(2π−i log(z)−α) as z → 0 or z → ∞ for some α > 1 q satisfies f (A) ∈ L(X). If X is isomorphic to a complemented subspace of a p-convex and q-concave Banach lattice then one may let α = 1 From Theorem 5.1 one obtains unconditionality of the functional calculus and p − 1 q . p− 1 square function estimates in the same manner as in [38, Theorem 12.2]. Corollary 5.3. Let A ∈ BIP(X) with θA < π, where X is a Banach space with type p ∈ [1, 2] and cotype q ∈ [2,∞). Let ψ ∈ (θA, π), f ∈ H∞0 (Sψ), and let (rk)k∈Z be a Rademacher sequence on a probability space (Ω, P). Then the following assertions hold: q ,1),X) n ∈ N, t > 0,ǫk = 1} < ∞; sup k=−n ǫkf (2ktA)kL(Dlog(A)( 1 ∞Xk=−∞ for all x ∈ Dlog(A)( 1 • sup{kPn p − 1 • there exists a constant C ≥ 0 such that t>0(cid:13)(cid:13)(cid:13) t>0(cid:13)(cid:13)(cid:13) ∞Xk=−∞ for all x∗ ∈ X∗. rkf (2ktA)x(cid:13)(cid:13)(cid:13)L2(Ω;X) ≤ C kxkDlog(A)( 1 rkf (2ktA)∗x∗(cid:13)(cid:13)(cid:13)L2(Ω;Dlog(A)( 1 q , 1), and p − 1 p − 1 sup p − 1 q ,1) q ,1)∗) ≤ C kx∗kX ∗ FUNCTIONAL CALCULUS FOR C0-GROUPS USING TYPE AND COTYPE 17 5.2. Cosine functions. For ω ≥ 0 let Πω := {z2 z ∈ Stω}. An operator A on a Banach space X is of parabola-type ω if σ(A) ⊆ Πω and if for all ω′ > ω there exists a Mω′ ≥ 0 such that kR(λ, A)k ≤ (λ ∈ C \ Πω′ ). Mω′ pλ(cid:16)Im(√λ) − ω′(cid:17) For operators of parabola-type ω ≥ 0 there is a natural functional calculus, con- structed similarly as the strip-type and sectorial functional calculi, and f (A) is defined as an unbounded operator for all f ∈ H∞(Πω′ ), ω′ > ω. For details see [27]. A cosine function Cos : R → L(X) on a Banach space X is a strongly continuous mapping such that Cos(0) = I and Cos(t + s) + Cos(t − s) = 2Cos(t)Cos(s) (s, t ∈ R). Then θ(Cos) := inf{ω ≥ 0 ∃M ≥ 0 : kCos(t)k ≤ M eωt for all t ∈ R} < ∞. The generator of a cosine function is the unique operator −A on X that satisfies λR(λ2,−A) =Z ∞ 0 e−λtCos(t) dt (λ > θ(Cos)). Then A is an operator of parabola-type θ(Cos). We now prove a version of Theorem 4.1 for generators of cosine functions. Theorem 5.4. Let −A generate a cosine function (Cos(s))s∈R ⊆ L(X) on a Ba- nach space X with type p ∈ [1, 2] and cotype q ∈ [2,∞). Let ω > θ(Cos). Then there exists a constant C ≥ 0 such that DA( 1 q ), 1) ⊆ D(f (A)) and p − 1 2 ( 1 kf (A)kL(DA( 1 2 ( 1 p − 1 q ),1),X) ≤ C kfkH∞(Πω) for all f ∈ H∞(Πω). Proof. The proof follows the same lines as that of [29, Proposition 5.5]. It suffices to assume that θ := 1 p − 1 q ∈ (0, 1). By [37, Theorem 2] there is a unique subspace V ⊆ X such that D(A) ⊆ V and such that −iA generates a C0-group (U (s))s∈R ⊆ L(V × X) on V × X, where A := i(cid:20) 0 IV −A 0 (cid:21) with domain D(A) := D(A) × V . Moreover, by [24, Theorem 6.2], θ(Cos) = θ(U ). Hence Theorem 4.1 yields a constant C ≥ 0 such that g(A) ∈ L(DA(θ, 1), V × X) with kg(A)kL(DA(θ,1),V ×X) ≤ C kgkH∞(Stω ) (5.1) for all g ∈ H∞(Stω). kfkH∞(Πω ). Moreover, it is straightforward to see that (5.2) Let f ∈ H∞(Πω). Then [z 7→ g(z) := f (z2)] ∈ H∞(Stω) and kgkH∞(Stω) = Now, f (AV ) ⊕ f (A) = g(A). A2 :=(cid:20) AV 0 A (cid:21) 0 18 JAN ROZENDAAL with D(A2) = D(AV ) × D(A). By [42, Proposition 3.1.4], D(A) × V ∈ K1/2(V × X, D(AV ) × D(A)) ∩ J1/2(V × X, D(AV ) × D(A)) , where K1/2 and J1/2 are as in [42, Definition 1.3.1]. Hence V ∈ K1/2(X, D(A)) ∩ J1/2(X, D(A)) . Now [42, Theorem 1.3.5] yields DA(θ, 1) = (V × X, D(A) × V )θ,1 = (V, D(A))θ,1 × (X, V )θ,1 2 , 1(cid:1) , X(cid:1) with Combining this with (5.1) and (5.2) yields f (A) ∈ L(cid:0)DA(cid:0) θ kf (A)kL(DA( θ 2 ,1),X) ≤ kg(A)kL(DA(θ,1),V ×X) ≤ C kgkH∞(Stω ) = C kfkH∞(Πω ) , 2 , 1(cid:1) × DA(cid:0) θ = DA(cid:0) 1+θ 2 , 1(cid:1) . as required. (cid:3) From Proposition 4.10 one deduces in a similar manner that, under the assump- q )) and u ∈ [1,∞], there exists tions of Proposition 5.4 and for all r ∈ (0, 1− 1 a constant C ≥ 0 such that p − 1 2 ( 1 kf (A)xkr,u ≤ C kfkH∞(Πω ) kxkr+ 1 2 ( 1 p − 1 q ),u 2 ( 1 p − 1 q ), u). We leave the formulation of the for all f ∈ H∞(Πω) and x ∈ DA(r + 1 obvious analogue of Proposition 4.6 for cosine functions to the reader. Remark 5.5. Let −A generate a cosine function (Cos(s))s∈R ⊆ L(X), where X is isomorphic to a complemented subspace of a p-convex and q-concave UMD Banach lattice, for p ∈ (1, 2] and q ∈ [2,∞). Let λ > ω > θ(Cos). Then there exists a constant C ≥ 0 such that kf (A)xkX ≤ C kfkH∞(Πω ) k(λ2 + A) 1 2 ( 1 p− 1 q )xkX q )). This follows from Theorem for all f ∈ H∞(Πω) and x ∈ D((λ2 + A) 4.3 as in the proof of Proposition 5.4, using the complex interpolation method and [23, Theorem 6.6.9] and [25, Theorem 5.5]. 1 2 ( 1 p− 1 It should be noted that generators of cosine functions on UMD spaces have a bounded sectorial H∞-calculus, by [25, Theorem 5.5]. Hence on UMD spaces The- orem 5.4 is only of use when it does not suffice to obtain an estimate for kf (A)xkX with respect to the supremum norm of f on a sector. The latter is e.g. the case if f (z) = g(z2) for a g ∈ H∞(Stω) which is unbounded on any double sector Sψ ∪−Sψ for ψ ∈ (0, π 2 ), such as g(z) = e−iz. 6. Operator-valued functional calculus In this section we extend our main functional calculus theorems to R-bounded operator-valued calculi. Since many of the ideas and proofs are similar to those in the sections before, we leave some details to the reader. Let A be a strip-type operator of height ω0 ≥ 0 on a Banach space X. Let A ⊆ L(X) be the algebra of bounded operators that commute with A, and let ω > ω0. For a bounded holomorphic f : Stω → A such that kf (z)kL(X) ≤ Cz−α for all z ∈ Stω and certain C ≥ 0 and α > 1, define f (A) as in (2.1). Regularization as in (2.2) extends this functional calculus to all bounded holomorphic f : Stω → A. FUNCTIONAL CALCULUS FOR C0-GROUPS USING TYPE AND COTYPE 19 For ω ≥ 0 let Mω(R;A) consist of the A-valued Borel measures µ on R such that eωsµ(ds) has bounded variation. If −iA generates a C0-group (U (s))s∈R ⊆ L(X) then one can define Uµ ∈ L(X) as in (2.3) for all ω > θ(U ) and µ ∈ Mω(R;A). Versions of Lemmas 2.1 and 2.2 hold for this operator-valued calculus, with the same proofs. For ω > 0 and T ⊆ L(X) let RH∞(Stω;T ) be the collection of bounded holo- morphic functions f : Stω → T such that {f (z) z ∈ Stω} ⊆ L(X) is R-bounded. If T is an algebra then RH∞(Stω;T ) is a Banach algebra with the norm (f ∈ RH∞(Stω;T )). kfkRH∞(Stω;T ) := RX ({f (z) z ∈ Stω}) The following result generalizes Theorem 4.1, since each f ∈ H∞(Stω) defines an element f ∈ RH∞(Stω;A) by ef (z) = f (z)IX for z ∈ Stω, and kefkRH∞(Stω ;A) = kfkH∞(Stω). Proposition 6.1. Let −iA generate a C0-group (U (s))s∈R ⊆ L(X) on a Banach space X with type p ∈ [1, 2] and cotype q ∈ [2,∞), and let ω > θ(U ). Then there exists a constant C ≥ 0 such that DA( 1 p − 1 kf (A)kL(DA( 1 p − 1 q , 1) ⊆ D(f (A)) and q ,1),X) ≤ C kfkRH∞(Stω ;A) for all f ∈ RH∞(Stω;A). Proof. First extend Proposition 3.1 to all µ ∈ Mω(R;A) for which RX ({F µω(s) s ∈ R}) < ∞. To this end, note that the abstract transference principle in [26, Section 2] extends to A-valued measures, and appeal to Proposition 2.4 instead of Corollary 2.5. Then proceed as in the proof of Theorem 4.1. (cid:3) The other results in the previous sections can be extended to statements about operator-valued calculi in a similar manner. In particular, if in Proposition 6.1 X is isomorphic to a complemented subspace of a p-convex and q-concave Banach lattice, then for each λ > ω there exists a constant C ≥ 0 such that D((λ + iA) q ) ⊆ D(f (A)) and p− 1 1 (6.1) kf (A)kL(D((λ+iA) 1 p − 1 q ),X) ≤ C kfkRH∞(Stω;A) for all f ∈ RH∞(Stω;A). Remark 4.4 applies to Proposition 6.1 and (6.1). Let (rk)k∈N and (r′j)j∈N be mutually independent Rademacher sequences on a probability space (Ω, P). We say that a Banach space X has property (α) if there exists a constant C ≥ 0 such that, for all m ∈ N, {xj,k}m j,k=1 ⊆ X and {αj,k}m 2(cid:17)1/2 j,k=1 ⊆ C with αj,k ≤ 1 for all j, k ∈ {1, . . . , m}, mXj,k=1 2(cid:17)1/2 mXj,k=1 (6.2) . ≤ C(cid:16)EE(cid:13)(cid:13)(cid:13) rkr′jxj,k(cid:13)(cid:13)(cid:13) (cid:16)EE(cid:13)(cid:13)(cid:13) rkr′jαj,kxj,k(cid:13)(cid:13)(cid:13) Each Banach lattice with finite cotype has property (α), and if X has property (α) then so do the closed subspaces of X and Lp(Ω; X) for each σ-finite measure space (Ω, µ) and each p ∈ [1,∞). the H∞(Stω)-calculus for A is R-bounded from DA( 1 kfkH∞(Stω) ≤ 1} ⊆ L(DA( 1 q , 1), X) is R-bounded. For T equal to the unit ball of C ⊆ L(X), the following theorem implies that q , 1) to X. That is, {f (A) p − 1 p − 1 20 JAN ROZENDAAL Theorem 6.2. Let −iA generate a C0-group (U (s))s∈R ⊆ L(X) on a Banach space X with property (α), type p ∈ [1, 2] and cotype q ∈ [2,∞). Let ω > θ(U ). Then there exists a constant C ≥ 0 such that RDA( 1 q ,1),X ({f (A) f ∈ RH∞(Stω;A ∩ T )}) ≤ CRX (T ) p − 1 p − 1 k=1 rkxk (xk)n k=1 ⊆ D(A)o for each R-bounded T ⊆ L(X). Proof. The proof is similar to that of [38, Theorem 12.8]. Write θ := 1 q and let (rk)k∈N be a Rademacher sequence on [0, 1]. Fix n ∈ N and let Radn(X) := {Pn k=1 ⊆ X} ⊆ L2([0, 1]; X). Let eA be the operator on Radn(X) with domain D(eA) =n nXk=1 k=1 rkxk) :=Pn eU (s)(cid:16) nXk=1 rkxk ∈ Radn(X)(cid:12)(cid:12)(cid:12)(xn)n k=1 rkAxk forPn such that eA(Pn k=1 rkxk ∈ D(eA). Then −ieA gener- ates the C0-group (eU (s))s∈R ⊆ L(Radn(X)) given by nXk=1 for s ∈ R and Pn k=1 rkxk ∈ Radn(X). Note that keU (s)kL(Radn(X)) = kU (s)kL(X) for all s ∈ R. Moreover, Radn(X) ⊆ L2([0, 1]; X) has type p and cotype q with τp,Radn(X) ≤ Cpτp,X and cq,Radn(X) ≤ Cqcq,X for constants Cp, Cq ≥ 0 depending only on p and q that come from the Kahane-Khintchine inequalities. By Proposition 6.1 and Remark 4.4 there exists a constant C1 ≥ 0 independent of n such that (6.3) rkxk(cid:17) = rkU (s)xk kf (eA)kL(D eA(θ,1),Radn(X)) ≤ C1 kfkRH∞(Stω; eA) for all f ∈ RH∞(Stω; eA), where eA ⊆ L(Radn(X)) is the algebra of operators commuting with eA. Let T ⊆ L(X) be R-bounded and let f1, . . . , fn ∈ RH∞(Stω;A ∩ T ). Define is well-defined. j=1 ⊆ Stω and (yj)m for z ∈ Stω and Pn k=1 rkxk ∈ Radn(X). We will now show that the range of f is R-bounded in L(Radn(X)), from which it will follow in particular that f : Stω → eA Let (r′j )j∈N be a Rademacher sequence on [0, 1], independent of (rk)k∈N. Let j=1 ⊆ Radn(X). Write yj = Pn m ∈ N, (zj)m k=1 rkxjk for j ∈ {1, . . . , m} and (xjk)n k=1 ⊆ X. Then [38, Lemma 4.11 and Remark 4.10] yield a constant C2 ≥ 0 depending only on X such that =Z 1 0 Z 1 0 (cid:13)(cid:13)(cid:13) r′j(v)rk(u)f (zj)xjk(cid:13)(cid:13)(cid:13) mXj=1 mXj=1 nXk=1 2 RX (T )2Z 1 0 Z 1 0 (cid:13)(cid:13)(cid:13) mXj=1 ≤ C2 r′jyj(cid:13)(cid:13)(cid:13) 2 RX (T )2(cid:13)(cid:13)(cid:13) mXj=1 r′j(v)rk(u)xjk(cid:13)(cid:13)(cid:13) r′j f (zj)yj(cid:13)(cid:13)(cid:13) L2([0,1];Radn(X)) L2([0,1];Radn(X)) nXk=1 = C2 (cid:13)(cid:13)(cid:13) dudv dudv 2 X 2 X 2 2 . f (z)(cid:16) nXk=1 rkxk(cid:17) = nXk=1 rkfk(z)xk FUNCTIONAL CALCULUS FOR C0-GROUPS USING TYPE AND COTYPE 21 (6.5) (6.4) (6.3) yields k=1 rkxkkRadn(D(A)) Hence f ∈ RH∞(Stω; eA) with kfkRH∞(Stω ; eA) ≤ C2RX (T ). Combining this with kf (eA)kL(D eA(θ,1),Radm(X)) ≤ C1C2RX (T ). Note that(cid:13)(cid:13)Pn k=1 rkxkkRadn(D(A)) ≤(cid:13)(cid:13)Pn k=1 rkxkkD( eA) ≤ 2(cid:13)(cid:13)Pn for all Pn k=1 rkxk ∈ D(eA), hence D eA(θ, 1) = Radn(DA(θ, 1)) with (cid:13)(cid:13)(cid:13) rkxk(cid:13)(cid:13)(cid:13)Radn(DA(θ,1)) ≤(cid:13)(cid:13)(cid:13) rkxk(cid:13)(cid:13)(cid:13)D eA(θ,1) ≤ 2(cid:13)(cid:13)(cid:13) rkxk(cid:13)(cid:13)(cid:13)Radn(DA(θ,1)) nXk=1 nXk=1 nXk=1 for allPn k=1 rkxk ∈ D eA(θ, 1). Also, it is straightforward to check (by regularization) k=1 rkfk(A)xk for all Pn k=1 rkxk) = Pn that f (eA)(Pn k=1 rkxk ∈ D(f (eA)). Hence (cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)f (eA)(cid:16) nXk=1 rkfk(A)xk(cid:13)(cid:13)(cid:13)L2([0,1];X) rkxk(cid:17)(cid:13)(cid:13)(cid:13)Radn(X) nXk=1 ≤ C1C2RX (T )(cid:13)(cid:13)(cid:13) rkxk(cid:13)(cid:13)(cid:13)D eA(θ,1) nXk=1 rkxk(cid:13)(cid:13)(cid:13)L2([0,1];DA(θ,1)) ≤ CRX (T )(cid:13)(cid:13)(cid:13) nXk=1 for all x1, . . . , xn ∈ DA(θ, 1), where C ≥ 0 is independent of T ⊆ L(X), n ∈ N, f1, . . . fn ∈ RH∞(Stω;A ∩ T ) and x1, . . . , xn ∈ DA(θ, 1). This concludes the proof. (cid:3) (6.4) and (6.5) yield In the same manner we deduce an R-bounded version of Theorem 4.3, under the extra assumption of p-convexity for some p > 1. By [41, Corollary 1.f.9] the latter is equivalent to the assumption of nontrivial type. Recall that any closed subspace of a Banach lattice with finite cotype has property (α). Theorem 6.3. Let X be isomorphic to a complemented subspace of a p-convex and q-concave Banach lattice, for p ∈ (1, 2] and q ∈ [2,∞). Let −iA generate a C0-group (U (s))s∈R ⊆ L(X), and let λ > ω > θ(U ). Then there exists a constant C ≥ 0 such that R D((λ+iA) 1 p − 1 q ),X ({f (A) f ∈ RH∞(Stω;A ∩ T )}) ≤ CRX (T ) for each R-bounded T ⊆ L(X). Proof. We use notation as in the proof of Theorem 6.2. It suffices to show that there exists a constant C ≥ 0 independent of n ∈ N such that (6.6) − 1 q ),Radn(X)) ≤ C kfkRH∞(Stω ; eA) kf (eA)kL(D((λ+i eA) 1 p for all f ∈ RH∞(Stω; eA). Indeed, once this has been established, the rest of the proof is identical to that of Theorem 6.2. To obtain (6.6) apply (6.1) to eA on Radn(X). To see that the constant C that one gets from this can be chosen to be independent of n, it suffices by Remark 4.4 to show that Radn(X) is complemented in L2([0, 1]; X) by a projection Pn ∈ L(L2([0, 1]; X)) with kPnkL(L2([0,1];X)) ≤ C′ for some C′ ≥ 0 independent of n. The latter in turn follows from the fact that 22 JAN ROZENDAAL X has nontrivial type and from Pisier's characterization in [45] of the K-convex Banach spaces as the spaces with nontrivial type. (cid:3) We do not know whether the assumption in Theorem 6.3 that X has nontrivial type is necessary. From Theorems 6.2 and 6.3 one can deduce R-bounded versions of Theorems 5.1 and 5.4 in the obvious manner. Also, as a corollary of our results, for C0-groups we improve Theorem [33, Theorem 6.1]. Corollary 6.4. Let −iA generate a C0-group (U (s))s∈R ⊆ L(X) on a Banach space X with property (α), type p ∈ [1, 2] and cotype q ∈ [2,∞). Let ω > θ(U ). Then {e−ωsU (s) s ∈ R} ⊆ L(DA( 1 If X is isomorphic to a complemented subspace of a p-convex and q-concave Banach lattice for p ∈ (1, 2] and q ∈ [2,∞), then {e−ωsU (s) s ∈ R} ⊆ L(D(λ + iA) Proof. Let ω′ ∈ (θ(U ), ω) and let T be the unit ball of C ⊆ L(X). Now apply Theorems 6.2 and 6.3 to {e−(ωs+is·) s ∈ R} ⊆ RH∞(Stω′ ;A ∩ T ). q , X) is R-bounded for each λ > ω. q , 1), X) is R-bounded. p − 1 (cid:3) 1 p− 1 As the following example shows, Corollary 6.4 and Theorem 6.3 are sharp. Example 6.5. Let p ∈ [1,∞) and let (U (s))s∈R ⊆ L(X) be the left translation group on X := Lp(R) with generator −iA, where Af := if′ for f ∈ D(A) = W1,p(R). Then D((iA)α) = Hα,p(R), where Hα,p(R) is a Bessel-potential space. It is shown in [33, Example 6.2] that {U (s) s ∈ [−1, 1]} ⊆ L(Hα,p(R), Lp(R)) is not R-bounded for α ∈ [0, 1 q ). Hence {e−ωsU (s) s ∈ R} ⊆ L(D((iA)α), X) is not R-bounded for w ∈ R and α ∈ [0, 1 p − 1 q ). p − 1 7. Rational approximation In this section we give an application of the results in previous sections to the theory of rational approximation of C0-groups. Note that the results in Section 6 can be used to replace the uniform bounds in this section by R-bounds. Recall that a C0-semigroup (T (t))t≥0 ⊆ L(X) on a Banach space X is expo- nentially stable if there exist M ≥ 1 and ω > 0 such that kT (t)kL(X) ≤ M e−ωt for all t ≥ 0. We note that, if −A generates an exponentially stable C0-semigroup (T (t))t≥0 ⊆ L(X) such that T (t) ∈ L(X) is invertible for each t ≥ 0, then −A in fact generates the C0-group (U (s))s∈R ⊆ L(X), where U (s) := T (s) for s ≥ 0 and U (s) := T (−s)−1 for s < 0. Then f (A) is defined as an unbounded operator for each f ∈ H∞(C+) by a shifted version of the strip-type calculus from Section 2.1. Lemma 7.1. Let −A generate an exponentially stable C0-semigroup (T (t))t≥0 on a Banach space X with type p ∈ [1, 2] and cotype q ∈ [2,∞). Suppose that T (t) is invertible for all t ≥ 0. Then there exists a constant C ≥ 0 such that (7.1) kf (A)kL(DA( 1 for all f ∈ H∞(C+). For each β > 1 (7.2) for all f ∈ H∞(C+). (cid:13)(cid:13)f (A)A−β(cid:13)(cid:13)L(X) ≤ C′ kfkH∞(C+) q ,1),X) ≤ C kfkH∞(C+) q there exists a constant C′ ≥ 0 such that p − 1 p − 1 For n ∈ N let pn and qn be the unique polynomials of degree n and n + 1, respectively, such that pn(0) = qn(0) = 1 and such that q ,1),X) < ∞. sup p− 1 n∈N(cid:13)(cid:13)(1 − A)n(1 + A)−n(cid:13)(cid:13)L(DA( 1 qn(z) − ez(cid:12)(cid:12)(cid:12)(cid:12) ≤ Cz2n+2 pn(z) (cid:12)(cid:12)(cid:12)(cid:12) FUNCTIONAL CALCULUS FOR C0-GROUPS USING TYPE AND COTYPE 23 Proof. Let f ∈ H∞(C+) and apply Theorem 4.1 and Proposition 4.5 to the strip- type operator −i(A − ω) and the function f (i · +ω) ∈ H∞(Stω) for ω > 0 large enough. Then use the composition rule f (i · +ω)(−i(A − ω)) = f (A), which is straightforward to prove in the same manner as [23, Theorem 4.2.4]. (cid:3) Lemma 7.1 applies to the important question of the power-boundedness of the Cayley transform (1− A)(1 + A)−1 of A, which in turn is equivalent to the stability of the Crank-Nicholson approximation scheme associated with A. Corollary 7.2. Let −A generate an exponentially stable C0-semigroup (T (t))t≥0 ⊆ L(X) on a Banach space X with type p ∈ [1, 2] and cotype q ∈ [2,∞). Suppose that T (t) is invertible for all t ≥ 0. Then for all z in a neighborhood of 0 ∈ C. Let rn := pn qn Pad´e approximation of the exponential function. . Then rn is the n-th subdiagonal For 1 p − 1 q ∈ [0, 1 p − 1 2 ) the following proposition improves convergence rates ob- tained in [17, Theorem 4.1] for uniformly bounded C0-semigroups on general Ba- nach spaces. Note that, on a Banach space X with type p ∈ [1, 2] and cotype q ∈ [2,∞), for α > 1 q we obtain strong convergence of rn(−tA) to T (t) on D(Aα) with rate ∩a<α O(n−a+ 1 Proposition 7.3. Let −A generate an exponentially stable C0-semigroup (T (t))t≥0 on a Banach space X with type p ∈ [1, 2] and cotype q ∈ [2,∞). Suppose that T (t) is invertible for all t ≥ 0. Let α > 1 q ). Then there exists a constant C ≥ 0 such that q and a ∈ (0, α − 1 q ), locally uniformly in t. p − 1 p + 1 p − 1 krn(−tA)x − T (t)xkX ≤ Cta(n + 1)−a kAαxkX for all t ∈ (0,∞), all n ∈ N with n > α (rn(−tA))n∈N converges strongly on DA( 1 Proof. Let t ∈ (0,∞) and n ∈ N with n ≥ α rn(−tz) − e−tz f (z) := 2 − 1 and all x ∈ D(Aα). Moreover, p − 1 q , 1) to T (t), locally uniformly in t. 2 − 1. Set za (z ∈ C+). Then Lemma 7.1 yields a constant C′ ≥ 0 such that (cid:13)(cid:13)(rn(−tA) − T (t))A−α(cid:13)(cid:13)L(X) =(cid:13)(cid:13)f (A)A−α+a(cid:13)(cid:13)L(X) ≤ C′ kfkH∞(C+) . By [17, Lemmas 3.3 and 3.5], kfkH∞(C+) = ta sup z∈C+ rn(−tz) − e−tz (tz)a ≤ 2ta(n + 1)−a. 24 JAN ROZENDAAL Therefore, with C := 2C′, krn(−tA)x − T (t)xkX ≤(cid:13)(cid:13)(rn(−tA) − T (t))A−α(cid:13)(cid:13)L(X) kAαxkX ≤ Cta(n + 1)−a kAαxkX for all x ∈ D(Aα), which proves the first statement. Since krnkH∞(C−) ≤ 1 for all n ∈ N by [18], Lemma 7.1 yields that q , 1), X) {rn(−tA) − T (t) n ∈ N, t ≥ 0} ⊆ L(DA( 1 p − 1 is uniformly bounded. The proof is now concluded by what we have already shown and by the fact that D(A2) is dense in DA( 1 (cid:3) p − 1 q , 1). p − 1 The same method that was used in Proposition 7.3 to yield strong convergence on DA( 1 q , 1) also works for other rational approximation methods. Recall that a rational function r ∈ H∞(C−) is said to be A-stable if krkH∞(C−) ≤ 1, and r is a rational approximation (of the exponential function) of order k ∈ N if there exists a constant C ≥ 0 such that r(z)− ez ≤ Czk+1 for all z in a complex neighborhood of 0. Corollary 7.4. Let r be an A-stable rational approximation of order k ∈ N. Let −A generate an exponentially stable C0-semigroup (T (t))t≥0 on a Banach space X with type p ∈ [1, 2] and cotype q ∈ [2,∞). Suppose that T (t) is invertible for all t ≥ 0. Then (r(− t q , 1) to T (t), locally uniformly in t ≥ 0. Proof. Lemma 7.1 yields a constant C ≥ 0 such that n A)n)n∈N converges strongly on DA( 1 p − 1 (cid:13)(cid:13)r(− t n A)n − T (t)(cid:13)(cid:13)L(DA( 1 p − 1 q ,1),X) ≤ C(cid:13)(cid:13)r(− t n·)n − e−t·(cid:13)(cid:13)H∞(C+) ≤ C(krnkH∞(C−) + 1) ≤ 2C for all n ∈ N and t ≥ 0. Since, by [10, Theorem 3], r(− t uniformly in t to T (t) on D(Ak+1), the uniform boundedness of n A)n − T (t) t ≥ 0, n ∈ N(cid:9) ⊆ L(DA( 1 (cid:8)r(− t p − 1 and the fact that D(Ak+1) is dense in DA( 1 q , 1), X) n A)n converges locally p − 1 q , 1) yield the desired statement. (cid:3) Remark 7.5. If X is isomorphic to a complemented subspace of a p-convex and q-concave Banach lattice, for p ∈ [1, 2] and q ∈ [2,∞), then the case β = 1 q is attained in Lemma 7.1. Hence, in the setting of Corollary 7.2, p − 1 sup n∈Nk(1 − A)n(1 + A)−n− 1 p + 1 qkL(X) < ∞. Moreover, one obtains rate O(n−α+ 1 on D(A q ) in Corollary 7.4. p − 1 1 p− 1 q ) in Proposition 7.3, and strong convergence Funding. This work was supported by the Netherlands Organisation for Scientific Research (NWO) [grant number 613.000.908 "Applications of Transference Princi- ples"]. Acknowledgements. The author thanks Mark Veraar for numerous helpful sugges- tions, and the referee for carefully checking the manuscript. FUNCTIONAL CALCULUS FOR C0-GROUPS USING TYPE AND COTYPE 25 References [1] H. Amann. Operator-valued Fourier multipliers, vector-valued Besov spaces, and applications. Math. Nachr., 186:5 -- 56, 1997. [2] W. Arendt. Semigroups and evolution equations: Functional calculus, regularity and kernel estimates. In C.M.Dafermos E. Feireisl, editor, Handbook of Differential Equations, pages 1 -- 85. Elsevier/North Holland, Amsterdam, 2004. [3] W. Arendt and N. Nikolski. Vector-valued holomorphic functions revisited. Math. Z., 234(4):777 -- 805, 2000. [4] P. Auscher, S. Hofmann, M. Lacey, A. McIntosh, and Ph. Tchamitchian. The solution of the Kato square root problem for second order elliptic operators on Rn. Ann. of Math. (2), 156(2):633 -- 654, 2002. [5] P. Auscher, A. McIntosh, and A. Morris. Calder´on reproducing formulas and applications to Hardy spaces. Rev. Mat. Iberoam., 31(3):865 -- 900, 2015. [6] A. Axelsson, S. Keith, and A. McIntosh. Quadratic estimates and functional calculi of per- turbed Dirac operators. Invent. Math., 163(3):455 -- 497, 2006. [7] C. Batty, M. Haase, and J. Mubeen. The holomorphic functional calculus approach to oper- ator semigroups. Acta Sci. Math. (Szeged), 79(1-2):289 -- 323, 2013. [8] E. Berkson, T. A. Gillespie, and P. S. Muhly. Generalized analyticity in UMD spaces. Ark. Mat., 27(1):1 -- 14, 1989. [9] K. Boyadzhiev and R. deLaubenfels. Spectral theorem for unbounded strongly continuous groups on a Hilbert space. Proc. Amer. Math. Soc., 120(1):127 -- 136, 1994. [10] P. Brenner and V. Thom´ee. On rational approximations of semigroups. SIAM J. Numer. Anal., 16(4):683 -- 694, 1979. [11] A. P. Calder´on. Ergodic theory and translation-invariant operators. Proc. Natl. Acad. Sci. USA, 59:349 -- 353, 1968. [12] A. Carbonaro and O. Dragicevi´c. Functional calculus for generators of symmetric contraction semigroups. Duke Math. J., 166(5):937 -- 974, 2017. [13] R. R. Coifman and G. Weiss. Transference methods in analysis. Expository lectures from the CBMS Regional Conference held at the University of Nebraska, May 31-June 4, 1976. Regional Conference Series in Mathematics. No.31. Providence, R.I.: American Mathematical Society (AMS) 59, 1977. [14] M. Cowling. Harmonic analysis on semigroups. Ann. of Math. (2), 117(2):267 -- 283, 1983. [15] J. Diestel, H. Jarchow, and A. Tonge. Absolutely summing operators, volume 43 of Cambridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge, 1995. [16] G. Dore. H∞ functional calculus in real interpolation spaces. II. Studia Math., 145(1):75 -- 83, 2001. [17] M. Egert and J. Rozendaal. Convergence of subdiagonal Pad´e approximations of C0- semigroups. J. Evol. Equ., 13(4):875 -- 895, 2013. [18] B. L. Ehle. A-stable methods and Pad´e approximations to the exponential. SIAM J. Math. Anal., 4:671 -- 680, 1973. [19] K. Engel and R. Nagel. One-parameter semigroups for linear evolution equations, volume 194 of Graduate Texts in Mathematics. Springer-Verlag, New York, 2000. With contributions by S. Brendle, M. Campiti, T. Hahn, G. Metafune, G. Nickel, D. Pallara, C. Perazzoli, A. Rhandi, S. Romanelli and R. Schnaubelt. [20] S. Flory, F. Neubrander, and L. Weis. Consistency and stabilization of rational approximation schemes for C0-semigroups. In Evolution equations: applications to physics, industry, life sci- ences and economics. Proceedings of the 7th international conference on evolution equations and their applications, EVEQ2000 conference, Levico Terme, Italy, October 30 -- November 4, 2000, pages 181 -- 193. Basel: Birkhauser, 2003. [21] A. Gomilko and Y. Tomilov. On convergence rates in approximation theory for operator semigroups. J. Funct. Anal., 266(5):3040 -- 3082, 2014. [22] B. Haak and M. Haase. Square function estimates and functional calculus. Online at https://arxiv.org/abs/1311.0453, 2013. [23] M. Haase. The functional calculus for sectorial operators, volume 169 of Operator Theory: Advances and Applications. Birkhauser Verlag, Basel, 2006. [24] M. Haase. Functional calculus for groups and applications to evolution equations. J. Evol. Equ., 7(3):529 -- 554, 2007. 26 JAN ROZENDAAL [25] M. Haase. A transference principle for general groups and functional calculus on UMD spaces. Math. Ann., 345(2):245 -- 265, 2009. [26] M. Haase. Transference principles for semigroups and a theorem of Peller. J. Funct. Anal., 261(10):2959 -- 2998, 2011. [27] M. Haase. The functional calculus approach to cosine operator functions. In Recent Trends in Analysis. Proceedings of the Conference in honor of N.K. Nikolski held in Bordeaux 2011, pages 123 -- 147. Theta Foundation, 2013. [28] M. Haase and J. Rozendaal. Functional calculus for semigroup generators via transference. J. Funct. Anal., 265(12):3345 -- 3368, 2013. [29] M. Haase and J. Rozendaal. Functional calculus on real interpolation spaces for generators of C0-groups. Math. Nachr., 289(2-3):275 -- 289, 2016. [30] M. Hieber and J. Pruss. Functional calculi for linear operators in vector-valued Lp-spaces via the transference principle. Adv. Differential Equations, 3(6):847 -- 872, 1998. [31] T. Hytonen, J. van Neerven, M. Veraar, and L. Weis. Analysis in Banach Spaces. Volume I: Martingales and Littlewood-Paley Theory, volume 63 of Ergebnisse der Mathematik und ihrer Grenzgebiete (3). Springer, 2016. [32] T. Hytonen, J. van Neerven, M. Veraar, and L. Weis. Analysis in Banach spaces. Volume II: Probabilistic Methods and Operator Theory, volume 67 of Ergebnisse der Mathematik und ihrer Grenzgebiete (3). Springer, 2017. [33] T. Hytonen and M. Veraar. R-boundedness of smooth operator-valued functions. Integral Equations Operator Theory, 63(3):373 -- 402, 2009. [34] P. Jara, F. Neubrander, and K. Ozer. Rational inversion of the Laplace transform. J. Evol. Equ., 12(2):435 -- 457, 2012. [35] N. J. Kalton and L. Weis. The H∞-calculus and sums of closed operators. Math. Ann., 321(2):319 -- 345, 2001. [36] N.J. Kalton and L.W. Weis. The H∞-calculus and square function estimates. In Nigel J. Kalton Selecta, Volume 1, pages 715 -- 764. Springer, 2016. [37] J. Kisy´nski. On cosine operator functions and one-parameter groups of operators. Studia Math., 44:93 -- 105, 1972. Collection of articles honoring the completion by Antoni Zygmund of 50 years of scientific activity. I. [38] P. C. Kunstmann and L. Weis. Maximal Lp-Regularity for Parabolic Equations, Fourier Multiplier Theorems and H∞-functional Calculus. In Functional Analytic Methods for Evo- lution Equations (Levico Terme 2001), volume 1855 of Lecture Notes in Math., pages 65 -- 312. Springer, Berlin, 2004. [39] S. Kwapie´n. Isomorphic characterizations of inner product spaces by orthogonal series with vector valued coefficients. Studia Math., 44:583 -- 595, 1972. Collection of articles honoring the completion by Antoni Zygmund of 50 years of scientific activity, VI. [40] C. Le Merdy. On square functions associated to sectorial operators. Bull. Soc. Math. Fr., 132(1):137 -- 156, 2004. [41] J. Lindenstrauss and L. Tzafriri. Classical Banach spaces. II, volume 97 of Ergebnisse der Mathematik und ihrer Grenzgebiete [Results in Mathematics and Related Areas]. Springer- Verlag, Berlin-New York, 1979. Function spaces. [42] A. Lunardi. Interpolation theory. Appunti. Scuola Normale Superiore di Pisa (Nuova Serie). [Lecture Notes. Scuola Normale Superiore di Pisa (New Series)]. Edizioni della Normale, Pisa, second edition, 2009. [43] A. McIntosh. Operators which have an H∞ functional calculus. In Miniconference on operator theory and partial differential equations (North Ryde, 1986), volume 14 of Proc. Centre Math. Anal. Austral. Nat. Univ., pages 210 -- 231. Austral. Nat. Univ., Canberra, 1986. [44] F. Neubrander, K. Ozer, and T. Sandmaier. Rational approximations of semigroups without scaling and squaring. Discrete Contin. Dyn. Syst., 33(11-12):5305 -- 5317, 2013. [45] G. Pisier. Holomorphic semigroups and the geometry of Banach spaces. Ann. of Math. (2), 115(2):375 -- 392, 1982. [46] J. Pruss and H. Sohr. On operators with bounded imaginary powers in Banach spaces. Math. Z., 203(3):429 -- 452, 1990. [47] J. Rozendaal and M. Veraar. Fourier multiplier theorems on Besov spaces under type and cotype conditions. Banach J. Math. Anal., 11(4):713 -- 743, 2017. [48] J. Rozendaal and M. Veraar. Fourier Multiplier Theorems Involving Type and Cotype. J. Fourier Anal. Appl., 24(2):583 -- 619, 2018. FUNCTIONAL CALCULUS FOR C0-GROUPS USING TYPE AND COTYPE 27 [49] H.-J. Schmeisser and W. Sickel. Vector-valued Sobolev spaces and Gagliardo-Nirenberg in- equalities. In Nonlinear elliptic and parabolic problems, volume 64 of Progr. Nonlinear Dif- ferential Equations Appl., pages 463 -- 472. Birkhauser, Basel, 2005. [50] J. van Neerven. γ-radonifying operators -- a survey. In The AMSI-ANU Workshop on Spectral Theory and Harmonic Analysis, volume 44 of Proc. Centre Math. Appl. Austral. Nat. Univ., pages 1 -- 61. Austral. Nat. Univ., Canberra, 2010. [51] J. van Neerven, M. Veraar, and L. Weis. Stochastic maximal Lp-regularity. Ann. Probab., 40(2):788 -- 812, 2012. [52] M. Veraar. Embedding results for γ-spaces. In Recent trends in analysis. Proceedings of the conference in honor of Nikolai Nikolski on the occasion of his 70th birthday, Bordeaux, France, August 31 -- September 2, 2011, pages 209 -- 219. Bucharest: The Theta Foundation, 2013. Mathematical Sciences Institute, Australian National University, Acton ACT 2601, Australia, and Institute of Mathematics, Polish Academy of Sciences, ul. ´Sniadeckich 8, 00-656 Warsaw, Poland E-mail address: [email protected]
1711.09479	1	1711	2017-11-26T22:50:34	On hypercyclic rank one perturbations of unitary operators	[ "math.FA" ]	Recently, S. Grivaux showed that there exists a rank one perturbation of a unitary operator in a Hilbert space which is hypercyclic. Another construction was suggested later by the first and the third authors. Here, using a functional model for rank one perturbations of singular unitary operators, we give yet another construction of hypercyclic rank one perturbation of a unitary operator. In particular, we show that any Carleson set on the circle can be the spectrum of a perturbed (hypercyclic) operator.	math.FA	math	ON HYPERCYCLIC RANK ONE PERTURBATIONS OF UNITARY OPERATORS ANTON BARANOV, VLADIMIR KAPUSTIN, ANDREI LISHANSKII Abstract. Recently, S. Grivaux showed that there exists a rank one perturbation of a unitary operator in a Hilbert space which is hypercyclic. Another construction was suggested later by the first and the third authors. Here, using a functional model for rank one perturbations of singular unitary operators, we give yet an- other construction of hypercyclic rank one perturbation of a unitary operator. In particular, we show that any Carleson set on the circle can be the spectrum of a perturbed (hypercyclic) operator. 1. introduction A continuous linear operator T in a Banach (or Fr´echet) space F is said to be hypercyclic if there exists a vector f ∈ F such that its orbit {T nf }∞ n=0 is dense in F . In this case the vector f is said to be hypercyclic for T . Though the property to have a vector with a dense orbit may look rather exotic and "pathological", it turns out that many natural operators on spaces of analytic functions possess it (e.g., some classes of Toeplitz and composition operators, pseudodifferential operators on the space of all entire functions, weighted shifts, etc.). The theory of hypercyclic operators and, more generally, dynamics of linear operators, were an active fields of research during the last three decades. For a comprehensive account of the theory we refer to the recent monographs [3, 9]. One of the natural questions about hypercyclicity is its stability (or instabil- ity) with respect to small (in some sense) perturbations. In 1991 K.C. Chan and J.H. Shapiro [5] showed that there exist hypercyclic operators in a Hilbert space of the form I +K, where I is the identity operator and the compact operator K belongs to any Schatten class. It is clear that I + R can not be hypercyclic when R is a finite rank operator (since in this case I + R∗ has an eigenvalue). But if we replace I by a unitary operator, hypercyclicity is possible. In 2010 S. Shkarin [16] produced an example of a unitary operator U such that U + R is hypercyclic for some rank two operator R. Shkarin then asked whether R can be taken to be of rank one. A positive answer was soon given by S. Grivaux [8]: 1991 Mathematics Subject Classification. 47A16, 30A76, 30H10. Key words and phrases. hypercyclic operator, rank one perturbation, inner function, model space, functional model. The results of Section 2 were obtained with the support of the RFBR grant 16-01-00635. The results of Section 3 were obtained with the support of the Russian Science Foundation grant 14-21- 00035. 1 Theorem 1.1 (S. Grivaux, [8]). There exists a unitary operator U in the space ℓ2 and a rank one operator R such that U + R is hypercyclic. In 2015 A. Baranov and A. Lishanskii [1] gave another proof of this theorem using a functional model of rank one perturbations of unitary operators due to V. Ka- pustin [11] and A.D. Baranov and D.V. Yakubovich [2]. The constructions of [8] and [1] were both based on the following interesting theorem (also due to S. Grivaux) which gives a condition sufficient for hypercyclicity in the case when the operator has sufficiently many eigenvectors corresponding to unimodular eigenvalues with a certain "continuity property". Theorem 1.2 (S. Grivaux, [7]). Let X be a complex separable infinite-dimensional Banach space, and let T be a bounded operator on X. Suppose that there exists a sequence {un}n≥1 of vectors in X having the following properties: (i) un is an eigenvector of T associated to an eigenvalue λn of T , with λn = 1 and λn are all distinct; (ii) span{un : n ≥ 1} is dense in X; (iii) for any n ≥ 1 and any ε > 0, there exists m 6= n such that kun − umk < ε. Then T is hypercyclic. Both in [8] and [1] the main technical difficulty was to prove the continuity prop- erty (iii). In this note we give yet another construction of hypercyclic rank one perturbation of a unitary operator. This construction is related to a class of func- tional spaces studied in [11, 12, 13]. One of the advantages of this approach is that the continuity property of eigenvectors becomes almost automatic. Another (and apparently more important) novelty is that now we have a lot of freedom in choosing the spectrum of the perturbed (hypercyclic) operator (see Theorem 2.1 below). 2. Carleson sets and spaces of analytic functions Let D = {z ∈ C : z < 1} be the unit disk and T = {z ∈ C : z = 1} be the unit circle. We denote by H 2 the standard Hardy space in D (identifying it with subspace − = L2(T) ⊖ H 2 = zH 2. We denote of L2(T) via its boundary values) and we put H 2 by S and S∗ the shift and backward shift operators (considered on the Hardy space or on the related spaces introduced below), Sf (z) = zf (z), S∗f (z) = f (z) − f (0) z . Let E be a closed subset of T with zero Lebesgue measure and let {Ij} be at most countable set of disjoint open arcs Ij ⊂ T such that T \ E =Sj Ij. The set E is said to be a Carleson set (or a set of finite entropy) if Xj Ij log 1 Ij < ∞ (by I we always denote the normalized Lebesgue measure of I ⊂ T). By the classical results of L. Carleson [4], sets of finite entropy are precisely those sets that may serve 2 as zero sets of smooth up to the boundary) analytic functions in D. Namely, a nonzero function f ∈ Hol (D) ∩ C 1(D) (or f ∈ Hol (D) ∩ C ∞(D)) vanishing on E ⊂ T exists if and only if E is a Carleson set. Recall that a set E is said to be perfect if it is closed and has no isolated points. One of the main results of this note is the following theorem: Theorem 2.1. For any perfect Carleson set E on the unit circle T there exists a hypercyclic operator of the form U + R, where U is unitary and R of rank one, such that σ(U + R) = σp(U + R) = E. Here σ(T ) and σp(T ) denote the spectrum and the point spectrum of the operator T , respectively. Appearance of Carleson sets in this context is not random. It is interesting to compare Theorem 2.1 with the following remarkable result of N.G. Makarov [13]: a set F ⊂ T is a point spectrum of some operator of the form U +K where U is unitary and K is of trace class if and only if F is a countable union of Carleson sets. We do not know for which F it is possible to find a hypercyclic operator U + R as above with point spectrum F (and, in particular, whether there exist countable unions of Carleson sets for which this is not true). Another interesting problem is to describe those unitary operators U for which there exists a hypercyclic rank one perturbation U + R. This seems to be a difficult problem. Our results imply that the essential spectrum of such operator U can be any perfect Carleson set. 2.1. Construction of the space H∗(ϕ, ψ). The following class of spaces of an- alytic functions was introduced in [11] in connection with a model for rank one perturbations of unitary operators. Let ϕ and ψ be the functions in H 2(T) with the properties ϕ = ψ a.e. on T and ψ(0) 6= 0. Let H∗ = H∗(ϕ, ψ) be the space of functions defined by H∗ = H 2 ϕ ∩ H 2 − ¯ψ = {f : f ϕ ∈ H 2, f ¯ψ ∈ H 2 −} with the norm kf kH∗ := kf ϕkH 2 = kf ¯ψkH 2 − . It is clear that the space H∗ is invariant with respect to S∗. It is also shown in [11] that the inclusion f ∈ H 2 −/ψ and the fact that ψ(0) 6= 0 imply that we can consider the values of f ∈ H∗ outside the unit disk and that there exists a finite limit (zf )∞ = lim z→∞ zf (z). Assume now that E ⊂ T is a perfect Carleson set. Let ϕ ∈ C 1(D) be such that ϕ(0) 6= 0 and {z ∈ T : ϕ(z) = 0} = E. Then it is clear that any function of the 3 form (z − λ)−1, λ ∈ E, belongs to the space H∗ = H∗(ϕ, ϕ) associated with the pair (ϕ, ϕ). Indeed, ϕ z − λ ∈ H 2, ϕ z − λ = 1 λ(cid:18) zϕ λ − z(cid:19) ∈ H 2 −. We will use the fact that the functions in the space H∗ have analytic continuation through T \ E. Lemma 2.2. Let µ ∈ T \ E. Then any function in H∗ admits analytic continuation into a neighborhood of µ and the mapping f 7→ f (µ) is a bounded functional on H∗. Proof. By definition, f ∈ H∗ if and only if ϕf ∈ H 2 and ϕ¯z ¯f ∈ H 2. The last condition can be rewritten as follows: there exists a function g meromorphic and of Nevanlinna class in {z > 1} such that g = f a.e. on T and If µ ∈ T \ E, then there exist an open neighborhood V of µ and δ > 0 such that 1 z g(cid:16)1 ¯z(cid:17)ϕ(z) ∈ H 2. ϕ(z) ≥ δ, z ∈ D ∩ V . It follows that, for some ε > 0, 1<r<1+εZ{eiθ−µ<ε} sup g(reiθ)2dθ < ∞. Hence, g is a usual analytic continuation of f through the arc {eiθ − µ < ε}. Moreover, f (µ)2 ≤ f (z)2dm2(z) ≤ 1 πε2 Z{z−µ<ε} 1 πε2δ2 kϕf k2 H 2. Here m2 is the planar Lebesgue measure. (cid:3) 3. Construction of a hypercyclic rank one perturbation. In this section we use the spaces of analytic functions from the previous section to construct a rank one perturbatrion of a unitary operator which satisfies the conditions of Grivaux' theorem (Theorem 1.2). As above let E ⊂ T be a perfect Carleson set and let ϕ ∈ C 1(D) be such that ϕ(0) 6= 0 and {z ∈ T : ϕ(z) = 0} = E. Consider the space H∗ associated with the pair (ϕ, ϕ). Then any function of the form (z − λ)−1 ∈ H∗ for λ ∈ E. It follows from the inclusion ϕ ∈ C 1(D) that whence, by the Lebesgue dominated convergence theorem, (1) → 0 as µ → λ, µ ∈ E. ϕ z − µ (cid:13)(cid:13)(cid:13) ϕ(z) z − λ(cid:12)(cid:12)(cid:12) < ∞, ϕ sup λ∈E,z∈T(cid:12)(cid:12)(cid:12) z − λ(cid:13)(cid:13)(cid:13)H 2 H0 = SpanH∗(cid:26) 1 − : j ∈ J(cid:27). z − λj 4 Now, let {λj}j∈J be a countable subset of E without isolated points and put Theorem 3.1. The operator S∗H0 is a hypercyclic operator which is a rank one perturbation of some unitary operator. Moreover, σ(S∗H0) = σp(S∗H0) = E where E = {¯λ : λ ∈ E}. z−λj z−λj cj z−λj cj that S∗ acts isometrically on H1. We have the following lemma: To prove the converse inclusion, let us note first that finite sums of the form ∈ H1 (we consider only finite sums), then g(0) = −Pj λj (z−λj ), so (zS∗g)∞ = 0. Since (zf )∞ is a continuous linear functional on H∗, we conclude that (zf )∞ = 0 whenever f ∈ S∗H1. Thus, S∗H1 ⊂ are dense in eH1. Let f ∈ H0 and (zf )∞ = 0. For any ε > 0 we can choose z−λj(cid:13)(cid:13)(cid:13) < ε. Since h 7→ (zh)∞ is Put H1 = {f ∈ H0 : f (0) = 0} and eH1 = {f ∈ H0 : (zf )∞ = 0}. Then it is clear Lemma 3.2. S∗H1 = eH1. Proof. If g =Pj In this case S∗g = Pj eH1. P cj such that (cid:13)(cid:13)(cid:13)f −P cj a finite linear combination P cj (cid:12)(cid:12)(cid:12)X cj(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12)(cid:12) (cid:18)z(cid:16)X cj f (z) =X cj (a finite sum) and (zf )∞ = P cj = 0. Put g = P λj cj z(cid:18)X λjcj a bounded functional on H0, there exists a constant A > 0 (independent of f ) such that − f(cid:17)(cid:19)∞(cid:12)(cid:12)(cid:12)(cid:12) < Aε. − Pj cj ∈ eH1 z − λ0 Now let f (z) = P cj Then g(0) = 0 (i.e., g ∈ H1) and z−λj Let λ0 ∈ E be fixed which is independent on f . Then it is easy to see that and kf − f k ≤ ε(1 + Ak(z − λ0)−1kH ∗). +X cj(cid:19) = f. z − λj z − λj z − λj cj λj = 0. . z−λj 1 S∗g = Thus, eH1 ⊂ S∗H1. (cid:3) 3.1. Proof of Theorem 3.1. First we verify that the operator S∗H0 satisfies the conditions of Theorem 1.2. It is clear that the functions fj = 1 are eigenvectors z−λj of the operator S∗H0 with eigenvalues 1 λj . Indeed, S∗fj(z) = 1 z(cid:16) 1 z − λj + 1 λj(cid:17) = 1 λj(z − λj) . By the definition of the space H0 the set {fj}j∈J is complete in H0. Since the set {λj}j∈J has no isolated points, the continuity property (iii) in Theorem 1.2 follows from (1). Thus, S∗H0 is hypercyclic. 5 Next we need to show that S∗H0 is a rank one perturbation of some unitary It is clear that H0 = H1 ⊕ H2, where operator. Let H1 and eH1 be as above. H2 = Lin{g} for some g ∈ H0 such that g(0) = 1. By Lemma 3.2 S∗H1 = eH1 and so the closed subspace S∗H1 has codimension 1 in H0. Then we can define a unitary operator U as follows: let UH1 = S∗H1 and let UH2 be an arbitrary unitary operator of one-dimensional subspace H2 onto one-dimensional subspace H0 ⊖ S∗H1. Clearly, S∗H0 will be a rank one perturbation of U. It remains to show that σ(S∗H0) = σp(S∗H0) = E. It is clear that E ⊂ σp(S∗H0). We show that for λ /∈ E the equation S∗f − λf = g has a unique solution f ∈ H0 for any g ∈ H0. Indeed, this equation is equivalent to f = zg + f (0) 1 − λz . If µ = λ−1 ∈ D or µ ∈ T\E, then by lemma 2.2 any function g ∈ H0 is analytic in the neighborhood of µ and so there exists a unique c ∈ C, namely c = µg(µ), such that λ(z−µ) belongs to ϕ−1H 2. The inclusion ϕ¯z ¯f ∈ H 2 follows immediately from the f = zg−c inclusion ϕ¯z¯g ∈ H 2. If λ ∈ D, then ϕf ∈ H 2 for any f of the form f = zg−c λ(z−µ) . Also, ϕ¯z ¯f = ϕ¯z ¯z¯g − ¯c 1 − ¯λ¯z = ϕ¯z¯g − ¯cϕ z − ¯λ ∈ H 2 if and only if ¯c = (ϕ¯z¯g)(¯λ)/ϕ(¯λ). Thus, we have seen that there is a unique solution f ∈ H∗ for any g ∈ H∗. If g is a finite linear combination of (z − λj)−1, then it is easy to see that f also is of this form. Hence, f ∈ H0 whenever g ∈ H0. This proves Theorem 3.1. (cid:3) 3.2. A link to nearly invariant subspaces and Clark measures. In this sub- section we discuss the relation of our construction with Hitt–Sarason theory of nearly invariant subspaces. A closed subspace X ⊂ H 2 is said to be nearly invariant, if for any f ∈ X such that f (0) = 0 we have S∗f ∈ X. Recall that by the Beurling theorem any S∗-invariant subspace is of the form KΘ = H 2 ⊖ ΘH 2 for some inner function Θ (see, e.g., [14]). The structure of nearly invariant subspaces was described by D. Hitt [10] and D. Sarason [15]. Theorem 3.3 (Hitt, Sarason, [10, 15]). If X is nearly invariant, then X = gKΘ, where Θ is an inner function, g is outer, and the mapping Mg : f 7→ gf is an isometry on KΘ. In particular, Mg maps KΘ isometrically onto X. Let us show that is a nearly-invariant subspace of H 2. Indeed, for finite sums ϕP cj dense in ϕH0) we have z−λj (which are ϕH0 = ϕSpanH∗(cid:26) 1 z − λj(cid:12)(cid:12)(cid:12)(cid:12)z=0 X cj 6 : λ ∈ E(cid:27). z − λ = − cj λj = 0, and so 1 z(cid:18)ϕX cj z − λj(cid:19) = 1 z(cid:18)ϕX(cid:16) cj z − λj + cj λj(cid:17)(cid:19) = ϕ(cid:18)X cj λj(z − λj)(cid:19) ∈ ϕH0. Therefore, by the Hitt–Sarason theorem ϕH0 = gKΘ for some g and Θ. If f ∈ H0, then ϕf g ∈ Kθ and kf kH0 = kϕf kH 2 =(cid:13)(cid:13)(cid:13) ϕf g (cid:13)(cid:13)(cid:13)H 2 and so the mapping Mϕ/g : f 7→ ϕf /g is a unitary operator from H0 onto KΘ. It is clear that operators S∗H1 (where H1 = {f ∈ H0 : f (0) = 0}) and S∗{f ∈KΘ: f (0)=0} are unitarily equivalent. Since S∗H1 coincides on H1 with the unitary operator U constructed in Theorem 3.1, we conclude that U = M −1 ϕ/gUMϕ/g is a unitary operator on KΘ which is a rank one perturbation of S∗KΘ. Unitary rank one perturbations of S∗KΘ were described by D.N. Clark in a seminal paper [6]. This description involves a construction of a special family of measures. Let α ∈ T. Then the function α+θ α−θ has positive real part in D, and so there exists a finite positive measure µα on T (singular with respect to the Lebesgue measure) such that Re α + θ(z) α − θ(z) = 1 π ZT 1 − z2 τ − z2 dµα(τ ), z ∈ D. Clark showed that the set of unitary rank one perturbations of S∗KΘ can be parametrized by α ∈ T and their spectral measures are exactly the measures µα. Corollary 3.4. Let E ⊂ T be a perfect Carleson set and let ϕ be the corresponding smooth function with boundary zero set E. Then the space ϕH0 is a nearly invariant subspace. Moreover, if Θ is the associated inner function from the Hitt–Sarason theorem. then the spectral measure of the unitary operator U which has a hypercyclic rank one perturbation is the Clark measure µα for Θ. It would be interesting to use this relation to characterize (at least a subclass of) unitary operators U which have a hypercyclic rank one perturbation. The difficulty here is that the proof of the Hitt–Sarason theorem is nonconstructive and, even if we have explicit knowledge of ϕ and H0, it is difficult to translate it into the information about Θ. In particular, it would be interesting to know whether the corresponding Clark measure µα can have a nontrivial singular continuous (i.e., without atoms) part. References [1] A. Baranov, A. Lishanskii, On S. Grivaux' example of a hypercyclic rank one perturbation of a unitary operator, Arch. Math. 104 (2015), 3, 223–235. [2] A. Baranov, D. Yakubovich, Completeness and spectral synthesis of nonselfadjoint one- dimensional perturbations of selfadjoint operators, Adv. Math. 302 (2016), 740–798. [3] F. Bayart, E. Matheron, Dynamics of Linear Operators, Cambridge University Press, 2009. [4] L. Carleson, Sets of uniqueness for functions regular in the unit circle, Acta Math. 87 (1952), 325–345. 7 [5] K.C. Chan, J.H. Shapiro, The cyclic behavior of translation operators on Hilbert spaces of entire functions, Indiana Univ. Math. J. 40, 4 (1991), 1421–1449. [6] D.N. Clark, One-dimensional perturbations of restricted shifts, J. Anal. Math. 25 (1972), 169–191. [7] S. Grivaux, A new class of frequently hypercyclic operators, Indiana Univ. Math. J. 60 (2011), 1177–1201. [8] S. Grivaux, A hypercyclic rank one perturbation of a unitary operator, Math. Nachr. 285 (2012), 5-6, 533–544. [9] K.-G. Grosse-Erdmann, A. Peris Manguillot, Linear Chaos, Universitext, Springer, London, 2011. [10] D. Hitt, Invariant subspaces of H 2 of an annulus, Pacific J. Math. 134 (1988), 101–120. [11] V.V. Kapustin, One-dimensional perturbations of singular unitary operators, Zapiski Nauchn. Sem. POMI 232 (1996), 118–122; English transl.: J. Math. Sci. (New York) 92 (1998), 1, 3619–3621. [12] V.V. Kapustin, Spectral analysis of almost unitary operators, Algebra i Analiz 13 (2001), 5, 44–68; English transl.: St. Petersburg Math. J. 13 (2002), 5, 739–756. [13] N.G. Makarov, Unitary point spectrum of almost unitary operators, Zap. Nauchn. Sem. LOMI 126 (1983), 143–149; English transl.: J. Soviet Math. 27 (1984), 1, 2517–2520. [14] N.K. Nikolski, Operators, Functions, and Systems: an Easy Reading, Math. Surveys Monogr., Vol. 92, AMS, Providence, RI, 2002. [15] D. Sarason, Nearly invariant subspaces of the backward shift, Oper. Theory: Adv. Appl. 35 (1988), 481–493. [16] S. Shkarin, A hypercyclic finite rank perturbation of a unitary operator, Math. Ann. 348 (2010), 379–393. Anton Baranov, Department of Mathematics and Mechanics, St. Petersburg State University, St. Petersburg, Russia, [email protected] Vladimir Kapustin, St. Petersburg Department of V.A. Steklov Mathematical Institute St. Petersburg, Russia, [email protected] Andrei Lishanskii, Chebyshev Laboratory, St. Petersburg State University, St. Petersburg, Russia, [email protected] 8
1111.5109	1	1111	2011-11-22T06:10:27	Boundedness of oscillatory integral operators and their commutators on weighted Morrey spaces	[ "math.FA" ]	It is proved that both oscillatory integral operators and fractional oscillatory integral operators are bounded on weighted Morrey spaces. The corresponding commutators generated by $BMO$ functions are also considered.	math.FA	math	BOUNDEDNESS OF OSCILLATORY INTEGRAL OPERATORS AND THEIR COMMUTATORS ON WEIGHTED MORREY SPACES ZUNWEI FU, SHAOGUANG SHI♯, AND SHANZHEN LU Abstract. It is proved that both oscillatory integral operators and fractional oscil- latory integral operators are bounded on weighted Morrey spaces. The corresponding commutators generated by BM O functions are also considered. 1. Introduction To investigate the local behavior of solutions to second order elliptic partial differential equations, Morrey [14] first introduced the classical Morrey space Mp,q(Rn) with the norm kf kMp,q(Rn) = sup B⊂Rn 1 B1− p q ZB f (x)pdx! 1 p , where f ∈ Lp Mp,q(Rn) was an expansion of Lp(Rn) in the sense that Mp,p(Rn) = Lp(Rn). loc(Rn) and 1 ≤ p ≤ q < ∞. Here and after, B denotes any balls in Rn. In [2], Chiarenza and Frasca obtained the boundedness of Hardy-Littlewood maximal function on Mp,q(Rn). Mf (x) = sup B∋x 1 BZB f (y)dy, (1.1) The Calder´on-Zygmund singular integral operator is defined by K(x − y)f (y)dy, (1.2) where K is a Calder´on-Zygmund kernel(CZK). We say a kernel K ∈ C 1(Rn/{0}) is a CZK if it satisfies eT f (x) = p.v.ZRn K(x) ≤ C xn , ∇K(x) ≤ C xn+1 (1.3) (1.4) 2010 Mathematics Subject Classification. Primary 42B20; Secondary 42B25. Key words and phrases. weighted Morrey space, oscillatory integral, commutator, Ap weights. ♯ Corresponding author: [email protected]. This work was partially supported by NSF of China (Grant Nos. 10901076, 10931001 and 11171345) and the Key Laboratory of Mathematics and Complex System (Beijing Normal University), Ministry of Education, China. 1 2 and ZUNWEI FU, SHAOGUANG SHI, AND SHANZHEN LU Za<x<b K(x)dx = 0, (1.5) for all a, b with 0 < a < b. Chiarenza and Frasca [2] showed the boundedness of eT on Mp,q(Rn). Here and subsequently, C will denote a positive constant which may vary from line to line but will remain independent of the relevant quantities. For 0 < α < n, the fractional integral operator Iα is defined by Iαf (x) =ZRn f (y) x − yn−α dy. The boundedness of Iα on Mp,q(Rn) was established by Adams in [1]. For some works on the boundedness for the multilinear singular integral operators on Morrey type spaces, we refer to [5], [21] and [23]. In [7], Komori and Shirai gave the definition of weighted Morrey space, which is a natural generalization of weighted Lebesgue space. Let 1 ≤ p < ∞, 0 < k < 1 and w be a function. Then the weighted Morrey space Mp,k(w) was defined by Mp,k(w) = {f ∈ Lp loc(w) : kf kMp,k(w) < ∞}, where kf kMp,k(w) = sup B (cid:18) 1 w(B)kZB f (x)pw(x)dx(cid:19) 1 p , and the supremum is taken over all balls B ⊂ Rn. It is obviously that if w = 1, k = 1 − p q , then Mp,k(w) = Mp,q(Rn). For w ∈ Ap(1 ≤ p < ∞), if k = 0, then Mp,0(w) = Lp(w) and if k = 1, Mp,1(w) = L∞(w). Also, Komori and Shirai [7] obtained the boundedness of Hardy- Littlewoood maximal operator M and Calder´on-Zygmund singular integral operator eT on Mp,k(w) with 1 < p < ∞. When p = 1, the corresponding weighted weak (1, 1) type boundedness was also true [7]. Here Ap and the following A(p,q) denote the Muckenhoupt classes [15] and Ap : sup w(x)1−p′ B (cid:18) 1 BZB B (cid:18) 1 BZB w(x)dx(cid:19)(cid:18) 1 BZB w(x)qdx(cid:19) 1 q(cid:18) 1 BZB A(p,q) : sup w(x)−p′ dx(cid:19)p−1 dx(cid:19) 1 p′ ≤ C, 1 < p < ∞ ≤ C, 1 < p, q < ∞ respectively. Here 1/p + 1/p′ = 1. In the fractional case, we need to consider a weighted Morrey space with two weights which also introduced by Komori and Shirai [7]. Let 1 ≤ p < ∞, 0 < k < 1. For two weights w1 and w2, Mp,k(w1, w2) =(cid:8)f : kf kMp,k(w1,w2) < ∞(cid:9) , BOUNDEDNESS OF OSCILLATORY INTEGRAL OPERATORS AND THEIR COMMUTATORS 3 with the norm kf kMp,k(w1,w2) = sup B (cid:18) 1 w2(B)kZB f (x)pw1(x)dx(cid:19) 1 p , and the supremum is taken over all balls B ⊂ Rn. If w1 = w2 = w, then we denote Mp,k(w1, w1) = Mp,k(w2, w2) = Mp,k(w). In [7], the authors considered the weighted estimates of the fractional maximal operator Mα Mαf (x) = sup B∋x 1 Bn−αZB f (y)dy (1.6) and the fractional integral operator Iα with 0 < α < n. It is worth pointing out that the kernel in (1.2) is convolution kernel. However, there were many kinds of operators with non-convolution kernels, such as Fourier transform and Radon transform [13] which both are versions of oscillatory integrals. The object we consider in this paper is a class of oscillatory integrals due to Ricci and Stein [17] T f (x) = p.v.ZRn eiP (x,y)K(x − y)f (y)dy, (1.7) where P (x, y) is a real valued polynomial defined on Rn × Rn, and K is a CZK. It is well known that the oscillatory factor eip(x,y) makes it impossible to establish the weighted norm inequalities of (1.7) by the method as in the case of Calder´on-Zygmund operators or fractional integrals. In [18], Sato established the weighted weak (1, 1) type estimate of T . The strong type weighted norm inequality of T was proved by Lu and Zhang in [11] with a more general case. For the other classical works about oscillatory integral, we refer to [9], [10] and [12]. Inspired by [7] and [18], we will study the weak type estimates for T on weighted morrey space, that is Theorem 1.1. If K is a CZK, w ∈ A1 and 0 < k < 1, then there exists a constant C independent on the coefficients of P such that λw ({x ∈ B : T f (x) > λ}) ≤ Ckf kM1,k(w)w(B)k. sup λ>0 A distribution kernel K is called a standard Calder´on-Zygmund kernel(SCZK) if it satisfies the following hypotheses K(x, y) ≤ C x − yn , x 6= y and ∇xK(x, y) + ∇yK(x, y) ≤ C x − yn+1 , x 6= y. (1.8) (1.9) The corresponding Calder´on-Zygmund integral operator eS and oscillatory integral oper- ator S are defined by K(x, y)f (y)dy (1.10) eSf (x) = p.v.ZRn ZUNWEI FU, SHAOGUANG SHI, AND SHANZHEN LU Sf (x) = p.v.ZRn eiP (x,y)K(x, y)f (y)dy. (1.11) where P (x, y) is a real valued polynomial defined on Rn × Rn. In 1992, Lu and zhang [11] proved that S was bounded on Lp(w) with 1 < p < ∞, w ∈ Ap by the methods of interpolation of operators with change of measures [22]. In [17], Ricci and Stein also introduced the standard fractional Calder´on-Zygmund kernel(SFCZK) Kα with 0 < α < n, where the condition (1.8) and (1.9) were replaced by 4 and and Kα(x, y) ≤ C x − yn−α , x 6= y (1.12) (1.13) (1.14) x − yn+1−α , x 6= y. The corresponding fractional oscillatory integral operator is defined by ∇xKα(x, y) + ∇yKα(x, y) ≤ C Sαf (x) =ZRn eiP (x,y)Kα(x, y)f (y)dy. where P (x, y) is also a real valued polynomial defined on Rn × Rn. Obviously, when α = 0, S0 = S and K0 = K. Recently, the authors of this paper obtained the weighted boundedness of Sα in [20]. Partly motivated by the idea from [7] and the results of [11] as well as [20], we now give another two results of this paper Theorem 1.2. Let 1 < p < ∞, 0 < k < 1 and w ∈ Ap. If the Calder´on-Zygmund singular integral operator eS is of type (L2(Rn), L2(Rn)), then for any real polynomial P (x, y), there exists constant C > 0 such that kSf (x)kMp,k(w) ≤ Ckf kMp,k(w), where its norm depends only on the total degree of P , but not on the coefficients of P . Theorem 1.3. Let 1 q = 1 p − α n , 1 < p < n α , 0 < k < p q , 0 < α < n and w ∈ A(p,q). For any real polynomial P (x, y), there exists constant C > 0 such that kSαf (x)kM q, qk p (wq) ≤ Ckf kMp,k(wp,wq), where its norm depends only on the total degree of P , but not on the coefficients of P . If p = 1 and w ∈ A(1,q) with q = n n−α , then for all λ > 0, there exists constant C > 0 such that wq ({x ∈ B : Sαf (x) > λ}) ≤ C λq kf kq M1,k(w,wq)(wq(B))kq. We prove Theorem 1.1-Theorem 1.3 in Section 2. In Section 3, we set up the weighted norm inequalities for the corresponding commutators of S and Sα, respectively. Through- out this paper all definitions and notations are standard. A weight w is a locally integrable function on Rn which takes values in (0, ∞) almost everywhere. B = B(x0, r) denotes the ball with center x0 and radius r. Given λ > 0, λB = B(x0, λr). BOUNDEDNESS OF OSCILLATORY INTEGRAL OPERATORS AND THEIR COMMUTATORS 5 2. Weighted estimates for oscillatory integral operators We begin this section with some properties of Ap weight classes which play important role in the proofs of our main results. Lemma 2.1. [4] Let 1 ≤ p < ∞, and w ∈ Ap. Then the following statements are true (1) There exists a constant C such that w(2B) ≤ Cw(B). When w satisfies this condition, we say w satisfies doubling condition. (2) There exists a constant C > 1 such that w(2B) ≥ Cw(B). (2.1) (2.2) When w satisfies this condition, we say w satisfies reverse doubling condition. (3) There exist two constant C and r > 1 such that the following reverse Holder inequality holds for every ball B ⊂ Rn (cid:18) 1 BZB (4) For all λ > 1, we have r w(x)rdx(cid:19) 1 ≤ C(cid:18) 1 BZB w(x)dx(cid:19) . w(λB) ≤ Cλnpw(B). (2.3) (2.4) (5) There exist two constant C and δ > 0 such that for any measurable set Q ⊂ B If w satisfies (2.5), we say w ∈ A∞. w(Q) w(B) B(cid:19)δ ≤ C(cid:18) Q . (6) For all 1 p + 1 p′ = 1 we have A∞ = [1<p<∞ Ap, w1−p′ ∈ Ap′ . (2.5) (2.6) Our argument based heavily on the following well-known results about T , S and Sα. Lemma 2.2. [18] If K is a CZK, w ∈ A1, then there exist constant C > 0 independent on the coefficients of P such that λw ({x ∈ Rn : T f (x) > λ}) ≤ Ckf kL1(w). sup λ>0 Lemma 2.3. [11] If K is a SCZK, w ∈ Ap(1 < p < ∞) and the Calder´on-Zygmund singular integral operator eS is of type (L2(Rn), L2(Rn)), then for any real polynomial P (x, y), there exists C > 0 such that kSf kLp(w) ≤ Ckf kLp(w), where its norm depends only on the total degree of P , but not on the coefficients of P . 6 ZUNWEI FU, SHAOGUANG SHI, AND SHANZHEN LU Lemma 2.4. [20], [7] Let w ∈ A(p,q) , 0 < α < n, 1 < p < n α and 1 p − 1 q = α n . Then there exists constant C > 0 independent of P such that such that kSαf kLq(wq) ≤ Ckf kLp(wp), and kIαf (x)kM q, qk p (wq) ≤ Ckf kMp,k(wp,wq). When p = 1 and w ∈ A(1,q) with q = n such that n−α , then for all λ > 0, there exists constant C > 0 wq ({x ∈ B : Iαf (x) > λ}) ≤ C λq kf kq M1,k(w,wq)(wq(B))kq. We first give the proof of Theorem 1.1. Decompose f = f χ2B + fχ(2B)c := f1 + f2. For any given λ > 0, we write w ({x ∈ B : T f (x) > λ}) ≤ w(cid:18)(cid:26)x ∈ B : T f1(x) > := I + II. An application of (2.1) and Lemma 2.2 yields that λ 2(cid:27)(cid:19) + w(cid:18)(cid:26)x ∈ B : T f2(x) > λ 2(cid:27)(cid:19) Next we turn to deal with the term II. An elementary estimate shows λ I ≤(cid:18)w(cid:26)x ∈ Rn : T f1(x) > w(cid:18)(cid:26)x ∈ B : T f2(x) > f (x)w(x)dx ≤ C λ Z2B 2(cid:27)(cid:19) ≤ 2(cid:27)(cid:19) =Z{x∈B:T f2(x)> λ λZ{x∈B:T f (x)> λ 2 } ≤ C λ 2 } w(x)dx T f2(x)w(x)dx. C λ kf kM1,k(w)w(B)k. (2.7) We note that for x ∈ B and y ∈ (2B)c, x0 − y < Cx − y. Applying (1.3), we conclude that eiP (x,y)K(x − y)f2(y)dy f2(y) x − yn dy T f2(x) = Z(2B)c ≤ CZRn ≤ CZx0−y>2r ∞Xj=1Z2j r<x0−y<2j+1r 2jBZ2j+1B ∞Xj=1 ≤ C ≤ C 1 f2(y) x0 − yn dy f (y) x0 − yn dy f (y)dy. BOUNDEDNESS OF OSCILLATORY INTEGRAL OPERATORS AND THEIR COMMUTATORS 7 Holder inequality and the Ap condition imply that II ≤ C λ ∞Xj=1 1 2jBZ2j+1B f (y)dyZ{x∈B:T f (x)> λ 2 } w(x)dx ≤ C λ Then, Theorem 1.1 is a by-product of (2.7)-(2.8). kf kM1,k(w)w(B)k. (2.8) We now give the proof of Theorem 1.2. As in [7], our method is adapted from [3] in the case of Lebesgue measure. Let 1 < p < ∞, 0 < k < 1, K be a SCZK. It suffices to show that 1 w(B)kZB Sf (x)pw(x)dx ≤ Ckf kp Mp,k(w). (2.9) For a fixed ball B = B(x0, r), we decompose f = f χ2B + fχ(2B)c := f1 + f2 and consider the corresponding splitting Sf (x) =Z2B eiP (x,y)K(x, y)f (y)dy +Z(2B)c =: Sf1(x) + Sf2(x). eiP (x,y)K(x, y)f (y)dy Since S is a linear operator, so we get 1 w(B)kZB Sf (x)pw(x)dx ≤ 1 w(B)kZB (Sf1(x)p + Sf2(x)p)w(x)dx := I + II. (2.10) It follows from Lemma 2.3 and (2.1) that I ≤ 1 w(B)kZRn Sf1(x)pw(x)dx ≤ C w(B)kZ2B f (x)pw(x)dx ≤ Ckf kp Mp,k(w). (2.11) We are now in a position to estimate the term II. We note that for x ∈ B and y ∈ (2B)c, x0 − y < Cx − y. Applying (1.8), we conclude that eiP (x,y)K(x, y)f2(y)dy f2(y) x − yn dy Sf2(x) = Z(2B)c ≤ CZRn ≤ CZx0−y>2r ∞Xj=1Z2j r<x0−y<2j+1r 2jBZ2j+1B ∞Xj=1 ≤ C ≤ C f2(y) x0 − yn dy f (y) x0 − yn dy 1 f (y)dy. 8 ZUNWEI FU, SHAOGUANG SHI, AND SHANZHEN LU Holder inequality and the Ap condition imply that Z2j+1B f (y)dy ≤(cid:18)Z2j+1B f (y)pw(y)dy(cid:19) 1 p(cid:18)Z2j+1B p(cid:18)Z2j+1B k w(y)− p′ p′ p dy(cid:19) 1 dy(cid:19)1− 1 p′ ≤ Ckf kMp,k(w)w(2j+1B) w(y)1−p′ ≤ Ckf kMp,k(w)2j+1Bw(2j+1B) 1 p (k−1). Then, using (2.2) we obtain II ≤ Ckf kp Mp,k(w) ∞Xj=1 w(B) 1−k p w(2j+1B) p !p 1−k ≤ Ckf kp Mp,k(w). (2.12) We get (2.9) by (2.10), (2.11) and (2.12). The proof of Theorem 1.2 is completed. For the last part of this section, we show the proof of Theorem 1.3. Theorem 1.3 is a byproduct of Lemma 2.4 and the following observation In fact, if 1 < p < ∞, 1 f (y) Sαf (x) ≤Z p Z Sαf (x)qwq ≤ qk w(B) x − yn−α dy = Iα(f )(x). 1 w(B) p Z Iαf (x)qwq qk In the same manner, we can obtain the result for p = 1. ≤ Ckf kMp,k(wp,wq). 3. Weighted estimates for the commutators The aim of this section is to set up the weighted boundedness for the commutators formed by S(Sα) and BMO(Rn) functions. A locally integrable function b is said to be in BMO(Rn) if for any ball B ⊂ Rn kbkBM O(Rn) = sup B 1 BZB b(x) − bBdx < ∞, where bB = 1 BRB b(x)dx. We next formulate some remarks about BMO(Rn). Lemma 3.1. [8],[24] Let 1 ≤ p < ∞, b ∈ BMO(Rn). Then for any ball B ⊂ Rn, the following statements are true (1) There exist constants C1, C2 such that for all α > 0 {x ∈ B : b(x) − bB > α} ≤ C1Be−C2α/kbkBM O(Rn ). (3.1) The inequality (3.1) is also called John-Nirenberg inequality. kbkBM O(Rn) ∼ sup kbkBM O(Rn) ∼ sup B p . b(x) − bBpdx(cid:19) 1 B (cid:18) 1 BZB BZB b(x) − adx. inf a∈R 1 (2) (3) (3.2) (3.3) (3.4) (3.5) BOUNDEDNESS OF OSCILLATORY INTEGRAL OPERATORS AND THEIR COMMUTATORS 9 (2) b2λB − bB ≤ 2nλkbkBM O(Rn). Lemma 3.2. [16] Let w ∈ A∞. Then the following statements are equivalent (1) kbkBM O(w) = sup B b(x) − bB,ww(x)dx. 1 w(B)ZB where BMO(w) = {b : kbkBM O(w) < ∞} and bB,w = 1 w(B)RB b(y)w(y)dy. For a locally integrable function b, the commutator formed by S(Sα) and b are defined by and Sb := [b, S]f (x) = b(x)Sf (x) − S(bf )(x) Sα,b := [b, Sα]f (x) = b(x)Sαf (x) − Sα(bf )(x). Our main results of this section are Theorem 3.3. Let p, k, w, K and eS be the same as Theorem 1.2. If b ∈ BMO(Rn), then for any real polynomial P (x, y), there exists constant C > 0 such that kSbf (x)kMp,k(w) ≤ Ckf kMp,k(w), where its norm depends only on the total degree of P , but not on the coefficients of P . Theorem 3.4. Let p, q, k, Kα and w be the same as in Theorem 1.3. Then for b ∈ BMO(Rn), there exists constant C > 0 such that kSα,bf (x)kM q, qk p (wq) ≤ Ckf kMp,k(wp,wq). The following results will play an important role in our analysis. Lemma 3.5. [19] Suppose K is a SCZK, w ∈ Ap(1 < p < ∞) and the operator eS is of type (L2(Rn), L2(Rn)). Then for any b ∈ BMO(Rn), there exists constants C > 0 independent on the coefficients of P such that kSbf kLp(w) ≤ Ckf kLp(w). 10 ZUNWEI FU, SHAOGUANG SHI, AND SHANZHEN LU Lemma 3.6. [7] Let 1 < p < n q , 0 < α < n, 1 for b ∈ BMO(Rn), there exists constant C > 0 such that α, 0 < k < p q = 1 p − α n and w ∈ A(p,q). Then kIα,bf (x)kM q, qk p (wq) ≤ Ckf kMp,k(wp,wq). Lemma 3.6 is the weighted version of Theorem 1 in [6]. The following Proposition is essential to the proof of Theorem 3.3. Proposition 3.7. Let B = B(x0, r), 0 < k < 1 and 1 < p < ∞. Then the inequality (cid:18)Zx0−y>2r f (y) x0 − yn bB,w − b(y)dy(cid:19)p w(B)1−k ≤ Ckf kp Mp,k(w)kbkp BM O(Rn). (3.6) holds for every y ∈ (2B)c, where (2B)c = Rn/2B. Proof. Using Holder's inequality to the left-hand-side of (3.6), we have w(B)1−k f (y) x0 − yn bB,w − b(y)dy(cid:19)p w(B)1−k 1 (cid:18)Zx0−y>2r ≤ ∞Xj=1Z2jr<x0−y<2j+1r ≤ ∞Xj=1 2jBZ2j+1B ≤ C" ∞Xj=1 2jB(cid:18)Z2j+1B Mp,k(w)" ∞Xj=1 ≤ Ckf kp 1 w(B)1−k f (y) x0 − yn bB,w − b(y)dy!p f (y)bB,w − b(y)dy!p f (y)pw(y)dy(cid:19) 1 (cid:18)Z2j+1B p(cid:18)Z2j+1B 2jB w(2j+1B) k p p′#p dy(cid:19) 1 w(B)1−k bB,w − b(y)p′ w(y)1−p′ w(y)1−p′ p′#p dy(cid:19) 1 bB,w − b(y)p′ w(B)1−k. For the simplicity of analysis, we denote A as (cid:18)Z2j+1B bB,w − b(y)p′ w(y)1−p′ p′ dy(cid:19) 1 . By an elementary estimate, we have (b2j+1B,w1−p′ − b(y) + b2j+1B,w1−p′ − bB,w)p′ w(y)1−p′ p′ dy(cid:19) 1 A ≤(cid:18)Z2j+1B ≤(cid:13)(cid:13)(cid:13)(cid:13) ≤(cid:18)Z2j+1B =: J + JJ. b2j+1B,w1−p′ − b(·) + b2j+1B,w1−p′ − bB,w w(·) b2j+1B,w1−p′ − b(y)w(y)1−p′ (cid:13)(cid:13)(cid:13)(cid:13)Lp′ (w) p′ dy(cid:19) 1 + b2j+1B,w1−p′ − bB,ww1−p′ (2j+1B) 1 p′ BOUNDEDNESS OF OSCILLATORY INTEGRAL OPERATORS AND THEIR COMMUTATORS 11 For the term J, Lemma 3.2 implies J ≤ CkbkBM O(w1−p′ )w1−p′ (2j+1B) 1 p′ ≤ Cw1−p′ (2j+1B) 1 p′ . (3.7) To deal with JJ, by (3.2), we have b2j+1B,w1−p′ − bB,w ≤ b2j+1B,w1−p′ − b2j+1B + b2j+1B − bB + bB − bB,w b(y) − b2j+1Bw(y)1−p′ dy + 2n(j + 1)kbkBM O(Rn) ≤ 1 w1−p′(2j+1B)Z2j+1B w(B)ZB + 1 := JJ1 + JJ2 + JJ3. b(y) − bBw(y)dy Combining (2.5) with (3.1), JJ3 = w({x ∈ B : b(y) − bB > α})dα 1 w(B)Z ∞ ≤ CZ ∞ 0 0 ≤ C. e−C2αδ/kbkBM O(Rn )dα In the same manner we can see that JJ1 ≤ C. It follows immediately that JJ ≤ C(2n(j + 1) + 2)w1−p′ (2j+1B) 1 p′ . (3.8) As a by-product of (3.7) and (3.8), we have A ≤ C(j + 1)w1−p′ (2j+1B) 1 p′ . Then, applying (2.2), the proof of (3.6) based on the following observation " ∞Xj=1 ≤ C" ∞Xj=1 w(2j+1B) k p 2jB (cid:18)Z2j+1B p #p 1−k 1−k p (j+1) w(B) w(2j+1B) b(y) − bB,wp′ w(y)1−p′ p′#p dy(cid:19) 1 w(B)1−k = C. (cid:3) Proof of Theorem 3.3. The task is now to find a constant C such that for fixed ball B = B(x0, r), we can obtain 1 w(B)kZB Sbf (x)p w(x)dx ≤ Ckf kp Mp,k(w). (3.9) 12 ZUNWEI FU, SHAOGUANG SHI, AND SHANZHEN LU We decompose f = f χ2B + fχ(2B)c := f1 + f2, and consider the corresponding splitting ZB Sbf (x)p w(x)dx ≤ C(cid:18)ZB Sbf1(x)pw(x)dx +ZB =: K + KK. Sbf2(x)pw(x)dx(cid:19) An application of Lemma 3.5 and w ∈ Ap yields K ≤ CZ2B f (x)pw(x)dx ≤ Ckf kp Mp,k(w)w(B)k. (3.10) To estimate the other term KK, we note that for x ∈ B and y ∈ (2B)c, x0 − y < Cx − y. Then a further use of (1.8) derives that eip(x,y)K(x, y)f2(y)(b(x) − b(y))dy(cid:12)(cid:12)(cid:12)(cid:12) f2(y)b(x) − b(y) p dy(cid:19)p Sbf2(x)p =(cid:12)(cid:12)(cid:12)(cid:12)ZRn ≤ C(cid:18)ZRn ≤ C(cid:18)Zx0−y>2r x − yn f (y) x0 − yn {b(x) − bB,w + bB,w − b(y)}dy(cid:19)p . where bB,w = 1 w(B)RB b(x)w(x)dx. Then, we have x0 − yn dy(cid:19)pZB KK ≤ C(cid:18)Zx0−y>2r f (y) b(x) − bB,wpw(x)dx + C(cid:18)Zx0−y>2r f (y) x0 − yn b(y) − bB,wdy(cid:19)p w(B) := KK1 + KK2. A further use of proposition 3.7, we get KK2 ≤ Ckf kp Mp,k(w)w(B)k. To get the desired estimate, we are led to estimate the term KK1. This estimate will be done via (2.1), (2.3) and Lemma 3.2. BOUNDEDNESS OF OSCILLATORY INTEGRAL OPERATORS AND THEIR COMMUTATORS 13 KK1 = ∞Xj=1Z2jr<x0−y<2j+1r 2jBZ2j+1B 2jB(cid:18) ≤ ∞Xj=1 ∞Xj=1 ≤ C 1 1 1 b(x) − bB,wpw(x)dx f (y) x0 − yn dy!pZB f (y)dy!pZB w(2j+1B)kZ2j+1B p(cid:18)Z2j+1B 2jB (cid:18) 1 2j+1B− 1 w(y)− 1 k p b(x) − bB,wpw(x)dx p f (y)pw(y)dy(cid:19) 1 p−1 dy(cid:19) p−1 p ZB w(y)dy(cid:19)− 1 2j+1BZ2j+1B p × w(2j+1B) ≤ Ckf kMp,k(w) ∞Xj=1 ×ZB b(x) − bB,wpw(x)dx w(2j+1B) k p!p b(x) − bB,wpw(x)dx 1−k p ∞Xj=1 w(B) w(2j+1B) p !p 1−k w(B)k ≤ Ckf kp Mp,k(w)kbkp BM O(Rn) ≤ Ckf kp Mp,k(w)w(B)k Hence KK ≤ Ckf kp Lp,k(w)w(B)k. (3.11) Combing (3.10), (3.11), we obtain (3.9), which is the desired conclusion. Proof of Theorem 3.4. As in the proof of Theorem 1.3. Theorem 3.4 can be deduced via Lemma 3.6 and the following observation Sbf (x) ≤Z f (y) x − yn−β dy = Iα(f )(x). References [1] Adams D.R., A note on Riesz potentials, Duke Math. J. 42(1975), 765 -- 778. [2] Chiarenza F. and Frasca M., Morrey spaces and Hardy-Littlewood maximal function, Rend. Math. Appl. 7(7)(1987), 273 -- 279. [3] Fan D.S., Lu S.Z. and Yang D.C., Regularity in Morry spaces of strong solutions to nondiver- gence elliptic equations with VMO coefficients, Georgian Math. J. 5(5)(1998), 425 -- 440. [4] Grafakos L., Classical and Modern Fourier Analysis, Pearson Education, Inc. Upper Saddle River, New Jersey, 2004. [5] Grafakos L. and Torres R., Multilinear Calder´on-Zygmund theory, Adv. Math, 165(2002), 124 -- 164. [6] Komori Y. and Mizuhara T., Notes on commutators and Morrey spaces. Hokkaido Math. J. 32(2003), 345 -- 353. 14 ZUNWEI FU, SHAOGUANG SHI, AND SHANZHEN LU [7] Komori Y. and Shirai S., Weighted Morrey spaces and a singular integral operator, Math. Nachr. 282(2009), 219 -- 231. [8] Lu S.Z., Four Lectures on Real H p spaces, World Scientific Publishing, Singapore, 1995. [9] Lu S.Z., Multilinar oscillatory integrals with Calder´on-Zygmund kernel, Sci. China(Ser.A) 42(1999), 1039 -- 1046. [10] Lu S.Z., A class of oscillatory integrals, Int. J. Appl. Math. Sci. 2(1)(2005), 42 -- 58. [11] Lu S.Z. and Zhang Y., Weighted norm inequality of a class of oscillatory singular operators, Chin. Sci. Bull. 37(1992), 9 -- 13. [12] Lu S.Z.and Xia X., Boundedness of commutators of an oscillatory singular operators, Studia Math. 186(2008), 15 -- 27. [13] Phong D.H. and Stein E.M., Singular integrals related to the Radon transform and boundary value problems, Proc. Nat. Acad. USA. 80(1983), 7697 -- 7701. [14] Morrey C.B., On the solutions of quasi-linear elliptic partial differential equations, Trans. Amer. Math. Soc. 43(1938), 126 -- 166. [15] Muckenhoupt B., Weighted norm inqualities for the Hardy maximal function, Trans. Amer. Math. Soc. 165(1972), 207 -- 226. [16] Muckenhoupt B. and Wheeden R., Weighted bounded mean oscillation and the Hilbert trans- form, Studia Math. 54(1976), 221 -- 237. [17] Ricci F. and Stein E.M., Harmonic analysis on Nilpotant groups and singular integrals I: Oscil- latory Integrals, J. Funct. Anal. 73(1987), 179 -- 194. [18] Sato S., Weighted weak type (1,1) estimates for oscillatory integrals, Studia Math. 47(2001), 1 -- 17. [19] Shi S.G., Weighted boundedness for the commutators of one class of oscillatory integral operators, J. Beijing Normal Univ. (Nat. Sci.) 47(2011), 344 -- 346. [20] Shi S.G., Fu Z.W. and Lu S.Z., Weighted norm inequalities for fractional oscillatory integrals and their commutators, arXiv:1111.3418v1. [21] Shi Y.L., Tao X.X., Multilinear Riesz potential operators on Herz-type spaces and generalized Morrey spaces, Hokkaido Math. J. 38 (2009), 635 -- 662. [22] Stein E.M. and Weiss G., Interpolation of operators with change of measures, Trans. Amer. Math. Soc. 87(1958), 159 -- 172. [23] Tao X.X., Yu X., Zhang H.H., Multilinear Calder´on-Zygmund operators on variable exponent Morrey spaces over domains, Appl. Math. J. Chinese Univ. 26(2)(2011), 187 -- 197. [24] Torchinsky A., Real variable Methods in Harmonic Analysis Academic Press, San DIego, 1986. School of Sciences Linyi University Linyi 276005 P. R. China E-mail address: [email protected] School of Mathematical Sciences Beijing Normal University Beijing 100875 and School of Sciences Linyi University BOUNDEDNESS OF OSCILLATORY INTEGRAL OPERATORS AND THEIR COMMUTATORS 15 Linyi 276005 P. R. China E-mail address: [email protected] School of Mathematical Sciences Beijing Normal University Beijing 100875 P. R. China E-mail address: [email protected]
1905.05971	1	1905	2019-05-15T06:39:36	Orlicz Modules over Coset Spaces of Compact Subgroups in Locally compact Groups	[ "math.FA" ]	Let $H$ be a compact subgroup of a locally compact group $G$ and let $m$ be the normalized $G$-invariant measure on homogeneous space $G/H$ associated with Weil's formula. Let $\varphi$ be a Young function satisfying $\Delta_2$-condition. We introduce the notion of left module action of $L^1(G/H, m)$ on the Orlicz spaces $L^\varphi(G/H, m).$ We also introduce a Banach left $L^1(G/H, m)$-submodule of $L^\varphi(G/H, m).$	math.FA	math	ORLICZ MODULES OVER COSET SPACES OF COMPACT SUBGROUPS IN LOCALLY COMPACT GROUPS VISHVESH KUMAR Abstract. Let H be a compact subgroup of a locally compact group G and let m be the normalized G-invariant measure on homogeneous space G/H associated with Weil's formula. Let ϕ be a Young function satisfying ∆2-condition. We introduce the notion of left module action of L1(G/H, m) on the Orlicz spaces Lϕ(G/H, m). We also introduce a Banach left L1(G/H, m)- submodule of Lϕ(G/H, m). 9 1 0 2 y a M 5 1 ] . A F h t a m [ 1 v 1 7 9 5 0 . 5 0 9 1 : v i X r a 1. Introduction The abstract theory of Banach modules or Banach algebras plays an important role in var- ious branches of Mathematics, for instance, abstract harmonic analysis, representation theory, operator theory; see [4, 10, 11, 8] and the references therein. In particular, convolution structure on the Orlicz spaces to be an Banach algebra or a Banach module over a locally compact group or hypergroup were studied by many researchers [1, 9, 21, 13, 14, 15]. In [13], the author defined and studied the notion of abstract Banach convolution algebra on Orlicz spaces over homogeneous spaces of compact groups. Recently, Ghaani Farashahi [8] introduced the notion of abstract Banach convolution function module on the Lp-space on coset spaces of compact subgroups in locally compact groups. The purpose of this article is to define and study a new class of abstract Banach module on Orlicz spaces over coset spaces of compact subgroups in locally compact groups. Let us remark that Orlicz spaces are genuine generalization of Lebesgue spaces. It is worth mentioning that an appropriate use of Jensen's inequality [19, pg. 62] plays a key role in this article. In the next section, we present some basics of Orlicz spaces and some classical harmonic analysis on a homogeneous space (the space of left cosets) of a locally compact group. Section 3 is devoted to the study of abstract convolution module structure on the Orlicz space Lϕ(G/H, m), where H is a compact subgroup of a locally compact group G, m is the normalized G-invariant measure on the homogeneous space G/H which satisfies Weil's formula and ϕ is a Young function 2010 Mathematics Subject Classification. Primary 43A85, 46E30; Secondary 43A15, 43A20. Key words and phrases. Homogeneous space, Convolution function modules, Orlicz spaces. 1 2 VISHVESH KUMAR satisfying ∆2-condition. In this section, we prove that Lϕ(G/H, m) is a Banach left L1(G/H, m)- module with respect to a generalized convolution. We also introduce a Banach left L1(G/H, m)- submodule of Lϕ(G/H). 2. Preliminaries A non-zero convex function ϕ : R → [0, ∞] is called a Young function if it is even, left continuous with ϕ(0) = 0 and lim x→∞ ϕ(x) = ∞. Here we note that every Young function is an integral of a non-decreasing left continuous function [19, Theorem 1]. Let Ω be a locally compact Hausdorff space and m be a positive Radon measure on Ω. Denote the space of all equivalence classes of m-measurable functions on Ω by L0(Ω). A Young function ϕ satisfies ∆2-condition if there exist a constant C > 0 and x0 > 0 such that ϕ(2x) ≤ Cϕ(x) for all x ≥ x0 if m(Ω) < ∞ and ϕ(2x) ≤ Cϕ(x) for all x ≥ 0 otherwise. Given a Young function ϕ, the modular function ρϕ : L0(Ω) → R is defined by ρϕ(f ) :=RΩ ϕ(f ) dm. We always assume that Young function ϕ satisfies ∆2-condition. For a given Young function ϕ, the Orlicz space Lϕ(Ω, m), in short Lϕ(Ω), is defined by Lϕ(Ω) :=(cid:8)f ∈ L0(Ω) : ρϕ(af ) < ∞ for some a > 0(cid:9) . Then the Orlicz space is a Banach space with respect to the norm k · k0 ϕ on Lϕ(Ω) called Luxemburg norm or gauge norm which is defined by kf k0 ϕ := inf(cid:26)k > 0 :ZG ϕ(cid:18) f k (cid:19) dm ≤ 1(cid:27) . If ϕ(x) = xp 1 ≤ p < ∞ then Lϕ(Ω) is usual Lp-spaces, 1 ≤ p < ∞. An example of Young function which satisfies ∆2-condition and gives an Orlicz space other than Lp-spaces is given by ϕ(x) = (e + x) log(e + x) − e. We denote the space of all continuous functions on Ω with compact support by Cc(Ω). It is well known that if ϕ satisfies ∆2-condition then Cc(Ω) is a dense subspace of Lϕ(Ω). If A is a measurable subset of Ω such that 0 < m(A) < ∞, then we have the Jensen's Inequality ϕ(cid:18)RA f dm m(A) (cid:19) ≤ RA ϕ(f ) dm m(A) . We make use of the above inequality several times in this article. In this paper, we also employ the notations of the author in [3, 8]. For a locally compact group G with the Haar measure dx and f, g ∈ Lϕ(G), define convolution ∗G on Lϕ(G, m) by f ∗G g(x) =ZG f (y) g(y−1x) dx (x ∈ G). ORLICZ MODULES OVER HOMOGENEOUS SPACES 3 It is well-known that (Lϕ(G), k · k0 ϕ) is a Banach algebra with respect to the convolution product ∗G (see [12]) that is, kf ∗G gk0 Φ ≤ kf k0 ϕ kgk0 ϕ for all f, g ∈ Lϕ(G). Also, if f ∈ L1(G) and f ∈ Lϕ(G) then the above convolution define a module action of L1(G) on Lϕ(G) which makes Lϕ(K) a Banach left L1(K)-module, that is, (1) kf ∗G gk0 ϕ ≤ kf k1 kgk0 ϕ for all f ∈ L1(G) and g ∈ Lϕ(G) (see [14, 18]). Let H be a compact subgroup of a locally compact group G with the normalized Haar measure dh. The left coset space G/H can be seen as a homogeneous space with respect to the action of G on G/H given by left multiplication. The canonical surjection q : G → G/H is given by q(x) = xH. Define TH(f )(xH) =RH f (xh) dh, then Cc(G/H) = {TH (f ) : f ∈ Cc(G)} The homogeneous space G/H has a unique normalized G-invariant positive Radon measure m that satisfies the Weil's formula (2) ZG/H TH (f )(xH) dm(xH) =ZG f (x) dx, and hence kTH(f )kL1(G/H,m) ≤ kf kL1(G) for all f ∈ L1(G). For more details on harmonic analysis on homogeneous spaces of locally compact groups see [3, 4, 5, 6, 7, 22, 8]. 3. Orlicz modules over Coset Spaces of compact subgroups of locally compact groups Throughout this section, we assume that the Young function ϕ satisfies ∆2- condition, G is a locally compact group with a Haar measure dx and H is a compact subgroup of G with the normalized Haar measure dh. It is also assumed that the homogeneous space G/H has the normalized G-invariant measure m satisfying the Weil's formula. In this section, we show that the space Lϕ(G/H) becomes a Banach left L1(G/H, m)-module with respect to the convolution on Cc(G/K) defined in [8]. We also define a subspace of Lϕ(G/H) and show that this subspace is a Banach left L1(G/H, m)-submodule of Lϕ(G/H, m). We begin this section with following result. Theorem 3.1. Let G be a locally compact group and let H be a compact subgroup of G. Let m be the normalized G-invariant measure on the coset space G/H. Then the linear map TH : 4 VISHVESH KUMAR Cc(G) → Cc(G/H) satisfies (3) kTH(f )k0 Lϕ(G/H,m) ≤ kf k0 Lϕ(G) for all f ∈ Cc(G). Further, if the Young function ϕ satisfies ∆2-condition. TH can be uniquely extended to a linear map from Lϕ(G) onto Lϕ(G/H, m). Proof. For f ∈ Cc(G) and k > 0, by using Weil's formula and Jensen's inequality, we have ρϕ(cid:18) TH (f ) k k k f (xh) f (xh) ϕ(cid:18) TH (f )(xH) (cid:19) dm(xH) k (cid:19) = ZG/H ϕ(cid:18)(cid:12)(cid:12)(cid:12)(cid:12) dh(cid:12)(cid:12)(cid:12)(cid:12) (cid:19) dm(xH) ZH = ZG/H ϕ(cid:18)ZH dh(cid:19) dm(xH) ≤ ZG/H ϕ(cid:18) f (xh) k (cid:19) dh dm(xH) ≤ ZG/HZH k (cid:19) (xh) dh(cid:19) dm(xH) ϕ(cid:18) f = ZG/H(cid:18)ZH TH(cid:18)ϕ(cid:18) f k (cid:19)(cid:19) (xH) dm(xH) = ZG/H ϕ(cid:18) f k (cid:19) dx = ρϕ(cid:18) f k(cid:19) . = ZG Now, kf k0 Lϕ(G) = inf{k > 0 : ρϕ(cid:18) f k(cid:19) ≤ 1} k (cid:19) ≤ 1} = kTH (f )k0 ≥ inf{k > 0 : ρϕ(cid:18) TH(f ) Lϕ(G/H,m). Therefore, we get kTH (f )k0 Lϕ(G/H,m) ≤ kf k0 Lϕ(G) for all f ∈ Cc(G). Since φ is ∆2-regular, Cc(G) and Cc(G/H) are dense in Lϕ(G) and Lϕ(G/H, m) respectively. Therefore, we can extend TH to a bounded linear map from Lϕ(G) onto Lϕ(G/H, m). We denote this extension of TH again by TH and it satisfies (3). (cid:3) Proposition 3.2. Let G be a locally compact group and let H be a compact subgroup of G. Let m be the normalized G-invariant measure on the coset space G/H. Suppose that ϕ satisfies ∆2-condition. If f ∈ Lϕ(G/H, m) and fq := f ◦ q then we have fq ∈ Lϕ(G) with (4) kfqk0 Lϕ(G) = kf k0 Lϕ(G/H,m). ORLICZ MODULES OVER HOMOGENEOUS SPACES 5 Proof. For f ∈ Lϕ(G/H, m) and k > 0, by Weil's formula and the fact that H is compact, we have ρϕ(cid:18) fq ϕ(cid:18) fq(x) k (cid:19) dx TH(cid:18)ϕ(cid:18) fq k (cid:19) = ZG = ZG/H = ZG/H(cid:18)ZH = ZG/H(cid:18)ZH = ZG/H(cid:18)ZH = ZG/H k (cid:19)(cid:19) (xH) dm(xH) (cid:19) dh(cid:19) dm(xH) (cid:19) dh(cid:19) dm(xH) (cid:19) dh(cid:19) dm(xH) (cid:19) dm(xH)!(cid:18)ZH k(cid:17) and consequently, we get kfqk0 ϕ(cid:18) fq(xh) ϕ(cid:18) f (xhH) ϕ(cid:18) f (xH) ϕ(cid:18) f (xH) k k k k dh(cid:19) = ρϕ(cid:18) f k(cid:19) . Lϕ(G) = kf k0 Lϕ(G/H,m). (cid:3) Therefore ρϕ(cid:16) fq k (cid:17) = ρϕ(cid:16) f Remark 1. Note that TH (fq) = f and therefore, it is clear from above corollary that for fq ∈ Lϕ(G), the equality in Theorem 3.1 holds. Here we set certain terminologies for further use. For any continuous function f define the left translation by Lhf (·) = f (h−1(·)) and the right translation by Rhf = f ((·)h). Let G be a locally compact group and let H be a compact subgroup of G. We set Cc(G : H) = {f ∈ Cc(G) : Rhf = f ∀ h ∈ H}, A(G : H) = {f ∈ Cc(G) : Lhf = f ∀ h ∈ H} and A(G/H) = {g ∈ Cc(G/H) : Lhg = g ∀ h ∈ H}. For a Young function ϕ ∈ ∆2, define Aϕ(G : H) = {f ∈ Lϕ(G) : Lhf = f ∀ h ∈ H}, and also Aϕ(G/H, m) = {g ∈ Lϕ(G/H, m) : Lhg = g ∀ h ∈ H}. Note that Aϕ(G/H, m) is the topological closure of A(G/H) in Lϕ(G/H, m) and therefore it is a closed subspace of Lϕ(G/H, m). Similarly Aϕ(G : H) is closed subspace of Lϕ(G). It is known that TH maps Cc(G : H) and A(G : H) onto Cc(G/H) and A(G/H) respectively (see [8, Proposition 4.2]). 6 VISHVESH KUMAR Since A(G : H) is dense in Lϕ(G : H) and A(G/H) is dense in Aϕ(G/H), the next lemma follows from the continuity of TH. Lemma 3.3. Let G be a locally compact group and let H be a compact subgroup of G. Let m be the normalized G-invariant measure on the coset space G/H. Suppose that ϕ satisfies ∆2-condition. Then the mapping TH maps Lϕ(G : H) onto Aϕ(G/H, m). For g ∈ Cc(G/H), note that the mapping xH 7→ RH g(hxH) dh is in Cc(G/H). Define J : Cc(G/H) → Cc(G/H) as It is clear that J is a linear operator. In addition, the following theorem says that the norm of Jg(xH) =ZH g(hxH) dh. J is bounded operator with the norm bounded by one. Theorem 3.4. For f ∈ Cc(G/H), we have kJf kLϕ(G/H,m) ≤ kf kLϕ(G/H,m). Proof. For f ∈ Cc(G/H) we have, ρϕ(cid:18)Jf 1 k ϕ(cid:18) Jf (xH) k (cid:19) = ZG/H (cid:19) dm(xH) ϕ(cid:18)(cid:12)(cid:12)(cid:12)(cid:12) f (hxH) dh(cid:12)(cid:12)(cid:12)(cid:12) = ZG/H k ZH ϕ(cid:18) f (hxH) ≤ ZG/HZH = ZH ZG/H ϕ(cid:18) f (hxH) = ZH ZG/H ϕ(cid:18) f (xH) = ZH ZG/H ϕ(cid:18) f (cid:19) dm(xH) (cid:19) dh dm(xH) (cid:19) dm(xH)! dh (cid:19) dm(xH)! dh k k k k (cid:19) (xH) dm(xH)! dh = ρϕ(cid:18) f k(cid:19) . Consequently, we get kJf kLϕ(G/H,m) ≤ kf kLϕ(G/H,m) for all f ∈ Cc(G/H). (cid:3) It is shown in [8, Theorem 4.5 (2)] that J maps Cc(G/H) onto A(G/H). Now, the following corollary is immediate. Corollary 3.5. Let G be a locally compact group and let H be a compact subgroup of G. Let m be the normalized G-invariant measure on the coset space G/H. Suppose that ϕ satisfies ORLICZ MODULES OVER HOMOGENEOUS SPACES 7 ∆2-condition. The bounded linear map J : Cc(G/H) → A(G/H) can be uniquely extended to a bounded linear map Jϕ : Lϕ(G/H, m) → Aϕ(G/H, m) which satisfies kJϕf kLϕ(G/H,m) ≤ kf kLϕ(G/H,m). Further, the linear operator Jϕ : Lϕ(G/H, m) → Aϕ(G/H, m) is an onto map. Proof. The extension of the map J from Cc(G/H) to Lϕ(G/H) follows from Theorem 3.4 and the density of Cc(G/H, m) and A(G/H) in Lϕ(G/H, m) and Aϕ(G/H, m), respectively. Further, for any f ∈ Lϕ(G/H) and z ∈ H, we have Lz(Jf )(xH) = Jf (z−1xH) =ZH f (hz−1xH) dh =ZH f (hxH) dh = Jf (xH). This shows that Jf ∈ Aϕ(G/H, m). Now we prove second part of the corollary. For any f ∈ Aϕ(G/H), we get, Jf (xH) =ZH f (hxH) dh =ZH f (xH) dh = f (xH), for all x ∈ G. Therefore, Jf = f. Hence Jϕ : Lϕ(G/H, m) → Aϕ(G/H, m) is a onto map. (cid:3) Remark 2. It can seen in the proof of Corollary 3.5 above that JAϕ(G/H,m) = IAϕ(G/H,m). Now, we are ready to define the convolution product '∗G/H ' on Cc(G/H) same as given in [8] as follows: let G be a compact group, H a closed subgroup and let m be the normalized G-invariant measure on G/H. For f, g ∈ Cc(G/H), the convolution f ∗G/H g : G/H → C is given by (5) f ∗G/H g(xH) =ZG/H f (yH)Jg(y−1xH) dm(yH), for all xH ∈ G/H. The convolution product '∗G/H ' has the following properties similar to the usual convolution in Cc(G) (see [8, Proposition 4.10]). (i) For any f, g ∈ Cc(G/H), (f, g) 7→ f ∗G/H g is a bilinear map from Cc(G/H) × Cc(G/H) to Cc(G/H) and (Cc(G/H), ∗G/H ) is an algebra. (ii) f ∗G/H g = TH(fq ∗G gq) and (f ∗G/H g)q = fq ∗G gq, where '∗G' is the usual convolution in C(G). (iii) Lx(f ∗G/H g) = (Lxf ) ∗G/H g. The following result says that Cc(G/H) is a normed algebra with respect to the norm k · k0 Lϕ(G/H,m). Lemma 3.6. If f, g ∈ Cc(G/H), then (6) kf ∗G/H gk0 Lϕ(G/H,m) ≤ kf kL1(G/H,m)kgk0 Lϕ(G/H,m). 8 VISHVESH KUMAR Proof. Let f, g ∈ Cc(G/H). Using Proposition 3.2 we get kf ∗G/H gk0 Lϕ(G/H,m) = k(f ∗G/H g)qk0 Lϕ(G) = kfq ∗G gqk0 Lϕ(G) Since Lϕ(G) is a L1(G)-module then by Proposition 3.2 we get kf ∗G/H gk0 Lϕ(G/H,m) ≤ kfqkL1(G)kgqk0 Lϕ(G) = kf kL1(G/H,m)kgk0 Lϕ(G/H,m). (cid:3) Theorem 3.7. Let G be a locally compact group and let H be a compact subgroup of G. Let m be the normalized G-invariant measure on the coset space G/H. Suppose that ϕ satisfies ∆2-condition. Then the convolution ∗G/H : Cc(G/H) × Cc(G/H) → Cc(G/H) given by (5) can be extended to a convolution ∗ϕ G/H : L1(G/H, m) × Lϕ(G/H, m) → Lϕ(G/H, m) such that Lϕ(G/H, m) is a Banach left L1(G/H, m)-module with respect to this extended convolution. Proof. Let f ∈ L1(G/H, m) and g ∈ Lϕ(G/H, m). Since Cc(G/H) is dense in L1(G/H, m) and in Lϕ(G/H, m) as ϕ ∈ ∆2, there exist {fn} and {gn} in Cc(G/H) such that fn → f and gn → g as n → ∞. Now, define f ∗ϕ G/H g = lim n→∞ fn ∗G/H gn. Note that ∗ϕ G/H : L1(G/H, m) × Lϕ(G/H, m) → Lϕ(G/H, m) is well-defined. By Lemma 3.6 we have kf ∗ϕ G/H gk0 Lϕ(G/H,m) ≤ kf kL1(G/H,m)kgk0 Lϕ(G/H,m). Thus Lϕ(G/H, m) is a Banach left L1(G/H, m)-module with respect to the extended convolu- tion. (cid:3) The above theorem claims the existence of convolution product ∗ϕ G/H but it does not reveal any explicit formula for the convolution product. The following corollary fulfils this objective whose proof is a consequence of the fact that Cc(G/H) is dense Lϕ(G/H, m). Corollary 3.8. Let G be a locally compact group and let H be a compact subgroup of G. Let m be the normalized G-invariant measure on the coset space G/H. Suppose that ϕ satisfies ∆2-condition. If f ∈ L1(G/H, m) and g ∈ Lϕ(G/H, m), then the convolution ∗ϕ G/H is given by (7) for all xH ∈ G/H. (f ∗ϕ G/H g)(xH) =ZG/H f (yH)Jϕg(y−1xH) dm(yH) In the next corollary we present a Banach left L1(G/H, m)-submodule of Lϕ(G/H, m) whose proof is a routine check. ORLICZ MODULES OVER HOMOGENEOUS SPACES 9 Corollary 3.9. Let G be a locally compact group and let H be a compact subgroup of G. Let m be the normalized G-invariant measure on the coset space G/H. Suppose that ϕ satisfies ∆2- condition. Then the space Aϕ(G/H, m) is a Banach left L1(G/H, m)-submodule of Lϕ(G/H, m). Acknowledgements. The author thanks Prof. V. Muruganandam for his support and encour- agement. References [1] H. P. Aghababa, I. Akbarbaglu and S. Maghsoudi, The space of multipliers and convolution of Orlicz spaces on a locally compact group, Studia Math. 291(1) (2013) 19-34. [2] I. Akabarbaglu and S. Maghsoudi, Banach-Orlicz algebras on a locally compact group, Mediterr. J. Math., 10 (2013) 1937-1947. [3] A. Ghaani Farashahi, Abstract convolution function algebra over homogeneous spaces of compact groups, Illinois J. Math., 59(4) (2012) 1025-1042. [4] A. Ghaani Farashahi, A class of Abstract linear representation for convolution function algebra over homogeneous space of compact groups, Canad. J. Math., DOI : 10.4153/CJM- 2016-043-9. [5] A. Ghaani Farashahi, Abstract Plancherel (trace) formulas over homogeneous spaces of compact groups, Canad. Math. Bull., 60(2) (2017) 111-121. [6] A. Ghaani Farashahi, Abstract operator-valued Fourier transforms over homogeneous spaces of compact groups, Groups Geom. Dyn. 11(4) (2017) 1437-1467. [7] A. Ghaani Farashahi, Abstract measure algebras over homogeneous spaces of compact groups, International Journal of Mathematics, 29(1) (2018) 1850005, 34 pp. [8] A. Ghani Farashahi, Abstract Banach convolution function modules over coset spaces on compact subgroups in locally compact spaces, Bull. Braz. Math. Soc. (New Series) (2019). DOI: https://doi.org/10.1007/s00574-018-00129-6 [9] H. G. Feichtinger, Some new subalgebras of L1(G), Indag. Math. 36 (1974) 44-47. [10] H. G. Feichtinger, On a class of convolution algebra of functions, Ann. Inst. Fourier (Greno- ble), 27(3) (1977) 135-162. [11] H. G. Feichtinger, Banach convolution algebra of functions II, Monatch. Math. 87 (1979) 181-207. [12] H. Hudzik, A. Kamiska and J. Musielak, On some Banach algebras given by a modular, in: Alfred Haar Memorial Conference, Budapest, Colloquia Mathematica Societatis J´anos Bolyai (North-Holland, Amsterdam) 49 (1987) 445-463. [13] V. Kumar, R. Sarma and N. Shravan Kumar, Orlicz algebras on homogeneous spaces of compact groups and their abstract linear representations, Mediterr. J. Math. 15(4) Art. 186 (2018) 13 pp. 10 VISHVESH KUMAR [14] V. Kumar, R. Sarma and N. Shravan Kumar, Orlicz spaces on hypergroups, Publ. Math. Debrecen, 94(1-2) (2019) 399-414. [15] , V. Kumar and R. Sarma, The Hausdorff-Young inequality for Orlicz spaces on compact hypergroups, (accepted in Colloq. Math.) (2019). [16] M. A. Krasnosel'skii and Ja. B. Rutickii, Convex functions and Orlicz spaces, Noordhoff, Groningen (1961). [17] A. Osancliol and S. Oztop, Weighted Orlicz algebra on locally compact groups, J. Aust. Math. Soc., 99 (2015) 399-414. [18] M. M. Rao, Convolution of vector field-I, Math. Z., 74 (1980) 63-79. [19] M.M. Rao and Z.D. Ren, Theory of Orlicz spaces, Dekker, New York (1991). [20] M. M. Rao, Convolution of vector fields-II : random walk models, Nonlinear Analysis, 47 (2001) 3599-3615. [21] M. M. Rao, Convolutions of vector fields-III: Amenability and spectral properties, Real and stochastic analysis, Trends Math., Birkhuser Boston, Boston, MA, (2004) 375-401. [22] H. Reiter and J. D. Stageman, Classical harmonic analysis and locally compact groups, Clarendon Press, Oxford, (2000). Vishvesh Kumar School of Mathematical Sciences National Institute of Science Education and Research Bhubaneshwar, HBNI At/Po- Jatni, District- Khurda, Odisha- 7520250, India. E-mail address: [email protected]
1608.02463	1	1608	2016-08-08T14:48:56	Dirichlet forms for singular diffusion on graphs	[ "math.FA" ]	We describe operators driving the time evolution of singular diffusion on finite graphs whose vertices are allowed to carry masses. The operators are defined by the method of quadratic forms on suitable Hilbert spaces. The model also covers quantum graphs and discrete Laplace operators.	math.FA	math	Dirichlet forms for singular diffusion on graphs Christian Seifert and Jurgen Voigt Abstract We describe operators driving the time evolution of singular diffusion on finite graphs whose vertices are allowed to carry masses. The operators are defined by the method of quadratic forms on suitable Hilbert spaces. The model also covers quantum graphs and discrete Laplace operators. MSC 2010: 47D06, 60J60, 47E05, 35Q99, 05C99 Keywords: gap diffusion, quantum graph, Dirichlet form, C0-semigroup, positive, submarkovian Introduction The present paper is a continuation and extension of [2]. We present suitable boundary or glueing conditions on graphs (quantum graphs) with singular sec- ond order differential operators on the edges. In particular, we describe those boundary conditions leading to positive and submarkovian C0-semigroups. The graph consists of finitely many bounded intervals, the edges, whose end points are connected with the vertices of the graph. On each of the edges e a finite Borel measure µe is given, determining where particles may be located. The particles move according to "Brownian motion" but are slowed down or accelerated by the "speed measure" µe. Further, each of the vertices v is provided with a weight µv > 0, and particles may also be located at those vertices v with µv > 0. The motivations for the treatment in [2] were twofold. The first issue was to treat singular diffusion, including gap diffusion, on the edges of the graph, in the framework of Dirichlet forms. The second aim was to describe glueing conditions on the vertices, in the spirit of [4], and investigate conditions under which the associated self-adjoint operator gives rise to a positive or submarkovian C0-semigroup. In the present paper, the extension with respect to [2] consists in two issues. On the one hand, the boundary conditions we describe are more general than glueing conditions. By glueing conditions or "local boundary conditions", we understand conditions where, for a given vertex, only the values of a function on adjacent edges and on the vertex itself can interact. In our treatment in Sections 2 1 2 C. Seifert, J. Voigt and 3, however, the graph structure does not intervene at all, and we only specify later the case of local boundary conditions, in Section 5. On the other hand, we include the general case of vertices with masses, whereas in [2; Section 4] only a special case was treated. These results have been obtained in [7]. The ultimate objective of the treatment is to obtain a semi-bounded (below) self-adjoint operator H on a Hilbert space HΓ over the graph Γ which can then be used in the initial value problem for the diffusion equation or heat equation u′ = −Hu, (0.1) thus governing the time evolution of a process, i.e., giving rise to a C0-semigroup on HΓ. For this equation it is of interest to obtain H in such a way that the associated C0-semigroup is positive or submarkovian. The self-adjointness of H is also of interest for the initial value problem for the Schrodinger equation u′ = −iHu. The part of the operator H acting on an edge e is of the form (Hf )e = −∂µe ∂fe, where ∂µe is the derivative with respect to µe; cf. Section 1. The domain of H is restricted by conditions on the boundary values of the functions on the edges and the values at the vertices. The Hilbert space HΓ is given by L2([ae, be], µe) ⊕ KV , HΓ =Me∈E where E is the set of edges, the interval [ae, be] corresponds to the edge e, and V is the set of vertices; cf. Section 2 for more details. The operator H is obtained by the method of forms. Avoiding all technicalities (which will be given in Section 2), the form τ giving rise to H is of the form τ (f, g) =Xe∈EZ be ae f ′ e(x)g′ e(x) dx + (L tr f tr g), with domain D(τ ) =(cid:8)f ∈ . . . ; tr f ∈ X(cid:9). Here, tr f denotes the boundary values of f on the edges and the values of f on the vertices, X is a subspace of the set of possible boundary values and values on the vertices, and L is a self-adjoint operator (matrix) on X. The boundary conditions for functions in the domain of H are encoded in the space X as well as in the operator L; cf. Theorem 3.1. Our treatment includes the case that some of the edges or vertices may have weight zero. For the discussion of positivity and the submarkovian property in connection with equation (0.1) we use the Beurling-Deny criteria for τ . These yield the Dirichlet forms on graphs 3 result that the subspace X should satisfy lattice properties and L should satisfy positivity properties; cf. Theorem 4.1. The investigations mentioned so far did not take into account the graph struc- ture of Γ. In the description of glueing conditions, allowing only interactions between vertices and adjacent edges, the space X and the operator L decompose into parts corresponding to single vertices; cf. Corollaries 5.1 and 5.2. In Section 1 we recall some notation and facts from the one-dimensional case on an interval. In Section 2 we define the form in the Hilbert space HΓ on the graph which then defines the operator driving the evolution. We show that the defined form τ constitutes a form that is bounded below and closed. Let us point out that our definition of the form looks somewhat different from the one given in [2; Section 3]. In fact, looking at the definition of τ in [2; Section 3], one realises that there is some interpretation needed in order to understand D(τ ) as a subset of the Hilbert space HΓ. This interpreation is made explicit in the present paper by the use of the mapping ι introduced in Sections 1 and 2. In Section 3 we describe the operator H associated with the form τ (Theorem 3.1). In Section 4 we indicate conditions for the C0-semigroup (e−tH )t>0 to be positive and submarkovian. In Section 5 we describe the case of local boundary conditions. 1 One-dimensional prerequisites In order to define the classical Dirichlet form we have to recall some notation and facts for a single interval [a, b] ⊆ R, where a, b ∈ R, a < b. Let µ be a finite Borel measure on [a, b], a, b ∈ spt µ, µ({a, b}) = 0. Our function spaces will consist of K-valued functions, where K ∈ {R, C}. We define Cµ[a, b] :=(cid:8)f ∈ C([a, b]; f affine linear on the components of [a, b] \ spt µ(cid:9), 2 (a, b) ∩ Cµ[a, b]. 2,µ(a, b) := W 1 W 1 For later use we recall the following inequalities. There exists a constant C > 0 such that kf k∞ 6 C(cid:0)kf ′k2 L2(a,b) + kf k2 for all f ∈ W 1 2 (a, b) ∩ C[a, b], and for all r ∈ (0, b − a] one has L2([a,b],µ)(cid:1)1/2 (1.1) f (a) 6 r1/2kf ′kL2(a,a+r) + kf kL2([a,a+r],µ) µ([a, a + r])−1/2, (1.2) and correspondingly for b; cf. [2; Lemma 1.4 and Remark 3.2(b)]. Let κ : W 1 be shown that R(κ) = R(κ W 1 2 (a, b) ∩ C[a, b] → L2([a, b], µ) be defined by κf := f . Then it can 2,µ(a,b) is [2; Lemma 1.2]), and that κ W 1 2,µ(a,b)) (cf. injective (cf. [2; lower part of p. 639]). We define ι := (cid:0)κ W 1 2,µ(a,b)(cid:1)−1. Thus, ι is 4 C. Seifert, J. Voigt an operator from L2([a, b], µ) to W 1 2,µ(a, b), D(ι) =(cid:8)f ∈ L2([a, b], µ); there exists g ∈ W 1 2 (a, b) ∩ C[a, b] such that g = f µ-a.e.(cid:9), 2,µ(a, b) such that g = f µ-a.e. and ιf is the unique element g ∈ W 1 In order to describe the operator associated with the form defined in the fol- lowing section we need some additional notions and facts concerning derivatives with respect to µ. If f ∈ L1,loc(a, b), g ∈ L1([a, b], µ) are such that f ′ = gµ (where f ′ = ∂f denotes the distributional derivative of f ), then we call g distributional derivative of f with respect to µ, and we write ∂µf := g. Note that then necessarily f ′ = 0 on [a, b] \ spt µ, i.e., f is constant on each of the components of [a, b] \ spt µ. It is easy to see that this definition is equivalent to f (x) = c +Z(a,x) g(y) dµ(y) a.e., (1.3) with some c ∈ K. Thus, the function f has representatives of bounded variation and these have one-sided limits (not depending on the representative) at all points of [a, b]. 2 The form on the graph Let Γ = (V, E, γ) be a finite directed graph. This means that V and E are finite sets, V ∩ E = ∅, V is the set of vertices (or nodes) of Γ, E the set of edges, and γ = (γ0, γ1) : E → V × V associates with each edge e a "starting vertex" γ0(e), and an "end vertex" γ1(e). We assume that each edge e ∈ E corresponds to an interval [ae, be] ⊆ R (where ae, be ∈ R, ae < be), and we assume that µe is a finite Borel measure on [ae, be] satisfying either µe = 0 or else ae, be ∈ spt µe, µe({ae, be}) = 0. We denote E0 := {e ∈ E; µe = 0}, E1 := E \ E0. We further assume that, for each v ∈ V , we are given a weight µv > 0, and we define V0 :=(cid:8)v ∈ V ; µv = 0(cid:9), V1 := V \ V0. 2.1 Remark. The sets E1 and V1 encode the parts of the graph Γ, where a particle driven by the diffusion can be localised. In the present section we describe general glueing conditions which do not take into account the correspondence of the edges to the vertices. In the case E1 = E, V1 = ∅ and µe the Lebesgue measure on [ae, be], the model will describe quantum graphs; cf. [3], [4], [5]. In the case E1 = ∅ we obtain (weighted) discrete diffusion on the vertices; cf.[1]. Dirichlet forms on graphs 5 We are going to describe the self-adjoint operator driving the evolution in the Hilbert space where on we use the scalar product HΓ := HE ⊕ KV1 , HE := Me∈E1 L2([ae, be], µe) ((fe)e∈E1 (ge)e∈E1)HΓ := Xe∈E1 (fe ge)L2([ae,be],µe), and on KV1 we use the scalar product ((fv)v∈V1 (gv)v∈V1)HΓ := Xv∈V1 fvgv µv In the following, the mapping ι defined in Section 1 will be applied in the situation of the edges e ∈ E1, and will then be denoted by ιe. We then define the (for f =(cid:0)(fe)e∈E1, (fv)v∈V1(cid:1), g =(cid:0)(ge)e∈E1, (gv)v∈V1(cid:1) ∈ HΓ). operator ι from HΓ to Qe∈E1 2,µe(ae, be) × KV1, by W 1 D(ι) :=(cid:8)f ∈ HΓ; fe ∈ D(ιe) (e ∈ E1)(cid:9), (e ∈ E1), (ιf )e := ιefe (ιf )v := fv (v ∈ V1). We define the trace mapping (or boundary value mapping) tr : Qe∈E1 1 := E1 × {0, 1}, by 1∪V1, where E′ KV1 → KE ′ C[ae, be]× tr f (e, j) :=(cid:26) fe(ae) fe(be) if e ∈ E1, j = 0, if e ∈ E1, j = 1, tr f (v) := fv (v ∈ V1). The space KE ′ 1∪V1 will be provided with the scalar product (ξ η) := X(e,j)∈E ′ 1 ξ(e, j)η(e, j) + Xv∈V1 ξ(v)η(v) µv. For the definition of the form we assume that X is a subspace of KE ′ 1∪V1 and that L is a self-adjoint operator in X. Then we define the form τ by D(τ ) :=(cid:8)f ∈ D(ι); tr(ιf ) ∈ X(cid:9), τ (f, g) := Xe∈E1Z be ae (ιefe)′(x)(ιege)′(x) dx +(cid:0)L tr(ιf )(cid:12)(cid:12) tr(ιg)(cid:1). 6 C. Seifert, J. Voigt 2.2 Remark. The subspace X encodes boundary conditions for the elements of D(τ ). One would expect boundary conditions to be in the form of some equation for tr(ιf ). Of course, if P denotes the orthogonal projection from KE ′ 1∪V1 onto X ⊥, then D(τ ) =(cid:8)f ∈ D(ι); P tr(ιf ) = 0(cid:9). Further boundary conditions for the elements of the associated operator H are encoded in the operator L; we refer to the description of H in Theorem 3.1. 2.3 Lemma. The form τ defined above is symmetric. D(τ ) is dense if and only if prV1(X) = KV1 , (2.1) where prV1 denotes the canonical projection prV1 : KE ′ Proof. The symmetry of τ is obvious. 1∪V1 → KV1. Assume that D(τ ) is dense. The image of the dense set D(τ ) under the orthogonal projection pr2 : HΓ → KV1 is dense in KV1, and therefore is equal to KV1. From the definition of D(τ ) it follows that pr2(D(τ )) is contained in prV1(X), and therefore prV1(X) = KV1. Now assume that (2.1) holds. For v ∈ V1 let ξv ∈ X be such that ξv(v) = 1 and ξv(w) = 0 for all w ∈ V1 \ {v}. Let gv ∈ D(ι) be defined by tr(ιgv) = ξv, and gv affine linear on the edges. The affine linear interpolation of the prescribed boundary values evidently yields an element of gv ∈ D(τ ). Let f ∈ HΓ, and define f := f − Xv∈V1 fvgv. Then fv = 0 for all v ∈ V1. Because C 1 the function f can be approximated by functions in c (ae, be) is dense in L2([ae, be], µe) (e ∈ E1), Therefore f can be approximated by functions in c (ae, be) (e ∈ E1), fv = 0 (v ∈ V1)(cid:9). Dc :=(cid:8)f ∈ D(τ ); fe ∈ C 1 Dc +Xv∈V1 fvgv ⊆ D(τ ). 2.4 Remarks. (a) For the special case that X = KE ′ 1∪V1 and L = 0 we denote the corresponding form by τN (the index N indicating Neumann boundary con- ditions). The form τN decomposes as the sum of the Neumann forms on each of the edges and the null form on KV1. Therefore the closedness of τN follows from the closedness in the one-dimensional cases; cf. [2; Section 1 and Remark 3.2]. (b) Condition (2.1) did not occur in the previous treatment [2]. The reason is that it is obviously satisfied if the vertices do not have masses, i.e. V1 = ∅. Also, in the case of vertices with masses, but with local boundary conditions of continuity (see Example 5.3), condition (2.1) is automatically satisfied. Dirichlet forms on graphs 7 2.5 Theorem. The form τ defined above is bounded below and closed. Proof. For f ∈ D(τ ) we obtain the estimate τ (f ) = τN(f ) +(cid:0)L tr(ιf )(cid:12)(cid:12) tr(ιf )(cid:1) > τN(f ) − kLk tr(ιf )2 (with τN defined in Remark 2.4(a)). From inequality (1.2) we obtain that the mapping f 7→ tr(ιf ) E ′ is infinitesimally form small with respect to τN. The remaining part of the trace, f 7→ tr(ιf ) V1, is bounded. These observations imply that τ is bounded below and the that the embedding DτN ∋ f 7→ ιf ∈ C[ae, be] × KV1 , k · k∞) is continuous. (Here, DτN denotes D(τN), provided 1 (Qe∈E1 with the form norm.) In order to obtain that τ is closed it is sufficient to show that D(τ ) is a closed subset of DτN. This, however, is immediate from the continuity of the mapping DτN ∋ f 7→ tr(ιf ) ∈ KE ′ 1∪V1 (and the fact that X is a closed subspace of KE ′ 1∪V1). 3 The operator H associated with the form τ We assume that the notation and the hypotheses are as in the previous section, and that (2.1) holds. Besides the trace mapping defined in the previous section we also need the signed trace (or signed boundary values) BV (ae, be) → KE ′ 1 ⊆ KE ′ 1∪V1 str : Ye∈E1 (where BV (ae, be) denotes the set of functions of bounded variation, with equiv- alence of functions coinciding a.e.), defined by str f (e, j) :=(cid:26) fe(ae+) −fe(be−) if e ∈ E1, j = 0, if e ∈ E1, j = 1. The inclusion KE ′ to be able to use str f also as an element of KE ′ 1 ⊆ KE ′ 1∪V1. 1∪V1 is to be understood in the canonical sense; we want For the description of the self-adjoint operator H associated with the form τ we use a maximal operator H for the differential part of the form. With the notation described in Section 1, we define D( H) :=(cid:8)f ∈ Ye∈E1 Hf := (−∂µe(ιefe)′)e∈E1 D(ιe); (ιefe)′ ∈ L1(ae, be), ∂µe(ιefe)′ exists, ∂µe(ιefe)′ ∈ L2([ae, be], µe) (e ∈ E1)(cid:9), (f ∈ D( H)). Thus, for f ∈ D( H), the signed trace str((ιefe)′)e∈E1 exists, and it describes the "ingoing derivatives" from the endpoints of the intervals. It is to be understood 8 C. Seifert, J. Voigt that for (ιefe)′ we choose representatives of bounded variation (which exist by the explanation given at the end of Section 1), in order to be able to apply the signed trace mapping. Let X0 :=(cid:8)x ∈ X; prV1 x = 0(cid:9), which could also be expressed as X0 := X ∩ KE ′ as a subspace of KE ′ onto X0. Also, for v ∈ V1, let ξv ∈ X be such that ξv Lemma 2.3). 1∪V1), and let Q0 be the orthogonal projection from KE ′ 1 (with our understanding of KE ′ 1∪V1 1 V1 = 1{v} (see the proof of In the following, for f ∈ D(ι) we will use the shorthand notation (ιf )′ := . (cid:0)(ιefe)′(cid:1)e∈E1 3.1 Theorem. The operator H associated with the form τ is given by D(H) =(cid:8)f ∈ HΓ; (fe)e∈E1 ∈ D( H), tr(ιf ) ∈ X, ((Hf )e)e∈E1 = H(fe)e∈E1, (Hf )v = 1 µv(cid:0)L tr(ιf ) − str(ιf )′(cid:12)(cid:12) ξv(cid:1) Q0 str(ιf )′ = Q0L tr(ιf )(cid:9), (v ∈ V1). Proof. (i) A preliminary step: Let f ∈ D( H), g ∈ D(τ ). For all e ∈ E1 one has Z be ae (ιefe)′(x)(ιege)′(x) dx = −Z(ae,be) ∂µe(ιefe)′(x)ge(x) dµe(x) + (ιefe)′(be−)ιege(be) − (ιefe)′(ae+)ιege(ae); cf. [2; equ. (1.2)]. Summing this equation over all the edges in E1 we obtain Xe∈E1Z be ae (ιefe)′(x)(ιege)′(x) dx = ( Hf g)HE −(cid:0)str((ιefe)′)e∈E1(cid:12)(cid:12) tr (ιg)(cid:1)KE ′ 1 . (ii) Let f ∈ D(H), g ∈ D(τ ). From D(H) ⊆ D(τ ) we conclude that tr(ιf ) ∈ X. As in [2; proof of Theorem 1.9] one obtains that (fe)e∈E1 ∈ D( H), H(fe)e∈E1 = . Using part (i) above we obtain (cid:0)(Hf )e(cid:1)e∈E1 (Hf g)HΓ = −Xe∈E1Z(ae,be) = Xe∈E1Z be ae ∂µe(ιefe)′(x)ge(x) dµe(x) + Xv∈V1 (Hf )vgv µv (ιefe)′(x)(ιege)′(x) dx +(cid:0)str(ιf )′(cid:12)(cid:12) tr (ιg)(cid:1)KE ′ 1 + Xv∈V1 (Hf )vgv µv. 9 (3.1) Dirichlet forms on graphs Because of (Hf g)HΓ = Xe∈E1Z be ae we therefore obtain (ιefe)′(x)(ιege)′(x) dx +(cid:0)L tr(ιf )(cid:12)(cid:12) tr(ιg)(cid:1) Xv∈V1 (Hf )vgv µv =(cid:0)L tr(ιf ) − str(ιf )′(cid:12)(cid:12) tr(ιg)(cid:1). For ξ ∈ X0 choose g ∈ D(τ ) satisfying tr(ιg) = ξ, and g affine linear on the edges e ∈ E1. Then equation (3.1) implies This shows that Q0L tr(ιf ) = Q0 str(ιf )′. 0 =(cid:0)L tr(ιf ) − str(ιf )′(cid:12)(cid:12) ξ(cid:1). Let v ∈ V1, and choose g ∈ D(τ ) satisfying tr(ιg) = ξv, and g affine linear on the edges e ∈ E1. Then equation (3.1) yields (Hf )v µv = Xw∈V1 (Hf )wξv(w) µw =(cid:0)L tr(ιf ) − str(ιf )′(cid:12)(cid:12) ξv(cid:1). This shows the second part of the formula for Hf . (iii) Now let H denote the operator indicated on the right hand side of the assertion, and let f ∈ D( H). Then f ∈ D(τ ). Let g ∈ D(τ ). Then ξ := tr(ιg) −Pv∈V1 gvξv ∈ X0, and therefore (cid:0)L tr(ιf ) − str(ιf )′(cid:12)(cid:12) ξ(cid:1) =(cid:0)Q0(L tr(ιf ) − str(ιf )′)(cid:12)(cid:12) ξ(cid:1) = 0. Using part (i) as well as the previous equality we obtain ( Hf g)HΓ =(cid:0) H(fe)e∈E1(cid:12)(cid:12) (ge)e∈E1(cid:1)HE = Xe∈E1Z be ae 1 µv +Xv∈V1 (L tr(ιf ) − str(ιf )′ ξv)gv µv (ιefe)′(x)(ιege)′(x) dx +(cid:0)str(ιf )′(cid:12)(cid:12) tr (ιg)(cid:1) + (L tr(ιf ) − str(ιf )′ tr (ιg)) = τ (f, g). The definition of H then yields that f ∈ D(H) and Hf = Hf . 3.2 Remarks. (a) For f ∈ D(H) and v ∈ V1, the expression given for (Hf )v given in Theorem 3.1 does not depend on the choice of ξv. (b) The case of a weight µv > 0 at a vertex leads to a case of Wentzell boundary condition at v. The expression of (Hf )v in Theorem 3.1 generalises the expression obtained at a boundary point in the case of a single interval; cf. [8; Poposition 4.3]. 10 C. Seifert, J. Voigt 4 Positivity and contractivity In this section we indicate conditions for the C0-semigroup (e−tH )t>0 to be positive or submarkovian. We assume that the hypotheses are as in Section 2 and that (2.1) holds. In the following we need the notion of a (Stonean) sublattice of Kn. We consider Kn as the function space C({1, . . . , n}), and accordingly use the notation x = (x1, . . . , xn), for x ∈ Kn, and x ∧ y = (x1 ∧ y1, . . . , xn ∧ yn), for x, y ∈ Rn. A sublattice X of Kn is a subspace for which x ∈ X implies that x ∈ X. A sublattice X is called Stonean if additionally x ∧ 1 ∈ X for all real x ∈ X. We refer to [2; Appendix] for the description of (Stonean) sublattices of Kn and of generators for positive (submarkovian) C0-semigroups on these sublattices. 4.1 Theorem. (a) Assume that X is a sublattice of KE ′ 1∪V1 and that the semi- group (e−tL)t>0 is positivity preserving. Then (e−tH )t>0 is positivity preserving. (b) Assume that X is a Stonean sublattice of KE ′ 1∪V1 and that the semigroup (e−tL)t>0 is a submarkovian semigroup. Then (e−tH)t>0 is submarkovian. This result was proved for the case of local boundary conditions (cf. Section 5) and no vertex masses in [2; Theorem 3.5], and for the case of vertices with masses and local boundary conditions of continuity (cf. Example 5.3) in [2; Theorem 4.2]. Its proof is completely analogous to [2; proof of Theorem 3.5]; so we refrain from giving a complete proof but rather only mention the main ingredients. The proof consists in an application of the Beurling-Deny criteria (cf. [6; Corollary 2.18]; see also [2; Remarks 1.6]). So, in order to prove part (a), it is equivalent to prove that the normal contraction f 7→ f acts on D(τ ), and that τ (f ) 6 τ (f ) for all f ∈ D(τ ). That the inequality works on the differential part is a one- dimensional issue which is taken care of in [2; Theorem 1.7]. For the trace part, the main observation is the equation tr ιf = tr ιf . This is less obvious than it might appear at the first glance since, in general, one does not have ιf = ιf . However, this equality holds on spt µe, and therefore at the end points of the intervals [ae, be], for all e ∈ E1. The reasoning for part (b) is analogous. 5 Local boundary conditions So far, the structure of the graph did not enter the considerations; in fact the function γ linking the edges to the vertices was not used at all. In order to explain what we understand by local boundary conditions, we need the following definitions. For v ∈ V , the sets describe the sets of all edges having mass and starting or ending at v, respectively, and the set (j = 0, 1) E1,v,j :=(cid:8)e ∈ E1; γj(e) = v(cid:9) E1,v :=(cid:0)E1,v,0 × {0}(cid:1) ∪(cid:0)E1,v,1 × {1}(cid:1) Dirichlet forms on graphs 11 is the set of all edges having mass connected with v (and where loops starting and ending at v yield two contributions). Note that then E′ Recall that the boundary conditions are specified by the choice of a subspace 1∪V1 and a self-adjoint operator L in X. The boundary conditions will X ⊆ KE ′ be called local if for each v ∈ V there exists a subspace 1 =Sv∈V E1,v. Xv ⊆ KE1,v if v ∈ V0, Xv ⊆ KE1,v ∪{v} if v ∈ V1, and a selfadjoint operator Lv in Xv, such that Xv, X =Mv∈V Lv. L =Mv∈V For v ∈ V , we define the "local trace mapping" by trv : Ye∈E1 C[ae, be] × KV1 →(cid:26) KE1,v KE1,v∪{v} if v ∈ V0, if v ∈ V1 trv f :=( tr f E1,v tr f E1,v∪{v} if v ∈ V0, if v ∈ V1. Then for the form τ we obtain D(τ ) =(cid:8)f ∈ D(ι); trv(ιf ) ∈ Xv (v ∈ V )(cid:9), τ (f, g) = Xe∈E1Z be (ιefe)′(x)(ιege)′(x) dx +Xv∈V (cid:0)Lv trv(ιf )(cid:12)(cid:12) trv(ιg)(cid:1). ae Xv,0 :=(cid:26) Xv if v ∈ V0, {ξ ∈ Xv; ξ(v) = 0} if v ∈ V1, With the condition (2.1) for D(τ ) to be dense then decomposes into Xv,0 6= Xv (v ∈ V1), or expressed differently, for all v ∈ V1 there exists ξv ∈ Xv such that ξv(v) = 1. It is an easy task to translate the description of the associated operator H, given in Theorem 3.1, to the present case of local boundary conditions, as follows. 5.1 Corollary. The operator H associated with τ is given by D(H) =(cid:8)f ∈ HΓ; (fe)e∈E1 ∈ D( H), trv(ιf ) ∈ Xv, Qv,0 strv(ιf )′ = Qv,0Lv trv(ιf ) (v ∈ V )(cid:9), ((Hf )e)e∈E1 = H(fe)e∈E1, (Hf )v = 1 µv(cid:0)Lv trv(ιf ) − strv(ιf )′(cid:12)(cid:12) ξv(cid:1) (v ∈ V1). 12 C. Seifert, J. Voigt Here, for v ∈ V the mapping strv : Qe∈E1 BV (ae, be) → KE1,v is defined by strv f := (str f ) E1,v, and Qv,0 is the orthogonal projection onto Xv,0 in KE1,v , for v ∈ V0, or in KE1,v∪{v}, for v ∈ V1. We will not put down further details here. Similarly, the conditions for (e−tH )t>0 to be positive and submarkovian, Theorem 4.1, can be spelled out in terms of the spaces Xv and the operators Lv. The statements are then analogous to [2; Theorem 3.5], where the case that E = E1 and V = V0 is treated. 5.2 Corollary. (a) Assume that Xv is a sublattice of KE1,v (v ∈ V0) or KE1,v∪{v} (v ∈ V1) and that (e−tLv)t>0 positivity preserving, for all v ∈ V . Then (e−tH)t>0 is a positivitiy preserving C0-semigroup on HΓ. (b) Assume that Xv is a Stonean sublattice of KE1,v (v ∈ V0) or KE1,v∪{v} (v ∈ V1) and that (e−tLv )t>0 is a submarkovian C0-semigroup on Xv, for all v ∈ V . Then (e−tH )t>0 is a submarkovian C0-semigroup on HΓ. 5.3 Example (local boundary conditions of continuity). This special case of local boundary conditions was studied in [2; Section 4]. In our framework, this example reads as follows. Let Xv = lin{1}, Lv ∈ R, lv := (Lv1 1) for v ∈ V . Then Xv,0 = {0} for v ∈ V1 (which makes it clear that condition (2.1) is satisfied) and hence Qv,0 =(cid:26) (· 1)1 if v ∈ V0, if v ∈ V1. 0 Functions f ∈ D(τ ) are continuous on Γ, i.e., for v ∈ V1 we have f (v) = (trv f )(e, j) for all (e, j) ∈ E1,v, and for v ∈ V0 there exists av(f ) ∈ K such that av(f ) = (trv f )(e, j) for all (e, j) ∈ E1,v (note that we cannot write f (v) in this case since f is not defined on V0). The second part of the boundary conditions for f ∈ D(H) translates to Xe∈E1,v,0 f ′ e(ae+) − Xe∈E1,v,1 f ′ e(be−) = lvav(f ) (v ∈ V0); see also [2; Theorem 4.3]. In the setup considered in [4], these boundary condi- tions are called δ-type conditions; cf. [4; Section 3.2.1]. References [1] F. Chung and L. Lu: Complex graphs and networks. American Mathematical Society, 2006 [2] U. Kant, T. Klauss, J. Voigt and M. Weber: Dirichlet forms for singular one- dimensional operators and on graphs. J. Evol. Equ. 9, 637 -- 659 (2009). [3] V. Kostrykin, J. Potthoff and R. Schrader: Contraction semigroups on met- ric graphs. In: Analysis on Graphs and its Applications, P. Exner (ed.) et Dirichlet forms on graphs 13 al., Amer. Math. Soc., Providence, RI, Proc. Symp. Pure Math. 77, 423 -- 458 (2008). [4] P. Kuchment: Quantum graphs: I. Some basic structures. Waves Random Me- dia 14, 107 -- 128 (2007). [5] P. Kuchment: Quantum graphs: an introduction and a brief survey. In: Anal- ysis on Graphs and its Applications, P. Exner (ed.) et al., Amer. Math. Soc., Providence, RI, Proc. Symp. Pure Math. 77, 291 -- 314 (2008). [6] E. M. Ouhabaz: Analysis of heat equations on domains. Princeton University Press, Princeton, 2005. [7] C. Seifert: Behandlung singularer Diffusion mit Hilfe von Dirichletformen. Diploma Thesis, 2009. http://www-user.tu-chemnitz.de/seifch/Forschung/Diplomarbeit.pdf [8] H. Vogt and J. Voigt: Wentzell boundary conditions on the context of Dirich- let forms. Adv. Differ. Equ. 8, No. 7, 821-842 (2003) Christian Seifert Fakultat Mathematik Technische Universitat Chemnitz 09107 Chemnitz, Germany [email protected] Jurgen Voigt Fachrichtung Mathematik Technische Universitat Dresden 01062 Dresden, Germany [email protected]
1307.6741	1	1307	2013-07-25T13:43:14	On eigenfunction expansions of first-order symmetric systems and ordinary differential operators of an odd order	[ "math.FA" ]	We study general (not necessarily Hamiltonian) first-order symmetric systems $J y'-B(t)y=\D(t) f(t)$ on an interval $\cI=[a,b\rangle $ with the regular endpoint $a$. It is assumed that the deficiency indices $n_\pm(\Tmi)$ of the minimal relation $\Tmi$ satisfy $n_+(\Tmi)< n_-(\Tmi)$. We define $\l$-depending boundary conditions which are analogs of separated self-adjoint boundary conditions for Hamiltonian systems. With a boundary value problem involving such conditions we associate an exit space self-adjoint extension $\wt T$ of $\Tmi$ and the $m$-function $m(\cd)$, which is an analog of the Titchmarsh-Weyl coefficient for the Hamiltonian system. By using $m$-function we obtain the eigenfunction expansion with the spectral function $\Si(\cd)$ of the minimally possible dimension and characterize the case when spectrum of $\wt T$ is defined by $\Si(\cd)$. Moreover, we parametrize all spectral functions in terms of a Nevanlinna type boundary parameter. Application of these results to ordinary differential operators of an odd order enables us to complete the results by Everitt and Krishna Kumar on the Titchmarsh-Weyl theory of such operators.	math.FA	math	ON EIGENFUNCTION EXPANSIONS OF FIRST-ORDER SYMMETRIC SYSTEMS AND ORDINARY DIFFERENTIAL OPERATORS OF AN ODD ORDER VADIM MOGILEVSKII Abstract. We study general (not necessarily Hamiltonian) first-order symmetric sys- tems J y′ − B(t)y = ∆(t)f (t) on an interval I = [a, bi with the regular endpoint a. It is assumed that the deficiency indices n±(Tmin) of the minimal relation Tmin satisfy n+(Tmin) < n−(Tmin). We define λ-depending boundary conditions which are analogs of separated self-adjoint boundary conditions for Hamiltonian systems. With a boundary value problem involving such conditions we associate an exit space self-adjoint extension eT of Tmin and the m-function m(·), which is an analog of the Titchmarsh-Weyl coefficient for the Hamiltonian system. By using m-function we obtain the eigenfunction expansion with the spectral function Σ(·) of the minimally possible dimension and characterize the case when spectrum of eT is defined by Σ(·). Moreover, we parametrize all spectral func- tions in terms of a Nevanlinna type boundary parameter. Application of these results to ordinary differential operators of an odd order enables us to complete the results by Everitt and Krishna Kumar on the Titchmarsh-Weyl theory of such operators. 1. Introduction Let H and bH be finite dimensional Hilbert spaces and let (1.1) H0 := H ⊕ bH, H := H0 ⊕ H = H ⊕ bH ⊕ H. In the paper we study first-order symmetric systems of differential equations defined on an interval I = [a, bi, −∞ < a < b ≤ ∞, with the regular endpoint a and regular or singular endpoint b. Such a system is of the form [3, 12] (1.2) J y′ − B(t)y = ∆(t)f (t), t ∈ I, where B(t) = B∗(t) and ∆(t) ≥ 0 are the [H]-valued functions on I and System (1.2) is called a Hamiltonian system if bH = {0}. any λ ∈ C each solution y(·) of the equation Throughout the paper we assume that system (1.2) is definite. The latter means that for (1.4) J y′ − B(t)y = λ∆(t)y satisfying ∆(t)y(t) = 0 (a.e. on I) is trivial, i.e., y(t) = 0, t ∈ I. 2010 Mathematics Subject Classification. 34B08, 34B20, 34B40,34L10,47A06,47B25. Key words and phrases. First-order symmetric system, boundary parameter, m-function, spectral func- tion of a boundary problem, Fourier transform. 1 (1.3) J = 0 0 IH 0 −IH iI bH 0 0 0  : H ⊕ bH ⊕ H → H ⊕ bH ⊕ H. 2 VADIM MOGILEVSKII In what follows we denote by H := L2 ∆(I) the Hilbert space of H-valued Borel functions f (·) on I (in fact, equivalence classes) satisfying f 2 (∆(t)f (t), f (t))H dt < ∞. ∆ :=RI Studying of symmetric systems is basically motivated by the fact that system (1.4) is a more general objet than a formally self-adjoint differential equation of an arbitrary order with matrix coefficients. In fact, such equation is reduced to the system (1.4) of a special form with J given by (1.3); moreover, this system is Hamiltonian precisely in the case when the differential equation is of an even order (we will concern this questions below). As is known, the extension theory of symmetric linear relations gives a natural framework for investigation of the boundary value problems for symmetric systems (see [4, 7, 20, 28, 36] and references therein). According to [20, 28, 36] the system (1.2) generates the minimal linear relation Tmin and the maximal linear relation Tmax in H. It turns out that Tmin is a closed symmetric relation with not necessarily equal deficiency indices n±(Tmin) and min. Since system (1.2) is definite, n±(Tmin) can be defined as a number of L2 Tmax = T ∗ ∆- solutions of (1.4) for λ ∈ C±. A description of various classes of extensions of Tmin (self-adjoint, m-dissipative, etc.) in terms of boundary conditions is an important problem in the spectral theory of symmetric systems. For Hamiltonian system (1.2) self-adjoint separated boundary conditions were described in [16]. Moreover, the Titchmarsh -- Weyl coefficient MT W (λ) of the boundary value problem for Hamiltonian system with self-adjoint separated boundary conditions was defined by various methods in [16, 25, 23]. Using MT W (·) one obtains the Fourier transform with the spectral function Σ(·) of the minimally possible dimension NΣ = dim H (see [7, 18, 20]). At the same time according to [34] non-Hamiltonian system (1.2) does not admit self-adjoint separated boundary conditions. Moreover, the inequality n+(Tmin) 6= n−(Tmin), and hence absence of self-adjoint boundary conditions is a typical situation for such systems. Therefore the following problems are of certain interest: • To find (may be λ-depending) analogs of self-adjoint separated boundary conditions for general (not necessarily Hamiltonian) systems (1.2) and describe such type conditions; • To describe in terms of boundary conditions all spectral matrix functions that have the minimally possible dimension and investigate the corresponding Fourier transforms. In the paper [2] these problems were considered for symmetric systems (1.2) satisfying n−(Tmin) ≤ n+(Tmin). In the present paper we solve these problems in the opposite case n+(Tmin) < n−(Tmin). It turns out that this case requires a somewhat another approach in comparison with [2], although the ideas of both the papers are similar. Moreover, we show that in the case n+(Tmin) < n−(Tmin) there is a class of exit space self-adjoint extensions of Tmin with special spectral properties. We apply also the obtained results to ordinary differential operators of an odd order with matrix valued coefficients and arbitrary deficiency indices. Let νb+ and νb− be indices of inertia of the skew-Hermitian bilinear form [y, z]b defined on dom Tmax by [y, z]b = lim t↑b (J y(t), z(t)), y, z ∈ dom Tmax, It turns out that system (1.2) with n+(Tmin) < n−(Tmin) has different properties depending on the sign of νb+ − νb−. Within this section we present the results of the paper assuming a simpler case νb+ − νb− ≤ 0. A crucial role in our considerations is played by a symmetric linear relation T (⊃ Tmin) in H Let a function y ∈ dom Tmax be represented as y(t) = {y0(t),by(t), y1(t)}(∈ H ⊕ bH ⊕ H). ON EIGENFUNCTION EXPANSIONS 3 given by means of boundary conditions as follows: T = {{y, f } ∈ Tmax : y1(a) = 0, by(a) = 0, eΓ0by = Γ1by = 0} and Γ1by are singular boundary values of a function y ∈ dom Tmax at the endpoint b. Here eΓ0b : dom Tmax → eHb and Γ1b : dom Tmax → Hb are linear mappings with special properties, eHb and Hb(⊂ eHb) are auxiliary finite dimensional Hilbert spaces. In fact , eΓ0by Recall that a linear relation eT = eT ∗ in a wider Hilbert space eH ⊃ H satisfying T ⊂ eT is called an exit space self-adjoint extension of T . Moreover, a generalized resolvent R(·) and a spectral function F (·) of T are defined by t ∈ R, R(λ) = PH(eT − λ)−1 ↾ H, λ ∈ C \ R, and F (t) = PHE(t) ↾ H, exit space self-adjoint extensions exist for symmetric linear relations with arbitrary (possibly unequal) deficiency indices. where E(·) is the orthogonal spectral function (resolution of identity) of eT . As is known [1] eR−(bH ⊕ eHb, Hb) of holomorphic operator pairs τ = {D0(λ), D1(λ)}. Such a pair is formed To describe the set of all generalized resolvents of T we use the Nevanlinna type class D0(λ) = (bD0(λ), eD0b(λ)) : bH ⊕ eHb → (bH ⊕ eHb) and D1(λ)(∈ [Hb, bH ⊕ eHb]) with special properties (see [35]). We show that each generalized resolvent y = R(λ)f, f ∈ H, is given as the L2 ∆-solution of the following boundary-value problem with λ-depending boundary conditions: by defined on C− holomorphic operator functions (1.5) (1.6) (1.7) y1(a) = 0, λ ∈ C−. J y′ − B(t)y = λ∆(t)y + ∆(t)f (t), t ∈ I, ibD0(λ)by(a) + eD0b(λ)eΓ0by + D1(λ)Γ1by = 0, Here (bD0(λ), eD0b(λ)) =: D0(λ) and D1(λ) are components of a pair τ = {D0(λ), D1(λ)} ∈ eR−(bH ⊕ eHb, Hb) (see (1.5)), so that the second equality in (1.7) is a Nevanlinna type bound- ary condition involving boundary values of a function y at both endpoints a and b. Thus, in- vestigation of boundary value problems for the system (1.2) in the case n+(Tmin) < n−(Tmin) require use of boundary conditions of another class in comparison with the case n−(Tmin) ≤ n+(Tmin)(cf.[2]). One may consider a pair τ as a boundary parameter, since R(λ) runs over the set of all generalized resolvents of T when τ runs over the set of all holomorphic operator and F (t) = Fτ (t) for the generalized resolvents and spectral functions of T respectively. pairs of the class eR−(bH ⊕ eHb, Hb). To indicate this fact explicitly we write R(λ) = Rτ (λ) Moreover, we denote by eT = eT τ the exit space self-adjoint extension of T generating Rτ (·) Next assume that ϕ(·, λ) and ψ(·, λ) are [H0, H]-valued operator solutions of the equation and Fτ (·). (1.4) satisfying the initial conditions ϕ(a, λ) =(cid:18)IH0 0 (cid:19) (∈ [H0, H0 ⊕ H]), ψ(a, λ) =(cid:18)− i −PH (cid:19) (∈ [H0, H0 ⊕ H]). 2 P bH (here PH and P bH are the orthoprojectors in H0 onto H and bH respectively). We show that, for each Nevanlinna boundary parameter τ = {D0(λ), D1(λ)} of the form (1.5), there exists a unique operator function mτ (·) : C \ R → [H0] such that the operator solution vτ (t, λ) := ϕ(t, λ)mτ (λ) + ψ(t, λ) bf (s) =ZI ϕ∗(t, s) ∆(t) f (t) dt satisfies (1.8) 4 VADIM MOGILEVSKII (1.4) has the following property: for every h0 ∈ H0 the function y = vτ (t, λ)h0 of Eq. belongs to L2 ∆(I) and satisfies the boundary condition We call mτ (·) the m-function corresponding to the boundary value problem (1.6), (1.7); re- ally, mτ (·) is an analog of the Titchmarsh-Weyl coefficient MT W (λ) for Hamiltonian systems. It turns out that mτ (·) is a Nevanlinna operator function satisfying the inequality bD0(λ)(iby(a) − P bH h0) + eD0b(λ)eΓ0by + D1(λ)Γ1by = 0, λ ∈ C−. (Im λ)−1 · Im mτ (λ) ≥ZI v∗ τ (t, λ)∆(t)vτ (t, λ) dt, λ ∈ C \ R. Next we study eigenfunction expansions of the boundary value problems for symmetric systems. Namely, let τ be a boundary parameter and let Fτ (·) be the spectral function of T generated by the boundary value problem (1.6), (1.7). A distribution operator-valued function Στ (·) : R → [H0] is called a spectral function of this problem if, for each function f ∈ H with compact support, the Fourier transform ((Fτ (β) − Fτ (α))f, f )H =Z[α,β) (dΣτ (s)bf (s),bf (s)) for any compact interval [α, β) ⊂ R. We show that for each boundary parameter τ there exists a unique spectral function Στ (·) and it is recovered from the m-function mτ (·) by means of the Stieltjes inversion formula (1.9) Στ (s) = − lim δ→+0 lim ε→+0 1 πZ s−δ −δ Im mτ (σ − iε) dσ. Below (within this section) we assume for simplicity that T is a (not necessarily densely defined) operator, i.e., mul T = {0}. one may define the inverse Fourier transform in the explicit form (see (5.30)). Moreover, if ous extension to a contractive linear mapping V : H → L2(Στ ; H0) (for the strict definition of the Hilbert space L2(Στ ; H0) see e.g. [8, Ch.13.5]. As in [2] one proves that V is an It follows from (1.8) that the mapping V f = bf (the Fourier transform) admits a continu- isometry (that is, the Parseval equality bfL2(Στ ;H0) = f H holds for every f ∈ H) if and only if the exit space extension eT τ in eH ⊃ H is an operator, i.e., muleT τ = {0}. In this case V is an isometry, then there exists a unitary extension eV ∈ [eH, L2(Στ ; H0)] of V such that the operator eT τ and the multiplication operator Λ in L2(Στ ; H0) are unitarily equivalent by means of eV . Hence, the operators eT τ and Λ have the same spectral properties; for instance, multiplicity of the spectrum of eT τ does not exceed dim H0(= dim H + dim bH). D1(λ) = (0, D1(λ))⊤(∈ [Hb, bH ⊕ eHb]). (1.10) D0(λ) = diag (I bH , D0(λ))(∈ [bH ⊕ eHb]), mτ (λ) =(cid:18)mτ,1(λ) mτ,2(λ) We show that in this case the corresponding m-function mτ (·) is of the triangular form Now assume that the boundary parameter τ = {D0(λ), D1(λ)} is (cf. (1.5)) 2 I bH (cid:19) : H ⊕ bH → H ⊕ bH, λ ∈ C−, 0 − i ON EIGENFUNCTION EXPANSIONS 5 so that the spectral function Στ (·) has the block matrix representation (1.11) Στ (s) =(cid:18)Στ,1(s) Στ,2(s) Στ,3(s) 1 2π sI bH(cid:19) : H ⊕ bH → H ⊕ bH, s ∈ R. Here Στ,1(·) is an [H]-valued distribution function, which can be defined by means of the Stieltjes inversion formula for mτ,1(λ). It follows from (1.11), that in the case when a boundary parameter τ = {D0(λ), D1(λ)} is of the form (1.10) and the Fourier transform is spectral properties (for more details see Theorem 5.14): an isometry, the corresponding exit space self-adjoint extension eT τ of T has the following (S1) σac(eT τ ) = R, where σac(eT τ ) is the absolutely continuous spectrum of eT τ . (S2) σs(eT τ ) = Ss(Στ,1), where σs(eT τ ) is the singular spectrum of eT τ and Ss(Στ,1) is multiplicity of the singular spectrum of eT τ does not exceed dim H, which yields the same estimate for multiplicity of each eigenvalue λ0 of eT τ . Next, we show that all spectral functions Στ (·) can be parametrized immediately in terms of the boundary parameter τ . More precisely, we show that there exists an operator function a closed support of the measure generated by the singular component of Στ,1. Hence the (1.12) X−(λ) =(cid:18)m0(λ) Φ−(λ) Ψ−(λ) M−(λ)(cid:19) : H0 ⊕ (bH ⊕ eHb) → H0 ⊕ Hb, λ ∈ C−, such that for each boundary parameter τ = {D0(λ), D1(λ)} the corresponding m-function mτ (·) is given by (1.13) mτ (λ) = m0(λ) + Φ−(λ)(D0(λ) − D1(λ) M−(λ))−1D1(λ)Ψ−(λ), λ ∈ C−. Thus, formula (1.13) together with the Stieltjes inversion formula (1.9) defines a (unique) spectral function Στ (·) of the boundary value problem (1.6), (1.7). Note that entries of the matrix (1.12) are defined in terms of the boundary values of respective operator solutions of Eq.(1.4). We also describe boundary parameters τ for which the Fourier transform V is an isometry and characterize the case when V is an isometry for every boundary parameter τ (see Theorem 5.15). Note that a description of spectral functions for various classes of boundary problems in the form close to (1.13), (1.9) can be found in [11, 13, 15, 22, 21, 33]. Clearly, all the foregoing results can be reformulated (with obvious simplifications) for Hamiltonian systems (1.2). We suppose that the above results about exit space extensions eT τ of T may be useful in the case when eT τ is a self-adjoint relation in eH = L2 on the whole line R. More precisely, we assume that spectral properties of such eT τ may be characterized in terms of the objects (m-function, boundary parameter etc.) associated with the restriction of this system onto the semi-axis I = [0, ∞). These problems will be considered elsewhere. ∆(R) induced by the symmetric system If n+(Tmin) takes on the minimally possible value n+(Tmin) = dim H, then n+(Tmin) ≤ n−(Tmin) and the above results can be rather simplified. Namely, in this case T is a maximal symmetric relation and hence there exists a unique generalized resolvent of T . Therefore there is a unique m-function m(·) which has the triangular form (1.14) m(λ) =(cid:18)M (λ) N−(λ) − i 0 2 I bH(cid:19) : H ⊕ bH → H ⊕ bH, λ ∈ C−. 6 VADIM MOGILEVSKII (M ∗(λ), N ∗ −(λ))⊤, where M (λ) and N−(λ) are taken from (1.14). Moreover, the spectral function Σ(·) of the corresponding boundary value problem is of the form (1.11) and the Fourier transform V is an isometry. Therefore a (unique) exit space self-adjoint extension T0 of T has the spectral properties (S1) and (S2). der [17], where the concept of the "rectangular" Titchmarsh-Weyl coefficient MT W (λ)(∈ Note that systems (1.2) with bH 6= {0} and both minimal deficiency indices n+(Tmin) = dim H and n−(Tmin) = dim H + dim bH were studied in the paper by Hinton and Schnei- [H, H ⊕ bH]), λ ∈ C+, was introduced. One can easily show that in fact MT W (λ) = In the final part of the paper we consider the operators generated by a differential ex- pression l[y] of an odd order 2m + 1 with [H]-valued coefficients (H is a Hilbert space with r := dim H < ∞) defined on an interval I = [a, bi (see (6.1)). In the particular case of scalar coefficients such differential operators have been investigated in the papers by Everitt and Krishna Kumar [9, 10, 26], where the limiting process from the compact intervals [a, β] ⊂ I was used for construction of the Titchmarsh-Weyl matrix MT W (λ) = (mjk(λ))m+1 j,k=1. Note that the results of these papers can not be considered to be completed; in particular, an attempt to define self-adjoint boundary conditions in [10] gave rise to hardly verifiable as- sumptions even in the case of minimally possible equal deficiency indices n±(Lmin) = m + 1 of the minimal operator Lmin. Our approach is based on the known fact [24] that the equation l[y] = λy is equivalent to a special symmetric non-Hamiltonian system (1.4). This enables us to extend the ob- tained results concerning symmetric systems to the expression l[y] with arbitrary (possibly unequal) deficiency indices of Lmin. In particular, we describe self-adjoint and λ-depending boundary conditions for l[y], which are analogs of self-adjoint separated boundary conditions for differential expressions of an even order. This makes it possible to construct eigenfunc- tion expansion with spectral matrix function Σ(·) of the dimension (m + 1)r × (m + 1)r and to describe all Σ(·) immediately in terms of boundary conditions (for operators of an even order and separated boundary conditions such a description was obtained in [33]). In conclusion note that the above results are obtained with the aid of the method of boundary triplets and the corresponding Weyl functions in the extension theory of symmetric linear relations (see [6, 14, 29, 31, 32] and references therein). 2. Preliminaries 2.1. Notations. The following notations will be used throughout the paper: H, H denote Hilbert spaces; [H1, H2] is the set of all bounded linear operators defined on the Hilbert space H1 with values in the Hilbert space H2; [H] := [H, H]; A ↾ L is the restriction of an operator A onto the linear manifold L; PL is the orthogonal projector in H onto the subspace L ⊂ H; C+ (C−) is the upper (lower) half-plane of the complex plane. Recall that a closed linear relation from H0 to H1 is a closed linear subspace in H0 ⊕ H1. The set of all closed linear relations from H0 to H1 (in H) will be denoted by eC(H0, H1) (eC(H)). A closed linear operator T from H0 to H1 is identified with its graph gr T ∈ eC(H0, H1). For a linear relation T ∈ eC(H0, H1) we denote by dom T, ran T, ker T and mul T the and adjoint linear relations of T are the relations T −1 ∈ eC(H1, H0) and T ∗ ∈ eC(H1, H0) domain, range, kernel and the multivalued part of T respectively. Recall also that the inverse ON EIGENFUNCTION EXPANSIONS 7 defined by (2.1) T ∗ = {{k1, k0} ∈ H1 ⊕ H0 : (k0, h0) − (k1, h1) = 0, {h0, h1} ∈ T }. T −1 = {{h1, h0} ∈ H1 ⊕ H0 : {h0, h1} ∈ T } In the case T ∈ eC(H0, H1) we write 0 ∈ ρ(T ) if ker T = {0} and ran T = H1, or equivalently if T −1 ∈ [H1, H0]; 0 ∈ bρ(T ) if ker T = {0} and ran T is a closed subspace in H1. For a linear relation T ∈ eC(H) we denote by ρ(T ) := {λ ∈ C : 0 ∈ ρ(T − λ)} and bρ(T ) = {λ ∈ C : 0 ∈ bρ(T − λ)} the resolvent set and the set of regular type points of T Recall also the following definition. respectively. Definition 2.1. A holomorphic operator function Φ(·) : C \ R → [H] is called a Nevanlinna function if Im λ · ImΦ(λ) ≥ 0 and Φ∗(λ) = Φ(λ), λ ∈ C \ R. (2.2) H = H0 ⊕ mul A, the following decompositions hold A = gr A′ ⊕ dmul A, E′(t)PH0 , where E′(·) : R → [H0] is the orthogonal spectral function of A′ operator in H0 (the operator part of A); moreover, A = A∗ if and only if A′ = (A′)∗. 2.2. Symmetric and self-adjoint linear relations. A linear relation A ∈ eC(H) is called symmetric (self-adjoint) if A ⊂ A∗ (resp. A = A∗). For each symmetric relations A ∈ eC(H) where dmul A = {0} ⊕ mul A and A′ is a closed symmetric (not necessarily densely defined) A spectral function E(·) : R → [H] of the relation A = A∗ ∈ eC(H) is defined as E(t) = For an operator A = A∗ we denote by σ(A), σac(A) and σs(A) the spectrum, the abso- lutely continuous spectrum and the singular spectrum of A respectively [19, Section 10.1]. Next assume that H is a finite dimensional Hilbert space. A non-decreasing operator function Σ(·) : R → [H] is called a distribution function if it is left continuous and satisfies Σ(0) = 0. For each distribution function Σ(·) there is a unique pair of distribution functions Σac(·) and Σs(·) such that Σac(·) is absolutely continuous, Σs(·) is singular and Σ(t) = Σac(t) + Σs(t) (the Lebesgue decomposition of Σ). For a distribution function Σ(·) we denote by S(Σ) the set of all t ∈ R such that Σ(t − δ) 6= Σ(t + δ) for any δ > 0. Moreover, we let Sac(Σ) = S(Σac) and Ss(Σ) = S(Σs). With each distribution function Σ(·) one associates the Hilbert space L2(Σ; H) of all [8, (dΣ(t)f (t), f (t)) < ∞ (for more details see e.g. functions f (·) : R → H such that RR Section 13.5]). The following theorem is well known. Theorem 2.2. Let Σ(·) be a [H]-valued distribution function. Then the relations dom ΛΣ = {f ∈ L2(Σ; H) : tf (t) ∈ L2(Σ; H)}, (ΛΣf )(t) = tf (t), f ∈ dom ΛΣ define a self-adjoint operator Λ = ΛΣ in L2(Σ; H) (the multiplication operator) and σ(ΛΣ) = S(Σ), σac(ΛΣ) = Sac(Σ), σs(ΛΣ) = Ss(Σ). 2.3. The class eR−(H0, H1). Let H0 be a Hilbert space, let H1 be a subspace in H0 and let τ = {τ+, τ−} be a collection of holomorphic functions τ±(·) : C± → eC(H0, H1). In the paper we systematically deal with collections τ = {τ+, τ−} of the special class eR−(H0, H1) introduced in [35]. In the case H0 = H1 =: H this class turns into the well known class 8 VADIM MOGILEVSKII eR(H) of Nevanlinna functions τ (·) : C \ R → eC(H) (see for instance [5]). If dim H0 < ∞, then according to [35] the collection τ = {τ+, τ−} ∈ eR−(H0, H1) admits the representation τ+(λ) = {(C0(λ), C1(λ)); H1}, λ ∈ C+; τ−(λ) = {(D0(λ), D1(λ)); H0}, λ ∈ C− (2.3) by means of holomorphic operator pairs (C0(λ), C1(λ)) : H0 ⊕ H1 → H1, λ ∈ C+; (D0(λ), D1(λ)) : H0 ⊕ H1 → H0, λ ∈ C− (more precisely, by equivalence classes of such pairs). The equalities (2.3) mean that (2.4) τ+(λ) = {{h0, h1} ∈ H0 ⊕ H1 : C0(λ)h0 + C1(λ)h1 = 0}, λ ∈ C+ τ−(λ) = {{h0, h1} ∈ H0 ⊕ H1 : D0(λ)h0 + D1(λ)h1 = 0}, λ ∈ C−. In [35] the class eR−(H0, H1) is characterized both in terms of eC(H0, H1)-valued functions τ±(·) and in terms of operator functions Cj(·) and Dj(·), j ∈ {0, 1}, from (2.3). 2.4. Boundary triplets and Weyl functions. Here we recall some definitions and results from our paper [35]. Let A be a closed symmetric linear relation in the Hilbert space H, let Nλ(A) = ker (A∗ − that H0 = H1 ⊕ H2. Denote by Pj the orthoprojector in H0 onto Hj, j ∈ {1, 2}. Next assume that H0 is a Hilbert space, H1 is a subspace in H0 and H2 := H0 ⊖ H1, so n±(A) := dim Nλ(A) ≤ ∞, λ ∈ C±, be deficiency indices of A. Denote by ExtA the set of Definition 2.3. A collection Π− = {H0 ⊕ H1, Γ0, Γ1}, where Γj : A∗ → Hj, j ∈ {0, 1}, are λ) (λ ∈ bρ(A)) be a defect subspace of A, let bNλ(A) = {{f, λf } : f ∈ Nλ(A)} and let all proper extensions of A, i.e., the set of all relations eA ∈ eC(H) such that A ⊂ eA ⊂ A∗. linear mappings, is called a boundary triplet for A∗, if the mapping Γ : bf → {Γ0bf , Γ1bf},bf ∈ holds for all bf = {f, f ′}, bg = {g, g′} ∈ A∗. (f ′, g) − (f, g′) = (Γ1bf , Γ0bg)H0 − (Γ0bf , Γ1bg)H0 − i(P2Γ0bf , P2Γ0bg)H2 Proposition 2.4. Let Π− = {H0 ⊕ H1, Γ0, Γ1} be a boundary triplet for A∗. Then: A∗, from A∗ into H0 ⊕ H1 is surjective and the following Green's identity (1) dim H1 = n+(A) ≤ n−(A) = dim H0. (2) ker Γ0 ∩ ker Γ1 = A and Γj is a bounded operator from A∗ into Hj, j ∈ {0, 1}. (3) The equality (2.5) defines the maximal symmetric extension A0 ∈ ExtA such that C− ⊂ ρ(A0). A0 := ker Γ0 = {bf ∈ A∗ : Γ0bf = 0} (2.6) (2.8) (2.9) Proposition 2.5. Let Π− = {H0 ⊕ H1, Γ0, Γ1} be a boundary triplet for A∗ (so that in view of Proposition 2.4, (1) n+(A) ≤ n−(A)). Denote also by π1 the orthoprojector in H ⊕ H onto H ⊕ {0}. Then: (1) The operators P1Γ0 ↾ bNλ(A), λ ∈ C+, and Γ0 ↾ bNz(A), z ∈ C−, isomorphically map bNλ(A) onto H1 and bNz(A) onto H0 respectively. Therefore the equalities (2.7) γ+(λ) = π1(P1Γ0 ↾ bNλ(A))−1, λ ∈ C+, γ−(z) = π1(Γ0 ↾ bNz(A))−1, z ∈ C−, M+(λ)h1 = (Γ1 − iP2Γ0){γ+(λ)h1, λγ+(λ)h1}, h1 ∈ H1, λ ∈ C+ M−(z)h0 = Γ1{γ−(z)h0, zγ−(z)h0}, h0 ∈ H0, z ∈ C− ON EIGENFUNCTION EXPANSIONS 9 correctly define the operator functions γ+(·) : C+ → [H1, H], γ−(·) : C− → [H0, H] and M+(·) : C+ → [H1, H0], M−(·) : C− → [H0, H1], which are holomorphic on their domains. Moreover, the equality M ∗ +(λ) = M−(λ), λ ∈ C−, is valid. (2) Assume that M+(λ) = (M (λ), N+(λ))⊤ : H1 → H1 ⊕ H2, M−(z) = (M (z), N−(z)) : H1 ⊕ H2 → H1, λ ∈ C+ z ∈ C− are the block representations of M+(λ) and M−(z) respectively and let (2.10) (2.11) 0 i M(λ) =(cid:18) M (λ) M(λ) =(cid:18)M (λ) N−(λ) 2 IH2(cid:19) : H1 ⊕ H2 → H1 ⊕ H2, 2 IH2(cid:19) : H1 ⊕ H2 → H1 ⊕ H2, N+(λ) − i 0 λ ∈ C+ λ ∈ C−. Then M(·) is a Nevanlinna operator function satisfying the identity (2.12) M(µ) − M∗(λ) = (µ − λ)γ∗ −(λ)γ−(µ), µ, λ ∈ C−. Definition 2.6. the Weyl functions, respectively, corresponding to the boundary triplet Π−. [35] The operator functions γ±(·) and M±(·) are called the γ-fields and It follows from (2.7) that for each h1 ∈ H1 and h0 ∈ H0 the following equalities hold (2.13) P1Γ0{γ+(λ)h1, λγ+(λ)h1} = h1, Γ0{γ−(z)h0, zγ−(z)h0} = h0. Proposition 2.7. Let Π− = {H0 ⊕ H1, Γ0, Γ1} be a boundary triplet for A∗ and let γ±(·) and M±(·) be the corresponding γ-fields and Weyl functions respectively. Moreover, let the spaces H0 and H1 be decomposed as (so that H0 = H1 ⊕ H2) and let be the block representations of the operators Γ0 and Γ1. Then: (1) The equalities H1 = bH ⊕ H1, Γ0 = (bΓ0, Γ0)⊤ : A∗ → bH ⊕ H0, eA = {bf ∈ A∗ :bΓ0bf = Γ0bf = Γ1bf = 0}, H0 = bH ⊕ H0 Γ1 = (bΓ1, Γ1)⊤ : A∗ → bH ⊕ H1 eA∗ = {bf ∈ A∗ :bΓ0bf = 0} define a closed symmetric extension eA ∈ ExtA and its adjoint eA∗. (2) The collection Π− = { H0 ⊕ H1, Γ0 ↾ eA∗, Γ1 ↾ eA∗} is a boundary triplet for eA∗. (3) The γ-fields γ±(·) and the Weyl functions M±(·) corresponding to Π− are given by γ+(λ) = γ+(λ) ↾ H1, γ−(λ) = γ−(λ) ↾ H0, M+(λ) = P H0 M−(λ) = P H1 M+(λ) ↾ H1, M−(λ) ↾ H0, λ ∈ C+ λ ∈ C−. The proof of Proposition 2.7 is similar to that of Proposition 4.1 in [5]. Remark 2.8. If H0 = H1 := H, then the boundary triplet in the sense of Definition 2.3 turns into the boundary triplet Π = {H, Γ0, Γ1} for A∗ in the sense of [14, 29]. In this case n+(A) = n−(A)(= dim H) and the γ-fields γ±(·) and Weyl functions M±(·) turn into the γ-field γ(·) and Weyl function M (·) respectively introduced in [6, 29]. Observe also that along with Π− we define in [32, 35] a boundary triplet Π+ = {H0 ⊕ H1, Γ0, Γ1} for A∗. Such a triplet is applicable to symmetric relations A with n−(A) ≤ n+(A). 10 VADIM MOGILEVSKII 2.5. Generalized resolvents and exit space extensions. The following definitions are well known. (2.14) (2.15) 1, H2 2]. Definition 2.9. Let eH be a Hilbert space and let H be a subspace in eH. A relation eA = eA∗ ∈ eC(eH) is called H-minimal if span{H, (eA − λ)−1H : λ ∈ C \ R} =eH. Definition 2.10. The relations Tj ∈ eC(Hj), j ∈ {1, 2}, are said to be unitarily equivalent (by means of a unitary operator U ∈ [H1, H2]) if T2 = eU T1 with eU = U ⊕ U ∈ [H2 spectral function of A respectively if there exist a Hilbert space eH ⊃ H and a self-adjoint relation eA ∈ eC(eH) such that A ⊂ eA and the following equalities hold: Definition 2.11. Let A be a symmetric relation in a Hilbert space H. The operator func- tions R(·) : C \ R → [H] and F (·) : R → [H] are called the generalized resolvent and the R(λ) = PH(eA − λ)−1 ↾ H, (in formula (2.15) E(·) is the spectral function of eA). The relation eA in (2.14) is called an exit space extension of A. space extension eA of A. Moreover, if the H-minimal exit space extensions eA1 ∈ eC(eH1) and eA2 ∈ eC(eH2) of A induce the same generalized resolvent R(λ), then there exists a unitary operator V ′ ∈ [eH1 ⊖ H,eH2 ⊖ H] such that eA1 and eA2 are unitarily equivalent by means of eV = IH ⊕ V ′. By using these facts we suppose in the following that the exit space extension eA in (2.14) is H-minimal, so that eA is defined by R(·) uniquely up to the unitary equivalence. Note that in this case the equality mul eA = mul A holds for any exit space extension eA = eA∗ of A if and only if mul A = mul A∗ or, equivalently, if and only if the operator A′ (the operator part of A) is densely defined. According to [27] each generalized resolvent of A is generated by some H-minimal exit F (t) = PHE(t) ↾ H, λ ∈ C \ R t ∈ R 3. Boundary triplets for symmetric systems 3.1. Notations. Let I = [a, bi (−∞ < a < b ≤ ∞) be an interval of the real line (the symbol i means that the endpoint b < ∞ might be either included to I or not). For a given finite-dimensional Hilbert space H denote by AC(I; H) the set of functions f (·) : I → H which are absolutely continuous on each segment [a, β] ⊂ I and let AC(I) := AC(I; C). Next assume that ∆(·) is a [H]-valued Borel functions on I integrable on each com- ∆(I) the semi-Hilbert (∆(t)f (t), f (t))H dt < ∞ (see pact interval [a, β] ⊂ I and such that ∆(t) ≥ 0. Denote by L2 space of Borel functions f (·) : I → H satisfying f 2 ∆ := RI that the latter condition is equivalent toRI 1 ∆ 2 (t)F (t)2 dt < ∞. e.g. (f, g)∆ = RI [8, Chapter 13.5]). The semi-definite inner product (·, ·)∆ in L2 (∆(t)f (t), g(t))H dt, f, g ∈ L2 ∆(I). Moreover, let L2 ∆(I) is defined by ∆(I) be the Hilbert space of the equivalence classes in L2 quotient map from L2 ∆(I) onto L2 ∆(I). ∆(I) with respect to the semi-norm · ∆ and let π be the For a given finite-dimensional Hilbert space K denote by L2 operator-functions F (·) : I → [K, H] such that F (t)h ∈ L2 ∆[K, H] the set of all Borel ∆(I) for each h ∈ K. It is clear ON EIGENFUNCTION EXPANSIONS 11 3.2. Symmetric systems. In this subsection we provide some known results on symmetric systems of differential equations. Let H and bH be finite-dimensional Hilbert spaces and let (3.1) In the following we put (3.2) ν+ := dim H, H0 = H ⊕ bH, H = H0 ⊕ H = H ⊕ bH ⊕ H. bν := dim bH, ν− := dim H0 = ν+ +bν, n := dim H = ν+ + ν−. Let as above I = [a, bi (−∞ < a < b ≤ ∞) be an interval in R . Moreover, let B(·) and ∆(·) be [H]-valued Borel functions on I integrable on each compact interval [a, β] ⊂ I and satisfying B(t) = B∗(t) and ∆(t) ≥ 0 a.e. on I and let J ∈ [H] be operator (1.3). A first-order symmetric system on an interval I (with the regular endpoint a) is a system of differential equations of the form (3.3) J y′(t) − B(t)y(t) = ∆(t)f (t), t ∈ I, where f (·) ∈ L2 ∆(I). Together with (3.3) we consider also the homogeneous system (3.4) J y′(t) − B(t)y(t) = λ∆(t)y(t), t ∈ I, λ ∈ C. A function y ∈ AC(I; H) is a solution of (3.3) (resp. (3.4)) if equality (3.3) (resp. (3.4) holds a.e. on I. Moreover, a function Y (·, λ) : I → [K, H] is an operator solution of equation (3.4) if y(t) = Y (t, λ)h is a (vector) solution of this equation for every h ∈ K (here K is a Hilbert space with dim K < ∞). In what follows we always assume that system (3.3) is definite in the sense of the following definition. Definition 3.1. each solution y of (3.4) the equality ∆(t)y(t) = 0 (a.e. on I) implies y(t) = 0, t ∈ I. [12, 24] Symmetric system (3.3) is called definite if for each λ ∈ C and As it is known [36, 20, 28] symmetric system (3.3) gives rise to the maximal linear relations Tmax and Tmax in L2 ∆(I) and L2 Tmax = {{y, f } ∈ (L2 ∆(I), respectively. They are given by ∆(I))2 : y ∈ AC(I; H) and (3.5) J y′(t) − B(t)y(t) = ∆(t)f (t) a.e. on I} and Tmax = {{πy, πf } : {y, f } ∈ Tmax}. Moreover the Lagrange's identity (3.6) (f, z)∆ − (y, g)∆ = [y, z]b − (J y(a), z(a)), {y, f }, {z, g} ∈ Tmax. holds with (3.7) [y, z]b := lim t↑b (J y(t), z(t)), y, z ∈ dom Tmax. Formula (3.7) defines the skew-Hermitian bilinear form [·, ·]b on dom Tmax, which plays a crucial role in our considerations. By using this form we define the minimal relations Tmin in L2 ∆(I) and Tmin in L2 ∆(I) via Tmin = {{y, f } ∈ Tmax : y(a) = 0 and [y, z]b = 0 for each z ∈ dom Tmax}. and Tmin = {{πy, πf } : {y, f } ∈ Tmin}. According to [36, 20, 28] Tmin is a closed symmetric linear relation in L2 ∆(I) and T ∗ min = Tmax. 12 VADIM MOGILEVSKII [28]) that the maximal relation Tmax induced by the Remark 3.2. It is known (see e.g. definite symmetric system (3.3) possesses the following property: there exists a unique function y ∈ AC(I; H) ∩ L2 for any f ∈ ef . Below we associate such a function y ∈ AC(I; H) ∩ L2 {ey,ef} ∈ Tmax. ∆(I). Definition (3.5) of Tmax implies (3.4) belonging to L2 For any λ ∈ C denote by Nλ the linear space of solutions of the homogeneous system for any {ey,ef } ∈ Tmax ∆(I) such that y ∈ ey and {y, f } ∈ Tmax ∆(I) with each pair Nλ = ker (Tmax − λ) = {y ∈ L2 ∆(I) : {y, λy} ∈ Tmax}, λ ∈ C, and hence Nλ ⊂ dom Tmax. As usual, denote by n±(Tmin) := dim Nλ(Tmin), λ ∈ C±, the deficiency indices of Tmin. Since the system (3.3) is definite, πNλ = Nλ(Tmin) and ker (π ↾ Nλ) = {0}, λ ∈ C. This implies that dim Nλ = n±(Tmin), λ ∈ C±. The following lemma is obvious. Lemma 3.3. If Y (·, λ) ∈ L2 (3.8) ∆[K, H] is an operator solution of Eq. (3.4), then the relation K ∋ h → (Y (λ)h)(t) = Y (t, λ)h ∈ Nλ. defines the linear mapping Y (λ) : K → Nλ and, conversely, for each such a mapping Y (λ) there exists a unique operator solution Y (·, λ) ∈ L2 ∆[K, H] of Eq. (3.4) such that (3.8) holds. Next assume that (3.9) U =(cid:18)u1 u2 u3 u4 u5 u6(cid:19) : H ⊕ bH ⊕ H → bH ⊕ H is the operator satisfying the relations (3.10) (3.11) (3.12) iu2u∗ 2 − u1u∗ 3 + u3u∗ iu5u∗ 1 = 0 1 = iI bH , 5 + u6u∗ 3 + u6u∗ 2 − u4u∗ 6 = 0 iu5u∗ 4 − u4u∗ ran U = bH ⊕ H  : H ⊕ bH ⊕ H → H ⊕ bH ⊕ H, One can prove that the operator (3.9) admits an extension to the J-unitary operator (3.13) (3.14) u7 u8 u9 u1 u2 u3 u4 u5 u6 In view of (3.1) each function y ∈ AC(I; H) admits the representation Using (3.13) and the representation (3.14) of y we introduce the linear mappings Γja : i.e. the operator satisfying eU ∗JeU = J. eU = y(t) = {y0(t), by(t), y1(t)}(∈ H ⊕ bH ⊕ H), AC(I; H) → H, j ∈ {0, 1}, andbΓa : AC(I; H) → bH by setting Γ0ay = u7y0(a) + u8by(a) + u9y1(a), Clearly, the mappingsbΓa and Γ1a are determined by the operator U , while Γ0a is determined by the extension eU . Moreover, the mapping Γa :=(cid:16)Γ0a, bΓa, Γ1a(cid:17)⊤ bΓay = u1y0(a) + u2by(a) + u3y1(a), Γ1ay = u4y0(a) + u5by(a) + u6y1(a). : AC(I; H) → H ⊕ bH ⊕ H y ∈ AC(I; H) (3.17) (3.16) (3.15) t ∈ I. ON EIGENFUNCTION EXPANSIONS 13 (3.18) satisfies Γay = eU y(a), y ∈ AC(I; H). Hence Γa is surjective and (J y(a), z(a)) = −(Γ1ay, Γ0az) + (Γ0ay, Γ1az) + i(bΓay,bΓaz), ϕ(·, λ) = ϕU (·, λ)(∈ [H0, H]), λ ∈ C, of Eq. (3.4) with the initial data In what follows we associate with each operator U (see (3.9)) the operator solution y, z ∈ AC(I; H). (3.19) (3.20) (3.21) . H {z sin B 0 − cos B 0 I bH cos B 0 0 0 I bH cos B 0 u∗ iu∗ 6 3 −iu∗ u∗ 5 2 −u∗ 4 −iu∗ 1 ϕU (a, λ) = where B = B∗ ∈ [H]. For such U the solution ϕU (·, λ) is defined by the initial data U =(cid:18) 0 ϕU (a, λ) =  : H ⊕ bH → H ⊕ bH ⊕ H } {z }H0 One can easily verify that for each J-unitary extension eU of U one has eU ϕU (a, λ) =(cid:18)IH0 0 (cid:19) : H0 → H0 ⊕ H. The particular case of the operator U and its J-unitary extension eU is (cf. [17]) eU = sin B  , sin B(cid:19) , − cos B 0 : H ⊕ bH → H ⊕ bH ⊕ H. n−(Tmin) = ν− + νb− = ν+ +bν + νb− Lemma 3.4. For any pair of finite-dimensional Hilbert spaces Hb and bHb with dim Hb = min{νb+, νb−} and dim bHb = νb+ − νb− there exists a surjective linear mapping [y, z]b = i · sign (νb+ − νb−)(bΓby,bΓbz) − (Γ1by, Γ0bz) + (Γ0by, Γ1bz). 3.3. Decomposing boundary triplets. According to [4, 34] the skew-Hermitian bilinear form (3.7) has finite indices of inertia νb+ and νb− and Γb = (Γ0b,bΓb, Γ1b)⊤ : dom Tmax → Hb ⊕ bHb ⊕ Hb Note that Γby is in fact a singular boundary value of a function y ∈ dom Tmax in the sense (for ν± see (3.2)). Moreover, the following lemma is immediate from [34, Lemma 5.1]. such that for all y, z ∈ dom Tmax the following equality is valid (3.22) n+(Tmin) = ν+ + νb+, sin B 0 0 I bH (3.23) (3.24) of [8, Chapter 13.2] (for more details see Remark 3.5 in [2]). the following two alternative cases may hold: Assume that n+(Tmin) < n−(Tmin). Then by (3.22) and (3.2) bν > νb+ − νb− and hence Case 1 . bν > νb+ − νb− > 0. Case 2 . bν ≥ 0 ≥ νb+ − νb− andbν 6= νb+ − νb−(6= 0). In Case 1 one has dim bH(= bν) > νb+ − νb− > 0. Let bH1 be a subspace in bH with dim bH1 = νb+ − νb− and let Hb be a Hilbert space with dim Hb = νb−. Then by Lemma 3.4 Below we construct a boundary triplet for Tmax separately in each of these cases. there exists a surjective linear mapping (3.25) Γb = (Γ0b,bΓb, Γ1b)⊤ : dom Tmax → Hb ⊕ bH1 ⊕ Hb 14 VADIM MOGILEVSKII (3.26) such that (3.24) holds for all y, z ∈ dom Tmax. Let bH be decomposed as bH = bH1 ⊕ bH2 with bH2 := bH ⊖ bH1 and let 0 := H ⊕ bH1, so be the block representation of the mappingbΓa (see (3.16)). Moreover, let H ′ bΓa = (bΓa1, bΓa2)⊤ : AC(I; H) → bH1 ⊕ bH2 that in view of (3.1) H0 admits the representation In Case 1 we let (3.27) (3.28) (3.29) (3.30) 0 H0 = H ⊕ bH1 0 ⊕ bH2 ⊕bH2 = H ′ {z }H ′ 0 ⊕ bH2 ⊕ Hb, H1 = H ′ 2 (bΓa1 +bΓb)y, −Γ1by}(∈ H ′ Γ0{ey,ef} = {−Γ1ay + i(bΓa1 −bΓb)y, ibΓa2y, Γ0by}(∈ H ′ Γ1{ey,ef} = {Γ0ay + 1 H0 = H ′ 0 ⊕ Hb, 0 ⊕ Hb), 0 ⊕ bH2 ⊕ Hb), {ey,ef } ∈ Tmax. Now assume that Case 2 holds. Let Hb and bHb be Hilbert spaces with dim Hb = νb+ and dim bHb = νb− − νb+. Then by Lemma 3.4 there is a surjective linear mapping (3.23) satisfying (3.24). Let eHb := Hb ⊕ bHb (so that Hb ⊂ eHb) and let eΓ0b : dom Tmax → eHb be the linear mapping given by eΓ0b = Γ0b +bΓb. (3.31) In Case 2 we put (3.32) (3.33) (3.34) H1 = H ⊕ Hb H0 = H ⊕ bH ⊕ eHb, Γ0{ey,ef } = {−Γ1ay, ibΓay,eΓ0by}(∈ H ⊕ bH ⊕ eHb) {ey,ef} ∈ Tmax. Γ1{ey,ef} = {Γ0ay, −Γ1by}(∈ H ⊕ Hb), Note that in both Cases 1 and 2 H1 is a subspace in H0 and Γj is an operator from Tmax to Hj, j ∈ {0.1}. Moreover, H2(= H0 ⊖ H1) = bH2 in Case 1 and H2 = bH ⊕ bHb in Case 2. Proposition 3.5. Assume that eU is J-unitary operator (3.13) and Γ0a, Γ1a and bΓa are linear mappings (3.15), (3.16). Moreover, let Γb be surjective linear mapping given either by(3.25) (in Case 1) or (3.23) (in Case 2) and satisfying (3.24). Then a collection Π− = {H0 ⊕ H1, Γ0, Γ1} defined either by (3.28) -- (3.30) (in Case 1) or (3.32) -- (3.34) (in Case 2) is a boundary triplet for Tmax. Proof. The immediate calculation with taking (3.18) and (3.24) into account gives (Γ1{ey,ef}, Γ0{ez,eg}) − (Γ0{ey,ef }, Γ1{ez,eg}) − i(P2Γ0{ey,ef }, P2Γ0{ez,eg}) = = [y, z]b − (J y(a), z(a)), This and the Lagrange's identity (3.6) yield identity (2.5) for Γ0 and Γ1. Moreover, the mapping Γ = (Γ0, Γ1)⊤ is surjective, because so are Γa (see (3.17)) and Γb. (cid:3) {ey,ef}, {ez,eg} ∈ Tmax. Definition 3.6. The boundary triplet Π− = {H0 ⊕ H1, Γ0, Γ1} constructed in Proposition 3.5 is called a decomposing boundary triplet for Tmax. ON EIGENFUNCTION EXPANSIONS 15 Proposition 3.7. Let in Case 1 U be operator (3.9) and letbΓa and Γ1a be linear mappings (3.16). Moreover, let bH1 be a subspace in bH, let bH2 = bH ⊖bH1, letbΓaj, j ∈ {1, 2}, be defined by (3.26) and let Γb be surjective linear mapping (3.25) satisfying (3.24). Then: (1) The equalities (3.35) (3.36) T = {{ey,ef} ∈ Tmax : Γ1ay = 0, bΓa1y =bΓby, bΓa2y = 0, Γ0by = Γ1by = 0} T ∗ = {{ey,ef} ∈ Tmax : Γ1ay = 0, bΓa1y =bΓby} define a symmetric extension T of Tmin and its adjoint T ∗. Moreover, the deficiency indices (3.38) Γ1{ey,ef} = −Γ1by, (3.37) is a boundary triplet for T ∗ and the (maximal symmetric) relation A0(= ker Γ0) is of T are n+(T ) = νb− and n−(T ) =bν + 2νb− − νb+. (2) The collection Π− = { H0 ⊕ Hb, Γ0, Γ1} with H0 = bH2 ⊕ Hb and the operators Γ0{ey,ef} = {ibΓa2y, Γ0by}(∈ bH2 ⊕ Hb), {ey,ef } ∈ T ∗, A0 = {{ey,ef } ∈ Tmax : Γ1ay = 0, bΓa1y =bΓby, bΓa2y = 0, Γ0by = 0}. {H0 ⊕ H1, Γ0, Γ1} be the decomposing boundary triplet (3.28) -- (3.30) for Tmax. Applying to this triplet Proposition 2.7 one obtains the desired statements. (cid:3) Proof. Let eU be J-unitary extension (3.13) of U , let Γ0a be operator (3.15) and let Π− = Proposition 3.8. Let in Case 2 U be operator (3.9) and letbΓa and Γ1a be linear mappings (3.16). Moreover, let Γb be surjective linear mapping (3.23) satisfying (3.24), let eHb = Hb ⊕ bHb and let eΓ0b be given by (3.31). Then: (1) The equalities (3.39) (3.40) T = {{ey,ef } ∈ Tmax : Γ1ay = 0, bΓay = 0, eΓ0by = Γ1by = 0} T ∗ = {{ey,ef} ∈ Tmax : Γ1ay = 0} define a symmetric extension T of Tmin and its adjoint T ∗. Moreover, n+(T ) = νb+ and n−(T ) =bν + νb−. (2) The collection Π− = { H0 ⊕ Hb, Γ0, Γ1} with H0 = bH ⊕ eHb and the operators {ey,ef} ∈ T ∗ (3.41) is a boundary triplet for T ∗ and the (maximal symmetric) relation A0(= ker Γ0) is Γ0{ey,ef} = {ibΓay,eΓ0by}(∈ bH ⊕ eHb), Γ1{ey,ef} = −Γ1by, A0 = {{ey,ef} ∈ Tmax : Γ1ay = 0, bΓay = 0, eΓ0by = 0}. We omit the proof of this proposition, because it is similar to that of Proposition 3.7. (3.42) 4. L2 ∆-solutions of boundary value problems 4.1. Basic assumption. In what follows we suppose (unless otherwise stated) that system (3.3) satisfies n−(Tmin) < n+(Tmin) and the following assumptions are fulfilled: mappings (3.16). (A1) U is the operator (3.9) satisfying (3.10) - (3.12) and bΓa and Γ1a are the linear (A2) In Case 1 bH is decomposed as bH = bH1 ⊕ bH2,bΓaj : AC(I; H) → bHj, j ∈ {1, 2}, are the operators given by (3.26), Hb is a finite-dimensional Hilbert space and Γb is a surjective operator (3.25) satisfying (3.24). 16 VADIM MOGILEVSKII (3.31). (A3) In Case 2 Hb and bHb are finite-dimensional Hilbert spaces, Γb is a surjective operator (3.23) satisfying (3.24), eHb = Hb ⊕ bHb and eΓ0b : dom Tmax → eHb is the operator Definition 4.1. In Case 1 a boundary parameter is a collection τ = {τ+, τ−} ∈ eR−(bH2 ⊕ Hb, Hb). A truncated boundary parameter is a boundary parameter τ = {τ+, τ−} satisfying τ−(λ) ⊂ ({0} ⊕ Hb) ⊕ Hb, λ ∈ C−. 4.2. Case 1. According to Subsection 2.3 a boundary parameter τ = {τ+, τ−} admits the representation (4.1) τ+(λ) = {(C0(λ), C1(λ)); Hb}, λ ∈ C+; {τ+, τ−} admits the representation (4.1) with τ−(λ) = {(D0(λ), D1(λ)); bH2 ⊕ Hb}, λ ∈ C− with holomorphic operator functions C0(λ)(∈ [bH2 ⊕ Hb, Hb]), C1(λ)(∈ [Hb]) and D0(λ)(∈ [bH2 ⊕ Hb]), D1(λ)(∈ [Hb, bH2 ⊕ Hb]). Moreover, a truncated boundary parameter τ = D1(λ)(cid:19) : Hb → bH2 ⊕Hb. (4.2) D0(λ) =(cid:18)I bH2 Let τ = {τ+, τ−} be a boundary parameter (4.1) and let 0 0 D0(λ)(cid:19) : bH2 ⊕Hb → bH2 ⊕Hb, D1(λ) =(cid:18) 0 C0(λ) = (bC02(λ), C0b(λ)) : bH2 ⊕ Hb → Hb, D0(λ) = (bD02(λ), D0b(λ)) : bH2 ⊕ Hb → (bH2 ⊕ Hb) be the block representations of C0(λ) and D0(λ). For a given function f ∈ L2 the following boundary value problem: ∆(I) consider (4.3) (4.4) (4.5) (4.6) (4.7) J y′ − B(t)y = λ∆(t)y + ∆(t)f (t), t ∈ I, Γ1ay = 0, λ ∈ C \ R, bΓa1y =bΓby, ibC02(λ)bΓa2y + C0b(λ)Γ0by + C1(λ)Γ1by = 0, λ ∈ C+ ibD02(λ)bΓa2y + D0b(λ)Γ0by + D1(λ)Γ1by = 0, λ ∈ C−. A function y(·, ·) : I × (C \ R) → H is called a solution of this problem if for each λ ∈ C \ R the function y(·, λ) belongs to AC(I; H) ∩ L2 ∆(I) and satisfies the equation (4.4) a.e. on I (so that y ∈ dom Tmax) and the boundary conditions (4.5) -- (4.7). Theorem 4.2. Let in Case 1 T be a symmetric relation in L2 τ = {τ+, τ−} is a boundary parameter (4.1), then for every f ∈ L2 (4.4) - (4.7) has a unique solution y(t, λ) = yf (t, λ) and the equality ∆(I) defined by (3.35). If ∆(I) the boundary problem R(λ)ef = π(yf (·, λ)), ef ∈ L2 ∆(I), f ∈ ef , defines a generalized resolvent R(λ) =: Rτ (λ) of T . Conversely, for each generalized resol- vent R(λ) of T there exists a unique boundary parameter τ such that R(λ) = Rτ (λ). λ ∈ C \ R, Proof. Let Π− be the boundary triplet for T ∗ defined in Proposition 3.7. Applying to this triplet [35, Theorem 3.11] we obtain the required statements. (cid:3) Remark 4.3. Let τ0 = {τ+, τ−} be a boundary parameter (4.1) with (4.8) C0(λ) ≡ PHb, C1(λ) ≡ 0, D0(λ) ≡ I bH2⊕Hb , D1(λ) ≡ 0, ON EIGENFUNCTION EXPANSIONS 17 and let A0 be a symmetric relation (3.38). Then (4.9) Rτ0(λ) = (A∗ 0 − λ)−1, λ ∈ C+ and Rτ0 (λ) = (A0 − λ)−1, λ ∈ C−. Proposition 4.4. Let in Case 1 PH , P bH1 and P bH2 and bH2 respectively (see (3.27)) and let eP bH2 bH2 (resp. Hb). Then: (3.4) satisfying be the orthoprojectors in H0 onto H, bH1 (PHb ) be the orthoprojector in bH2 ⊕ Hb onto ∆[H0, H] of Eq. (1) For any λ ∈ C \ R there exists a unique operator solution v0(·, λ) ∈ L2 Γ1av0(λ) = −PH , , Γ0bv0(λ) = 0, λ ∈ C \ R (2) For any λ ∈ C+ (λ ∈ C−) there exists a unique operator solution u+(·, λ) ∈ L2 ∆[Hb, H] (resp. u−(·, λ) ∈ L2 , λ ∈ C− i(bΓa1 −bΓb)v0(λ) = P bH1 ibΓa2v0(λ) = P bH2 ∆[bH2 ⊕ Hb, H]) of Eq. (3.4) such that i(bΓa1 −bΓb)u±(λ) = 0, ibΓa2u−(λ) = eP bH2 Γ0bu+(λ) = IHb , λ ∈ C+ Γ0bu−(λ) = PHb , Γ1au±(λ) = 0, , λ ∈ C± λ ∈ C−. (4.10) (4.11) (4.12) (4.13) (4.14) (4.16) (4.18) (4.19) In formulas (4.10) -- (4.14) v0(λ) and u±(λ) are linear mappings from Lemma 3.3 corre- sponding to the solutions v0(·, λ) and u±(·, λ) respectively. Proof. Let eU be the J-unitary extension (3.13) of U , let Γ0a be the operator (3.15) and let Π− = {H0⊕H1, Γ0, Γ1} be the decomposing boundary triplet (3.28)-(3.30) for Tmax. Assume also that γ±(·) are the γ-fields of Π−. Since the quotient mapping π isomorphically maps Nλ onto Nλ(Tmin), it follows that for every λ ∈ C+ (λ ∈ C−) there exists an isomorphism Z+(λ) : H1 → Nλ (resp. Z−(λ) : H0 → Nλ) such that (4.15) γ+(λ) = πZ+(λ), λ ∈ C+; γ−(λ) = πZ−(λ), λ ∈ C−. Let Γ′ 0 and Γ′ 1 be the linear mappings given by Γ′ Γ′ 0 = −Γ1a + i(bΓa1 −bΓb) ibΓa2 1 =(cid:18)Γ0a + 1 2 (bΓa1 +bΓb) −Γ1b Γ0b  : dom Tmax → H ′ (cid:19) : dom Tmax → H ′ 0 ⊕ bH2 ⊕ Hb, 0 ⊕ Hb. Then by (3.29) and (3.30) one has Γj{πy, λπy} = Γ′ this equality with (4.15) and (2.13) gives jy, y ∈ Nλ, j ∈ {0, 1}. Combining of (4.17) PH ′ 0⊕HbΓ′ 0Z+(λ) = IH1 , λ ∈ C+; Γ′ 0Z−(λ) = IH0 , λ ∈ C−, which in view of (4.16) can be written as Γ0b (cid:18)−Γ1a + i(bΓa1 −bΓb)  −Γ1a + i(bΓa1 −bΓb) (cid:19) Z+(λ) =(cid:18)IH ′  Z−(λ) = ibΓa2 IH ′ 0 0 Γ0b 0 0 0 0 IHb(cid:19) , λ ∈ C+  , λ ∈ C− 0 0 IHb 0 I bH2 0 18 VADIM MOGILEVSKII It follows from (4.18) and (4.19) that (4.20) Γ1aZ+(λ) = (−PH , 0), Γ1aZ−(λ) = (−PH , 0, 0), bΓa2Z−(λ) = (0, −iI bH2 Next assume that 2 P bH1 , 0), Γ0bZ+(λ) = (0, IHb ) 1 2 (bΓa1 −bΓb)Z+(λ) = (− i 1 2 P bH1 , 0), Γ0bZ−(λ) = (0, 0, IHb ) 2 (bΓa1 −bΓb)Z−(λ) = (− i , 0, 0) are the block representations of Z±(λ) and let Z+(λ) = (r(λ), u+(λ)) : H ′ 0 ⊕ Hb → Nλ, λ ∈ C+ v0(λ) := (r(λ), 0) : H ′ v0(λ) := (r(λ), ω−(λ)) : H ′ Z−(λ) = (r(λ), ω−(λ), eu−(λ)) : H ′ 0 ⊕ bH2 ⊕ Hb → Nλ, λ ∈ C− 0 ⊕ bH2 → Nλ, λ ∈ C+ 0 ⊕ bH2 → Nλ, λ ∈ C− u−(λ) = (ω−(λ), eu−(λ)) : bH2 ⊕ Hb → Nλ, λ ∈ C−. ∆[H0, H], u+(·, λ) ∈ L2 (4.21) (4.22) (4.23) (4.24) (4.25) (4.26) (4.27) (4.28) (4.29) where (4.30) (4.31) (4.32) (4.33) Then in view of (3.27) the equalities (4.25) and (4.26) define the operator v0(λ) ∈ [H0, Nλ], λ ∈ C \ R. Moreover, (4.23) induces the operator u+(λ) ∈ [Hb, Nλ]. Introduce also the operator Now assume that v0(·, λ) ∈ L2 Hb, H] are the operator solutions of Eq. (3.4) corresponding to v0(λ), u+(λ) and u−(λ) in accordance with Lemma 3.3. Then combining of (4.23) -- (4.27) with (4.20) -- (4.22) yields the relations (4.10) -- (4.14) for v0(·, λ) and u±(·, λ). Finally, by using uniqueness of the solution of the boundary value problem (4.4) -- (4.7) (with f = 0) one proves uniqueness of v0(·, λ) and u±(·, λ) in the same way as in [2, Proposition 4.3]. (cid:3) ∆[Hb, H] and u−(·, λ) ∈ L2 ∆[bH2 ⊕ the solutions v0(·, λ) and u±(·, λ) we define the operator functions Let eU be a J-unitary extension (3.13) of U and let Γ0a be the mapping (3.15). By using Ψ+(λ) X+(λ) =(cid:18)m0(λ) Φ+(λ) X−(λ) =(cid:18)m0(λ) Φ−(λ) M+(λ)(cid:19) : H0 ⊕ Hb → H0 ⊕ (bH2 ⊕ Hb), M−(λ)(cid:19) : H0 ⊕ (bH2 ⊕ Hb) → H0 ⊕ Hb, Ψ−(λ) λ ∈ C+ λ ∈ C−, m0(λ) = (Γ0a +bΓa)v0(λ) + i Φ±(λ) = (Γ0a +bΓa)u±(λ), λ ∈ C± , 2 P bH , λ ∈ C \ R Ψ+(λ) = (bΓa2 − Γ1b)v0(λ) + iP bH2 Ψ−(λ) = −Γ1bv0(λ), M+(λ) = (bΓa2 − Γ1b)u+(λ), λ ∈ C+ M−(λ) = −Γ1bu−(λ), λ ∈ C− In the following proposition we specify a connection between the operator functions X±(·) and the Weyl functions M±(·) of the decomposing boundary triplet Π− for Tmax. ON EIGENFUNCTION EXPANSIONS 19 Proposition 4.5. Let Π− = {H0 ⊕ H1, Γ0, Γ1} be the decomposing boundary triplet (3.28) -- (3.30) for Tmax and let M1(λ) M2+(λ) N1+(λ) N2+(λ) M3+(λ) M4+(λ) : H ′ M+(λ) = M−(λ) =(cid:18) M1(λ) N1−(λ) M2−(λ) M3−(λ) N2−(λ) M4−(λ)(cid:19) : H ′ 0 ⊕ Hb → H ′ 0 ⊕ bH2 ⊕ Hb, λ ∈ C+ 0 ⊕ Hb, λ ∈ C− 0 ⊕ bH2 ⊕ Hb → H ′ be the block representations of the corresponding Weyl functions M±(·). Then the entries of the operator matrices (4.28) and (4.29) are holomorphic on their domains and have the following block matrix representations: (4.34) (4.35) (4.36) (4.37) (4.38) (4.39) (4.40) i 0 2 I bH2(cid:19) : H ′ 0 ⊕ bH2 {z }H0 2 I bH2(cid:19) : H ′ 0 ⊕ bH2 {z } H0 N1+(λ) m0(λ) =(cid:18) M1(λ) m0(λ) =(cid:18)M1(λ) N1−(λ) Φ+(λ) =(cid:18)M2+(λ) Ψ+(λ) =(cid:18)N1+(λ) 0(cid:19) , M3+(λ) − i iI 0 N2+(λ)(cid:19) , λ ∈ C+; Φ−(λ) =(cid:18)N1−(λ) M2−(λ) 0 (cid:19) , −iI λ ∈ C+; Ψ−(λ) = (M3−(λ), N2−(λ)), λ ∈ C− λ ∈ C− λ ∈ C+ λ ∈ C− → H ′ , H0 0 ⊕ bH2 {z } 0 ⊕ bH2 {z } H0 , → H ′ M+(λ) = (N2+(λ), M4+(λ))⊤, λ ∈ C+; M−(λ) = (N2−(λ), M4−(λ)), λ ∈ C−. Moreover, the following equalities hold (4.41) m∗ 0(λ) = m0(λ), Φ∗ +(λ) = Ψ−(λ), Ψ∗ +(λ) = Φ−(λ), M ∗ +(λ) = M−(λ), λ ∈ C−. Proof. Let Γ′ Then by (4.15) and (2.8), (2.9) one has 0 and Γ′ 1 be given by (4.16) and let Z±(·) be the same as in Proposition 4.4. (Γ′ 1 − iP bH2 Γ′ 0)Z+(λ) = M+(λ), λ ∈ C+; Γ′ 1Z−(λ) = M−(λ), λ ∈ C−, which in view of (4.34) and (4.35) can be represented as  (cid:18)Γ0a + 1 Γ0a + 1 −Γ1b 2 (bΓa1 +bΓb) bΓa2 2 (bΓa1 +bΓb) −Γ1b M1(λ) M2+(λ) N1+(λ) N2+(λ)  Z+(λ) = M3+(λ) M4+(λ) , λ ∈ C+ (cid:19) Z−(λ) =(cid:18) M1(λ) N1−(λ) M2−(λ) M3−(λ) N2−(λ) M4−(λ)(cid:19) , λ ∈ C−. 2 (bΓa1 +bΓb)Z+(λ) = P bH1 1 bΓa2Z+(λ) = (N1+(λ), N2+(λ)), Γ1bZ+(λ) = (−M3+(λ), −M4+(λ)) Γ0aZ−(λ) = PH (M1(λ), N1−(λ), M2−(λ)) (M1(λ), N1−(λ), M2−(λ)) 1 2 (bΓa1 +bΓb)Z−(λ) = P bH1 Γ1bZ−(λ) = (−M3−(λ), −N2−(λ), −M4−(λ)). Γ0aZ+(λ) = PH (M1(λ), M2+(λ)), (M1(λ), M2+(λ)) This implies that (4.42) (4.43) (4.44) (4.45) (4.46) 20 VADIM MOGILEVSKII Summing up the second equality in (4.20) with the equalities (4.42) and the first equality in (4.43) one obtains Similarly, summing up the second equality in (4.21), the first equality in (4.22) and the equalities (4.44) and (4.45) one gets , M2+(λ) + N2+(λ)) : H ′ 0 ⊕ Hb → H0. Combining (4.47) and (4.43) with the block representation (4.23) of Z+(λ) and taking defi- nition (4.25) of v0(λ) into account we obtain , N1−(λ) − iI bH2 , M2−(λ)). )PH ′ 0 , λ ∈ C+ λ ∈ C+ , bΓa2u+(λ) = N2+(λ), , Γ1bu+(λ) = −M4+(λ), λ ∈ C+ λ ∈ C+. (4.49) (4.48) 2 P bH1 2 P bH1 (4.47) (Γ0a +bΓa)Z+(λ) = (M1(λ) + N1+(λ) − i (Γ0a +bΓa)Z−(λ) = (M1(λ) − i (Γ0a +bΓa)v0(λ) = (M1(λ) + N1+(λ) − i (Γ0a +bΓa)u+(λ) = M2+(λ) + N2+(λ), bΓa2v0(λ) = N1+(λ)PH ′ (Γ0a +bΓa)v0(λ) = (M1(λ) − i (Γ0a +bΓa)u−(λ) = (N1−(λ) − iI bH2 Γ1bv0(λ) = −M3+(λ)PH ′ 2 P bH1 (4.50) (4.51) (4.52) (4.53) (4.54) (4.55) 2 P bH1 0 0 Combining of (4.30) with (4.49) and (4.53) yields Moreover, (4.48) and (4.46) together with (4.24), (4.26) and (4.27) give Γ1bv0(λ) = −(M3−(λ), N2−(λ)), Γ1bu−(λ) = −(N2−(λ), M4−(λ)), )PH ′ 0 + (N1−(λ) − iI bH2 )P bH2 , λ ∈ C− , M2−(λ)) : bH2 ⊕ Hb → H0, λ ∈ C− λ ∈ C−. m0(λ) = (M1(λ) + N1+(λ) − i 2 P bH1 )PH ′ 0 + i + i 2 (P bH1 m0(λ) = (M1(λ) − i − iP bH2 2 P bH1 − i 2 P bH1 + i ) = (M1(λ) + N1+(λ))PH ′ + i 0 + (N1−(λ) − iI bH2 + P bH2 ) = M1(λ)PH ′ 2 P bH = (M1(λ) + N1+(λ))PH ′ 0 + i , λ ∈ C+ 2 P bH = M1(λ)PH ′ 2 P bH2 0 + N1−(λ)P bH2 2 P bH2 )P bH2 − i + P bH2 )PH ′ 2 (P bH1 0 + N1−(λ)P bH2 , λ ∈ C−, − 0 − i 2 P bH1 + which is equivalent to the block representations (4.36) and (4.37) of m0(λ). Moreover, (4.31) together with (4.50) and (4.54) gives the block representations (4.38) of Φ+(λ) and Φ−(λ). Combining (4.32) and (4.33) with (4.51), (4.52) and (4.55) we arrive at the block M±(λ). Finally, holomorphy of the operator representations (4.39) of Ψ±(λ) and (4.40) of functions m0(·), Φ±(·), Ψ±(·) and M±(·) and the equalities (4.41) are implied by (4.34) -- (4.40) and the equality M ∗ (cid:3) +(λ) = M−(λ), λ ∈ C−. Theorem 4.6. Let in Case 1 τ = {τ+, τ−} be a boundary parameter (4.1), (4.3). Then: (1) For each λ ∈ C \ R there exists a unique operator solution vτ (·, λ) ∈ L2 ∆[H0, H] of Eq. (3.4) satisfying the boundary conditions (4.56) (4.57) (4.58) (4.59) bC02(λ)(ibΓa2vτ (λ) − P bH2 bD02(λ)(ibΓa2vτ (λ) − P bH2 Γ1avτ (λ) = −PH , λ ∈ C \ R i(bΓa1 −bΓb)vτ (λ) = P bH1 , λ ∈ C \ R ) + C0b(λ)Γ0bvτ (λ) + C1(λ)Γ1bvτ (λ) = 0, λ ∈ C+ ) + D0b(λ)Γ0bvτ (λ) + D1(λ)Γ1bvτ (λ) = 0, λ ∈ C− (here vτ (λ) is the linear map from Lemma 3.3 corresponding to the solution vτ (·, λ)). ON EIGENFUNCTION EXPANSIONS 21 (2) vτ (·, λ) is connected with the solutions v0(·, λ) and u±(·, λ) from Proposition 4.4 by (4.60) (4.61) −(λ) + M+(λ))−1Ψ+(λ), vτ (t, λ) = v0(t, λ) − u+(t, λ)(τ ∗ vτ (t, λ) = v0(t, λ) − u−(t, λ)(τ−(λ) + M−(λ))−1Ψ−(λ), λ ∈ C+, λ ∈ C−, where Ψ±(λ) and M±(λ) are the operator functions defined in (4.32) and (4.33). Proof. It follows from (4.40) and (4.35) that Therefore by Proposition 2.7, (3) M−(λ) is the Weyl function of the boundary triplet Π− for T ∗ and in view of [35, Theorem 3.11] one has 0 ∈ ρ(τ−(λ)+ M−(λ)), λ ∈ C−. Hence for each λ ∈ C \ R the equalities (4.60) and (4.61) correctly define the solution vτ (·, λ) ∈ L2 ∆[H0, H] of Eq. (3.4) and to prove the theorem it is sufficient to show that such vτ (·, λ) is a unique solution of (3.4) belonging to L2 M−(λ) = PHb M−(λ) ↾ bH2 ⊕ Hb, λ ∈ C−. ∆[H0, H] and satisfying (4.56) -- (4.59). Combining (4.60) and (4.61) with (4.10) and (4.12) one gets the equalities (4.56) and (4.57). To prove (4.58) and (4.59) we let (4.62) T+(λ) = (τ ∗ −(λ) + M+(λ))−1, λ ∈ C+, so that τ ∗ −(λ) and τ−(λ) can be written as T−(λ) = (τ−(λ) + M−(λ))−1, λ ∈ C−, −(λ)}. M+(λ)T+(λ))h, (−PHb + PHb M+(λ)T+(λ))h}}, This and (4.63) yield (4.63) (4.64) τ ∗ τ+(λ) = {{−h1 + iP2h0, −P1h0} : {h1, h0} ∈ τ ∗ τ−(λ) = {{T−(λ)h, (I − M−(λ)T−(λ))h} : h ∈ Hb}. −(λ) = {{T+(λ)h, (I − M+(λ)T+(λ))h} : h ∈ bH2 ⊕ Hb}, By using definition of the class eR−(H0, H1) in [35] one can easily show that (4.65) τ+(λ) = {{(−T+(λ) + ieP bH2 where eP bH2 and PHb are the same as in Proposition 4.4 and h runs over bH2 ⊕ Hb. bΓa2v0(λ) = eP bH2 bΓa2u+(λ) = eP bH2 (ibΓa2vτ (λ) − P bH2 Combining (4.60) with (4.66), (4.67), (4.13) and the last equality in (4.10) one gets M+(λ)T+(λ))Ψ+(λ), ) + Γ0bvτ (λ) = (−T+(λ) + ieP bH2 Γ1bvτ (λ) = (−PHb + PHb M+(λ), Γ1bu+(λ) = −PHb M+(λ), , Γ1bv0(λ) = −PHbΨ+(λ), λ ∈ C+ M+(λ)T+(λ))Ψ+(λ), It follows from (4.32) that − ieP bH2 − ieP bH2 (4.66) (4.67) (4.68) λ ∈ C+. λ ∈ C+. Ψ+(λ) − iP bH2 Moreover, (4.61) together with (4.11), (4.14), (4.33) and the last equality in (4.10) yields (ibΓa2vτ (λ) − P bH2 (4.69) ) + Γ0bvτ (λ) = −eP bH2 Hence by (4.64) and (4.65) one has T−(λ)Ψ−(λ) − PHb T−(λ)Ψ−(λ) = −T−(λ)Ψ−(λ), Γ1bvτ (λ) = −(I − M−(λ)T−(λ))Ψ−(λ), λ ∈ C−. {(ibΓa2vτ (λ) − P bH2 )h0 + Γ0bvτ (λ)h0, Γ1bvτ (λ)h0} ∈ τ±(λ), h0 ∈ H0, λ ∈ C±, which in view of (2.4) and the block representations (4.3) yields (4.58) and (4.59). Finally, uniqueness of vτ (·, λ) is implied by uniqueness of the solution of the boundary value problem (4.4) -- (4.7) (see Theorem 4.2). (cid:3) 22 VADIM MOGILEVSKII 4.3. Case 2. Definition 4.7. A boundary parameter in Case 2 is a collection τ = {τ+, τ−} ∈ eR−(bH ⊕ eHb, Hb). A truncated boundary parameter is a boundary parameter τ = {τ+, τ−} satisfying τ−(λ) ⊂ ({0} ⊕ eHb) ⊕ Hb, λ ∈ C−. According to Subsection 2.3 a boundary parameter τ = {τ+, τ−} admits the representation (4.70) admits the representation (4.70) with τ+(λ) = {(C0(λ), C1(λ)); Hb}, λ ∈ C+; τ−(λ) = {(D0(λ), D1(λ)); bH ⊕ eHb}, λ ∈ C− with holomorphic operator functions C0(λ)(∈ [bH ⊕ eHb, Hb]), C1(λ)(∈ [Hb]) and D0(λ)(∈ [bH ⊕eHb]), D1(λ)(∈ [Hb, bH ⊕eHb]). Moreover, a truncated boundary parameter τ = {τ+, τ−} D1(λ)(cid:19) : Hb → bH ⊕ eHb. (4.71) D0(λ) =(cid:18)I bH Let τ = {τ+, τ−} be a boundary parameter (4.70) and let 0 D0(λ)(cid:19) : bH ⊕ eHb → bH ⊕ eHb, D1(λ) =(cid:18) 0 C0(λ) = (bC0(λ), eC0b(λ)) : bH ⊕ eHb → Hb, D0(λ) = (bD0(λ), eD0b(λ)) : bH ⊕ eHb → (bH ⊕ eHb) be the block representations of C0(λ) and D0(λ). For a given function f ∈ L2 the following boundary value problem: ∆(I) consider (4.72) 0 J y′ − B(t)y = λ∆(t)y + ∆(t)f (t), t ∈ I, Γ1ay = 0, λ ∈ C \ R, (4.73) (4.74) (4.75) (4.76) ibC0(λ)bΓay + eC0b(λ)eΓ0by + C1(λ)Γ1by = 0, ibD0(λ)bΓay + eD0b(λ)eΓ0by + D1(λ)Γ1by = 0, λ ∈ C−. λ ∈ C+, One proves the following theorem in the same way as Theorem 4.2. Theorem 4.8. Let in Case 2 T be a symmetric relation in L2 ∆(I) defined by (3.39). Then statements of Theorem 4.2 hold with a boundary parameter τ = {τ+, τ−} of the form (4.70) and the boundary value problem (4.73) - (4.76) in place of (4.4) - (4.7). Remark 4.9. Let τ0 = {τ+, τ−} be a boundary parameter (4.70) with (4.77) C0(λ) ≡ PHb (∈ [bH ⊕ eHb, Hb]), C1(λ) ≡ 0, D0(λ) ≡ I bH⊕ eHb and let A0 be the symmetric relation (3.42). Then Rτ0(λ) is of the form (4.9). , D1(λ) ≡ 0 Proposition 4.10. Let in Case 2 PH (resp. P bH ) be the orthoprojector in H0 onto H (resp. bH) and let eP bH (resp. P eHb (3.4) satisfying (1) For any λ ∈ C \ R there exists a unique operator solution v0(·, λ) ∈ L2 ) be the orthoprojector in bH ⊕ eHb onto bH (resp. eHb). Then: ∆[H0, H] of Eq. (4.78) (4.79) Γ1av0(λ) = −PH, λ ∈ C \ R Γ0bv0(λ) = 0, λ ∈ C+; ibΓav0(λ) = P bH , eΓ0bv0(λ) = 0, λ ∈ C− ON EIGENFUNCTION EXPANSIONS 23 (2) For any λ ∈ C+ (λ ∈ C−) there exists a unique operator solution u+(·, λ) ∈ L2 ∆[Hb, H] (resp. u−(·, λ) ∈ L2 ∆[bH ⊕ eHb, H]) of Eq. (3.4) satisfying ibΓau−(λ) = eP bH , Γ1au±(λ) = 0, λ ∈ C± Γ0bu+(λ) = IHb , λ ∈ C+; Proof. let Π− be the decomposing boundary triplet for Tmax constructed with the aid of some J-unitary extension eU of U (see Proposition 3.5). By using γ-fields of this triplet one proves in the same way as in Proposition 4.4 the existence of isomorphisms Z+(λ) : H1 → Nλ, λ ∈ C+, and Z−(λ) : H0 → Nλ, λ ∈ C−, such that eΓ0bu−(λ) = P eHb , λ ∈ C−. −Γ1a 0 Γ0b (cid:19) Z+(λ) =(cid:18)IH (cid:18)−Γ1a  Z−(λ) = ibΓa eΓ0b IH 0 0  0 IHb(cid:19) , 0 I bH 0 0 0 I eHb λ ∈ C+  , λ ∈ C− Γ1aZ+(λ) = (−IH , 0), Γ0bZ+(λ) = (0, IHb) It follows from (4.82) and (4.83) that (4.80) (4.81) (4.82) (4.83) (4.84) (4.85) (4.88) (4.89) (4.90) Assume that (4.86) (4.87) Γ1aZ−(λ) = (−IH , 0, 0), ibΓaZ−(λ) = (0, I bH , 0), eΓ0bZ−(λ) = (0, 0, I eHb ). Z+(λ) = (r(λ), u+(λ)) : H ⊕ Hb → Nλ, λ ∈ C+ Z−(λ) = (r(λ), ω−(λ), eu−(λ)) : H ⊕ bH ⊕ eHb → Nλ, λ ∈ C− are the block representations of Z±(λ) (see (3.32) ) and let v0(λ) : H0 → Nλ, λ ∈ C \ R, be the operator given by It is clear that (4.86) induces the operator u+(λ) ∈ [Hb, Nλ]. Introduce also the operator v0(λ) := (r(λ), 0) : H ⊕ bH → Nλ, λ ∈ C+ v0(λ) := (r(λ), ω−(λ)) : H ⊕ bH → Nλ, λ ∈ C−. u−(λ) = (ω−(λ), eu−(λ)) : bH ⊕ eHb → Nλ, λ ∈ C−. ∆[Hb, H] and u−(·, λ) ∈ L2 ∆[H0, H], u+(·, λ) ∈ L2 Let v0(·, λ) ∈ L2 solutions of Eq. (3.4) corresponding to v0(λ), u+(λ) and u−(λ) respectively (see Lemma 3.3). Then combining of (4.86) -- (4.90) with (4.84) and (4.85) gives the relations (4.78) -- (4.81). Finally, uniqueness of v0(·, λ) and u±(·, λ) follows from uniqueness of the solution of the boundary value problem (4.73) -- (4.76). (cid:3) ∆[bH ⊕eHb, H] be the operator the operator functions Let eU be a J-unitary extension (3.13) of U and let Γ0a be the mapping (3.15). Introduce Ψ+(λ) X+(λ) =(cid:18)m0(λ) Φ+(λ) X−(λ) =(cid:18)m0(λ) Φ−(λ) M+(λ)(cid:19) : H0 ⊕ Hb → H0 ⊕ (bH ⊕ eHb), M−(λ)(cid:19) : H0 ⊕ (bH ⊕ eHb) → H0 ⊕ Hb, Ψ−(λ) λ ∈ C+ λ ∈ C−, (4.91) (4.92) 24 VADIM MOGILEVSKII with the entries defined in terms of the solutions v0(·, λ) and u±(·, λ) as follows: m0(λ), Φ±(λ), Ψ−(λ) and M−(λ) are given by (4.30), (4.31) and (4.33), while (4.93) Ψ+(λ) = (bΓa − Γ1b − ibΓb)v0(λ) + iP bH , M+(λ) = (bΓa − Γ1b − ibΓb)u+(λ), λ ∈ C+. Proposition 4.11. Let Π− = {H0 ⊕H1, Γ0, Γ1} be the decomposing boundary triplet (3.32) -- (3.34) for Tmax and let M1(λ) M2+(λ) N1+(λ) N2+(λ) M+(λ) = M−(λ) =(cid:18) M1(λ) N1−(λ) M2−(λ) M3+(λ) M4+(λ) : H ⊕ Hb → H ⊕ bH ⊕ eHb, λ ∈ C+ M3−(λ) N2−(λ) M4−(λ)(cid:19) : H ⊕ bH ⊕ eHb → H ⊕ Hb, λ ∈ C− be the block matrix representations of the corresponding Weyl functions. Then the entries of the operator matrices (4.91) and (4.92) are holomorphic on their domains and satisfy (4.94) (4.95) (4.96) (4.97) N1+(λ) m0(λ) =(cid:18) M1(λ) m0(λ) =(cid:18)M1(λ) N1−(λ) − i 0 i 0 2 I bH(cid:19) : H ⊕ bH {z }H0 2 I bH (cid:19) : H ⊕ bH {z }H0 , → H ⊕ bH {z }H0 → H ⊕ bH {z }H0 λ ∈ C+ , λ ∈ C− and (4.38) -- (4.40). Moreover, the equalities (4.41) are valid. Proof. Let Z±(·) be the same as in Proposition 4.10. By using the reasonings similar to those in the proof of Proposition 4.5 one proves the equalities (4.98) (4.99) Γ0a   Z+(λ) = bΓa −Γ1b − ibΓb −Γ1b(cid:19) Z−(λ) =(cid:18) M1(λ) N1−(λ) M2−(λ) M3−(λ) N2−(λ) M4−(λ)(cid:19) , (cid:18) Γ0a M3+(λ) M4+(λ) , λ ∈ C+, M1(λ) M2+(λ) N1+(λ) N2+(λ) λ ∈ C−. It follows from (4.98) that Moreover, (4.99) and the second equality in (4.85) yield (Γ0a +bΓa)Z+(λ) = (M1(λ) + N1+(λ), M2+(λ) + N2+(λ)), (bΓa − Γ1b − ibΓb)Z+(λ) = (N1+(λ) + M3+(λ), N2+(λ) + M4+(λ)), (Γ0a +bΓa)Z−(λ) = (M1(λ), N1−(λ) − iI bH , M2−(λ)), −Γ1bZ−(λ) = (M3−(λ), N2−(λ), M4−(λ)), λ ∈ C−. λ ∈ C+. ON EIGENFUNCTION EXPANSIONS 25 Combining these equalities with the block representations (4.86) and (4.87) of Z±(λ) and taking (4.88) -- (4.90) into account we obtain (Γ0a +bΓa)v0(λ) = (M1(λ) + N1+(λ))PH , (Γ0a +bΓa)u+(λ) = M2+(λ) + N2+(λ), λ ∈ C+ (bΓa − Γ1b − ibΓb)v0(λ) = (N1+(λ) + M3+(λ))PH , λ ∈ C+ (bΓa − Γ1b − ibΓb)u+(λ) = N2+(λ) + M4+(λ), λ ∈ C+ (Γ0a +bΓa)v0(λ) = M1(λ)PH + N1−(λ)P bH − iP bH , (Γ0a +bΓa)u−(λ) = (N1−(λ) − iI bH , M2−(λ)) : bH ⊕ eHb → H0, −Γ1bu−(λ) = (N2−(λ), M4−(λ)), λ ∈ C− −Γ1bv0(λ) = (M3−(λ), N2−(λ)), This and definitions (4.30), (4.31), (4.33) and (4.93) of m0(·), Φ±(·), Ψ±(·) and M±(·) yield the equalities (4.96), (4.97) and (4.38) -- (4.40). Finally one proves (4.41) in the same way as in Proposition 4.5. (cid:3) λ ∈ C− λ ∈ C− Theorem 4.12. Let in Case 2 τ = {τ+, τ−} be a boundary parameter defined by (4.70) and (4.72). Then: (1) Statement (1) of Theorem 4.6 holds with the boundary condition (4.56) and the fol- lowing boundary conditions in place of (4.57) -- (4.59): (4.100) (4.101) bC0(λ)(ibΓavτ (λ) − P bH ) + eC0b(λ)eΓ0bvτ (λ) + C1(λ)Γ1bvτ (λ) = 0, bD0(λ)(ibΓavτ (λ) − P bH ) + eD0b(λ)eΓ0bvτ (λ) + D1(λ)Γ1bvτ (λ) = 0, (2) The solution vτ (·, λ) is of the form (4.60), (4.61), where v0(·, λ) and u±(·, λ) are defined in Proposition 4.10 and Ψ±(λ) and M±(λ) are given by (4.33) and (4.93). λ ∈ C+ λ ∈ C− Proof. As in Theorem 4.6 one proves that (4.60) and (4.61) correctly define the solution vτ (·, λ) ∈ L2 ∆[H0, H] of Eq. (3.4). Therefore it remains to show that such vτ (·, λ) is a unique solution of (3.4) belonging to L2 ∆[H0, H] and satisfying (4.56), (4.100) and (4.101). First observe that (4.78) and (4.80) yield (4.56). Next assume that T±(λ) are defined by (4.62). Then in the same way as in Theorem 4.6 one gets (4.64) and the equality (4.102) τ+(λ) = {{(−T+(λ) + iP bH⊕ bHb − iP bH⊕ bHb M+(λ)T+(λ))h, (−PHb + PHb M+(λ)T+(λ))h} : h ∈ bH ⊕ eHb}, (4.103) (4.104) where P bH⊕ bHb It follows from (4.93) that is the orthoprojector in bH ⊕ eHb onto bH ⊕ bHb(= (bH ⊕ eHb) ⊖ Hb). bΓav0(λ) = eP bH Ψ+(λ) − iP bH , Γ1bv0(λ) = −PHb Ψ+(λ), λ ∈ C+ bΓbu+(λ) = iP bHb bΓbv0(λ) = iP bHb Ψ+(λ), λ ∈ C+ M+(λ), M+(λ), Γ1bu+(λ) = −PHb (4.105) bΓau+(λ) = eP bH where eP bH , PHb and P bHb Moreover, (4.104), the last equality in (4.105) and the first equalities in (4.79) and (4.81) give are the orthoprojectors in bH ⊕ eHb onto bH, Hb and bHb respectively. M+(λ), (4.106) Now combining (4.60) with (4.103) -- (4.106) one gets (4.68) and the equality eΓ0bv0(λ) = iP bHb Ψ+(λ), eΓ0bu+(λ) = IHb + iP bHb M+(λ), λ ∈ C+. − iP bH⊕ bHb M+(λ)T+(λ))Ψ+(λ), λ ∈ C+. (ibΓavτ (λ) − P bH ) +eΓ0bvτ (λ) = (−T+(λ) + iP bH⊕ bHb 26 VADIM MOGILEVSKII Moreover, (4.61) together with (4.79), (4.81) and (4.33) leads to (4.69) and the equality (ibΓavτ (λ)−P bH )+eΓ0bvτ (λ) = −eP bHT−(λ)Ψ−(λ)−P eHb Therefore in view of (4.64) and (4.102) one has T−(λ)Ψ−(λ) = −T−(λ)Ψ−(λ), λ ∈ C−. {(ibΓavτ (λ) − P bH )h0 +eΓ0bvτ (λ)h0, Γ1bvτ (λ)h0} ∈ τ±(λ), which implies (4.100) and (4.101). Finally, uniqueness of vτ (·, λ) follows from uniqueness of the solution of the boundary value problem (4.73) -- (4.76). (cid:3) h0 ∈ H0, λ ∈ C±, 5. m-functions and eigenfunction expansions U and Γ0a is the mapping (3.15). 5.1. m-functions. In this subsection we assume that eU is a J-unitary extension (3.13) of ∆[H0, H] be the operator solution of Eq. (3.4) defined in Theorems 4.6 and 4.12. Similarly to the case n−(Tmin) ≤ n+(Tmin) (see [2]) we introduce the following definition. Let τ be a boundary parameter and let vτ (·, λ) ∈ L2 Definition 5.1. The operator function mτ (·) : C \ R → [H0] defined by (5.1) is called the m-function corresponding to the boundary parameter τ or, equivalently, to the boundary value problem (4.4) -- (4.7) (in Case 1 ) or (4.73) -- (4.76) (in Case 2 ). mτ (λ) = (Γ0a +bΓa)vτ (λ) + i 2 P bH , λ ∈ C \ R, From (4.56) it follows from that mτ (·) satisfies the equality (5.2) eU vτ (a, λ)(cid:18)=(cid:18)Γ0a +bΓa Γ1a (cid:19) vτ (λ)(cid:19) =(cid:18)mτ (λ) − i −PH 2 P bH (cid:19) : H0 → H0 ⊕ H, It is easily seen that Proposition 5.3 in [2] remains valid in the case n+(Tmin) < n−(Tmin). This means that for a given operator U (see (3.9)) and a boundary parameter τ the m- functions m(1) connected by m(1) τ (·) corresponding to J-unitary extensions eU1 and eU2 of U are A definition of the m-function mτ in somewhat other terms ia given in the following τ (λ) + B, λ ∈ C \ R, with some B = B∗ ∈ [H0]. τ (·) and m(2) τ (λ) = m(2) λ ∈ C \ R. proposition, which directly follows from Theorems 4.6 and 4.12. Proposition 5.2. Let τ = {τ+, τ−} be a boundary parameter (4.1), (4.3) in Case 1 (resp. (4.70), (4.72) in Case 2), let ϕU (·, λ)(∈ [H0, H]), λ ∈ C, be the operator solution of Eq. (3.4) defined by (3.19) and let ψ(·, λ)(∈ [H0, H]), λ ∈ C, be the operator solutions of Eq. (3.4) with eU ψ(a, λ) =(cid:18)− i 2 P bH −PH (cid:19) : H0 → H0 ⊕ H. Then there exists a unique operator function m(·) = mτ (·) : C \ R → [H0] such that for any λ ∈ C \ R the operator solution v(·, λ) = vτ (·, λ) of Eq. (3.4) given by (5.3) v(t, λ) := ϕU (t, λ)m(λ) + ψ(t, λ) belongs to L2 (4.100), (4.101) in Case 2. ∆[H0, H] and satisfies the boundary conditions (4.57) -- (4.59) in Case 1 and In the following theorem we give a description of all m-functions immediately in terms of the boundary parameter τ . ON EIGENFUNCTION EXPANSIONS 27 Theorem 5.3. Assume the following hypothesis: (i) τ0 is a boundary parameter from Remark 4.3 in Case 1 and Remark 4.9 in Case 2. (ii) m0(·), Φ−(·), Ψ−(·) and M−(·) are the operator functions given by (4.30), (4.31) and (4.33) (this functions form the operator matrix X−(·) given by (4.29) in Case 1 and (4.92) in Case 2). (iii) M1(·) and Mj−(·), j ∈ {2, 3, 4}, are the operator functions defined by the block repre- sentations (4.35) (in Case 1) and (4.95)(in Case 2) of the Weyl function M−(·) (according to Propositions 4.5 and 4.11 these functions can be also defined as the entries of the block matrix representations (4.36) -- (4.40) of m0(·), Φ−(·), Ψ−(·) and M−(·)). Then: (1) m0(λ) = mτ0(λ), λ ∈ C \ R, and for any boundary parameter τ = {τ+, τ−} defined by (4.1) in Case 1 and (4.70) in Case 2 the corresponding m-function mτ (·) is (5.4) mτ (λ) = m0(λ) + Φ−(λ)(D0(λ) − D1(λ) M−(λ))−1D1(λ)Ψ−(λ), λ ∈ C−. (2) For each truncated boundary parameter τ = {τ+, τ−} defined by (4.1), (4.2) in Case 1 and (4.70), (4.71) in Case 2 the corresponding m-function mτ (·) has the triangular block matrix representation (5.5) mτ (λ) =(cid:18)mτ,1(λ) mτ,2(λ) 2 I (cid:19) , − i 0 λ ∈ C−, with respect to the decomposition H0 = H ′ Case 2. Moreover, the operator function mτ,1(·) in (5.5) is (5.6) mτ,1(λ) = M1(λ) + M2−(λ)(D0(λ) − D1(λ)M4−(λ))−1D1(λ)M3−(λ), 0 ⊕ bH2 in Case 1 (see (3.27)) and H0 = H ⊕ bH in λ ∈ C−. Proof. We prove the theorem only for Case 1, because in Case 2 the proof is similar. (1) It is easily seen that v0(t, λ) = vτ0 (t, λ), where v0(·, λ) is defined in Proposition 4.4. Hence by (4.30) one has m0(λ) = mτ0(λ). Next, applying the operator Γ0a +bΓa to the equalities (4.60) and (4.61) with taking (4.30) and (4.31) into account one gets −(λ) + M+(λ))−1Ψ+(λ), mτ (λ) = m0(λ) − Φ+(λ)(τ ∗ mτ (λ) = m0(λ) − Φ−(λ)(τ−(λ) + M−(λ))−1S−(λ), λ ∈ C+, λ ∈ C−. (5.7) (5.8) Moreover, according to [30, Lemma 2.1] 0 ∈ ρ(D0(λ) − D1(λ) M−(λ)) and −(τ−(λ) + M−(λ))−1 = (D0(λ) − D1(λ) M−(λ))−1D1(λ), λ ∈ C−, which together with (5.8) yields (5.4). (2) It follows from (4.2) and the second equality in (4.40) that (D0(λ)−D1(λ) M−(λ))−1 = −D1(λ)N2−(λ) D0(λ) − D1(λ)M4−(λ)(cid:19)−1 (cid:18) 0 I =(cid:18)I ∗ (D0(λ) − D1(λ)M4−(λ))−1(cid:19) 0 (the entry ∗ does not matter). Combining this equality with (5.4) and taking (4.37) -- (4.39) into account one gets the equalities (5.5) and (5.6). (cid:3) By using the reasonings similar to those in the proof of Proposition 5.7 in [2] one proves the following proposition. Proposition 5.4. The m-function mτ (·) is a Nevanlinna operator function satisfying (Im λ)−1 · Im mτ (λ) ≥ZI v∗ τ (t, λ)∆(t)vτ (t, λ) dt, λ ∈ C−. 28 VADIM MOGILEVSKII 5.2. Green's function. In the sequel we put H := L2 ∆(I) and denote by Hb the set of all ef ∈ H with the following property: there exists β ef ∈ I such that for some (and hence for all) function f ∈ ef the equality ∆(t)f (t) = 0 holds a.e. on (β ef , b). Let ϕU (·, λ) be the operator-valued solution (3.19), let τ be a boundary parameter and ∆[H0, H] be the operator-valued solution of Eq. (3.4) defined in Theorems let vτ (·, λ) ∈ L2 4.6 and 4.12. Definition 5.5. The operator function Gτ (·, ·, λ) : I × I → [H] given by Gτ (x, t, λ) =(vτ (x, λ) ϕ∗ ϕU (x, λ) v∗ U (t, λ), x > t τ (t, λ), x < t , λ ∈ C \ R (5.9) will be called the Green's function corresponding to the boundary parameter τ . (5.10) RI Theorem 5.6. Let τ be a boundary parameter and let Rτ (·) be the corresponding generalized resolvent of the relation T (see Theorems 4.2 and 4.8). Then (5.11) f ∈ ef . ∆(I) the inequality ef ∈ H, Gτ (x, t, λ)∆(t)f (t) dt < ∞ is valid. This implies that formula Proof. As in [2, Theorem 6.2] one proves that for each f ∈ L2 Gτ (·, t, λ)∆(t)f (t) dt(cid:19) , Rτ (λ)ef = π(cid:18)ZI yf = yf (x, λ) :=ZI following statement: for each ef ∈ H To prove (5.12) first assume that ef ∈ Hb. We show that the function yf (·, λ) given by (5.11) is a solution of the boundary problem (4.4) -- (4.7) in Case 1 (resp. (4.73) -- (4.76) in Case 2 ). As in [2, Theorem 6.2] one proves that yf (·, λ) belongs to AC(I; H) ∩ L2 ∆(I) and satisfies (4.4) a.e. on I. Now it remains to show that yf satisfies the boundary conditions (4.5) -- (4.7) in Case 1 (resp. (4.74) -- (4.76) in Case 2 ). ∆(I) and Rτ (λ)ef = π(yf (·, λ)), correctly defines the function yf (·, ·) : I × C \ R → H. Therefore (5.10) is equivalent to the f ∈ ef , λ ∈ C \ R. Gτ (x, t, λ)∆(t)f (t) dt, yf (·, λ) ∈ L2 λ ∈ C \ R (5.12) It follows from (5.11) and (5.9) that (5.13) yf (a, λ) = ϕU (a, λ)ZI v∗ τ (t, λ)∆(t)f (t) dt, ϕ∗ (5.14) x ∈ (β ef , b). U (t, λ)∆(t)f (t) dt, yf (x, λ) = vτ (x, λ)ZI Let eU be a J-unitary extension (3.13) of U and let Γa be the operator (3.17). Then Γayf = eU yf (a, λ) and in view of (5.13) one has Γayf = eU ϕU (a, λ)ZI bΓayf = P bHZI v∗ τ (t, λ)∆(t)f (t) dt, v∗ τ (t, λ)∆(t)f (t) dt. This and (3.20) imply Γ1ayf = 0. (5.15) ON EIGENFUNCTION EXPANSIONS 29 Next assume that (5.16) h ef :=ZI ϕ∗ U (t, λ)∆(t)f (t) dt(∈ H0) Then in view of (5.14) one has (see [2, Remark 3.5 (1)]) (5.17) Let us also prove the equality bΓbyf =bΓbvτ (λ)h ef , Γ0byf = Γ0bvτ (λ)h ef , Γ1byf = Γ1bvτ (λ)h ef . (5.18) It follows from (5.2) that bΓayf = (bΓavτ (λ) + iP bH )h ef . eU vτ (a, λ) ↾ bH =(cid:18)(mτ (λ) − i 0 2 IH0 ) ↾ bH (cid:19) : bH → H0 ⊕ H, λ ∈ C \ R. Therefore by (3.20) one has vτ (t, λ) ↾ bH = ϕU (t, λ)(mτ (λ)− i τ (λ) = mτ (λ). This implies that to Proposition 5.4 m∗ 2 IH0 ) ↾ bH. Moreover, according P bH v∗ τ (t, λ) = P bH (mτ (λ) + i 2 IH0 )ϕ∗ U (t, λ) = (P bH mτ (λ) + i 2 P bH )ϕ∗ U (t, λ) and (5.15), (5.16) yield (5.19) bΓayf = (P bH mτ (λ) + i 2 P bH )ZI ϕ∗ U (t, λ)∆(t)f (t) dt = (P bH mτ (λ) + i 2 P bH )h ef . Since in view of (5.1) P bH mτ (λ) =bΓavτ (λ) + i Next assume Case 1 and the hypothesis (A2). Then by (3.26) and (5.18) The second equality in (5.15) gives the first condition in (4.5) and the condition (4.74). 2 P bH , the equality (5.19) gives (5.18). (5.20) bΓa1yf = (bΓa1vτ (λ) + iP bH1 )h ef , bΓa2yf = (bΓa2vτ (λ) + iP bH2 )h ef and combining of (5.20) and (5.17) with (4.57) -- (4.59) shows that the second condition in (4.5) and the conditions (4.6) and (4.7) are fulfilled for yf . fulfillment of the conditions (4.75) and (4.76) for yf . Thus yf (·, λ) is a solution of the boundary value problem (4.4) -- (4.7) in Case 1 (resp. Now assume Case 2 and the hypothesis (A3). Then by (3.31) and (5.17) one haseΓ0byf = eΓ0bvτ (λ)h ef . Combining of this equality and (5.17), (5.18) with (4.100) and (4.101) implies (4.73) -- (4.76) in Case 2 ) and by Theorems 4.2 and 4.8 relations (5.12) hold (for ef ∈ Hb). Finally, one proves (5.12) for arbitrary ef ∈ H in the same way as in [2, Theorem 6.2]. Hamiltonian systems (i.e., for systems (3.3) with bH = {0}). Moreover, for non-Hamiltonian systems (3.3) in the case of minimally possible deficiency indices n±(Tmin) = ν± formulas (5.9) and (5.10) were proved in [17]. Remark 5.7. Theorem 5.6 is a generalization of similar results obtained in [7, 16, 25] for (cid:3) 5.3. Spectral functions and the Fourier transform. Let T be a symmetric relation in H defined by (3.35) in Case 1 and (3.39) in Case 2 and let τ = {τ+, τ−} be a boundary parameter given by (4.1) in Case 1 and (4.70) in Case 2 . According to Theorems 4.2 and 4.8 the boundary value problems (4.4) -- (4.7) and (4.73) -- (4.76) establish a bijective correspondence between boundary parameters τ and generalized resolvents R(·) = Rτ (·) of 30 VADIM MOGILEVSKII bf (s) =ZI f ∈ ef . (5.21) Rτ (λ) = PH(eT τ − λ)−1 ↾ H, λ ∈ C \ R. T . Denote by eT τ the (H-minimal) exit space self-adjoint extension of T in the Hilbert space eH ⊃ H generating Rτ (·): Clearly, formula (5.21) gives a parametrization of all such extensions eT = eT τ by means of a boundary parameter τ . Denote by Fτ (·) the spectral function of T corresponding to eT τ . In the following we assume that a certain J-unitary extension eU of U is fixed and hence a choice of eU does not matter in our further considerations. For each ef ∈ Hb introduce the Fourier transform bf (·) : R → H0 by setting the m-function mτ (·) is defined (see (5.1)). Note that in view of the assertion just after (5.2) U (t, s)∆(t)f (t) dt, (5.22) ϕ∗ (5.23) Definition 5.8. A distribution function Στ (·) : R → [H0] is called a spectral function of the boundary value problem (4.4) -- (4.7) in Case 1 (resp. (4.73) -- (4.76) in Case 2 ) if, for each ef ∈ Hb and for each finite interval [α, β) ⊂ R, the Fourier transform (5.22) satisfies the equality ((Fτ (β) − Fτ (α))ef ,ef )H =Z[α,β) (dΣτ (s)bf (s),bf (s)). In the following "the boundary value problem" means either the boundary value problem (4.4) -- (4.7) (in Case 1 ) or the boundary value problem (4.73) -- (4.76) (in Case 2 ). Let Στ (·) be the spectral function of the boundary value problem. Then in view of (5.23) bf ∈ L2(Στ ; H0) and the same reasonings as in [2] give the Bessel inequality bf L2(Στ ;H0) ≤ ef H (for the Hilbert space L2(Σ; H) see Subsection 2.2). Therefore for each ef ∈ H there exists a function bf ∈ L2(Στ ; H0) (the Fourier transform of ef ) such that f ∈ ef , and the equality V ef = bf , ef ∈ H, defines the contraction V : H → L2(Στ ; H0) Theorem 5.9. For each boundary parameter τ there exists a unique spectral function Στ (·) of the boundary value problem. This function is defined by the Stieltjes inversion formula bf −Z β U (t, ·)∆(t)f (t) dtL2(Στ ;H0) = 0, lim β↑b ϕ∗ a (5.24) Στ (s) = − lim δ→+0 lim ε→+0 Im mτ (σ − iε) dσ. If in addition τ is a truncated boundary parameter, then the corresponding spectral function Στ (·) has the block matrix representation 1 πZ s−δ −δ (5.25) 2π sI (cid:19) , Στ (s) =(cid:18)Στ,1(s) Στ,2(s) Στ,3(s) 1 s ∈ R with respect to the decomposition H0 = H ′ In (5.25) Στ,1(·) is an [H ′ 0]-valued (resp. [H]-valued) distribution function defined by 0 ⊕ bH2 in Case 1 (resp. H0 = H ⊕ bH in Case 2). πZ s−δ Im mτ,1(σ − iε) dσ −δ 1 (5.26) Στ,1(s) = − lim δ→+0 with mτ,1(·) taken from (5.5) lim ε→+0 ON EIGENFUNCTION EXPANSIONS 31 Proof. In the case of an arbitrary boundary parameter τ one proves the required statement by using Theorem 5.6 and the Stieltjes-Livsic inversion formula in the same way as in [2, Theorem 6.5]. The statements of the theorem for a truncated boundary parameter τ are implied by (5.24) and the block matrix representation (5.5) of mτ (·). (cid:3) Let H be decomposed as (5.27) H = H0 ⊕ mul T (5.28) and let T ′ be the operator part of T . Moreover, let τ be a boundary parameter, let eT τ = (eT τ )∗ be an exit space extension of T in the Hilbert space eH ⊃ H, let eH be decomposed as and let T τ be the operator part of eT τ (so that T τ is a self-adjoint operator in eH0). Assume also that Στ (·) is a spectral function of the boundary value problem and V ∈ [H, L2(Στ ; H0)] is the Fourier transform. As in [2] one proves that V ↾ mul T = 0. Therefore by (5.27) H0 is the maximally possible subspace of H on which V may by isometric. eH =eH0 ⊕ muleT τ Definition 5.10. A spectral function Στ (·) is referred to the class SF0 if the operator V0 := V ↾ H0 is an isometry from H0 to L2(Στ ; H0). Similarly to [2] the following equivalence is valid: Στ (·) ∈ SF0 ⇐⇒ muleT τ = mul T. Hence all spectral functions Στ (·) of the boundary value problem belong to SF0 if and only if mul T = mul T ∗ or, equivalently, if and only if the operator T ′ is densely defined. inverse Fourier transform is As in [2] one proves the following statement: if Στ (·) ∈ SF0, then for each ef ∈ H0 the (5.29) (5.30) ef = π(cid:18)ZR ϕU (·, s) dΣτ (s)bf (s)(cid:19) , ∆(I). where the integral converges in the semi-norm of L2 The following theorem can be proved in the same way as Theorem 6.9 in [2]. Theorem 5.11. Let τ be a boundary parameter such that Στ (·) ∈ SF0 and let V be the corresponding Fourier transform. Then H0 ⊂ eH0 and there exists a unitary extension eV ∈ [eH0, L2(Στ ; H0)] of the isometry V0(= V ↾ H0) such that the operator T τ and the multiplication operator Λ = ΛΣτ in L2(Στ ; H0) are unitarily equivalent by means of eV . Moreover, if mul T = mul T ∗, then the statements of the theorem hold for arbitrary bound- ary parameter τ (that is, for any spectral function Στ (·)). Since in the case of a densely defined T one has mul T = mul T ∗ = {0}, the following theorem is implied by Theorem 5.11. Theorem 5.12. If T is a densely defined operator, then for each boundary parameter τ and the corresponding spectral function Στ (·) the following hold: (i) eT τ is an operator, that is, eT τ = T τ ; (ii) the Fourier transform V is an isometry; (iii) there exists a unitary extension eV ∈ [eH, L2(Στ ; H0)] of V such that eT τ and Λ are unitarily equivalent by means of eV . Since n+(Tmin) 6= n−(Tmin), the equality σ(T τ ) = R holds for any boundary parameter τ . Moreover, Theorem 5.11 yields the following corollary. 32 VADIM MOGILEVSKII Corollary 5.13. (1) If τ is a boundary parameter such that Στ (·) ∈ SF0, then the spectral multiplicity of the operator T τ does not exceed ν−(= dim H0). (2) If mul T = mul T ∗, then the above statement about the spectral multiplicity of T τ holds for any boundary parameter τ . In the case of the truncated boundary parameter τ there is a somewhat more information about the spectrum of T τ . Namely, the following theorem is valid. Theorem 5.14. (1) Let τ be a truncated boundary parameter such that Στ (·) ∈ SF0 and let Στ,1(·) be a distribution function given by the block matrix representation (5.25) of Στ (·). Moreover, let E(·) be the orthogonal spectral measure of the operator T τ and let Es(·) be the singular part of E(·). Then σac(T τ ) = R, σs(T τ ) = Ss(Στ,1) and the multiplicity of the orthogonal spectral measure Es(·) does not exceed ν+ +νb+ −νb−(= dim H ′ 0) in Case 1 and ν+(= dim H) in Case 2 (this implies that the same estimates are valid for multiplicity of each eigenvalue λ0 of T τ ). (2) If mul T = mul T ∗, then the above statements about the spectrum of T τ hold for each truncated boundary parameter τ . Proof. (1) let Στ,ac(·), Στ,s(·) and Στ 1,ac(·), Στ 1,s(·) be absolutely continuous and singular parts of Στ (·) and Στ,1(·) respectively. Then in view of (5.25) the block matrix representa- tions 2π sI (cid:19) , Στ,ac(s) =(cid:18)Στ 1,ac(s) Στ,2(s) Στ,3(s) 1 Στ,s(s) =(cid:18)Στ 1,s(s) 0 0(cid:19) 0 hold with respect to the same decompositions of H0 as in Theorem 5.9. This and Theorems 5.11 and 2.2 give the required statements about the spectrum of T τ . The statement (2) of the theorem follows from the fact that in the case mul T = mul T ∗ the inclusion Στ (·) ∈ SF0 holds for any boundary parameter τ and hence for any truncated boundary parameter τ . (cid:3) In the next theorem we give a parametrization of all spectral functions Στ (·) immediately in terms of the boundary parameter τ . Theorem 5.15. Let X−(·) be the operator matrix (4.29) in Case 1 (resp. (4.92) in Case 2). Then for each boundary parameter τ = {τ+, τ−} of the form (4.1) in Case 1 (resp. (4.70) in Case 2) the equality (5.31) mτ (λ) = m0(λ) + Φ−(λ)(D0(λ) − D1(λ) M−(λ))−1D1(λ)Ψ−(λ), λ ∈ C− together with (5.24) defines a (unique) spectral function Στ (·) of the boundary value problem. If in addition τ = {τ+, τ−} is a truncated boundary parameter defined by (4.1), (4.2) in Case 1 and (4.70), (4.71) in Case 2, then the distribution function Fτ (·) in (5.25) is defined by (5.6) and (5.26). Moreover, the following statements are valid: (1) Let in Case 1 bC02(λ), C0b(λ) and M4+(λ), N2+(λ) be defined by (4.3) and (4.40) and let PHb be the orthoprojector in bH2 ⊕ Hb onto Hb. Then there exist the limits Bτ := lim y→+∞ 1 1 iy PHb (D0(iy) − D1(iy) M−(iy))−1D1(iy) iy (C0b(iy) − C1(iy)M4+(iy) + ibC02(iy)N2+(iy))−1C1(iy) = iy M4+(iy)(C0b(iy) − C1(iy)M4+(iy) + ibC02(iy)N2+(iy))−1C0b(iy) 1 = lim y→−∞ bBτ := lim y→+∞ ON EIGENFUNCTION EXPANSIONS 33 and the following equivalence holds: (5.32) (2) Let in Case 2 C0(λ) and M+(λ) has the block representations (cf. (4.72) and (4.40)) Στ (·) ∈ SF0 ⇐⇒ Bτ = bBτ = 0 C0(λ) = (C02(λ), C0b(λ)) : (bH ⊕ bHb) ⊕ Hb → Hb, λ ∈ C+ M+(λ) = (N2+(λ), N4+(λ), M4(λ))⊤ : Hb → bH ⊕ bHb ⊕ Hb, and let N+(λ) = (N2+(λ), N4+(λ))⊤ ∈ [Hb, bH ⊕ bHb], λ ∈ C+. Then there exist limits iy (C0b(iy) − C1(iy)M4(iy) + iC02(iy)N+(iy))−1C1(iy) = iy PHb (D0(iy) − D1(iy) M−(iy))−1D1(iy) Bτ := lim λ ∈ C+ y→+∞ 1 1 = lim y→−∞ bBτ := lim y→+∞ 1 iy M4(iy)(C0b(iy) − C1(iy)M4(iy) + iC02(iy)N+(iy))−1C0b(iy) and the equivalence (5.32) holds. (3) Each spectral function Στ (·) belongs to SF0 if and only if lim y→+∞ 1 iy M4+(iy) = 0 in Case 1 (resp. lim y→+∞ 1 iy M4(iy) = 0 in Case 2) and lim y→−∞ y(Im( M−(iy)h, h) − 1 2 P2h2) = +∞, h 6= 0, where h ∈ bH2 ⊕ Hb and P2 is the orthoprojector in bH2 ⊕ Hb onto bH2 in Case 1 (resp. h ∈ bH2 ⊕ eHb and P2 is the orthoprojector in bH2 ⊕ eHb onto bH2 ⊕ bHb in Case 2). Proof. The main statement of the theorem directly follows from Theorems 5.3 and 5.9. Next, consider the boundary triplet Π− = { H0⊕Hb, Γ0, Γ1} for T ∗ defined in Propositions 3.7 and 3.8. Since the Weyl functions M±(·) of the decomposing boundary triplet Π− for Tmax have the block representations (4.34), (4.35) in Case 1 and (4.94), (4.95) in Case 2 , it follows from (4.40) and Proposition 2.7, (3) that the Weyl functions of the triplet Π− are M+(λ) and M−(λ). Now applying to the boundary triplet Π− Theorems 4.12 and 4.13 from [35] and taking equivalence (5.29) into account one obtains statements (1) -- (3) of the theorem. (cid:3) 5.4. The case of minimal deficiency indices. It follows from (3.22) that for a given system (3.3) minimally possible deficiency indices of the linear relation Tmin are (5.33) n+(Tmin) = ν+, n−(Tmin) = ν− and the first (second) equality in (5.33) holds if and only if νb+ = 0 (resp. νb− = 0). If n+(Tmin) = ν+, then by (3.22) one has n+(Tmin) ≤ n−(Tmin); moreover, n+(Tmin) < In the case of the minimal deficiency index n+(Tmin) the above results can be somewhat simplified. Namely, assume the first equality in (5.33) and let n+(Tmin) < n−(Tmin). Then the equality νb+ = 0 implies Case 2. Moreover, Lemma 3.4 gives Hb = {0} and hence n−(Tmin) if in additionbν > 0(⇔ bH 6= {0}). Γ0b = Γ1b = 0. Therefore eHb = bHb and (3.31) yieldseΓ0b =bΓb. If the assumption (A1) from Subsection 4.1 is satisfied, then by Proposition 3.8 the equality (5.34) T = {{ey,ef} ∈ Tmax : Γ1ay = 0, bΓay = 0, bΓby = 0}. (5.35) (5.36) 34 VADIM MOGILEVSKII defines a maximal symmetric relation T in H(= L2 ∆(I)) with the deficiency indices n+(T ) = 0 and n−(T ) = bν +bνb−. Moreover, in this case T coincides with the relation A0 defined by (3.42). Hence T has a unique generalized resolvent R(λ) of the form (4.9), which in view of Theorem 4.8 and Remark 4.9 is given by the boundary value problem J y′ − B(t)y = λ∆(t)y + ∆(t)f (t), t ∈ I, Γ1ay = 0, λ ∈ C+; Γ1ay = 0, bΓay = 0, bΓby = 0, λ ∈ C−. m(·) of the problem (5.35), (5.36) is given by the relations If eU is a J-unitary extension (3.13) of U , then according to Proposition 5.2 the m-function and ibΓav(λ) = P bH , bΓbv(λ) = 0, λ ∈ C−. Since Hb = {0}, the decomposing boundary triplet (3.32) -- (3.34) for Tmax takes the form v(t, λ) := ϕU (t, λ)m(λ) + ψ(t, λ) ∈ L2 Π− = {H0 ⊕ H, Γ0, Γ1}, where ∆[H0, H], λ ∈ C \ R Assume that the Weyl function M−(·) of Π− has the block matrix representation (cf. (4.95)) H0 = H ⊕ bH ⊕ bHb = H0 ⊕ bHb, (5.37) Γ0{ey,ef } = {−Γ1ay, ibΓay,bΓby} (∈ H ⊕ bH ⊕ bHb), Γ1{ey,ef} = Γ0ay, M−(λ) = (M (λ), N−(λ), M−(λ)) : H ⊕ bH ⊕ bHb → H, m(λ) =(cid:18)M (λ) N−(λ) According to Theorem 5.3 m(λ) = m0(λ) and by (4.97) one has (5.38) − i 0 , λ ∈ C−. {ey,ef } ∈ Tmax. λ ∈ C−. Let M(·) be the operator function (2.10), (2.11) corresponding to the decomposing boundary triplet Π−. Then by (5.37) and (5.38) (5.39) , λ ∈ C−, 2 I bH(cid:19) : H ⊕ bH {z }H0 2 I bHb(cid:19) : H0 ⊕ bHb {z }H0 → H ⊕ bH {z }H0 → H0 ⊕ bHb {z }H0 0 − i M(λ) =(cid:18)m(λ) M1(λ) m(µ) − m∗(λ) = (µ − λ)ZI Σ(s) =(cid:18) F (s) Σ1(s) Σ2(s) where M1(λ) = (M−(λ), 0)⊤(∈ [bHb, H ⊕ bH]). Using (5.39) and (2.12) one can show that v∗(t, λ)∆(t)v(t, µ) dt, µ, λ ∈ C−. The spectral function Σ(·) of the problem (5.35), (5.36) has the block matrix representation 1 2π sI bH(cid:19) : H ⊕ bH → H ⊕ bH, s ∈ R, where F (s) is an [H]-valued distribution function defined by the Stieltjes formula (5.26) with M (λ) in place of mτ,1(λ). Moreover, since T is maximal symmetric, it follows that mul T = mul T ∗ and, consequently, Σ(·) ∈ SF0. Hence the operator part T ′ 0 of the (unique) exit space self-adjoint extension T0 of T satisfies statements of Theorem 5.14 with T ′ 0 and F (·) in place of T τ and Fτ (·) respectively. If in addition to the above assumptions n−(Tmin) = ν− (i.e., both the deficiency in- dices n±(Tmin) are minimal), then bHb = {0} and bΓb = 0. This implies the corresponding modification of the results of this subsection. In particular, formula (5.39) takes the form m(λ) = M(λ), λ ∈ C \ R. ON EIGENFUNCTION EXPANSIONS 35 6. Differential operators of an odd order In this section we apply the above results to ordinary differential operators of an odd order on an interval I = [a, bi (−∞ < a < b ≤ ∞) with the regular endpoint a. Assume that H is a finite dimensional Hilbert space and (6.1) l[y] = (−1)k(cid:16) i mXk=0 2 [(qn−ky(k))(k+1) + (qn−ky(k+1))(k)] + (pn−ky(k))(k)(cid:17) is a differential expression of an odd order 2m + 1 with sufficiently smooth operator valued coefficients pk(·), qk(·) : I → [H] such that pk(t) = p∗ k(t) and 0 ∈ ρ(q0(t)). Denote by y[k](·), k ∈ {0, . . . , 2m + 1}, the quasi-derivatives of y ∈ AC(I; H) and let dom l be the set of all functions y ∈ AC(I; H) for which l[y] := y[2m+1] makes sense [37, 24]. k(t), qk(t) = q∗ As is known H = H + t ⊕ H − t , where H + t (H − t ) is an invariant subspace of the operator q0(t), on which q0(t) is strictly positive (resp. negative). We put ν0+ = dim H + t , ν0− = dim H − t (this numbers does not depend on t); moreover, we assume for definiteness that ν0− ≤ ν0+. By using formula (1.27) in [24] one can easily show that there exist finite dimensional Hilbert spaces H ′ and bH and an absolutely continuous operator function t ∈ I, such that 0 ∈ ρ(Q(t)) and the following holds: iq0(t) = −Q∗ Q(t) = (Q1(t), bQ(t), Q2(t))⊤ : H → H ′ ⊕ bH ⊕ H ′, 2(t)Q1(t) + ibQ∗(t)bQ(t), dimbH = ν0+ − ν0−. 1(t)Q2(t) + Q∗ dim H ′ = ν0−, H = H ⊕ · · · ⊕ H ⊕H ′(= H m ⊕ H ′) Introduce the finite dimensional Hilbert spaces (cf. (3.1)) (6.2) (6.3) t ∈ I m terms {z } Clearly, the space H admits the representation H0 = H ⊕ bH = H m ⊕ H ′ ⊕ bH, H = H0 ⊕ H = H ⊕ bH ⊕ H (6.4) H = H ⊕ · · · ⊕ H m terms For each function y ∈ dom l we let z H } {z { ⊕H ′ ⊕bH ⊕ z } H ⊕ · · · ⊕ H H } {z m terms } ⊕H ′ { (6.5) (6.6) (6.7) y0(t) = {y(t), . . . , y[m−1](t), Q1(t)y(m)(t)}(∈ H) y1(t) = {y[2m](t), . . . , y[m+1](t), Q2(t)y(m)(t)}(∈ H) y(t) = {y0(t), bQ(t)y(m)(t), y1(t)}(∈ H ⊕ bH ⊕ H = H). Let K be a finite dimensional Hilbert space. For an operator valued solution Y (·) : I → [K, H] of the equation (6.8) l[y] = λy (λ ∈ C) we define the operator function Y(·) : I → [K, H] by the following relations: if h ∈ K and Y (t)h = y(t), then Y(t)h = y(t). 36 VADIM MOGILEVSKII on I satisfyingRI Next assume that H′ := L2(I) is the Hilbert space of all Borel H-valued functions f (·) f (t)2 dt < ∞. Denote also by L2[K, H] the set of all operator functions Y (·) : I → [K, H] such that Y (t)h ∈ H′, h ∈ K. Moreover, by using (6.4) we associate with a function f (·) ∈ H′ the H-valued functions f (·) on I given by f (t) = {f (t), 0, . . . , 0}, t ∈ I. According to [37] expression (6.1) induces in H′ the maximal operator Lmax and the minimal operator Lmin. Moreover, Lmin is a closed densely defined symmetric operator and L∗ min = Lmax. It turns out that the expression l[y] is equivalent in fact to a certain symmetric system. More precisely, the following proposition is implied by the results of [24]. Proposition 6.1. Let l[y] be the expression (6.1) and let J = 0 0 IH 0 −IH iI bH 0 0 0  ∈ [H ⊕ bH ⊕ H], ∆(t) = P 0 0 0 0 0 0 0 0 ∈ [H ⊕ bH ⊕ H], where P is the orthoprojector in H onto the first term in the right hand side of (6.3). Then there exists a continuous operator function B(t) = B∗(t)(∈ [H]), t ∈ I, (defined in terms of pj and qj) such that the first-order symmetric system (6.9) Jy′(t) − B(t)y(t) = ∆(t)f (t), t ∈ I and the corresponding homogeneous system (6.10) Jy′(t) − B(t)y(t) = λ∆(t)y(t), t ∈ I, λ ∈ C possess the following properties: (2) Let Tmax be the maximal linear relation in L2 of the system (6.10). Moreover, Y (·, λ) ∈ L2[K, H] if and only if Y(·, λ) ∈ L2 (1) There is a bijective correspondence Y (·, λ) ↔ eY (·, λ) = Y(·, λ) between all [K, H]- valued operator solutions Y (·, λ) of Eq. (6.8) and all [K, H]-valued operator solutions eY (·, λ) ∆(I) induced by the system (6.9). Then the equality (V1y)(t) = y(t), y(·) ∈ dom Lmax, defines a bijective linear mapping V1 from dom Lmax onto dom Tmax satisfying (V1y, V1z)∆ = (y, z)H′, y, z ∈ dom Lmax. (3) Let Tmin and Tmax be minimal and maximal relations in H := L2 ∆(I) induced by system (6.9). Then Tmin is a densely defined operator and the equality V2f = π( f (·)), f = f (·) ∈ H′, defines a unitary operator V2 ∈ [H′, H] such that (V2 ⊕ V2) (gr Lmin) = gr Tmin and (V2 ⊕ V2) (gr Lmax) = gr Tmax (i.e., the operators Lmin and Tmin as well as Lmax and Tmax are unitarily equivalent by means of V2). ∆[K, H]. By using Proposition 6.1 one can easily translate all the results of [2] and the present paper to the expression (6.1). Below we specify only the basic points in this direction. respectively. Using (6.5) -- (6.7) introduce the linear mappings Γja : dom l → H, j ∈ {0, 1}, Let eU ∈ [H] be a J-unitary operator given by (3.13) with H and bH in place of H and bH andbΓa : dom l → bH by setting y ∈ dom l (6.11) (6.12) Γ0ay = u7y0(a) + u8bQ(a)y(m)(a) + u9y1(a), bΓay = u1y0(a) + u2bQ(a)y(m)(a) + u3y1(a), Γ1ay = u4y0(a) + u5bQ(a)y(m)(a) + u6y1(a). (Jy(t), z(t)), y, z ∈ dom Lmax. [y, z]b := lim t↑b Next, assume that νb+ and νb− are indices of inertia of the bilinear form ON EIGENFUNCTION EXPANSIONS 37 Combining (3.22) with (6.3) and (6.2) and taking Proposition 6.1, (3) into account one gets the following equality for deficiency indices d± = n±(Lmin) of the operator Lmin: (6.13) d+ = m · dim H + ν0− + νb+, d− = m · dim H + ν0+ + νb−. Therefore d+ < d− if and only if one of the following two alternative cases holds: Case 1. ν0+ − ν0− > νb+ − νb− > 0. Case 2. ν0+ − ν0− ≥ 0 ≥ νb+ − νb− and ν0+ − ν0− 6= νb+ − νb−(6= 0) Proposition 6.1, (2) enables one to identify dom Lmax and dom Tmax. Therefore in the case d− ≤ d+ we may assume that the linear mapping Γb in [2, Lemma 3.4] is of the form (6.14) Γb = (Γ0b,bΓb, Γ1b)⊤ : dom Lmax → H0b ⊕ bH ⊕ H1b, where H0b is a finite dimensional Hilbert space and H1b is a subspace in H0b. Similarly in the case d+ < d− we define on dom Lmax the linear mappings Γ0b, Γ1b,bΓb (Case 1 ) and eΓ0b, Γ1b (Case 2 ) in the same way as in Subsection 3.3. In the following we suppose that: (i) U is the operator (3.9) (with H and bH in place of H and bH) satisfying (3.10) -- (3.12); (ii) bΓa and Γ1a are the mappings (6.12). Then the same boundary conditions as in the right hand sides of (3.35), (3.39) and [2, (3.40)] give a symmetric operator T (⊃ Lmin) in H′. Moreover, we define a boundary parameter τ in the same way as for symmetric systems in [2, Definition 5.1] (the case d− ≤ d+) and Definitions 4.1 and 4.7 (the case d+ < d−). A boundary parameter τ gives a parametrization of all generalized resolvents Rτ (·) and, consequently, all spectral functions Fτ (·) of T . Such a parametrization is generated by means of a boundary value problem involving the equation (6.15) l[y] − λy = f (t), t ∈ I and the boundary conditions [2, (4.2) and (4.3)] (d− ≤ d+), (4.5) -- (4.7) (d+ < d−, Case 1 ) or (4.74) -- (4.76) (d+ < d−, Case 2). In the following "the boundary value problem for l[y]" means one of the listed above boundary value problems for the expression l[y]. If d+ = d−, then in (6.14) H0b = H1b =: Hb and Rτ (·) is a canonical resolvent of T if and only if τ = {(C0, C1); Hb} is a self-adjoint operator pair (this means that C0, C1 ∈ [Hb], ImC1C ∗ 0 = 0 and 0 ∈ ρ(C0 ± iC1)). In this case the corresponding boundary problem involves equation (6.15) and self-adjoint boundary conditions (6.16) Γ1ay = 0, bΓay =bΓby, C0Γ0by + C1Γ1by = 0, (6.17) T τ = Lmax ↾ dom T τ is a self-adjoint extension of Lmin with the domain where Γ0b, Γ1b and bΓb are taken from (6.14). Moreover, Rτ (λ) = (T τ − λ)−1, where Next assume that eU ∈ [H ⊕ bH ⊕ H] is a J-unitary extension (3.13) of U and Γ0a is the dom T τ = {y ∈ dom Lmax : Γ1ay = 0, bΓay =bΓby, C0Γ0by + C1Γ1by = 0}. Definition 6.2. The m-function of the expression l[y] corresponding to the boundary pa- rameter τ is the m-function mτ (·) of the equivalent system (6.9). mapping (6.11). Definition 6.2 means that mτ (·) : C \ R → [H0] is a unique operator function such that for any λ ∈ C \ R the operator solution vτ (·, λ) of Eq. (6.8) given by (6.18) belongs to L2[H0, H] and satisfies the boundary conditions [2, (4.41) -- (4.43)] (d− ≤ d+), In (6.18) (4.57) -- (4.59) (d+ < d−, Case 1 ) or (4.100) and (4.101) (d+ < d−, Case 2 ). vτ (t, λ) := ϕU (t, λ)mτ (λ) + ψ(t, λ) 38 VADIM MOGILEVSKII bf (s) =ZI ϕU (·, λ)(∈ [H0, H]) and ψ(·, λ)(∈ [H0, H]) are the operator solutions of Eq. (6.8) with the initial data 0 (cid:19) : H0 → H0 ⊕ H, eUϕϕϕU (a, λ) =(cid:18)IH0 (note that ϕU (·, λ) does not depend on a choice of the extension eU ⊃ U ). eUψψψ(a, λ) =(cid:18)− i b the Fourier transform bf (·) : R → H0 is −PH(cid:19) : H0 → H0 ⊕ H b be the set of all functions f (·) ∈ H′ such that f (t) = 0 on some interval (β, b) (depending on f ). For a function f (·) ∈ H′ ϕ∗ U (t, s)f (t) dt. Let H′ (6.19) 2 P bH Definition 6.3. Let τ be a boundary parameter. A distribution function Στ (·) : R → [H0] is called a spectral function of the boundary value problem for l[y] if for each f (·) ∈ H′ b the Fourier transform (6.19) satisfies (5.23) (with H′ in place of H). V : H′ → L2(Στ ; H0). f H′ . Therefore for each f ∈ H′ there exists the fourier transform (6.19) (the integral Since Lmin is densely defined, formula (5.23) yields the Parseval equality bf L2(Στ ;H0) = converges in the norm of L2(Στ ; H0)) and the equality V f = bf , f ∈ H′, define an isometry In view of Proposition 6.1 each spectral function Στ (·) of the expression l[y] is a spectral function of the equivalent symmetric system (6.9) and vice versa. Hence by [2, Theorem 6.5] and Theorem 5.9 for each boundary parameter τ there exits a unique spectral function Στ (·) of the boundary value problem for l[y] and this function is defined by the Stieltjes inversion formula (5.24). Moreover, for each f ∈ H′ the inverse Fourier transform is f (t) =ZR ϕU (t, s) dΣτ (s)bf (s). For a given boundary parameter τ denote by T τ the (H-minimal) exit space self-adjoint extension of T generating Rτ (λ). Since T is densely defined, T τ is a self-adjoint operator in The following theorem is implied by Theorem 5.12 and [2, Theorem 6.11]. Theorem 6.4. Let τ be a boundary parameter and let Στ (·) and V be the corresponding spectral function and Fourier transform respectively. Then there exists a unitary extension the Hilbert space eH ⊃ H′. eV ∈ [eH, L2(Στ ; H0)] of V such that T τ and the multiplication operator Λ in L2(Στ ; H0) are unitarily equivalent by means of eV . Moreover, the following statements are equivalent: (1) d+ = d− and τ = {(C0, C1); Hb} is a self-adjoint operator pair, so that T τ is the canonical self-adjoint extension (6.17) of T ; (2) V H′ = L2(Στ ; H0), that is the fourier transform V is a unitary operator. If the statement (1) (and hence (2)) is valid, then the operators T τ and Λ are unitarily equivalent by means of V . Theorem 6.4 immediately implies that the spectral multiplicity of T τ does not exceed m · dim H + ν0+(= dim H0). Similarly one can translate to the expression l[y] Theorem 5.14 and parametrization of all spectral functions Στ (·) by means of (5.31), [35, (5.22)] and the Stieltjes inversion formula. It follows from (6.13) that for a given expression (6.1) minimally possible deficiency indices of the operator Lmin are n+(Lmin) = m · dim H + ν0−, n−(Lmin) = m · dim H + ν0+. ON EIGENFUNCTION EXPANSIONS 39 The routine reformulation of the results of Subsection 5.4 to the case of the expression l[y] with minimal deficiency index n+(Lmin) of the operator Lmin is left to the reader. Remark 6.5. Let H = C, so that l[y] is a scalar expression with real valued coefficients pk(·) and qk(·). Then q0(t) > 0, bQ(t) = q H = Cm, 0 (t), bH = C and the equalities (6.3) -- (6.7) take the form H = Cm+1 ⊕ Cm = C2m+1 H0 = Cm ⊕ C = Cm+1, y0(t) = {y(t), . . . , y[m−1](t)}(∈ Cm), y1(t) = {y[2m](t), . . . , y[m+1](t)}(∈ Cm), 1 2 y(t) = {y0(t), q 0 (t)y(m)(t), y1(t)}(∈ C2m+1). 1 2 In this case ϕU (·, λ) is the (m + 1)-component operator solution ϕU (t, λ) = (ϕ1(t, λ), ϕ2(t, λ), . . . , ϕm(t, λ), ϕm+1(t, λ)) : Cm+1 → C of Eq. (6.8) with the initial data eUϕϕϕU (a, λ) =(cid:18)Im+1 0 (cid:19) : Cm+1 → Cm+1 ⊕ Cm and (6.19) defines the Fourier transform bf (·) : R → Cm+1. Moreover, in a fixed basis of Cm+1 the m-function mτ (·) and spectral function Στ (·) can be represented as (m + 1) × (m + 1)- matrix valued functions mτ (λ) = (mij (λ))m+1 i,j=1 and Στ (λ) = (Σij(λ))m+1 i,j=1 respectively. In conclusion note that for a scalar expression (6.1) one has ν0+ = 1 and ν0− = 0. Therefore for a scalar expression l[y] with n+(Lmin) < n−(Lmin) Case 1 is impossible. In other words, for such an expression either n−(Lmin) ≤ n+(Lmin) or n+(Lmin) < n−(Lmin) and Case 2 holds. References [1] N.I. Akhiezer, I.M. Glazman, Theory of linear operators in Hilbert space, Dover Publications, New York, 1993. [2] S. Albeverio,M.M. Malamud, V.I.Mogilevskii,On Titchmarsh-Weyl functions and eigenfunction ex- pansions of first-order symmetric systems, arXiv:1303.6153[math.FA] 25 Mar 2013. [3] F.V. Atkinson, Discrete and continuous boundary problems, Academic Press, New York, 1963. [4] J. Behrndt, S. Hassi, H. de Snoo, R. Wiestma, Square-integrable solutions and Weyl functions for singular canonical systems, Math. Nachr. 284 (2011), no 11-12, 1334 -- 1383. [5] V.A. Derkach, S. Hassi, M.M. Malamud, H.S.V. de Snoo, Generalized resolvents of symmetric oper- ators and admissibility, Methods of Functional Analysis and Topology 6 (2000), no 3 , 24 -- 55. [6] V.A. Derkach, M.M. Malamud, Generalized resolvents and the boundary value problems for Hermit- ian operators with gaps, J. Funct. Anal. 95 (1991),1 -- 95. [7] A. Dijksma, H. Langer, H.S.V. de Snoo, Eigenvalues and pole functions of Hamiltonian systems with eigenvalue depending boundary conditions, Math. Nachr. 161 (1993), 107 -- 153. [8] N. Dunford and J.T. Schwartz, Linear operators. Part2. Spectral theory, Interscience Publishers, New York-London, 1963. [9] W.N. Everitt and V. Krishna Kumar, On the Titchmarsh-Weyl theory of ordinary symmetric differ- ential expressions 1 : The general theory , Nieuw Arch. voor Wiskunde (3) 24 (1976), 1 -- 48. [10] W.N. Everitt and V. Krishna Kumar, On the Titchmarsh-Weyl theory of ordinary symmetric differ- ential expressions 2 : The odd-order case, Nieuw Arch. voor Wiskunde (3) 24 (1976), 109 -- 145. [11] Ch. T. Fulton, Parametrizations of Titchmarsh's m(λ)-functions in the limit circle case, Trans. Amer. Math. Soc. 229 (1977), 51 -- 63. [12] I. Gohberg, M.G. Krein, Theory and applications of Volterra operators in Hilbert space, Transl. Math. Monographs, 24, Amer. Math. Soc., Providence, R.I., 1970 [13] M.L. Gorbachuk, On spectral functions of a differential equation of the second order with operator- valued coefficients, Ukrain. Mat. Zh. 18 (1966), no 2 , 3 -- 21. 40 VADIM MOGILEVSKII [14] V.I. Gorbachuk, M.L. Gorbachuk, Boundary problems for differential-operator equations, Kluver Acad. Publ., Dordrecht-Boston-London, 1991. (Russian edition: Naukova Dumka, Kiev, 1984). [15] D.B. Hinton, J.K. Shaw, Parameterization of the M (λ) function for a Hamiltonian system of limit circle type, Proc. Roy. Soc. Edinburgh Sect. A 93 (1982/83), no 3-4, 349-360. [16] D.B. Hinton, A. Schneider, On the Titchmarsh-Weyl coefficients for singular S-Hermitian systems I, Math. Nachr. 163 (1993), 323 -- 342. [17] D.B. Hinton, A. Schneider, Titchmarsh-Weyl coefficients for odd order linear Hamiltonian systems, J. Spectral Mathematics 1(2006),1-36. [18] D.B. Hinton, A. Schneider, On the spectral representation for singular selfadjoint boundary eigen- function problems, Oper. Theory: Advances and Applications, 106, Birkhauser, Basel, 1998. [19] T. Kato,Perturbation theory of linear operators,Springer-Verlag, Berlin,1966. [20] I.S. Kats, Linear relations generated by the canonical differential equation of phase dimension 2, and eigenfunction expansion, St. Petersburg Math. J. 14, 429 -- 452 (2003). [21] I.S. Kac, M.G. Krein, On spectral functions of a string, Supplement to the Russian edition of F.V. Atkinson , Discrete and continuous boundary problems, Mir, Moscow, 1968. [22] A.M. Khol'kin, Description of selfadjoint extensions of differential operators of an arbitrary order on the infinite interval in the absolutely indefinite case, Teor. Funkcii Funkcional. Anal. Prilozhen. 44 (1985), 112 -- 122. [23] V.I. Khrabustovsky, On the characteristic operators and projections and on the solutions of Weyl type of dissipative and accumulative operator systems. 3. Separated boundary conditions, J. Math. Phis. Anal. Geom., 2 (2006), no 4, 449 -- 473. [24] V.I. Kogan, F.S. Rofe-Beketov, On square-integrable solutions of symmetric systems of differential equations of arbitrary order, Proc. Roy. Soc. Edinburgh Sect. A 74 (1974/75), 5 -- 40. [25] A.M. Krall, M (λ)-theory for singular Hamiltonian systems with one singular endpoint, SIAM J. Math. Anal. 20 (1989), 664 -- 700. [26] V. Krishna Kumar, On the Titchmarsh-Weyl theory of ordinary symmetric odd-order differential expressions and a direct convergence theorem, Quaestiones Math. 5 (1982), 165-185. [27] H. Langer, B. Textorious, On generalized resolvents and Q-functions of symmetric linear relations (subspaces) in Hilbert space, Pacif. J. Math. 72(1977), No.1 , 135-165. [28] M. Lesch, M.M. Malamud, On the deficiency indices and self-adjointness of symmetric Hamiltonian systems, J. Differential Equations 189 (2003), 556 -- 615. [29] M. M. Malamud, On the formula of generalized resolvents of a nondensely defined Hermitian oper- ator, Ukr. Math. Zh. 44(1992), no 12, 1658-1688. [30] M. M. Malamud, V. I. Mogilevskii, Krein type formula for canonical resolvents of dual pairs of linear relations, Methods of Funct. Anal. and Topology 8, No.4 (2002), 72-100. [31] M. Malamud, H. Neidhardt, Sturm-Liouville boundary value problems with operator potentials and unitary equivalence, J. Differential Equations, 252 (2012), 5875-5922. [32] V.I.Mogilevskii, Boundary triplets and Krein type resolvent formula for symmetric operators with unequal defect numbers, Methods Funct. Anal. Topology 12 (2006), no 3, 258 -- 280. [33] V.I.Mogilevskii, Description of spectral functions of differential operators with arbitrary deficiency indices, Math. Notes 81 (2007), no 4, 553 -- 559. [34] V.I.Mogilevskii, Boundary pairs and boundary conditions for general (not necessarily definite) first- order symmetric systems with arbitrary deficiency indices, Math. Nachr.285 (2012), no14 -- 15, 1895 -- 1931 [35] V.I.Mogilevskii, On exit space extensions of symmetric operators with applications to first order symmetric systems, Methods Funct. Anal. Topology 19 (2013), to appear [36] B.C. Orcutt, Canonical differential equations, Dissertation, University of Virginia, 1969. [37] J. Weidmann, Spectral theory of ordinary differential operators, Lecture notes in mathematics, 1258, Springer-Verlag, Berlin, 1987. Department of Mathematical Analysis, Lugans'k National University, Oboronna Str. 2, 91011 Lugans'k, Ukraine E-mail address: [email protected]
1703.01891	1	1703	2017-03-06T14:29:06	Coarse embeddings into $c_0(\Gamma)$	[ "math.FA" ]	Let $\lambda$ be a large enough cardinal number (assuming GCH it suffices to let $\lambda=\aleph_\omega$). If $X$ is a Banach space with $\text{dens}(X)\ge\lambda$, which admits a coarse (or uniform) embedding into any $c_0(\Gamma)$, then $X$ fails to have nontrivial cotype, i.e. $X$ contains $\ell_\infty^n$ $C$-uniformly for every $C>1$. In the special case when $X$ has a symmetric basis, we may even conclude that it is linearly isomorphic with $c_0(\text{dens}X)$.	math.FA	math	Coarse embeddings into c0(Γ) Petr H´ajek∗1 and Th. Schlumprecht† 2 1Institute of Mathematics, Academy of Science of the Czech Republic, Zitn´a 25 115 67 Prague 1, Czech Republic, and Department of Mathematics, Faculty of Electrical Engineering, Czech Technical University in Prague, Zikova 4, 160 00 Prague, Czech Republic. email: [email protected] 2Department of Mathematics, Texas A&M University, College Station, TX 77843, USA and Department of Mathematics, Faculty of Electrical Engineering, Czech Technical University in Prague, Zikova 4, 160 00 Prague, Czech Republic. email: [email protected] July 19, 2018 Abstract Let λ be a large enough cardinal number (assuming GCH it suffices to let λ = ℵω). If X is a Banach space with dens(X) ≥ λ, which admits a coarse (or uniform) embedding into any c0(Γ), then X fails to have nontrivial cotype, i.e. X contains ℓn ∞ C-uniformly for every C > 1. In the special case when X has a symmetric basis, we may even conclude that it is linearly isomorphic with c0(densX). 1 Introduction The classical result of Aharoni states that every separable metric space (in particular every separable Banach space) can be bi-Lipschitz embedded (the definition is given below) into c0. The natural problem of embeddings of metric spaces into c0(Γ), for an arbi- trary set Γ, has been treated by several authors, in particular Pelant and Swift. ∗The first named author was supported by GA CR 16-073785 and RVO: 67985840. †The second named author was supported by the National Science Foundation under the Grant Number DMS -- 1464713. 2000 Mathematics Subject Classification Primary 46B26; Secondary 54E15, 54D20 1 1 INTRODUCTION 2 The characterizations that they obtained, and which play a crucial role in our argument, are described below. Our main interest, motivated by some problems posed in [3], lies in the case of embeddings of Banach spaces into c0(Γ). We now state the main results of this paper. We first define the following cardinal numbers inductively. We put λ0 = ω0, and, assuming for n ∈ N0, λn has been defined, we put λn+1 = 2λn. Then we let λ = lim n→∞ λn. (1) It is clear that assuming the generalized continuum hypothesis (GCH) λ = ℵω. Theorem A. If X is a Banach space with density dens(X) ≥ λ, which admits a coarse (or uniform) embedding into any c0(Γ), then X fails to have nontrivial cotype, i.e. X contains ℓn ∞ C-uniformly for some C > 1 (equivalently, every C > 1). Our method of proof gives a much stronger result for Banach spaces with a symmetric basis. Namely, under the assumptions of Theorem A, such spaces are linearly isomorphic with c0(Γ) (Theorem (4.2)). Theorem A will follow from the following combinatorial result which is of independent interest. Theorem B. Assume that Λ is a set whose cardinality is at least λ, n ∈ N, and σ : [Λ]n → C is a map into an arbitrary set C. Then (at least) one of the following conditions holds: 1. There is a sequence (Fj )∞ j=1 of pairwise disjoint elements of [Λ]n, so that σ(Fi) = σ(Fj ), for all i, j ∈ N. 2. There is an F ∈ [Λ]n−1 so that σ(cid:0)(cid:8)F ∪ {γ} : γ ∈ Λ(cid:9)(cid:1) is infinite. The above Theorem B was previously deduced in [6] from a combinatorial result of Baumgartner, provided Λ is a weakly compact cardinal number (whose existence is not provable in ZFC, as it is inaccessible [5] p. 325, p. 52). The au- thors in [6] pose a question whether assuming that Γ is uncountable is sufficient in Theorem B. Theorem B is used in order to obtain a scattered compact set K of height ω0, such that C(K) does not uniformly embed into c0(Γ). It is easy to check that our version of Theorem B implies a ZFC example of such a C(K) space. It is further shown in [6] that the space C[0, ω1] does not uniformly embed into any c0(Γ). Let us point out that a special case of Theorem A was obtained by Pelant and Rodl [8], namely it was shown there that ℓp(λ), 1 ≤ p < ∞, spaces (which are well known to have nontrivial cotype) do not uniformly embed into any c0(Γ). The paper is organized as follows. In Section 2 we recall Pelant's [7, 6] amd Swift's [10] conditions for Lipschitz, uniform, and coarse embeddability into 2 PELANT'S AND SWIFT'S CRITERIA 3 c0(Γ). In Section 3 we provide a proof for Theorem B. Finally, in Section 4, we provide a proof of Theorem A as well as the symmetric version of the result. All set theoretic concepts and results used in our note can be found in [5], whereas for facts concerning nonseparable Banach spaces [4] can be consulted. 2 Pelant's and Swift's criteria for Lipschitz, uni- form, and coarse embeddability into c0(Γ) In this section we recall some of the notions and results by Pelant [7, 6] and Swift [10] about embeddings into c0(Γ). For a metric space (M, d) a cover is a set U of subsets of M such that M = SU∈U U . A cover U of M is called uniform if there is an r > 0 so that for all x ∈ M there is a U ∈ U, so that Br(x) = {x′ ∈ M : d(x′, x) < r} ⊂ U . It is called uniformly bounded if the diameters of the U ∈ U are uniformly bounded, and it is called point finite if every x ∈ M lies in only finitely many U ∈ U. A cover V of M is a refinement of a cover U, if for every V ∈ V there is a U ∈ U, for which V ⊂ U . Definition 2.1. [6] A metric space (M, d) is said to have the Uniform Stone Property if every uniform cover U of M has a point finite uniform refinement. Definition 2.2. [10] A metric space (M, d) is said to have the Coarse Stone Property if every bounded cover is the refinement of a point finite uniformly bounded cover. Definition 2.3. Let (M1, d1) and (M2, d2) be two metric spaces. For a map f : M1 → M2 we define the modulus of uniform continutiy wf : [0, ∞) → [0, ∞], and the modulus of expansion ρ : [0, ∞) → [0, ∞] as follows wf (t) = sup(cid:8)d2(cid:0)f (x), f (y)(cid:1) : x, y ∈ M1, d1(x, y) ≤ t(cid:9) and ρf (t) = inf (cid:8)d2(cid:0)f (x), f (y)(cid:1) : x, y ∈ M1, d1(x, y) ≥ t(cid:9). The map f is called uniform continuous if limt→0 wf (t) = 0, and it is called a uniform embedding if moreover ρf (t) > 0 for every t > 0. It is called coarse if wf (t) < ∞, for all 0 < t < ∞ and it is called a coarse embedding, if limt→∞ ρf (t) = ∞. The map f is called Lischitz continuous if Lip(f ) = sup x6=y d2(f (x), f (y)) d1(x, y) < ∞, and a bi-Lipschitz embedding, if f is injective and Lip(f −1) is also finite. The following result recalls results from [6](for (i) ⇐⇒ (ii)) and [10] (for (ii) ⇐⇒ (iii) ⇐⇒ (iv) ⇐⇒ (v)). Theorem 2.4. For a Banach space X the following properties are equivalent. (i) X has the uniform Stone Property. 3 A COMBINATORIAL ARGUMENT 4 (ii) X is uniformly embeddable into c0(Γ), for some set Γ. (iii) X has the coarse Stone Property. (iv) X is coarsely embeddable into c0(Γ), for some set Γ. (v) X is bi-Lipschitzly embeddable into c0(Γ), for some set Γ. It is easy to see, and was noted in [6, 10], that the uniform Stone property and the coarse Stone property are inherited by subspaces. The equivalence (i) ⇐⇒ (ii) was used in [6] to show that C[0, ω1] does not uniformly embed in any c0(Γ). It was also used to prove that certain other C(K)-spaces do not uniformly embed into c0(Γ): Let Λ be any set and denote for n ∈ N by [Λ]≤n and [Λ]n the subsets of Λ which have cardinality at most n and exactly n, respectively. Endow [Λ]≤n with the restriction of the product topology on {0, 1}Λ (by identifying each set with its characteristic function). Then define KΛ to be the one-point Alexandroff compactification of the topological sum of the spaces [Λ]≤n, n ∈ N. It was shown in [6] that if Λ satisfies Theorem B then C(KΛ) is not uniformly Stone and thus does not embed uniformly into any c0(Γ). 3 A combinatorial argument We start by introducing property P (α) for a cardinal α as follows. For every n ∈ N and any map σ : [α]n → C, C being an arbitrary set, (P (α)) (at least) one of the following two conditions hold: There is a sequence (Fj) of pairwise disjoint elements of [λ]n, with σ(Fi) = σ(Fj ), (2) for any i, j ∈ N. There is an F ∈ [λ]n−1, so that σ(cid:0)(cid:8)F ∪ {γ} : γ ∈ λ \ F(cid:9)(cid:1) is infinite. As remarked in Section 2, if κ is an uncountable weakly compact cardinal number, then P (κ) holds. But the existence of weakly compact cardinal numbers requires further set theoretic axioms, beyond ZFC [5]. In [6, Question 3] the authors ask if P (ω1) is true. (3) Theorem 3.1. For λ defined by (1), P (λ) holds. For our proof of Theorem 3.1 it will be more convenient to reformulate it into a statement about n-tuples, instead of sets of cardinality n. We will first introduce some notation. Let n ∈ N and Γ1, Γ2, . . . Γn be sets of infinite cardinality, and put Γ = Qn i=1 Γi. For a ∈ Γ and 1 ≤ i ≤ n we denote the i-the coordinate of a by a(i). We say that two points a and b in Γ are diagonal, if a(i) 6= b(i), for all i ∈ {1, 2, . . . , n}. 3 A COMBINATORIAL ARGUMENT 5 Let a ∈ Γi for i ∈ N. For i ∈ {1, 2 . . . , n} we call the set H(a, i) = (cid:8)(b1, b2, . . . , bi−1, a(i), bi+1, . . . , αn) : bj ∈ Γj, for j ∈ {1, 2 . . . n} \ {i}(cid:9), the Hyperplane through the point a orthogonal to i. We call the set L(a, i) = (cid:8)(a(1), . . . , a(i − 1), bi, a(i + 1), . . . a(n)) : bi ∈ Γi(cid:9), the Line through the point a in direction of i. For a cardinal number β, we define recursively the following sequence of car- dinal numbers (cid:0) exp+(β, n) : n ∈ N0(cid:1): exp+(β, 0) = β, and, assuming exp+(β, n) has been defined for some n ∈ N0, we put exp+(β, n + 1) = (cid:0)2exp+(β,n)+(cid:1)+ . Here γ+ denotes the successor cardinal, for a cardinal γ, i.e., the smallest car- dinal number γ′ with γ′ > γ. Note that since exp+(γ, 1) ≤ 222γ , it follows for the above defined cardinal number λ, that λ = lim n→∞ exp+(ω0, n). Secondly, successor cardinals are regular [5], and thus every set of cardinality γ, with γ being a successor cardinal, can be partitioned for n ∈ N into n disjoint sets Γ1, Γ2, . . . , Γn, all of them having also cardinality γ, and the map Γ1 × Γ2 × . . . × Γn → (cid:2)Sn , (a1, a2, . . . , an) 7→ {a1, a2, . . . an}, is injective. We therefore deduce that the following statement implies Theorem 3.1. i=1 Γi(cid:3)n Theorem 3.2. Let n ∈ N and a assume that the sets Γ1, Γ2, . . . Γn have cardi- nality at least exp+(ω1, n2). For any function σ : Γ := n Y i=1 Γi → C, where C is an arbitrary set, at least one of the following two conditions hold There is a sequence (a(j))∞ σ(a(i)) = σ(a(j)), for any i, j ∈ N. There is a line L ⊂ Γ, for which σ(L) is infinite. j=1, of pairwise diagonal elements in Γ, so that (4) (5) We will make the following observation before proving Theorem 3.2. Lemma 3.3. Let n ∈ N and Γ1, Γ2, . . . Γn be non empty sets. Let σ : Γ := n Y i=1 Γi → C, be a function that fails both conditions (4) and (5). 3 A COMBINATORIAL ARGUMENT 6 Then there is a set C and a function σ : Γ := n Y i=1 Γi → C, that fails both (4) and (5) and moreover has the property that for every c ∈ C there is a hyperplane Hc ⊂ Γ so that {b ∈ Γ : σ(b) = c} ⊂ Hc. (6) Proof. We may assume without loss of generality that σ is surjective. Since (4) is not satisfied for each c ∈ C there exists an m(c) ∈ N and a (finite) sequence (a(c,j))m(c) j=1 ⊂ σ−1({c}), which is pairwise diagonal, and maximal, with this property. Hence σ−1({c}) ⊂ m(c) n [ j=1 [ i=1 H(a(c,j), i). Indeed, from the maximality of (a(c,j))m(c) j=1 ⊂ σ−1({c}), it follows that each b ∈ σ−1({b}) must have at least one coordinate in common with at least one element of (a(c,j))m(c) j=1 ⊂ σ−1({c}). We define C = [ c∈C {1, 2, . . . , m(c)} × {1, 2 . . . n} × {c}, and σ : Γ → C, c = σ(b), b 7→ (c, j, i), where n j = minnj′ : b ∈ [ i′=1 H(a(c,j′), i′)o, and i = min{i′ : b ∈ H(a(c,j), i)}. It is clear that σ satisfies (6). Since for every c ∈ C, {b ∈ Γ : σ(b) = c} = m(c) n [ j=1 [ i=1 {b ∈ Γ : σ(b) = (c, j, i)}, σ does not satisfy (4). In order to verify that (5) is not satisfied, assume L ⊂ Γ is a line, and let {c1, c2, . . . , cp} be the image of L under σ. By construction, σ(L) ⊂ {(j, i, ck), k ≤ p, j ≤ m(ck), ı ≤ n}, which is also finite. Proof of Theorem 3.2. We assume that σ : Γ = Γ1 × Γ2 × . . . × Γn → C is a map which fails both (4) and (5). By Lemma 3.3 we may also assume that σ satisfies (6). For each a ∈ Γ we fix an i(a) ∈ {1, 2, . . . , n} so that σ−1(cid:0){σ(a)}(cid:1) ⊂ H(a, i(a)). It is important to note that, since (5) is not satisfied, it follows that 3 A COMBINATORIAL ARGUMENT 7 each line L, whose direction is some j ∈ {1, 2, . . . , n}, can only have finitely many elements b for which i(b) = j. Indeed, if i(b) = j then b is uniquely determined by the value σ(b). To continue with the the proof the following Reduction Lemma will be essential. Lemma 3.4. Let β be an uncountable regular cardinal. Assume that Γ1 ⊂ Γ1, Γ2 ⊂ Γ2, . . . , Γn ⊂ Γn are such that Γi ≥ exp+(β, n), for all i ∈ {1, . . . , n}. Then, for any i ∈ {1, 2 . . . , n} there are a number Ki ∈ N, and subsets 1 ⊂ Γ1, Γ′ Γ′ 2 ⊂ Γ2, . . . , Γ′ n ⊂ Γn, with Γ′ j ≥ β, so that ∀(a1, a2, . . . , ai−1, ai+1, . . . an) ∈ n Y j=1,j6=i Γ′ i (7) Proof. We assume without loss of generality that i = n. Abbreviate βj = exp+(β, j), for j = 1, 2 . . . n. We choose subsets Γ(0) βn+1−j. i : i(a1, a2, . . . ai−1, a, ai+1, . . . an) = i}(cid:12)(cid:12) ≤ Ki. Γ(0) (cid:12)(cid:12) = j j ⊂ Γj, for which (cid:12)(cid:12) (cid:12)(cid:12){a ∈ Γ′ Since the βj's are regular, it follows for each j = 1, 2 . . . , n − 1 that Γ(0) (cid:12)(cid:12) j (cid:12)(cid:12) = βn+1−j > 2βn−j = 2Γ(0) = (cid:12)(cid:12){f : Γ(0) j+1×Γ(0) j+1×...×Γ(0) n j+1 × Γ(0) j+1 × . . . × Γ(0) n → N}(cid:12)(cid:12). Abbreviate for i = 1, . . . , n. Fj = {f : Γ(0) j × Γ(0) j+1 × . . . × Γ(0) n → N} and consider the function φ1 : n−1 Y j=1 Γ(0) j → N, (a1, a2, . . . , an−1) 7→ (cid:12)(cid:12){a ∈ Γ(0) n : i(a1, a2, . . . , an−1, a) = n}(cid:12)(cid:12). For fixed a1 ∈ Γ(0) 1 , φ1(a1, ·) ∈ F2, and the cardinality of F2 is by the above estimates smaller than the cardinality of Γ(0) 1 , which is regular. Therefore we 1 ⊂ Γ(0) can find a function Φ2 ∈ F2 and a subset Γ′ of cardinality βn so that φ1(a1, ) = φ2 for all a1 ∈ Γ′ j ⊂ Γ(0) , for j = 1, 2 . . . , n − 2 of cardinality βn+1−j and functions φj ∈ Fj, for j = 1, 2, . . . , n − 1, so that for all (a1, a2, . . . , an−2) ∈ Qn−2 n−1, we 1. We continue the process and find Γ′ j and an−1 ∈ Γ(0) j=1 Γ′ j i have φ1(a1, a2, . . . , an−1) = φ2(a2, . . . , an−1) = . . . = φn−1(an−1). (8) Then, since φn−1 is N valued, we can finally choose an Kn ∈ N and a subset Γ′ n−1, of cardinality at least β, so that φn−1(a1) ≤ Kn, which finishes our argument. 4 PROOF OF THEOREM A 8 Continuation of the proof of Theorem 3.2. We apply Lemma 3.4 successively to all i ∈ {1, 2 . . . n}, and the cardinals β(i) = exp+(ω1, n(n − i)). We obtain numbers K1, K2, . . . Kn in N and infinite sets Λj ⊂ Γi, for j ≤ n, so that for all i = 1, 2 . . . n and all a ∈ Qn j=1 Λj (cid:12)(cid:12){a ∈ Λi : i(a1, a2, . . . , ai−1, a, ai+1, . . . an) = i}(cid:12)(cid:12) ≤ Ki. In order to deduce a contradiction choose for each j = 1, . . . n a subset Aj of Λj of cardinality lj = (n + 1)Kj. Then it follows that Y j=1 Aj(cid:12)(cid:12)(cid:12) n Y j=1 lj = (cid:12)(cid:12)(cid:12) X = n X i=1 a∈Qn j=1,j6=i Aj (cid:12)(cid:12){a ∈ Ai : i(a1, a2, . . . , ai−1, a, ai+1, . . . an) = i}(cid:12)(cid:12) ≤ n X i=1 Ki n Y j=1,j6=i lj ≤ n n + 1 n Y j=1 lj which is a contradiction and finishes the proof of the Theorem. We can now state the ZFC version of Theorem 4.1. in [6], in which it was shown that for weakly compact cardinalities κ0 the space C(Kκ0 ), where Kκ0 was defined at the end of Section 2, cannot be uniformly (or coarsely) embedded into any c0(Γ), where Γ has any cardinality. Since the only property of κ0, which is needed in [6], is the fact that (P (κ0)) holds, we deduce Corollary 3.5. C(Kλ) does not coarsely (or uniformly) embed into c0(Γ), for any cardinality Γ. 4 Proof of Theorem A In this section we use our combinatorial Theorem B from Section 3 to show Theorem A. γ=0 such that x = PΓ Recall that a long Schauder basis of a Banach space X is a transfinite se- quence {eγ}Γ γ=0 such that for every x ∈ X there exists a unique transfinite se- quence of scalars {aγ}Γ γ=0 aγ eγ. Similarly, a long Schauder basic sequence in a Banach space X is a transfinite sequence {eγ}Γ γ=0 which is a long Schauder basis of its closed linear span. Recall that the w∗−dens(X ∗) is the smallest cardinal such that there exists a w∗-dense subset of X ∗. Analogously to the classical Mazur construction of a Schauder basic sequence in a separable Banach space we have the following result, proved e.g. in [4, p.135] (the fact that the basis is normalized, i.e. keγk = 1, is not a part of the statement in [4], but it is easy to get it by normalizing the existing basis). Theorem 4.1. Let X be a Banach space with Γ = w∗ − densX ∗ > ω0. Then X contains a monotone normalized long Schauder basic sequence of length Γ. 4 PROOF OF THEOREM A 9 Proof of Theorem A. Using the Hahn-Banach theorem it is easy to see that w∗ − densX ∗ ≤ densX. On the other hand, since every x ∈ X is uniquely determined by its values on a w∗-dense subset of X ∗, it is clear that densX ≤ cardX ≤ 2w∗−densX ∗ It follows that for λ defined in (1) we get that λ = w∗ − densX ∗ holds if and only if λ = densX. In order to prove Theorem A we may assume without loss of generality that X has a long normalized and monotone Schauder basis (eµ)µ<λ, of length λ, i.e. Γ = λ. Set Dn = {F ⊂ λ : F = n}, n ∈ N Suppose that F = {γ1, . . . , γn} where γ1 < · · · < γn are elements of [0, λ) arranged in an increasing order. Consider the corresponding finite set MF = {Pn i=1 εieγi : εi ∈ {−1, 1}}, containing 2n distinct vectors of X, and put a linear order ≺ on this set according to the arrangement of the signs εi, setting n X i=1 εieγi ≺ n X i=1 εieγi if and only if for the minimal i, such that εi 6= εi, it holds εi < εi. In order to prove Theorem A it suffices to show that if M = ∪F ∈Dn ,n∈NMF ⊂ X has the coarse Stone property then X fails to have nontrivial cotype. To this end, starting with U = {B2(x) : x ∈ M } we find a uniform bounded cover V, which is point finite and so that U refines V, i.e., for all x ∈ M there is a Vx ∈ V with B2(x) ⊂ Vx. Let r > 0 be such that each V ∈ V is a subset of a ball of radius r. Let C be the set consisting of all finite tuples (V 1, . . . , V m), where V j ∈ V. We now define the function σ : M → C as follows. If F ∈ Dn, F = {γ1, . . . , γn} where γ1 < · · · < γn, we let σ(F ) = (Vy1 , . . . , Vy2n ), (9) where y1 ≺ · · · ≺ y2n are the elements of MF arranged in the increasing order. Applying Theorem B to the function σ, for a fixed n ∈ N, yields one of two possibilities. Either there is an F = {γ1, . . . , γn−1}, where γ1 < · · · < γn−1, so that σ(cid:0)(cid:8)F ∪ {τ } : τ ∈ λ \ F(cid:9)(cid:1) is infinite. In this case, pick an infinite sequence of distinct {τj}∞ j=1 witnessing the desired property. By passing to a subsequence, we may assume without loss of generality that either there exists k, 1 ≤ k ≤ n − 1, so that for all j ∈ N, γk < τj < γk+1, or τj < γ1 for all j ∈ N, or γn−1 < τj for all j ∈ N. For simplicity of notation, assume the last case, i.e. γ1 < · · · < γn−1 < τj holds for all j ∈ N. Denoting F j = {γ1, . . . , γn−1, τj}, we conclude that there exists a fixed selection of signs ε1, . . . , εn such that the set B = nVy : y = n−1 X i=1 εieγi + εnτj , j ∈ No 4 PROOF OF THEOREM A 10 is infinite. Indeed, otherwise the set of values {σ({γ1, . . . , γn−1, τj}), j ∈ N}, which are determined by the definition (9), would have only a finite set of options for each coordinate, and would therefore have to be finite. This is a contradiction with the point finiteness of the system V, because n−1 X i=1 εieγi ∈ Vy, for all Vy ∈ B. It remains to consider the other case when there is a sequence (Fj) of pairwise disjoint elements of [λ]n, with σ(Fi) = σ(Fj ), for any i, j ∈ N. In fact, it suffices to choose just a pair of such disjoint elements (written in an increasing order of ordinals) F = {γ1, . . . , γn}, G = {β1, . . . , βn}, such that σ(F ) = σ(G). This means, in particular, that for every fixed selection of signs ε1, . . . , εn, By our assumption, the elements of V are contained in a ball of radius r, hence VPn i=1 εieγi = VPn i=1 εieβi k n X i=1 εieγi − n X i=1 εieβik ≤ 2r (10) holds for any selection of signs ε1, . . . , εn. Let uj = eγj − eβj , j ∈ {1, . . . , n}. Because {eγ} is a monotonne normalized long Schauder basis, we have the trivial estimate 1 ≤ kujk ≤ 2. The equation (10) means that 1 ≤ k n X i=1 εiuik ≤ 2r (11) holds for any selection of signs ε1, . . . , εn. Since norm functions are convex, this means that for the unite vector ball BE of of E = span(ui : i ≤) it follows that n n X j=1 ajuj : aj ≤ 1 2ro ⊂ BE ⊂ n n X j=1 ajuj : aj ≤ 2o, which means that (uj)n j=1 is 4r-equivalent to the unit vector basis of ℓn ∞. In fact, our proof gives a much stronger condition than just failing cotype, because our copies of ℓk ∞ are formed by vectors of the type eα − eβ. This fact can be used to obtain much stronger structural results for spaces with special bases. Recall that a long Schauder basis {eγ}Λ γ=1 is said to be symmetric if k n X i=1 aieγik = k n X i=1 aieβik for any selection of ai ∈ R, and any pair of sets {γi}n [1, Λ). It is well-known (c.f. automatically unconditional, i.e. there exists K > 0 such that i=1 ⊂ [9, Prop. II.22.2]), that each symmetric basis is i=1 ⊂ [1, Λ), {βi}n 1 K k n X i=1 aieγik ≤ k n X i=1 aieγik ≤ Kk n X i=1 aieγik. 5 FINAL COMMENTS AND OPEN PROBLEMS 11 In particular, whenever A ⊂ B. 1 K kX i∈A aieγik ≤ kX i∈B aieγik Theorem 4.2. Let X be a Banach space of density Λ ≥ λ, with a symmetric basis {eγ}Λ γ=1, which coarsely (or uniformly) embeds into some c0(Γ). Then X is linearly isomorphic with c0(Λ). Proof. By the proof of the above results, if X embeds into c0(Γ), there exists an C > 0, such that for each k ∈ N there are some vectors {vi}k i=1 of the form vi = eγi − eβi satisfying the conditions 1 2 max j ai ≤ (cid:13)(cid:13)(cid:13) k X i=1 aivi(cid:13)(cid:13)(cid:13) ≤ C max j aj. (12) Using the fact that the basis {eγ} is unconditional, (and symmetric) we obtain by an easy manipulation that there exist some constants A, B > 0 such that k X i=1 A(cid:13)(cid:13)(cid:13) aieγi(cid:13)(cid:13)(cid:13) ≤ (cid:13)(cid:13)(cid:13) k X i=1 aivi(cid:13)(cid:13)(cid:13) ≤ B(cid:13)(cid:13)(cid:13) k X i=1 aieγi(cid:13)(cid:13)(cid:13) (13) Combining (12) and (13) we finally obtain that for some D ≥ 1, and any k ∈ N, 1 D max j ai ≤ k k X i=1 aieγik ≤ D max j aj. (14) for all {γ1, . . . , γk} ⊂ [1, Λ), which proves our claim. 5 Final comments and open problems Let us mention in this final section some problems of interest. First of all, we do not know whether or not Theorem A is true if we replace λ by smaller cardinal numbers. Problem 1. Assume that X is a Banach space with dens(X) ≥ ω1, and assume that X coarsely embeds into c0(Γ) for some cardinal number Γ. Does X have trivial co-type? If moreover X has a symmetric basis, must it be isomorphic to c0(ω1)? Of course Problem 1 would have a positive answer if the following is true. Problem 2. Is Theorem B true for ω1? Connected to Problems 1 and 2 is the following REFERENCES 12 Problem 3. Does ℓ∞ coarsely embed into c0(κ) for some uncountable cardinal number κ. Another line of interesting problems asks which isomorphic properties do non separable Banach spaces have which coarsely embed into c0(Γ) Problem 4. Does a non separable Banach space which coarsely embeds into some c0(Γ), Γ being uncountable, contain copies c0, or even c0(ω1). References [1] J. E. Baumgartner, Canonical partition relations, J. Symbolic Logic 40 (1975), 541 -- 554. [2] J. Diestel, H. Jarchow, and A. Tonge, Absolutely Summing Operators, Cambridge University Press (1993). [3] G. Godefroy, G. Lancien, and V. Zizler, The non-linear geometry of Banach spaces after Nigel Kalton, Rocky Mtn. J. Math. 44 (2014), 1529 -- 1583. [4] P. H´ajek, V. Montesinos, J. Vanderwerff, and V. Zizler, Biorthogonal Systems in Banach Spaces, CMS Books in Mathematics, Springer 2008. [5] Th. Jech, Set Theory, Springer Monographs in Mathematics, fourth edition, Springer Verlag, Berlin, Heidelberg, New York, 2006. [6] J. Pelant, P. Holick´y, and O. Kalenda, C(K) spaces which cannot uni- formly embedded into c0(Λ). Fund.Math. 192 (2006), 245 -- 254. [7] J. Pelant, Embeddings into c0, Topol. Appl. 57 (1994), 259 -- 269. [8] J. Pelant, and V. Rodl, On coverings of infinite-dimensional metric spaces. Discrete .Math. 108 (1992), 75 -- 81. [9] I. Singer, Bases in Banach spaces I, Grundlehren Math. Wiss. 154, Springer-Verlag, Berlin (1970). [10] A. Swift, On coarse Lipschitz embeddability into c0(κ), preprint, arX- ive:1611.04623v1.
1204.0934	2	1204	2012-05-07T11:52:54	Coherent states quantization of generalized bergman spaces on the unit ball of cn with a new formula for their associated berezin transforms	[ "math.FA", "math-ph", "math-ph" ]	While dealing with a class of generalized Bergman spaces on the unit ball, we construct for each of these spaces a set of coherent states to apply a coherent states quantization method. This provides us with another way to recover the Berezin transforms attached to these spaces. Finally, a new formula representing these transforms a functions of the Laplace-Beltrami operator is established in terms ofWilson polynomials by using the Fourier-Helgason transform.	math.FA	math	COHERENT STATES QUANTIZATION OF GENERALIZED BERGMAN SPACES ON THE UNIT BALL OF Cn WITH A NEW FORMULA FOR THEIR ASSOCIATED BEREZIN TRANSFORMS ABDELHAMID BOUSSEJRA & ZOUHAIR MOUAYN ABSTRACT. While dealing with a class of generalized Bergman spaces on the unit ball, we construct for each of these spaces a set of coherent states to apply a coherent states quantization method. This provides us with another way to recover the Berezin transforms attached to these spaces. Finally, a new formula representing these transforms a functions of the Laplace-Beltrami operator is estab- lished in terms of Wilson polynomials by using the Fourier-Helgason transform. 1 INTRODUCTION The Berezin transform introduced in [3] for certain classical bounded symmetric domains in Cn is a transform linking the Berezin symbols and symbols for Toeplitz operators. It is present in the study of the correspondence principle. The formula representing the Berezin transform as a function of the Laplace operators ∆1, ..., ∆r ( r being the rank of the domain) plays a key role in the Berezin quantization [4]. In this paper, we deal with the rank one symmetric domains. Namely the unit ball Bn in (Cn,(cid:104),(cid:105)) endowed with its Bergman metric. We are precisely concerned with the L2-eigenspaces (1.1) ϕ ∈ L2(Bn, (1 − ξ2)−n−1dµ), Hν ϕ = ν,n m ϕ A2,ν m (Bn) = (cid:110) (cid:111) associated to the discrete spectrum m = 4ν(2m + n) − 4m(m + n), m = 0, 1, 2, ..., [ν − n/2] ν,n of the Schrödinger operator with uniform magnetic field on Bn given by (cid:32) n∑ (cid:0)δij−zizj (cid:1) (1.2) (cid:33) Hν = −4(1 − z2) (1.3) provided that ν > n/2. Above [x] denotes the greatest integer not exceeding x. For m ∈ Z+, the Berezin transform associated with the space in (1.1) was obtained in [13] via the well known formalism of Toeplitz operators as ∂zi∂zj + ν i,j=1 j=1 (zj ) + ν2 + 4ν2 ∂2 n∑ ∂ ∂zj − zj ∂ ∂zj Bν,n m [ϕ] (z) = m! (2ν − 2m − n) Γ (2ν − m) Γ(n) n!Γ (2ν − m − n + 1) Γ(n + m) P(n−1,2(ν−m)−n) 1 − 2ξ2(cid:17)(cid:17)2 ×(cid:16) (cid:16) m (cid:33)2(ν−m) (cid:90) (cid:16) B (1 − z2 (1 − ξ2) (cid:32) 1 − ξ2(cid:17)n+1 dµ (ξ) 1 − (cid:104)z, ξ(cid:105)2 ϕ (ξ) (1.4) m where P(α,β) (.) denotes the Jacobi polynomial [16]. Moreover this transform have been expressed as a function f (∆Bn ) of the Laplace-Beltrami operator ∆Bn in terms of an 3(cid:122)2-sum, see (5.24) be- low. Our aim here is to construct for each of the eigenspaces in (1.1) a set of coherent states by following a generalized formalism [11] in order to apply a coherent states quantization method. Date: May 4, 2021. 1 2 A. BOUSSEJRA, Z. MOUAYN This provides us with another way to recover the Berezin transforms in (1.4) attached to the L2- eigenspace spaces in (1.1). Finally, we add a new formula expressing the transform (1.4) as a function of the Laplace-Beltrami operator. The idea is to make the integral (1.4) appear as "con- volution product" of the function ϕ with a specific radial function given in terms of the square of a Jacobi polynomial. Next, a straightforward computation of the spherical transform of this radial function with the use of a Clebsh-Gordon type linearisation [8] for the square of a Jacobi polynomial amounts to a finite sum containing some integrals whose general form was given by Koornwinder [17] in terms of Wilson polynomials. This paper is summarized as follows. In Section 2, we recall briefly the formalism of coherent states quantization we will be using. Section 3 deals with some needed facts on the generalized Bergman spaces. In Section 4, we construct for each of these spaces a set of coherent states and we apply the corresponding quantization scheme in order to recover their associated Berezin trans- forms. In Section 5, we present the formula expressing these Berezin transforms as functions of the Laplace-Beltrami operator by a different way and in a new form. 2 COHERENT STATES QUANTIZATION Coherent states are mathematical tools which provide a close connection between classical and quantum formalism. In general, they are a specific overcomplete set of vectors in a Hilbert space satisfying a certain resolution of the identity condition. Here, we review a coherent states for- malism starting from a measure space "as a set of data" as presented in [11]. Let X = {x x ∈ X} (cid:9)+∞ be a set equipped with a measure dµ and L2(X, dµ) the space of dµ−square integrable functions j=0. Let H be another j=0. Then consider the family of states on X. Let A2 be a subspace of L2(X, dµ) with an orthonormal basis(cid:8)Φj (functional) space with a given orthonormal basis (cid:8)φj (cid:9)+∞ { x >}x∈X in H, through the following linear superposition: where x >:= (N (x)) − 1 2 Φj (x) φj >, +∞ ∑ j=0 N (x) = +∞ ∑ j=0 Φj (x) Φj (x). These coherent states obey the normalization condition (cid:104)x x(cid:105)H = 1 and the following resolution of the identity of H 1H = x >< x N (x) dµ (x) (2.1) (2.2) (2.3) (2.4) (2.5) (cid:90) X (cid:90) which is expressed in terms of Dirac's bra-ket notation x >< x meaning the rank-one -operator ϕ (cid:55)→ (cid:104)ϕ x(cid:105)H . x > . The choice of the Hilbert space H define in fact a quantization of the space X by the coherent states in (2.1), via the inclusion map x (cid:55)→ x >∈ H and the property (2.4) is cru- cial in setting the bridge between the classical and the quantum mechanics. The Klauder-Berezin coherent states quantization consists in associating to a classical observable that is a function f (x) on X having specific properties the operator-valued integral A f := x >< x f (x) N (x) dµ (x) The function f (x) ≡ (cid:98)A f (x) is called upper (or contravariant) symbol of the operator A f and is nonunique in general. On the other hand, the expectation value(cid:10)x A f x(cid:11) of A f with respect X to the set of coherent states { x >}x∈X is called lower ( or covariant) symbol of A f . Finally, asso- ciating to the classical observable f (x) the obtained mean value(cid:10)x A f x(cid:11) , we get the Berezin 3 transform of this observable. That is, B [ f ] (x) :=(cid:10)x A f x(cid:11) , x ∈ X. (2.6) (cid:18)a (cid:19) For all aspect of the theory of coherent states and their genesis, we refer to the survey [9] by Dodonov or to the book by Gazeau [11]. 3 THE SPACES A2,ν m (Bn) In this section, we review some results on the L2-concrete spectral analysis of the Schrödinger operator Hν in (1.3) and acting in the Hilbert space L2(Bn, dµn),see [7], for more details. Let Bn = {z ∈ Cn; z < 1} be the unit ball in Cn with the Lebesgue measure dµ normalized so that µ(Bn) and let ∂Bn = {ω ∈ Cn, ω = 1} be the unit sphere with dσ the normalized measure on it. Let G = SU(n, 1) be the group of all C-linear transforms g on Cn+1 that preserve the indefinite j=1 zj 2 − zn+1 2, with det g = 1. Then G acts transitively on the unit ball by hermitian form ∑n G (cid:51) g = (3.1) As a homogeneous space we have the identification Bn = G/K where K = S(U(n) × U(1)) is i,j ∂j∂j(Log(1 − the stabilizer of 0. It is endowed with its usual Khaler-Bergman metric ds2 = − ∑n z2))dzi ⊗ dzj. The Bergman distance and the volume element on Bn are given respectively by : z → g.z = (az + b)(cz + d)−1. b c d cosh2 d (z, w) = 1 − (cid:104)z, w(cid:105)2 (1− z 2)(1− w 2) (3.2) and dµn (z) = (1− z 2)−(n+1)dµ (z). The group G acts unitarily on the space L2(Bn, dµn), via U(g)F(z) = F(g−1.z). Let consider the magnetic gauge vector potential given through the canonical 1-form on Bn: θ = −i(∂− ∂)Log(1− z 2), to which the Schrodinger operator Hν = − (d + iνext (θ)) ∗ (3.3) can be associated. Here ν ≥ 0 is a fixed number d denotes the usual exterior derivative on differential forms on Bn and ext(θ) is the exterior multiplication by θ while the symbol ∗ stands for the adjoint operation with respect to the Hermitian scalar product induced by the Bergman metric ds2 on differential forms. Note that when ν = 0, the operator in (3.3) reduces to (d + iext (θ)) + 4ν2 which is the Laplace-Beltrami operator of the Bergman ball Bn. For general ν ≥ 0, the Schrodinger operator Hν in (3.3) can be expressed in the complex coordinates (z1, ..., zn) by the formula (1.3) see [2],[6] and [12]. Now, for an arbitrary complex number λ, a fundamental family of eigenfunctions of Hν with eigenvalue λ2 + 4ν2 + n2 is given by the Poisson kernels : (cid:18) 2 (iλ+1)(cid:32) (cid:19) 1 (cid:33)ν z (cid:55)→ Pν λ(z, θ) = 1− z 2 1− < z, θ >2 1 − < z, θ > 1− < z, θ > , z ∈ Bn. (3.5) Moreover, a complete description of the expansion of an eigenfunction f of Hν with eigenvalue λ2 + 4ν2 + n2, in terms of the appropriate Fourier series in Bn have been given in [7, Proposition H0 ≡ ∆Bn = 4(1− z 2) ∂2 ∂zi∂zj (3.4) (cid:0)δij − zizj (cid:1) n∑ i,j=1 A. BOUSSEJRA, Z. MOUAYN 4 2.2]. Precisely, f (z) = (1 − ρ2) iλ+n 2 (cid:18) iλ + n 2 +∞ ∑ p,q=0 ρp+q ·2 (cid:122)1 + ν + p, iλ + n 2 − ν + q, p + q + n; ρ2 (cid:19) aλ,ν p,q.hp,q(θ), (3.6) in C∞([0, 1[×∂Bn), z = ρθ, ρ ∈ [0, 1[ and θ = 1. Above 2F1 denotes the Gauss hypergeometric function [14] and aλ,ν p,q,j) ∈ Cd(n,p,q) are complex numbers, where p,q = (aλ,ν d(n, p, q) := (p + q + n − 1)(p + n − 2)!(q + n − 2)! p!q!(n − 1)!(n − 2)! (3.7) is the dimension of the space H(p, q) of restrictions to the unit sphere ∂Bn of harmonic polynomi- als h(z) on Cn, which are homogeneous of degree p in z and degree q in z, see [10] or [19] for more details. The notation "." in (3.6) means the following finite sum aλ,ν p,q.hp,q(θ) = d(n,p,q)∑ j=1 p,q,jhj aλ,ν p,q(θ), (3.8) where {hj p,q}1≤j≤d(n,p,q) is an orthonormal basis of H(p, q). The spectral analysis of Hν have been studied by many authors, see [7] and references therein. Actually, Hν is an elliptic densely de- fined operator on the Hilbert space L2(Bn, (1 − (cid:104)z, z(cid:105))−(n+1)dµ) admitting a unique self-adjoint realization also denoted by Hν. Its spectrum consists of a continuous part given by(cid:2)n2, +∞(cid:2) (cor- responding to scattering states) and a finite number of infinitely degenerate eigenvalues ν,n m given by (1.2) (characterizing bound states) provided that 2ν > n. More precisely, ν,n m + 4ν2 + n2, with λm = i(2m + n − 2ν), m = 0, 1, ..., [ν − n/2]. Here, we focus on the discrete part of the spec- trum, which is labeled by the integer m and the corresponding eigenspace A2,ν m (Bn) defined in (1.1). Taking into account (3.6) and expressing the involved hypergeometric in terms of Jacobi polynomial, an orthonormal basis of A2,ν m (Bn) can be given explicitly by m = λ2 (cid:16) 1 − z2(cid:17)ν−m (cid:16) (cid:18) nΓ(2ν − m − n − q + 1)Γ(p + n + m) P(n+p+q−1,2(ν−m)−n) m−q 1 − 2z2(cid:17) (cid:19)− 1 (m − q)!(2(ν − m) − n)Γ(2ν − m + p) . 2 hj p,q (z, z) (3.9) (3.10) with Φν,m,j p,q (z) = κν,m,n p,q κν,m,n p,q = for varying p = 0, 1, 2, ..., q = 0, 1, ...., m and j = 1, ...., d(n; p, q). Furthermore, the space A2,ν m (Bn) is a reproducing kernel Hilbert space. That is, there exists a unique complex valued function Kν,m on Bn × Bn such that, denoting Kν,m m (Bn) for any z ∈ Bn and for all functions f in A2,ν the Bergman geodesic distance as m (Bn) and all z ∈ Bn.Its expression can be given explicitly as function of (w) = Kν,m(w, z), Kν,m belongs to A2,ν f (z) =< f , Kν,m z >, z z (cid:18) (1 − < z, w >) 1− < z, w > (cid:19)ν (3.11) (1 − 2 tanh2 d(z, w)) Kν,m(z, w) = (2 (ν − m) − n) Γ (2ν − m) n!Γ (2ν − m − n + 1) Remark 3.1. For m = 0, the space A2,ν Bergman space of holomorphic function ψ on Bn satisfying the growth condition 0 (Bn) reduces further to be isomorphic to the weighted ×(cosh d(z, w))−2(ν−m))P(n−1,2(ν−m)−n) (cid:90) ψ (z)2 ((1 − (cid:104)z, z(cid:105))2ν−n−1dµ (z) < +∞. m Bn This fact justify why the eigenspace A2,ν index m. m (Bn) have been also called a generalized Bergman spaces of 4 COHERENT STATES QUANTIZATION 5 Now, to adapt the defintion (2.1) of coherent states for the context of the generalized Bergman spaces in (1.1) we first list the following notations. 1 − z2(cid:17)−(n+1) (cid:16) (cid:111) m (Bn) ⊂ L2(Bn, is the orthonormal basis of A2,ν (cid:19) 1 − z2(cid:17)−n−1 dµ). dµ Bn, (cid:18) • (X, dη) := • x ≡ z ∈ Bn. • A2 := A2,ν (cid:16) • {Φk (x)} ≡(cid:110)Φν,m • {ϕk} ≡(cid:8)ϕp,q,j • N (x) ≡ N (z) is a normalization factor. (cid:9) is an orthonormal basis of another (functional) Hilbert space H. m (Bn) in (3.8) p,q,j (z) , dη ≡ dµn is the volume element on Bn. Definition 4.1. For each fixed integer m = 0, 1, ..., [n − ν/2] , a class of generalized coherent states asso- ciated with the space A2,ν m (Bn) is defined according to (2.1) by the form z ≡ z, ν, m >:= (N (z)) φν,m ∑ − 1 2 Φν,m p,q,j (z) ϕp,q,j 0≤q≤m,0≤p<+∞ 1≤j≤d(n,p,q) where N (z) is a normalization factor. Proposition 4.1. The factor in (4.1) is given by N (z) = for every z ∈ Bn. (2(ν − m) − n) Γ (2ν − m) n!Γ (2ν − m − n + 1) Γ (m + n) m!Γ (n) Proof. To calculate this factor, we start by writing the condition (cid:104)φν,m z , φν,m z (cid:105)H = 1. Equation (4.3) is equivalent to (N (z)) −1 +∞ ∑ p=0 m∑ q=0 d(n,p,q)∑ j=1 Φν,m p,q,j (z) Φν,m p,q,j (z) = 1 Making use of (3.9) and (3.11) for the particular case z = w, we get that P(n−1,2(ν−m)−n) (2(ν − m) − n) Γ (2ν − m) n!Γ (2ν − m − n + 1) Next, by the following fact on Jacobi polynomial [14]: N (z) = m (1) (4.1) (4.2) (4.3) (4.4) (4.5) P(α,β) m (1) = Γ (m + α + 1) m!Γ (α + 1) (4.6) z ≡ z, ν, m > satisfy the resolution for α = n − 1 to arrive at the announced result.The states φν,m of the identity (cid:90) 1H = z, ν, m >< z, ν, m N (z) dν (4.7) and with the help of them we can achieve the coherent states quantization scheme described in Sec.2 to rederive the Berezin transform Bν,n m in (1.4) which was defined by Toeplitz operators Bn 6 A. BOUSSEJRA, Z. MOUAYN formalism in [13]. For this let us associate to any arbitrary function ϕ ∈ L2(Bn, (1 − ξ2)−n−1dµ) the operator-valued integral Aϕ := z, ν, m >< z, ν, m ϕ (z) N (z) (1 − z2)−n−1dµ (4.8) (cid:90) Bn The function ϕ (z) is a upper symbol of the operator Aϕ. On the other hand, we need to calculate the expectation value E{z,ν,m>}(cid:0)Aϕ (cid:1) :=< z, ν, m Aϕ z, ν, m > (4.9) of Aϕ with respect to the set of coherent states { z, ν, m >}z∈Bn defined in (4.1). This will consti- tute a lower symbol of the operator Aϕ. Proposition 4.2. Let ϕ ∈ L2(Bn, (1 − ξ2)−n−1dµ). Then, the expectation value in (4.9) has the follow- ing expression Γ (n) m! (2 (ν − m) − n) Γ (2ν − m) n!Γ (n + m) Γ (2ν − m − n + 1) (cid:16) P(n−1,2(ν−m)−n) 1 − 2ξ2(cid:17)(cid:17)2 ×(cid:16) (cid:16) m (cid:33)2(ν−m) (4.10) (1 − z2 (1 − ξ2) (cid:32) (cid:90) 1 − ξ2(cid:17)n+1 dµ (ξ) 1 − (cid:104)z, ξ(cid:105)2 ϕ (ξ) B E{z,ν,m>}(cid:0)Aϕ (cid:1) = for every z ∈ Bn. Proof. We first write the action of the operator Aϕ in (4.8) on an arbitrary coherent state z, ν, m > in terms of Dirac's bra-ket notation as Aϕ z, ν, m >= (cid:90) Bn Therefore, the expectation value reads < z, ν, m Aϕ z, ν, m > = < z, ν, m w, ν, m > < z, ν, m w, ν, m > w, ν, m >< w, ν, m z, ν, m > (cid:90) (cid:90) < z, ν, m w, ν, m >2 Bn ϕ (w) = Bn N (w) (1 − w2)n+1 dµ (w) (4.11) N (w) (1 − w2)n+1 dµ (w) N (w) (1 − w2)n+1 dµ (w) . (4.12) (4.13) Now, we need to evaluate the quantity < z, ν, m w, ν, m >2 in (4.13). For this, we write the scalar product as < z, ν, m w, ν, m >= +∞ ∑ p=0 m∑ q=0 d(n;p,q)∑ m∑ +∞ ∑ r=0 s=0 j=1 (cid:10)ϕp,q,j, ϕp,q,l l=1 d(n;p,q)∑ (cid:11) Recalling that since(cid:8)ϕp,q,j (cid:9) is an orthonormal basis of H, the above sum in (4.14) reduces to H = δj,lδp,rδq,s Φν,m p,q,j (z) Φν,m (cid:112)N (z) N (w) p,q,j (w) (cid:10)ϕp,q,j, ϕp,q,l (cid:11) H (4.14) (4.15) < z, ν, m w, ν, m >= (N (z) N (w)) − 1 2 ∑ 0≤q≤m,0≤p<+∞ 1≤j≤d(n,p,q) p,q,j (z) Φν,m Φν,m p,q,j (w). (4.16) (cid:32) − 1 2 1 − (cid:104)z, w(cid:105) 1 − (cid:104)z, w(cid:105) (cid:33)ν 1 − 2 tanh2 (d (z, w)) 7 (4.17) (cid:17) . < z, ν, m w, ν, m > = (2 (ν − m) − n) Γ (2ν − m) Now, taking account (3.9) and (3.11), Equation ( 4.16) takes the form (N (z) N (w)) (cid:16) −2(ν−m) P(n−1,2(ν−m)−n) So that the square modulus of the scalar product in (4.17) reads n!Γ (2ν − m − n + 1) × (cosh (d (z, w))) m < z, ν, m w, ν, m >2 = n!Γ (2ν − m − n + 1) (cid:18) (2 (ν − m) − n) Γ (2ν − m) −4(ν−m)(cid:16) (cid:18) (2 [ν − m] − n) Γ (2ν − m) (cid:19)2 P(n−1,2(ν−m)−n) (cid:19)2 × (cosh (d (z, w))) m (cid:16) (N (z) N (w)) −1 (4.18) (cid:17)(cid:17)2 . (4.19) 1 − 2 tanh2 (d (z, w)) −1 (N (z) N (w)) (cid:16) 1 − 2 tanh2 (d (z, w)) (1 − w2)n+1 N (w) (cid:17)(cid:17)2 dµ (w) , Returning back to (4.12) , we get E{z,ν,m>}(cid:0)Aϕ (cid:1) = (cid:90) Bn ϕ (w) n!Γ (2ν − m − n + 1) −4(ν−m)(cid:16) × (cosh (d (z, w))) (cid:90) (cid:18) (2 (ν − m) − n) Γ (2ν − m) P(n−1,2(ν−m)−n) (cid:19)2 m which can be also written as E{z,ν,m>}(cid:0)Aϕ (cid:1) = Bn ϕ (w) n!Γ (2ν − m − n + 1) −4(ν−m)(cid:16) −4(ν−m)(cid:16) m × (cosh (d (z, w))) (2(ν − m) − n) Γ (2ν − m) m!Γ (n) n!Γ (2ν − m − n + 1) Γ (m + n) × (cosh (d (z, w))) (cid:90) Bn = −1 (N (z)) (cid:16) (1 − w2)n+1 ϕ (w) (1 − w2)n+1 (cid:16) P(n−1,2(ν−m)−n) 1 − 2 tanh2 (d (z, w)) (4.20) dµ (w) (4.21) (cid:17)(cid:17)2 (cid:17)(cid:17)2 1 − 2 tanh2 (d (z, w)) Finally, we summarize the above discussion by considering the following definition. Definition 4.3. The Berezin transform of the classical observable ϕ ∈ L2(Bn, (1 − ξ2)−n−1dµ) con- structed according to the quantization by the coherent states { z, ν, m >} in (4.1) is obtained by associat- ing to ϕ the obtained mean value in (4.10) . That is, P(n−1,2(ν−m)−n) dµ (w) . m m [ϕ] (z) = E{z,ν,m>}(cid:0)Aϕ (cid:1) Bν,n for every z ∈ Bn. Remark 4.4. For m = 0, the transform (4.10) reduces to the well known Berezin transform attached to the weighted Bergman space A2,ν 0 (Bn) of holomorphic function ψ on Bn satisfying the growth condition (3.12) and given by Bν,n 0 [ϕ] (z) = (2ν − n) Γ (2ν) n!Γ (2ν − n + 1) (cosh d (z, ξ)) −4ν The latter one have also been written as a function of the Bergman Laplacian ∆Bn as ϕ (ξ) 1 − ξ2(cid:17)n+1 dµ (ξ) (cid:16) (cid:19)(cid:12)(cid:12)(cid:12)(cid:12)2 (cid:112)−∆Bn − n2 (4.22) (4.23) (4.24) (cid:90) Bn (cid:18) (cid:12)(cid:12)(cid:12)(cid:12)Γ Bν,n 0 = 1 Γ (α + 1) Γ (α + n + 1) α + 1 + n 2 + i 2 firstly by Berezin. The above form, involving gamma factors, was derived by Peetre in [18, p. 182], so that α there occurring in the weight of the Bergman space, corresponds to 2ν − n − 1. 8 A. BOUSSEJRA, Z. MOUAYN Then Berezin transform Bν,n Bν,n m [ϕ] (z) = cν,n m with Bn cν,n m = where Wk(.) are Wilson polynomials, γν,n,m k = and the coefficients Aν,n,m k 2m!Γ (n) (2 (ν − m) − n) Γ (2ν − m) (−1)k Γ (n + m) Γ (2ν − m − n + 1) k!Γ2 (2 (ν − m) + k) are given by (5.10) below. × Aν,n,m k , Proof. Recall that Bν,n where m [ϕ] (z) = cν,n (cid:16) By this way, we have to compute the spherical transform F [hν,n 1 − ξ2(cid:17)2(ν−m)(cid:16) (cid:90) m (ϕ ∗ hν,n m )(g), P(n−1,2(ν−m)−n) hν,n m (ξ) := (cid:16) m z = g.0, 1 − 2ξ2(cid:17)(cid:17)2 m ] of hν,n m (z)φ−λ(z)dµn(z), λ ∈ R hν,n F [hν,n m ](λ) := , ξ ∈ Bn, m . Namely where φλ is the spherical function associated to ∆Bn, given by Bn φλ(z) = (1− z ) iλ+n 2 2F1( Using Pfaff's transformation [14] iλ + n , iλ + n 2 2 (cid:18) 2(cid:122)1 (a, b, c; x) = (1 − x) −b 2(cid:122)1 b, c − a, c; we rewrite φ−λ as φ−λ(z) = 2F1 (cid:18)−iλ + n 2 , iλ + n 2 , n; z 2 z 2 −1 , n; z 2). (cid:19) x x − 1 (cid:19) 5 AN EXPRESSION OF Bν,n m AS FUNCTION OF ∆Bn m associated the generalized Bergman space A2,ν P(n−1,2(ν−m)−n) 1 − 2 tanh2 d (z, ξ) m (cid:17)(cid:17)2 (cid:16) (cid:16) (cid:90) (cosh d (z, ξ))4(ν−m) m (Bn) is given by (cid:16) dµ (ξ) 1 − ξ2(cid:17)n+1 , ϕ (ξ) Γ (n) m! (2 (ν − m) − n) Γ (2ν − m) n!Γ (n + m) Γ (2ν − m − n + 1) m (z, w) be the kernel function of the above integral operator and set hν,n Let Bν,n z = g.0. Then the integral operator (5.1) can be written as a convolution product over G: m (g) = Bν,n m (z, 0), Bν,n m [ϕ] (z) = cν,n m (ϕ ∗ hν,n m )(g), z = g.0, from which it follows easily that the Berezin operator is an L2-bounded operator. Since Bν,n U(g) ◦ Bν,n function of the G-invariant Laplacian ∆Bn of the unit ball. Below we give it explicitly. m (z, w) is a G bi-invariant function it follows that Bν,n m ◦ U(g), for every g ∈ G. Therefore Bν,n m is a G-invariant operator. That is m is, in the spectral theoretic sense, a m = Bν,n Theorem 5.1.The Berezin transform Bν,n ∆Bn as m can be expressed as a function of the Laplace-Beltrami operator (cid:18) (cid:12)(cid:12)(cid:12)(cid:12)Γ 2m∑ k=0 Bν,n m = + i 2 2 (ν − m) − n 2 ∆Bn − n2 4 k Wk(−1 γν,m,n 4 (cid:112)−∆Bn − n2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12)2 ; 2(ν − m) − n 2 , 1 + n 2 , n 2 , n 2 ) (5.1) (5.2) (5.3) (5.4) (5.5) (5.6) (5.7) So that returning back to (5.5) we get F [hν,n m ](λ) = 2n (cid:16) 1(cid:90) 0 ρ2n−1 (1 − ρ2)n+1−2(ν−m) n − iλ (cid:18) n + iλ , 2 2 ×2(cid:122)1 P(n−1,2(ν−m)−n) m (cid:19) , n; ρ2 ρ2 − 1 dρ (cid:0)1 − 2ρ2(cid:1)(cid:17)2 9 (5.8) To calculate this last integral, we first use a linearisation of the square of Jacobi polynomial in (5.8) by making appeal to the following Clebsh-Gordon type formula see [8, p. 611], P(κ,) s (u) P(τ,η) (5.9) for the particular case of parameters s = l = m, κ = τ = α = n − 1, = η = 2 (ν − m) − n and δ = 2 (ν − m) . In our setting, the linearisation coefficients As,l (k) are of the form (u) = (u) k=0 k l As,l (k) P(α,δ) s+l∑ Aν,n,m k = 2:1 ×(cid:122)2:2 −2m,−2m − n + 1 : 1 − 2ν, 1 − 2ν (2 (ν − m) + n)k (n)2m (2k + 2 (ν − m) + n) (−1)k (2m)! ((2 (ν − m))2m)2 (cid:19) (cid:18) −2m + k,−2ν − k − n : −m,−n − m + 1;−m,−m − n + 1 (n)k (2 (ν − m) + n)2m+k+1 (m!)2 (2m − k)! ((2 (ν − m))m)2 (cid:2)cp(cid:48)(cid:3) (cid:18) (cid:0)ap (cid:1) :(cid:0)bp(cid:48)(cid:1) ,(cid:0)cp(cid:48)(cid:1) (cid:1) (cid:0)aj (5.11) (dl) : (κl(cid:48) ) , (l(cid:48) ) s in which (x)s = x (x + 1) ... (x + s − 1) is the Pochhammer symbol. l:l(cid:48) (.) denotes the Kampé de Fériet double hypergeometric function defined by [20, p. 63] [dl]q+s [κl(cid:48) ]q [l(cid:48) ]s (cid:2)bp(cid:48)(cid:3) +∞ ∑ q,s=0 (cid:2)ap (cid:122)p:p(cid:48) l:l(cid:48) x, y 1, 1 (5.10) (cid:19) xq q! ys s! (cid:3) q+s = q s Here (cid:122)p:p(cid:48) (cid:3) where (cid:2)ap Therefore, inserting(cid:16) s = ∏p j=1 (cid:0)1 − 2ρ2(cid:1)(cid:17)2 P(n−1,2(ν−m)−n) m = 2m∑ k=0 Aν,n,m k P(n−1,2(ν−m)) k (cid:0)1 − 2ρ2(cid:1) into (5.8) the Fourier-Helgason transform of hν,n m takes the form 2m∑ Aν,n,m m ](λ) = Iν,n,m k k (λ) , k=0 where the last term in this sum is defined by the integral F [hν,n 1(cid:90) Iν,m k (λ) = 2nρ2n−1 (1 − ρ2)n+1−2(ν−m) P(n−1,2(ν−m)) (cid:18)1 (n − iλ) , n; ×2F1 (n + iλ) , 0 k 1 2 2 (cid:0)1 − 2ρ2(cid:1) (cid:19) dρ ρ2 ρ2 − 1 (cid:17) (5.12) (5.13) (5.14) To calculate this last integral we make the change of variable ρ = tanh t. Therefore (5.14) reads Iν,m k (λ) = +∞(cid:90) 2n (sinh t)2n−1 P(n−1,2(ν−m)) k 0 × (cosh t) −4(ν−m)+1 .2F1 (cid:16) (cid:18) n + iλ 1 − 2 tanh2 t n − iλ , 2 2 , n;− sinh2 t (5.15) (cid:19) dt 10 A. BOUSSEJRA, Z. MOUAYN Now, we make use of the result established by Koornwinder [17] 0 = k 2 2 dt (cid:17) × Wk ×2 F1 1 − 2 tanh2 t , α + 1;− sinh2 t (cid:16) +∞(cid:90) (cosh t)−α+β−δ−µ(cid:48)−1 (sinh t)2α+1 P(α,δ) (cid:19) (cid:18) α + β + 1 + iλ , α + β + 1 − iλ 2 (δ + µ(cid:48) + 1 − iλ)(cid:1) 2 (δ + µ(cid:48) + 1 + iλ)(cid:1) Γ(cid:0) 1 Γ (α + 1) (−1)k Γ(cid:0) 1 2 (α − β + δ + µ(cid:48) + 2) + k(cid:1) 2 (α + β + δ + µ(cid:48) + 2) + k(cid:1) Γ(cid:0) 1 k!Γ(cid:0) 1 (cid:18)1 (cid:48) + 1(cid:1) , (cid:0)δ + µ (cid:48) + 1(cid:1) , (cid:0)δ − µ (cid:0)x2, a, b, c, d(cid:1) := (a + b)k (a + c)k (a + d)k (cid:12)(cid:12)(cid:12)(cid:12)Γ (cid:18) −k, k + a + b + c + d − 1, a + ix, a − ix (cid:19)(cid:12)(cid:12)(cid:12)(cid:12)2 2nΓ (n) (−1)k a + b, a + c, a + d (α + β + 1) , ×4 F3 + i λ 2 4 λ2; Iν,m k (λ) = (cid:18) Wk 1 2 1 2 1 2 1 2 (cid:19) 1 where β, δ, λ ∈ R, α, δ > −1 , δ + (cid:60)(µ)(cid:48) > −1 and Wk (.) is the Wilson polynomial given in terms of the 4F3-sum as ([1],p. 158): (5.16) (cid:19) (α − β + 1) (5.17) (5.18) (5.19) for the parameters α = n − 1, δ = 2(ν − m) − n, β = 0 and µ(cid:48) = 2(ν − m) − n − 1.We find that Summarizing the above calculations F [hν,n m ] (λ) = k!Γ2 (2 (ν − m) + k) ×Wk (cid:18)1 4 λ2; 2 (ν − m) − n 2 , 1 + , , n 2 n 2 n 2 2 (ν − m) − n (cid:19) 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12)2 + i λ 2 2 (ν − m) − n 2 (cid:18) (cid:12)(cid:12)(cid:12)(cid:12)Γ (cid:32) (cid:33) × 2m∑ k=0 γν,n,m k Wk 2 λ 4 ; 2 (ν − m) − n 2 , 1 + n 2 , n 2 , n 2 with the constants γν,n,m k where the constants Aν,n,m announced result. k := 2m!Γ (n) (2 (ν − m) − n) Γ (2ν − m) (−1)k Aν,n,m Γ (n + m) Γ (2ν − m − n + 1) k!Γ2 (2 (ν − m) + k) is given by (5.10). Finally, replacing λ by(cid:112)−∆Bn − n2, we arrive at the (5.20) , k Remark 5.1. Setting m = 0 in the formula (5.3) in Theorem 5.1, we recover the result of Peetre [18]. Remark 5.2. We should note that the transform Bν,n the Laplace-Beltrami operator ∆Bn in terms of the 3(cid:122)2-sum as m have been expressed in [13] as a function of Γ(cid:16) (cid:16) n − i(cid:112)−∆Bn − n2(cid:17)(cid:17) Γ(cid:16) (cid:16) n + i(cid:112)−∆Bn − n2(cid:17)(cid:17) 2 (ν − m) − 1 n + i(cid:112)−∆Bn − n2(cid:17) (cid:16)  1 2 (ν − m) + j + 1 n + i(cid:112)−∆Bn − n2(cid:17) (cid:16) (ν − m) + j + 1 , n + j, 1 2 (cid:16) 2 2 2 n + i(cid:112)−∆Bn − n2(cid:17) , n 2 Bν,n m = 2m∑ j=0 Cν,n,m j ×3 (cid:122)2 (5.21)  1 where Cν,n,m j = (2 (ν − m) − n) Γ (n + m) (−1)j Γ (n + j) m!Γ (2ν − n − m + 1) Γ (2ν − n) × min(m,j)∑ p=max(0,j−m) (m!)2 Γ (2ν − m) Γ (2ν − m + j − p) (j − p)! (m + p − j)!p! (m − p)!Γ (n + j − p) Γ (n + p) REFERENCES 11 (5.22) . [1] G. E. Andrews, R. Askey and R. Roy, Special Functions, Encyclopedia of Mathematics and its Applications, Cam- [2] Ayaz, K.and Intissar, A.: Selberg trace formula for the heat and wave kernels of Maass Lpalacians of the complex bridge University Press, 1999 hyperbolic space Hn (C) , n ≥ 2. Diff. Geom. App. 15, 1-31 (2001) [3] Berezin, F.A.: General concept of quantization. Commun. Math. Phys. 40, 153-174 (1975) [4] Berezin, F.A.: Quantization. Math. USSR, Izv. 8, 1109-1165 (1974) [5] Berezin, F.A.: Quantization in complex symmetric spaces. Izv. Akad. Naud.SSSR. Ser.Mat(2) 39, 1109-1165 (1974) [6] Bonami, N. Bruna, J. and Grellier, S.: On Hardy, BMO and Lipschitz spaces of invariantly harmonic functions in the unit ball. Proc. London Math. Soc. 3, 665-696 (1998) [7] Boussejra, A. and Intissar, A. L2-concrete Spectral Analysis of the Invariant Laplacian ∆ α,β in the Unit complex ball Bn, J. Funct. Anal 160, 115-140 (1998) [8] Chaggara, H. and Koepf, W.: On linearization coefficients of Jacobi polynomials. App. Math. Lett. 23, 609-614 (2010) [9] Dodonov, V. V. : 'Noncalssical' states in quantum optics: a 'squeezed review of the first 75 years, J.opt.B:Quantum [10] Folland, G. B.: Spherical harmonic expansion of the Poisson-Szegö kernel for the ball. Proc.Am. Math. Soc. 47 (2) , Semiclass.opt.4, R1-R33 (2002) 401-407 (1975) [11] Gazeau, J.P.: Coherent states in quantum physics. Wiley, Weinheim (2009) [12] Geller, D.: Some results in Hp theory for the Heisenberg group, Duke Math. J. 47, 365-391 (1980) [13] Ghanmi, A. and Mouayn, Z.: A formula representing magnetic Berezin transforms on the Bergman ball of Cn as functions of the Laplace-Beltrami operator, arxiv: 1101.1248v1 (2011) [14] Gradshteyn I S and Ryzhik I M, "Table of Integrals, Series and Products", Academic Press, INC, Seven Edition [15] Helgason, S.: A duality for symmetric spaces with applications to group representations. Advances in math., 5, 2007 1-154 (1970) [16] Mourad E.H.Ismail.: Classical and Quantum Orthogonal Polynomials in one variable, Encyclopedia of Mathemat- ics and its applications, Cambridge university press (2005) [17] Koornwinder T.H.: Special orthogonal polynomial systems mapped onto each other by the Fourier-Jacobi trans- form. Orthogonal polynomials and applications (Bar-le-Duc, 1984), 174-183, Lecture Notes in Math. 1171, Springer, Berlin, 1985 [18] Peetre, J.: The Berezin transform and Ha-plitz operator. J. Operator Theory. 24, 165-186 (1990) [19] Rudin, W.: Function theory in the unit ball of Cn. Grundlehern der mathematischen Wissenschften, vol.241. [20] Srivastava, H. M. and Manocha A.: A treatise on Generating functions, John Willey and Sons, New York, Chich- [21] Unterberger, A., Upmeir, H.: The Berezin transform and invariant differential operators. Commun. Math. Phys. 164, Springer, New York (1980) ester, Brisbane, Toraonto (1984) 563-597 (1994) [22] Zernik, F. and Brinkman, H.C.: Hyperspharische Functionen und die in spahrischen bereichen orthogonalen poly- nome. Proc. Kon. Akad. v. wet. Amsterdam 38, 161-170 (1935) (A. B.) DEPARTMENT OF MATHEMATICS, FACULTY OF SCIENCES, IBN TOFAIL UNIVERSITY, KENITRA, MOROCCO E-mail address: [email protected] (Z. M.) DEPARTMENT OF MATHEMATICS, FACULTY OF SCIENCES AND TECHNICS, SULTAN MOULAY SLIMANE UNIVERSITY,BP. 123, BENI MELLAL, MOROCCO E-mail address: [email protected]
1702.03201	1	1702	2017-02-10T15:12:23	Kernel theorems for modulation spaces	[ "math.FA", "math.NA" ]	We deal with kernel theorems for modulation spaces. We completely characterize the continuity of a linear operator on the modulation spaces $M^p$ for every $1\leq p\leq\infty$, by the membership of its kernel to (mixed) modulation spaces. Whereas Feichtinger's kernel theorem (which we recapture as a special case) is the modulation space counterpart of Schwartz' kernel theorem for temperate distributions, our results do not have a couterpart in distribution theory. This reveals the superiority, in some respects, of the modulation space formalism upon distribution theory, as already emphasized in Feichtinger's manifesto for a post-modern harmonic analysis, tailored to the needs of mathematical signal processing. The proof uses in an essential way a discretization of the problem by means of Gabor frames. We also show the equivalence of the operator norm and the modulation space norm of the corresponding kernel. For operators acting on $M^{p,q}$ a similar characterization is not expected, but sufficient conditions for boundedness can be sated in the same spirit.	math.FA	math	KERNEL THEOREMS FOR MODULATION SPACES ELENA CORDERO AND FABIO NICOLA Abstract. We deal with kernel theorems for modulation spaces. We completely characterize the continuity of a linear operator on the modulation spaces M p for every 1 ≤ p ≤ ∞, by the membership of its kernel to (mixed) modulation spaces. Whereas Feichtinger's kernel theorem (which we recapture as a special case) is the modulation space counterpart of Schwartz' kernel theorem for temperate distributions, our results do not have a couterpart in distribution theory. This reveals the superiority, in some respects, of the modulation space formalism upon distribution theory, as already emphasized in Feichtinger's manifesto for a post-modern harmonic analysis, tailored to the needs of mathematical signal processing. The proof uses in an essential way a discretization of the problem by means of Gabor frames. We also show the equivalence of the operator norm and the modulation space norm of the corresponding kernel. For operators acting on M p,q a similar characterization is not expected, but sufficient conditions for boundedness can be sated in the same spirit. 1. Introduction Schwartz' kernel theorem certainly represents one of the most important results in modern Functional Analysis. It states, in the framework of temperate distribu- tions, that every linear continuous operator A : S(Rd) → S ′(Rd) can be regarded as an integral operator in a generalized sense, namely hAf, gi = hK, g ⊗ fi f, g ∈ S(Rd), in the sense of distributions, for some kernel K ∈ S ′(R2d), and viceversa [20]. We write, formally, (1) Af (x) =ZRd K(x, y) f (y)dy f ∈ S(Rd). This result provides a general big framework including most operators of interest in Harmonic Analysis. However, for most applications in Time-frequency Analysis and Mathematical Signal Processing the spaces S(Rd) and S ′(Rd) can be fruit- fully replaced by the modulation spaces M 1(Rd) and M ∞(Rd), respectively, intro- duced by H. Feichtinger in [13] and nowadays widely used as a fundamental tool in 2010 Mathematics Subject Classification. 42B35, 42C15, 47G30, 81Q20. Key words and phrases. Time-frequency analysis, Gabor frames, modulation spaces, 1 2 ELENA CORDERO AND FABIO NICOLA Time-frequency Analysis (see [18] and also Feichtinger's manifesto for a postmod- ern Harmonic Analysis [17]). One of the obvious advantages is that M 1(Rd) and M ∞(Rd) are Banach spaces; secondly their norm gives a direct and transparent information on the time-frequency concentration of a function or temperate distri- bution. The same holds, more generally, for a whole scale of intermediate spaces M p(Rd), 1 ≤ p ≤ ∞, as well as the more general mixed norm spaces M p,q(Rd), 1 ≤ p, q ≤ ∞. In short, for (x, ξ) ∈ Rd × Rd, define the time-frequency shifts π(x, ξ) by π(x, ξ)f (t) = e2πitξf (t − x), f ∈ S ′(Rd). Then M p,q(Rd) is the space of temperate distributions f in Rd such that the function (x, ξ) 7→ hf, π(x, ξ)gi, for some (and therefore any) window g ∈ S(Rd)\{0}, is in Lq(Rd x)), endowed with the obvious norm (see Section 2 below for details). We also set M p(Rd) = M p,p(Rd). Weighted versions are also used in the literature but here we limit ourselves, for simplicity, to the unweighted case. ξ; Lp(Rd Now, a kernel theorem in the framework of modulation spaces was announced in [9] and proved in [14]; see also [18, Theorem 14.4.1]. It states that linear continuous operators A : M 1(Rd) → M ∞(Rd) are characterized by the memenbership of thier distribution kernel K to M ∞(R2d). In this note we provide a similar characterization of linear continuous operators acting on the following spaces: • M 1(Rd) → M p(Rd), for a fixed 1 ≤ p ≤ ∞ • M p(Rd) → M ∞(Rd), for a fixed 1 ≤ p ≤ ∞ • M p(Rd) → M p(Rd), for every 1 ≤ p ≤ ∞. In all cases the characterization is given in terms of the membership of their distribution kernels to certain mixed modulation spaces, first introduced and studied in [3]. The proof uses in an essential way a discretization of the problem by means of Gabor frames. We also show the equivalence of the operator norm and the modulation space norm of the corresponding kernel. These results seem remarkable, because no such a characterization is known, for example, for linear continuous operators from S(Rd) into itself and similarly from S ′(Rd) into itself. This again shows the superiority, in some respects, of the modulation space framework upon distribution theory. For operators acting on mixed-norm modulation spaces M p,q(Rd) → M p,q(Rd) such a characterization is not expected, as we will see in Section 3 below. We have however sufficient conditions for boundedness in the same spirit. KERNEL THEOREMS FOR MODULATION SPACES 3 Let us also observe that in the literature there are plenty of sufficient conditions for operators in certain classes (Fourier multipliers, localization operators, pseudo- differential operators, Fourier integral operators) to be bounded M p(Rd) → M p(Rd) for every 1 ≤ p ≤ ∞, or M p,q(Rd) → M p,q(Rd) for every 1 ≤ p, q ≤ ∞. For ex- ample it is known that pseudodifferential operators with Weyl symbol in M ∞,1 are in fact bounded M p,q(Rd) → M p,q(Rd) for every 1 ≤ p, q ≤ ∞ [6, 18, 19, 26, 27] (see also [24]). On the other hand the Fourier multipier with symbol eiξ2 is still bounded on these spaces [2], although that symbol does not belong to M ∞,1(R2d). Similary, it is easy to see that any linear continuous operator A : S(Rd) → S ′(Rd) whose Gabor matrix enjoys decay estimates such as hAπ(z)g, π(w)gi ≤ C(1 + w − χ(z))−s z, w ∈ R2d for some s > 2d, g ∈ S(Rd) \ {0}, and bi-Lipschtiz mapping χ : R2d → R2d, is bounded M p(Rd) → M p(Rd) for every 1 ≤ p ≤ ∞ [4] (see also [4, 7, 8]). But once again this condition is not necessary at all. The very simple characterization provided in the present paper when p = q therefore includes, at least implicitly, all these continuity results on M p. 2. Backgrounds on time-frequency analysis 2.1. Mixed norm spaces. We summarize the main properties of mixed norm spaces, we refer to [1] for a full treatment of this subject. Consider measure spaces (Xi, µi), indices pi ∈ [1,∞], with i = 1, . . . , d. Then the mixed norm space Lp1,p2,...,pd(X1 × X2 × · · · × Xd; µ1 × µ2 × · · · × µd) consists of the measurable functions F : X1 × X2 × · · · × Xd → C such that the following norm is finite: (2) kFkLp1,p2,...,pd = ZXd · · ·(cid:18)ZX1 F (x1, . . . , xd)p1 dµ1(x1)(cid:19) 1 pd p2 p1 · · · dµd(xd)! with standard modification when pi = ∞ for some i. If Xi = R and µi is the Lebesgue measure on R for every i = 1, . . . , d, we simply write Lp1,...,pd. If each Xi is a countable set and the corresponding measure µi is the counting measure, we use the notation ℓp1,...,pd(X1 × X2 × · · · × Xd). Banach spaces, generalizations of the classical Lp and ℓp spaces, cf. [1]. 2.2. Modulation spaces. For x, ξ ∈ Rd, define the translation operator Tx and modulation operator Mξ by The mixed norm spaces Lp1,p2,...,pd(X1 × X2 × · · · × Xd; µ1 × µ2 × · · · × µd) are Txf (t) = f (t − x) Mξf (t) = e2πitxf (t). For z = (x, ξ), the composition operator π(z) = MξTx is called a time-frequency shift of the phase space R2d. 4 ELENA CORDERO AND FABIO NICOLA For a fixed non-zero g ∈ S(Rd), the short-time Fourier transform (STFT) of f ∈ S ′(Rd) with respect to the window g is given by (3) f (t) g(t − x) e−2πiξt dt = hf, π(z)gi Vgf (x, ξ) =ZRd z = (x, ξ), where the integral is intended as the (anti-)duality between S ′ and S. The short- time Fourier transform gives information about the time-frequency content of the signal f : indeed, roughly speaking, it can be seen as a localized Fourier transform of f in a neighbourhood of x. The STFT is the basic tool in the definition of modulation spaces. Indeed, modulation space norms are a measure of the joint time-frequency distribution of f ∈ S ′. For their basic properties we refer [10, 11, 12] (see also [3, 21, 22, 23]) and the textbook [18]. Definition 2.1. Fix g ∈ S(Rd), p = (p1, . . . , pd), q = (q1, . . . , qd), pi, qi ∈ [1,∞], i = 1, . . . , d. Then M p,q(Rd) = {f ∈ S ′(Rd) : kfkM p,q(Rd) < ∞}, where kfkM p,q(Rd) = kVgfkLp1,...,pd ,q1,...,qd . q = q1, . . . , qd, we come back to the classical modulation If p = p1, . . . , pd, spaces of distributions f ∈ S ′(Rd) such that kfkM p,q(Rd) = ZRd(cid:18)ZRd Vgf (x, ξ)p dx(cid:19) q p 1 q dξ! < ∞ [13]). The extension of the original modulation spaces in Definition 2.1 was (cf. widely studied and applied to the investigation of boundedness and sampling prop- erties for pseudodifferential operators in [21, 22, 23]. To state our kernel theorems we need a further extension of Definition 2.1, that was introduced by S. Bishop in [3], in her study of Schatten p-class properties for integral operators. This is a natural generalization of the classical modulation spaces, as will be clear in the sequel. We shall mainly use the notation in [3]. From now on we assume that c is a permutation of the set {1, . . . , 2d}. We identify c with the linear bijection c : R2d → R2d, c(x1, . . . , x2d) = (xc(1), . . . , xc(2d)). Definition 2.2. Consider g ∈ S(Rd) and let c be a permutation corresponding to the map c as above. Then M(c)p1,p2,...,p2d is the mixed modulation space of tempered distributions f ∈ S ′(Rd) for which (4) kfkM (c)p1,p2 ,...,p2d = kVgf ◦ ckLp1,p2 ,...,p2d < ∞. KERNEL THEOREMS FOR MODULATION SPACES 5 If p = p1 = p2 = · · · = pd, q = pd+1 = pd+2 = · · · = p2d we simply write M(c)p,q; if p = p1 = p2 = · · · = pd = pd+1 = pd+2 = · · · = p2d, we shorten to M(c)p. Remark 2.3. (i) If c is the identity permutation and the indices satisfy p = p1 = p2 = · · · = pd, q = pd+1 = pd+2 = · · · = p2d then M(c)p1,p2,...,p2d = M p,q(Rd) is the original modulation space introduced by Feichtinger in [13]. (ii) If c is the identity permutation and p = (p1, . . . , pd), q = (q1, . . . , qd), we come back to the modulation spaces in Definition 2.1. (iii) If the indices satisfy p = p1 = p2 = · · · = pd = pd+1 = pd+2 = · · · = p2d, given any permutation c, a simple change of variables yields M(c)p = M p(Rd). (iv) In general the mixed modulation spaces differ from the classical ones of Definition 2.1. A simple example is provided by the permutation (5) c0(1, 2, . . . , d, d + 1, . . . , 2d) = (d + 1, d + 2, . . . , 2d, 1, 2, . . . , d). If we choose p = p1 = p2 = · · · = pd, then it is easy to see that q = pd+1 = pd+2 = · · · = p2d M p,q(c0) = W (F Lp, Lq)(Rd), that is the space of distributions f ∈ S ′(Rd) such that kfkW (F Lp,Lq)(Rd) = ZRd(cid:18)ZRd Vgf (x, ξ)p dξ(cid:19) dx! q p 1 q < ∞ for some window function g ∈ S(Rd) \ {0}. It can be shown (cf. [3]) that the definition of mixed modulation spaces is in- dependent of the choices of the window g in S(Rd), with different windows giv- ing equivalent norms. Furthermore, this fact also holds for g in the larger space M 1(Rd), as follows by the subsequent Theorem 2.4, which gives the inversion for- mula on M(c)p1,...,p2d. Theorem 2.4. Suppose c is a permutation of {1, . . . , 2d}, and for a measurable function ψ : Rd → C define the operator Υψ by (6) F (x)π(c(x))ψ(t) dx. ΥψF (t) =ZR2d 6 ELENA CORDERO AND FABIO NICOLA (i) Given ψ, g ∈ M 1(Rd), p1, . . . , p2d ∈ [1,∞], and f ∈ M(c)p1,...,p2d, we have Υψ(Vgf ◦ c) = hψ, gif. (ii) Different windows in M 1(Rd) define equivalent norms on M(c)p1,...,p2d. Corollary 2.5. For any p1, . . . , p2d ∈ [1,∞], M(c)p1,...,p2d is a Banach space. We also have the expected duality result. Theorem 2.6. If p1, . . . , p2d ∈ [1,∞), M(c)p′ where p′ i ∈ (1,∞] satisfies 1/pi + 1/p′ i = 1. 1,...,p′ 2d is the dual space of M(c)p1,...,p2d, 2.3. Gabor frames. Given a lattice Λ in R2d and a non-zero square integrable function g (called window) on Rd the system G(g, Λ) = {π(λ) : λ ∈ Λ} is called a Gabor frame if it is a frame for L2(Rd), that is there exist constants 0 < A ≤ B such that Akfk2 (7) hf, π(λ)gi2 ≤ Bkfk2 2, ∀f ∈ L2(Rd). 2 ≤Xλ∈Λ If (7) holds, then there exists a γ ∈ L2(Rd) (so-called dual window), such that G(γ, Λ) is a frame for L2(Rd) and every f ∈ L2(Rd) can be expanded as (8) f =Xλ∈Λ hf, π(λ)giπ(λ)γ =Xλ∈Λ hf, π(λ)γiπ(λ)g, with unconditional convergence in L2(Rd). From now on we shall work with lattices Λ = αZ2d, for suitable α > 0. If the window function belongs to the Feichtinger's algebra M 1(Rd), then a Gabor frame for L2(Rd) is a frame also for mixed modulation spaces (cf. [3, Theorem 4.6]). Theorem 2.7. Let g be a window function in M 1(Rd) such that G(g, Λ) is a frame for L2(Rd) with a dual frame G(γ, Λ). Consider p1, . . . , p2d ∈ [1,∞]. Then (i) There exist constants 0 < A ≤ B independent of p1, . . . , p2d such that AkfkM (c)p1 ,...,p2d ≤ kVgf ◦ cΛkℓp1,...,p2d ≤ BkfkM (c)p1,...,p2d , (ii) We have ∀f ∈ M(c)p1,...,p2d. (9) f =Xλ∈Λ hf, π(λ)giπ(λ)γ =Xλ∈Λ hf, π(λ)γiπ(λ)g, with unconditional convergence in M(c)p1,...,p2d if p1, . . . , p2d ∈ [1,∞), weak* convergence in M ∞(Rd) if p1, . . . , p2d ∈ [1,∞]. KERNEL THEOREMS FOR MODULATION SPACES 7 This generalizes the basic fact, valid for the identity permutation, that the coef- ficient operator Cg and the reconstruction operator Dγ, defined by (10) satisfy (Cgf )λ := hf, π(λ)gi Dγc :=Xλ∈Λ cλπ(λ)γ, Cg : M p,q(Rd) → ℓp,q(Λ), Dγ : ℓp,q(Λ) → M p,q(Rd), (11) with DγCg = I, the identity on M p,q(Rd), cf. (9), and similarly if we interchange the windows g and γ. 3. Kernel Theorems In this section we will present the characterization of continuity properties on modulation spaces of integral operators (1). We need the following result (see e.g. Tao's Lecture Notes [25, Proposition 5.2]). Proposition 3.1. Assume that (X, νX), (Y, µY ) are measure space and p ∈ [1,∞]. Suppose that K : X ×Y → C, and kK(·, y)kLp(X) is uniformly bounded on Y . Then the operator A with kernel K maps L1(X) into Lp(Y ) and (12) kAkL1(X)→Lp(Y ) = ess sup y∈Y kK(·, y)kLp(X). Similarly, we have a dual statment (see e.g. [25, Proposition 5.4]). Proposition 3.2. Assume that (X, νX), (Y, µY ) are measure space and p ∈ [1,∞]. Suppose that K : X × Y → C, and kK(x,·)kLp′ (Y ) is uniformly bounded on Y . Then the operator A with kernel K maps Lp(X) into L∞(Y ) and (13) kAkLp(X)→L∞(Y ) = ess sup x∈X kK(x,·)kLp′ (Y ). Let us introduce the permutation c1 of the set {1, . . . , 4d} with related c1 given Recall that we only consider lattices of the form Λ = αZ2d. Theorem 3.3. Suppose p ∈ [1,∞]. Let g be a window function in S(Rd) such that G(g, Λ) is a frame for L2(Rd) with a dual frame G(γ, Λ). A linear continuous operator A : S(Rd) → S ′(Rd) is bounded M 1(Rd) → M p(Rd) if and only if its distrinution kernel K satisfies (15) VGK ◦ c1 ∈ ℓp,∞(Λ × Λ), by, for xi ∈ Rd, i = 1, . . . , 4, c1(x1, x2, x3, x4) =  (14) Id 0d 0d 0d Id 0d 0d 0d Id 0d 0d 0d 0d 0d 0d Id   x1 x2 x3 x4     = (x1, x3, x2, x4). 8 ELENA CORDERO AND FABIO NICOLA where G = g ⊗ ¯γ. Moreover kAkM 1→M p ≍ kVGK ◦ c1kℓp,∞(Λ×Λ). Proof. Assume first (15). For λ = (λ1, λ2), µ = (µ1, µ2) ∈ Λ, observe that the Gabor matrix of the operator A is given by (16) Kλ,µ := hAπ(µ)γ, π(λ)gi = hK, π(λ1, λ2)g ⊗ π(µ1,−µ2)¯γi = VGK ◦ c1(λ1, λ2, µ1,−µ2). This matrix turns out to be the integral kernel of the operator A := CgADγ, acting on sequences on Λ, where the coefficient operator Cg and synthesis operator Dγ are defined in (10). By (15) and Proposition 3.1, A is a bounded operator ℓ1(Λ) → ℓp(Λ). Since Cg : M p(Rd) → ℓp(Λ) and Dγ : ℓp(Λ) → M p(Rd) continuously, we have A = Dγ ACγ : M 1(Rd) → M p(Rd), with kAkM 1→M p ≤ C sup µ∈Λ kK·,µkℓp(Λ). Vice-versa, consider the finite dimensional space EN = {(aλ)λ∈Λ : aλ = 0 for λ > N} and the natural projection PN : ℓ∞(Λ) → EN . Consider the composition operator (17) PN CgADγEN : (EN ,k · kℓ1) → (EN ,k · kℓp). This operator is represented by the (finite) matrix {Kλ,µ}λ,µ∈Λ,λ≤N,µ≤N . Using formula (12) in Proposition 3.1 (we work on finite sets, so that the assump- tions are trivially satisfied) we have sup µ∈Λ,µ≤N kKλ,µkℓp(Λ,λ≤N ) = kPN CgADγENkℓ1→ℓp ≤ kCgkkDγkkAkM 1→M p. Since the right-hand side is independent of N, taking the supremum with respect to N, we obtain This concludes the proof. µ∈Λ kK·,µkℓp(Λ) ≤ CkAkM 1→M p. sup We observe that, given a frame G(g, Λ) for L2(Rd), Λ = αZ2d, with dual frame G(γ, Λ), then G(¯γ, Λ) is still a frame for L2(Rd) and G(g ⊗ ¯γ, Λ × Λ) is a frame for L2(R2d). Using Theorem 2.7 we can therefore reformulate the previous result in terms of mixed modulation spaces. KERNEL THEOREMS FOR MODULATION SPACES 9 Corollary 3.4. Suppose p ∈ [1,∞]. A linear continuous operator A : S(Rd) → S ′(Rd) is bounded from M 1(Rd) into M p(Rd) if and only if its distribution kernel K satisfies (18) where c1 is identified with (14). Moreover K ∈ M(c1)p,∞, kAkM 1→M p ≍ kKkM (c1)p,∞. Remark 3.5. When p = ∞, condition (18) becomes K ∈ M ∞(R2d) and we recap- ture Feichtinger's kernel Theorem [14]. A similar characterization of bounded operators A : M p(Rd) → M ∞(Rd) can be obtained arguing as above, using Proposition 3.2, (16) and Theorem 2.7. Precisely, define the permutation c2 of the set {1, . . . , 4d} with related c2 by (19) = (x3, x1, x4, x2), c2(x1, x2, x3, x4) =  0d 0d Id 0d Id 0d 0d 0d Id 0d 0d 0d 0d Id 0d 0d   x1 x2 x3 x4     for xi ∈ Rd, i = 1, . . . , 4. Theorem 3.6. Suppose p ∈ [1,∞]. A linear continuous operator A : S(Rd) → S ′(Rd) is bounded from M p(Rd) into M ∞(Rd) if and only if its distribution kernel K satisfies (20) where 1/p + 1/p′ = 1. Moreover K ∈ M(c2)p′,∞, kAkM p→M ∞ ≍ kKkM (c2)p′ ,∞. As a consequence of Theorems 3.3 and 3.6 we have the following result. Corollary 3.7. Let A : S(Rd) → S ′(Rd) be an operator with distribution kernel K ∈ S ′(R2d). The following conditions are equivalent: (i) A : M p(Rd) → M p(Rd), for every 1 ≤ p ≤ ∞. (ii) A : M 1(Rd) → M 1(Rd) and A : M ∞(Rd) → M ∞(Rd). (iii) K ∈ M(c1)1,∞ ∩ M(c2)1,∞. Proof. All the equivalences follow trivially from Theorems 3.3 and 3.6 but (ii) ⇒ (i), which follows by complex interpolation for modulation spaces [13]. 10 ELENA CORDERO AND FABIO NICOLA 4. Operators acting on M p,q Let us now consider the problem of boundedness of an integral operator on M p,q(Rd), 1 ≤ p, q ≤ ∞. We do no longer expect a characterization as in the case p = q, studied in the previous section (see Remark 4.2 below). However sufficient conditions in the same spirit can be stated. Consider the permutations c3 and c4 of the set {1, . . . , 4d} with related c3, c4 defined by (21) and (22) c3(x1, x2, x3, x4) =  c4(x1, x2, x3, x4) =  Id 0d 0d 0d 0d 0d Id 0d Id 0d 0d 0d 0d 0d 0d Id Id 0d 0d 0d 0d 0d 0d Id Id 0d 0d 0d 0d 0d Id 0d     x1 x2 x3 x4 x1 x2 x3 x4         = (x2, x3, x1, x4), = (x1, x4, x2, x3), for xi ∈ Rd, i = 1, . . . , 4. contained in [5, Proposition 2.4]). The following result was proved in [7, Proposition 5.1] (a continuous version is Proposition 4.1. Consider an at most countable index set J and the operator defined on sequences (aλ1,λ2) on the set Λ = J × J by (Aa)λ1,λ2 = Xµ1,µ2∈J Kλ1,λ2,µ1,µ2aµ1,µ2. kAkℓ∞,1→ℓ∞,1 ≤ kK ◦ c3kℓ1,∞,1,∞. (i) If K ◦ c3 ∈ ℓ1,∞,1,∞(J × J × J × J), then A is continuous on ℓ∞,1(Λ). Moreover, (23) (ii) If K ◦ c4 ∈ ℓ1,∞,1,∞(J × J × J × J), then A is continuous on ℓ1,∞(Λ), with (24) (iii) If K satisfies assumptions (i) and (ii) and moreover K ◦ c0, K ∈ ℓ1,∞(Λ × Λ), where the permutation c0 is defined in (5), then the operator A is continuous on ℓp,q(Λ), for every 1 ≤ p, q ≤ ∞. Remark 4.2. Observe that the reverse inequalites in (23), (24) do not hold, even if one allows a multiplicative constant. Consider for example (23), with J = {0, 1, . . . , N − 1}, N ∈ N, where kAkℓ1,∞→ℓ1,∞ ≤ kK ◦ c4kℓ1,∞,1,∞. Kλ1,λ2,µ1,µ2 = Fλ2,µ1 = e−2πiλ2µ1 1 √N KERNEL THEOREMS FOR MODULATION SPACES 11 is the Fourier matrix. We have so that but (K ◦ c3)µ1,λ1,λ2,µ2 = Kλ1,λ2,µ1,µ2 kK ◦ c3kℓ1,∞,1,∞ = kFkℓ1 = N 3/2 (25) which blows-up, as N → +∞, at a lower rate. kAkℓ∞,1→ℓ∞,1 ≤ N, Let us verify (25). We have λ1kℓ1 λ2 kk Xµ1,µ2∈J Kλ1,λ2,µ1,µ2aµ1,µ2kℓ∞ = kXµ1∈J = kXµ1∈J with aµ1 :=Pµ2∈J aµ1,µ2. Now, using the embeddings kbkℓ2(J) ≤ √Nkbkℓ2(J), kbkℓ1(J) ≤ aµ1,µ2kℓ1 λ2 Fλ2,µ1 Xµ2∈J Fλ2,µ1aµ1kℓ1 λ2 √Nkbkℓ∞(J) and the fact that F is a unitary transformation of ℓ2(J) we get (25). As a consequence, we obtain the following boundedness result. Consider the permutations c5 and c6 of the set {1, . . . , 4d} with related c5, c6 defined by (26) and (27) c5 = c1c3 =  c6 = c1c4 =  0d Id 0d 0d Id 0d 0d 0d 0d 0d Id 0d Id 0d 0d 0d Id 0d 0d 0d Id 0d 0d 0d Id 0d 0d 0d 0d 0d Id 0d     . Theorem 4.3. (i) A linear continuous operator A : S(Rd) → S ′(Rd) is bounded on M ∞,1(Rd) if its distribution kernel K satisfies K ∈ M(c5)1,∞,1,∞. (28) (ii) A linear continuous operator A : S(Rd) → S ′(Rd) is bounded on M 1,∞(Rd) if its distribution kernel K satisfies (29) K ∈ M(c6)1,∞,1,∞. 12 ELENA CORDERO AND FABIO NICOLA Proof. The result is a consequence of (16), Proposition 4.1 and Theorem 2.7, by using the same arguments as in the first part of the proof of Theorem 3.3. By interpolation we also obtain the following result. Corollary 4.4. Let A : S(Rd) → S ′(Rd) be an operator with distribution kernel K ∈ M(c5)1,∞,1,∞ ∩ M(c6)1,∞,1,∞. Then A : M p′,p(Rd) → M p′,p(Rd) continuously, for every 1 ≤ p ≤ ∞, with 1/p + 1/p′ = 1. By combining Corollaries 3.7 and 4.4, we finally obtain the following result. Corollary 4.5. Let A : S(Rd) → S ′(Rd) be an operator with distribution kernel K ∈ M(c1)1,∞ ∩ M(c2)1,∞ ∩ M(c5)1,∞,1,∞ ∩ M(c6)1,∞,1,∞. Then A : M p,q(Rd) → M p,q(Rd) continuously, for every 1 ≤ p, q ≤ ∞. Acknowledgments This research was partially supported by the Gruppo Nazionale per l'Analisi Matematica, la Probabilit`a e le loro Applicazioni (GNAMPA) of the Istituto Nazionale di Alta Matematica (INdAM), Italy. References [1] A. Benedek and R. Panzone. The space Lp with mixed-norm. Duke Math. J., 28:301–324, 1961. [2] A. B´enyi, K. Grochenig, K.A. Okoudjou and L.G. Rogers. Unimodular Fourier multipliers for modulation spaces. J. Funct. Anal., 246(2):366-384, 2007. [3] S. Bishop. Mixed modulation spaces and their applications to pseudodifferential operators. J. Math. Anal. Appl., 363:255–264, 2010. [4] E. Cordero, K. Grochenig, F. Nicola and L. Rodino. Wiener algebras of Fourier integral operators. J. Math. Pures Appl., 99:219–233, 2013. [5] E. Cordero and F. Nicola. Boundedness of Schrodinger Type Propagators on Modulation Spaces. J. Fourier Anal. Appl., 16:311-339, 2010. [6] E. Cordero, F. Nicola. Pseudodifferential operators on Lp, Wiener amalgam and modulation spaces. Intern. Math. Res. Not., 10:1860–1893, 2010. [7] E. Cordero, F. Nicola and L. Rodino. Time-frequency analysis of Fourier integral operators. Commun. Pure Appl. Anal., 9(1):1–21, 2010. [8] E. Cordero, F. Nicola and L. Rodino. Sparsity of Gabor representation of Schrodinger prop- agators. Appl. Comput. Harmon. Anal., 26(3):357–370, 2009. [9] H.G. Feichtinger. Un espace de Banach de distributions temp´er´ees sur les groupes localement compacts ab´eliens. Compt. Rend. Acad. Sci. Paris, Ser. A, 290(17):791–794, 1980. [10] H. G. Feichtinger. Banach spaces of distributions of Wiener's type and interpolation, in "Proc. Conf. Functional Analysis and Approximation," Oberwolfach August 1980, Internat. Ser. Numer. Math., Birkhauser, Boston, 69:153–165, 1981. KERNEL THEOREMS FOR MODULATION SPACES 13 [11] H. G. Feichtinger. Banach convolution algebras of Wiener type, in "Functions, series, opera- tors," Vol. I, II, (Budapest, 1980), North-Holland, Amsterdam, 509–524, 1983. [12] H. G. Feichtinger. Generalized amalgams, with applications to Fourier transform, Canad. J. Math., 42:395–409, 1990. [13] H.G. Feichtinger. Modulation spaces on locally compact abelian groups, Technical Report, University Vienna, 1983, and also in Wavelets and Their Applications, M. Krishna, R. Radha, S. Thangavelu, editors, Allied Publishers, 99–140, 2003. [14] H.G. Feichtinger, K. Grochenig. Gabor frames and time-frequency analysis of distributions, J. Funct. Anal., 146(2):464–495, 1997. [15] H.G. Feichtinger and T. Strohmer (eds.). Gabor Analysis and Algorithms: Theory and Ap- plications. Birkhuser Boston, Boston, 1998. [16] H.G. Feichtinger and T. Strohmer (eds.). Advances in Gabor analysis. Applied and Numerical Harmonic Analysis. Birkhauser Boston Inc., Boston, 2003. [17] H.G. Feichtinger. Elements of Postmodern Harmonic Analysis, in Operator-related Function Theory and Time-Frequency Analysis. The Abel Symposium 2012, Oslo, Norway, August 20–24, 2012, Springer, 77–105, 2015. [18] K. Grochenig. Foundations of Time-Frequency Analysis. Applied and Numerical Harmonic Analysis, Birkhauser, Boston, MA, 2001. [19] K. Grochenig. Time-frequency analysis of Sjstrand's class. Revista Mat. Iberoam., 22(2):703- 724, 2006. [20] L. Hormander. The Analysis of Linear Partial Differential Operators I, Distribution Theory and Fourier Analysis, Springer-Verlag, Berlin, second edition 1990. [21] S. Molahajloo, K.A. Okoudjou and G. Pfander. Boundedness of Multilinear Pseudo- differential Operators on Modulation Spaces. J. Fourier Anal. Appl., 22(6):1381–1415, 2016. [22] S. Molahajloo and G. Pfander. Boundedness of pseudo-differential operators on Lp, Sobolev and modulation spaces. Math. Model. Nat. Phenom., 8(1):175–192, 2013. [23] G. Pfander. Sampling of operators. J. Fourier Anal. Appl., 19(3):612–650, 2013. [24] J. Sjostrand. An algebra of pseudodifferential operators. Math. Res. Lett., 1(2):185–192, 1994. [25] T. Tao. Lecture Notes 2 for 247 A: Fourier Analysis. [26] J. Toft. Continuity properties for modulation spaces, with applications to pseudo-differential www.math.ucla.edu/∼tao/247a.1.06f/notes2.pdf calculus. I. J. Funct. Anal., 207(2):399–429, 2004. [27] J. Toft. Continuity properties for modulation spaces, with applications to pseudo-differential calculus. II. Ann. Global Anal. Geom., 26(1):73–106, 2004. Universit`a di Torino, Dipartimento di Matematica, via Carlo Alberto 10, 10123 Torino, Italy E-mail address: [email protected] Dipartimento di Scienze Matematiche, Politecnico di Torino, corso Duca degli Abruzzi 24, 10129 Torino, Italy E-mail address: [email protected]
1911.10255	2	1911	2019-12-10T14:41:23	On $C$-compact orthogonally additive operators	[ "math.FA" ]	We consider $C$-compact orthogonally additive operators in vector lattices. After providing some examples of $C$-compact orthogonally additive operators on a vector lattice with values in a Banach space we show that the set of those operators is a projection band in the Dedekind complete vector lattice of all regular orthogonally additive operators. In second part of the article we introduce a new class of vector lattices, called $C$-complete, and show that any laterally-to-norm continuous $C$-compact orthogonally additive operator from a $C$-complete vector lattice to a Banach space is narrow, which generalizes a result of Pliev and Popov.	math.FA	math	ON C-COMPACT ORTHOGONALLY ADDITIVE OPERATORS MARAT PLIEV and MARTIN R. WEBER Abstract. We consider C-compact orthogonally additive operators in vector lattices. After providing some examples of C-compact orthogonally additive operators on a vector lattice with values in a Banach space we show that the set of those operators is a projection band in the Dedekind complete vector lattice of all regular orthogonally additive operators. In the second part of the article we introduce a new class of vector lattices, called C-complete, and show that any laterally-to-norm continuous C-compact orthogonally additive operator from a C-complete vector lattice to a Banach space is narrow, which generalizes a result of Pliev and Popov. 1. Introduction Orthogonally additive operators in vector lattices first were investigated in [12]. Later these results were extended in [1, 2, 6, 7, 8, 20, 22, 23]). Recently, some connections with problems of the convex geometry were revealed [24, 25]. Orthogonally additive operators in lattice-normed spaces were studied in [3]. In this paper we continue this line of research. We analyze the notion of C-compact orthogonally additive operator. In the first part of the article we show that set of all C-compact orthogonally additive operators from a vector lattice E to an order continuous Banach lattice F is a projection band in the vector lattice of all regular orthogonally additive operators from E to F (Theorem 3.9). In the final part of the paper we introduce a new class of vector lattices which we call C-complete (the precise definition is given in section 4) and prove that any laterally-to-norm continuous C-compact orthogonally additive operator from an atomless C-complete vector lattice E to a Banach space X is narrow (Theorem 4.7). This is a generalization of the result of the article [18, Theorem 3.2]. Note that linear narrow operators in function spaces first appeared in [17]. Nowadays the theory of narrow operators is a well-studied object of functional 2010 Mathematics Subject Classification. Primary 47H30; Secondary 47H99. Key words and phrases. C-compact operator, orthogonally additive operator, narrow oper- ator, Urysohn operator, C-complete vector lattice, Banach lattice. 1 2 MARAT PLIEV and MARTIN R. WEBER analysis and is presented in many research articles ([9, 11, 14, 15, 18, 23]) and in the monograph [21]. 2. Preliminaries All necessary information on vector lattices one can find in [4]. In this article all vector lattices are assumed to be Archimedean. Two elements x, y of a vector lattice E are called disjoint (written as x ⊥ y), if x ∧ y = 0. An element a > 0 of E is an atom if 0 ≤ x ≤ a, 0 ≤ y ≤ a and x⊥y imply that either x = 0 or y = 0. The equality x = n xi means that n Fi=1 x = xi and xi⊥xj for all i 6= j. In the case of n = 2 we write x = x1 ⊔ x2. An Pi=1 element y of a vector lattice E is called a fragment (or a component) of x ∈ E, if y ⊥ (x − y). The notation y ⊑ x means that y is a fragment of x. Two fragments x1 and x2 of an element x are said to be mutually complemented if x = x1 ⊔ x2. The set of all fragments of an element x ∈ E is denoted by Fx. Definition 2.1. Let be E a vector lattice and X a real vector space. An operator T : E → X is said to be orthogonally additive if T (x + y) = T x + T y for any disjoint elements x, y ∈ E. It is not hard to check that T (0) = 0. The set of all orthogonally additive operators from E to X is a real vector space with respect to the natural linear operations. Definition 2.2. Let E and F be vector lattices. An orthogonally additive (in general, nonlinear) operator T : E → F is said to be: • positive if T x ≥ 0 for all x ∈ E, • regular, if T = S1 − S2 for two positive, orthogonally additive operators S1 and S2 from E to F . The sets of all positive, regular orthogonally additive operators from E to F are denoted by OA+(E, F ), OAr(E, F ), respectively, where the order in OAr(E, F ) is introduced as follows: S ≤ T whenever (T − S) ≥ 0. Then OAr(E, F ) becomes an ordered vector space. For a Dedekind complete vector lattice F we have the following property of OAr(E, F ). Proposition 2.3 ([19, Theorem 3.6]). Let E and F be a vector lattices, and assume that F is Dedekind complete.Then OAr(E, F ) is a Dedekind complete ON C-COMPACT OPERATORS 3 vector lattice. Moreover, for every S, T ∈ OAr(E, F ) and every x ∈ E the following formulas1 hold (T ∨ S)(x) = sup{T y + Sz : x = y ⊔ z}; (T ∧ S)(x) = inf{T y + Sz : x = y ⊔ z}; (1) (2) (3) T +(x) = sup{T y : y ⊑ x}; (4) T −(x) = − inf{T y : y ⊑ x}; (5) T x ≤ T (x). 3. The projection band of C-compact orthogonally additive operators In this section we show that the set of all C-compact regular orthogonally additive operators from a vector lattice E to a Banach lattice F with order continuous norm is a band in the vector lattice of all orthogonally additive regular operators from E to F . Consider some examples. Example 3.1. Assume that (A, Ξ, µ) and (B, Σ, ν) are σ-finite measure spaces. We say that a map K : A × B × R → R is a Carath´eodory function if there hold the conditions: (1) K(·, ·, r) is µ × ν-measurable for all r ∈ R; (2) K(s, t, ·) is continuous on R for µ × ν-almost all (s, t) ∈ A × B. If additionally there holds K(s, t, ·) = 0 for µ × ν-almost all (s, t) ∈ A × B then with the function K a nonlinear integral operator: T : Dom(K) → L0(µ) (T f )(s) := ZB is associated, where K(s, t, f (t)) dν(t) for µ-almost all s ∈ A Dom(K) := {f ∈ L0(ν) : ZB K(s, t, f (t)) dν(t) ∈ L0(µ)}. Assume that E is a Banach function space in L0(ν) and E ⊆ Dom(K). Then T defines an orthogonally additive integral operator acting from E to L0(µ). The operator T is called Urysohn (integral) operator and, the function K is called the kernel of this operator. If L0(µ) is replaced by R we say that T is an Urysohn integral functional. 1 In the literature these formulas are known as the Riesz-Kantorovich formulas. 4 MARAT PLIEV and MARTIN R. WEBER Example 3.2. Assume that (Ω, Σ, µ) is a measure space. Then the functional N : L1(µ) → R defined by N (f ) = kf kL1(µ) := ZΩ f (t) dµ, f ∈ L1(µ) is positive and orthogonally additive. Let E be a vector lattice. A linear operator S : E → E is called band preserving if Sx ∈ {x}⊥⊥ for all x ∈ E. Clearly, every band preserving operators preserves disjointness. Example 3.3. Let (Ω, Σ, µ) be a measure space and S : L0(µ) → L0(µ) be a band preserving linear operator. Consider the new operator T : L0(µ) → L0(µ) defined by T f = f S(f ), f ∈ L0(µ). Observe that T can be treated as the multiplication operator by the "variable" function. It is not hard to verify that T is an orthogonally additive operator. Indeed, due to Sf ∈ {f }⊥⊥, for every disjoint f1, f2 ∈ L0(µ) we have T (f1 + f2) = (f1 + f2)S(f1 + f2) = f1S(f1 + f2) + f2S(f1 + f2) = f1Sf1 + f2Sf2 = T f1 + T f2. Definition 3.4. Let (Ω, Σ, µ) be a σ-finite measure space. We say that a function N : Ω × R → R belongs to the class S (or N is a S-function) if the following conditions hold: (1) N(t, 0) = 0 for µ-almost all t ∈ Ω; (2) N(·, g(·)) is µ-measurable for any g ∈ L0(µ). If a function N satisfies only the condition (2), then it is called a superpositionally measurable function or sup-measurable function. Example 3.5. Let N : Ω × R → R be a S-function. Then with N there is associated an orthogonally additive operator TN : L0(µ) → L0(µ) defined by TN (g)(t) = N(t, g(t)) for µ-almost all t ∈ Ω and g ∈ L0(µ). This class of operators (sometimes called nonlinear superposition operators or Nemytskii operators) is widely represented in the literature (see e.g. [5]). Definition 3.6. Let be E be a vector lattice and Y a normed space. An orthog- onally additive operator T : E → Y is called ON C-COMPACT OPERATORS 5 • AM-compact provided T maps order bounded subsets of E into relatively compact subsets in Y , • C-compact, if T (Fx) is a relatively compact in Y for any x ∈ E. For a Banach lattice F by COAr(E, F ) is denoted the space of all C-compact regular orthogonally additive operators from E to F . Example 3.7. We note that OAr(R, R) is exactly the set of all real-valued func- tions such that f (0) = 0. Define an orthogonally additive operator T : R → R by 1 x2 , 0, if x 6= 0 if x = 0. T x =   Since any element x ∈ R is an atom one has Fx = {0, x} for any x ∈ R. It follows that T is a C-compact operator. On the other hand T ([0, 1]) is an unbounded set in R and therefore T is not AM-compact. Proposition 3.8. Let (B, Σ, ν) be σ-finite measure spaces E be a Banach func- tion space in L0(ν) and T : E → R be the Urysohn integral functional defined by T f = ZB K(t, f (t)) dν(t), f ∈ E with the kernel K. Then T is C-compact. Proof. Take f ∈ E. We note that Ff coincides with the set {f 1D : D ∈ Σ}. Now for every D ∈ Σ we write T f 1D = ZB K(t, f 1D(t)) dν(t) = ZD K(t, f (t)) dν(t) ≤ ZB K(t, f (t)) dν(t) = M. Hence the set T (Ff ) is order bounded in R and therefore the operator T is C-compact. We mention that a C-compact order bounded orthogonally additive operator T : E → F from a Banach lattices E to a σ-Dedekind complete Banach lattice F is AM-compact if, in addition, T is uniformly continuous on order bounded subsets of E [13, Theorem 3.4]. The norm in a normed vector lattice is order continuous if xα ↓ 0 implies kxαk ↓ 0. We point out that a Banach lattice with order continuous norm is Dedekind complete (see [4, Theorem 12.9]). Now we are ready to present the first main result of the article. 6 MARAT PLIEV and MARTIN R. WEBER Theorem 3.9. Let be E a vector lattice and F a Banach lattice with order contin- uous norm. Then the set of all C-compact regular orthogonally additive operators from E to F is a projection band in OAr(E, F ). In order to prove Theorem 3.9 we need some auxiliary propositions. Proposition 3.10. [16, Proposition 3.9] Let E be a vector lattice and x, y ∈ E. Then x ⊑ y if and only if x+ ⊑ y+ and x− ⊑ y−. Proposition 3.11. Let E be a vector lattice and yk for some (xi)n i=1 k=1 ⊂ E. Then there exist a family of pairwise disjoint elements (zik) ⊂ and (yk)m E, where i ∈ {1, . . . , n} and k ∈ {1, . . . , m} such that n m xi = Fk=1 Fi=1 (i) xi = (ii) yk = zik for any i ∈ {1, . . . , n}; zik for any k ∈ {1, . . . , m}; (iii) n Fi=1 zik = xi = n Fi=1 yk. m Fk=1 m n Fk=1 Fi=1 Fk=1 m Proof. Using Proposition 3.10, and applying induction arguments we have n Gi=1 Gi=1 n x+ i = (cid:16) i = (cid:16) x− n Gi=1 Gi=1 n xi(cid:17)+ xi(cid:17)− = (cid:16) = (cid:16) m Gk=1 Gk=1 m yk(cid:17)+ yk(cid:17)− = = m Gk=1 Gk=1 m y+ k ; y− k . Now, by the Riesz Decomposition property there exist z+ ik, z− ik ∈ E+, i ∈ {1, . . . , n} and k ∈ {1, . . . , m} such that m m x+ i = y+ k = Gk=1 Gi=1 n Gk=1 Gi=1 n z+ ik, x− i = z− ik, i ∈ {1, . . . , n}, z+ ik, y− k = z− ik, k ∈ {1, . . . , m}. Set zik = z+ ik − z− ik, i ∈ {1, . . . , n} and k ∈ {1, . . . , m}. Hence n m Gi=1 Gk=1 zik = n Gi=1 xi = m Gk=1 yk and the proof is completed. ON C-COMPACT OPERATORS 7 Lemma 3.12. Let E and F be vector lattices with F Dedekind complete, S, T ∈ OAr(E, F ) and x ∈ E. Then the following equalities hold: (1) (2) (3) (T ∨ S)x = supn n Pi=1 (T ∧ S)x = inf n n Pi=1 T x = supn n Pi=1 T xi ∨ Sxi : x = T xi ∧ Sxi : x = n n xi, n ∈ No; Fi=1 xi, n ∈ No; Fi=1 T xi : x = n Fi=1 xi, n ∈ No. Proof. Since x ∧ y = −(−x) ∨ (−y) and x = x ∨ (−x) for proving the lemma it is sufficient to establish only the equation (1). Put A(x) := n n Xi=1 T xi ∨ Sxi : x = n Gi=1 xi, n ∈ No. We show that A(x) is an upward directed set. Indeed, take two disjoint decompo- sitions x = n Fi=1 decomposition x = xi and x = n m Fi=1 Fk=1 yk of x. By Proposition 3.11 there exists a disjoint m Fk=1 zik such that m xi = yk = Gk=1 Gi=1 n zik, i ∈ {1, . . . , n}, zik, k ∈ {1, . . . , m}. We observe that Sxi ∨ T xi ≤ m Xk=1 T zik ∨ Szik for any i ∈ {1, . . . , n}. Indeed m m m Szik ≤ T zik ∨ Szik, Sxi = S(cid:16) T xi = T(cid:16) zik(cid:17) = Gk=1 zik(cid:17) = Gk=1 Xk=1 Xk=1 m m Xk=1 Xk=1 m T zik ≤ T zik ∨ Szik =⇒ Sxi ∨ T xi ≤ m Xk=1 T zik ∨ Szik. Similar arguments show that Syk ∨ T yk ≤ n Xi=1 T zik ∨ Szik 8 MARAT PLIEV and MARTIN R. WEBER for any k ∈ {1, . . . , m}. Now we may write n Xi=1 T xi ∨ Sxi ≤ n m Xi=1 Xk=1 T zik ∨ Szik and m Xk=1 T yk ∨ Syk ≤ n m Xi=1 Xk=1 T zik ∨ Szik, and deduce that A(x) is an upward directed subset of F . By Proposition 2.3 OAr(E, F ) is a Dedekind complete vector lattice and (T ∨ S)x = sup{T y + Sz : x = y ⊔ z}. Clearly T x ∨ Sx ≤ (T ∨ S)x for any x ∈ E. Thus for any decomposition x = n xi, n ∈ N we have that Fi=1 n Xi=1 T xi ∨ Sxi ≤ n Xi=1 (T ∨ S)xi = (T ∨ S)x and therefore the set A(x) is upper order bounded. Put fx := sup A(x). Then fx ≤ (T ∨ S)x. On the other hand for any disjoint decomposition x = y ⊔ z we have T y + Sz ≤ supnT y ∨ Sy + T z ∨ Sz : x = y ⊔ zo ≤ n n supn T xi ∨ Sxi : x = Xi=1 Gi=1 xi, n ∈ No = fx. Passing to the supremum in the last inequality over all disjoint decompositions x = y ⊔ z of x, due to formula (1) of Proposition 2.3, we deduce (T ∨ S)x ≤ fx. It follows (T ∨ S)x = fx which completes the proof. Below we shall use the following elementary observation. Since n Xi=1 n Sxi ∨ T xi : x = n Gi=1 xi, n ∈ No is an upward downward directed subset of F the set n Xi=1 n Sxi ∧ T xi : x = n Gi=1 xi, n ∈ No is downward directed. Lemma 3.13. Let be E a vector lattice, F a Banach lattice with order continuous norm and T ∈ COAr(E, F ). Then FT ⊂ COAr(E, F ). Proof. We recall that by definition S ∈ FT if S ∧ T − S = 0. We show that S(Fx) is a relatively compact set2 in F for any S ∈ FT and x ∈ E. Indeed, fix 2 Due to the norm-completeness of F this is equivalent to its totally boundedness. ON C-COMPACT OPERATORS 9 x ∈ E and ε > 0. Since n n Pi=1 Sxi ∧ T − Sxi : x = n Fi=1 xi, n ∈ No is a downward n directed subset of F by Lemma 3.12 and the order continuity of the norm in the Banach lattice F there exists a disjoint decomposition x = xi of x such that Fi=1 n Xi=1 (cid:13)(cid:13)(cid:13) Sxi ∧ T − Sxi(cid:13)(cid:13)(cid:13) < ε 4 . Since T (Fxi) is a relatively compact set finite subset Di ⊂ Fxi such that for any w ∈ Fxi there is u ∈ Di satisfying for any i ∈ {1, . . . , n} there exists a kT w − T uk < ε 2n . Moreover, by Proposition 3.11 for any y ∈ Fx there exists a disjoint decomposition n y = yi, where yi ⊑ xi, i.e. yi ∈ Fxi. Hence, for some ui ∈ Di the following Fi=1 inequality kT yi − T uik < ε 2n holds. We remark that for any positive orthogonally additive operator G : E → F one has3 Gw ≤ Gxi for all w ∈ Fxi. Now, by taking into account the inequality x ≤ x + y + x ∧ y, which holds in any vector lattice4 we have Syi − Sui ≤ Syi − Sui + (T − S)yi − (T − S)ui + Syi − Sui ∧ (T − S)yi − (T − S)ui ≤ T yi − T ui + (Syi + Sui) ∧ ((T − S)yi + (T − S)ui) ≤ T yi − T ui + (Syi + Sui) ∧ (T − Syi + T − Sui) ≤ T yi − T ui + (Sxi + Sxi) ∧ (T − Sxi + T − Sxi) = T yi − T ui + 2(cid:16)Sxi ∧ T − Sxi(cid:17). Then we may write n Xi=1 Syi − Sui ≤ n Xi=1 T yi − T ui + 2 n Xi=1 (cid:16)Sxi ∧ T − Sxi(cid:17). 3 This immediately follows from y ∧ x − y = 0, the positivity of G and G(x) = G(y) + G(x − y) ≥ G(y). 4 Indeed, since (see [26], Theorems 5.1 and 5.5) x + y ≥ x − y ≥ (x − y)+ we have x − x ∧ y = x + (−x) ∨ (−y) = 0 ∨ (x − y) = (x − y)+ ≤ x + y. 10 MARAT PLIEV and MARTIN R. WEBER n Put D := {u = ui : ui ∈ Di}. Clearly D is a finite subset of Fx. Thus for any y ∈ Fx and u ∈ D we have Fi=1 n n n kSy − Suk = (cid:13)(cid:13)(cid:13)(cid:12)(cid:12)S(cid:16) Syi − Sui(cid:13)(cid:13)(cid:13) Xi=1 ≤ (cid:13)(cid:13)(cid:13) yi(cid:17) − S(cid:16) Gi=1 T yi − T ui(cid:13)(cid:13)(cid:13) n Xi=1 = (cid:13)(cid:13)(cid:13)(cid:12)(cid:12) Syi − Sui(cid:12)(cid:12)(cid:13)(cid:13)(cid:13) ui(cid:17)(cid:12)(cid:12)(cid:13)(cid:13)(cid:13) Gi=1 + 2(cid:13)(cid:13)(cid:13) Xi=1 (cid:16)Sxi ∧ T − Sxi(cid:17)(cid:13)(cid:13)(cid:13) n ≤ ≤ n (cid:13)(cid:13)(cid:13) Xi=1 Xi=1 n kT yi − T uik + ≤ ε. ε 2 Hence {Su : u ∈ D} is a finite ε-net for S(Fx), and therefore the proof is com- pleted. Let V be a vector lattice and v ∈ V+. The order ideal generated by v is denoted by Iv, i.e. Iv = {x ∈ V : ∃α > 0 such that x ≤ αv}. A v-step function is any vector s ∈ V for which there exist pairwise disjoint fragments v1, . . . , vn of vn i and real numbers λ1, . . . , λn satisfying s = λivi. The next v with v = Fi=1 n Pi=1 proposition is known as the Freudenthal Spectral Theorem ([4], Theorem 2.8). Proposition 3.14. Let V be a vector lattice with the principal projection property and let v ∈ V+. Then for every u ∈ Iv there exists a sequence (sn)n∈N of v-step functions satisfying 0 ≤ u − sn ≤ 1 n v for each n and sn ↑ u. Moreover, 0 ≤ sn if 0 ≤ u. Now we ready to prove the first main result. Proof of Theorem 3.9. We prove some properties of COAr(E, F ): a) Clearly, COAr(E, F ) is a vector subspace of OAr(E, F ). b) We show that COAr(E, F ) is even a vector sublattice of OAr(E, F ). Take S, T ∈ COAr(E, F ). Then T − S ∈ COAr(E, F ). By Lemma 3.13 FT ⊂ COAr(E, F ) and therefore T + ∈ COAr(E, F ). Therefore due to the equalities S + (T − S)+ = S + (T − S) ∨ 0 = T ∨ S, S ∧ T = −(−S) ∨ (−T ) which are valid in OAr(E, F ) we obtain that COAr(E, F ) is a sublattice of OAr(E, F ). c) Now we show that T ∈ COAr(E, F ) if 0 ≤ Tλ ↑ T in OAr(E, F ) and any Tλ ∈ COAr(E, F ). Indeed, take x ∈ E and ε > 0. Since the Banach lattice F is order continuous it follows from Tλx ↑ T x that kT x − Tλ0xk < ε 4 for some λ0. ON C-COMPACT OPERATORS 11 Then actually kT y − Tλ0yk < ε some z ∈ E. Then 4 for any y ∈ Fx. Indeed, consider x = y ⊔ z for 0 ≤ T x − Tλ0x = T (y ⊔ z) − Tλ0(y ⊔ z) = T y − Tλ0y + T z − Tλ0 z ≥ T y − Tλ0y implies kT y − Tλ0yk ≤ kT x − Tλ0xk. Since Tλ0 ∈ COAr(E, F ) there exists a finite subset D of Fx with the property that for any y ∈ Fx there exists u ∈ D satisfying So we obtain kTλ0u − Tλ0yk < ε 2 . kT u − T yk ≤ kT u − Tλ0u + Tλ0u − T y + Tλ0y − Tλ0yk ≤ kT u − Tλ0 uk + kT y − Tλ0 yk + kTλ0u − Tλ0yk < ε, what establishes the relative compactness of T (FT ) in F . d) Finally we prove that COAr(E, F ) is an order ideal in OAr(E, F ). Let 0 ≤ R ≤ T , where R ∈ OAr(E, F ) and T ∈ COAr(E, F ). Then R ∈ IT and by Proposition 3.14 there5 exists a sequence (Sn)n∈N in OAr(E, F ) of T - step-functions with 0 ≤ Sn ↑ R. Taking into account that6 Sn ∈ COAr(E, F ), n ∈ N and what has been established in c) we deduce that R ∈ COAr(E, F ). So, COAr(E, F ) is a band in OAr(E, F ). e) Due to the Dedekind completeness of OAr(E, F ) it is a projection band. 4. C-compact and narrow orthogonally additive operators In this section we consider a new class of vector lattices, where the condition of Dedekind completeness is replaced by a much weaker property. For laterally-to- norm continuous, C-compact orthogonally additive operators from a C-complete vector lattice E to a Banach space X we show their narrowness. Definition 4.1. A vector lattice E is said to be C-complete, if for each x ∈ E+ any subset D ⊂ Fx has a supremum and an infimum. Clearly, every Dedekind complete vector lattice E is C-complete. The reverse statement, in general, is not true. 5 This proposition can be applied since by Proposition 2.3 the vector lattice OAr(E, F ) is Dedekind complete. Any T -step-function S ∈ OAr(E, F ) has the form S = m λiTi, where Ti m Pi=1 are disjoint fragments of T such that T = Ti. Fi=1 6 This follows from the fact that together with T each fragment Ti of T belongs to COAr(E, F ). 12 MARAT PLIEV and MARTIN R. WEBER Proposition 4.2. The vector lattice E = C[0, 1] of all continuous functions on the interval [0, 1] is C-complete. ∞ Si=1 Proof. Fix f ∈ E+. Since the set Sf := {t ∈ [0, 1] : f (t) > 0} is open, by [10, Chap.2, Theorem 5] there is the decomposition Sf = (ai, bi) where (ai, bi) ∩ (aj, bj) = ∅, i 6= j. Take g ∈ Ff . We claim that for every i ∈ N either g(t) = 0 or g(t) = f (t) for any t ∈ (ai, bi). Indeed, denote by fi and gi the restrictions of f and g on the closed interval [ai, bi], respectively. It is clear that gi ∈ Ffi. Assume that there exists a nonzero fragment vi ∈ Ffi such that vi ⊥ gi and fi = gi + vi. Put Gi := {t ∈ (ai, bi) : g(t) > 0} and Vi = {t ∈ (ai, bi) : v(t) > 0}. Since fi is strictly positive on (ai, bi) we deduce that (ai, bi) = Gi ⊔ Vi and therefore Gi and Vi are open-closed subsets of (ai, bi). But the interval [ai, bi] is a connected set and we come to the contradiction. Thus fi has no fragment 0 < vi < fi and therefore, either fi(t) = gi(t) or gi(t) = 0 for any t ∈ (ai, bi). With each g ∈ Ff there is associated the sequence (gi)i∈N, where 1, 0, gi =   if g(t) = f (t) if g(t) = 0 t ∈ (ai, bi). Clearly, in this way a one-to-one correspondence between Ff and the set of all 0 - 1 sequences is established. Let D be any fixed subset of Ff . Put D+ := {i ∈ N : ∃ g ∈ D such that gi = 1}; D− := {i ∈ N : such that gi = 1 for ∀ g ∈ D}. Consider the pair of sequences (ui)i∈N and (vi)i∈N, where 1, 0, 1, 0, ui =   vi =   if i ∈ D+ otherwise, if i ∈ D− otherwise. With (ui)i∈N and (vi)i∈N there is associated the pair {u, v} of fragments of f . Clearly u = sup D and v = inf D. Since the vector lattice C[0, 1] is Archimedean but not Dedekind complete the previous proposition shows that the set of all C-complete vector lattices is a new subclass of vector lattices which strictly contains the class of all Dedekind complete vector lattices. We note that in general a C-complete vector lattice is not Archimedian. ON C-COMPACT OPERATORS 13 Example 4.3. Let E = R2 equipped with the lexicographic order. That is, we consider E as a vector lattice with the following order7 (x1, x2) ≥ (y1, y2), when- ever either x1 > y1 or else x1 = y1 and x2 ≥ y2. The vector lattice E is not Archimedian. On the other hand it is not hard to verify that the Boolean algebra of all fragments Fx of an arbitrary element x = (x1, x2) ∈ E+ contains only two elements: Hence E is C-complete. Fx = {(x1, x2), (0, 0)}. We say that a set D ⊂ E is laterally bounded, if there exits x ∈ E such that D ⊂ Fx. We say that a laterally bounded set D has a lateral supremum (infimum) if there exists u ∈ E (v ∈ E) such that u = sup D (v = inf D) with respect to the partial order ⊑ in Fx. Taking into account Proposition 3.10 we deduce that a vector lattice E is C-complete if and only if every laterally bounded subset of E has the lateral supremum and infimum. Definition 4.4. Let E be a vector lattice and X be a normed space. An or- thogonally additive operator T : E → X is called narrow, if for any x ∈ E and ε > 0 there exists a pair x1, x2 of mutually complemented fragments of v, such that kT x1 − T x2k < ε. In particular, if X = R, we call T a narrow functional. Observe that the image of an atom under a narrow operator T is zero. Indeed. The only disjoint fragments of an atom a are 0 and a. So, due to the narrowness of T , for any ε > 0 one has kT uk < ε, what means T u = 0. This is the reason for supposing the vector lattice E to be atomless in the Theo- rems 4.7 and in most of the propositions of the current section. Example 4.5. Let (Ω, Σ, µ) be a σ-finite measure space. Consider a map N : L1(µ) → R defined by N (f ) = kf kL1(µ), f ∈ L1(µ). In [23] (Proposition 2.5) it was shown that N is a narrow orthogonally additive functional on L1(µ). A net (xα)α∈Λ in a vector lattice E laterally converges to x ∈ E if xα ⊑ xβ ⊑ x for all α ≤ β and xα order converges to x. This is written as xα lat−→ x. 7 The cone E+ in this vector lattice consists of the open right half-space {(x1, x2) : x1 > 0} joint with the half-ray {(x1, x2) : x1 = 0, x2 ≥ 0}. 14 MARAT PLIEV and MARTIN R. WEBER Definition 4.6. An orthogonally additive operator T from a vector lattice E to a normed space X is called laterally-to-norm continuous whenever for each lat−→ x the net (T xα)α∈Λ converges with laterally convergent net (xα)α∈Λ with xα respect to the norm to in X to T x. The following theorem is the second main result of the article. Theorem 4.7. Let E be an atomless C-complete vector lattice and X be a Banach space. Then every orthogonally additive laterally-to-norm continuous C-compact operator T : E → X is narrow. The next auxiliary proposition is well known (see e.g. [21, Lemma 10.20]). Proposition 4.8. Let (vi)n normed space V and (λi)n i. Then there exists a set (θi)n and i=1 be a finite subset of elements in a finite dimensional i=1 be non-negative numbers such that 0 ≤ λi ≤ 1 for each i=1 of numbers such that θi ∈ {0, 1}, i ∈ {1, . . . , n} n Xi=1 (cid:13)(cid:13)(cid:13) (λi − θi) vi(cid:13)(cid:13)(cid:13) ≤ dim V 2 max i∈{1,...,n} kvik. Proposition 4.9. Let be E an atomless vector lattice, x ∈ E, X a Banach space and T : E → X an orthogonally additive laterally-to-norm continuous operator. Then for any ε > 0 there exists a decomposition x = y ⊔ z, where y, z are nonzero fragments of x such that kT zk < ε. Proof. Since E is an atomless vector lattice the Boolean algebra A := Fx has infinite cardinality. We note that the set of all fragments of x is the net (xα)α∈A laterally converges to x. The laterally-to-norm continuity of T implies the exis- tence of α0 ∈ A such that kT x − T xαk < ε for all α ≥ α0. Consider for α0 the disjoint decomposition x = (x − xα0) ⊔ xα0. Then T x = T(cid:0)(x − xα0) ⊔ xα0(cid:1) = T (x − xα0) + T xα implies kT (x − xα0)k < ε. Assign y := xα0 and z = x − xα0. Then kT zk < ε and x = y ⊔ z is the desirable disjoint decomposition. Proposition 4.10. Let be E an atomless vector lattice, x ∈ E, X a Banach space and T : E → X an orthogonally additive laterally-to-norm continuous operator. Assume that (yn)n∈N is a sequence of fragments of x such that y1 = x, yn ⊑ ym for n ≥ m; m, n ∈ N and Tn∈N Fyn = {0}. Then lim n→∞ kT ynk = 0. ON C-COMPACT OPERATORS 15 Proof. By definition ym = yn ⊔ (ym − yn) for m, n ∈ N with n ≥ m. For xn := x − yn, n ∈ N with yn = ym − (ym − yn) we can write xn = x − ym + (ym − yn) = x − ym + ((x − yn) − (x − ym)) = xm + (xn − xm). Clearly xm ⊥ (xn − xm) for n, m ∈ N and n ≥ m and therefore xn ⊥ (x − xn) for Fyn = {0}. Hence all n ∈ N. Thus x = xn ⊔ (x − xn), n ∈ N and Tn∈N we deduce that the sequence (xn)n∈N laterally converges8 to x. Now the relation xn ⊥ (x − xn) for all n ∈ N implies Fx−xn = Tn∈N T x = T(cid:0)xn ⊔ (x − xn)(cid:1) = T xn + T (x − xn) and T (x − xn) = T x − T xn. Thus by laterally-to-norm continuity of T we have that T xn is norm convergent to T x in X and lim n→∞ kT x − T xnk = lim n→∞ kT ynk = 0. Proposition 4.11. Let be E an atomless C-complete vector lattice, X a Banach space, T : E → X an orthogonally additive laterally-to-norm continuous operator, x ∈ E and ε > 0. Then for some n ∈ N there exists a decomposition x = n where xi are nonzero fragments of x such that kT xik ≤ ε for any i ∈ {1, . . . , n}. xi, Fi=1 Proof. By Proposition 4.9 the set Dx,T,ε := {z ∈ Fx : z 6= 0, kT zk ≤ ε} is not empty. We note that Dx,T,ε is a partially ordered set with respect to the relation ⊑. Let (uλ)λ∈Λ ⊆ Dx,T,ε be a chain, where Λ is a some linearly ordered index set. Clearly uµ ⊑ uλ for all µ, λ ∈ Λ, µ ⊑ λ. By the C-completeness of (uλ), where u ∈ Fx. Due to kT uλk ≤ ε vector lattice E there exists u = (o)-lim λ∈Λ and the laterally-to-norm continuity of T we have that then T u = lim (T uλ) λ∈Λ and therefore u ∈ Dx,T,ε. By Zorn's Lemma, there is a maximal element9 z ∈ Dx,T,ε with kT zk ≤ ε. Put y = x − z. If kT yk ≤ ε then we got the required decomposition of x. Otherwise we apply Proposition 4.9 to y and get y = y1 ⊔ y2, where y1 is a maximal element in Dy,T,ε with kT y1k ≤ ε. In case of necessity, i.e. if kT y2k > ε, by further continuing in the same way with the corresponding fragments y2, y4, . . . , y2k, . . . of y we construct a sequence of decompositions y2k = y2k+1 ⊔ y2k+2, where y2k+1 is a maximal element in Dy2k,T,ε and satisfying the 8 Indeed, if some vector 0 6= u is a fragment of each vector x − xn then u ⊑ (x − xn) and x − xn ≤ un for some sequence un ↓ 0 is impossible. 9 The element z is maximal in Dx,T,ε, if ∄ u ∈ Dx,T,ε such that z ⊑ u and z 6= u. 16 MARAT PLIEV and MARTIN R. WEBER conditions kT y2k+1k ≤ ε and kT y2k+2k > ε, k ∈ N. We claim that there exists l ∈ N such that y2l = y2l+1 ⊔ y2l+2 and both kT y2l+1k, kT y2l+2k ≤ ε. Assume the contrary. Then the sequence (y2k)k∈N of fragments of y is such that kT y2kk > ε Fy2k = {0}. Indeed, assume that there for all k ∈ N. Nevertheless we show Tk∈N exists a nonzero element v ∈ Tk∈N 2 = y2 − v, y′ y′ 2m for m, n ∈ N with n ≥ m and Tn∈N 4 = y4 − v, . . . , y′ we have y′ 2n ⊑ y′ Fy2k. Then for the sequence 2k = y2k − v, . . . Fy′ 2n = {0}. According to Proposition 4.10 there exists n0 ∈ N such that kT y′ y′ 2n0 ∈ Dy2n0 ,T,ε and y2n0+1 is a maximal element of Dy2n0 ,T,ε. We have 2n0k < ε. Thus y2n0 = y′ 2n0 ⊔ v, y2n0 = y2n0+1 ⊔ y2n0+2 = y′ 2n0 ⊔ v. Consider now the two cases: Fy2n0+1 ∩ Fv = {0} and Fy2n0+1 ∩ Fv 6= {0}. 2n0, y2n0+1 6= y′ In the first case, since no non-zero fragment of v is a fragment of y2n0+1, we get y2n0+1 ⊑ y′ 2n0. Observe that - due to the maximality of y2n0+1 in Dy2n0 ,T,ε - the relation y2n0+1 ⊑ y′ 2n0 = y2n0+1, i.e. y2n0+2 = v and, therefore v is not a fragment of y2n0+4. In the second one there is a nonzero fragment w ∈ Fy2n0+1 ∩ Fv. Thus v = (v − w) ⊔ w and w ⊑ y2n0+1, i.e. the non-zero fragment w of v belongs to Fy2n0+1. Therefore the element v can not be a fragment of y2n0+2. 2n0 is impossible. Hence y′ Fy2k = {0} and so, again by applying Proposition 4.10, we get Hence Tk∈N kT y2kk = 0. This is a contradiction and therefore the desirable l ∈ N exists. lim k→∞ Consider the following elements x1 = y1, x2 = y3, . . . , xl = y2l−1, xl+1 = y2l+1, xl+2 = y2l+2, xl+3 = z. Then x = n Fi=1 xi with n = l + 3 is the desirable decomposition of x. Proposition 4.12. Let E be an atomless C-complete vector lattice and V a finite dimensional Banach space. Then every orthogonally additive laterally-to-norm continuous operator G : E → V is narrow. Proof. Fix any x ∈ E and ε > 0. According to Proposition 4.11 there is a disjoint dim V for any i ∈ {1, . . . , n}. Then by xi such that kGxik < ε decomposition x = n Fi=1 ON C-COMPACT OPERATORS 17 using Proposition 4.8 for λi = 1 such that 2 there exist numbers θi ∈ {0, 1} for i ∈ {1, . . . , n} n 2 Xi=1(cid:16)1 2(cid:13)(cid:13)(cid:13) Observe that for I0 = (cid:8)i ∈ {1, . . . , n(cid:9) : θi = 0(cid:9) and I1 = (cid:8)i ∈ {1, . . . , n} : θi = 1(cid:9) the vectors yk = Fi∈Ik xi for k ∈ {0, 1} are mutually complemented fragments of x − θi(cid:17) Gxi(cid:13)(cid:13)(cid:13) (4.1) ≤ dim V max kGxik < ε. i∈{1,...,n} and by (4.1), kGy1 − Gy2k = (cid:13)(cid:13)(cid:13) Xi∈I0⊔I1 hence the operator G is narrow. (1 − 2θi) Gxi(cid:13)(cid:13)(cid:13) < ε, Definition 4.13. Let E be a vector lattice and F a vector space. An orthogonally additive operator T : E → F is called of finite rank if the set T (E) generates a finite-dimensional subspace in F . Now we are in the position to prove the second main result (Theorem 4.7). Before, however, we notice that a Banach space X can be considered as a closed subspace of the Banach space W := l∞(BX ∗) = {q : BX ∗ → R, sup q(f ) < ∞, f ∈ BX ∗} (equipped with the supremum-norm) of all real-valued bounded functions on the closed unit ball BX ∗ of the dual space X ∗ according to X ֒→ X ∗∗ ֒→ W, where the notation ֒→ means the isometric embedding X ∋ x 7→ Fx ∈ X ∗∗ given by Fx(f ) := f (x), f ∈ BX ∗ . Observe that, if H is a relatively compact subset of the Banach space W and ε > 0, then there exists a linear finite rank operator R ∈ L(W ) such10 that kw − Rwk ≤ ε for every w ∈ H [21, Lemma 10.25]. Proof of Theorem 4.7. Fix an arbitrary x ∈ E and ε > 0. Due to the C- compactness of T the set K = T (Fx) is relatively compact in X and therefore in W . It follows that there exists a finite rank operator S ∈ L(W ) such that ky − Syk ≤ ε 4 for every y ∈ K. Then G = S ◦ T is an orthogonally additive laterally-to-norm continuous finite rank operator. By Lemma 4.12 G is narrow 10 By L(W ) there is denoted the space of all linear bounded operators on W . 18 MARAT PLIEV and MARTIN R. WEBER and consequently, there exist mutually complemented fragments x1, x2 of x such that kGx1 − Gx2k < ε 2. Therefore, kT x1 − T x2k = kT x1 − T x2 + S(T x1) − S(T x2) − S(T x1) + S(T x2)k = kT x1 − T x2 + Gx1 − Gx2 − Gx1 + Gx2k ≤ kGx1 − Gx2k + kT x1 − Gx1k + kT x2 − Gx2k < ε 2 + ε 2 = ε since kT xi − Gxik < ε 4 for i ∈ {1, 2}. This completes the proof. We remark that Theorem 4.7 generalizes the result of the article [18, Theo- rem 3.2]. Acknowledgments. Marat Pliev was supported by the Russian Foundation for Basic Research (grant number 17-51-12064). Martin Weber was supported by the Deutsche Forschungsgemeinschaft (grant number CH 1285/5-1, Order pre- serving operators in problems of optimal control and in the theory of partial differential equations). References [1] N. Abasov, M. Pliev, On extensions of some nonlinear maps in vector lattices, J. Math. Anal. Appl. 455 (2017), 516 -- 527. [2] N. Abasov, M. Pliev, Disjointness preserving orthogonally additive operators in vector lattices, Banach of Math. Anal. 12 (2018), 3, 730 -- 750. [3] N. Abasov, M. Pliev, Dominated orthogonally additive operators in lattice-normed spaces, Advances in Operator Theory, 4 (2019), 1, 251 -- 264. [4] C. D. Aliprantis, O. Burkinshaw, Positive Operators, Springer, Dordrecht. (2006). [5] J. A. Appell, P. P. Zabrejko, Nonlinear superposition operators, Cambridge University Press, 2008. [6] W. A. Feldman, Lattice preserving maps on lattices of continuous functions, J. Math. Anal. Appl., 404 (2013), 310 -- 316. [7] W. A. Feldman, A characterization of non-linear maps satisfying orthogonality properties, Positivity 21 (2017), 1, 85 -- 97. [8] W. A. Feldman, A factorization for orthogonally additive operators on Banach lattices, J. Math. Anal. and Appl., 472, (2019), 1, 238 -- 245. [9] O. Fotiy, A. Gumenchuk, I. Krasikova, M. Popov, On sums of narrow and compact operators, Positivity, doi.org/10.1007/s11117-019-00666-4. [10] A. N. Kolmogorov, S. V. Fomin, Introductory Real Analysis, Dover Publication. (1975). [11] O. Maslyuchenko, V. Mykhaylyuk, M. Popov, A lattice approach to narrow operators, Positivity, 13 (2009), 459 -- 495. [12] J. M. Maz´on, S. Segura de Le´on, Order bounded ortogonally additive operators, Rev. Roumane Math. Pures Appl., 35 (1990), 4, 329 -- 353. ON C-COMPACT OPERATORS 19 [13] J. M. Maz´on, S. Segura de Le´on, Uryson operators, Rev. Roumane Math. Pures Appl. 35 (1990), 5, 431 -- 449. [14] V. Mykhaylyuk, M. Popov, On sums of narrow operators on Kothe function space, J. Math. Anal. Appl. 404 (2013), 554 -- 561. [15] V. Mykhaylyuk, On the sum of narrow and a compact operators, J. Funct. Anal. 266 (2014), 5912 -- 5920. [16] V. Orlov, M. Pliev, D. Rode, Domination problem for AM -compact abstract Uryson oper- ators, Archiv der Mathematik, 107, 5, (2016), 543 -- 552. [17] Plichko A., M.Popov M. Symmetric function spaces on atomless probability spaces, Disser- tationes Math. (Rozprawy Mat.), 306 (1990), 1 -- 85. [18] M. Pliev, M. Popov, Narrow orthogonally additive operators, Positivity, 18 (2014), 4, 641 -- 667. [19] M. Pliev, K. Ramdane Order unbounded orthogonally additive operators in vector lattices, Mediterranean J. Math. 15 (2018), 2, 20 p. [20] M. Pliev, X. Fang, Narrow orthogonally additive operators in lattice-normed spaces, Siberian Math. J., 58 (2017), 1, 134 -- 141. [21] M. Popov, B. Randrianantoanina, Narrow Operators on Function Spaces and Vector Lat- tices, De Gruyter Studies in Mathematics 45, De Gruyter (2013). [22] M. A. Pliev, M. R. Weber, Disjointness and order projections in the vector lattices of ab- stract Uryson operators, Positivity, 20 (2016), 3, 695 -- 707. [23] M. A. Pliev, F. Polat, M. R. Weber, Narrow and C-compact orthogonally additive operators in lattice-normed spaces, Results in Math., 74 (2019), 4, 19 pp. [24] Tradacete, P., and Villanueva, I. Continuity and representation of valuations on star bodies, Adv. Math. 329 (2018), 361 -- 391. [25] Tradacete, P., and Villanueva, I. Valuations on Banach lattices, Int. Math. Res. Not., (2019), to appear. [26] A. C. Zaanen, Introduction to Operator Theory in Riesz Spaces, Springer, Berlin Heidelberg New York. (1997). Southern Mathematical Institute of the Russian Academy of Sciences, Vladikavkaz, 362027 Russia E-mail address: [email protected] Technical University Dresden, Department of Mathematics, Institute of Anal- ysis, 01062 Dresden, Germany E-mail address: [email protected]
1112.5262	2	1112	2012-07-18T12:27:28	Nonstationary Gabor Frames - Existence and Construction	[ "math.FA" ]	Nonstationary Gabor frames were recently introduced in adaptive signal analysis. They represent a natural generalization of classical Gabor frames by allowing for adaptivity of windows and lattice in either time or frequency. In this paper we show a general existence result for this family of frames. We then give a perturbation result for nonstationary Gabor frames and construct nonstationary Gabor frames with non-compactly supported windows from a related painless nonorthogonal expansion. Finally, the theoretical results are illustrated by two examples of practical relevance.	math.FA	math	Nonstationary Gabor Frames - Existence and Construction Monika Dorfler, Ewa Matusiak∗ Department of Mathematics, NuHAG, University of Vienna, Austria Abstract Nonstationary Gabor frames were recently introduced in adaptive signal analysis. They represent a natural generalization of classical Gabor frames by allowing for adaptivity of windows and lattice in either time or frequency. In this paper we show a general existence result for this family of frames. We then give a perturbation result for nonstationary Gabor frames and construct nonstationary Gabor frames with non-compactly supported windows from a related painless nonorthogonal expansion. Finally, the theoretical results are illus- trated by two examples of practical relevance. Keywords: adaptive representations, nonorthogonal expansions, irregular Gabor frames, existence 1. Introduction The principal idea of Gabor frames was introduced in [14] with the aim to represent signals in a time-frequency localized manner. Since the work of Gabor himself, a lot of research has been done on the topic of atomic time-frequency representation. While it turned out that the original model proposed by Gabor does not yield stable representations in the sense of frames [4, 7, 10], the existence of Gabor frames was first established in the so called painless case, [7], which requires the use of compactly supported analysis windows. The existence of Gabor frames in more general situations was proved later [19, 23] and the proof often uses an argument invoking the invertibility of diagonally dominant matrices. Various irregular and adaptive versions of Gabor frames have been introduced over the years, cf. [1, 5, 12, 21]. In these approaches, the irregularity usually concerns either the sampling set, which is allowed to deviate from a lattice, or the window, which is allowed to be modified. In [1], varying windows as well as irregular sampling points are allowed, however, existence of a local frame is assumed, from which a global frame is constructed. Nonstationary Gabor frames give up the strict regularity of the classical Gabor setting, but, as opposed to irregular frames, maintain enough structure to guarantee efficient implementa- tion and, possibly approximate, efficient reconstruction. In analogy to the classical, regular case [7], painless nonstationary Gabor frames were introduced in [2], where the principal idea is described and illustrated in detail. The construction of painless nonstationary Gabor ∗Corresponding author Email addresses: [email protected] (Monika Dorfler), [email protected] (Ewa Matusiak) Preprint submitted to Elsevier April 30, 2014 frames is similar to, but more flexible than the construction of windowed modified cosine transforms and other lapped transforms [17, 24] that allow for adaptivity of the window length. An efficient and perfectly invertible constant-Q transform was recently introduced using nonstationary Gabor transforms [22]. In this and similar situations, redundancy of the transform is crucial, since non-redundant versions of the constant-Q transform lead to dyadic wavelet transforms, which are often inappropriate for audio signal processing. Redundancy allows for good localization of both the analysis and synthesis windows, and their respective Fourier transforms and often promote sparse representations in adaptive processing. Painless non-orthogonal expansions can only be devised if the involved analysis windows are either compactly supported or band limited. This requirement may sometimes be too re- strictive. For instance, one may be interested in designing frequency-adaptive nonstationary Gabor frames with windows that are compactly supported in time, i.e. can be implemented as FIR filters, cf. [11] and lend themselves to real-time implementation, cp. [9]. The present contribution addresses the case of nonstationary Gabor frames with more general windows than used in the painless case. In Theorem 3.4, we derive the existence of nonstationary Gabor frames directly from a generalized Walnut representation: under mild uniform decay conditions on all windows involved, we show an existence result of nonstationary Gabor frames in parallel to the result given in [23] for regular Gabor frames, also compare [15, Theorem 6.5.1]. Note that the existence of a different class of nonstationary Gabor frames, the quilted Gabor frames, [8], was recently proved in the general context of spline type spaces in the remarkable paper [18]. This paper is organized as follows. In the next section, we introduce notation and state some auxiliary results. In Section 3, we first define nonstationary Gabor frames and recall known results for the painless case. In Section 3.2 a Walnut representation and a correspond- ing bound of the frame operator in the general setting is derived and Section 3.3 provides the existence of nonstationary Gabor frames. In Section 3.4 we pursue two basic approaches for the construction of nonstationary Gabor frames . Using tools from the theory of perturbation of frames, we construct nonstationary frames from an existing frame in Proposition 3.7. In Corollary 3.8 we design nonstationary Gabor frames by exploiting knowledge about a related painless frame, to obtain ”almost painless nonstationary Gabor frames”. In Section 4 we provide examples based on the two introduced construction principles. 2. Notation and Preliminaries Given a non-zero function g ∈ L2(R), let gk,l(t) = MblTakg(t) := e2πibltg(t − ak). Mbl is a modulation operator, or frequency shift, and Tak is a time shift. The set G(g, a, b) = {gk,l : k, l ∈ Z} is called a Gabor system for any real, positive a, b. G(g, a, b) is a Gabor frame for L2(R), if there exist frame bounds 0 < A ≤ B < ∞ such that for every f ∈ L2(R) we have To every Gabor system, we associate the analysis operator Cg given by (Cgf )k,l = (cid:104)f, gk,l(cid:105), k,l∈Z ck,lγk,l for c ∈ (cid:96)2. The operator and the synthesis operator Uγ = C∗ A(cid:107)f(cid:107)2 (cid:104)f, gk,l(cid:105)2 ≤ B(cid:107)f(cid:107)2 2 . (1) 2 ≤ (cid:88) γ , given by Uγc =(cid:80) k,l∈Z 2 Sg,γ associated to G(g, a, b) and G(γ, a(cid:48), b(cid:48)), where Sg,γ = UγCg reads (cid:88) k,l∈Z Sg,γf = (cid:104)f, gk,l(cid:105)γk,l . The inequality (1) is equivalent to the invertibility and boundedness of the frame operator Sg,g of G(g, a, b). window g ∈ L2(R), the STFT of f ∈ L2(R) is The analysis operator is the sampled short-time Fourier transform (STFT). For a fixed Vgf (x, ω) = f (t)e−2πiωtg(t − x) dt = (cid:104)f, MωTxg(cid:105) . Setting (x, ω) = (ak, bl), leads to Vgf (ak, bl) = (Cgf )k,l. When working with irregular grids, we assume that the sampling points form a separated set: a set of sampling points {ak : k ∈ Z} is called δ-separated, if ak − am > δ for ak, am, whenever k (cid:54)= m. χI will denote the characteristic function of the interval I. A convenient class of window functions for time-frequency analysis on L2(R) is the Wiener space. Definition 2.1. A function g ∈ L∞(R) belongs to the Wiener space W (L∞, (cid:96)1) if (cid:107)g(cid:107)W (L∞,(cid:96)1) := ess supt∈Qg(t + k) < ∞ , Q = [0, 1] . (cid:90) R (cid:88) (cid:88) k∈Z k∈Z For g ∈ W (L∞, (cid:96)1) and δ > 0 we have [15] ess supt∈R g(t − δk) ≤ (1 + δ−1)(cid:107)g(cid:107)W (L∞,(cid:96)1) . (2) In dealing with polynomially decaying windows, we will repeatedly use the following lemma. Lemma 2.2. For p > 1 the following estimates hold: (a) Let δ > 0, then (b) Let {ak : k ∈ Z} ⊂ R be a δ−separated set. Then (1 + δk)−p ≤ (1 + δ)−p(δ−1 + p)(p − 1)−1 . ∞(cid:88) (cid:88) (1 + t − ak)−p ≤ 2(cid:0)1 + (1 + δ)−p(δ−1 + p)(p − 1)−1(cid:1) . k=1 ess supt∈R k∈Z Proof. To show (a) we write (1 + δk)−p = (1 + δ)−p + ∞(cid:88) k=1 ∞(cid:88) k=2 (1 + δk)−p = (1 + δ)−p + 3 (cid:90) ∞(cid:88) k=2 [0,1]+k (1 + δk)−p dt . for t ∈ [k, k + 1], we have δt ≤ δ(k + 1) which implies that 1 + δ(t − 1) ≤ 1 + δk. Therefore, (cid:90) ∞(cid:88) (cid:90) ∞ (1 + δk)−p dt ≤ (1 + δ(t − 1))−p dt = (1 + δ(t − 1))−p dt k=2 [0,1]+k k=2 [0,1]+k = (1 + δ)−p+1δ−1(p − 1)−1 , 2 (cid:90) ∞(cid:88) (cid:88) k∈Z and the estimate follows. To prove (b), fix t ∈ R. Since ak − al > δ for k (cid:54)= l, each interval of length δ contains at most one point t − ak, k ∈ Z. Therefore we may write t − ak = δnk + xk for unique nk ∈ Z and xk ∈ [0, δ), and by the choice of δ, we have nk (cid:54)= nl for k (cid:54)= l. We assume, without loss of generality, that nk = 0 for k = 0, and we find that (1 + t − ak)−p = (1 + δnk + xk)−p (cid:88) k∈Z ≤ 1 + ≤ 1 + ≤ 1 + (cid:32) k∈Z ; nk>0 (cid:88) (cid:88) ∞(cid:88) ∞(cid:88) k=1 k∈Z ; nk>0 (1 + δk)−p + = 2 1 + (1 + δk)−p (cid:88) (cid:88) k∈Z ; nk>0 (1 + δnk + xk)−p + (1 + δnk − xk)−p (1 + δnk)−p + (1 + δnk − δ)−p k∈Z ; nk>0 (1 + δ(k − 1))−p ∞(cid:88) (cid:33) ≤ 2(cid:0)1 + (1 + δ)−p(δ−1 + p)(p − 1)−1(cid:1) . k=1 k=1 The last expression is independent of t, and the claim follows. Remark 1. When the set A = {ak : k ∈ Z} ⊂ R is relatively δ−separated, meaning then the estimate (b) in Lemma 2.2 becomes rel(A) := max t∈R #{A ∩ ([0, δ] + t)} < ∞ , (1 + t − ak)−p ≤ 2 rel(A)(cid:0)1 + (1 + δ)−p(δ−1 + p)(p − 1)−1(cid:1) . (cid:88) k∈Z ess supt∈R Notice, that for a separated set, rel(A) = 1. 3. Nonstationary Gabor frames Nonstationary Gabor systems provide a generalization of the classical Gabor systems of time-frequency-shifted versions of a single window function. Definition 3.1. Let g = {gk ∈ L2(R) : k ∈ Z} be a set of window functions and let b = {bk : k ∈ Z} be a corresponding sequence of frequency-shift parameters. Set gk,l = Mbklgk. Then, the set G(g, b) = {gk,l : k, l ∈ Z} is called a nonstationary Gabor system. 4 Note that, conceptually, we assume that the windows gk are centered at points {ak : k ∈ Z}, in direct generalization of the regular case, where gk(t) = g(t − ak) for some time- shift parameter a. In this sense, we have a two-fold generalization: the sampling points can be irregular and the windows can change for every sampling point. We are interested in conditions under which a nonstationary Gabor system forms a frame. We first recall the case of nonstationary Gabor frames with compactly supported windows, see [2] and http://www.univie.ac.at/nonstatgab/ for further information. 3.1. Compactly supported windows: the painless case Based on the support length of the windows gk, we can easily determine frequency-shifts parameters bk, for which we obtain a frame. The following result is the nonstationary version of the result given in [7]. Proposition 3.2 ([2]). Let g = {gk ∈ L2(R) : k ∈ Z} be a collection of compactly supported functions with supp gk ≤ 1/bk. Then G(g, b) is a frame for L2(R) if there exist constants A > 0 and B < ∞ such that A ≤ G0(t) = k gk(t)2 ≤ B a.e. . b−1 k∈Z The dual atoms are then γk,l(t) = MlbkG−1 Remark 2. An analogous theorem holds for bandlimited functions gk. 0 (t)gk(t). The above theorem follows from the fact that the frame operator associated to the col- lection of atoms described in the theorem can be written as Sg,gf (t) = k gk(t)2f (t) a.e. . b−1 The diagonality of the frame operator in the painless case is derived from a generalized Walnut representation for the frame operator Sg,g of nonstationary Gabor frames. In the next section we will see that this representation immediately implies diagonality of Sg,g under the assumptions of Proposition 3.2. 3.2. A Walnut representation for nonstationary Gabor Frames Let us now consider nonstationary Gabor systems G(g, b) and G(γ, b), with all windows gk and γk in W (L∞, (cid:96)1). The operator associated to G(g, b) and G(γ, b) reads Sg,γf = (cid:104)f, Mlbkgk(cid:105)Mlbkγk . Proposition 3.3. The operator Sg,γ in (3) admits a Walnut representation (cid:88) k,l∈Z Sg,γf = k,l · Tlb−1 Gg,γ k f , where Gg,γ k,l (t) = b−1 k gk(t − lb−1 k )γk(t) , 5 (cid:88) (cid:88) k∈Z (cid:88) k,l∈Z (3) (4) for f ∈ L2(R). Moreover, its operator norm can be bounded by (cid:104)Sg,γf, h(cid:105) ≤(cid:16) ·(cid:16) (cid:17)1/2(cid:16) k )(cid:107)γk(cid:107)W (L∞,(cid:96)1) (cid:17)1/2(cid:16) k )(cid:107)gk(cid:107)W (L∞,(cid:96)1) (1 + b−1 sup k∈Z (1 + b−1 sup k∈Z ess supt∈R ess supt∈R k∈Z gk(t)(cid:17)1/2 (cid:88) γk(t)(cid:17)1/2(cid:107)f(cid:107)2(cid:107)h(cid:107)2 (cid:88) k∈Z (5) for all f, h ∈ L2(R). Proof. First assume that f, h ∈ L2(R) are compactly supported. Since (cid:104)f, Mlbkgk(cid:105) = (cid:91)(f ¯gk)(lbk) we can write Sg,γ as Sg,γf (t) = (cid:91)(f ¯gk)(lbk)Mlbkγk(t) = mk(t)γk(t) , (6) (cid:88) k,l∈Z (cid:88) k∈Z where mk(t) =(cid:80) l∈Z and by the Poisson formula can be written as (cid:91)(f ¯gk)(lbk)e2πilbkt, for every k ∈ Z. The functions mk are b−1 k periodic mk(t) = b−1 k (f ¯gk)(t − lb−1 k ) . (7) (cid:88) l∈Z Therefore, substituting (7) in (6) yields the Walnut representation. We next prove the boundedness (5). In the following chain of inequalities, we will use Cauchy-Schwartz inequality for sums and integrals and, since all summands have absolute value, Fubini’s theorem to justify changing the order of summation and integral. We thus find (cid:104)Sg,γf, h(cid:105) = k )γk(·)f (· − lb−1 k gk(· − lb−1 b−1 k,l∈Z k R k,l∈Z b−1 (cid:12)(cid:12)(cid:12)(cid:68)(cid:88) (cid:90) ≤ (cid:88) (cid:20)(cid:90) ≤ (cid:88) (cid:34)(cid:88) (cid:90) (cid:34)(cid:90) f (t)2(cid:88) b−1 b−1 k,l∈Z k,l∈Z ≤ = R R k k R k,l∈Z gk(t − lb−1 k )γk(t)f (t − lb−1 gk(t − lb−1 k )γk(t)f (t − lb−1 gk(t)γk(t + lb−1 k )f (t)2 dt k gk(t)γk(t − lb−1 b−1 k ) dt k ), h (cid:69)(cid:12)(cid:12)(cid:12) k )h(t) dt (cid:21)1/2(cid:20)(cid:90) (cid:35)1/2(cid:34)(cid:88) (cid:35)1/2(cid:34)(cid:90) k )2 dt k,l∈Z k gk(t − lb−1 (cid:21)1/2 k )γk(t)h(t)2 dt (cid:35)1/2 k )h(t)2 dt (cid:35)1/2 k ) dt k γk(t)gk(t − lb−1 b−1 γk(t)gk(t − lb−1 R b−1 (cid:90) h(t)2(cid:88) R k,l∈Z . (8) R 6 The first term in the last expression can be bounded as follows (cid:90) R f (t)2(cid:88) ≤(cid:88) k,l∈Z (cid:88) k R b−1 k ess supt∈R k gk(t)γk(t − lb−1 k ) dt = b−1 (cid:32) (cid:88) (cid:32) (cid:88) (cid:32) (cid:88) b−1 k ess supt∈R b−1 k ess supt∈R l∈Z b−1 (cid:88) k∈Z γk(t − lb−1 k ) (cid:90) (cid:33)(cid:90) (cid:33)(cid:88) k ) γk(t − lb−1 (cid:33)(cid:32) k ) γk(t − lb−1 k∈Z l∈Z R (cid:90) R k∈Z ≤ sup k∈Z ≤ sup k∈Z γk(t − lb−1 k )f (t)2gk(t) dt l∈Z f (t)2gk(t) dt f (t)2gk(t) dt (cid:33) (cid:88) k∈Z Using relation (2) for ess supt∈R(cid:80) l∈Z l∈Zγk(t − lb−1 k ) and substituting (9) into (8) yields (5). The estimate for the second term in (8) is obtained analogously. By the density of compactly supported functions in L2(R), the estimate holds for all of L2(R). ess supt∈R gk(t) (cid:107)f(cid:107)2 2 . (9) Remark 3. From (8) in the proof of Proposition 3.3, it follows that the operator Sg,γ is also bounded by Remark 4. In the case of a frame operator Sg,g, the Walnut representation (4) becomes and the above bounds reduce to (cid:104)Sg,gf, h(cid:105) ≤(cid:16) (cid:32) (cid:104)Sg,gf, h(cid:105) ≤ (1 + b−1 sup k∈Z ess supt∈R k )(cid:107)gk(cid:107)W (L∞,(cid:96)1) (cid:88) k gk(t − lb−1 b−1 k,l∈Z (10) k ) a.e. , k )gk(t)f (t − lb−1 (cid:17)(cid:16) gk(t)(cid:17)(cid:107)f(cid:107)2(cid:107)h(cid:107)2 (cid:88) (cid:33) k )gk(t) k∈Z (cid:107)f(cid:107)2(cid:107)h(cid:107)2 ess supt∈R multiplication operator Sg,gf =(cid:80) k∈Z Gg,g k,0 · f = G0 · f . Remark 5. Note that in the painless case of Theorem 3.2, the frame operator reduces to the Remark 6. In the standard setting of Gabor frames, i.e. gk(t) = g(t − ak) for fixed a > 0, and bk = b for all k ∈ Z, the above bound reduces to the well know bound (cid:104)Sg,gf, h(cid:105) ≤ (1 + a−1)(1 + b−1)(cid:107)g(cid:107)2 W (L∞,(cid:96)1)(cid:107)f(cid:107)2(cid:107)h(cid:107)2 7 (cid:33)1/2 k )γk(t) (cid:33)1/2 k )gk(t) (cid:104)Sg,γf, h(cid:105) ≤ ess supt∈R k gk(t − lb−1 b−1 (cid:32) (cid:32) (cid:88) (cid:88) k,l∈Z k,l∈Z (cid:88) k,l∈Z Sg,gf (t) = k gk(t − lb−1 b−1 · ess supt∈R k γk(t − lb−1 b−1 (cid:107)f(cid:107)2(cid:107)h(cid:107)2 . 3.3. Existence of nonstationary Gabor frames For windows gk that are neither compactly supported nor bandlimited, we are interested in the existence of frames of the form G(g, b) and in the construction of the involved param- eters. The following theorem derives a sufficient condition for the existence of nonstationary Gabor frames and shows that this condition can be satisfied. In this and the subsequent sections, [bL, bU ], [pL, pU ], [CL, CU ] are compact intervals of positive real numbers. Theorem 3.4. Let g = {gk ∈ W (L∞, (cid:96)1) : k ∈ Z} be a set of windows such that i) for some positive constants A0, B0 0 < A0 ≤(cid:88) k∈Z gk(t)2 ≤ B0 < ∞ a.e. ; (11) ii) for all k ∈ Z, the windows decay polynomially around a δ-separated set {ak : k ∈ Z} of time-sampling points ak gk(t) ≤ Ck(1 + t − ak)−pk , (12) where pk ∈ [pL, pU ] ⊂ R, pL > 2 and Ck ∈ [CL, CU ]. k}k∈Z, such that for bk ≤ b0 (cid:90) Then there exists a sequence {b0 system G(g, b) forms a frame for L2(R). Proof. Let f ∈ L2(R). Applying (10), we write (cid:104)Sg,gf, f(cid:105) = (cid:88) k gk(t)2f (t)2 dt + b−1 (cid:88) (cid:88) (cid:90) k∈Z R R (cid:88) l∈Z\{0} (cid:88) k∈Z Using similar arguments as in the derivation of (8), we obtain (cid:12)(cid:12)(cid:12) ≤ k )f (t) dt (cid:12)(cid:12)(cid:12)(cid:90) R l∈Z\{0} k∈Z k, k ∈ Z, the nonstationary Gabor k gk(t)gk(t − lb−1 b−1 k )f (t − lb−1 k )f (t) dt ≤ k∈Z l∈Z\{0} ≤ max k∈Z {b−1 k )f (t − lb−1 k gk(t)gk(t − lb−1 b−1 ess supt∈R (cid:88) (cid:88) (cid:32) k } (cid:88) (cid:124) (cid:32) (cid:32) 2 ≥ min ess inf t∈R l∈Z\{0} k } k∈Z{b−1 k∈Z {b−1 k } ess supt∈R ess supt∈R k gk(t)gk(t − lb−1 k ) b−1  (cid:107)f(cid:107)2 (cid:33) k ) gk(t)gk(t − lb−1 (cid:125) 2 (cid:88) (cid:123)(cid:122) k∈Z R (cid:88) k } gk(t)2 − maxk∈Z{b−1 k } R mink∈Z{b−1 (cid:88) (cid:33) gk(t)2 + R k∈Z , k∈Z (cid:104)Sg,gf, f(cid:105)(cid:107)f(cid:107)−2 (cid:104)Sg,gf, f(cid:105)(cid:107)f(cid:107)−2 2 ≤ max (cid:107)f(cid:107)2 2 . (cid:33) (13) Therefore, lower and upper frame bounds are obtained from 8 We need to construct a sequence of bk, k ∈ Z, such that for all f ∈ L2(R), (13) is bounded away from zero. )1/pk. Then mink∈Z{b−1 (CL−1)1/p2, maxk∈Z{b−1 Let < CL and consider the sequence of frequency shifts bk = ( Ck k } ≥ k } ≤ (CU −1)1/p1 and maxk∈Z{b−1 k } k } ≤ C 1/pL mink∈Z{b−1 U C −1/pU L 1/pU−1/pL . (14) Since (1 + x + y)−p ≤ (1 + x)p(1 + y)−p for x, y ∈ R and p ≥ 0, using (12), we have, for some µ with pL − 2 > µ > 0: gk(t)gk(t − lb−1 k (cid:125) k )−pk+(1+µ) k )−pk+(1+µ) (1 + t − ak)−(1+µ) l−pk+(1+µ) 1−(1+µ)/pk . k ) ≤ C 2 ≤ C 2 ≤ C 2 = C 1+(1+µ)/pk k(1 + t − ak)−pk(1 + t − ak − lb−1 k(1 + t − ak)(1+µ)(1 + lb−1 k(1 + t − ak)−(1+µ) l−pk+(1+µ) bpk−(1+µ) k(cid:124) (cid:123)(cid:122) k ) ≤(cid:88) k∈Z Ek l−pL+(1+µ) 1−(1+µ)/pL (cid:88) k∈Z Ek l−pL+(1+µ) 1−(1+µ)/pL 2(cid:0)1 + (1 + δ)−(1+µ)(δ−1 + 1 + µ)µ−1(cid:1) , Ek(1 + t − ak)−(1+µ) l−pk+(1+µ) 1−(1+µ)/pk (1 + t − ak)−(1+µ) k∈Z k∈Z Ek Hence,(cid:88) k∈Z gk(t)gk(t − lb−1 ≤ max ≤ max (15) where the last estimate follows from Lemma 2.2 (b). Summing the expression (15) over l ∈ Z \ {0} using Lemma 2.2 ¡ , we see that R, as a function of , behaves like 1−(1+µ)/pL, i.e. R ≈ 1−(1+µ)/pL, and R tends to 0 for → 0. Moreover, maxk∈Z{b−1 k } k } R ≈ 1−(2+µ)/pL+1/pU mink∈Z{b−1 can be made arbitrarily small by choosing small since 1− (2 + µ)/pL + 1/pU > 0. Therefore, )1/pk, maxk∈Z{b−1 k } k } R < A0, then {Mlbkgk}k,l∈Z is a frame for all if 0 is such that for b0 mink∈Z{b−1 bk ≤ b0 k. k := ( 0 Ck For completeness, we state the equivalent result for analysis windows gk with polynomial decay in the frequency domain. Corollary 3.5. Let g = {gk ∈ L2(R) : gk ∈ W (L∞, (cid:96)1), k ∈ Z} be a set of windows such that i) for some positive constants A0, B0 gk(t)2 ≤ B0 < ∞ a.e. ; (16) 0 < A0 ≤(cid:88) k∈Z 9 ii) for all k ∈ Z, the windows decay polynomially around a δ-separated set {bk : k ∈ Z} of frequency-sampling points bk: gk(t) ≤ Ck(1 + t − bk)−pk , (17) where pk and Ck are chosen as in Theorem 3.4. Then there exists a sequence {a0 system {Tlakgk : k, l ∈ Z} forms a frame for L2(R). k}k∈Z, such that for ak ≤ a0 k, k ∈ Z, the nonstationary Gabor 3.3.1. Nonstationary Gabor frames on modulation spaces Modulation spaces, cf. [13, 15], are considered as the appropriate function spaces for time-frequency analysis and in particular, for the study of Gabor frames. By their definition, modulation spaces require decay in both time and frequency. Under additional assumptions on the windows gk, the collection G(g, b) is a frame for all modulation spaces M p, 1 ≤ p ≤ ∞. Proposition 3.6. Let G(g, b) be a frame for L2(R) satisfying the uniform estimate Vφgk(x, ω) ≤ C(1 + x − ak)−r−2(1 + ω)−r−2 , r > 2 (18) where φ is a Gaussian window. Then the frame operator S is invertible simultaneously on all modulation spaces M p for 1 ≤ p ≤ ∞. Proof. Notice, that Vφgk,l(x, ω) ≤ C(1 + (x − ak, ω − lbk))−r−2 , since (1+x+ω)−r ≥ (1+x)−r(1+ω)−r and the weights (1+(x, ω))r and (1+x+ω)r are equivalent. then the frame operator Sf =(cid:80) A result on Gabor molecules [16] states that, if an L2-frame {gz : z = (z1, z2) ∈ Z ⊆ R2}, where Z is separable, satisfies the uniform estimate Vφgz(x, ω) ≤ C(1+(x−z1, ω−z2))−r−2, z∈Z(cid:104)f, gz(cid:105)gz is invertible simultaneously on all M p for each 1 ≤ p ≤ ∞. The result hence follows from condition (18). 3.4. Constructing nonstationary Gabor frames Theorem 3.4 shows that for windows with sufficient uniform decay, nonstationary Gabor frames can always be constructed by choosing sufficient density in the frequency samples. In the present section we assume the existence of a certain nonstationary Gabor frame and explicitly construct a new frame by exploiting the prior information about the original one. This is a situation of practical relevance, since we may often be interested in using windows that decay fast and are negligible outside a support of interest. In particular, we will use the fact that painless nonstationary Gabor frames are easily constructed and deduce new frames from painless frames. The new frames thus obtained will be called almost painless nonstationary Gabor frames. We will subsequently assume bk ∈ [bL, bU ] for frequency-shift parameters and we let Ck ∈ [CL, CU ] and pk ∈ [pL, pU ] with pL > 1 for the constants involved in the decay assumptions 10 for the analysis windows. We then work with the following constants that depend on the separation of the sampling points, the decay of the windows and the frequency-sampling parameters: E1 = 1 + δ−1 + pL (1 + δ)pL(pL − 1) and E2 = 1 + (1 + b−1 bU + pU U )pL(pL − 1) (19) We first consider nonstationary Gabor frames obtained by perturbation of a known frame. This result is in the spirit of similar results for regular Gabor frames [3, 6]. Proposition 3.7. Let {ak : k ∈ Z} be a δ-separated set and G(h, b) a nonstationary Gabor frame with frame bounds Ah, Bh and frame operator Sh,h. Let gk ∈ L2(R) be a set of windows such that for all k ∈ Z and for almost all t ∈ R gk(t) − hk(t) ≤ Ck(1 + t − ak)−pk. (20) √ (cid:113) If CU < Ahλ−1 for (21) h )2. λ = 4b−1 U λB−1 C 2 L · E1 · E2, Bh(1 + Proof. By applying Cauchy-Schwartz inequality, it is easy to see that, for G(g, b) to be a 2 for some R < Ah, then G(g, b) is a frame for L2(R) with frame bounds A = Ah(1 −(cid:113) frame for L2(R), it suffices to show that(cid:80) also cf. [6, Proposition 4.1.]. Then, frame bounds of G(g, b) can be taken as Ah(1−(cid:112)R/Ah)2 and Bh(1 +(cid:112)R/Bh)2. in (9):(cid:88) To obtain the required error bound, we let ψk(t) := gk − hk and use the estimate given k,l∈Z(cid:104)f, gk,l − hk,l(cid:105)2 ≤ R(cid:107)f(cid:107)2 h )2 and B = U λA−1 C 2 ψk(t)(cid:17)(cid:107)f(cid:107)2 (cid:88) ψk(t − lb−1 k )(cid:17)(cid:16) b−1 k ess supt∈R ess supt∈R 2 . (22) (cid:16) (cid:104)f, ψk,l(cid:105)2 ≤ sup k∈Z k,l∈Z (cid:88) l∈Z k∈Z The first term of the above in the above inequality is bounded by assumption (20) and we can simplify ess supt∈R (cid:88) l∈Z k )−pk is b−1 l∈Z(1 + t − ak − lb−1 (cid:88) ψk(t − lb−1 k ) ≤ Ckess supt∈[0,b−1 (1 + t − lb−1 k )−pk . k ] l∈Z k −periodic, therefore (cid:88) Lemma 2.2 (b):(cid:88) To bound the second term, note that(cid:80) ψk(t) ≤ CU k∈Z k∈Z (cid:88) k∈Z (1 + t − ak)−pk ≤ CU (1 + t − ak)−p1 ≤ 2CU E1 . 11 Hence, by Lemma 2.2, we obtain for t ∈ [0, b−1 k ]: (cid:88) l∈Z (1 + t − lb−1 (1 + (l − 1)b−1 ∞(cid:88) k )−pk ≤ 1 + ∞(cid:88) k )−pk l=1 = 2 ∞(cid:88) k )−pk + (cid:16) ∞(cid:88) l=1 1 + (1 + t − lb−1 (1 + t + lb−1 k )−pk (1 + lb−1 k )−pk = 2 (1 + lb−1 k )−pk l=0 l=1 (cid:17) ≤ 2E2. Gathering all the estimates, we obtain (cid:104)f, ψk,l(cid:105)2 ≤ C 2 U 4b−1 1 E1 · E2(cid:107)f(cid:107)2 2 = C 2 U λ(cid:107)f(cid:107)2 2 . (cid:88) k,l∈Z By assumption C 2 U λ < Ah, and the proof is complete. Using Proposition 3.7 we next construct a special class of nonstationary Gabor frames by relying on knowledge of a painless nonstationary Gabor frame. We construct new windows which are no more compactly supported, but coincide with the known, compact windows on their support. We call the resulting new systems almost painless nonstationary Gabor frames. Corollary 3.8. Let g = {gk ∈ W (L∞, (cid:96)1) : k ∈ Z} be a set of windows, and let Ik be the intervals Ik = [ak − (2bk)−1, ak + (2bk)−1] where {ak : k ∈ Z} forms a δ-separated set. Assume that G(h, b), where hk = gkχIk, is a Gabor frame with lower frame bound Ah, and that for ψk = gk − hk, all k ∈ Z and almost all tinR  Ck(1 + t − ak − (2bk)−1)−pk , ψk(t) ≤ 0 , Ck(1 − t + ak − (2bk)−1)−pk , L · δ−1 · E1 · E2, then G(g, b) forms a nonstationary Gabor frame (23) t > ak + (2bk)−1 ; t ∈ Ik ; t < ak − (2bk)−1 . √ If CU < for L2(R). Ahλ−1 for λ = 4b−2 Proof. The proof follows the steps of the proof of Proposition 3.7 with small changes on how to approximate the terms in (22). First observe, that for any t ∈ R and all k, we have (cid:2)(1 + t − ak − (2bk)−1)−pk + (1 + t − ak + (2bk)−1)−pk(cid:3) ψk(t) ≤ Ck Since the frequency shifts bk are taken from a finite interval and the set {ak : k ∈ Z} is δ−separated, the sets Γ+ = {ak + (2bk)−1 : k ∈ Z} and Γ− = {ak − (2bk)−1 : k ∈ Z} are relatively δ−separated with rel(Γ) = rel(Γ+) = rel(Γ−) = (cid:98)(2bLδ)−1(cid:99). Therefore, by Remark 1, it follows that ψk(t) ≤ CU (1 + t − ak − (2bk)−1)−pL + CU (1 + t − ak + (2bk)−1)−pL (cid:88) k∈Z (cid:88) k∈Z (cid:88) k∈Z ≤ 4 CU rel(Γ) (1 + (1 + δ)−pL(δ−1 + pL)(pL − 1)−1) . 12 l∈Zψk(t − lb−1 k ) is b−1 k −periodic. Let t ∈ Ik, then, by (23) (1 + ak − (2bk)−1 − t + lb−1 k )−pk (cid:35) (1 − ak − (2bk)−1 + t − lb−1 k )−pk Now, the expression(cid:80) (cid:34)(cid:88) (cid:88) k ) ≤ Ck (cid:88) ψk(t − lb−1 l∈Z l>0 + l<0 ≤ Ck2 ∞(cid:88) (1 + lb−1 k )−pk ≤ 2Ck(1 + (1 + b−1 k )−pk(bk + pk)(pk − 1)−1) , where the last estimate follows from Lemma 2.2(a). l=0 4. Examples We illustrate our theory with two examples. In both examples, we consider a basic window and dilations by 2 and 1 2, respectively. Since the dilation parameters take only three different values, there are three kinds of windows, with support size 1/2, 1 and 2, respectively. Note that, while theoretically possible, sudden changes in the shape and width of adjacent windows turn out to be undesirable for applications, hence we only allow for stepwise change in dilation parameters. In the first example we consider a frame that arises as perturbation of a painless non- stationary Gabor frame. The perturbation consists in the application of a bandpass filter in order to obtain windows with compact support in the frequency domain. Example 4.1. Let h be a Hann or raised cosine window, i.e. h(t) = 0.5 + 0.5 cos(2πt) for t ∈ [−1/2, 1/2], and zero otherwise. We construct a painless nonstationary Gabor frame 2, respectively: Let {sk}k∈Z be a sequence with values from the set by dilating h by 2 or 1 {−1, 0, 1} with the restriction that sk − sk−1 ∈ {0, 1} to avoid sudden changes between adjacent windows. We then define corresponding shift-parameters by setting a0 = 0 and ak+1 = ak + 2−sk · 5 6 ak+1 = ak + 2−sk+1 · 1 3 ak+1 = ak + 2−sk+1 · 5 6 if if sk > sk+1 , sk = sk+1 , if sk < sk+1 . The points ak, k ∈ Z, form a separated set with minimum separation δ = 1/3. Setting 2skh(2skt), the system {Mlbkhk : k, l ∈ Z} forms a painless bk = 2sk and hk(t) = Tak nonstationary Gabor frame with lower frame bound Ah = 0.5. √ Let Ω = 0.02 and φ be a bandlimited filter given by (cid:98)φ(ω) = 0.5 + 0.5 cos(2πΩ−1ω) on its support [−Ω/2, Ω/2] . We build new windows gk by convolving φ with hk. The windows gk := φ ∗ hk are no more compactly supported. Since φ(t) ≤ Ω(1 + t)−3, we rely on [20, Theorem 9.9] to deduce 13 the following bound, with C(cid:48) = (cid:107)hk(cid:107)∞ Ω  (1 + (t − ak) − 2−s(cid:48) 2 , Ik = [ak − 2−s(cid:48) k, ak + 2−s(cid:48) k)−2 − (1 + (t − ak) + 2−s(cid:48) k)−2 − (1 − (t − ak) + 2−s(cid:48) k] and s(cid:48) k)−2 k)−2 − (1 − (t − ak) + 2−s(cid:48) k)−2 gk(t) − hk(t) ≤ C(cid:48) 2 − (1 + (t − ak) + 2−s(cid:48) (1 − (t − ak) − 2−s(cid:48) t > ak + 2−s(cid:48) t ∈ Ik t < ak − 2−s(cid:48) We obtain the joint bound, gk(t)− hk(t) ≤ CU (1 +t− ak)−2 by setting Ck = C(cid:48)· (1 + 2−s(cid:48) √ k)2 for all k ∈ Z and CU = maxk∈Z Ck = 0.0282 < Ahλ−1 = 0.0768, with λ as defined in (21). Thus, by Proposition 3.7, {Mlbkgk : k, l ∈ Z} is a Gabor frame with a lower frame bound A = 0.2. k = sk + 1: k)−2 k k Remark 7. Note that the construction presented in the Example 4.1 is of particular interest for constructing frequency-adaptive frames with windows that are compactly supported in time. This is a situation of considerable interest in applications, since it allows for real-time implementation with finite impulse response filters, cp. [11]. In the second example we construct a nonstationary Gabor frame by applying Corol- lary 3.8. The windows of the new frame coincide with the windows of a painless frame on their support. The windows in this example are constructed in analogy to the windows used in scale frames, introduced in [2] to automatically improve the resolution of transients in audio signals. Example 4.2. As in the previous example, let sk ∈ {−1, 0, 1} with sk − sk−1 ∈ {0, 1} for all k ∈ Z. We consider a sequence of windows gk that are translated and dilated versions of the Gaussian window g(t) = e−π(2.5t)2: gk(t) = Tak ak+1 = ak + 2−sk+1−1 · 2−sk+1 · 2−sk 2skg(2skt) with a0 = 0 and ak+1 = ak + sk = sk+1 , sk > sk+1 , ak+1 = ak + sk < sk+1 . √ if if if 1 3 1 3 Here, the {ak : k ∈ Z} are separated with minimum distance δ = 1/4. We arrange the windows as follows: after each change of window size, no change is allowed in the next step; in other words, each window has at least one neighbor of the same size. An example of the arrangement is shown in Figure 1. Let Ik = [ak − 2−sk−1, ak + 2−sk−1] and define a new set of windows by hk(t) = gk(t)χIk. Then {Mlbkhk : k, l ∈ Z} with bk = 2sk is a painless nonstationary Gabor frame. By numerical calculations, its lower frame bound is Ah = 0.1609 and ψk(t) = gk(t) − hk(t) can be bounded by  √ √ 0 ψk(t) ≤ 2skg(1/2)(1 + t − ak − 2−sk−1)−19 2skg(1/2)(1 + ak − 2−sk−1 − t)−19 t > ak + 2−sk−1 t ∈ Ik t < ak − 2−sk−1 . From Proposition 3.8 it follows that 4(δb2 k, l ∈ Z} is a nonstationary Gabor frame with a lower frame bound A = 0.1538. L)−1 · C 2 U · E1 · E2 = 0.0071 < Ah, and {MlbkTakgk : 14 Figure 1: An example for the arrangement of dilated windows in Example 2 5. Acknowledgement This work was supported by the WWTF project Audiominer (MA09-24) and the Austrian Science Fund (FWF):[T384-N13] Locatif. References [1] A. Aldroubi, C. A. Cabrelli, and U. Molter. Wavelets on Irregular Grids with Arbitrary Dilation Matrices, and Frame Atoms for L2(Rd). Appl. Comput. Harmon. Anal., Special Issue on Frames II.:119–140, 2004. [2] P. Balazs, M. Dorfler, F. Jaillet, N. Holighaus, and G. A. Velasco. Theory, imple- mentation and applications of nonstationary Gabor Frames. J. Comput. Appl. Math., 236:1481–1496, 2011. [3] I. Carrizo and S. Favier. Perturbation of wavelet and Gabor frames. Anal. Theory Appl., 19(3):238–254, 2003. 15 [4] O. Christensen. An Introduction to Frames and Riesz Bases. Applied and Numerical Harmonic Analysis. Birkhauser, 2003. [5] O. Christensen, S. Favier, and F. Z´o. Irregular wavelet frames and Gabor frames. Approx. Theory Appl., 17(3):90–101, 2001. [6] O. Christensen and C. Heil. Perturbations of Banach frames and atomic decompositions. Math. Nachr., 185:33–47, December 1997. [7] I. Daubechies, A. Grossmann, and Y. Meyer. Painless nonorthogonal expansions. J. Math. Phys., 27(5):1271–1283, May 1986. [8] M. Dorfler. Quilted Gabor frames - A new concept for adaptive time-frequency repre- sentation. Advances in Applied Mathematics, 47(4):668 – 687, Oct. 2011. [9] N. Holighaus, M. Dorfler, G. Velasco, and T. Grill. A framework for invertible, real-time constant-Q transforms. preprint, submitted, http://www.univie.ac.at/nonstatgab/slicq, 2012. [10] R. J. Duffin and A. C. Schaeffer. A class of nonharmonic Fourier series. Trans. Amer. Math. Soc., 72:341–366, 1952. [11] G. Evangelista, M. Dorfler, and E. Matusiak. Phase Vocoders With Arbitrary Frequency Band Selection. Proceedings of the 9th Sound and Music Computing Conference, July 11-14th 2012 Kopenhagen, 2012. [12] S. J. Favier and R. A. Zalik. On the stability of frames and Riesz bases. Appl. Comput. Harmon. Anal., 2(2):160–173, 1995. [13] H. G. Feichtinger. On a new Segal algebra. Monatsh. Math., 92:269–289, 1981. [14] D. Gabor. Theory of communication. J. IEE, 93(26):429–457, 1946. [15] K. Grochenig. Foundations of Time-Frequency Analysis. Appl. Numer. Harmon. Anal. Birkhauser Boston, 2001. [16] K. Grochenig. Localization of Frames, Banach Frames, and the invertibility of the frame operator. Proceedings of SPIE, San Diego, 2004. [17] H. Malvar. Signal Processing with Lapped Transforms. Boston, MA: Artech House. xvi, 1992. [18] J. L. Romero. Surgery of spline-type and molecular frames. J. Fourier Anal. Appl., 17:135 – 174, 2011. [19] A. Ron and Z. Shen. Weyl-Heisenberg frames and Riesz bases in L2(Rd). Duke Math. J., 89(2):237–282, 1997. [20] W. Rudin. Real and Complex Analysis. McGraw-Hill Book Company, New York, 1966. 16 [21] W. Sun and X. Zhou. Irregular wavelet/Gabor frames. Appl. Comput. Harmon. Anal., 13(1):63–76, 2002. [22] G. A. Velasco, N. Holighaus, M. Dorfler, and T. Grill. Constructing an invertible constant-Q transform with non-stationary Gabor frames. In Proceedings of DAFX11, Paris, September 2011. [23] D. F. Walnut. Continuity properties of the Gabor frame operator. J. Math. Anal. Appl., 165(2):479–504, 1992. [24] Z. Xiong and H. Malvar. A nonuniform modulated complex lapped transform. IEEE Signal Processing Letters, 8(9):257–260, September 2001. 17
1812.01419	3	1812	2019-05-30T18:59:57	On existence of shift-type invariant subspaces for polynomially bounded operator	[ "math.FA" ]	A particular case of results from [K2] is as follows. Let the unitary asymptote of a contraction $T$ contain the bilateral shift (of finite or infinite multiplicity). Then there exists an invariant subspace $\mathcal M$ of $T$ such that $T\|_{\mathcal M}$ is similar to the unilateral shift of the same multiplicity. The proof is based on the Sz.-Nagy--Foias functional model for contractions. In the present paper this result is generalized to polynomially bounded operators, but in the simplest case. Namely, it is proved that if the unitary asymptote of a polynomially bounded operator $T$ contains the bilateral shift of multiplicity $1$, then there exists an invariant subspace $\mathcal M$ of $T$ such that $T\|_{\mathcal M}$ is similar to the unilateral shift of multiplicity $1$. The proof is based on a result from [B].	math.FA	math	ON EXISTENCE OF SHIFT-TYPE INVARIANT SUBSPACES FOR POLYNOMIALLY BOUNDED OPERATORS MARIA F. GAMAL' Abstract. A particular case of results from [K2] is as follows. Let the unitary asymptote of a contraction T contain the bilateral shift (of finite or infinite multiplicity). Then there exists an invariant subspace M of T such that T M is similar to the unilateral shift of the same multi- plicity. The proof is based on the Sz.-Nagy -- Foias functional model for contractions. In the present paper this result is generalized to polynomi- ally bounded operators, but in the simplest case. Namely, it is proved that if the unitary asymptote of a polynomially bounded operator T contains the bilateral shift of multiplicity 1, then there exists an invari- ant subspace M of T such that T M is similar to the unilateral shift of multiplicity 1. The proof is based on a result from [B]. 1. Introduction Let H be a (complex, separable) Hilbert space, and let L(H) be the algebra of all (bounded, linear) operators acting on H. A (closed) subspace M of H is called invariant for an operator T ∈ L(H), if TM ⊂ M. The complete lattice of all invariant subspaces of T is denoted by Lat T . For a (closed) subspace M of a Hilbert space H, by PM and IM the orthogonal projection from H onto M and the identity operator on M are denoted, respectively. For Hilbert spaces H and K, let L(H,K) denote the space of (bound- ed, linear) transformations acting from H to K. Suppose that T ∈ L(H), R ∈ L(K), X ∈ L(H,K), and X intertwines T and R, that is, XT = RX. If X is unitary, then T and R are called unitarily equivalent, in notation: T ∼= R. If X is invertible, that is, X−1 ∈ L(K,H), then T and R are called similar, in notation: T ≈ R. If X is a quasiaffinity, that is, ker X = {0} and clos XH = K, then T is called a quasiaffine transform of R, in notation: T ≺ R. If clos XH = K, we write T ≺ R. If T ≺ R and R ≺ T , then T and R are called quasisimilar, in notation: T ∼ R. Clearly, T ≺ R if and only if R∗ ≺ T ∗. An operator T is called power bounded, if supn≥0 kT nk < ∞. An operator T is called polynomially bounded, if there exists a constant C such that kp(T )k ≤ C max{p(z) : z ≤ 1} for every (analytic) polynomial p. The smallest such constant is called the polynomial bound of T and is denoted here by Cpol,T . d 2010 Mathematics Subject Classification. Primary 47A15, 47A60. Key words and phrases. Shift-type invariant subspace, polynomially bounded operator, similarity, unilateral shift. 1 2 M.F. GAMAL' Every polynomially bounded operator can be represented as a direct sum of an operator similar to a singular unitary operator and of an absolutely continuous (a.c.) polynomially bounded operator, that is, an operator which admits an H ∞-functional calculus, see [M] or [K3, Theorems 13, 17 and 23]. In the present paper, absolutely continuous polynomially bounded operators are considered. (Although many results on polynomially bounded operators that will be used in the present paper were originally proved by Mlak [M], we will refer to [K3] for the convenience of references.) An operator T is called a contraction, if kTk ≤ 1. Every contraction is polynomially bounded with the constant 1 by the von Neumann inequality (see, for example, [SFBK, Proposition I.8.3]). Clearly, every polynomially bounded operator is power bounded. (It is well known that the converse is not true, see [F] for the first example of a power bounded but not poly- nomially bounded operator, and [P] for the first example of a polynomially bounded operator which is not similar to a contraction.) Let T ∈ L(H) be a power bounded operator. It is easy to see that the space HT,0 = {x ∈ H : kT nxk → 0} is invariant for T (sf. [SFBK, Theorem II.5.4]). Classes Cab, a, b = 0, 1,·, of power bounded operators are defined as follows (see [SFBK, Sec. II.4] and [K1]). If HT,0 = H, then T is of class C0·, while if HT,0 = {0}, then T is of class C1·. Furthermore, T is of class C·a, if T ∗ is of class Ca·, and T is of class Cab, if T is of classes Ca· and C·b, a, b = 0, 1. Let T and R be power bounded operators, and let T ≺ R. It easily follows from the definition that if R is of class C1· or of class C·0, then T is of class C1· or of class C·0, too. Clearly, any isometry is of class C1·, a unitary operator is of class C11, and the unilateral shift is of class C10. The notions of isometric and unitary asymptotes will be used. Although there exist further studies of these notions ([K3], [K4]), we restrict ourselves to the case of power bounded operators. For a power bounded operator T ∈ L(H) the isometric asymptote (X+, T (a) + ) can be defined using a Banach limit, see [K1]. Here T (a) + is an isometry (on a Hilbert space), and X+ is the canonical intertwining mapping: X+T = T (a) + X+. Recall that ker X+ = d ≺ T (a) HT,0, and the range of X+ is dense. Thus, T + . In particular, if T is cyclic, then T (a) + is cyclic, too. A power bounded operator T is of class C11 if and only if T is quasisimilar to a unitary operator; then T (a) + is unitary and T ∼ T (a) The unitary asymptote (X, T (a)) of a power bounded operator T ∈ L(H) is a pair where T (a) ∈ L(H(a)) (here H(a) is a some Hilbert space) is the minimal unitary extension of T (a) + , and X is an extension of X+. Therefore, + (see [SFBK, Propositions II.3.4 and II.5.3]). (1.1) _n≥0 (T (a))−nXH = H(a). SHIFT-TYPE INVARIANT SUBSPACES 3 In particular, if T is cyclic, then T (a) is cyclic, too. If T (a) of course, (X+, T (a) unitarily equivalence. + ) = (X, T (a)). Note that T (a) + is unitary, then, + and T (a) are defined up to Set T0· = THT ,0 and T1· = PHT ,0THT ,0. By [K1, Lemma 1], T0· and T1· are of classes C0· and C1·, respectively. Thus, every power bounded operator has the triangulations of the form (cid:18)C0· O C1·(cid:19) and (cid:18)C·1 O C·0(cid:19) . ∗ ∗ By [K1, Theorem 3], T (a) ∼= T (a) 1· . d T A particular case of [K2] is the following (see also [SFBK, Sec. other hand, it follows from (1.1) that T realized by the canonical intertwining mapping in general). Let T be a power bounded operator, let U be a unitary operator, and let d ≺ U . By [K1, Theorem 2], T (a) contains U as an orthogonal summand, that is, there exists a unitary operator V such that T (a) ∼= U ⊕ V . On the ≺ T (a) (but this relation is not IX.3] and references in [K2] and [SFBK] for the history of a question). Let T be a contraction, and let T (a) contain the bilateral shift of finite or infinite multiplicity as an orthogonal summand. Then there exists M ∈ Lat T such that TM is similar to the unilateral shift of the same multiplicity. The proof is based on the Sz.-Nagy -- Foias functional model for contractions, see [SFBK]. In the present paper this result is generalized to polynomially bounded operators, but in the simplest case. Namely, it is proved that if T is a polynomially bounded operator and T (a) contains the bilateral shift of multiplicity 1, then there exists M ∈ Lat T such that TM is similar to the unilateral shift of multiplicity 1. The proof is based on a result from [B]. Before formulating the main result of the paper, we introduce some no- tation. T and D denote the unit circle and the open unit disc, respectively. S is the unilateral shift of multiplicity 1, that is, S is the operator of mul- tiplication by the independent variable on the Hardy space H 2 on T. The normalized Lebesgue measure on T is denoted by m. For a measurable set σ ⊂ T denote by Uσ the operator of multiplication by the independent vari- able on L2(σ) := L2(σ, m). Clearly, UT is the bilateral shift of multiplicity 1. It is well known and easy to see that S and Uσ are a.c. contractions. Theorem 1.1. Suppose that T is a cyclic a.c. polynomially bounded oper- ator of class C1·, and T (a) ∼= UT. Then there exists M ∈ Lat T such that TM ≈ S. Corollary 1.2. Suppose that T is a polynomially bounded operator such that ≺ UT. Then for every c > 0 there exist M ∈ Lat T and W ∈ L(H 2,M) T such that W is invertible, W S = TMW and d kWkkW −1k ≤ (1 + c)(cid:0)√2(K 2 + 2)KCpol,T + 1(cid:1)qK 2C 2 Remark 1.3. Examples of polynomially bounded operators satisfying The- orem 1.1 and not similar to contractions can be found in [G]. Moreover, it follows from Theorem 1.1 and Corollary 1.2 that if T is a polynomially pol,T + 1KC 2 pol,T . 4 M.F. GAMAL' bounded operator such that T ∼ UT and the product of intertwining quasi- affinities is an analytic function of UT, or if T ≺ S, then T is similar to an operator constructed in [G, Proposition 2.7 or Corollary 2.3], respectively. Remark 1.4. Let T satisfy the assumption of Corollary 1.2. By [K1, Theo- rem 4], T ⊂ σ(T ). By [R], T either has a nontrivial hyperinvariant subspace or is reflexive. Therefore, the result on existence of nontrivial invariant sub- spaces of T is not new. The reflexivity of contractions satisfying the assump- tions of Corollary 1.2 is proved in [T1] and [T2], see also [SFBK, Theorem IX.3.8]. This proof can be generalized to polynomially bounded operators. We expect to give a detailed proof later elsewhere. Remark 1.5. Suppose that N ∈ N and T is a polynomially bounded oper- ≺ ⊕N n=1UT. It is possible to prove by induction that there ator such that T exists M ∈ Lat T such that TM ≈ ⊕N n=1S. In the case of infinite sum the question remains open. d The paper is organized as follows. In Sec. 2, Bourgain's result [B] is cited, and its corollaries which hold true for arbitrary a.c. polynomially bounded operators are given. In Sec. 3 and 4, auxiliary results for some functions and for operators intertwined with unitaries, respectively, are given. The main part of the paper is Sec. 5, where Theorem 1.1 is proved. In Sec. 6 Corollary 1.2 is proved. 2. Bourgain's result A denotes the disc algebra. For a positive finite Borel measure µ on T set P 2(µ) = closL2(µ) A, and denote by Sµ the operator of multiplication by the independent variable on P 2(µ). Theorem A ([B]). There exists a universal constant K with the following property. Let H be a Hilbert space, and let W ∈ L(A,H) (that is, W is a bounded linear transformation from A to H). Then there exists a positive Borel measure µ on T such that µ(T) = 1 and kWϕk ≤ KkWk(cid:16)ZT ϕ2dµ(cid:17)1/2 for every ϕ ∈ A. The following corollary is [BP, Lemma 2.1] emphasized for a.c. polyno- mially bounded operators. Corollary 2.1. Suppose that H is a Hilbert space, T ∈ L(H) is an a.c. polynomially bounded operator, Cpol,T is the polynomial bound of T , and x ∈ H. Then there exist ψ ∈ L2(T, m) such that RT ψ2dm ≤ 1 and for every ϕ ∈ H ∞ kϕ(T )xk ≤ KCpol,Tkxk(cid:16)ZT ϕ2ψ2dm(cid:17)1/2 and W ∈ L(P 2(ψ2m),H) such that W Sψ2m = T W, and W ϕ = ϕ(T )x for every ϕ ∈ H ∞. kWk ≤ KCpol,Tkxk SHIFT-TYPE INVARIANT SUBSPACES 5 Proof. Define W ∈ L(A,H) by the formula Wϕ = ϕ(T )x, ϕ ∈ A. Let µ be the measure from Theorem A. Then µ = ψ2m + µs, where ψ ∈ L2(T, m) and µs is a positive Borel measure on T singular with respect to m. We have P 2(µ) = P 2(ψ2m) ⊕ L2(µs) ([C, Proposition III.12.3] or [N, Corollary A.2.2.1]), Sµs is a singular uni- tary operator, and W has a continuous extension on P 2(µ) denoted by W . Clearly, W Sµ = T W . Therefore, WL2(µs)Sµs = T WL2(µs). Since T is a.c., WL2(µs) = O ( [K3, Proposition 15] or [M]). Thus, µ can be replaced by ψ2m. Let ϕ ∈ H ∞. Set ϕr(ζ) = ϕ(rζ), ζ ∈ clos D, 0 < r < 1. Then ϕr ∈ A, and ϕr → ϕ when r → 1 in the weak-∗ topology in H ∞ and in the norm in L2(ψ2m) simultaneously. The conclusion of the corollary follows from these convergences. (cid:3) Remark 2.2. For contractions T , Cpol,T = 1, and Corollary 2.1 is proved in [BT, Lemma 3] with K = 1. The proof is based on the existence of isometric dilations for contractions, see [SFBK, Theorem I.4.1]. Then Remark 2.3. Let ψ ∈ L2(T). If RT log ψdm = −∞, then P 2(ψ2m) = L2(ψ2m). If RT log ψdm > −∞, we accept that ψ is an outer function. P 2(ψ2m) = = khkH 2 , h ∈ H 2, H 2 ψ h =n h ψ , h ∈ H 2o, (cid:13)(cid:13)(cid:13) ψ(cid:13)(cid:13)(cid:13)P 2(ψ2m) see [C, Ch. III.12, VII.10] or [N, Ch. A.4.1]. Lemma 2.4. Suppose that H is a Hilbert space, T ∈ L(H) is an a.c. polyno- mially bounded operator, Cpol,T is the polynomial bound of T . Given N ∈ N and {ϕn}N n=1 ⊂ H ∞, define A ∈ L(H,⊕N n=1H), Ax = ⊕N n=1ϕn(T )x, x ∈ H. Then kAk ≤ KCpol,T ess sup ζ∈T (cid:16) NXn=1 ϕn(ζ)2(cid:17)1/2 . 6 M.F. GAMAL' Proof. For x ∈ H, let ψx ∈ L2(T) be the function from Corollary 2.1 (applied to T ). We have kϕn(T )xk2 kAxk2 = ≤ NXn=1 NXn=1 K 2C 2 pol,Tkxk2ZT ϕn2ψx2dm pol,Tkxk2ZT NXn=1 ϕn2ψx2dm ϕn2ZT ψx2dm NXn=1 pol,Tkxk2 ess sup ϕn2(cid:17) · kxk2. NXn=1 pol,T ess sup T T = K 2C 2 ≤ K 2C 2 ≤(cid:16)K 2C 2 (cid:3) Theorem 2.5. Suppose that H is a Hilbert space, T ∈ L(H) is an a.c. polynomially bounded operator, Cpol,T is the polynomial bound of T . Then (cid:13)(cid:13)(cid:13)Xn∈N ϕn(T )xn(cid:13)(cid:13)(cid:13) ≤ KCpol,T ess sup ζ∈T (cid:16)Xn∈N ϕn(ζ)2(cid:17)1/2(cid:16)Xn∈N kxnk2(cid:17)1/2 for every {ϕn}n∈N ⊂ H ∞ and every {xn}n∈N ⊂ H such that the right part of the above inequality is finite. Proof. Let {ϕn}n∈N ⊂ H ∞ be such that ess sup ζ∈T Xn∈N ϕn(ζ)2 := a < ∞. that Let N ∈ N. Let A be the transformation from Lemma 2.4 applied to T ∗ and n=1, where eϕ(ζ) = ϕ(ζ) (ζ ∈ D, ϕ ∈ H ∞). Taking into account that {fϕn}N kAk = kA∗k and ϕ(T )∗ = eϕ(T ∗) ([K3, Proposition 24] or [M]), we obtain ϕn(T )xn(cid:13)(cid:13)(cid:13) = kA∗(cid:0)⊕N (cid:13)(cid:13)(cid:13) NXn=1 ≤ KCpol,T ess sup n=1xn(cid:1)k ζ∈T (cid:16) NXn=1 ≤ KCpol,T a1/2(cid:16) NXn=1 kxnk2(cid:17)1/2 ϕn(ζ)2(cid:17)1/2(cid:16) NXn=1 kxnk2(cid:17)1/2 . theorem follows. Therefore, if Pn∈N kxnk2 < ∞, then Pn∈N ϕn(T )xn converges, and the Lemma 2.6. Suppose that H is a Hilbert space, T ∈ L(H) is an a.c. poly- nomially bounded operator, x ∈ H, ψk ∈ L2(T, m), k = 1, 2, are such that (cid:3) SHIFT-TYPE INVARIANT SUBSPACES 7 RT log ψkdm > −∞, and kϕ(T )xk ≤(cid:16)ZT ϕ2ψk2dm(cid:17)1/2 for every ϕ ∈ H ∞, k = 1, 2. (2(1 + ε2))1/2 Set ψ(ζ) = min(ψ1(ζ),ψ2(ζ)), ζ ∈ T. Then for every 0 < ε < 1 there exists an inner function ω ∈ H ∞ such that for every ϕ ∈ H ∞. kω(T )ϕ(T )xk ≤ Proof. We have T = τ1 ∪ τ2, where τ1 ∩ τ2 = ∅ and ψ = ψk a.e. on τk, k = 1, 2. There exist outer functions ηk ∈ H ∞, k = 1, 2, such that on τ1, on τ2. (cid:16)ZT ϕ2ψ2dm(cid:17)1/2 η2 =(ε ψ1 η1 =(1 on τ1, on τ2, 1 − ε ε ψ2 ψ1 and ψ2 1 Clearly, 1 − ε ≤ η1 + η2 ≤ 1 + ε a.e. on T. Set ωη = η1 + η2, where ω, η ∈ H ∞, ω is inner, and η is outer. Let ϕ ∈ H ∞. We have 2 2 2 2 ϕ kω(T )ϕ(T )xk2 =(cid:13)(cid:13)(cid:13)ω(T )η(T )(cid:16) ϕ =(cid:13)(cid:13)(cid:13)(η1 + η2)(T )(cid:16) ϕ η(cid:17)(T )x(cid:13)(cid:13)(cid:13) η(cid:17)(T )x(cid:13)(cid:13)(cid:13) ≤ 2(cid:16)(cid:13)(cid:13)(cid:13)η1(T )(cid:16) ϕ η(cid:17)(T )x(cid:13)(cid:13)(cid:13) η(cid:17)(T )x(cid:13)(cid:13)(cid:13) +(cid:13)(cid:13)(cid:13)η2(T )(cid:16) ϕ 2(cid:17) ψ12dm +ZT η22(cid:12)(cid:12)(cid:12) ≤ 2(cid:16)ZT η12(cid:12)(cid:12)(cid:12) η(cid:12)(cid:12)(cid:12) η(cid:12)(cid:12)(cid:12) ψ22dm(cid:17) (1 − ε)2(cid:16)ZT η12ϕ2ψ12dm +ZT η22ϕ2ψ22dm(cid:17) (1 − ε)2(cid:16)Zτ1 ϕ2ψ2dm + ε2Zτ2 ϕ2ψ2dm + ε2Zτ1 ϕ2ψ2dm +Zτ2 ϕ2ψ2dm(cid:17) (1 − ε)2 ZT ϕ2ψ2dm. 2(1 + ε2) ≤ = = ϕ 2 2 2 (cid:3) Remark 2.7. Applying Theorem 2.5, it is possible to prove an analog of Lemma 2.6 for a finite family {ψk}N k=1 with a constant not depended on N , and for a countable family {ψk}k∈N under additional assumption RT log ψdm > −∞, where ψ(ζ) = inf k∈N ψk(ζ), ζ ∈ T. We do not prove these statements, because we will not apply them. 3. Preliminaries: function theory Lemma 3.1. Suppose that ψ ∈ H 2 is an outer function. For t > 0 let ψt be an outer function such that ψt = max(ψ, t). Then (cid:13)(cid:13)(cid:13) ψ ψt − 1(cid:13)(cid:13)(cid:13)H 2→ 0 when t → 0. 8 M.F. GAMAL' ψ(ζ) < t}. Since ψ 6= 0 a.e. on H 2 ≥ m(T \ σt) → 1 when t → 0. Proof. For 0 < t < 1 set σt = {ζ ∈ T : T, m(σt) → 0 when t → 0. Therefore, (3.1) ψ 2 1 ≥(cid:13)(cid:13)(cid:13) ψt(cid:13)(cid:13)(cid:13) Furthermore, 0 ≤ (− log t)m(σt) ≤ −Zσt because m(σt) → 0 andRT log ψdm > −∞. Therefore, log ψdm → 0 when t → 0, log ψ t log ψ t dm → 0 when t → 0. Zσt (0) = exp(cid:16)Zσt dm(cid:17) → 1 when t → 0. =(cid:13)(cid:13)(cid:13) H 2 −(cid:12)(cid:12)(cid:12) (0)(cid:13)(cid:13)(cid:13) (0)(cid:12)(cid:12)(cid:12) ψt(cid:13)(cid:13)(cid:13) (0) − 1(cid:12)(cid:12)(cid:12) → 0 when t → 0. (cid:3) +(cid:12)(cid:12)(cid:12) (0)(cid:13)(cid:13)(cid:13)H 2 → 0 when t → 0. ψ ψt − ψ ψt ψ ψt ψ ψt ψ ψt H 2 ψ 2 2 2 Consequently, (3.2) ψ ψt By (3.1) and (3.2), (cid:13)(cid:13)(cid:13) (3.3) ψ ψt − By (3.2) and (3.3), ψ ψt − 1(cid:13)(cid:13)(cid:13)H 2≤(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13) Pn∈N ξ2 Proposition 3.2. Suppose that σ ⊂ T is a measurable set, {τn}n∈N is a family of measurable sets such that σ = ∪n∈Nτn and τn ∩ τk = ∅, if n 6= k. Suppose that C > 0 and {ξn}n∈N is a family of positive numbers such that n < ∞. Suppose that {ηn}n∈N, {ϕn}n∈N, and {ψn}n∈N are families of functions from L2(σ) such that ηn ≤(1 ξn on τn, on σ \ τn, and ϕn ≤ Cψn a.e. on σ. For every t > 0 and n ∈ N let ψnt ∈ L2(σ) be such that ψn ≤ ψnt a.e. on σ and Set (cid:13)(cid:13)(cid:13) ψn ψnt − 1(cid:13)(cid:13)(cid:13)L2(σ) → 0 when t → 0 αt =Xn∈N and α =Xn∈N ϕn ψnt η2 n η2 n ϕn ψn . for every n ∈ N. ψnt − 1 ≤ 2 a.e. on σ. Therefore, Then supt>0 kαtkL∞(σ) < ∞ and kαt − αkL2(σ) → 0 when t → 0. Proof. Clearly, ψn ψn(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ηn2(cid:12)(cid:12)(cid:12) αt − α ≤Xn∈N n < ε and m(cid:16) [n≥N +1 Xn≥N +1 ψnt − 1(cid:12)(cid:12)(cid:12) ≤ 2C(cid:16)1 +Xk6=n τn(cid:17) < ε. Take ε > 0. There exists N ∈ N such that k(cid:17) on τn. ψn ϕn ξ2 ξ2 SHIFT-TYPE INVARIANT SUBSPACES 9 We have ξ2 ε. ξ2 ψk ZSn≥N +1 τn αt − α2dm ≤ 4C 2(cid:16)1 +Xn∈N ψnt − 1(cid:12)(cid:12)(cid:12) + Xk6=n,k≤N ψnt − 1(cid:12)(cid:12)(cid:12) k(cid:1) NXk=1(cid:12)(cid:12)(cid:12) n(cid:17)2 ψkt − 1(cid:12)(cid:12)(cid:12) + 2 Xn≥N +1 k(cid:12)(cid:12)(cid:12) ψkt − 1(cid:12)(cid:12)(cid:12) + Xk6=n,k≤N ψkt − 1(cid:12)(cid:12)(cid:12) + 8C 2ε2 on k(cid:12)(cid:12)(cid:12) ψn ψk ψk ξ4 ξ4 2 Therefore, Let 1 ≤ n ≤ N . We have ψn αt − α ≤ C(cid:16)(cid:12)(cid:12)(cid:12) αt − α2 ≤ C 2(cid:16)2N(cid:16)(cid:12)(cid:12)(cid:12) ≤ 2N C 2 max(cid:0)1, sup Zσ(cid:12)(cid:12)(cid:12) ψnt − 1(cid:12)(cid:12)(cid:12) ψn k∈N There exists tε > 0 such that 2 dm ≤ ε 2N 2C 2 max(1, supk∈N ξ4 k) for every 0 < t < tε and n = 1, . . . , N. Therefore, 2 ξ2 n(cid:17) on τn. 2(cid:17) + 8ε2(cid:17) N[n=1 τn. ZSN n=1 τn αt − α2dm ≤ ε + 8C 2ε2 for every 0 < t < tε. (cid:3) 4. Preliminaries: operators intertwined with unitaries Lemma 4.1. Suppose that T ∈ L(H) is an a.c. polynomially bounded op- erator, Cpol,T is the polynomial bound of T , x ∈ H, and ψ ∈ L2(T) is from Corollary 2.1. Furthermore, suppose that σ ⊂ T is a measurable set, X ∈ L(H, L2(σ)), XT = UσX, and f = Xx. Then f ≤ kXkKCpol,Tkxkψ a.e. on σ. Proof. We have Zσ ϕ2f2dm = kϕ(Uσ)Xxk2 = kXϕ(T )xk2 ≤ kXk2kϕ(T )xk2 ≤ kXk2K 2C 2 pol,Tkxk2ZT ϕ2ψ2dm for every ϕ ∈ H ∞. It remains to apply standard reasons based on the fact that for every measurable set τ ⊂ T and every a,b > 0 there exists ϕ ∈ H ∞ such that ϕ = a a.e. on τ and ϕ = b a.e. on T \ τ . Lemma 4.2. Suppose that Lim is a Banach limit, H is a Hilbert space, T ∈ L(H) is a power bounded operator, X is the canonical intertwining mapping constructed using Lim. If M ∈ Lat T is such that TM is not of class C0·, then kXMk ≥ 1. Proof. We do not recall the construction of the canonical intertwining map- ping from [K1] here. We recall only that kXxk2 = Limn kT nxk2 for every x ∈ H. By assumption, there exists x ∈ M such that inf n≥0 kT nxk > 0. Therefore, kXxk > 0. Since (T (a))nX = XT n and T (a) is unitary, we have (cid:3) kXxk = k(T (a))nXxk = kXT nxk ≤ kXMkkT nxk for every n ≥ 0. 10 M.F. GAMAL' Consequently, kXxk2 ≤ kXMk2 Lim n kT nxk2 = kXMk2kXxk2. (cid:3) Since kXxk > 0, the conclusion of the lemma follows. Lemma 4.3. Suppose that T ∈ L(H) is an a.c. polynomially bounded oper- ator, Cpol,T is the polynomial bound of T , X∗ is the canonical intertwining mapping for T ∗ constructed using a Banach limit. Let X be any transfor- mation acting from H to some Hilbert space. For x ∈ H let ψ ∈ L2(T) be a function from Corollary 2.1. IfRT log ψdm = −∞, then Proof. Denote by H∗ the space on which (T ∗)(a) acts. Let W be from Corollary 2.1. Since Sψ2m is unitary (see Remark 2.3), there exists B∗ ∈ L(H∗, P 2(ψ2m)) such that W ∗ = B∗X∗ and kB∗k ≤ kW ∗k (see [K1, The- orem 2]). Set u = B∗ kXxk ≤ KCpol,TkXX∗ ∗kkxk. kuk ≤ kB∗ ∗ 1, where 1(ζ) = 1 (ζ ∈ T). We have ∗kk1kP 2(ψ2m) ≤ kWk(ZT ψ2dm)1/2 ≤ KCpol,Tkxk, ∗ B∗ ∗ 1 = XX∗ ∗ u, Xx = XW 1 = XX∗ ∗kKCpol,Tkxk. becauseRT ψ2dm ≤ 1. Taking into account that we conclude that kXxk ≤ kXX∗ ∗kkuk ≤ kXX∗ Proposition 4.4. Suppose that T ∈ L(H) is an a.c. polynomially bounded operator, σ ⊂ T is a measurable set, X ∈ L(H, L2(σ)), XT = UσX, x ∈ H is such that ∨n≥0T nx = H, f = Xx, ψ0 ∈ L2(T), andRT ψ02dm ≤ 1. Furthermore, suppose that 0 < ε, δ < 1, τ ⊂ σ is a measurable set, 0 6= u ∈ H is such that Xu ∈ L2(τ ) and kXuk ≥ δkuk. 0(cid:17)1/2 (we assume that ε is so small that γ0,δ0 > 0). Let 0 < γ < γ0. By assump- tion on x, there exists ϕ0 ∈ H ∞ such that kϕ0(T )x − uk ≤ ε. Then γ0 =(cid:16)1 − (kuk − ε)2δ2 δkuk − εkXk ε2kXk2 kuk + ε δ0 = and m(τ ) > 0, Set (cid:3) τ 6⊂ {ζ ∈ σ : ϕ0(ζ)f (ζ) ≤ γδ0kϕ0(T )xkψ0(ζ)}. Proof. We have ϕ0f = Xϕ0(T )x, and (ϕ0f )σ\τ = PL2(σ\τ )Xϕ0(T )x = PL2(σ\τ )X(ϕ0(T )x − u). Therefore, Also, Thus, Zσ\τ ϕ0f2dm ≤ kXk2kϕ0(T )x − uk2 ≤ ε2kXk2. kXϕ0(T )xk ≥ kXuk − εkXk ≥ δkuk − εkXk and δ0kϕ0(T )xk ≤ (kuk + ε)δ0. kXϕ0(T )xk ≥ δ0kϕ0(T )xk. δ2 0kϕ0(T )xk2 ≤ kXϕ0(T )xk2 =Zσ ϕ0f2dm =Zσ\τ ϕ0f2dm +Zτ ϕ0f2dm ≤ ε2kXk2 +Zτ ≤ ε2kXk2 + γ2δ2 (becauseRτ ψ02dm ≤RT ψ02dm ≤ 1). Thus, ε2kXk2 0kϕ0(T )xk2 ≤ ε2kXk2 0(kuk − ε)2 , 0kϕ0(T )xk2 1 − γ2 ≤ δ2 δ2 γ2δ2 0kϕ0(T )xk2ψ02dm SHIFT-TYPE INVARIANT SUBSPACES 11 Let us assume that τ ⊂ {ζ ∈ σ : ϕ0(ζ)f (ζ) ≤ γδ0kϕ0(T )xkψ0(ζ)}. Then (cid:3) a contradiction with assumption on γ. Proposition 4.5. Suppose that T ∈ L(H) is an a.c. polynomially bounded operator, σ ⊂ T is a measurable set, T (a) ∼= Uσ, and X is the canonical in- tertwining mapping for T and Uσ constructed using a Banach limit. Suppose that V ∈ L(K) is a unitary operator, Y ∈ L(K,H) is such that Y V = T Y and clos XY K = L2(σ). Furthermore, suppose that x ∈ H and {ϕn}n∈N ⊂ H ∞ are such that {ϕn(T )x}n∈N is dense in H. Set f = Xx. Suppose that {ψn}n∈N ⊂ L2(T) are such that RT ψn2dm ≤ 1 for all n ∈ N. For n ∈ N and 0 < c < 1 set ϕn(ζ)f (ζ) ≤ ckϕn(T )xkψn(ζ)}. σn,c = {ζ ∈ σ : For 0 < ε0, δ0, γ0 < 1 set Nε0,δ0 = {n ∈ N : kXϕn(T )xk ≥ δ0kϕn(T )xk and 1 − ε0 ≤ kϕn(T )xk ≤ 1 + ε0} and τε0,δ0,γ0 = \n∈Nε0,δ0 σn,γ0δ0. Finally, let 0 < ε00, δ0, γ0 < 1 be given. Then there exists 0 < ε0 ≤ ε00 such that m(τε0,δ0,γ0) = 0. PK⊖E V K⊖E ≺ Uσ. Proof. Set E = ker XY . Then E ∈ Lat V and XY K⊖E realizes the relation (4.1) Therefore, V ∗K⊖E is of class C·1. Consequently, V ∗K⊖E is unitary. Thus, K ⊖ E ∈ Lat V , and V K⊖E ∼= Uσ by (4.1) and [SFBK, Proposition II.3.4]. Without loss of generality, K ⊖ E = L2(σ) and V K⊖E = Uσ. We have ker XY L2(σ) = {0}. Therefore, there exists g ∈ L∞(σ) such that a.e. on σ and XY h = gh for every h ∈ L2(σ). (4.2) g 6= 0 12 M.F. GAMAL' For 0 < ε0, δ < 1 set δ1 = δ − ε0kXk 1 + ε0 and γ1 =(cid:16)1 − 0kXk2 ε2 (1 − ε0)2δ2 1(cid:17)1/2 . Then γ1 → 1 and δ1 → 1 when δ → 1 and ε0 → 0. Choose 0 < ε0, δ < 1 such that 0 < ε0 ≤ ε00, γ0 < γ1 and δ0 ≤ δ1. We claim that m(τε0,δ0,γ0) = 0. Set τ = τε0,δ0,γ0. Let us assume that m(τ ) > 0. Set M = clos Y L2(τ ). By (4.2), XM 6≡ O. By Lemma 4.2, kXMk ≥ 1. Therefore, there exists u ∈ M such that kuk = 1 and kXuk ≥ δ. We have Xu ∈ clos XM = clos XY L2(τ ) = L2(τ ). Since {ϕn(T )x}n∈N is dense in H, there exists n ∈ N such that kϕn(T )x − uk ≤ ε0. Therefore, n ∈ Nε0,δ0. By Proposition 4.4 applied to x and ϕn with ε0 and δ, taking into account that γ0 < γ1, we obtain τ 6⊂ σn,γ0δ1. By the definition of τ , we have τ ⊂ σn,γ0δ0 ⊂ σn,γ0δ1, a contradiction. (cid:3) The following lemma is a version of [K1, Theorem 3]. Lemma 4.6. Suppose that Lim is a Banach limit, T is a power bounded operator, M ∈ Lat T , X∗ and Z are canonical intertwining mappings for T ∗ and (TM)∗ from their unitary asymptotes, respectively, constructed using Lim. Set n kPM(T ∗)nxk2 ≤ Lim n k(T ∗)nxk2 =kX∗xk2 Following the proof of [K1, Theorem 3(b)] (applied to T ∗) we obtain a L = _n≥0(cid:0)(T ∗)(a)(cid:1)−nX∗M⊥. Then there exists a transformation B such that kBk ≤ Z = BPL⊥X∗M and √2kX∗k. Proof. By the constructions of X∗ and Z, kZxk2 = Lim for every x ∈ M. Therefore, kZk ≤ kX∗k. transformation(cid:18)B1 B2 (cid:18)B1 B2 O B(cid:19)(cid:18)X∗M⊥ n k(cid:0)(TM)∗(cid:1)nxk2 = Lim O B(cid:19) such that PL⊥X∗M(cid:19) =(cid:18)X∗M⊥ PLX∗M (cid:19) and(cid:13)(cid:13)(cid:13)(cid:18)B1 B2 O B(cid:19)(cid:13)(cid:13)(cid:13) ≤(cid:13)(cid:13)(cid:13)(cid:18)X∗M⊥ PLX∗M (cid:19)(cid:13)(cid:13)(cid:13) ≤ kBk ≤(cid:13)(cid:13)(cid:13)(cid:18)B1 B2 O B(cid:19)(cid:13)(cid:13)(cid:13) ≤ √2kX∗k. (cid:3) Proposition 4.7. Set χ(ζ) = ζ and 1(ζ) = 1 (ζ ∈ T). Set UTH 2 Therefore, BPL⊥X∗M = Z and H 2 − = L2(T) ⊖ H 2 and S∗ = PH 2 PLX∗M Z . − Z O O O − Suppose that N and K are Hilbert spaces, Lim is a Banach limit, T0 ∈ L(N ) is power bounded, √2kX∗k. SHIFT-TYPE INVARIANT SUBSPACES 13 V0 ∈ L(K), Z0 ∈ L(N ,K), (Z0, V0) is the unitary asymptote of T ∗ structed using Lim, Y0 ∈ L(H 2,N ), and Y0S = T0Y0. Set Z0 O Y ∗ O 0 O IH 2 S∗ (cid:19) and X∗ = T =(cid:18)T0  . (·, χ)Y01 O − Then 0 con- (Y0 ⊕ IH 2 (4.3) T) is the unitary asymptote of T ∗ con- T is power bounded, and (X∗, V0 ⊕ U ∗ structed using Lim. Furthermore, if T0 is polynomially bounded, then T is polynomially bounded, and if T0 is a.c., then T is a.c.. )UT = T (Y0 ⊕ IH 2 ), − − Moreover, let (X0, UT) be the unitary asymptote of T0 constructed using some Banach limit. Since (X0Y0)S = UT(X0Y0), there exists g1 ∈ L∞(T) such that X0Y0h = g1h for every h ∈ H 2. Let be defined by the formula X ∈ L(N ⊕ H 2 −, L2(T)) X(x ⊕ h) = X0x + g1h (x ∈ N , h ∈ H 2 −). Then (X, UT) is the unitary asymptote of T constructed using the same Banach limit as X0. Proof. The equality (4.3) easy follows from the definition of T . Let p be an (analytic) polynomial. It easy follows from (4.3) that = Y0PH 2p(UT)H 2 PN p(T )H 2 (4.4) . − − The conclusions on power boundedness and polynomial boundedness of T follow from (4.4). Since the right part of (4.4) is defined for every p ∈ H ∞, the conclusion on a.c. polynomial boundedness of T follows from (4.4) and [K3, Theorem 23]. Let x1, x2 ∈ N , let h1, h2 ∈ H 2 ((T ∗)n(x1 ⊕ h1), (T ∗)n(x2 ⊕ h2)) χnY ∗ = (h1, h2) + (PH 2 − −, and let n ∈ N. It follows from (4.4) that 0 x1, PH 2 − χnY ∗ 0 x2) + ((T ∗ 0 )nx1, (T ∗ 0 )nx2). We have lim n (PH 2 − χnY ∗ 0 x1, PH 2 − χnY ∗ 0 x2) = (Y ∗ 0 x1, Y ∗ 0 x2). By the construction of Z0, (Z0x1, Z0x2) = Lim n and by the construction of X∗, ((T ∗ 0 )nx1, (T ∗ 0 )nx2), (X∗(x1 ⊕ h1), X∗(x2 ⊕ h2)) = Lim n ((T ∗)n(x1 ⊕ h1), (T ∗)n(x2 ⊕ h2)). Therefore, (X∗(x1 ⊕ h1), X∗(x2 ⊕ h2)) = (h1, h2) + (Y ∗ 0 x1, Y ∗ 0 x2) + (Z0x1, Z0x2). From the latest equality and [K1, Theorem 3] we obtain the conclusion on the unitary asymptote of T ∗. Suppose that (X0, UT) is the unitary asymptote of T0 constructed using some Banach limit, and X is defined as above. Clearly, XT = UTX. Denote 14 M.F. GAMAL' by XT the canonical intertwining mapping for T and T (a) constructed using the same Banach limit as X0. By [K1, Theorem 3], T (a) ∼= (T0)(a) and X0 = XTN . Therefore, T (a) ∼= UT, and it is easy to see from the definitions of T and X and the intertwining properties of X and XT that XT = X. (cid:3) 5. Proof of Theorem 1.1 For convenience of applying Proposition 4.7, we denote the operator from Theorem 1.1 by T0. Suppose that N is a Hilbert space, T0 ∈ L(N ) is a power bounded oper- ator of class C1·, and (T0)(a) ∼= UT. Let ∗ ∗ O T·0(cid:19) O C·0(cid:19), see Introduction and [K1]. T0 =(cid:18)T11 be the triangulation of T0 of the form(cid:18)C·1 Since T0 is of class C1·, we have T11 is of class C11. By [K1, Theorem 3] and 0 )(a) ∼= ((T11)(a))∗, [SFBK, Propositions II.3.4 and II.5.3], T11 ∼ (T11)(a), (T ∗ and there exists a measurable set τ ⊂ T such that (T11)(a) ∼= Uτ . First case: m(τ ) = 1. Then T11 ∼ UT. Suppose that T0 is polynomially bounded. We can consider T11 instead of T0. The proof becomes simpler. Proposition 4.7 is not used, because we need not to represent T11 as a restriction of some operator of class C·1 on its invariant subspace N . Lemma 2.6 is not used, because we need not to request that the needed shift-type invariant subspace is contained in the given subspace N . The detailed proof can be deduced from the proof of second case and is not given here. Two cases are possible. Second case: m(τ ) < 1. Suppose that T0 is a cyclic a.c. polynomially bounded operator. By Corollary 2.1, there exists a quasiaffinity Y0 ∈ L(H 2,N ) such that Y0S = T0Y0. Denote by X0 the canonical intertwining mapping for T0 and UT constructed using a Banach limit. We have (5.1) X0T0 = UTX0 and _n≥0 U −n T X0N = L2(T). −. Let T ∈ L(H), X ∈ L(H, L2(T)), g1 ∈ L∞(T) be from Set H = N ⊕ H 2 Proposition 4.7 applied to T0 and Y0. Since Y0 is a quasiaffinity, it follows from the second equality in (5.1) that g1 6= 0 a.e. on T, and it follows from (4.3) that T is of class C·1. Let X∗ ∈ L(H, L2(τ ) ⊕ L2(T)) be the canonical intertwining mapping for T ∗ constructed in Proposition 4.7 with V0 = U ∗ X∗ ∗ (L2(τ ) ⊕ H 2) ⊂ N . (5.2) Furthermore, there exists g2 ∈ L∞(τ ) such that τ . Then (XX∗ ∗ )(h2 ⊕ h1) = g2h2 + g1h1 for every h2 ∈ L2(τ ) and h1 ∈ L2(T). Set θ1 = g1 (g12 + g22)1/2 and θ2 = g2 (g12 + g22)1/2 , SHIFT-TYPE INVARIANT SUBSPACES 15 and set Clearly, E0 = ker XX∗ E0 = {θ1h ⊕ (−θ2h) : h ∈ L2(τ )}. ∗ and E0 is a reducing subspace for Uτ ⊕ UT, that is, E0 ∈ Lat(Uτ ⊕ UT) and (L2(τ ) ⊕ L2(T)) ⊖ E0 ∈ Lat(Uτ ⊕ UT). Define by the formula Set (5.3) J ∈ L(cid:0)L2(T), (L2(τ ) ⊕ L2(T)) ⊖ E0(cid:1) (h ∈ L2(T)). Jh = θ2hτ ⊕ (θ1hτ + hT\τ ) g =((g12 + g22)1/2 g1 on τ, on T \ τ. ∗ Jh = gh for every h ∈ L2(T) and L2(τ ) ⊕ L2(T) = JL2(T) ⊕ E0. Then g 6= 0 a.e. on T, J is unitary, JUT = (Uτ ⊕ UT)(L2(τ )⊕L2(T))⊖E0 J, (5.4) XX∗ Take σ ⊂ T such that (5.5) KCpol,T√2kX∗k ess sup Let h0 ∈ H 2 be such that T\σ g > 0 and m(τ ∪ σ) < 1. g < 1, ess inf σ (5.6) There exists θ0 ∈ L∞(τ ) such that θ0 = 1 a.e. on τ and (5.7) h0 ≍ 1 a.e. on T. Re θ1θ2θ0 = 0 a.e. on τ. Set (5.8) h01 = (θ2θ0 + θ1)h0χτ + h0χT\τ and h02 = (−θ2 + θ1θ0)h0τ . h01 = h0 a.e. on T. Fτ,σ = θ2h01τ ∩σ + θ1h02, FT,σ = θ1h01τ ∩σ + h01σ\τ − θ2h02, Fτ,T\σ = θ2h01τ \σ, FT,T\σ = θ1h01τ \σ + h01(T\τ )\σ. By (5.7), (5.9) Set (5.10) and (5.13) Put (5.14) Then (5.11) Fτ,σ ⊕ FT,σ = Jh01σ + θ1h02 ⊕ (−θ2h02), (5.12) Fτ,T\σ ⊕ FT,T\σ = Jh01T\σ, θ1h02 ⊕ (−θ2h02) ∈ E0, Fτ,σ ⊕ FT,σ + Fτ,T\σ ⊕ FT,T\σ = θ0h0τ ⊕ h0 ∈ L2(τ ) ⊕ H 2. x0 = X∗ ∗ (Fτ,σ ⊕ FT,σ). 16 M.F. GAMAL' It follows from (5.11) and (5.4) that Xx0 = gh01σ. M0 = _k≥0 T kx0 and K0 = _k≥0 (Uτ ⊕ UT)k(Fτ,σ ⊕ FT,σ). By (5.10), Fτ,σ ⊕ FT,σ ∈ L2(τ ) ⊕ L2(τ ∪ σ). It follows from third inequality in (5.5) that (Uτ ⊕ UT)K0 is unitary. It follows from (5.6), (5.9), (5.15), and the equality XT = UTX that By [K1, Theorem 3], (XM0 , Uσ) is the unitary asymptote of TM0 con- structed using the same Banach limit as X. clos XM0 = L2(σ). (5.15) Set Set (5.16) L = _n≥0 (U ∗ τ ⊕ U ∗ T)−nX∗M⊥ 0 . Denote by Z the canonical intertwining mapping for (TM0)∗ constructed using the same Banach limit as X∗. Let B be the transformation from Lemma 4.6 applied to T and M0. We have XM0 Z ∗ = XPM0X∗ Therefore, kXM0 Z ∗k ≤ It follows from (5.16) that X∗ ∗ PL⊥B∗. √2kX∗kkXPM0 X∗ ∗L⊥ ⊂ M0. Therefore, and XX∗ ∗L⊥k. XPM0X∗ (5.17) Let u ∈ L⊥. By (5.4), there exist h ∈ L2(T) and v ∈ E0 such that u = Jh+v. By (5.4) and (5.17), ∗L⊥ ⊂ L2(σ). ∗L⊥ = XX∗ ∗L⊥ XX∗ Therefore, hT\σ = 0 and ∗ uk ≤ ess sup = ess sup kXX∗ σ ∗ u = XX∗ ∗ Jh = gh ∈ L2(σ). gkhkL2(σ) ≤ ess sup gkuk σ σ g(khk2 L2(T) + kvk2)1/2 (because J is unitary). We obtain that Consequently, By (5.5), Take kXM0 Z ∗k ≤ σ kXX∗ ∗L⊥k ≤ ess sup g. √2kX∗k ess sup KCpol,TkXM0 Z ∗k < 1. σ g. KCpol,TkXM0 Z ∗k < δ0 < 1 and 0 < ε00, γ0 < 1. Let {ϕn}n∈N ⊂ H ∞ be such that {ϕn(T )x0}n∈N is dense in M0. Let ψn be constructed by Corollary 2.1 applied to ϕn(T )x0 for every n ∈ N. Set (5.18) f = Xx0. SHIFT-TYPE INVARIANT SUBSPACES 17 It is easy to see that TM0, XM0, (Uτ ⊕ UT)K0, X∗ Proposition 4.5. Let 0 < ε0 ≤ ε00 be such that m(τε0,δ0,γ0) = 0, ∗K0, x0, {ψn}n∈N satisfy where τε0,δ0,γ0 is from Proposition 4.5. For convenience, relabel {ϕn}n∈Nε0,δ0 as {ϕn}n∈N. Then kXϕn(T )x0k ≥ δ0kϕn(T )x0k for every n ∈ N. By Lemma 4.3 applied to TM0 and the choice of δ0 we have (5.19) log ψndm > −∞ for every n ∈ N. ZT We accept that ψn are outer functions (and ψn ∈ H 2), see Remark 2.3. Set σn = σn,γ0δ0, n ∈ N. By (5.18) and Lemma 4.1 applied to ϕn(T )x0, (5.20) ϕnf ≤ kXkKCpol,Tkϕn(T )x0kψn ≤ (1 + ε0)kXkKCpol,Tψn a.e. on σ for every n ∈ N. By Proposition 4.5, ϕnf > γ0δ0kϕn(T )xkψn ≥ (1 − ε0)γ0δ0ψn a.e. on σ \ σn for every n ∈ N and m(\n∈N σn) = 0. Set Then and (5.21) τ1 = σ \ σ1, τn = (σ \ σn) \ ∪n−1 [n∈N τn ∩ τk = ∅, τn = σ, k=1τk, n ≥ 2. if n 6= k, n, k ≥ 1, ϕnf > (1 − ε0)γ0δ0ψn a.e. on τn, n ∈ N. By (5.14), ϕ(T )ϕn(T )x0 = ϕ(T )ϕn(T )X∗ ∗ (Fτ,σ ⊕ FT,σ) = X∗ ∗ (ϕϕnFτ,σ ⊕ ϕϕnFT,σ). Therefore, kϕ(T )ϕn(T )x0k2 ≤ kX∗ = kX∗ = kX∗ ≤ kX∗ Easy calculation show that ∗k2(kϕϕnFτ,σk2 + kϕϕnFT,σk2) ∗k2(cid:16)Zτ ϕϕnFτ,σ2dm +ZT ϕϕnFT,σ2dm(cid:17) ∗k2ZT ϕϕn2(χτFτ,σ2 + FT,σ2)dm (χτFτ,σ2 + FT,σ2)ZT ϕϕn2dm. ∗k2 ess sup (τ \σ)∪(σ\τ ) h02). h02, ess sup τ ∩σ T ess sup T (χτFτ,σ2 + FT,σ2) = max(2 ess sup 18 Set Then M.F. GAMAL' C1 = kX∗ kϕ(T )ϕn(T )x0k ≤ KCpol,T (1 + ε0)(cid:16)ZT ϕ2C 2 ∗k(cid:0)max(2 ess supτ ∩σ h02, ess sup(τ \σ)∪(σ\τ ) h02)(cid:1)1/2 1ϕn2dm(cid:17)1/2 KCpol,T (1 + ε0) . Take 0 < c1 < 1. By Lemma 2.6, there exist inner functions ωn ∈ H ∞, for every ϕ ∈ H ∞ and every n ∈ N. n ∈ N, such that kϕ(T )ωn(T )ϕn(T )x0k ≤ KCpol,T (1 + ε0) (2(1 + c2 1 − c1 for every ϕ ∈ H ∞ and every n ∈ N. ψ1n =(ψn (5.22) There exist outer functions ψ1n ∈ H 2 such that min(ψn, C1ϕn) on σ, on T \ σ. 1))1/2 (cid:16)ZT ϕ2 min(ψn, C1ϕn)2dm(cid:17)1/2 Since min(ψn, C1ϕn) ≤ ψ1n a.e. on T, we have (5.23) (2(1 + c2 1 − c1 kϕ(T )ωn(T )ϕn(T )x0k ≤KCpol,T (1 + ε0) 1))1/2 (cid:16)ZT ϕ2ψ1n2dm(cid:17)1/2 for every ϕ ∈ H ∞ and every n ∈ N. It follows from (5.22) that Let {ξn}n∈N be a sequence of positive numbers such that (cid:12)(cid:12)(cid:12) ϕn 1 C1 ψ1n(cid:12)(cid:12)(cid:12) ≥ ξ :=Xn∈N on T \ σ for every n ∈ N. ξ2 n < ∞, Xn≥2 ξ2 n < 1 (5.24) (5.25) and (5.26) (1 + ε0)kXkKCpol,T ξ < (1 − ε0)γ0δ0. For n ∈ N, define outer functions ηn ∈ H ∞ as follows: (5.27) 1/2 1 ξ1 on T \ σ, on τ1, on σ \ τ1, ηn = η1 = ϕn(cid:12)(cid:12)(cid:12) ϕ1(cid:12)(cid:12)(cid:12) ξn(cid:12)(cid:12)(cid:12) ψ1n (cid:12)(cid:12)(cid:12) ψ11 Xn∈Nηn(ζ)2 ≤((1 +Pn≥2 ξ2 n)C1 1 ξn 1 + ξ By (5.24) and (5.25), for a.e. ζ ∈ T \ σ, for a.e. ζ ∈ σ. 1/2 on T \ σ, on τn, on σ \ τn, n ≥ 2. For every t > 0 and every n ∈ N let ψnt be the outer function such that (5.28) ψnt = max(ψ1n, t) a.e. on T. SHIFT-TYPE INVARIANT SUBSPACES 19 Put (5.29) κnt(ζ) = ηn(ζ) ψnt(ζ) , ζ ∈ T, n ∈ N, t > 0. Since ψnt ≥ t a.e. on T, we have κnt ∈ H ∞ for every n ∈ N and t > 0. Set xnt = (κntωnϕn)(T )x0 = (κntωnϕn)(T )X∗ ∗ (κntωnϕnFτ,T\σ ⊕ κntωnϕnFT,T\σ) and ynt = X∗ ∗ (Fτ,σ ⊕ FT,σ) = (κntωnϕn)(T )X∗ ∗ (Fτ,T\σ ⊕ FT,T\σ) for every n ∈ N and t > 0. (5.30) We have xnt + ynt = (κntωnϕn)(T )X∗ ∗ (Fτ,σ ⊕ FT,σ + Fτ,T\σ ⊕ FT,T\σ). By (5.13) and (5.2), (5.31) By (5.23) and (5.28), xnt + ynt ∈ N . (5.32)Xn∈N kϕ(T )xntk2 =Xn∈N and by (5.29), (5.33) (cid:17)2Xn∈NZT ϕκnt2ψnt2dm, 1))1/2 k(ϕκntωnϕn)(T )x0k2 (2(1 + c2 1 − c1 ≤(cid:16)KCpol,T (1 + ε0) Xn∈NZT ϕκnt2ψnt2dm =Xn∈NZT ϕ2ηn2dm =ZTXn∈N T Xn∈N ηn2ϕ2dm ≤ ess sup ηn2ZT ϕ2dm. By Theorem 2.5, (5.32) and (5.33), (5.34) kXn∈N (ηnϕ)(T )xntk ≤ K 2C 2 pol,T (1 + ε0) 1))1/2 (2(1 + c2 1 − c1 ess sup T Xn∈N ηn2(cid:16)ZT ϕ2dm(cid:17)1/2 for every t > 0. For t > 0 put (5.35) βt =Xn∈N η2 n ωnϕn ψnt on T \ σ. It follows from (5.27) and (5.28) that the series (5.35) converges in k·kL∞(T\σ) and (5.36) kβtkL∞(T\σ) ≤ 1 +Xn≥2 ξ2 n for all t > 0. 20 M.F. GAMAL' For t > 0 and ϕ ∈ H ∞ let Wtϕ =Xn∈N (5.37) (ηnϕ)(T )xnt + X∗ ∗ (βtϕFτ,T\σ ⊕ βtϕFT,T\σ). It follows from the convergence in k · kL∞(T\σ) of the series (5.35) that (5.38) X∗ ∗ (βtϕFτ,T\σ⊕βtϕFT,T\σ) X∗ ωnϕn ψnt n ∗(cid:16)η2 =Xn∈N ϕFτ,T\σ ⊕ η2 n ωnϕn ψnt ϕFτ,T\σ ⊕ η2 ϕFT,T\σ(cid:17). ϕFT,T\σ(cid:17) = (ϕηn)(T )ynt. ∗ and (5.30) that It follows from the intertwining property of X∗ It follows from (5.37), (5.38), (5.39), and (5.31) that X∗ n ωnϕn ψnt ωnϕn ψnt ∗(cid:16)η2 Wtϕ ∈ N for every ϕ ∈ H ∞ and t > 0. n By (5.36) and (5.10), By (5.9) and (5.10), T\σ (cid:0)χτFτ,T\σ2 + FT,T\σ2(cid:1)ZT\σ ϕ2dm. ∗ (βtϕFτ,T\σ ⊕ βtϕFT,T\σ)k2 n(cid:1)2 ess sup ≤ kX∗ ξ2 ess sup ∗k2(cid:0)1 +Xn≥2 T\σ (cid:0)χτFτ,T\σ2 + FT,T\σ2(cid:1) = ess sup T Xn∈N (2(1 + c2 1 − c1 pol,T (1 + ε0) 1))1/2 ess sup T\σ C2 =K 2C 2 h02. ηn2 + kX∗ ∗k(cid:0)1 +Xn≥2 ξ2 n(cid:1) ess sup T\σ h0. It follows from (5.34) and (5.40) that the mapping Wt defined on H ∞ by (5.37) is extended onto H 2 for every t > 0. Denoting the extension by the same letter, we obtain Wt ∈ L(H 2,N ) such that WtS = TN Wt and kWtk ≤ C2 for every t > 0. (5.39) (5.40) kX∗ Set (5.41) Put (5.42) αt =Xn∈N η2 n ωnϕngh01 ψnt ωnϕngh01 η2 n and α =Xn∈N for every t > 0 and h ∈ H 2. ψ1n on T. By (5.37), (5.30), (5.29), (5.11), (5.12), (5.4), (5.35), (5.43) XWth = αth By (5.27), (5.20), (5.22), (5.18), (5.15), (5.28) and Lemma 3.1, αt and α satisfy Proposition 3.2. By Proposition 3.2, (5.44) t>0 kαtkL∞(σ) < ∞ and kαt − αkL2(σ) → 0 when t → 0. sup SHIFT-TYPE INVARIANT SUBSPACES 21 An analog of (5.44) for T \ σ follows from (5.28), (5.27), (5.5), (5.6), (5.9) and Lemma 3.1. Therefore, (5.45) t>0 kαtkL∞(T) < ∞ and kαt − αkL2(T) → 0 when t → 0. sup It follows from (5.41) that there exist a sequence {tj}j and W ∈ L(H 2,N ) such that tj →j 0 and Wtj →j W in the weak operator topology, W S = TN W and kWk ≤ C2. It follows from (5.43) and (5.45) that (5.46) XW h = αh for every h ∈ H 2. We will show that (5.47) We have σ = ∪n∈Nτn. Fix n ∈ N. By (5.20), (5.22), (5.25), (5.27), taking into account that f = gh01 a.e. on σ by (5.18) and (5.15), we obtain 1/α ∈ L∞(T). ηk2ϕkf ψk ≤ (1 + ε0)kXkKCpol,T Xk∈N,k6=n ξ2 k ≤ (1 + ε0)kXkKCpol,T ξ a.e. on τn. From the latest estimate, (5.27) and (5.21) we conclude that η2 k (cid:12)(cid:12)(cid:12) Xk∈N,k6=n ωkϕkf ψ1k (cid:12)(cid:12)(cid:12) ≤ Xk∈N,k6=n ψn −(cid:12)(cid:12)(cid:12) Xk∈N,k6=n α ≥ ϕnf η2 k ϕkf ψk (cid:12)(cid:12)(cid:12) ≥ (1 − ε0)γ0δ0 − (1 + ε0)kXkKCpol,T ξ a.e. on τn. It follows from (5.26) that (5.48) It follows from (5.27) that 1/α ∈ L∞(σ). (cid:12)(cid:12)(cid:12)Xn∈N η2 n ωnϕn ψ1n (cid:12)(cid:12)(cid:12) ≥(cid:12)(cid:12)(cid:12)η2 1 ω1ϕ1 ψ11 (cid:12)(cid:12)(cid:12) −Xn≥2(cid:12)(cid:12)(cid:12)η2 n ωnϕn ψ1n (cid:12)(cid:12)(cid:12) = 1 −Xn≥2 ξ2 n on T \ σ. It follows from (5.25), (5.6), (5.9) and the second inequality in (5.5) that (5.49) 1/α ∈ L∞(T \ σ). Now (5.47) follows from (5.48) and (5.49). Define A ∈ L(L2(T)) by the formula Ah = h/α, h ∈ L2(T). Set Z = AX. It follows from (5.46) that ZW h = h for every h ∈ H 2. Set M = W H 2. Then M ⊂ N and TM ≈ S. The conclusion of Theorem 1.1 for T0 follows from the relation T0 = TN . 22 M.F. GAMAL' 6. Proof of Corollary 1.2 Lemma 6.1. Suppose that T ∈ L(H) is an a.c. polynomially bounded op- erator, X ∈ L(H, H 2), Y ∈ L(H 2,H), XT = SX, Y S = T Y , clos XH = H 2 and clos Y H 2 = H. Then ker X = {0} and ker Y = {0}. Proof. Since XY S = SXY and clos XY H 2 = H 2, there exists an outer function g ∈ H ∞ such that XY = g(S). Therefore, ker Y = {0}. Since Y XY = Y g(S) = g(T )Y and clos Y H 2 = H, we obtain that Y X = g(T ). By [M], ker g(T ) = {0}. Therefore, ker X = {0}. Lemma 6.2. Suppose that T is a cyclic a.c. polynomially bounded operator (cid:3) such that T d ≺ S. Then T ∼ S. d Proof. Since T is cyclic, it follows from Corollary 2.1 that S 6.1, T ∼ S. Lemma 6.3. Suppose that T is a cyclic a.c. polynomially bounded operator ≺ T . By Lemma (cid:3) d ≺ UT. Then T (a) ∼= UT. such that T Proof. Since T is an a.c. polynomially bounded operator, T (a) is a.c. by [K3, Proposition 15]. By [K1, Theorem 2], T (a) contains UT as an orthogonal summand. Since T (a) is a.c. and cyclic, we conclude that T (a) ∼= UT. (cid:3) Lemma 6.4. Suppose that T ∈ L(H) is an a.c. polynomially bounded op- erator, T =(cid:18)T0 ∗O T1(cid:19) with respect to some decomposition of H, and there exists M1 ∈ Lat T1 such that T1M1 ∼ S. Then there exists M ∈ Lat T such that TM ∼ S. Proof. Let x be a cyclic vector for T1M1. Set M = ∨n≥0T nx. It is easy to see that PM1M realizes the relation TM ≺ T1M1. By Lemma 6.2, TM ∼ S. Lemma 6.5. Suppose that T ∈ L(H) is an a.c. polynomially bounded op- erator, σ ⊂ T is a measurable set, X ∈ L(H, L2(σ)), XT = UσX and Wk≥0 U −k σ XH = L2(σ). Then there exists x ∈ H such that Xx =: f 6= 0 a.e. on σ. Proof. Let {xn}N n=1 be such that xn ∈ H, kxnk = 1 for all n, and (cid:3) d where 1 ≤ N ≤ ∞. Set fn = Xxn. Then T kxn = H, N_n=1_k≥0 {ζ ∈ σ : fn(ζ) = 0}(cid:17) = 0. m(cid:16) N\n=1 Take {an}N n=1 such that an > 0 for all n, andPN σ1n = {ζ ∈ σ : fn(ζ) > 1}. n=1 an < ∞. Set SHIFT-TYPE INVARIANT SUBSPACES 23 Let {ϕn}N n=1 ⊂ H ∞ be such that ϕn =(an an fn on T \ σ1n, on σ1n. Set x1n = ϕn(T )xn and f1n = Xx1n. Then kx1nk ≤ Cpol,T an and f1n ≤ an a.e. on σ. Set f0 = supn f1n. Then 0 < f0 < ∞ a.e. on σ. Take 0 < δ < 1. Set Set Then σ2n = {ζ ∈ σ : δf0(ζ) ≤ f1n(ζ)}. τ1 = σ21, τn = σ2n \ ∪n−1 k=1τk, n ≥ 2. τn = σ and τn ∩ τk = ∅, N[n=1 n=1 such that εn > 0 for all n, andPN Take {εn}N H ∞ be such that if n 6= k, n, k ≥ 1. n=1 εn < δ. Let {ηn}N n=1 ⊂ f ≥ f1n −Xk6=n NXk=1 εk(cid:17)f0 a.e. on τn. Thus, x satisfies the conclusion of the lemma. (cid:3) Theorem 6.6. Suppose that T1 is a polynomially bounded operator such that T1 ∼ S. Then for every c > 0 there exist M ∈ Lat T1 and W ∈ L(H 2,M) such that W is invertible, W S = T1MW and kWkkW −1k ≤ (1 + c)(cid:0)√2(K 2 + 2)KCpol,T1 + 1(cid:1)qK 2C 2 + 1KC 2 pol,T1 . pol,T1 Proof. Applying [K1, Theorem 1], it is easy to see that (T1)(a) + ∼= S. Denote by X+ the canonical intertwining mapping for T1 and S constructed using a Banach limit. Denote by H1 the space in which T1 acts. Let 1+c ≤ c3 < c2 < 1. By Lemma 4.2, there exists x1 ∈ H1 such that kx1k = 1 and kX+x1k ≥ c2. Set f1 = X+x1. Denote by ψx1 the function from Corollary 2.1 applied to T1 and x1. By Lemma 4.1, 1 (6.1) f1 ≤ kX+kKCpol,T1ψx1 a.e. on T. Since f1 ∈ H 2, we haveRT log ψx1dm > −∞. Taking into account Remark 2.3, we can assume that ψx1 is outer. Put g1 = f1/ψx1. Then g1 ∈ H ∞. Put Then We have on τn, on T \ τn. ηn(T )x1n. ηnf1n. εn x = ηn =(1 NXn=1 NXn=1 εkf1k ≥(cid:16)δ − f = 24 M.F. GAMAL' Acting as in Proposition 4.4, we obtain that (6.2) Namely, if f1 ≤ c3ψx1 a.e. on T, then kg1k∞ > c3. c2 2 ≤ kX+x1k2 = kf1k2 =ZT f12dm ≤ZT 3ψx12dm ≤ c2 c2 3, becauseRT ψx12dm ≤ 1, a contradiction with the choice of c2 and c3. Set N =Wk≥0 T k 1 x1, T0 = T1N , and X0 = X+N . Taking into account that Sclos X+N ∼= S and deleting on the inner factor of f1, we can assume that g1 is outer and we accept that X0 realizes the relation T0 ≺ S. By Corollary 2.1, Remark 2.3, and Lemma 6.1 there exists Y0 ∈ L(H 2,N ) such that kY0k ≤ KCpol,T1, Y0 realizes the relation S ≺ T0, and X0Y0 = g1(S). Clearly T0 ∈ L(N ) satisfies the assumption of Theorem 1.1. We prove Theorem 1.1 for T0 and obtain estimates of kWk and kZk for W and Z constructed in the proof of Theorem 1.1. (S)X0 and If g1 ≥ c3 a.e. on T, then set W = Y0. We have W −1 = 1 g1 . Therefore, we assume that kWkkW −1k ≤ (1 + c)KC 2 (6.3) pol,T1 m({ζ ∈ T : g1(ζ) ≥ c3}) < 1. Applying Proposition 4.7 to T0, Y0, X0, and g1, we obtain T ∈ L(N ⊕ H 2 −) and X ∈ L(N ⊕ H 2 −, L2(T)). Taking into account the estimates of kY0k, of kg1k∞ by (6.1), and evident inequality Cpol,T0 ≤ Cpol,T1, we obtain that (6.4) + 1. and kXk ≤ Cpol,T1qK 2C 2 pol,T1 Cpol,T ≤pK 2 + 2Cpol,T1 Since T0 is of class C10, we have τ = ∅, X∗ ∗ = g1(UT). Therefore, g = g1 and h01 = h0, where g and h01 are defined in (5.3) and (5.8), respectively (h0 will be chosen below; g and h01 will be used in (5.42)). We replace (5.5) by ∗ = Y0⊕IH 2 , and XX∗ − (6.5) It allows to improve the estimate of kZk constructed in the end of Theorem 1.1. By (6.3) and (6.2), 0 < m(σ) < 1. g1(ζ) < c3}. σ = {ζ ∈ T : Set h0(ζ) = 1 (ζ ∈ T). Define x0 as in (5.14). Namely, set x0 = X∗ Take 0 < ε00, γ0, δ0 < 1 and apply Proposition 4.5 as in the proof of Theorem 1.1. The first inequality in (5.5) and the choice of δ0 in the proof of Theorem 1.1 was used to apply Lemmas 4.6 and 4.3 and obtain that ψn constructed by Corollary 2.1 applied to ϕn(T )x0 satisfy (5.19). In the case T0 ∼ S considered now (5.20) implies ∗ χσ. Take t0 > 0 and set log ψndm > −∞ for every n ∈ N. Zσ ψ0n =(ψn max(ψn, t0) on σ, on T \ σ. SHIFT-TYPE INVARIANT SUBSPACES 25 Then ψ0n satisfy (5.19) for all n ∈ N. Repeat the remaining part of the proof of Theorem 1.1 with ψ0n instead of ψn. We obtain the constants C1, C2, the function α defined as in (5.42) which satisfies (5.47) and intertwining transformations W and Z with the following estimates: C1 = kX∗ ∗k KCpol,T (1 + ε0) ≤ 1 (1 + ε0) , because kX∗k ≤ KCpol,T1 ≤ KCpol,T ; (2(1 + c2 1 − c1 ξ2 C2 ≤ K 2C 2 pol,T (1 + ε0) + KCpol,T1(cid:0)1 +Xn≥2 n(cid:1); 1))1/2 max(cid:0)(1 +Xn≥2 ξ2 n)C1, 1 + ξ(cid:1) kWk ≤ C2; kZk ≤(cid:13)(cid:13)(cid:13) 1 α(cid:13)(cid:13)(cid:13)∞kXk. Choosing c1, ε00, {ξn}n∈N close to 0 and c3, γ0, δ0 close to 1, taking into account that ε0 ≤ ε00, the definition (5.25) of ξ, the estimates before (5.48) and (5.49), the choice of σ by (6.5), and (6.4), we obtain the conclusion of (cid:3) the theorem. Remark 6.7. There exist operators T which satisfy the assumptions of Theorem 1.1 and have non-trivial factorization of the form (cid:18)C·1 see [SFBK, Sec. IX.2]. Therefore, to obtain estimates of the norms of intertwining transformations in terms of Cpol,T , a separate consideration of operators of class C10 is needed in the present construction. O C·0(cid:19), ∗ Proof of Corollary 1.2. We have T = Ta ∔ Ts, where Ta is a.c. and Ts is similar to a singular unitary operator, see [M] or [K3]. Since there is no nonzero transformation intertwining a.c. and singular unitaries, we conclude d that Ta ≺ UT. Let x be from Lemma 6.5 applied to Ta (with σ = T). Set N0 = _k≥0 T k a x. d Then N0 ∈ Lat T , TN0 is a cyclic a.c. polynomially bounded operator and TN0 ≺ UT. By Lemma 6.3,(cid:0)TN0(cid:1)(a) ∼= UT. Let ∗O T1(cid:19) , TN0 =(cid:18)T0 1 ∼=(cid:0)TN0(cid:1)(a). Since TN0 is cyclic and a.c., where T0 and T1 are of classes C0· and C1·, respectively, see Introduction and [K1]. By [K1, Theorem 3], T (a) we have T1 is cyclic and a.c., too. By Theorem 1.1 applied to T1, there exists M1 ∈ Lat T1 such that T1M1 ≈ S. By Lemma 6.4, there exists M ∈ Lat T such that TM ∼ S. It remains to apply Theorem 6.6 to TM. (cid:3) 26 M.F. GAMAL' References [BP] H. Bercovici and B. Prunaru, Quasiaffine transforms of polynomially bounded oper- ators, Arch. Math. (Basel), 71 (1998), 384 -- 387. [BT] H. Bercovici and K. Takahashi, On the reflexivity of contractions on Hilbert spaces, J. London Math. Soc. (2), 32 (1985), 149 -- 156. [B] J. Bourgain, New Banach space properties of the disc algebra and H ∞, Acta Math., 152 (1984), 1 -- 48. [C] J. B. Conway, The theory of subnormal operators, Math. Surveys and Monographs, 36, AMS, 1991. [F] S. R. Foguel, A counterexample to a problem of Sz.-Nagy, Proc. Amer. Math. Soc., 15 (1964), 788-790. [G] M. F. Gamal', Examples of cyclic polynomially bounded operators that are not similar to contractions, Acta Sci. Math. (Szeged), 82 (2016), 597 -- 628. [K1] L. K´erchy, Isometric asymptotes of power bounded operators, Indiana Univ. Math. J., 38 (1989), 173 -- 188. [K2] L. K´erchy, Shift-type invariant subspaces of contractions, J. Funct. Anal., 246 (2007), 281 -- 301. [K3] L. K´erchy, Quasianalytic polynomially bounded operators, in: Operator Theory: the State of the Art, Theta, Bucharest, 2016, 75 -- 101. [K4] L. K´erchy, Quasianalytic n-tuples of Hilbert space operators, J. Operator Theory, 81 (2019), 3 -- 20. [M] W. Mlak, Algebraic polynomially bounded operators, Ann. Polon. Math., 29 (1974), 133 -- 139. [N] N. K. Nikolski, Operators, functions, and systems: an easy reading. Volume I: Hardy, Hankel, and Toeplitz, Math. Surveys and Monographs, 92, AMS, 2002. [P] G. Pisier, A polynomially bounded operator on Hilbert space which is not similar to a contraction, J. Amer. Math. Soc., 10 (1997), no. 2, 351 -- 369. [R] O. R´ejasse, Factorization and reflexivity results on polynomially bounded operators, J. Oper. Theory, 60(2008), 219 -- 238. [SFBK] B. Sz.-Nagy, C. Foias, H. Bercovici and L. K´erchy, Harmonic analysis of operators on Hilbert space, Springer, New York, 2010. [T1] K. Takahashi, The reflexivity of contractions with non-reductive ∗-residual parts, Michigan Math. J., 34(1987), 153 -- 159. [T2] K. Takahashi, On the reflexivity of contractions with isometric parts, Acta Sci. Math. (Szeged), 53(1989), 147 -- 152. M.F. GAMAL', St. Petersburg Branch, V. A. Steklov Institute of Math- ematics, Russian Academy of Sciences, Fontanka 27, St. Petersburg, 191023, Russia E-mail address: [email protected]
1102.3998	1	1102	2011-02-19T16:20:16	Almost periodic generalized functions	[ "math.FA", "math.CA" ]	The aim of this paper is to introduce and to study an algebra of almost periodic generalized functions containing the classical Bohr almost periodic functions as well as almost periodic Schwartz distributions	math.FA	math	ALMOST PERIODIC GENERALIZED FUNCTIONS CHIKH BOUZAR AND MOHAMMED TAHA KHALLADI Abstract. The aim of this paper is to introduce and to study an algebra of almost periodic generalized functions containing the classical Bohr almost periodic functions as well as almost periodic Schwartz distributions. 1. Introduction The theory of uniformly almost periodic functions was introduced and studied by H. Bohr, since then, many authors contributed to the development of this theory. There exist three equivalent definitions of uniformly almost periodic functions, the first definition of H. Bohr, S. Bochner's definition and the definition based on the approximation property, see [1]. S. Bochner's definition is more suitable for exten- sion to Schwartz distributions. L. Schwartz in [6] introduced the basic elements of almost periodic distributions. The new generalized functions of [2], [3], give a solution to the prob- lem of multiplication of distributions, these generalized functions are currently the subject of many scientific works, see [4] and [5]. The aim of this work is to introduce and to study an algebra of almost periodic generalized functions containing the classical Bohr al- most periodic functions as well almost periodic Schwartz distributions. 2. Almost periodic functions and distributions We consider functions and distributions defined on the whole one dimensional space R. Recall Cb the space of bounded and continuous complex valued functions on R endowed with the norm k k∞ of uniform convergence on R, (Cb, k k∞) is a Banach algebra. Definition 1. (S. Bochner) A complex valued function f defined and continuous on R is called almost periodic, if for any sequence of real numbers (hn)n one can extract a subsequence (hnk)k such that 1991 Mathematics Subject Classification. Primary 46F30; Secondary 42A75. Key words and phrases. Almost periodic functions; Almost periodic distribu- tions; Colombeau algebra; Almost periodic generalized functions. 1 2 CHIKH BOUZAR AND MOHAMMED TAHA KHALLADI (f (. + hnk))k converges in (Cb, k k∞). Denoted by Cap the space of al- most periodic functions. To recall Schwartz almost periodic distributions, we need some func- tion spaces, see [6]. Let p ∈ [1, +∞] , the space endowed with the topology defined by the countable family of norms DLp :=(cid:8)ϕ ∈ C∞ : ϕ(j) ∈ Lp, ∀j ∈ Z+(cid:9) ϕk,p :=Xj≤k(cid:13)(cid:13)ϕ(j)(cid:13)(cid:13)Lp , k ∈ Z+, is a differential Frechet subalgebra of C∞. The topological dual of DL1, denoted by D′ L∞, is called the space of bounded distributions. Let h ∈ R and T ∈ D′, the translate of T by h, denoted by τhT , is defined as : hτhT , ϕi = hT , τ−hϕi , ϕ ∈ D, where τ−hϕ (x) = ϕ (x + h) . The definition and characterizations of an almost periodic distribu- tion are summarized in the following results. Theorem 1. For any bounded distribution T ∈ D′ statements are equivalent : L∞, the following i) The set {τhT, h ∈ R} is relatively compact in D′ ii) T ∗ ϕ ∈ Cap, ∀ ϕ ∈ D. L∞. iii) ∃ (fj)j≤k ⊂ Cap, T = Pj≤k T ∈ D′ f (j) j . L∞ is said almost periodic if it satisfies any (hence every) of the above conditions. Definition 2. The space of almost periodic distributions is denoted by B′ ap. Let recall the space of regular almost periodic functions. Definition 3. The space of almost periodic infinitely differentiable functions on R is defined and denoted by Bap =(cid:8)ϕ ∈ DL∞ : ϕ(j) ∈ Cap , ∀j ∈ Z+(cid:9) . Some, easy to prove, properties of Bap are given in the following assertions. Proposition 1. We have i) Bap is a closed differential subalgebra of DL∞. ii) If T ∈ B′ iii) Bap ∗ L1 ⊂ Bap . iv) Bap = DL∞ ∩ Cap . ap and ϕ ∈ Bap , then ϕT ∈ B′ ap. ALMOST PERIODIC GENERALIZED FUNCTIONS 3 As a consequence of (iv) , we have the following result. Corollary 1. If v ∈ DL∞ and v ∗ ϕ ∈ Cap, ∀ϕ ∈ D, then v ∈ Bap. Remark 1. It is important to mention that Bap$C∞ ∩ Cap. 3. Almost periodic generalized functions Let I = ]0, 1] and ML∞ =n(uε)ε ∈ (DL∞)I , ∀k ∈ Z+, ∃m ∈ Z+, NL∞ =n(uε)ε ∈ (DL∞)I , ∀k ∈ Z+, ∀m ∈ Z+, uεk,∞ = O(cid:0)ε−m(cid:1) , ε −→ 0o uεk,∞ = O (εm) , ε −→ 0o Definition 4. The algebra of bounded generalized functions, denoted by GL∞, is defined by the quotient GL∞ = ML∞ NL∞ Define (3.1) Map =n(uε)ε ∈ (Bap)I , ∀k ∈ Z+, ∃m ∈ Z+, uεk,∞ = O (ε−m) , ε −→ 0o Nap =n(uε)ε ∈ (Bap)I , ∀k ∈ Z+, ∀m ∈ Z+, uεk,∞ = O (εm) , ε −→ 0o The properties of Map and Nap are summarized in the following proposition. Proposition 2. i) The space Map is a subalgebra of (Bap)I . ii) the space Nap is an ideal of Map. Proof. i) It follows from the fact that Bap is an differential algebra. ii) Let (uε)ε ∈ Nap and (vε)ε ∈ Map, we have ∀k ∈ Z+, ∃m′ ∈ Z+, ∃c1 > 0, ∃ε0 ∈ I, ∀ε < ε0, vεk,∞ < c1ε−m′ . Take m ∈ Z+, then for m′′ = m+m′, ∃c2 > 0 such that uεk,∞ < c2εm′′ . Since the family of the norms uεk,∞ is compatible with the algebraic structure of DL∞, then ∀k ∈ Z+, ∃ck > 0 such that uεvεk,∞ ≤ ck uεk,∞ vεk,∞ , consequently uεvεk,∞ < ckc2εm ′′ c1ε−m′ ≤ Cεm, where C = c1c2ck. Hence (uεvε)ε ∈ Nap. (cid:3) The following definition introduces the algebra of almost periodic generalized functions. 4 CHIKH BOUZAR AND MOHAMMED TAHA KHALLADI Definition 5. The algebra of almost periodic generalized functions is the quotient algebra Gap = Map Nap We have a characterization of elements of Gap similar to the result (ii) of theorem (1) for almost periodic distributions. Theorem 2. Let u = [(uε)ε] ∈ GL∞, the following assertions are equiv- alent : i) u is almost periodic. ii) uε ∗ ϕ ∈ Bap, ∀ε ∈ I, ∀ϕ ∈ D. Proof. i) =⇒ ii) If u ∈ Gap, so for every ε ∈ I we have uε ∈ Bap, then uε ∗ ϕ ∈ Bap, ∀ε ∈ I, ∀ϕ ∈ D. ii) =⇒ i) Let (uε)ε ∈ ML∞ and uε ∗ ϕ ∈ Bap, ∀ε ∈ I, ∀ϕ ∈ D, therefor uε ∈ Bap follows from theorem (1) (ii); it suffices to show that ∀k ∈ Z+, ∃m ∈ Z+, uεk,∞ = O(cid:0)ε−m(cid:1) , ε −→ 0, which follows from the fact that u ∈ GL∞. (cid:3) Remark 2. The characterization (ii) does not depend on representa- tives. Definition 6. Denote by Σ the subset of functions ρ ∈ S satisfying Z ρ (x) dx = 1 and Z xkρ (x) dx = 0, ∀k = 1, 2, ... Set ρε (.) = 1 ε ρ(cid:0) . ε(cid:1) , ε > 0. Proposition 3. Let ρ ∈ Σ, the map iap : B′ ap −→ u −→ (u ∗ ρε)ε + Nap, Gap is a linear embedding which commutes with derivatives. Proof. Let u ∈ B′ we have u = Pβ≤m ap, by characterization of almost periodic distributions f (β) β , where fβ ∈ Cap, so ∀α ∈ Z, 1 consequently, there exists c > 0 such that (cid:12)(cid:12)(cid:0)u(α) ∗ ρε(cid:1) (x)(cid:12)(cid:12) ≤Xβ≤m x∈R(cid:12)(cid:12)(cid:0)u(α) ∗ ρε(cid:1) (x)(cid:12)(cid:12) ≤Xβ≤m εα+βZR (cid:12)(cid:12)fβ (x − εy) ρ(α+β) (y)(cid:12)(cid:12) dy, εα+β kfβkL∞(R)ZR (cid:12)(cid:12)ρ(α+β) (y)(cid:12)(cid:12) dy ≤ sup 1 c , εα+m ALMOST PERIODIC GENERALIZED FUNCTIONS 5 i.e. u ∗ ρεm′,∞ = Xα≤m′ sup x∈R(cid:12)(cid:12)(cid:0)u(α) ∗ ρε(cid:1) (x)(cid:12)(cid:12) ≤ c′ εm+m′ , c′ = Xα≤m′ c εα , u∗ρε = u in D′ L∞, but lim ε−→0 u∗ρε = 0 this shows that (u ∗ ρε)ε ∈ Map. Let (u ∗ ρε)ε ∈ Nap, then lim ε−→0 in D′ L∞, this shows that iap is an embedding. Finally we note that iap is linear, this results from the fact that the ε = (cid:3) convolution is linear and that iap(cid:0)w(j)(cid:1) =(cid:0)w(j) ∗ ρε(cid:1)ε = (w ∗ ρε)(j) (iap (w))(j) . The space Bap is embedded into Gap canonically, i.e. σap : Bap −→ Gap f −→ [(f )ε] = (f )ε + Nap There is two ways to embed f ∈ Bap into Gap. Actually we have the same result. Proposition 4. The following diagram Bap −→ B′ ap σap ց ↓ iap Gap is commutative. Proof. Let f ∈ Bap, we prove that (f ∗ ρε − f )ε ∈ Nap. By Taylor's formula and the fact that ρ ∈ Σ, we obtain kf ∗ ρε − f kL∞ ≤ εmsup (−y)m m! x∈RZR (cid:12)(cid:12)(cid:12)(cid:12) f (m) (x − θ (x) εy) ρ (y) dy(cid:12)(cid:12)(cid:12)(cid:12) , then ∃Cm > 0, such that kf ∗ ρε − f kL∞ ≤ εmCm(cid:13)(cid:13)f (m)(cid:13)(cid:13)L∞ kymρkL1 . The same result can be obtained for all the derivatives of f. Hence (f ∗ ρε − f )ε ∈ Nap. (cid:3) The Colombeau algebra of tempered generalized functions on C is denoted GT (C) , for more details on GT (C) see [2] or [4]. Proposition 5. Let u ∈ Gap and F ∈ GT (C), then F ◦ u = [(F ◦ uε)ε] is a well defined element of Gap. 6 CHIKH BOUZAR AND MOHAMMED TAHA KHALLADI Proof. It follows from the classical case of composition, in context of Colombeau algebra, we have F ◦uε ∈ Bap in view of the classical results of composition and convolution. (cid:3) We recall a characterization of integrable distributions. Definition 7. A distribution v ∈ D′ is said an integrable distribution, denoted v ∈ D′ , where fi ∈ L1. f (i) i Proposition 6. If u = [(uε)ε] ∈ Gap and v ∈ D′ u ∗ v defined by L1, if and only if v =Pi≤l (u ∗ v) (x) =ZR uε (x − y) v (y) dyε is a well defined almost periodic generalized function. + N [C] L1, then the convolution Proof. Let (uε)ε ∈ Map be a representative of u, then ∀k ∈ Z+, ∃m ∈ Z+, ∃C > 0, ∃ε0 ∈ I, ∀ε ≤ ε0, uεk,∞ < Cε−m, since v ∈ D′ , where fi ∈ L1. For each ε ∈ I, uε ∗v is an almost periodic infinitely differentiable function. By Young inequality there exists C > 0 such that i f (i) L1 then v =Pi≤l (cid:13)(cid:13)(cid:13)(uε ∗ v)(j)(cid:13)(cid:13)(cid:13)L∞ kfikL1(cid:13)(cid:13)u(i+j) uε ∗ vk,∞ = O(cid:0)ε−m(cid:1) , ε −→ 0, ≤ CXi≤l ε (cid:13)(cid:13)L∞ , consequently this shows that (uε ∗ v)ε ∈ Map. Suppose that (wε)ε ∈ Map is another representative of u, then there exists C > 0 such that k(uε ∗ v − wε ∗ v)kL∞ ≤ Xi≤l ≤ CXi≤l as (uε − wε)ε ∈ Nap, so ∀m ∈ Z+, sup R ZR (cid:12)(cid:12)(cid:12)(uε − wε)(i) (x − y)(cid:12)(cid:12)(cid:12) fi (y) dy kfikL1(cid:13)(cid:13)(cid:13)(uε − wε)(i)(cid:13)(cid:13)(cid:13)L∞ , (uε ∗ v − wε ∗ v) (x) = O (εm) , ε −→ 0. We obtain the same result for (uε ∗ v − wε ∗ v)(j) Nap. ε . Hence (uε ∗ v − wε ∗ v)ε ∈ (cid:3) If u = [(uε)ε] ∈ Gap, taking the integral of each element uε on a compact, we obtain an element of CI . ALMOST PERIODIC GENERALIZED FUNCTIONS 7 Definition 8. Let u = [(uε)ε] ∈ Gap and x0 ∈ R, define the primitive of u by U (x) = xZx0 uε (t) dtε + N [C] . We give a generalized version of the classical Bohl-Bohr theorem. Proposition 7. The primitive of an almost periodic generalized func- tion is almost periodic if and only if it is bounded generalized function. Proof. (=⇒) : It follows from the fact that Gap ⊂ GL∞. (⇐=) : Let u = [(uε)ε] ∈ Gap and let U = [(Uε)ε] be its primitive, since for each uε (t) dt ∈ DL∞ then by the classical result ε ∈ I, uε ∈ Bap and Uε = xRx0 uε (t) dt ∈ Bap. We xRx0 of Bohl-Bohr and Bap, for every ε ∈ I we have shows that (Uε)ε ∈ Map, i.e. ∀k ∈ Z+, ∃m ∈ Z+, Uεk,∞ =Xj≤k ε sup x∈R(cid:12)(cid:12)U (j) (x)(cid:12)(cid:12) = O(cid:0)ε−m(cid:1) , ε −→ 0. If j = 0, since (Uε)ε ∈ ML∞. By hypothesis, we have sup x∈R Uε (x) = O(cid:0)ε−m(cid:1) , ε −→ 0, (x)(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12)u(j−1) which shows that (Uε)ε ∈ Map. If j ≥ 1, we have(cid:12)(cid:12)(cid:12)U (j) which gives Xj≤k x∈R(cid:12)(cid:12)U (j) (x)(cid:12)(cid:12) , x∈R(cid:12)(cid:12)u(j−1) Uεk,∞ = O(cid:0)ε−m(cid:1) , ε −→ 0, (x)(cid:12)(cid:12) ≤Xj≤k i.e. (Uε)ε ∈ Map. consequently sup sup ε ε ε ε (x)(cid:12)(cid:12)(cid:12) , (cid:3) As in the classical theory, we introduce the notion of mean value within the algebra Gap. Definition 9. Let u ∈ Gap, the generalized mean value of u, denoted by Mg (u) , is defined by Mg (u) = lim X−→+∞ 1 X XZ0 uε (x) dxε + N [C] , where (uε)ε is a representative of u. 8 CHIKH BOUZAR AND MOHAMMED TAHA KHALLADI The definition of Mg (u) is correct and does not depend on represen- tatives. uε (x) = O (ε−m) , ε −→ 0, hence ≤ sup x∈R uε (x) , lim X→+∞ 1 X XZ0 Proof. i) Let ε ∈ I, we have as (uε)ε ∈ Map, so ∃m ∈ Z+, sup x∈R Proposition 8. i) If u = [(uε)ε] ∈ Gap, then Mg (u) ∈eC. ii) If (uε)ε ∈ Nap, then Mg (u) = 0 in eC. (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) uε (x) dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) uε (x) dxε  lim XZ0 uε (x) dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) XZ0 then Mg (u) = 0 in eC. ii) If (uε)ε ∈ Nap, i.e. ∈ EM [C] . lim X→+∞ 1 X (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 1 X X→+∞ distribution as stated in the following. = O (εm) , ε −→ 0, ∀m ∈ Z+, We have compatibility of the generalized mean value with that of a (cid:3) Proposition 9. If T ∈ B′ ap, then Mg (iap (T )) = M (T ) in C. Proof. We have Mg (iap (T )) = lim X−→+∞ 1 X where ρ ∈ Σ and ρε (.) = 1 iap (T ) (h) dh = lim XZ0 ε(cid:1) . Let ϕ ∈ D andRR ε ρ(cid:0) . X−→+∞ 1 X XZ0 (T ∗ ρε) (h) dhε , ϕ (x) dx = 1, then M (T ) = lim X−→+∞ 1 X (T ∗ ϕ) (h) dh. XZ0 We have Mg (iap (T )) − M (T ) = M (T ∗ (ρε − ϕ)) , ∀ε ∈ I. ALMOST PERIODIC GENERALIZED FUNCTIONS 9 In view of formula (V I.9.2) of [6], we obtain Mg (iap (T )) − M (T ) = M (T )ZR (ρε (x) − ϕ (x)) dx, as ∀ε ∈ I,RR C. (ρε (x) − ϕ (x)) dx = 0, then Mg (iap (T )) = M (T ) in (cid:3) Remark 3. We can introduce a new association within Gap with the aid of the generalized mean value Mg. Definition 10. A generalized trigonometric polynomial is a generalized function [(Pε)ε] , where Pε (x) = cε,neiλε,nx , lXn=1 (cε,n)ε ∈eC and (λε,n)ε ∈eR, n = 1, ..., l. Proposition 10. Every generalized trigonometric polynomial is an al- most periodic generalized function. Proof. Let Pε (x) = n = 1, ..., l, we have lPn=1 cε,neiλε,nx where (cε,n)ε ∈ eC and (λε,n)ε ∈ eR, ∀ε ∈ I, cε,neiλε,nx ∈ Bap, lXn=1 λε,nj cε,n 0. Consequently ∀k ∈ Z+, ∃C > 0 such that moreover ∃m ∈ Z+, ∃m′ ∈ Z+, λε,n = O (ε−m) and cε,n = O(cid:0)ε−m′(cid:1) , ε −→ lXn=1 Pεk,∞ ≤ Xj≤k λε,1j + λε,2j + ... + λε,lj! c′ε−m′ ≤ Xj≤k ≤ k(cid:0)cε−m + c2ε−2m + ... + ckε−km(cid:1) c′ε−m′ , where C = c′kl(cid:0)c + c2 + ... + ck(cid:1) , m′′ = km + m′, Let u ∈ Gap and eλ = [(λε)ε] ∈ eR, then ue−ieλx =(cid:0)uεe−iλεx(cid:1)ε ∈ Gap, so the generalized mean value Mg(cid:16)ue−ieλx(cid:17) is a well defined element of this show that (Pε)ε ∈ Map. In a similar way we show that if (λε,n)ε ∈ N [R] and (cε,n)ε ∈ N [C] , then (Pε (x))ε ∈ Nap. (cid:3) ≤ Cε−m′′ 10 CHIKH BOUZAR AND MOHAMMED TAHA KHALLADI eC. Define and the generalized spectra of u, aeλ (u) = Mg(cid:16)ue−ieλx(cid:17) , Λg (u) =neλ ∈eR : aeλ (u) 6= 0 in eCo . Remark 4. If a function or a distribution u is almost periodic, then its spectra is at most countable, see [1] and [6]. Proposition 11. Let P =(cid:20)(cid:18) lPn=1 metric polynomial, then cε,neiλε,nx(cid:19)ε(cid:21) be a generalized trigono- Proof. A direct computation gives the result. Λg (P ) =(cid:8)(cid:2)(λε,n)ε(cid:3) : n = 1, ..., l(cid:9) . (cid:3) Remark 5. We are indebted to the referee for the following example of an almost periodic generalized function having uncountable generalized spectra. Let u = [(uε)ε] , uε (x) = 1 + eix, ∀ε ∈ I, by proposition (11) , Λg (u) = {[(Aε)ε] : Aε = {0, 1}} , as the set {(λε)ε : λε = {0, 1} , ε ∈ I} ⊂ Λg (u) is uncountable, then the generalized spectra of u is too. References [1] H. Bohr. Almost periodic functions. Chelsea Publishing Company, 1947. [2] J. F. Colombeau. New generalized functions and multiplication of distributions. North-Holland, 1984. [3] J. F. Colombeau. Elementary Introduction to New Generalized Functions. North-Holland, 1985. [4] M. Grosser, M. Kuzinger, M. Oberguggenberger,R. Steinbauer. Geometric the- ory of generalised functions. Kluwer, 2001. [5] A. Kaminski, M. Oberguggenberger, S. Pilipovic (Eds). Linear and non-linear theory of generalized functions and its applications. Banach Center Publica- tions, Vol. 88, 2010. [6] L. Schwartz. Th´eorie des distributions. Hermann, 1966. Oran Essenia University, Algeria E-mail address: [email protected] University of Adrar, Algeria E-mail address: [email protected]
1801.06884	3	1801	2018-03-22T14:39:19	On the Existence of Normal Square and Nth Roots of Operators	[ "math.FA", "math.OA" ]	The primary purpose of this paper is to show the existence of normal square and nth roots of some classes of bounded operators on Hilbert spaces. Two interesting simple results hold. Namely, under simple conditions we show that if any operator $T$ is such that $T^2=0$, then this implies that $T$ is normal and so $T=0$. Also, we will see when the square root of an arbitrary bounded operator is normal.	math.FA	math	ON THE EXISTENCE OF NORMAL SQUARE AND NTH ROOTS OF OPERATORS MOHAMMED HICHEM MORTAD Abstract. The primary purpose of this paper is to show the existence of normal square and nth roots of some classes of bounded operators on Hilbert spaces. Two interesting simple results hold. Namely, under simple conditions we show that if any operator T is such that T 2 = 0, then this implies that T is normal and so T = 0. Also, we will see when the square root of an arbitrary bounded operator is normal. 1. Introduction Let H be a complex Hilbert space, and let B(H) denote the algebra of all bounded linear operators on H. By a square root of A ∈ B(H), we mean a B ∈ B(H) such that B2 = A. An A ∈ B(H) is called positive if < Ax, x >≥ 0, ∀x ∈ H. 1 It is known that any positive operator A has a unique positive square root (which is denoted by √A or A 2 and these notations are exclusively reserved to the unique positive square root). However, there are non normal N such that N 2 is normal. As an extreme example, just consider any non-normal N such that N 2 = 0. In fact, the identity 2× 2 matrix I has an infinite number of self-adjoint square roots. Indeed, the self-adjoint x Ax = (cid:18) √1 − x2 −x represents a square root of I for each x ∈ [−1, 1]. There are some quite known research papers on this topic. For instance, readers may wish to consult: [2], [4], [7], [8], [9] and [12]. In particular, Putnam [7] gave a condition guaranteeing that the square root of a normal operator be normal. The problem considered in this paper is of this sort. √1 − x2 (cid:19) self-adjoint. As it is customary, we denote A by ReT and B by ImT . Recall that any T ∈ B(H) is expressible as T = A + iB where A, B ∈ B(H) are It is also known (cf. [6]) that T = 0 iff A = B = 0. Therefore, if A+iB = C +iD, It is also easily then A = C and B = D whenever A, B, C, D are self-adjoint. verifiable that T is normal iff AB = BA. We also recall some known results which will be called on below (these are standard facts, see [6] for proofs). 2010 Mathematics Subject Classification. Primary 47A62, Secondary 47A05. Key words and phrases. Square roots. Normal, self-adjoint and positive operators. Real and imaginary parts of an operator. Spectrum. Cartesian decomposition. 1 2 M. H. MORTAD Theorem 1.1. Let A, B ∈ B(H) be self-adjoint. Then: √B 0 ≤ A ≤ B =⇒ √A ≤ Proposition 1.2. Let A, B ∈ B(H) be self-adjoint. Then where A = √A∗A. A ≤ B =⇒ −B ≤ A ≤ B, In the end, we assume that readers are familiar with other notions and results on B(H). 2. Existence of Normal Roots Theorem 2.1. Let C ∈ B(H) be self-adjoint and let T = A + iB ∈ B(H) be such that T 2 = C. (1) If σ(A) ∩ σ(−A) = ∅, then T is self-adjoint and invertible. (2) If σ(B) ∩ σ(−B) = ∅, then T is skew symmetric (that is, T ∗ = −T ) and invertible. The proof is based upon the following known result: Lemma 2.2. ([10] or [1]) Let A, B ∈ B(H) be such that σ(A) ∩ σ(B) = ∅. Then the equation AX − XB = S has a unique solution X (in B(H)) for each S ∈ B(H). Now, we prove Theorem 2.1: Proof. We have T 2 = C ⇐⇒ T 2 = A2 − B2 + i(AB + BA) = C. By the self-adjointness of A and B, we obtain the self-adjointness of A2 − B2 and AB + BA as well. Hence (cid:26) A2 − B2 = C, AB + BA = 0. (1) If σ(A) ∩ σ(−A) = ∅, then Lemma 2.2 says that the equation AB − B(−A) = AB + BA = 0 has a unique solution which is necessarily B = 0. Hence T = A is self- adjoint. If 0 ∈ σ(T ), then 0 ∈ σ(A) and so 0 ∈ σ(−A) too. This, however, would violate the assumption σ(A)∩σ(−A) = ∅. Therefore, T is invertible. (2) When σ(B) ∩ σ(−B) = ∅, a similar method yields A = 0, i.e. T = iB. Hence T is invertible because σ(B) ∩ σ(−B) = ∅. (cid:3) A result by Embry [3] may be readjusted as follows: Lemma 2.3. Let A, B ∈ B(H) be normal and such that AB = −BA. Designate the numerical range of A by W (A). If σ(A) ∩ σ(−A) = ∅ or 0 6∈ W (A) (resp. σ(B) ∩ σ(−B) = ∅ or 0 6∈ W (B)), then B = 0 (resp. A = 0). Using the same method as above and the foregoing lemma, we may establish: Proposition 2.4. Let C ∈ B(H) be self-adjoint and let T = A + iB ∈ B(H) be such that T 2 = C. If 0 6∈ W (A) (or 0 6∈ W (B)), then T is self-adjoint. As above, in the first case T = A and in the second T = iB. ROOTS OF OPERATORS 3 We know that if T ∈ B(H), then T 2 = 0 does not, in general, imply that T = 0. We also know that if T satisfies kT 2k = kTk2 (for example, if T is self-adjoint or normal or normaloid in general i.e. kT nk = kTkn for all n), then T 2 = 0 does imply that T = 0. The following result is therefore of interest. Proposition 2.5. Let T ∈ B(H) be such that T 2 = 0. If ReT ≥ 0 (or ImT ≥ 0), then T is normal and so T = 0. Proof. Write T = A + iB where A, B ∈ B(H) are self-adjoint where A = ReT and B = ImT . Then clearly So, if T 2 = 0, then T 2 = A2 − B2 + i(AB + BA). A2 − B2 + i(AB + BA) = 0 =⇒ (cid:26) A2 = B2, AB = −BA. Hence, if A ≥ 0 (a similar argument works when B ≥ 0), then AB = −BA =⇒ A2B = −ABA = BA2 =⇒ AB = BA. Therefore, T is normal. Accordingly, kTk2 = kT 2k = 0 =⇒ T = 0, as suggested. Corollary 2.6. Let T ∈ B(H) be such that T 2 = 0. If any of σ(ReT ) or σ(ImT ) is a subset of either R+ or R−, then T is normal and so T = 0. Remark. The condition ReT ≥ 0 (or ImT ≥ 0) in Proposition 2.5 is, in general, not sufficient to make T normaloid. Indeed, if V is the Volterra operator on L2[0, 1] say, then (cf. [6]) V is not normaloid as (cid:3) r(V ) = 0 6= 2 π = kV k where r(V ) denotes the spectral radius of V . It can, however, easily be checked that ReV ≥ 0. We finish this section with a general criterion guaranteeing the normality of the square root (this generalizes Proposition 2.5): Theorem 2.7. Let S = C + iD ∈ B(H) be such that T 2 = S where T = A + iB ∈ B(H). Then (if [·,·] denotes the usual commutator) [B, C] = [A, D] and [A, C] = [B, D]. (Consequently, BC = CB ⇔ AD = DA and AC = CA ⇔ BD = DB). If A ≥ 0 (or A ≤ 0), then In particular, if S is self-adjoint (i.e. D = 0) and A ≥ 0 or A ≤ 0 or B ≥ 0 or B ≤ 0, then T is always normal. If B ≥ 0 (or B ≤ 0), then T is normal ⇐⇒ AD = DA. T is normal ⇐⇒ BD = DB. 4 M. H. MORTAD Proof. By assumption, A2 − B2 + i(AB + BA) = C + iD =⇒ (cid:26) A2 − B2 = C, AB + BA = D. Hence and so AB + BA = D =⇒ A2B + ABA = AD and ABA + BA2 = DA Since A2 = B2 + C, we get Also, A2B − BA2 = AD − DA. BC − CB = AD − DA. AB + BA = D =⇒ AB2 + BAB = DB and BAB + B2A = BD and so as above and by invoking B2 = A2 − C, we obtain B2A − AB2 = BD − DB To show the last two assertions, we have from above that AC − CA = BD − DB. A2B = BA2 ⇐⇒ AD = DA and B2A = AB2 ⇐⇒ BD = DB. So if A ≥ 0, then clearly A2B = BA2 ⇐⇒ AB = BA, and the previous holds iff T = A + iB is normal. A similar reasoning applies when A ≤ 0. Finally, argue similarly in the event B ≥ 0 or B ≤ 0 and this completes the proof. (cid:3) 3. Explicit Construction of Roots It is known (cf. [11]) that if N is normal with a spectral integral Rσ(N ) λdE √λdE is square root of N where √λ is where E is a spectral measure, then Rσ(N ) a complex square root of λ. Under an extra condition on the normal operator, we can have an even more explicit formula. Theorem 3.1. Let N = C + iD ∈ B(H) be normal with either D ≥ 0 or D ≤ 0 (equivalently, σ(D) ⊂ R+ or σ(D) ⊂ R−). When D ≥ 0, then 1 2 1 2 1 2 1 2 T = (cid:18)N + C 2 (cid:19) + i(cid:18)N − C 2 (cid:19) is a normal square root of N . If D ≤ 0, then T = (cid:18)N + C 2 (cid:19) is another normal square root of N . − i(cid:18)N − C 2 (cid:19) ROOTS OF OPERATORS 5 Proof. The hardest part of the proof is the meticulousness! First C and D are self-adjoint. Besides, CD = DC as N is normal. Then N2 = N ∗N = (C − iD)(C + iD) = C2 + D2 ≥ C2 as D2 ≥ 0 because D is self-adjoint. By Theorem 1.1, we get N ≥ C. Hence, by Proposition 1.2 and so −N ≤ C ≤ N Therefore, it makes sense to define their positive square roots. Consider (the self- adjoint!) N − C ≥ 0 and N + C ≥ 0. A = (cid:18)N + C 2 (cid:19) 1 2 and B = (cid:18)N − C 2 (cid:19) 1 2 . Since N commutes with C, it follows that A commutes with B. Consequently, the operator M := A + iB is normal. Finally, (1) If D ≥ 0, then M 2 = (A + iB)(A + iB) = A2 − B2 + i(AB + BA) = N + C − N − C 2 2 + 2i(cid:18)N2 − C2 4 1 2 (cid:19) 1 2 = C + i(D2) = C + iD = N, that is M is a normal square root of N . (2) A similar argument applies when D ≤ 0. In this case, (M ∗)2 = N, that is, M ∗ is a normal square root of N . This marks the end of the proof. (cid:3) This approach, besides its explicit construction, does apply for higher powers of the type 2n. In other language, the algorithm prescribed in the previous proof may be applied to deal with biquadratic equations or in general equations of the form T 2n = N where n ∈ N. We have: Corollary 3.2. Let N = C + iD ∈ B(H) be normal with either D ≥ 0 (or D ≤ 0). Let T ∈ B(H) be such that T 4 = N . Then a normal 4th root of T is given by where T = (cid:18)S + ReS 2 (cid:19) S = (cid:18)N + C 2 (cid:19) 1 2 1 2 + i(cid:18)S − ReS 2 (cid:19) 1 2 + i(cid:18)N − C 2 (cid:19) 1 2 6 M. H. MORTAD Proof. Put S = T 2. Then T 4 = N becomes S2 = N . So, if e.g. N is such that σ(ImN ) ⊂ R+, then S is normal. Moreover, by Theorem 3.1, we know that S = (cid:18)N + C 2 (cid:19) and it is normal. Since clearly 1 2 + i(cid:18)N − C 2 (cid:19) 1 2 ImS = (cid:18)N − C 2 (cid:19) 1 2 ≥ 0, it follows that T 2 = S has a normal solution T given by T = (cid:18)S + ReS 2 (cid:19) 1 2 + i(cid:18)S − ReS 2 (cid:19) 1 2 . Remark. By induction, we know how to find a root of order 2n of normal operators very explicitly. We finish with the case of general nth roots. First, the following result should (cid:3) be readily verified. Lemma 3.3. Let A ∈ B(H). If k ∈ Z, then ei(A+2kπI) = eiA. Theorem 3.4. Let N ∈ B(H) be normal. Let n ∈ N. Then N has always an nth root which is also normal and given by n ei( A+2kπI for some self-adjoint A ∈ B(H) and where k ∈ Z. Proof. Since N is normal, it may be expressed as (see e.g. [11]) n = N N ) n 1 1 where U is unitary and P = √N ∗N = N. Since U is unitary, U = eiA for some self-adjoint A ∈ B(H) (cf. Proposition 18.20 in [5]). Hence N = U P = P U, N = eiAP = P eiA. Set for k ∈ Z M = P 1 n ei( A+2kπI n ) 1 n is the unique positive nth root of P . Then M is normal (cf. [6]) because where P it is a product of two commuting normal operators (in fact, a product of a positive operator and a unitary one). By the commutativity of the factors and simple results, we obtain M n = (cid:16)P as needed. 1 n ei( A+2kπI n )(cid:17)n = P ei(A+2kπI) = P eiA = N, (cid:3) Remark. As is the case of many results on normal operators, the previous two theorems too are inspired by results about complex numbers. This corroborates and strengthens the ressemblance which is already known to readers. ROOTS OF OPERATORS 7 References [1] R. Bhatia, P. Rosenthal, How and why to solve the operator equation AX − XB = Y . Bull. London Math. Soc., 29/1 (1997) 1-21. [2] J. B. Conway, B. B. Morrel, Roots and logarithms of bounded operators on Hilbert space, J. Funct. Anal., 70/1 (1987) 171-193. [3] M. R. Embry, Similarities Involving Normal Operators on Hilbert Space, Pacific J. Math., 35 (1970) 331-336. [4] S. Kurepa, On n-th roots of normal operators, Math. Z., 78 (1962) 285-292. [5] R. Meise, D. Vogt, Introduction to Functional Analysis, Oxford G.T.M. 2, Oxford University Press 1997. [6] M. H. Mortad, An Operator Theory Problem Book, World Scientific Publishing Co., (to appear in 2018). [7] C. R. Putnam, On Square Roots of Normal Operators, Proc. Amer. Math. Soc., 8 (1957) 768-769. [8] C. R. Putnam, On Square Roots and Logarithms of Self-adjoint Operators, Proc. Glasgow Math. Assoc., 4 (1958) 1-2. [9] H. Radjavi, P. Rosenthal, On Roots of Normal Operators, J. Math. Anal. Appl., 34 (1971) 653-664. [10] M. Rosenblum, On the Operator Equation BX −XA = Q, Duke Math. J., 23 (1956) 263-269. [11] W. Rudin, Functional Analysis, McGraw-Hill Book Co., Second edition, International Series in Pure and Applied Mathematics, McGraw-Hill, Inc., New York, 1991. [12] J. G. Stampfli, Roots of Scalar Operators, Proc. Amer. Math. Soc., 13/5 (1962) 796-798. Department of Mathematics, University of Oran 1, Ahmed Ben Bella, B.P. 1524, El Menouar, Oran 31000, Algeria. Mailing address: Pr Mohammed Hichem Mortad BP 7085 Seddikia Oran 31013 Algeria E-mail address: [email protected], [email protected].
1307.6208	1	1307	2013-07-19T17:06:12	Difference Sequence Spaces Derived by Using Generalized Means	[ "math.FA" ]	This paper deals with new sequence spaces $X(r, s, t ;\Delta) $ for $X\in \{l_\infty, c, c_0\}$ defined by using generalized means and difference operator. It is shown that these spaces are complete normed linear spaces and the spaces $X(r, s, t ;\Delta)$ for $X\in \{c, c_0\}$ have Schauder basis. Furthermore, the $\alpha$-, $\beta$-, $\gamma$- duals of these sequence spaces are computed and also established necessary and sufficient conditions for matrix transformations from $X(r, s, t ;\Delta)$ to $X$.	math.FA	math	Difference Sequence Spaces Derived by Using Generalized Means Atanu Manna∗, Amit Maji†, P. D. Srivastava ‡ Department of Mathematics, Indian Institute of Technology, Kharagpur Kharagpur 721 302, West Bengal, India Abstract This paper deals with new sequence spaces X(r, s, t; ∆) for X ∈ {l∞, c, c0} defined by using generalized means and difference operator. It is shown that these spaces are complete normed linear spaces and the spaces X(r, s, t; ∆) for X ∈ {c, c0} have Schauder basis. Furthermore, the α-, β-, γ- duals of these sequence spaces are computed and also established necessary and sufficient conditions for matrix trans- formations from X(r, s, t; ∆) to X. 2010 Mathematics Subject Classification: 46A45, 46A35. Keywords: Sequence spaces; Difference operator; Generalized means; α-, β-, γ- duals; Matrix transfor- mations. 1 Introduction The study of sequence spaces has importance in the several branches of analysis, namely, the structural theory of topological vector spaces, summability theory, Schauder basis theory etc. Besides this, the theory of sequence spaces is a powerful tool for obtaining some topological and geometrical results using Schauder basis. Let w be the space of all real or complex sequences x = (xn), n ∈ N0. For an infinite matrix A and a sequence space λ, the matrix domain of A, which is denoted by λA and defined as λA = {x ∈ w : Ax ∈ λ} [17]. Basic methods, which are used to determine the topologies, matrix transformations and inclusion relations on sequence spaces can also be applied to study the matrix domain λA. In recent times, there is an approach of forming new sequence spaces by using matrix domain of a suitable matrix and characterize the matrix mappings between these sequence spaces. Kizmaz first introduced and studied the difference sequence spaces in [8]. Later on, several authors including Ahmad and Mursaleen [1], C¸ olak and Et [5], Ba¸sar and Altay [3], Orhan [13], Polat and Altay [15], Aydin and Ba¸sar [4] etc. have introduced and studied new sequence spaces defined by using difference operator. ∗Corresponding author e-mail: [email protected] †Author's e-mail: [email protected] ‡Author's e-mail: [email protected] 1 On the other hand, sequence spaces are also defined by using generalized weighted means. Some of them can be viewed in Malkowsky and Sava¸s [11], Altay and Ba¸sar [2]. Mursaleen and Noman [12] introduced a sequence space of generalized means, which includes most of the earlier known sequence spaces. But till 2011, there was no such literature available in which a sequence space is generated by combining both the weighted means and the difference operator. This was firstly initiated by Polat et al. [14]. The authors in [14] have introduced the sequence spaces λ(u, v; ∆) for λ ∈ {l∞, c, c0} defined as where u, v ∈ w such that un, vn 6= 0 for all n and the matrices G(u, v) = (gnk), ∆ = (δnk) are defined by λ(u, v; ∆) =(cid:8)x ∈ w : (G(u, v).∆)x ∈ λ(cid:9), gnk =( unvk 0 if 0 ≤ k ≤ n, if k > n respectively. 0 (−1)n−k 0 if 0 ≤ k < n − 1 if n − 1 ≤ k ≤ n, if k > n, δnk =  The aim of this article is to introduce new sequence spaces defined by using both the generalized means and the difference operator. We investigate some topological properties as well as α-, β-, γ- duals and bases of the new sequence spaces are obtained. Further, we characterize some matrix transformations between these new sequence spaces. 2 Preliminaries Let l∞, c and c0 be the spaces of all bounded, convergent and null sequences x = (xn) respectively, with norm kxk∞ = sup xn. Let bs and cs be the sequence spaces of all bounded and convergent series n respectively. We denote by e = (1, 1, · · · ) and en for the sequence whose n-th term is 1 and others are zero and N0 = N ∪ {0}, where N is the set of all natural numbers. A sequence (bn) in a normed linear space (X, k.k) is called a Schauder basis for X if for every x ∈ X there is a unique sequence of scalars (µn) such that x − (cid:13)(cid:13)(cid:13) k Xn=0 µnbn(cid:13)(cid:13)(cid:13) → 0 as k → ∞, µnbn. ([17], [9]). i.e., x = ∞ Xn=0 For any subsets U and V of w, the multiplier space M (U, V ) of U and V is defined as M (U, V ) = {a = (an) ∈ w : au = (anun) ∈ V for all u ∈ U }. In particular, U α = M (U, l1), U β = M (U, cs) and U γ = M (U, bs) are called the α-, β- and γ- duals of U respectively [10]. Let A = (ank)n,k be an infinite matrix with real or complex entries ank. We write An as the sequence of the n-th row of A, i.e., An = (ank)k for every n. For x = (xn) ∈ w, the A-transform of x is defined as the sequence Ax = ((Ax)n), where 2 An(x) = (Ax)n = ankxk, ∞ Xk=0 provided the series on the right side converges for each n. For any two sequence spaces U and V , we denote by (U, V ), the class of all infinite matrices A that map U into V . Therefore A ∈ (U, V ) if and only if Ax = ((Ax)n) ∈ V for all x ∈ U . In other words, A ∈ (U, V ) if and only if An ∈ U β for all n [17]. An infinite matrix T = (tnk)n,k is said to be triangle if tnk = 0 for k > n and tnn 6= 0, n ∈ N0. 3 Sequence spaces X(r, s, t; ∆) for X ∈ {l∞, c, c0} In this section, we first begin with the notion of generalized means given by Mursaleen et al. [12]. We denote the sets U and U0 as U =nu = (un)∞ n=0 ∈ w : un 6= 0 for all no and U0 =nu = (un)∞ n=0 ∈ w : u0 6= 0o. Let r, t ∈ U and s ∈ U0. The sequence y = (yn) of generalized means of a sequence x = (xn) is defined by n yn = 1 rn sn−ktkxk (n ∈ N0). Xk=0 The infinite matrix A(r, s, t) of generalized means is defined by (A(r, s, t))nk =( sn−ktk rn 0 0 ≤ k ≤ n, k > n. Since A(r, s, t) is a triangle, it has a unique inverse and the inverse is also a triangle [7]. Take 0 = 1 s0 and D(s) D(s) n = 1 sn+1 0 s1 s2 ... sn−1 sn (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) s0 s1 ... sn−2 sn−1 0 s0 ... sn−3 sn−2 0 0 0 · · · 0 · · · ... sn−4 · · · sn−3 · · · s0 s1 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) for n = 1, 2, 3, · · · Then the inverse of A(r, s, t) is the triangle B = (bnk)n,k which is defined as bnk =( (−1)n−k D(s) 0 n−k tn rk 0 ≤ k ≤ n, k > n. We now introduce the sequence spaces X(r, s, t; ∆) for X ∈ {l∞, c, c0} as X(r, s, t; ∆) =nx = (xn) ∈ w :(cid:16) 1 rn n Xk=0 sn−ktk∆xk(cid:17)n ∈ Xo, which is a combination of the generalized means and the difference operator ∆ such that ∆xk = xk−xk−1, x−1 = 0. By using matrix domain, we can write X(r, s, t; ∆) = XA(r,s,t;∆) = {x ∈ w : A(r, s, t; ∆)x ∈ X}, where A(r, s, t; ∆) = A(r, s, t).∆, product of two triangles A(r, s, t) and ∆. 3 These sequence spaces include many known sequence spaces studied by several authors. For examples, I. if rn = 1 un , tn = vn and sn = 1 ∀ n, then the sequence spaces X(r, s, t; ∆) reduce to X(u, v; ∆) for X ∈ {l∞, c, c0} introduced and studied by Polat et al. [14] II. if tn = 1, sn = 1 ∀ n and rn = n + 1, then the sequence space X(r, s, t; ∆) for X = l∞ reduces to ces∞(∆) studied by Orhan [13]. III. if rn = 1 n! , tn = αn X ∈ {l∞, c, c0} reduce to eα n! , sn = (1−α)n ∞(∆), eα n! , where 0 < α < 1, then the sequence spaces X(r, s, t; ∆) for c (∆) and eα 0 (∆) respectively [15]. IV. if rn = n + 1, tn = 1 + αn, where 0 < α < 1 and sn = 1 ∀n, then the sequence spaces X(r, s, t; ∆) 0 (∆) studied by Aydin and Ba¸sar for X ∈ {c, c0} reduce to the spaces of sequences aα [4]. For X = l∞, the sequence space X(r, s, t; ∆) reduces to aα ∞(∆) studied by Djolovi´c [6]. c (∆) and aα 4 Main results In this section, we begin with some topological results of the newly defined sequence spaces. Theorem 4.1. The sequence spaces X(r, s, t; ∆) for X ∈ {l∞, c, c0} are complete normed linear spaces under the norm defined by = sup (A(r, s, t; ∆)x)n. n Proof. Let u, v ∈ X(r, s, t; ∆) and α, β be any two scalars. Then 1 rn kxkX(r,s,t;∆) = sup n (cid:12)(cid:12)(cid:12) sn−ktk∆(αuk + βvk)(cid:12)(cid:12)(cid:12)(cid:12) n sn−ktk∆xk(cid:12)(cid:12)(cid:12) Xk=0 n (cid:12)(cid:12)(cid:12)(cid:12) Xk=0 1 rn n 1 rn n Xk=0 sup n (cid:12)(cid:12)(cid:12)(cid:12) ≤ α sup + β sup sn−ktk∆uk(cid:12)(cid:12)(cid:12)(cid:12) 1 rn n Xk=0 n (cid:12)(cid:12)(cid:12)(cid:12) sn−ktk∆vk(cid:12)(cid:12)(cid:12)(cid:12) and hence αu + βv ∈ X(r, s, t; ∆). Therefore X(r, s, t; ∆) is a linear space. It is easy to show that the functional k.kX(r,s,t;∆) defined above gives a norm on the linear space X(r, s, t; ∆). To show completeness, let (xm) be a Cauchy sequence in X(r, s, t; ∆), where xm = (xm ∈ X(r, s, t; ∆), for each m ∈ N0. Then for every ǫ > 0 there exists n0 ∈ N such that k ) = (xm 0 , xm 1 , xm 2 , . . .) The above implies that for each k ∈ N0, kxm − xlkX(r,s,t;∆) < ǫ for m, l ≥ n0. A(r, s, t; ∆)(xm k − xl k) < ǫ for all m, l ≥ n0. (4.1) Therefore ((A(r, s, t).∆)xm converges for each k. We write k ) is a Cauchy sequence of scalars for each k ∈ N0 and hence ((A(r, s, t).∆)xm k ) lim m→∞ ((A(r, s, t).∆)x(m) k ) = ((A(r, s, t).∆)xk), k ∈ N0. Letting l → ∞ in (4.1), we obtain A(r, s, t; ∆)(xm < ǫ for all m ≥ n0 and each k ∈ N0. (4.2) Hence by definition kxm − xkX(r,s,t;∆) < ǫ for all m ≥ n0. Next we show that x ∈ X(r, s, t; ∆). Consider (cid:12)(cid:12)(cid:12) k − xk)(cid:12)(cid:12)(cid:12) 4 kxkX(r,s,t;∆) ≤ kxmkX(r,s,t;∆) + kxm − xkX(r,s,t;∆), which is finite for m ≥ n0 and hence x ∈ X(r, s, t; ∆). This completes the proof. Theorem 4.2. The sequence spaces X(r, s, t; ∆), where X ∈ {l∞, c, c0} are linearly isomorphic to the spaces X ∈ {l∞, c, c0} respectively i.e. l∞(r, s, t; ∆) ∼= l∞, c(r, s, t; ∆) ∼= c and c0(r, s, t; ∆) ∼= c0. Proof. We prove the theorem only for the case X = l∞. To prove this, we need to show that there exists a bijective linear map from l∞(r, s, t; ∆) to l∞. We define a map T : l∞(r, s, t; ∆) → l∞ by x 7−→ T x = y = (yn), where yn = 1 rn n Xk=0 sn−ktk∆xk. Since ∆ is a linear operator, so the linearity of T is trivial. It is clear from the definition that T x = 0 implies x = 0. Thus T is injective. To prove T is surjective, let y = (yn) ∈ l∞. Since y = (A(r, s, t).∆)x, i.e., we can get a sequence x = (xn) as x =(cid:0)A(r, s, t).∆(cid:1)−1 y = ∆−1.A(r, s, t)−1y, xn = Then (−1)k D(s) tk+j k n−j Xk=0 rjyj, n ∈ N0. (4.3) kxkl∞(r,s,t;∆) = sup = sup yn = kyk∞ < ∞. n n Xj=0 n (cid:12)(cid:12)(cid:12)(cid:12) 1 rn n Xk=0 sn−ktk∆xk(cid:12)(cid:12)(cid:12)(cid:12) Thus x ∈ l∞(r, s, t; ∆) and this shows that T is surjective. Hence T is a linear bijection from l∞(r, s, t; ∆) to l∞. Also T is norm preserving. So l∞(r, s, t; ∆) ∼= l∞. In the same way, we can prove that c0(r, s, t; ∆) ∼= c0, c(r, s, t; ∆) ∼= c. This completes the proof. Since X(r, s, t; ∆) ∼= X for X ∈ {c0, c}, the Schauder bases of the sequence spaces X(r, s, t; ∆) are the inverse image of the bases of X for X ∈ {c0, c}. So, we have the following theorem without proof. Theorem 4.3. Let µk = (A(r, s, t; ∆)x)k for all k ∈ N0. Define the sequences b(j) = (b(j) b(−1) n as n ), j ∈ N0 and (−1)k D(s) tk+j k n−j Xk=0 0 rj if 0 ≤ j ≤ n if j > n. and b(−1) n = n Xj=0 n−j Xk=0 (−1)k D(s) tk+j k rj. b(j) n =  Then the followings are true: (i) The sequence (b(j))∞ unique representation of the form j=0 is a basis for the space c0(r, s, t; ∆) and any sequence x ∈ c0(r, s, t; ∆) has a x = ∞ Xj=0 µjb(j). 5 (ii) The sequence (b(j))∞ representation of the form j=−1 is a basis for the space c(r, s, t; ∆) and any x ∈ c(r, s, t; ∆) has a unique x = ℓb(−1) n + (µj − ℓ)b(j), ∞ Xj=0 where ℓ = lim n→∞ (A(r, s, t; ∆)x)n. Remark 4.1. In particular, if we choose rn = 1 un X(r, s, t; ∆) reduce to X(u, v; ∆) for X ∈ {l∞, c, c0} [14]. With this choice of (sn), we have D(s) D(s) , tn = vn, sn = 1, ∀ n then the sequence spaces 0 = n = 0 for n ≥ 2. Thus the sequences b(j) = (b(j) n ), j ∈ N0 and b(−1) 1 = 1 and D(s) reduce to n b(j) n =  The sequences (b(j))∞ [14]. 1 uj(cid:16) 1 1 vj unvn 0 − 1 vj+1(cid:17) if if if 0 ≤ j < n j = n j > n. and b(−1) n = n−1 Xj=0 1 uj(cid:16) 1 vj − 1 vj+1(cid:17) + 1 unvn . j=0 and (b(j))∞ j=−1 are the bases for the spaces c0(u, v; ∆) and c(u, v; ∆) respectively Let F be the collection of all nonempty finite subsets of the set of all natural numbers and A = (ank)n,k be an infinite matrix satisfying the conditions: ∞ < ∞ ank < ∞ sup K∈F sup n ank(cid:12)(cid:12)(cid:12) ∞ Xn=0(cid:12)(cid:12)(cid:12) Xk∈K Xk=0 lim n lim n lim n lim n lim n lim n ank = 0 ∞ Xk=0 ank = 0 for all k ank = 0 ∞ Xk=0 ank exists for all k ank − lim n ank = 0 ank exists ∞ ∞ Xk=0 Xk=0 (4.4) (4.5) (4.6) (4.7) (4.8) (4.9) (4.10) (4.11) We now state some results of Stieglitz and Tietz [16] which are required to obtain the duals and to characterize some matrix transformations. Theorem 4.4. [16] (a) A ∈ (c0, l1), A ∈ (c, l1), A ∈ (l∞, l1) if and only if (4.4) holds. (b) A ∈ (c0, l∞), A ∈ (c, l∞), A ∈ (l∞, l∞) if and only if (4.5) holds. (c) A ∈ (c0, c0) if and only if (4.5) and (4.7) hold. 6 (d) A ∈ (l∞, c0) if and only if (4.6) holds. (e) A ∈ (c, c0) if and only if (4.5), (4.7) and (4.8) hold. (f ) A ∈ (c0, c) if and only if (4.5) and (4.9) hold. (g) A ∈ (l∞, c) if and only if (4.5), (4.9) and (4.10) hold. (h) A ∈ (c, c) if and only if (4.5), (4.9) and (4.11) hold. 4.1 The α-dual, γ-dual of X(r, s, t; ∆) for X ∈ {l∞, c, c0} Theorem 4.5. The α-dual of the space X(r, s, t; ∆) for X ∈ {l∞, c, c0} is the set Λ =na = (an) ∈ w : sup K∈FXn (cid:12)(cid:12)(cid:12)Xj∈K (−1)k D(s) tk+j k n−j Xk=0 < ∞o. rj an(cid:12)(cid:12)(cid:12) Proof. Let a = (an) ∈ w, x ∈ X(r, s, t; ∆) and y ∈ X for X ∈ {l∞, c, c0}. Then for each n ∈ N0, we have anxn = n Xj=0 n−j Xk=0 (−1)k D(s) tk+j k rjanyj = (Cy)n, where the matrix C = (cnj) is defined as (−1)k D(s) tk+j k n−j Xk=0 0 rjan if 0 ≤ j ≤ n if j > n cnj =  and xn is given by (4.3). Thus for each x ∈ X(r, s, t; ∆), (anxn)n ∈ l1 if and only if (Cy)n ∈ l1 where y ∈ X for X ∈ {l∞, c, c0}. Therefore a = (an) ∈ [X(r, s, t; ∆)]α if and only if C ∈ (X, l1). By using Theorem 4.4(a) , we have [X(r, s, t; ∆)]α = Λ. Theorem 4.6. The γ-dual of X(r, s, t; ∆) for X ∈ {l∞, c, c0} is the set Γ =na = (an) ∈ w : sup m ∞ Xn=0 emn < ∞o, where the matrix E = (emn) is defined by E = (emn) =  Note: We mean +(cid:16) D(s) 0 tn − D(s) 1 tn+1(cid:17) m Xj=n+1 aj + m Xj=n+2 (−1)j−n D(s) j−n tj (cid:16) m Xk=j ak(cid:17)(cid:21) 0 ≤ n ≤ m, n > m. rn(cid:20) an s0tn 0 m = 0 if n > m. Pj=n Proof. Let a = (an) ∈ w, x ∈ X(r, s, t; ∆) and y ∈ X, where X ∈ {l∞, c, c0}. Then by using (4.3), we 7 have m Xn=0 anxn = k rjyjan tk+j n n m n n−j (−1)k D(s) (−1)k D(s) = n−j n−1 m−1 Xk=0 Xj=0 Xn=0 Xk=0 Xj=0 Xn=0 =(cid:20) m−1 (−1)k D(s) Xk=0 Xn=0 k tk +(cid:20) m−1 Xn=0 Xk=0 a0 +(cid:16) D(s) =(cid:20) D(s) +(cid:20) D(s) a1 +(cid:16) D(s) +(cid:16) D(s) rn(cid:20) an Xn=0 s0tn 0 tn 0 t1 0 t1 0 t0 0 t0 − = (−1)k D(s) tk+1 − m k (−1)k D(s) tk+1 k k rj yjan tk+j m an + m m−j + Xk=0 Xj=0 (−1)k D(s) k tk D(s) 1 m m−1 aj + an + Xk=0 Xk=0 t1 (cid:17) Xj=1 t2 (cid:17) Xj=2 tn+1(cid:17) Xj=n+1 D(s) 1 D(s) 1 aj + m m − m Xj=2 Xj=3 m (−1)k D(s) k rj yjam tk+j am(cid:21)r0y0 am(cid:21)r1y1 + · · · + D(s) m (−1)j D(s) 0 tm rmymam j tj (cid:16) m D(s) j−1 ak(cid:17)(cid:21)r0y0 Xk=j ak(cid:17)(cid:21)r1y1 + · · · + tj (cid:16) Xk=j (−1)j−n D(s) ak(cid:17)(cid:21)yn tj (cid:16) Xk=j j−n m (−1)j−1 aj + m Xj=n+2 rmam tm D(s) 0 ym m Xn=0 anxn(cid:17)m ∈ = (Ey)m, where E = (emn) is the matrix defined above. Thus a ∈(cid:2)X(r, s, t; ∆)(cid:3)γ if and only if ax = (anxn) ∈ bs for x ∈ X(r, s, t; ∆) if and only if(cid:16) l∞, i.e. (Ey)m ∈ l∞, for y ∈ X. Hence by Theorem 4.4(b), we have (cid:2)X(r, s, t; ∆)(cid:3)γ = Γ. 4.2 β-dual and Matrix transformations We now first discuss about the β-dual and then characterize some matrix transformations. Let T be a triangle and XT be the matrix domain of T . Theorem 4.7. ([7], Theorem 2.6) Let X be a BK space with AK property and R = St, the transpose of S, where S = (sjk) is the inverse of the matrix T . Then a ∈ (XT )β if and only if a ∈ (X β)R and W ∈ (X, c0), where the triangle W is defined by wmk = ajsjk. Moreover if a ∈ (XT )β, then ∞ Pj=m akzk = ∞ Xk=0 ∞ Xk=0 Rk(a)Tk(z) ∀ z ∈ XT . Remark 4.2. ([7], Remark 2.7) The conclusion of the Theorem 4.7 is also true for X = l∞. Remark 4.3. ([10], [7]) We have a ∈ (cT )β if and only if Ra ∈ l1 and W ∈ (c, c). Moreover, if 8 a ∈ (cT )β then we have for all z ∈ cT ∞ akzk = Rk(a)Tk(z) − ηγ, ∞ Xk=0 Xk=0 Xk=0 m where η = lim k→∞ Tk(z) and γ = lim m→∞ wmk. To find β-duals of the sequence spaces X(r, s, t; ∆) for X ∈ {l∞, c, c0}, we list the following sets: ∞ ∞ m m lim m→∞ lim m→∞ Xk=0 Rk(a) < ∞o B1 =na ∈ w : wmk = 0 for all ko B2 =na ∈ w : B3 =na ∈ w : sup wmk < ∞o Xk=0 wmk = 0o B4 =na ∈ w : Xk=0 exists for all ko B5 =na ∈ w : wmk exists o, B6 =na ∈ w : Xk=0 (−1)l D(s) − D(s) tk+1(cid:17) Pl=2 Pj=k+1 Pj=k+l aj(cid:21). (−1)l D(s) Pl=m−k+1 Pj=k+l lim m→∞ lim m→∞ wmk tl+k m ∞ ∞ ∞ l l l 1 ∞ ∞ ∞ 0 tk s0tk aj + +(cid:16) D(s) Pj=m where Rk(a) = rk(cid:20) ak wmk = rk(cid:20) m−k (−1)l D(s) Pl=0 Theorem 4.8. We have [c0(r, s, t; ∆)]β = B1T B2T B3, [l∞(r, s, t; ∆)]β = B1T B4 and [c(r, s, t; ∆)]β = B1T B3T B5T B6. aj(cid:21), R(a) = (Rk(a))k and Proof. Here the matrix T = A(r, s, t).∆ = (t′ aj + tl+k tl+k nk), where So, T −1 = (A(r, s, t).∆)−1 = ∆−1.A(r, s, t)−1. Let S = (sjk) be the inverse of T . Then we easily get t′ 1 rn s0tn rn 0 nk =  sjk =  [sn−ktk − sn−k+1tk+1] if 0 ≤ k < n if k = n if k > n. (−1)l D(s) l tl+k j−k Pl=0 0 rk if if 0 ≤ k ≤ j k > j. 9 To compute β-duals, we first determine W = (wmk) and R(a) = (Rk(a)), where R = St. ∞ ajsjk rkak + Rk(a) = = = Xj=k D(s) 0 tk D(s) 0 tk (−1)l D(s) tl+k l rkaj ∞ j−k 1 Xl=0 Xj=k+1 (−1)l D(s) Xl=0 tl+k +(cid:16) D(s) tk+1(cid:17) D(s) 1 0 tk − l rkak + rkak+1 + rkak+2 + · · · 2 l (−1)l D(s) Xl=0 tl+k (−1)l D(s) Xl=2 tl+k ∞ l ∞ Xj=k+1 aj + ∞ Xj=k+l aj(cid:21) and = rk(cid:20) ak s0tk ∞ wmk = ajsjk ∞ = j−k Xj=m Xl=0 Xj=m = rk(cid:20) m−k Xl=0 l (−1)l D(s) tl+k (−1)l D(s) tl+k l rkaj ∞ Xj=m aj + ∞ Xl=m−k+1 (−1)l D(s) tl+k l ∞ Xj=k+l aj(cid:21). Using Theorem 4.7 and Remark 4.2 & 4.3, we have [c0(r, s, t; ∆)]β = B1T B2T B3, [l∞(r, s, t; ∆)]β = B1T B4 and [c(r, s, t; ∆)]β = B1T B3T B5T B6. Theorem 4.9. ([7], Theorem 2.13) Let X be a BK space with AK property, Y be an arbitrary subset of w and R = St. Then A ∈ (XT , Y ) if and only if BA ∈ (X, Y ) and W An ∈ (X, c0) for all n = 0, 1, 2, · · · , where BA is the matrix with rows BA n = R(An), An are the rows of A and the triangles W An are defined by ∞ anjsjk Pj=m 0 : : 0 ≤ k ≤ m k > m. wAn mk =  Theorem 4.10. ([7]) Let Y be any linear subspace of w. Then A ∈ (cT , Y ) if and only if Rk(An) ∈ m (c0, Y ) and W An ∈ (c, c) for all n and Rk(An)e − (γn) ∈ Y , where γn = lim m→∞ and e = (1, 1, 1, · · · ). Moreover, if A ∈ (cT , Y ) then we have wAn mk for n = 0, 1, 2 · · · Xk=0 Az = Rk(An)(T (z)) − η(γn) for all z ∈ cT , where η = lim k→∞ Tk(z). To characterize the matrix transformations A ∈ (X(r, s, t; ∆), Y ) for X, Y ∈ {l∞, c, c0}, we consider the following conditions: 10 ∞ Xk=0 sup n lim n→∞ Rk(An) < ∞ Rk(An) = 0 for all k wAn mk < ∞ for all n wAn mk = 0 for all n, k Rk(An) exists for all k ∞ Rk(An) = 0 wAn mk = 0 for all n sup m m Xk=0 lim m→∞ lim n→∞ lim n→∞ lim m→∞ ∞ m Xk=0 Xk=0 Xk=0(cid:12)(cid:12)(cid:12) Xk=0 wAn = 0 Rk(An)(cid:12)(cid:12)(cid:12) lim n→∞ Rk(An) − lim n→∞ lim m→∞ lim m→∞ mk exists for all k, n m wAn mk exists for all n Rk(An)e − (γn) ∈ c0 for all γn, n = 0, 1, 2, · · · Rk(An)e − (γn) ∈ l∞ for all γn, n = 0, 1, 2, · · · Rk(An)e − (γn) ∈ c for all γn, n = 0, 1, 2, · · · , m (4.12) (4.13) (4.14) (4.15) (4.16) (4.17) (4.18) (4.19) (4.20) (4.21) (4.22) (4.23) (4.24) where γn = lim m→∞ s0tk Rk(An) = rkh ank mk = rk(cid:20) m−k Pl=0 wAn 0 tk wAn mk, Xk=0 +(cid:16) D(s) Pj=m tl+k ∞ (−1)l D(s) l 1 − D(s) tk+1(cid:17) anj + anj + ∞ l (−1)l D(s) Pl=2 (−1)l D(s) tl+k ∞ l tl+k ∞ ∞ Pj=k+1 Pl=m−k+1 Pj=k+l anji and ∞ Pj=k+l anj(cid:21). Theorem 4.11. (a) A ∈ (c0(r, s, t; ∆), c0) if and only if the conditions (4.12), (4.13), (4.14) and (4.15) hold. (b) A ∈ (c0(r, s, t; ∆), c) if and only if the conditions (4.12), (4.14), (4.15) and (4.16) hold. (c) A ∈ (c0(r, s, t; ∆), l∞) if and only if the conditions (4.12), (4.14) and (4.15) hold. Proof. We prove only the part (a) of this theorem. The other parts follow in a similar way. For this, we first compute the matrices BA = Rk(An) and W An = (wAn mk) for n = 0, 1, 2, · · · of Theorem 4.9 to determine the conditions BA ∈ (c0, c0) and W An ∈ (c0, c0). Using the same lines of proof as used in 11 Theorem 4.8, we have ∞ Rk(An) = ∞ = j−k sjkanj Xj=k (−1)l D(s) Xj=k+1 Xl=0 tl+k +(cid:16) D(s) = rkh ank s0tk 0 tk − l rkanj + D(s) 0 tk rkank D(s) 1 tk+1(cid:17) ∞ Xj=k+1 anj + (−1)l D(s) tl+k l ∞ Xl=2 ∞ Xj=k+l anji and ∞ wAn mk = anjsjk Xj=m = rk(cid:20) m−k Xl=0 (−1)l D(s) tl+k l ∞ Xj=m anj + ∞ Xl=m−k+1 (−1)l D(s) tl+k l ∞ Xj=k+l anj(cid:21). Using Theorem 4.9, we have A ∈ (c0(r, s, t; ∆), c0) if and only if the conditions (4.12), (4.13), (4.14) and (4.15) hold. We can also obtain the following results. Corollary 4.1. (a) A ∈ (l∞(r, s, t; ∆), c0) if and only if the conditions (4.17) and (4.18) hold. (b) A ∈ (l∞(r, s, t; ∆), c) if and only if the conditions (4.12), (4.16), (4.18) and (4.19) hold. (c) A ∈ (l∞(r, s, t; ∆), l∞) if and only if the conditions (4.12) and (4.18) hold. Corollary 4.2. (a) A ∈ (c(r, s, t; ∆), c0) if and only if the conditions (4.12), (4.13), (4.14), (4.20), (4.21) and (4.22) hold. (b) A ∈ (c(r, s, t; ∆), c) if and only if the conditions (4.12), (4.14), (4.16), (4.20), (4.21) and (4.24) hold. (c) A ∈ (c(r, s, t; ∆), l∞) if and only if the conditions (4.12), (4.14), (4.20), (4.21) and (4.23) hold. 5 Some applications In this section, we justify our results in some special cases. Also, we illustrate the results related to matrix transformations given by Djolovic[6], Polat and Altay [15], Aydin and Ba¸sar [4] etc. (I) In particular, if we choose rn = 1 spaces X(r, s, t; ∆) for X ∈ {l∞, c0, c} reduce to the Euler difference sequence spaces eα 0 = 1, D(s) and eα D(s) difference sequence spaces is the set , where 0 < α < 1 then the sequence ∞(∆), eα 0 (∆) 1 = (1 − α), , k ∈ N0. Thus the α-dual of the Euler c (∆) respectively [15]. By the above choice of r, s and t, we have D(s) and so on. Therefore D(s) n! and sn = (1−α)n 2 = (1−α)2 3 = (1−α)3 k = (1−α)k n! , tn = αn , D(s) n! k! 3! 2! n K∈FXn (cid:12)(cid:12)(cid:12)(cid:12)Xj∈K na ∈ w : sup Xk=j K∈FXn (cid:12)(cid:12)(cid:12)(cid:12)Xj∈K =na ∈ w : sup (−1)k−j (1 − α)k−j (k − j)! n Xk=j (−1)k−j(cid:18)k . 1 j! α−kk!an(cid:12)(cid:12)(cid:12)(cid:12) < ∞o j(cid:19)(1 − α)k−j α−kan(cid:12)(cid:12)(cid:12)(cid:12) < ∞o. 12 Here we illustrate that how the characterization of matrix transformation A ∈ (eα tained with the help of Corollary 4.1(c) ∞(∆), l∞) can be ob- ∞ Rk(An) = anjsjk Xj=k +(cid:16) D(s) = rk(cid:20) ank +(cid:16) 1 αk − ank αk s0tk 0 tk = − 1 αk+1(cid:17) ∞ ∞ Xj=k+1 anj + D(s) 1 tk+1(cid:17) Xj=k+1 ∞ anj + l ∞ (−1)l D(s) tk+l Xl=2 (−1)l(cid:18)l + k Xj=k+l l (cid:19) (1 − α)l αk+l anj(cid:21) Xj=k+l ∞ ∞ Xl=2 anj. ∞ wAn mk = l anjsjk (−1)l D(s) tk+l Xj=m = rk(cid:20) m−k Xl=0 k!(cid:20) m−k Xl=0 =(cid:20) m−k (−1)l(cid:18)l + k Xl=0 = 1 (−1)l (1 − α)l(l + k)! l!αl+k anj + l (cid:19) (1 − α)l αl+k anj + (−1)l(cid:18)l + k l (cid:19)) (1 − α)l αl+k (−1)l D(s) tk+l l ∞ anj(cid:21) Xj=k+l (−1)l (1 − α)l(l + k)! l!αl+k anj ∞ Xj=m ∞ ∞ Xl=m−k+1 Xj=m Xj=m ∞ ∞ Xm−k+1 Xm−k+1 ∞ ∞ Xj=k+l Xj=k+l ∞ anj(cid:21) anj(cid:21). So A ∈ (eα ∞(∆), l∞) if and only if and sup n ∞ Xk=0 Rk(An) < ∞ lim m→∞ m Xk=0 wAn mk = 0 for all n. (II) We choose sn = 1 ∀ n, rn = (n + 1), tn = 1 + αn, where 0 < α < 1 then the sequence 0 (∆) respectively. With this choice mk) become spaces X(r, s, t; ∆) for X ∈ {l∞, c, c0} reduce to aα D(s) k = 0 for all k ≥ 2. Therefore the matrices Rk(An) and W An = (wAn 0 = 1 = D(s) c (∆) and aα and D(s) ∞(∆), aα 1 Rk(An) = (k + 1)(cid:20) ank 1 + αk +(cid:16) 1 1 + αk − 1 1 + αk+1(cid:17) ∞ Xj=k+1 anj(cid:21), and wAn 0 tk mk = rk(cid:20)(cid:16) D(s) = (k + 1)(cid:20)(cid:16) D(s) 1 − tk+1(cid:17) 1 + αk − 1 ∞ anj − Xj=m 1 + αk+1(cid:17) 1 D(s) 1 tm+1 ∞ Xj=m ∞ Xj=m+1 anj(cid:21) anj − 1 1 + αm+1 ∞ Xj=m+1 anj(cid:21). 13 By evaluating, we have wAn mk = m Xk=0 m (k + 1)(cid:16) Xk=0 1 1 + αk − Therefore by Corollary 4.1(a), we have A ∈ (aα ∞(∆), c0) if and only if 1 1 + αk+1(cid:17)(cid:12)(cid:12)(cid:12) ∞ Xj=m + anm m + 1 1 + αm+1 . anj(cid:12)(cid:12)(cid:12) lim n→∞ ∞ Xk=0 Rk(An) = 0 lim m→∞ m Xk=0 wAn mk = 0 for all n. and References [1] Z. U. Ahmad, M. Mursaleen, Kothe-Toeplitz duals of some new sequence spaces and their matrix maps, Publ. Inst. Math.(Beograd), 42(56)(1987), 57-61. [2] B. Altay, F. Ba¸sar, Generalization of the sequence space ℓ(p) derived by weighted mean, J. Math. Anal. Appl., 330(2007), 174-185. [3] B. Altay , F. Ba¸sar, The fine spectrum and matrix domain of the difference operator ∆ on the sequece space ℓp, (0 < p < 1), Commun. Math. Anal., 2 (2)(2007), 1-11. [4] C. Aydin, F. Ba¸sar, Some new difference sequence spaces, Appl. Math. Comput., 157(3)(2004), 677-693. [5] R. C¸ olak, M. Et, On some generalized difference sequence spaces and related matrix transformations, Hokkaido Math. J., 26(3)(1997), 483-492. [6] I. Djolovi´c , On the space of bounded Euler difference sequences and some classes of compact operators, Appl. Math. Comput., 182(2) (2006),1803-1811. [7] A. M. Jarrah, E. Malkowsky, Ordinary, absolute and strong summability and matrix transforma- tions, Filomat, 17(2003), 59-78. [8] H. Kizmaz, On certain sequence spaces, Canad. Math. Bull., 24(2)(1981), 169-176. [9] I. J. Maddox, Elements of Functioal Analysis, 2nd Edition, The University Press, Cambridge, 1988. [10] E. Malkowsky, V. Rakocevi´c, On matrix domains of triangles, Appl. Math. Comput., 189(2)(2007), 1146-1163. [11] E. Malkowsky, E. Sava¸s, Matrix transformations between sequence spaces of generalized weighted means, Appl. Math. Comput., 147(2004), 333-345. [12] M. Mursaleen, A. K. Noman, On generalized means and some related sequence spaces, Comput. Math. Appl., 61(4)(2011), 988-999. [13] C. Orhan, Matrix transformations on Cesro difference sequence spaces, Comm. Fac. Sci. Univ. Ankara Sr. A1 Math., 33(1)(1984), 1-8. 14 [14] H. Polat, V. Karakaya, N. S¸im¸sek, Difference sequence spaces derived by using a generalized weighted mean, Appl. Math. Lett., 24(5)(2011), 608-614. [15] H. Polat, B. Atay, On some new Euler difference sequence spaces, Southeast Asian Bulletin Math, 30(2006), 209-220. [16] M. Stieglitz, H. Tietz, Matrix trasformationnen von Folenraumen Eine Erebisubersicht, Mathema- tische Zeitschrift(Math. Z.), 154(1977), 1-16. [17] A. Wilansky, Summability through Functional Analysis, North-Holland Math. Stud., vol. 85, Else- vier Science Publishers, Amsterdam, New York, Oxford, 1984. 15
1912.06846	1	1912	2019-12-14T13:45:55	A class of sectorial relations and the associated closed forms	[ "math.FA" ]	Let $T$ be a closed linear relation from a Hilbert space ${\mathfrak H}$ to a Hilbert space ${\mathfrak K}$ and let $B \in \mathbf{B}({\mathfrak K})$ be selfadjoint. It will be shown that the relation $T^{}(I+iB)T$ is maximal sectorial via a matrix decomposition of $B$ with respect to the orthogonal decomposition ${\mathfrak H}={\rm d\overline{om}\,} T^ \oplus {\rm mul\,} T$. This leads to an explicit expression of the corresponding closed sectorial form. These results include the case where ${\rm mul\,} T$ is invariant under $B$. The more general description makes it possible to give an expression for the extremal maximal sectorial extensions of the sum of sectorial relations. In particular, one can characterize when the form sum extension is extremal.	math.FA	math	A CLASS OF SECTORIAL RELATIONS AND THE ASSOCIATED CLOSED FORMS S. HASSI AND H.S.V. DE SNOO Dedicated to V.E. Katsnelson on the occasion of his 75th birthday Abstract. Let T be a closed linear relation from a Hilbert space H to a Hilbert space K and let B ∈ B(K) be selfadjoint. It will be shown that the relation T ∗(I + iB)T is maximal sectorial via a matrix decomposition of B with respect to the orthogonal decomposition H = dom T ∗ ⊕ mul T . This leads to an explicit expression of the corresponding closed sectorial form. These results include the case where mul T is invariant under B. The more general description makes it possible to give an expression for the extremal maximal sectorial extensions of the sum of sectorial relations. In particular, one can characterize when the form sum extension is extremal. 1. Introduction A linear relation H in a Hilbert space H is said to be accretive if Re (h′, h) ≥ 0, {h, h′} ∈ H. Note that the closure of an accretive relation is also accretive. An accretive relation H in H is said to be maximal accretive if the existence of an accretive relation H ′ in H with H ⊂ H ′ implies H ′ = H. A maximal accretive relation is automatically closed. In a similar way, a linear relation H in a Hilbert space H is said to be sectorial with vertex at the origin and semi-angle α, α ∈ [0, π/2), if (1.1) Im (h′, h) ≤ (tan α) Re (h′, h), {h, h′} ∈ H. The closure of a sectorial relation is also sectorial. A sectorial relation H in a Hilbert space H is said to be maximal sectorial if the existence of a sectorial relation H ′ in H with H ⊂ H ′ implies H ′ = H. A maximal sectorial relation is automatically closed. Note that a sectorial relation is maximal sectorial if and only if it is maximal as an accretive relation; see [6]. A sesquilinear form t = t[·, ·] in a Hilbert space H is a mapping from dom t ⊂ H to C which is linear in its first entry and antilinear in its second entry. The adjoint t∗ is defined by t∗[h, k] = t[k, h], h, k ∈ dom t; for the diagonal of t the notation t[·] will be used. A (sesquilinear) form is said to be sectorial with vertex at the origin and semi-angle α, α ∈ [0, π/2), if (1.2) ti[h] ≤ (tan α) tr[h], h ∈ dom t, 2010 Mathematics Subject Classification. Primary 47B44; Secondary 47A06, 47A07, 47B65. Key words and phrases. Sectorial relation, Friedrichs extension, Kreın extension, extremal extension, form sum. The second author thanks the University of Vaasa for its hospitality during the preparation of this work. 1 2 S. HASSI AND H.S.V. DE SNOO where the real part tr and the imaginary part ti are defined by (1.3) tr = t + t∗ 2 , ti = t − t∗ 2i , dom tr = dom ti = dom t. A sesquilinear form will be called a form in the rest of this note. Observe that the form tr is nonnegative and that the form ti is symmetric, while t = tr + i ti. A sectorial form t is said to be closed if hn → h, t[hn − hm] → 0 ⇒ h ∈ dom t and t[hn − h] → 0. A sectorial form t is closed if and only if its real part tr is closed; see [7]. The connection between maximal sectorial relations and closed sectorial forms is given in the so-called first representation theorem; cf. [1], [4], [7], [8]. Theorem 1.1. Let t be a closed sectorial form in a Hilbert space H with vertex at the origin and semi-angle α, α ∈ [0, π/2). Then there exists a unique maximal sectorial relation H in H with vertex at the origin and semi-angle α in H such that (1.4) dom H ⊂ dom t, and for every {h, h′} ∈ H and k ∈ dom t one has t[h, k] = (h′, k). (1.5) Conversely, for every maximal sectorial relation H with vertex at the origin and semi-angle α, α ∈ [0, π/2), there exists a unique closed sectorial form t such that (1.4) and (1.5) are satisfied. This result contains as a special case the connection between nonnegative selfad- joint relations and closed nonnegative forms. The nonnegative selfadjoint relation Hr corresponding to the real part tr of the form t is called the real part of H; this notion should not to be confused with the real part introduced in [6]. In the theory of sectorial operators one encounters expressions T ∗(I +iB)T where T is a linear operator from a Hilbert space H to a Hilbert space K and B ∈ B(K) is a selfadjoint operator. In the context of sectorial relations the operator T may be replaced by a linear relation T . A frequently used observation is that when T is a closed linear relation and the multivalued part mul T is invariant under B, then the product is a maximal sectorial relation; cf. [4]. However, in fact, the relation (1.6) T ∗(I + iB)T is maximal sectorial for any closed linear relation T . This will be shown in this note via a matrix decomposition of B with respect to the orthogonal decomposition H = dom T ∗ ⊕mul T . In addition the closed sectorial form corresponding to T ∗(I +iB)T will be determined. The main argument consists of a reduction to the case where T is an operator. For the convenience of the reader the arguments in the operator case are included. Note that if T is not closed, then T ∗(I + iB)T is a sectorial relation which may have maximal sectorial extensions, such as T ∗(I + iB)T ∗∗ and some of these extensions have been determined in [5]; cf. [10]. It is clear that the sum of two sectorial relations is a sectorial relation and there will be maximal sectorial extensions. In [5] the Friedrichs extension has been determined in general, while the Kreın extension has been determined only under additional conditions. As an application of the above results for the relation in (1.6) the Kreın extension and, in fact, all extremal maximal sectorial extensions A CLASS OF SECTORIAL RELATIONS AND THE ASSOCIATED CLOSED FORMS 3 of the sum of two sectorial relations will be characterized in general. With this characterization one can determine when the form sum extension is extremal. 2. A preliminary result The first case to be considered is the linear relation T ∗(I + iB)T , where T a closed linear operator, which is not necessarily densely defined, and B ∈ B(K) is selfadjoint. In this case one can write down a natural closed sectorial form and verify that T ∗(I + iB)T is the maximal sectorial relation corresponding to the form via Theorem 1.1. Theorem 2.1. Let T be a closed linear operator from a Hilbert space H to a Hilbert space K and let the operator B ∈ B(K) be selfadjoint. Then the form t in H defined by (2.1) t[h, k] = ((I + iB)T h, T k), h, k ∈ dom t = dom T, is closed and sectorial with vertex at the origin and semi-angle α ≤ arctan kBk and the maximal sectorial relation H corresponding to the form t is given by (2.2) H = T ∗(I + iB)T, with mul H = mul T ∗ = (dom T )⊥. A subset of dom t = dom T is a core of the form t if and only if it is a core of the operator T . Moreover, the nonnegative selfadjoint relation Hr corresponding to the real part (tH )r of the form t is given by (2.3) Hr = T ∗T. Proof. It is straightforward to check that t in (2.1) is a closed sectorial form as indicated, since tr[h, k] = (T h, T k), ti[h, k] = (BT h, T k). Therefore, ti[h] = (BT h, T h) ≤ kBkkT hk2 = kBktr[h], so that t is closed and sectorial with vertex at the origin and semi-angle α ≤ arctan kBk. Moreover, since T is closed, it is clear that tr and hence t is closed. Now let {h, h′} ∈ T ∗(I + iB)T , then there exists ϕ ∈ K such that {h, ϕ} ∈ T, {(I + iB)ϕ, h′} ∈ T ∗, from which it follows that Consequently, one sees that (h′, h) = (ϕ, ϕ) + i(Bϕ, ϕ). Im (h′h) = (Bϕ, ϕ) ≤ kBk kϕk2 = kBk Re (h′, h), which implies that T ∗(I + iB)T is a sectorial relation with vertex at the origin and semi-angle α ≤ arctan kBk. Furthermore, observe that the above calculation also shows that mul T ∗(I + iB)T = mul T ∗. To see that T ∗(I +iB)T is closed, let {hn, h′ n} ∈ T ∗(I +iB)T converge to {h, h′}. Then there exist ϕn ∈ K such that {hn, ϕn} ∈ T, n, hn) = kϕnk2 shows that (ϕn) is a Cauchy sequence in K, {(I + iB)ϕn, h′ n} ∈ T ∗, and the identity Re (h′ so that ϕn → ϕ with ϕ ∈ K. Thus {hn, ϕn} → {h, ϕ}, {(I + iB)ϕn, h′ n} → {(I + iB)ϕ, h′}. 4 S. HASSI AND H.S.V. DE SNOO Since T and T ∗ are closed, one concludes that {h, ϕ} ∈ T and {(I + iB)ϕ, h′} ∈ T ∗, which implies that {h, h′} ∈ T ∗(I + iB)T . Hence T ∗(I + iB)T is closed. Now let H be the maximal sectorial relation corresponding to t in (2.1). Assume that {h, h′} ∈ H, then for all k ∈ dom t = dom T t[h, k] = (h′, k) or ((I + iB)T h, T k) = (h′, k), which implies that {(I + iB)T h, h′} ∈ T ∗ or {h, h′} ∈ T ∗(I + iB)T. Consequently, it follows that H ⊂ T ∗(I + iB)T . Since T ∗(I + iB)T is sectorial and H is maximal sectorial, it follows that H = T ∗(I + iB)T . In particular, one sees that the closed relation T ∗(I + iB)T is maximal sectorial. (cid:3) With the closed linear operator T from H to K and the selfadjoint operator B ∈ B(K), consider the following matrix decomposition of B (2.4) B∗ B =(cid:18)Baa Bab ba Bbb(cid:19) : (cid:18)ker T ∗ ran T (cid:19) →(cid:18)ker T ∗ ran T (cid:19) . Then it is clear that (2.5) t[h, k] = ((I + iB)T h, T k) = ((I + iBbb)T h, T k), h, h ∈ dom t = dom T, which shows that only the compression of B to ran T plays a role in (2.1). In applications involving Theorem 2.1, it is therefore useful to recall the following corollary. Corollary 2.2. Let T ′ be a closed linear operator from the Hilbert space H to a Hilbert space K′ and let the operator B′ ∈ B(K′) be selfadjoint. Assume that the form t in Theorem 2.1 is also given by t[h, k] = ((I + iB′)T ′h, T ′k), h, k ∈ dom t = dom T ′. Then there is a unitary mapping U from ran T onto ran T ′, such that T ′ = U T, B′ bb = U BbbU ∗, where Bbb and B′ respectively. bb stand for the compressions of B and B′ to ran T and ran T ′, Proof. By assumption ((I + iB′)T ′h, T ′k) = ((I + iB)T h, T k) for all h, k ∈ dom t. This leads to (T ′h, T ′k) = (T h, T k) and (B′T ′h, T ′k) = (BT h, T k) for all h, k ∈ dom t. Hence the mapping T h 7→ T ′h is unitary, and denote it by U . Then T ′ = U T and it follows that (B′T ′h, T ′k) = (BU ∗T ′h, U ∗T ′k). (cid:3) 3. A matrix decomposition for T ∗(I + iB)T Let T be a linear relation from H to K which is closed; observe that then the subspace mul T is closed. The adjoint T ∗ of T is the set of all {h, h′} ∈ K × H for which (h′, f ) = (h, f ′) for all {f, f ′} ∈ T. Hence, the definition of T ∗ depends on the Hilbert spaces H and K in which T is assumed to act. Let K have the orthogonal decomposition (3.1) K = dom T ∗ ⊕ mul T, A CLASS OF SECTORIAL RELATIONS AND THE ASSOCIATED CLOSED FORMS 5 and let P be the orthogonal projection onto dom T ∗. Observe that P T ⊂ T , since {0} × mul T ⊂ T . Therefore T ∗ ⊂ (P T )∗ = T ∗P , where the last equality holds since P ∈ B(K). Then one has (3.2) (P T )∗ = T ∗ b⊕ (mul T × {0}). The orthogonal operator part Ts of T is defined as Ts = P T . Hence Ts is an operator from the Hilbert space H to the Hilbert space K and Ts ⊂ T . Note that ran Ts ⊂ dom T ∗ = K ⊖ mul T . Thus one may interpret Ts as an operator from the Hilbert space H to the Hilbert space dom T ∗ and one may also consider the adjoint (Ts)× of Ts with respect to these spaces. It is not difficult to see the connection between these adjoints: if {h, h′} ∈ K × H, then (3.3) {h, h′} ∈ T ∗ ⇔ {h, h′} ∈ (Ts)×. The identity (3.2) shows the difference between (Ts)∗ and (Ts)×. Let T be a closed linear relation from a Hilbert space H to a Hilbert space K and let B ∈ B(K) be selfadjoint. In order to study the linear relation T ∗(I + iB)T, decompose the Hilbert space K as in (3.1) and decompose the selfadjoint operator B ∈ B(K) accordingly: (3.4) B =(cid:18)B11 B12 12 B22(cid:19) : (cid:18)dom T ∗ mul T (cid:19) →(cid:18)dom T ∗ mul T (cid:19) . B∗ Here the operators B11 ∈ B(dom T ∗) and B22 ∈ B(mul T ) are selfadjoint, while B12 ∈ B(mul T, dom T ∗) and B∗ 12 ∈ B(dom T ∗, mul T ). By means of the decomposition (3.4) the following auxiliary operators will be introduced. First, define the operator C0 ∈ B(dom T ∗) by (3.5) C0 = I + B12(I + B2 22)−1B∗ 12. Observe that C0 ≥ I and that (C0)−1 belongs to B(dom T ∗) and is a nonnegative operator. Next, define the operator C ∈ B(dom T ∗) by (3.6) − 1 C = C 0 2 22)− 1 2 B22(I + B2 22)− 1 2 B∗ hB11 − B12(I + B2 12i C − 1 0 2 , which is clearly selfadjoint. Lemma 3.1. Let T be a closed linear relation from a Hilbert space H to a Hilbert space K, let Ts be the orthogonal operator part of T , and let the selfadjoint operator B ∈ B(K) be decomposed as in (3.4). Let the operators C0 and C be defined by (3.5) and (3.6). Then (3.7) T ∗(I + iB)T = (Ts)×C 1/2 0 (I + iC)C 1/2 0 Ts, and, consequently, T ∗(I + iB)T is maximal sectorial and (3.8) mul T ∗(I + iB)T = mul T ∗ = mul (Ts)×. Proof. In order to prove the equality in (3.7), assume that {h, h′} ∈ T ∗(I + iB)T . This means that (3.9) {h, ϕ} ∈ T and {(I + iB)ϕ, h′} ∈ T ∗ for some ϕ ∈ K. Decompose the element ϕ as (3.10) ϕ = ϕ1 + ϕ2, ϕ1 ∈ dom T ∗, ϕ2 ∈ mul T. 6 S. HASSI AND H.S.V. DE SNOO Since {0, ϕ2} ∈ T , it is clear that (3.11) {h, ϕ} ∈ T ⇔ {h, ϕ1} ∈ Ts. Using (3.10) and the above decomposition (3.4) of B, one observes that {(I + iB)ϕ, h′} =(cid:26)(cid:18)(I + iB11)ϕ1 + iB12ϕ2 12ϕ1 + (I + iB22)ϕ2(cid:19) , h′(cid:27) , iB∗ which implies that the condition {(I + iB)ϕ, h′} ∈ T ∗ is equivalent to (cid:26) {(I + iB11)ϕ1 + iB12ϕ2, h′} ∈ T ∗, 12ϕ1 + (I + iB22)ϕ2 = 0, iB∗ or, what is the same thing, (3.12) (cid:26) {[I + iB11 + B12(I + iB22)−1B∗ ϕ2 = −i(I + iB22)−1B∗ 12ϕ1. 12]ϕ1, h′} ∈ T ∗, Due to the definitions (3.5) and (3.6) and the identity (I + iB22)−1 = (I + B2 22)− 1 2 (I − iB22)(I + B2 22)− 1 2 , observe that I + iB11 + B12(I + iB22)−1B∗ 12 = C0 + i[B11 − B12(I + B2 22)− 1 2 B22(I + B2 22)− 1 2 B∗ 12] 1/2 = C 0 1/2 (I + iC)C 0 . Therefore, it follows from (3.12), via the equivalence in (3.3), that (3.13) {(I + iB)ϕ, h′} ∈ T ∗ ⇔ (cid:26) {C 1/2 0 (I + iC)C 1/2 0 ϕ1, h′} ∈ (Ts)×, ϕ2 = −i(I + iB22)−1B∗ 12ϕ1. Combining (3.11) and (3.13), one sees that {h, h′} ∈ (Ts)×C 1/2 0 (I + iC)C 1/2 0 Ts. Conversely, if this inclusion holds, then there exists ϕ1 ∈ dom T ∗, such that {h, ϕ1} ∈ Ts 1/2 and {C 0 (I + iC)C 1/2 0 ϕ1, h′} ∈ (Ts)×. Then define ϕ2 = −i(I + iB22)−1B∗ ϕ = ϕ1 + ϕ2. Hence {h, ϕ} ∈ T , and it follows from (3.13) that 12ϕ1, so that ϕ2 ∈ mul T . Furthermore, define Therefore one can rewrite T ∗(I + iB)T in the form (3.7). 1 {h, h′} ∈ T ∗(I + iB)T. Observe that C 2 0 Ts is a closed linear operator from the Hilbert space H to the Hilbert space dom T ∗ whose adjoint is given by (3.14) (C 1/2 0 Ts)× = (Ts)× C 1/2 0 . Hence, by Theorem 2.1 (Ts)×C H and by the identity (3.7) the same is true for T ∗(I + iB)T . (I + iC)C 1/2 0 1/2 0 Ts is a maximal sectorial relation in The statement in (3.8) follows by tracing the above equivalences for an element (cid:3) {0, h′}. A CLASS OF SECTORIAL RELATIONS AND THE ASSOCIATED CLOSED FORMS 7 Remark 3.2. Let ϕ = ϕ1 + ϕ2 ∈ K be decomposed as in (3.10). Then one has the following equivalence: (I + iB)ϕ ∈ dom T ∗ ⇔ (I + iB)ϕ = C 1/2 0 (I + iC)C 1/2 0 ϕ1. To see this, let η = (I + iB)ϕ. Then η ∈ dom T ∗ if and only if (cid:18)I + iB11 iB∗ 12 iB12 I + iB22(cid:19)(cid:18)ϕ1 ϕ2(cid:19) =(cid:18)η 0(cid:19) , where dom T ∗ is interpreted as the subspace dom T ∗ × {0} of K. Now apply (3.13). 4. A class of maximal sectorial relations and associated forms The linear relation T ∗(I +iB)T is maximal sectorial for any selfadjoint B ∈ B(K) and any closed linear relation T from H to K. Now the corresponding closed sectorial form will be determined. This gives the appropriate version of Theorem 2.1 in terms of relations. In fact, the general result is based on a reduction via Lemma 3.1 to Theorem 2.1. Theorem 4.1. Let T be a closed linear relation from a Hilbert space H to a Hilbert space K and let the selfadjoint operator B ∈ B(K) be decomposed as in (3.4). Let the operators C0 and C be defined by (3.5) and (3.6). Then the form t defined by 1 1 (4.1) t[h, k] = ((I + iC) C 2 0 Ts h, C 2 0 Ts k), h, k ∈ dom t = dom T, is closed and sectorial with vertex at the origin and semi-angle γ ≤ arctan kCk. Moreover, the maximal sectorial relation H corresponding to the form t is given by (4.2) H = (Ts)× C 1/2 0 (I + iC)C 1/2 0 Ts = T ∗(I + iB)T. A subset of dom t = dom T is a core of the form t if and only if it is a core of the operator Ts. Moreover, the nonnegative selfadjoint relation Hr corresponding to the real part (tH)r of the form t is given by Hr = (Ts)×C0Ts. 1 2 Proof. Since C 0 Ts is a closed linear operator from the Hilbert space H to the Hilbert 1/2 space dom T ∗, Theorem 2.1 (with K replaced by dom T ∗, B by C, and T by C 0 Ts) shows that the form t in (4.1) is closed and sectorial with vertex at the origin and semi-angle γ ≤ arctan kCk. Moreover, the maximal sectorial relation associated with the form t is given by (C 1/2 0 Ts)×(I + iC)C 1/2 0 Ts = (Ts)×C 1/2 0 (I + iC)C 1/2 0 Ts, cf. (2.1), (2.2), and (3.14). The identities in (4.2) are clear from Lemma 3.1. The assertion concerning the core holds, since the factor C0 is bounded with bounded inverse. The formula (4.2) shows that 1 1 (tH )r[h, k] = (C and hence Hr = (C 1/2 0 Ts)×C 2 2 0 Tsk), 0 Tsh, C 1/2 0 Ts = (Ts)×C0Ts (cf. the discussion above). h, k ∈ dom t = dom T, (cid:3) Recall that if {h, h′} ∈ T ∗(I + iB)T , then {h, ϕ} ∈ T and {(I + iB)ϕ, h′} ∈ T ∗. The last inclusion implies the condition (I + iB)ϕ ∈ dom T ∗ ⊂ dom T ∗, giving rise to ϕ2 = −i(I + iB22)−1B∗ 12ϕ1. Thus, for instance, when B = diag (B11, B22), it follows that ϕ2 = 0, so that it is immediately clear that γ ≤ arctan kB11k, independent of B22. Note that the following assertions are equivalent: 8 S. HASSI AND H.S.V. DE SNOO (i) B = diag (B11, B22); (ii) B12 = 0; (iii) C0 = I; (iv) mul T is invariant under B, in which case C = B11. Hence, if mul T is invariant under B, i.e., if any of the assertions (i) -- (iv) hold, then Theorem 4.1 gives the following corollary, which co- incides with [4, Theorem 5.1]. In the case where mul T = {0} the corollary reduces to Theorem 2.1. Corollary 4.2. Let T be a closed linear relation from a Hilbert space H to a Hilbert space K, let Ts be the orthogonal operator part of T , and let mul T be invariant under the selfadjoint operator B ∈ B(K), so that B = diag (B11, B22). Then the form t defined by t[h, k] = ((I + iB11)Tsh, Tsk), h, k ∈ dom t = dom T, is closed and sectorial with vertex at the origin and semi-angle γ ≤ arctan kB11k. Moreover, the maximal sectorial relation H corresponding to the form t is given by H = (Ts)×(I + iB11)Ts = T ∗(I + iB)T. In the case that mul T is not invariant under B, one has C0 6= I, and the formulas are different: for instance, the real part (tH )r in Theorem 4.1 is of the form 1 1 (tH )r[h, k] = (C 2 0 Tsh, C 2 0 Tsk), h, k ∈ dom t = dom Ts = dom T. Example 4.3. Assume that B11 6= 0 and B11 = B12(I + B2 22)− 1 2 B22(I + B2 22)− 1 2 B∗ 12, so that C = 0. In this case the maximal sectorial relation H = T ∗(I + iB)T in Theorem 4.1 is selfadjoint, i.e., H = Hr and the associated form t is nonnegative. On the other hand, with such a choice of B the operator part of T determines the maximal sectorial relation (Ts)∗(I + iB)Ts with semi-angle arctan kB11k > 0, while T ∗(I + iB)T has semi-angle γ = 0. 5. Maximal sectorial relations and their representations Let H be a maximal sectorial relation in H and let the closed sectorial form tH correspond to H; cf. Theorem 1.1. Since the closed form tH is sectorial, one has the inequality (5.1) (tH )i[h] ≤ (tan α)(tH )r[h], h ∈ dom t, and in this situation the real part (tH )r is a closed nonnegative form. Hence by the first representation theorem there exists a nonnegative selfadjoint relation Hr, the so-called real part of H, such that dom Hr ⊂ dom (tH )r = dom tH and (tH )r[h, k] = (h′, k), {h, h′} ∈ Hr, k ∈ dom (tH )r = dom tH. This real part Hr, not to be confused with the real part introduced in [6], will play an important role in formulating the second representation theorem below. First the case where H is a maximal sectorial operator will be considered, in which case H is automatically densely defined; see [7]. A CLASS OF SECTORIAL RELATIONS AND THE ASSOCIATED CLOSED FORMS 9 Lemma 5.1. Let H be an maximal sectorial operator in H, let the closed sectorial form tH correspond to H via Theorem 1.1, and let Hr be the real part of H. Then there exists a unique selfadjoint operator G ∈ B(H) with kGk = tan α, of the form G =(cid:18)0 0 Gbb(cid:19) : (cid:18)ker Hr ran Hr(cid:19) →(cid:18)ker Hr ran Hr(cid:19) , 0 (5.2) such that (5.3) tH[h, k] = ((I + iG)(Hr) 1 2 h, (Hr) 1 2 k), h, k ∈ dom tH = dom H 2 r . 1 Moreover, the corresponding maximal sectorial operator H is given by H = (Hr) 1 2 (I + iG)(Hr) 1 2 , with mul H = mul Hr. Proof. The inequality (tH )i[h, k]2 ≤ Ctr[h]tr[k] = CkH 1 1 2 r hkkH 2 r kk, h, k ∈ dom , shows the existence of a selfadjoint operator G in H ⊖ ker H such that (5.4) (tH )i[h, k] = (G(Hr) 1 2 h, (Hr) 1 2 k), h, k ∈ dom (Hr) 1 2 . Extend G to all of H in a trivial way, so that the same formula remains valid; see Corollary 2.2. It follows from the decomposition t = tr + iti, cf. (1.3), and the identities (5.5) and (5.4), that so that 1 1 1 1 tH = (tH )r + i(tH )i, tH = [h, k] = ((Hr) 2 s h, (Hr) 2 s k) + i(G(Hr) 2 s h, (Hr) 2 s k). This last identity immediately gives (5.3). The rest follows from Corollary 4.2. (cid:3) Now let H be a maximal sectorial relation, let Hr be its real part, and let (Hr)s be its orthogonal operator part. Then one obtains the representation (5.5) (tH )r[h, k] = (((Hr)) 2 s h, ((Hr)s) 1 1 2 k), h, k ∈ dom (tH )r = dom ((Hr)s) 1 2 , cf. Theorem 2.1. Now apply Corollary 4.2 and therefore one may formulate the second representation theorem as follows. Theorem 5.2. Let H be a maximal sectorial relation in H, let the closed sectorial form tH correspond to H via Theorem 1.1, and let Hr be the real part of H. Then there exists a selfadjoint operator G ∈ B(H) with kGk = tan α, such that G is trivial on ker Hr ⊕ mul Hr, and (5.6) tH [h, k] = ((I + iG)((Hr)s) 1 2 h, ((Hr)s) 1 2 k), h, k ∈ dom tH = dom H 2 r . 1 Moreover, the maximal sectorial relation H is given by (5.7) H = (((Hr)s) 1 2 )×(I + iG)((Hr)s) 1 2 , with mul H = mul Hr. 10 S. HASSI AND H.S.V. DE SNOO Next, it is assumed that H is a maximal sectorial relation of the form H = T ∗(I + iB)T , where T is a closed linear relation from a Hilbert space H to a Hilbert space K and the operator B ∈ B(K) is selfadjoint. Let the operators C0 and C be defined by (3.5) and (3.6), then H = (Ts)× C 1/2 0 (I + iC)C 1/2 0 Ts, while the corresponding closed sectorial form is given 1 1 t[h, k] = ((I + iC) C 2 0 Ts h, C 2 0 Ts k), h, k ∈ dom t = dom T. To compare these expressions with (5.6) and (5.7), observe that 1 1 ( C 2 0 Ts h, C 2 0 Ts k) = ((Hr)s) 1 2 h, ((Hr)s) 1 2 k) and 1 1 (C C 2 0 Ts h, C 2 0 Ts k) = (G((Hr)s) 1 2 h, ((Hr)s) 1 2 k). 1/2 It is clear from (4.1) that only the (selfadjoint) compression of C to ran C 0 Ts contributes to the form (4.2), so that it is straightforward to set up a unitary mapping; cf. Corollary 2.2. 6. Extremal maximal sectorial extensions of sums of maximal sectorial relations Let H1 and H2 be maximal sectorial relations in a Hilbert space H. Then the sum H1 + H2 is a sectorial relation in H with dom (H1 + H2) = dom H1 ∩ dom H2, so that the sum is not necessarily densely defined. In particular, H1 + H2 and its closure need not be operators, since (6.1) mul (H1 + H2) = mul H1 + mul H2. To describe the class of extremal maximal sectorial extensions of H1 + H2 some basic notations are recalled from [5], together with the description of the Friedrichs and Kreın extensions (H1 + H2)F and (H1 + H2)K of H1 + H2, respectively. In order to describe the whole class of extremal extensions of H1 + H2 and the corresponding closed forms a proper description of the closed sectorial form tK is essential. The results in Sections 3 and 4 allow a general treatment that will relax the additional conditions in [5]. 6.1. Basic notions. Let H1 and H2 be maximal sectorial relations and decompose them as follows 1 1 (6.2) Hj = A 2 j (I + iBj)A 2 j , 1 ≤ j ≤ 2, where Aj (the real part of Hj), 1 ≤ j ≤ 2, are nonnegative selfadjoint relations in H and Bj, 1 ≤ j ≤ 2, are bounded selfadjoint operators in H which are trivial on ker Aj ⊕ mul Aj; cf. Theorem 5.2. Furthermore, if A1 and A2 are decomposed as Aj = Ajs ⊕ Aj∞, 1 ≤ j ≤ 2, A CLASS OF SECTORIAL RELATIONS AND THE ASSOCIATED CLOSED FORMS 11 where Aj∞ = {0} × mul Aj , 1 ≤ j ≤ 2, and Ajs, 1 ≤ j ≤ 2, are densely defined nonnegative selfadjoint operators (defined as orthogonal complements in the graph sense), then the uniquely determined square roots of Aj, 1 ≤ j ≤ 2 are given by 1 1 A 2 j = A 2 js ⊕ Aj∞, 1 ≤ j ≤ 2. Associated with H1 and H2 is the relation Φ from H × H to H, defined by (6.3) Φ =n {{f1, f2}, f ′ 1 + f ′ 2} : {fj, f ′ j} ∈ A 2 1 j , 1 ≤ j ≤ 2o . Clearly, Φ is a relation whose domain and multivalued part are given by dom Φ = dom A 2 1 × dom A 2 2 , mul Φ = mul H1 + mul H2. 1 1 The relation Φ is not necessarily densely defined in H × H, so that in general Φ∗ is a relation as mul Φ∗ = (dom Φ)⊥. Furthermore, the adjoint Φ∗ of Φ is the relation from H to H × H, given by Φ∗ =n{h, {h′ 1, h′ 2}} : {h, h′ j} ∈ A 2 1 j , 1 ≤ j ≤ 2o . The identity (6.4) shows that the (orthogonal) operator part (Φ∗)s of Φ∗ is given by: (6.4) (6.5) (6.6) (6.7) (6.8) (Φ∗)s = n{h, {h′ = nnh, {A 1, h′ 2}} : {h, h′ j} ∈ A 1 2 1sh, A 2 1 2sh}o : h ∈ dom A 2 1 js, 1 ≤ j ≤ 2o 1 2 1 ∩ dom A 2 1 2o . The identities (6.4) and (6.5) show that dom Φ∗ = dom A 2 1 ∩ dom A 2 2 , mul Φ∗ = mul H1 × mul H2, ran (Φ∗)s = F0, 1 1 where the subspace F0 ⊂ H × H is defined by F0 =nnA 1 2 1sh, A 2 1 2sho : h ∈ dom A 1 2 1 ∩ dom A 2 1 2o . The closure of F0 in H × H will be denoted by F. Define the relation Ψ from H to H × H by Ψ =nnh,nA 1 2 1sh, A 2 1 2shoo : h ∈ dom H1 ∩ dom H2o ⊂ H × (H × H). It follows from this definition that dom Ψ = dom H1 ∩ dom H2, mul Ψ = {0}, ran Ψ = E0, where the space E0 ⊂ H × H is defined by E0 =nnA 1 2 1sf, A 2 1 2sfo : f ∈ dom H1 ∩ dom H2o . Observe that E0 ⊂ F0. The closure of E0 in H × H will be denoted by E. Hence, (6.9) E ⊂ F. Comparison of (6.5) and (6.7) shows (6.10) Ψ ⊂ (Φ∗)s, and thus the operator Ψ is closable and Ψ∗∗ ⊂ (Φ∗)s. It follows from dom Ψ∗ = (mul Ψ∗∗)⊥ and mul Ψ∗ = (dom Ψ)⊥, that dom Ψ∗ = H, mul Ψ∗ = (dom H1 ∩ dom H2)⊥. 12 S. HASSI AND H.S.V. DE SNOO Next, define the relation K from H × H to H by (6.11) K = (cid:8){{(I + iB1)A 1 2 1sf, (I + iB2)A 2 2sf }, f ′ 1 + f ′ 2} : 1 1 {(I + iB1)A 1 2 1sf, f ′ 1} ∈ A 2 1 , {(I + iB2)A 2 2sf, f ′ 2} ∈ A 1 2 1 2(cid:9) ⊂ (H × H) × H. Clearly, the domain and multivalued part of K are given by dom K = D0, mul K = mul (H1 + H2), where (6.12) D0 =n {(I + iB1)A 1/2 1s f, (I + iB2)A 1/2 2s f } : f ∈ dom H1 ∩ dom H2o . The closure of D0 in H × H will be denoted by D. Lemma 6.1. The relations K, Φ, and Ψ satisfy the following inclusions: (6.13) K ⊂ Φ ⊂ Ψ∗, Ψ ⊂ Φ∗ ⊂ K ∗. Proof. To see this note that K ⊂ Φ follows from (6.3) and (6.11), and that Ψ ⊂ Φ∗ follows from (6.4) and (6.7). Therefore, also Φ∗ ⊂ K ∗ and Φ ⊂ Φ∗∗ ⊂ Ψ∗. (cid:3) 6.2. The Friedrichs and the Kreın extensions of H1 + H2. The descriptions of the Friedrichs extension and the Kreın extension (H1 + H2)F and (H1 + H2)K of H1 + H2 are now recalled from [5]. For this, define the orthogonal sum of the operators B1 and B2 in H × H by B⊕ := B1 ⊕ B2 =(cid:18)B1 0 B2(cid:19) . 0 The descriptions of (H1 + H2)F and (H1 + H2)K incorporate the initial data on the factorizations (6.2) of H1 and H2 via the mappings Φ, Ψ, and K in Subsection 6.1. The construction of the Friedrichs extension was given in [5, Theorem 3.2], where some further details and a proof of the following result can be found. The new additions in the next theorem are the second representations for (H1 + H2)F and tF that will be needed in the rest of this paper. Theorem 6.2. Let H1 and H2 be maximal sectorial and let Ψ be defined by (6.7). Then the Friedrichs extension of H1 + H2 has the expression (6.14) (H1 + H2)F = Ψ∗(I + iB⊕)Ψ∗∗ = Ψ∗C 1/2 0 (I + iC)C 1/2 0 PD(Ψ∗∗)s. The closed sectorial form tF associated with (H1 + H2)F is given by (6.15) tF [f, g] = ((I + iB⊕)Ψ∗∗f, Ψ∗∗g) = (C 1/2 0 (I + iC)C 1/2 0 PD(Ψ∗∗)sf, PD(Ψ∗∗)sg), for all f, g ∈ dom tF = dom Ψ∗∗. Proof. As indicated the first expressions for (H1 + H2)F in (6.14) and tF in (6.15) have been proved in [5, Theorem 3.2] and, hence, it suffices to derive the second expressions in (6.14) and (6.15). By definition, one has ran Ψ = E0 (see (6.7), (6.8)), and by Lemma 6.1 one has Ψ ⊂ Ψ∗∗ ⊂ K ∗, which after projection onto D = dom K yields PDΨ∗∗ ⊂ PDK ∗ = (K ∗)s. A CLASS OF SECTORIAL RELATIONS AND THE ASSOCIATED CLOSED FORMS 13 Notice that D0 = dom K = (I + iB⊕)E0 (see (6.8), (6.12)). Since the operator I + iB⊕ is bounded with bounded inverse, one has the equality (6.16) D = (I + iB⊕)E. It follows that the range of (I + iB⊕)Ψ∗∗ belongs to D = dom K. Now by Remark 3.2 this implies that for all f ∈ dom Ψ∗∗ one has the equality (6.17) This leads to (I + B⊕)(Ψ∗∗)sf = C 1/2 0 (I + iC)C 1/2 0 PD(Ψ∗∗)sf. Ψ∗(I + iB⊕)Ψ∗∗ = Ψ∗C 1/2 0 (I + iC)C 1/2 0 PD(Ψ∗∗)s, which proves (6.14). Similarly by substituting (6.17) into the first formula for tF and noting that PDC (cid:3) 1/2 0 = PD, one obtains the second formula in (6.15). Also the construction of the Kreın extension for the sum H1 + H2 can be found in [5, Theorem 3.2]. However, the corresponding form tK was described only under additional conditions to prevent the difficulty that appears by the fact that the multivalued part of (H1 + H2)K is in general not invariant under the mapping B⊕. Theorem 4.1 allows a removal of these additional conditions and leads to a description of the form tK in the general situation. For this purpose, decompose the Hilbert space H × H as follows (6.18) H × H = dom K ⊕ mul K ∗, and let P be the orthogonal projection onto dom K. Moreover, decompose the selfadjoint operator B⊕ ∈ B(H × H) accordingly: (6.19) B⊕ =(cid:18)B11 B12 12 B22(cid:19) : (cid:18)dom K mul K ∗(cid:19) →(cid:18) dom K mul K ∗(cid:19) . B∗ Next define the operator C0 ∈ B(dom K ∗) by (6.20) C0 = I + B12(I + B2 22)−1B∗ 12, and the operator C ∈ B(dom K ∗) by (6.21) − 1 C = C 0 2 which is clearly selfadjoint. hB11 − B12(I + B2 22)− 1 2 B22(I + B2 22)− 1 2 B∗ 12i C − 1 0 2 , Theorem 6.3. Let H1 and H2 be maximal sectorial relations in a Hilbert space H, let K be defined by (6.11), and let C0 and C be given by (6.20) and (6.21), respectively. Then the Kreın extension of H1 + H2 has the expression (H1 + H2)K = K ∗∗(I + iB⊕)K ∗ = ((K ∗)s)× C 1/2 0 1/2 (I + iC)C 0 (K ∗)s. The closed sectorial form tK associated with (H1 + H2)K is given by tK[f, g] = ((I + iC)C 1/2 0 (K ∗)sf, C 1/2 0 (K ∗)sg), f, g ∈ dom tK = dom K ∗. Proof. The first equality in the first statement is proved in [5, Theorem 3.2]. The second equality is obtained by applying Theorem 4.1 to the sectorial relation K ∗∗(I + iB⊕)K ∗. The statement concerning the form tK is a consequence of this second represen- (K ∗)s is a closed operator and hence one can apply (cid:3) tation of (H1 + H2)K , since C Theorem 2.1 to get the desired expression for the corresponding form tK. 1/2 0 14 S. HASSI AND H.S.V. DE SNOO The form tK described in Theorem 6.3 can be used to give a complete descrip- tion of all extremal maximal sectorial extensions of the sum H1 + H2. Namely, a maximal sectorial extension eH of a sectorial relation S is extremal precisely when the corresponding closed sectorial form t eH is a restriction of the closed sectorial form tK generated by the Kreın extension SK of S; see e.g. [4, Definition 7.7, Theorems 8.2, 8.4, 8.5]. Therefore, Theorem 6.3 implies the following description of all extremal maximal sectorial extensions of H1 + H2. Theorem 6.4. Let H1 and H2 be maximal sectorial relations in H, let Ψ and K be defined by (6.7) and (6.11), respectively, and let PD be the orthogonal projection from H × H onto D = dom K. Then the following statements are equivalent: (i) eH is an extremal maximal sectorial extension of H1 + H2; (ii) eH = R∗(I + iC)R, where R is a closed linear operator satisfying 1/2 0 PDΨ∗∗ ⊂ R ⊂ C (K ∗)s. 1/2 0 C Proof. For comparison with the abstract results this statement will be proved by means of the constructions used in [4]. Let S = H1 + H2 then the sectorial relation S gives rise to a Hilbert space HS and a selfadjoint operator BS ∈ B(HS) such that the Friedrichs extension SF and the Kreın extension SK of S are given by SF = Q∗(I + iBS)Q∗∗, tF = J ∗∗(I + iBS)J ∗, with corresponding forms tF [f, g] = ((I + iBS)Q∗∗f, Q∗∗g), f, g ∈ dom Q∗∗, and tK[f, g] = ((I + iBS)J ∗f, J ∗g), f, g ∈ dom J ∗; see [4, Theorem 8.3]. Here Q : H → HS is an operator and J : HS → H is a densely defined linear relation such that J ⊂ Q∗, Q ⊂ J ∗; in particular, the adjoint J ∗ is an operator. Recall from Theorem 6.3 that tK[f, g] = ((I + iC)C 1/2 0 (K ∗)sf, C 1/2 0 (K ∗)sg), while Theorem 6.2 gives tF [f, g] = ((I + iB⊕)Ψ∗∗f, Ψ∗∗g) = (C 1/2 0 (I + iC)C 1/2 0 PD(Ψ∗∗)sf, PD(Ψ∗∗)sg). Now apply [4, Theorem 8.4] and Corollary 2.2. (cid:3) 6.3. The form sum construction. The maximal sectorial relations H1 and H2 generate the following closed sectorial form 1 1 1 1 1 1 (6.22) ((I + iB1)A 2 1sh, A 2 1sk) + ((I + iB2)A 2 2sh, A 2 2sk), h, k ∈ dom A 2 1 ∩ dom A 2 2 . Observe that the restriction of this form to dom Ψ∗∗ is equal to (6.23) (Ψ∗∗h, Ψ∗∗k) = ((I + iB1)A 1sk) + ((I + iB2)A 1sh, A 2sh, A 2 2 2sk), 2 2 1 1 1 1 h, k ∈ dom Ψ∗∗, since Ψ∗∗ ⊂ (Φ∗)s, cf. (6.5). Thus, the form in (6.22) has a natural domain which is in general larger than dom Ψ∗∗. A CLASS OF SECTORIAL RELATIONS AND THE ASSOCIATED CLOSED FORMS 15 Theorem 6.5. Let H1 and H2 be maximal sectorial relations in H, let Φ be given by (6.3), and let E = clos E0 and F = clos F0 be defined by (6.8) amd (6.6). Then the maximal sectorial relation Φ∗∗(I + iB⊕)Φ∗ is an extension of the relation H1 + H2, which corresponds to the closed sectorial form in (6.22). Moreover, the following statements are equivalent: (i) Φ∗∗(I + iB⊕)Φ∗ is extremal; (ii) E = F. Proof. The first statement is proved in [5, Theorem 3.5]. For the proof of the equivalence of (i) and (ii) appropriate modifications are needed in the arguments used in the proof of [5, Theorem 3.5]. The special case treated there was based on the additional assumption that D = E, where D = dom K; a condition which implies the invariance of mul K ∗ under the operator B⊕. In the present general case such an invariance property cannot be assumed. Now for simplicity denote the the inclusion F ⊂ E. By Theorem 6.4 and mul Φ∗ = mul H1 × mul H2 one sees form sum extension of H1 + H2 briefly by bH = Φ∗∗(I + iB⊕)Φ∗. (i) ⇒ (ii) Assume that bH is extremal. Since E ⊂ F by (6.9), it is enough to prove bH = ((Φ∗)s)∗(I + iB⊕)(Φ∗)s = R∗(I + iC)R, for some closed operator R satisfying (6.24) (6.25) 1/2 0 PDΨ∗∗ ⊂ R ⊂ C 1/2 0 C (K ∗)s, where PD is the orthogonal projection of H × H onto D = dom K. Recall that (Φ∗)s ⊂ Φ∗ ⊂ K ∗ and hence PD(Φ∗)s ⊂ PDK ∗ = (K ∗)s. Moreover, one has dom PD(Φ∗)s = dom (Φ∗)s = dom R, since by assumption these two domains co- R, one has incide with the corresponding joint form domain. Denoting bR = C −1/2 dom bR = dom PD(Φ∗)s and (6.24) can be rewritten as bR, where bR satisfies PDΨ∗∗ ⊂ bR ⊂ (K ∗)s. One concludes that PD(Φ∗)s = bR, since bH = ((Φ∗)s)∗(I + iB⊕)(Φ∗)s = bR∗C both operators are restrictions of (K ∗)s, and thus 1/2 (I + iC)C 0 (6.26) 1/2 0 0 (6.27) Now one obtains from (6.26) the equalities ((Φ∗)s)∗PD = bR∗. ((Φ∗)s)∗(I + iB⊕)(Φ∗)s = bR∗C = ((Φ∗)s)∗PDC 1/2 = ((Φ∗)s)∗C 0 Hence, for every f ∈ dom bH one has (I + iB⊕)(Φ∗)sf − C 1/2 0 1/2 (I + iC)C 0 1/2 0 1/2 (I + iC)C 0 bR bR 1/2 0 1/2 (I + iC)C 0 1/2 (I + iC)C 0 bR. bRf ∈ ker ((Φ∗)s)∗. 1/2 (I + iC)C 0 1/2 Here C 0 1/2 Therefore, there exists ϕ ∈ E such that C 0 bRf ∈ D = dom K and D = dom K = (I + iB⊕)E; see (6.16). bRf = (I + iB⊕)ϕ. On 1/2 (I + iC)C 0 16 S. HASSI AND H.S.V. DE SNOO the other hand, (Φ∗)sf ∈ F = ran (Φ∗)s = (ker ((Φ∗)s)∗)⊥, see (6.5), (6.6). Since ϕ ∈ E ⊂ F, this yields ((I + iB⊕)((Φ∗)sf − ϕ), (Φ∗)sf − ϕ) = 0, and thus (Φ∗)sf − ϕ = 0. Consequently, for all f ∈ dom bH one has Since dom bH is a core for the corresponding closed form, or equivalently, the closure of (Φ∗)s↾ dom bH is equal to (Φ∗)s, the claim follows: F = ran (Φ∗)s ⊂ E. (ii) ⇒ (i) Assume that E = F. Then F0 = ran (Φ∗)s ⊂ E and hence for all f ∈ dom (Φ∗)s one has (I + B⊕)(Φ∗)sf ∈ dom K. By Remark 3.2 this implies that (Φ∗)sf ∈ E. (6.28) (I + B⊕)(Φ∗)sf = C 1/2 0 (I + iC)C 1/2 0 PD(Φ∗)sf. On the other hand, as shown above PD(Φ∗)s ⊂ PDK ∗ = (K ∗)s. Let bR be the closure of (K ∗)s↾ dom (Φ∗)s. Then bR∗ satisfies the identity (6.27). Since Ψ∗∗ ⊂ (Φ∗)s (see (6.10)) one obtains PDΨ∗∗ ⊂ bR. The identities (6.27) and (6.28) imply that for all f ∈ dom bH the equalities ((Φ∗)s)∗(I + B⊕)(Φ∗)sf = ((Φ∗)s)∗PDC 1/2 0 PD(Φ∗)sf (I + iC)C 1/2 0 1/2 0 1/2 (I + iC)C 0 = bR∗C bRf 1/2 hold. Then the closed operator R = C 0 bR satisfies the inclusions (6.25) as well as the desired identity ((Φ∗)s)∗(I + B⊕)(Φ∗)s = R∗(I + iC)R, and thus bH is extremal, cf. Theorem 6.4. (cid:3) Theorem 6.5 is a generalization of [5, Theorem 3.5], where an additional invari- ance of mul K ∗ under the operator B⊕ was used. Moreover, Theorem 6.5 generalizes a corresponding result for the form sum of two closed nonnegative forms established earlier in [3, Theorem 4.1]. The present result relies on Theorem 4.1, where the description of the closed sectorial form generated by a general maximal sectorial relation of the form H = T ∗(I + iB)T where T is a closed relation. This generality implies that with special choices of B the relation H can be taken to be nonnegative and selfadjoint, i.e., the corresponding closed form t becomes nonnegative; see Example 4.3. References [1] Yu.M. Arlinskiı, "Maximal sectorial extensions and closed forms associated with them", Ukrainian Math. J., 48 (1996), 723 -- 739. [2] M. Haase, The functional calculus for sectorial operators, Oper. Theory Adv. Appl., 169, Birkhauser Verlag, 2006. [3] S. Hassi, A. Sandovici, H.S.V. de Snoo, and H. Winkler, "Extremal extensions for the sum of nonnegative selfadjoint relations", Proc. Amer. Math. Soc., 135 (2007), 3193 -- 3204. [4] S. Hassi, A. Sandovici, H.S.V. de Snoo, and H. Winkler, "Extremal maximal sectorial exten- sions of sectorial relations", Indag. Math., 28 (2017), 1019 -- 1055. [5] S. Hassi, A. Sandovici, and H.S.V. de Snoo, "Factorized sectorial relations, their maximal sectorial extensions, and form sums", Banach J. Math. Anal., 13 (2019), 538 -- 564. [6] S. Hassi, H.S.V. de Snoo, and F.H. Szafraniec, "Componentwise and Cartesian decomposi- tions of linear relations", Dissertationes Mathematicae, 465 (2009), 59 pages. [7] T. Kato, Perturbation theory for linear operators, Springer-Verlag, Berlin, 1980. [8] F.S. Rofe-Beketov, "Numerical range of a linear relation and maximal relations", Teor. Funkt- sii Funktsional. Anal. i Prilozhen, 44 (1985), 103 -- 112 (Russian) [English translation: J. Math. Sciences, 48 (1990), 329 -- 336]. A CLASS OF SECTORIAL RELATIONS AND THE ASSOCIATED CLOSED FORMS 17 [9] A. Sandovici, Contributions to the extension theory in Hilbert spaces, Doctoral dissertation, University of Groningen, 2006. [10] Z. Sebesty´en and Zs. Tarcsay, "T ∗T always has a positive selfadjoint extension", Acta Math. Hungar., 135 (2012), 116 -- 129. Department of Mathematics and Statistics, University of Vaasa, P.O. Box 700, 65101 Vaasa, Finland E-mail address: [email protected] Bernoulli Institute for Mathematics, Computer Science and Artificial Intelligence, University of Groningen, P.O. Box 407, 9700 AK Groningen, Nederland E-mail address: [email protected]

Subsets and Splits